Implement syncOnAddress ABIs (perf improvement)

Co-authored-by: DeathWhispers <56944463+Death-Whispers-U@users.noreply.github.com>

.github/workflows/build.yml

@@ -94,6 +94,7 @@ jobs:
           cmake -S . -B _Build/windows ^
             -G Ninja ^
             -DCMAKE_BUILD_TYPE=Release ^
+            -DCMAKE_INTERPROCEDURAL_OPTIMIZATION=ON ^
             -DKYTY_BUILD_ORIGIN=%KYTY_BUILD_ORIGIN% ^
             -DKYTY_BUILD_REPOSITORY=%KYTY_BUILD_REPOSITORY% ^
             -DKYTY_RELEASE_TAG=${{ needs.metadata.outputs.release_tag }} ^
@@ -172,6 +173,7 @@ jobs:
           cmake -S . -B _Build/macos \
             -G Ninja \
             -DCMAKE_BUILD_TYPE=Release \
+            -DCMAKE_INTERPROCEDURAL_OPTIMIZATION=ON \
             -DKYTY_BUILD_ORIGIN="$KYTY_BUILD_ORIGIN" \
             -DKYTY_BUILD_REPOSITORY="$KYTY_BUILD_REPOSITORY" \
             -DKYTY_RELEASE_TAG="${{ needs.metadata.outputs.release_tag }}" \
@@ -299,6 +301,7 @@ jobs:
           cmake -S . -B _Build/linux \
             -G Ninja \
             -DCMAKE_BUILD_TYPE=Release \
+            -DCMAKE_INTERPROCEDURAL_OPTIMIZATION=ON \
             -DKYTY_BUILD_ORIGIN="$KYTY_BUILD_ORIGIN" \
             -DKYTY_BUILD_REPOSITORY="$KYTY_BUILD_REPOSITORY" \
             -DKYTY_RELEASE_TAG="${{ needs.metadata.outputs.release_tag }}" \

CMakeLists.txt

@@ -4,6 +4,12 @@ if(CMAKE_VERSION VERSION_GREATER_EQUAL 4.0)
 	set(CMAKE_POLICY_VERSION_MINIMUM 3.5 CACHE STRING "Minimum policy version for bundled dependencies")
 endif()
 
+get_property(KYTY_GENERATOR_IS_MULTI_CONFIG GLOBAL PROPERTY GENERATOR_IS_MULTI_CONFIG)
+if(NOT KYTY_GENERATOR_IS_MULTI_CONFIG AND NOT CMAKE_BUILD_TYPE)
+	set(CMAKE_BUILD_TYPE Release CACHE STRING "Choose the type of build" FORCE)
+	set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS Debug Release RelWithDebInfo MinSizeRel)
+endif()
+
 project(Kyty)
 
 set(KYTY_SOURCE_DIR "${CMAKE_CURRENT_SOURCE_DIR}/src")
@@ -444,6 +450,13 @@ add_executable(event_queue_lifetime_tests EXCLUDE_FROM_ALL
 target_link_libraries(event_queue_lifetime_tests common fmt::fmt)
 target_include_directories(event_queue_lifetime_tests PRIVATE ${inc_headers})
 
+add_executable(sync_on_address_tests EXCLUDE_FROM_ALL
+	"${KYTY_TESTS_DIR}/SyncOnAddressTests.cpp"
+	"${KYTY_SOURCE_DIR}/kernel/syncOnAddress.cpp"
+)
+target_link_libraries(sync_on_address_tests common)
+target_include_directories(sync_on_address_tests PRIVATE ${inc_headers})
+
 add_executable(image_page_table_tests EXCLUDE_FROM_ALL
 	"${KYTY_TESTS_DIR}/ImagePageTableTests.cpp"
 )
@@ -497,6 +510,7 @@ if(BUILD_TESTING)
 	add_test(NAME resource_tracking COMMAND $<TARGET_FILE:resource_tracking_tests>)
 	add_test(NAME resource_mutex COMMAND $<TARGET_FILE:resource_mutex_tests>)
 	add_test(NAME event_queue_lifetime COMMAND $<TARGET_FILE:event_queue_lifetime_tests>)
+	add_test(NAME sync_on_address COMMAND $<TARGET_FILE:sync_on_address_tests>)
 	add_test(NAME audio_out2_port COMMAND $<TARGET_FILE:audio_out2_port_tests>)
 	add_test(NAME shader_recompiler_compute COMMAND $<TARGET_FILE:shader_recompiler_compute_tests>)
 	add_test(NAME virtual_memory_allocation
@@ -507,6 +521,8 @@ if(BUILD_TESTING)
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --stream-buffer-only)
 	add_test(NAME gpu_command_lane
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --gpu-command-lane-only)
+	add_test(NAME pm4_context_state
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --context-state-only)
 	add_test(NAME gpu_tiler
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --gpu-tiler-only)
 
@@ -543,6 +559,7 @@ if(BUILD_TESTING)
 		resource_tracking_tests
 		resource_mutex_tests
 		event_queue_lifetime_tests
+		sync_on_address_tests
 		shader_recompiler_compute_tests
 		virtual_memory_allocation_tests
 	)

src/common/emulatorConfig.cpp

@@ -101,4 +101,8 @@ bool ReadbackLinearImagesEnabled() {
 	return g_config->readback_linear_images;
 }
 
+const Keymap& GetKeymap() {
+	return g_config->keymap;
+}
+
 } // namespace Config

src/common/emulatorConfig.h

@@ -5,6 +5,8 @@
 #include "common/subsystems.h"
 
 #include <filesystem>
+#include <string>
+#include <vector>
 
 namespace Config {
 
@@ -18,6 +20,8 @@ enum class ProfilerDirection { None, Network };
 
 enum class OutputDirection { Silent, Console, File };
 
+using Keymap = std::vector<std::string>;
+
 constexpr uint32_t DEFAULT_CONSOLE_LANGUAGE = 1;
 constexpr uint32_t MAX_CONSOLE_LANGUAGE     = 29;
 
@@ -41,6 +45,7 @@ struct ConfigOptions {
 	bool                   spirv_debug_printf_enabled  = false;
 	bool                   renderdoc_enabled           = false;
 	bool                   readback_linear_images      = false;
+	Keymap                 keymap;
 };
 
 void Load(const ConfigOptions& cfg);
@@ -72,6 +77,8 @@ bool SpirvDebugPrintfEnabled();
 bool RenderDocEnabled();
 bool ReadbackLinearImagesEnabled();
 
+const Keymap& GetKeymap();
+
 } // namespace Config
 
 #endif /* KYTY_COMMON_EMULATOR_CONFIG_H_ */

src/graphics/guest_gpu/command_processor/commandProcessor.h

@@ -15,6 +15,13 @@ bool TestWaitRegMemValue(uint64_t value, uint64_t ref, uint64_t mask, uint32_t f
 
 enum class Pm4ProcessResult { Complete, Blocked };
 
+enum class ContextStateOperation : uint32_t {
+	Clear     = 0,
+	Push      = 1,
+	Pop       = 2,
+	PushClear = 3,
+};
+
 class Pm4Execution {
 public:
 	[[nodiscard]] bool MadeProgress() const noexcept { return m_made_progress; }
@@ -49,6 +56,7 @@ public:
 	KYTY_CLASS_NO_COPY(CommandProcessor);
 
 	void Reset();
+	void ApplyContextStateOperation(ContextStateOperation operation);
 
 	void            BufferInit();
 	void            BufferFlush();
@@ -149,6 +157,8 @@ private:
 
 	RenderContext&   m_renderer;
 	HW::Context      m_ctx;
+	HW::Context      m_saved_ctx;
+	bool             m_context_state_pushed = false;
 	HW::UserConfig   m_ucfg;
 	HW::Shader       m_sh_ctx;
 	HW::UserSgprType m_user_data_marker                 = HW::UserSgprType::Unknown;

src/graphics/guest_gpu/command_processor/pm4Dispatch.cpp

@@ -254,6 +254,7 @@ constexpr auto MakeCustomOpcodeDispatchTable() {
 	g_cp_op_custom_func[Pm4::R_WAIT_MEM_64]    = CpOpWaitRegMem64;
 	g_cp_op_custom_func[Pm4::R_FLIP]           = CpOpFlip;
 	g_cp_op_custom_func[Pm4::R_RELEASE_MEM]    = CpOpReleaseMem;
+	g_cp_op_custom_func[Pm4::R_CONTEXT_STATE]  = CpOpContextState;
 
 	return g_cp_op_custom_func;
 }

src/graphics/guest_gpu/command_processor/pm4Dispatch.h

@@ -121,6 +121,7 @@ uint32_t HwUcSetGeIndexOffset(CommandProcessor&, uint32_t, uint32_t, const uint3
 uint32_t HwUcSetGdsOaRegisters(CommandProcessor&, uint32_t, uint32_t, const uint32_t*, uint32_t);
 
 uint32_t CpOpNop(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
+uint32_t CpOpContextState(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpSetBase(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpClearState(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpIndexBase(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);

src/graphics/guest_gpu/command_processor/pm4Handlers.cpp

@@ -2107,6 +2107,19 @@ KYTY_CP_OP_PARSER(CpOpClearState) {
 	return 1;
 }
 
+KYTY_CP_OP_PARSER(CpOpContextState) {
+	KYTY_PROFILER_FUNCTION();
+
+	const auto packet_size_dw = KYTY_PM4_LEN(cmd_id);
+	EXIT_NOT_IMPLEMENTED(KYTY_PM4_R(cmd_id) != Pm4::R_CONTEXT_STATE);
+	EXIT_NOT_IMPLEMENTED(packet_size_dw != 3u && packet_size_dw != 5u);
+	EXIT_NOT_IMPLEMENTED(buffer[0] > static_cast<uint32_t>(ContextStateOperation::PushClear));
+
+	cp.ApplyContextStateOperation(static_cast<ContextStateOperation>(buffer[0]));
+
+	return packet_size_dw - 1u;
+}
+
 KYTY_CP_OP_PARSER(CpOpDumpConstRam) {
 	KYTY_PROFILER_FUNCTION();
 

src/graphics/guest_gpu/graphicsRun.cpp

@@ -312,6 +312,8 @@ void CommandProcessor::Reset() {
 	m_sh_ctx.Reset();
 	m_ucfg.Reset();
 	m_ctx.Reset();
+	m_saved_ctx.Reset();
+	m_context_state_pushed = false;
 	m_index_type_and_size              = 0;
 	m_index_buffer_size                = 0;
 	m_user_data_marker                 = HW::UserSgprType::Unknown;
@@ -321,6 +323,30 @@ void CommandProcessor::Reset() {
 	std::memset(m_const_ram, 0, sizeof(m_const_ram));
 }
 
+void CommandProcessor::ApplyContextStateOperation(ContextStateOperation operation) {
+	switch (operation) {
+		case ContextStateOperation::Clear: m_ctx.Reset(); break;
+		case ContextStateOperation::Push:
+			EXIT_IF(m_context_state_pushed);
+			m_saved_ctx            = m_ctx;
+			m_context_state_pushed = true;
+			break;
+		case ContextStateOperation::Pop:
+			EXIT_IF(!m_context_state_pushed);
+			m_ctx                  = m_saved_ctx;
+			m_saved_ctx            = {};
+			m_context_state_pushed = false;
+			break;
+		case ContextStateOperation::PushClear:
+			EXIT_IF(m_context_state_pushed);
+			m_saved_ctx            = m_ctx;
+			m_context_state_pushed = true;
+			m_ctx.Reset();
+			break;
+		default: EXIT("unknown context state operation: %u\n", static_cast<uint32_t>(operation));
+	}
+}
+
 void CommandProcessor::BufferInit() {
 	GetScheduler().Begin(m_ctx, m_ucfg, m_sh_ctx);
 }

src/graphics/guest_gpu/pm4.cpp

@@ -30,6 +30,7 @@ constexpr auto MakeRegisterNames() {
 	names[R_FLIP]           = "R_FLIP";
 	names[R_RELEASE_MEM]    = "R_RELEASE_MEM";
 	names[R_DMA_DATA]       = "R_DMA_DATA";
+	names[R_CONTEXT_STATE]  = "R_CONTEXT_STATE";
 	return names;
 }
 

src/graphics/guest_gpu/pm4.h

@@ -87,6 +87,7 @@ constexpr uint32_t R_WAIT_MEM_64    = 0x16;
 constexpr uint32_t R_FLIP           = 0x17;
 constexpr uint32_t R_RELEASE_MEM    = 0x18;
 constexpr uint32_t R_DMA_DATA       = 0x19;
+constexpr uint32_t R_CONTEXT_STATE  = 0x1A;
 
 constexpr uint32_t R_NUM = 0x3F + 1;
 

src/graphics/presentation/window/hostInput.cpp

@@ -0,0 +1,256 @@
+#include "graphics/presentation/window/hostInput.h"
+
+#include "SDL_keyboard.h"
+#include "SDL_keycode.h"
+#include "SDL_mouse.h"
+#include "common/assert.h"
+#include "common/emulatorConfig.h"
+#include "libs/controller.h"
+
+#include <algorithm>
+#include <array>
+#include <cstddef>
+#include <string>
+#include <string_view>
+#include <utility>
+
+namespace Libs::Graphics {
+
+namespace {
+
+struct Binding {
+	SDL_Keycode key          = SDLK_UNKNOWN;
+	uint8_t     mouse_button = 0;
+	uint32_t    pad_button   = 0;
+};
+
+static constexpr std::array PAD_BUTTONS = {
+    std::pair {std::string_view("L3"), Controller::PAD_BUTTON_L3},
+    std::pair {std::string_view("R3"), Controller::PAD_BUTTON_R3},
+    std::pair {std::string_view("Options"), Controller::PAD_BUTTON_OPTIONS},
+    std::pair {std::string_view("Up"), Controller::PAD_BUTTON_UP},
+    std::pair {std::string_view("Right"), Controller::PAD_BUTTON_RIGHT},
+    std::pair {std::string_view("Down"), Controller::PAD_BUTTON_DOWN},
+    std::pair {std::string_view("Left"), Controller::PAD_BUTTON_LEFT},
+    std::pair {std::string_view("L2"), Controller::PAD_BUTTON_L2},
+    std::pair {std::string_view("R2"), Controller::PAD_BUTTON_R2},
+    std::pair {std::string_view("L1"), Controller::PAD_BUTTON_L1},
+    std::pair {std::string_view("R1"), Controller::PAD_BUTTON_R1},
+    std::pair {std::string_view("Triangle"), Controller::PAD_BUTTON_TRIANGLE},
+    std::pair {std::string_view("Circle"), Controller::PAD_BUTTON_CIRCLE},
+    std::pair {std::string_view("Cross"), Controller::PAD_BUTTON_CROSS},
+    std::pair {std::string_view("Square"), Controller::PAD_BUTTON_SQUARE},
+    std::pair {std::string_view("TouchPad"), Controller::PAD_BUTTON_TOUCH_PAD},
+};
+
+uint32_t PadButtonFromName(std::string_view name) {
+	const auto button = std::find_if(PAD_BUTTONS.begin(), PAD_BUTTONS.end(),
+	                                 [name](const auto& item) { return item.first == name; });
+	return button != PAD_BUTTONS.end() ? button->second : 0;
+}
+
+SDL_Keycode NormalizeKey(SDL_Keycode key) {
+	switch (key) {
+		case SDLK_RSHIFT: return SDLK_LSHIFT;
+		case SDLK_RCTRL: return SDLK_LCTRL;
+		case SDLK_RALT: return SDLK_LALT;
+		case SDLK_RGUI: return SDLK_LGUI;
+		default: return key;
+	}
+}
+
+uint8_t MouseButtonFromName(std::string_view name) {
+	static constexpr std::array buttons = {
+	    std::pair {std::string_view("Mouse:Left"), uint8_t {SDL_BUTTON_LEFT}},
+	    std::pair {std::string_view("Mouse:Middle"), uint8_t {SDL_BUTTON_MIDDLE}},
+	    std::pair {std::string_view("Mouse:Right"), uint8_t {SDL_BUTTON_RIGHT}},
+	    std::pair {std::string_view("Mouse:X1"), uint8_t {SDL_BUTTON_X1}},
+	    std::pair {std::string_view("Mouse:X2"), uint8_t {SDL_BUTTON_X2}},
+	};
+
+	const auto button = std::find_if(buttons.begin(), buttons.end(),
+	                                 [name](const auto& item) { return item.first == name; });
+	return button != buttons.end() ? button->second : 0;
+}
+
+bool Conflicts(const Binding& first, const Binding& second) {
+	return first.pad_button == second.pad_button ||
+	       (first.key != SDLK_UNKNOWN && first.key == second.key) ||
+	       (first.mouse_button != 0 && first.mouse_button == second.mouse_button);
+}
+
+class InputMap {
+public:
+	InputMap(): m_custom(!Config::GetKeymap().empty()) {
+		for (const auto& value: Config::GetKeymap()) {
+			const std::string_view entry = value;
+			const auto             split = entry.find('=');
+
+			Binding binding;
+			if (split != std::string_view::npos) {
+				binding.pad_button    = PadButtonFromName(entry.substr(0, split));
+				const auto host_input = entry.substr(split + 1);
+				binding.mouse_button  = MouseButtonFromName(host_input);
+				if (binding.mouse_button == 0) {
+					binding.key = NormalizeKey(SDL_GetKeyFromName(std::string(host_input).c_str()));
+				}
+			}
+
+			const bool reserved =
+			    binding.key == SDLK_ESCAPE || binding.key == SDLK_SPACE || binding.key == SDLK_F1;
+			if (binding.pad_button == 0 || reserved ||
+			    (binding.key == SDLK_UNKNOWN && binding.mouse_button == 0)) {
+				EXIT("Invalid input mapping: %s\n", value.c_str());
+			}
+			Add(binding);
+		}
+	}
+	[[nodiscard]] bool Custom() const { return m_custom; }
+
+	[[nodiscard]] uint32_t FindKey(int key_code) const {
+		key_code = NormalizeKey(static_cast<SDL_Keycode>(key_code));
+		const auto binding =
+		    std::find_if(m_bindings.begin(), m_bindings.begin() + m_size,
+		                 [key_code](const auto& item) { return item.key == key_code; });
+		return binding != m_bindings.begin() + m_size ? binding->pad_button : 0;
+	}
+
+	[[nodiscard]] uint32_t FindMouseButton(uint8_t mouse_button) const {
+		const auto binding = std::find_if(
+		    m_bindings.begin(), m_bindings.begin() + m_size,
+		    [mouse_button](const auto& item) { return item.mouse_button == mouse_button; });
+		return binding != m_bindings.begin() + m_size ? binding->pad_button : 0;
+	}
+
+private:
+	void Add(const Binding& binding) {
+		for (std::size_t index = 0; index < m_size;) {
+			if (Conflicts(m_bindings[index], binding)) {
+				m_bindings[index] = m_bindings[--m_size];
+			} else {
+				index++;
+			}
+		}
+		EXIT_IF(m_size >= m_bindings.size());
+		m_bindings[m_size++] = binding;
+	}
+
+	std::array<Binding, PAD_BUTTONS.size()> m_bindings {};
+	std::size_t                             m_size = 0;
+	bool                                    m_custom;
+};
+
+const InputMap& GetInputMap() {
+	static const InputMap map;
+	return map;
+}
+
+void SetButton(uint32_t button, bool down) {
+	if (button == Controller::PAD_BUTTON_L2) {
+		Controller::ControllerAxis(Controller::HOST_INPUT_CONTROLLER_ID,
+		                           Controller::Axis::TriggerLeft, down ? 255 : 0);
+	} else if (button == Controller::PAD_BUTTON_R2) {
+		Controller::ControllerAxis(Controller::HOST_INPUT_CONTROLLER_ID,
+		                           Controller::Axis::TriggerRight, down ? 255 : 0);
+	} else if (button != 0) {
+		Controller::ControllerButton(Controller::HOST_INPUT_CONTROLLER_ID, button, down);
+	}
+}
+
+uint32_t DefaultKeyboardButton(int key_code) {
+	switch (NormalizeKey(static_cast<SDL_Keycode>(key_code))) {
+		case SDLK_UP: return Controller::PAD_BUTTON_UP;
+		case SDLK_LEFT: return Controller::PAD_BUTTON_LEFT;
+		case SDLK_DOWN: return Controller::PAD_BUTTON_DOWN;
+		case SDLK_RIGHT: return Controller::PAD_BUTTON_RIGHT;
+		case SDLK_j: return Controller::PAD_BUTTON_CROSS;
+		case SDLK_i: return Controller::PAD_BUTTON_TRIANGLE;
+		case SDLK_k: return Controller::PAD_BUTTON_SQUARE;
+		case SDLK_l: return Controller::PAD_BUTTON_CIRCLE;
+		case SDLK_q: return Controller::PAD_BUTTON_L1;
+		case SDLK_e: return Controller::PAD_BUTTON_R1;
+		case SDLK_LSHIFT: return Controller::PAD_BUTTON_L3;
+		case SDLK_LCTRL: return Controller::PAD_BUTTON_R3;
+		case SDLK_RETURN:
+		case SDLK_RETURN2: return Controller::PAD_BUTTON_OPTIONS;
+		case SDLK_BACKSPACE:
+		case SDLK_TAB: return Controller::PAD_BUTTON_TOUCH_PAD;
+		default: return 0;
+	}
+}
+
+struct StickKeys {
+	bool left  = false;
+	bool right = false;
+	bool up    = false;
+	bool down  = false;
+};
+
+void SetStickAxis(Controller::Axis axis, bool negative, bool positive) {
+	const int value = negative == positive ? 128 : negative ? 0 : 255;
+	Controller::ControllerAxis(Controller::HOST_INPUT_CONTROLLER_ID, axis, value);
+}
+
+void DefaultKeyboardInput(int key_code, bool down) {
+	static StickKeys left;
+	static StickKeys right;
+
+	switch (NormalizeKey(static_cast<SDL_Keycode>(key_code))) {
+		case SDLK_a:
+			left.left = down;
+			SetStickAxis(Controller::Axis::LeftX, left.left, left.right);
+			return;
+		case SDLK_d:
+			left.right = down;
+			SetStickAxis(Controller::Axis::LeftX, left.left, left.right);
+			return;
+		case SDLK_w:
+			left.up = down;
+			SetStickAxis(Controller::Axis::LeftY, left.up, left.down);
+			return;
+		case SDLK_s:
+			left.down = down;
+			SetStickAxis(Controller::Axis::LeftY, left.up, left.down);
+			return;
+		case SDLK_f:
+			right.left = down;
+			SetStickAxis(Controller::Axis::RightX, right.left, right.right);
+			return;
+		case SDLK_h:
+			right.right = down;
+			SetStickAxis(Controller::Axis::RightX, right.left, right.right);
+			return;
+		case SDLK_t:
+			right.up = down;
+			SetStickAxis(Controller::Axis::RightY, right.up, right.down);
+			return;
+		case SDLK_g:
+			right.down = down;
+			SetStickAxis(Controller::Axis::RightY, right.up, right.down);
+			return;
+		default: SetButton(DefaultKeyboardButton(key_code), down); return;
+	}
+}
+} // namespace
+
+void HostInputInit() {
+	GetInputMap();
+}
+
+void HostInputKey(int key_code, bool down) {
+	const auto& map = GetInputMap();
+	if (map.Custom()) {
+		SetButton(map.FindKey(key_code), down);
+	} else {
+		DefaultKeyboardInput(key_code, down);
+	}
+}
+
+void HostInputMouseButton(uint8_t mouse_button, bool down) {
+	const auto& map = GetInputMap();
+	if (map.Custom() && mouse_button != 0) {
+		SetButton(map.FindMouseButton(mouse_button), down);
+	}
+}
+
+} // namespace Libs::Graphics

src/graphics/presentation/window/hostInput.h

@@ -0,0 +1,14 @@
+#ifndef KYTY_GRAPHICS_PRESENTATION_WINDOW_HOST_INPUT_H_
+#define KYTY_GRAPHICS_PRESENTATION_WINDOW_HOST_INPUT_H_
+
+#include <cstdint>
+
+namespace Libs::Graphics {
+
+void HostInputInit();
+void HostInputKey(int key_code, bool down);
+void HostInputMouseButton(uint8_t mouse_button, bool down);
+
+} // namespace Libs::Graphics
+
+#endif /* KYTY_GRAPHICS_PRESENTATION_WINDOW_HOST_INPUT_H_ */

src/graphics/presentation/window/swapchain.cpp

@@ -789,12 +789,14 @@ void Presenter::Present(Frame& frame, bool reuse) {
 		    [&window] {
 			    window.UpdateIcon();
 			    SDL_ShowWindow(window.window);
+			    SDL_RaiseWindow(window.window);
 		    },
 		    true);
 #else
 		window.UpdateIcon();
 
 		SDL_ShowWindow(window.window);
+		SDL_RaiseWindow(window.window);
 #endif
 
 		window.window_hidden = false;

src/graphics/presentation/window/window.cpp

@@ -32,6 +32,7 @@
 #include "graphics/host_gpu/vma.h"
 #include "graphics/host_gpu/vulkanCommon.h"
 #include "graphics/presentation/renderDoc.h"
+#include "graphics/presentation/window/hostInput.h"
 #include "graphics/presentation/window/windowInternal.h"
 #include "kytyGitVersion.h"
 #include "libs/controller.h"
@@ -59,8 +60,6 @@
 
 namespace Libs::Graphics {
 
-constexpr int KEYBOARD_CONTROLLER_ID = -1000;
-
 struct EventKeyboard {
 	bool     down;
 	bool     up;
@@ -73,26 +72,6 @@ struct EventKeyboard {
 	double   timestamp_seconds;
 };
 
-static uint32_t KeyboardKeyToPadButton(int key_code) {
-	switch (key_code) {
-		case SDLK_w: return Controller::PAD_BUTTON_UP;
-		case SDLK_a: return Controller::PAD_BUTTON_LEFT;
-		case SDLK_s: return Controller::PAD_BUTTON_DOWN;
-		case SDLK_d: return Controller::PAD_BUTTON_RIGHT;
-		case SDLK_j: return Controller::PAD_BUTTON_CROSS;
-		case SDLK_i: return Controller::PAD_BUTTON_TRIANGLE;
-		case SDLK_k: return Controller::PAD_BUTTON_SQUARE;
-		case SDLK_l: return Controller::PAD_BUTTON_CIRCLE;
-		case SDLK_q: return Controller::PAD_BUTTON_L1;
-		case SDLK_e: return Controller::PAD_BUTTON_R1;
-		case SDLK_RETURN:
-		case SDLK_RETURN2: return Controller::PAD_BUTTON_OPTIONS;
-		case SDLK_BACKSPACE:
-		case SDLK_TAB: return Controller::PAD_BUTTON_TOUCH_PAD;
-		default: return 0;
-	}
-}
-
 static uint32_t ControllerButtonToPadButton(int button) {
 	switch (button) {
 		case SDL_CONTROLLER_BUTTON_A: return Controller::PAD_BUTTON_CROSS;
@@ -261,16 +240,11 @@ static void GameEventKeyboard(WindowLoopState& game, const EventKeyboard& key) {
 		}
 	}
 
-	const auto button = KeyboardKeyToPadButton(key.key_code);
-	if (button != 0 && (key.down || key.up) && !key.repeat) {
-		static bool keyboard_connected = false;
-		if (!keyboard_connected) {
-			Controller::ControllerConnect(KEYBOARD_CONTROLLER_ID);
-			keyboard_connected = true;
-		}
-		Controller::ControllerButton(KEYBOARD_CONTROLLER_ID, button, key.down);
-	}
 #endif
+
+	if ((key.down || key.up) && !key.repeat) {
+		HostInputKey(key.key_code, key.down);
+	}
 }
 
 static void GameEventMouse([[maybe_unused]] const EventMouse& mb) {
@@ -292,6 +266,23 @@ static void GameEventMouse([[maybe_unused]] const EventMouse& mb) {
 		     mb.x, mb.y);
 	}
 #endif
+
+	if (mb.down || mb.up) {
+		uint8_t mouse_button = 0;
+		if (mb.left) {
+			mouse_button = SDL_BUTTON_LEFT;
+		} else if (mb.middle) {
+			mouse_button = SDL_BUTTON_MIDDLE;
+		} else if (mb.right) {
+			mouse_button = SDL_BUTTON_RIGHT;
+		} else if (mb.x1) {
+			mouse_button = SDL_BUTTON_X1;
+		} else if (mb.x2) {
+			mouse_button = SDL_BUTTON_X2;
+		}
+
+		HostInputMouseButton(mouse_button, mb.down);
+	}
 }
 
 static void GameEventFinger([[maybe_unused]] const EventFinger& f) {
@@ -795,6 +786,7 @@ static void WindowCreate(WindowContext& context) {
 	if (SDL_InitSubSystem(SDL_INIT_VIDEO | SDL_INIT_GAMECONTROLLER) < 0) {
 		EXIT("%s\n", SDL_GetError());
 	}
+	HostInputInit();
 
 	LOGF("WindowCreate(): width = %d, height = %d\n", width, height);
 

src/kernel/syncOnAddress.cpp

@@ -0,0 +1,287 @@
+#include "kernel/syncOnAddress.h"
+
+#include "common/threads.h"
+#include "libs/errno.h"
+
+#include <algorithm>
+#include <chrono>
+#include <climits>
+#include <cstdint>
+#include <list>
+#include <memory>
+#include <mutex>
+#include <unordered_map>
+
+#if KYTY_PLATFORM == KYTY_PLATFORM_LINUX && !defined(__APPLE__)
+#include <cerrno>
+#include <linux/futex.h>
+#include <sys/syscall.h>
+#include <unistd.h>
+#endif
+
+namespace Libs::LibKernel::SyncOnAddress {
+
+namespace {
+
+constexpr uint32_t SIGNAL_POLL_MICROS = 10000;
+
+using Clock = std::chrono::steady_clock;
+
+template <typename T>
+[[nodiscard]] bool IsValidWaitAddress(const volatile T* address) {
+	return address != nullptr && (reinterpret_cast<uintptr_t>(address) & (alignof(T) - 1u)) == 0;
+}
+
+[[nodiscard]] bool IsValidWakeAddress(const volatile void* address) {
+	return address != nullptr &&
+	       (reinterpret_cast<uintptr_t>(address) & (alignof(uint32_t) - 1u)) == 0;
+}
+
+template <typename T>
+[[nodiscard]] T ReadWord(const volatile T* address) {
+	return __atomic_load_n(address, __ATOMIC_ACQUIRE);
+}
+
+struct WaitDeadline {
+	bool              finite = false;
+	Clock::time_point end {};
+};
+
+[[nodiscard]] WaitDeadline MakeDeadline(const uint32_t* timeout_micros) {
+	if (timeout_micros == nullptr) {
+		return {};
+	}
+	return {true, Clock::now() + std::chrono::microseconds(*timeout_micros)};
+}
+
+[[nodiscard]] uint32_t GetWaitSliceMicros(const WaitDeadline& deadline, bool first_wait) {
+	if (!deadline.finite) {
+		return SIGNAL_POLL_MICROS;
+	}
+
+	const auto now = Clock::now();
+	if (now >= deadline.end) {
+		return first_wait ? 0u : UINT32_MAX;
+	}
+
+	const auto remaining =
+	    std::chrono::duration_cast<std::chrono::microseconds>(deadline.end - now).count();
+	return static_cast<uint32_t>(std::min<int64_t>(remaining, SIGNAL_POLL_MICROS));
+}
+
+void PollSignals(signal_poll_func_t signal_poll) {
+	if (signal_poll != nullptr) {
+		signal_poll();
+	}
+}
+
+#if KYTY_PLATFORM == KYTY_PLATFORM_LINUX && !defined(__APPLE__)
+
+template <typename T>
+int WaitLinux(volatile T* address, T expected, const uint32_t* timeout_micros,
+              signal_poll_func_t signal_poll) {
+	const auto deadline   = MakeDeadline(timeout_micros);
+	bool       first_wait = true;
+
+	for (;;) {
+		if (ReadWord(address) != expected) {
+			return OK;
+		}
+		const auto slice_micros = GetWaitSliceMicros(deadline, first_wait);
+		if (slice_micros == UINT32_MAX) {
+			return ReadWord(address) == expected ? KERNEL_ERROR_ETIMEDOUT : OK;
+		}
+		const timespec timeout = {
+		    .tv_sec  = static_cast<time_t>(slice_micros / 1000000u),
+		    .tv_nsec = static_cast<long>(slice_micros % 1000000u) * 1000L,
+		};
+
+		long result     = 0;
+		int  wait_error = 0;
+		result          = syscall(SYS_futex, const_cast<T*>(address), FUTEX_WAIT_PRIVATE,
+		                          static_cast<uint32_t>(expected), &timeout, nullptr, 0);
+		if (result != 0) {
+			wait_error = errno;
+		}
+		if (result == 0 || wait_error == EAGAIN) {
+			return OK;
+		}
+		if (wait_error != ETIMEDOUT && wait_error != EINTR) {
+			return KERNEL_ERROR_EINVAL;
+		}
+
+		PollSignals(signal_poll);
+		if (deadline.finite && Clock::now() >= deadline.end) {
+			return ReadWord(address) == expected ? KERNEL_ERROR_ETIMEDOUT : OK;
+		}
+		first_wait = false;
+	}
+}
+
+int WakeLinux(volatile void* address, int32_t count) {
+	const auto result = syscall(SYS_futex, const_cast<void*>(address), FUTEX_WAKE_PRIVATE, count,
+	                            nullptr, nullptr, 0);
+	return result < 0 ? KERNEL_ERROR_EINVAL : OK;
+}
+
+#endif
+
+struct PortableWaiter {
+	Common::CondVar condition;
+	bool            wake_requested = false;
+};
+
+struct PortableAddressEntry {
+	Common::Mutex              mutex;
+	std::list<PortableWaiter*> waiters;
+};
+
+struct PortableAddressRegistry {
+	std::mutex                                                           mutex;
+	std::unordered_map<uintptr_t, std::shared_ptr<PortableAddressEntry>> entries;
+};
+
+PortableAddressRegistry& GetPortableRegistry() {
+	static PortableAddressRegistry registry;
+	return registry;
+}
+
+std::shared_ptr<PortableAddressEntry> RegisterPortableWaiter(volatile void*  address,
+                                                             PortableWaiter* waiter) {
+	auto&           registry = GetPortableRegistry();
+	std::lock_guard registry_lock(registry.mutex);
+	auto&           entry = registry.entries[reinterpret_cast<uintptr_t>(address)];
+	if (!entry) {
+		entry = std::make_shared<PortableAddressEntry>();
+	}
+	entry->mutex.Lock();
+	entry->waiters.push_back(waiter);
+	return entry;
+}
+
+void UnregisterPortableWaiter(volatile void*                               address,
+                              const std::shared_ptr<PortableAddressEntry>& entry,
+                              PortableWaiter*                              waiter) {
+	entry->waiters.remove(waiter);
+	const bool empty = entry->waiters.empty();
+	entry->mutex.Unlock();
+	if (!empty) {
+		return;
+	}
+
+	auto&           registry = GetPortableRegistry();
+	std::lock_guard registry_lock(registry.mutex);
+	entry->mutex.Lock();
+	const auto it = registry.entries.find(reinterpret_cast<uintptr_t>(address));
+	if (it != registry.entries.end() && it->second == entry && entry->waiters.empty()) {
+		registry.entries.erase(it);
+	}
+	entry->mutex.Unlock();
+}
+
+template <typename T>
+int WaitPortable(volatile T* address, T expected, const uint32_t* timeout_micros,
+                 signal_poll_func_t signal_poll) {
+	PortableWaiter waiter;
+	auto           entry      = RegisterPortableWaiter(address, &waiter);
+	const auto     deadline   = MakeDeadline(timeout_micros);
+	bool           first_wait = true;
+	int            result     = OK;
+
+	while (ReadWord(address) == expected && !waiter.wake_requested) {
+		const auto slice_micros = GetWaitSliceMicros(deadline, first_wait);
+		if (slice_micros == UINT32_MAX) {
+			result = KERNEL_ERROR_ETIMEDOUT;
+			break;
+		}
+		if (slice_micros == 0) {
+			result = KERNEL_ERROR_ETIMEDOUT;
+			break;
+		}
+
+		(void)waiter.condition.WaitFor(&entry->mutex, slice_micros);
+		entry->mutex.Unlock();
+		PollSignals(signal_poll);
+		entry->mutex.Lock();
+
+		if (deadline.finite && Clock::now() >= deadline.end && ReadWord(address) == expected &&
+		    !waiter.wake_requested) {
+			result = KERNEL_ERROR_ETIMEDOUT;
+			break;
+		}
+		first_wait = false;
+	}
+
+	UnregisterPortableWaiter(address, entry, &waiter);
+	return result;
+}
+
+int WakePortable(volatile void* address, int32_t count) {
+	if (count == 0) {
+		return OK;
+	}
+	auto&            registry = GetPortableRegistry();
+	std::unique_lock registry_lock(registry.mutex);
+	const auto       it = registry.entries.find(reinterpret_cast<uintptr_t>(address));
+	if (it == registry.entries.end()) {
+		return OK;
+	}
+	auto entry = it->second;
+	entry->mutex.Lock();
+	registry_lock.unlock();
+
+	int32_t remaining = count;
+	for (auto* waiter: entry->waiters) {
+		if (!waiter->wake_requested) {
+			waiter->wake_requested = true;
+			waiter->condition.Signal();
+			if (remaining != INT_MAX && --remaining == 0) {
+				break;
+			}
+		}
+	}
+	entry->mutex.Unlock();
+	return OK;
+}
+
+template <typename T>
+int WaitImpl(volatile T* address, T expected, const uint32_t* timeout_micros,
+             signal_poll_func_t signal_poll) {
+	if (!IsValidWaitAddress(address)) {
+		return KERNEL_ERROR_EINVAL;
+	}
+
+	int result = OK;
+#if KYTY_PLATFORM == KYTY_PLATFORM_LINUX && !defined(__APPLE__)
+	result = WaitLinux(address, expected, timeout_micros, signal_poll);
+#else
+	result = WaitPortable(address, expected, timeout_micros, signal_poll);
+#endif
+	PollSignals(signal_poll);
+	return result;
+}
+
+} // namespace
+
+int Wait32(volatile uint32_t* address, uint32_t expected, const uint32_t* timeout_micros,
+           signal_poll_func_t signal_poll) {
+	return WaitImpl(address, expected, timeout_micros, signal_poll);
+}
+
+int Wait64(volatile uint64_t* address, uint64_t expected, const uint32_t* timeout_micros,
+           signal_poll_func_t signal_poll) {
+	return WaitImpl(address, expected, timeout_micros, signal_poll);
+}
+
+int Wake(volatile void* address, int32_t count) {
+	if (!IsValidWakeAddress(address) || count < 0) {
+		return KERNEL_ERROR_EINVAL;
+	}
+
+#if KYTY_PLATFORM == KYTY_PLATFORM_LINUX && !defined(__APPLE__)
+	return WakeLinux(address, count);
+#endif
+	return WakePortable(address, count);
+}
+
+} // namespace Libs::LibKernel::SyncOnAddress

src/kernel/syncOnAddress.h

@@ -0,0 +1,19 @@
+#ifndef EMULATOR_INCLUDE_EMULATOR_KERNEL_SYNC_ON_ADDRESS_H_
+#define EMULATOR_INCLUDE_EMULATOR_KERNEL_SYNC_ON_ADDRESS_H_
+
+#include "common/abi.h"
+#include "common/common.h"
+
+namespace Libs::LibKernel::SyncOnAddress {
+
+using signal_poll_func_t = void (*)();
+
+int Wait32(volatile uint32_t* address, uint32_t expected, const uint32_t* timeout_micros,
+           signal_poll_func_t signal_poll = nullptr);
+int Wait64(volatile uint64_t* address, uint64_t expected, const uint32_t* timeout_micros,
+           signal_poll_func_t signal_poll = nullptr);
+int Wake(volatile void* address, int32_t count);
+
+} // namespace Libs::LibKernel::SyncOnAddress
+
+#endif /* EMULATOR_INCLUDE_EMULATOR_KERNEL_SYNC_ON_ADDRESS_H_ */

src/launcher/CMakeLists.txt

@@ -16,6 +16,7 @@ file(GLOB launcher_forms
 add_executable(launcher WIN32
 	${launcher_src}
 	${launcher_forms}
+	resources.qrc
 )
 
 

src/launcher/forms/configuration_list_widget.ui

@@ -69,6 +69,31 @@
            </property>
           </widget>
          </item>
+         <item>
+          <widget class="QToolButton" name="input_mapping_button">
+           <property name="minimumSize">
+            <size>
+             <width>36</width>
+             <height>36</height>
+            </size>
+           </property>
+           <property name="toolTip">
+            <string>Edit input mapping</string>
+           </property>
+           <property name="text">
+            <string/>
+           </property>
+           <property name="iconSize">
+            <size>
+             <width>32</width>
+             <height>32</height>
+            </size>
+           </property>
+           <property name="autoRaise">
+            <bool>true</bool>
+           </property>
+          </widget>
+         </item>
          <item>
           <widget class="QToolButton" name="edit_button">
            <property name="minimumSize">

src/launcher/icons/edit-configuration.svg

@@ -0,0 +1,7 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 32 32">
+  <g fill="none" stroke="#e8f1ff" stroke-linecap="round" stroke-linejoin="round" stroke-width="2">
+    <path d="M8 4h11l5 5v7M24 9h-5V4M16 27H8a2 2 0 0 1-2-2V6a2 2 0 0 1 2-2"/>
+    <path d="m15 24 1-4 8-8 4 4-8 8-4 1Z"/>
+    <path d="m22 14 4 4" opacity=".8"/>
+  </g>
+</svg>

src/launcher/icons/global-settings.svg

@@ -0,0 +1,8 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 32 32">
+  <g fill="none" stroke="#e8f1ff" stroke-linecap="round" stroke-linejoin="round" stroke-width="2">
+    <path d="M5 8h3M14 8h13M5 16h13M24 16h3M5 24h6M17 24h10"/>
+    <circle cx="11" cy="8" r="3"/>
+    <circle cx="21" cy="16" r="3"/>
+    <circle cx="14" cy="24" r="3"/>
+  </g>
+</svg>

src/launcher/icons/input-mapping.svg

@@ -0,0 +1,9 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 32 32">
+  <g fill="none" stroke="#e8f1ff" stroke-linecap="round" stroke-linejoin="round" stroke-width="2">
+    <path d="M10 10h12c3.6 0 5.8 2.3 6.6 6.8l.9 5.1c.4 2.4-.8 4.1-2.8 4.1-1.4 0-2.5-.7-3.6-2.3L21.3 21H10.7l-1.8 2.7C7.8 25.3 6.7 26 5.3 26c-2 0-3.2-1.7-2.8-4.1l.9-5.1C4.2 12.3 6.4 10 10 10Z"/>
+    <path d="M8 16h6M11 13v6"/>
+    <circle cx="22" cy="15" r="1" fill="#e8f1ff" stroke="none"/>
+    <circle cx="25" cy="18" r="1" fill="#e8f1ff" stroke="none"/>
+    <path d="M13 7h6M16 4v6" opacity=".8"/>
+  </g>
+</svg>

src/launcher/icons/remove-configuration.svg

@@ -0,0 +1,6 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 32 32">
+  <g fill="none" stroke="#e8f1ff" stroke-linecap="round" stroke-linejoin="round" stroke-width="2">
+    <path d="M6 9h20M12 9V5h8v4M9 9l1 18h12l1-18"/>
+    <path d="M14 14v8M18 14v8" opacity=".8"/>
+  </g>
+</svg>

src/launcher/include/configuration.h

@@ -99,6 +99,7 @@ public:
 	QString                printf_output_file          = "_kyty.txt";
 	ProfilerDirection      profiler_direction          = ProfilerDirection::None;
 	bool                   renderdoc_enabled           = false;
+	QStringList            host_input_mapping;
 
 	QString elf = QStringLiteral("eboot.bin");
 
@@ -118,6 +119,7 @@ public:
 		printf_output_file          = other.printf_output_file;
 		profiler_direction          = other.profiler_direction;
 		renderdoc_enabled           = other.renderdoc_enabled;
+		host_input_mapping          = other.host_input_mapping;
 	}
 
 	void CopyFrom(const Configuration& other) {
@@ -154,6 +156,7 @@ public:
 		KYTY_CFG_SET(printf_output_file);
 		KYTY_CFG_SET(profiler_direction);
 		KYTY_CFG_SET(renderdoc_enabled);
+		s->setValue("host_input_mapping", host_input_mapping);
 		KYTY_CFG_SET(elf);
 	}
 
@@ -180,7 +183,8 @@ public:
 		KYTY_CFG_GET(printf_output_file);
 		KYTY_CFG_GET(profiler_direction);
 		KYTY_CFG_GET(renderdoc_enabled);
-		elf = s->value("elf", elf).toString();
+		host_input_mapping = s->value("host_input_mapping", host_input_mapping).toStringList();
+		elf                = s->value("elf", elf).toString();
 	}
 };
 

src/launcher/include/configurationListWidget.h

@@ -38,7 +38,10 @@ public:
 
 	[[nodiscard]] const QString&     GetSettingsFile() const { return m_settings_file; }
 	[[nodiscard]] const QStringList& GetGameDirectories() const { return m_game_dirs; }
-	[[nodiscard]] bool               CanViewSelectedTrophies() const;
+	[[nodiscard]] const QStringList& GetHostInputMapping() const {
+		return m_global_info.host_input_mapping;
+	}
+	[[nodiscard]] bool CanViewSelectedTrophies() const;
 
 	bool EnsureGameDirectory();
 	void ScanGameDirectory();
@@ -58,6 +61,7 @@ protected slots:
 	void edit_configuration();
 	void delete_configuartion();
 	void edit_global_settings();
+	void edit_input_mapping();
 	void run_configuration();
 	void list_currentItemChanged(QTreeWidgetItem* current, QTreeWidgetItem* previous);
 	void list_itemDoubleClicked(QTreeWidgetItem* witem, int column);

src/launcher/include/inputMappingDialog.h

@@ -0,0 +1,31 @@
+#ifndef LAUNCHER_INCLUDE_INPUT_MAPPING_DIALOG_H_
+#define LAUNCHER_INCLUDE_INPUT_MAPPING_DIALOG_H_
+
+#include <QDialog>
+#include <QStringList>
+
+class QPushButton;
+class QTreeWidget;
+class QTreeWidgetItem;
+class QWidget;
+
+class InputMappingDialog final: public QDialog {
+public:
+	explicit InputMappingDialog(const QStringList& mapping, QWidget* parent = nullptr);
+
+	[[nodiscard]] QStringList Mapping() const;
+
+private:
+	void ChangeBinding();
+	void ClearBinding();
+	void RestoreDefaults();
+	void SetBinding(QTreeWidgetItem* item, const QString& binding);
+	void UpdateButtons();
+
+	QTreeWidget* m_bindings      = nullptr;
+	QPushButton* m_change_button = nullptr;
+	QPushButton* m_clear_button  = nullptr;
+	bool         m_customized    = false;
+};
+
+#endif /* LAUNCHER_INCLUDE_INPUT_MAPPING_DIALOG_H_ */

src/launcher/resources.qrc

@@ -0,0 +1,8 @@
+<RCC>
+ <qresource prefix="/icons">
+  <file alias="edit-configuration.svg">icons/edit-configuration.svg</file>
+  <file alias="global-settings.svg">icons/global-settings.svg</file>
+  <file alias="input-mapping.svg">icons/input-mapping.svg</file>
+  <file alias="remove-configuration.svg">icons/remove-configuration.svg</file>
+ </qresource>
+</RCC>

src/launcher/src/configurationListWidget.cpp

@@ -6,6 +6,7 @@
 #include "configurationEditDialog.h"
 #include "configurationItem.h"
 #include "gameListTreeWidget.h"
+#include "inputMappingDialog.h"
 #include "mainDialog.h"
 #include "patchesDialog.h"
 #include "trophyViewerDialog.h"
@@ -169,6 +170,7 @@ static void ApplyGameListStyle(Ui::ConfigurationListWidget* ui) {
 	    "QToolButton:disabled { background: transparent; }"));
 	ui->edit_button->setStyleSheet(ui->global_settings_button->styleSheet());
 	ui->delete_button->setStyleSheet(ui->global_settings_button->styleSheet());
+	ui->input_mapping_button->setStyleSheet(ui->global_settings_button->styleSheet());
 }
 
 static void AddSaveDataDir(QStringList* dirs, QSet<QString>* seen, const QString& root,
@@ -226,10 +228,12 @@ ConfigurationListWidget::ConfigurationListWidget(QWidget* parent)
 	m_ui->setupUi(this);
 	ApplyGameListStyle(m_ui);
 
-	m_ui->global_settings_button->setIcon(style()->standardIcon(QStyle::SP_FileDialogDetailedView));
+	m_ui->global_settings_button->setIcon(QIcon(QStringLiteral(":/icons/global-settings.svg")));
 	m_ui->global_settings_button->setToolTip(tr("Edit global settings and game folders"));
-	m_ui->edit_button->setIcon(style()->standardIcon(QStyle::SP_FileIcon));
-	m_ui->delete_button->setIcon(style()->standardIcon(QStyle::SP_DialogDiscardButton));
+	m_ui->input_mapping_button->setIcon(QIcon(QStringLiteral(":/icons/input-mapping.svg")));
+	m_ui->input_mapping_button->setToolTip(tr("Edit global input mapping"));
+	m_ui->edit_button->setIcon(QIcon(QStringLiteral(":/icons/edit-configuration.svg")));
+	m_ui->delete_button->setIcon(QIcon(QStringLiteral(":/icons/remove-configuration.svg")));
 
 	m_ui->delete_button->setEnabled(false);
 	m_ui->edit_button->setEnabled(false);
@@ -249,6 +253,8 @@ ConfigurationListWidget::ConfigurationListWidget(QWidget* parent)
 
 	connect(m_ui->global_settings_button, &QToolButton::clicked, this,
 	        &ConfigurationListWidget::edit_global_settings);
+	connect(m_ui->input_mapping_button, &QToolButton::clicked, this,
+	        &ConfigurationListWidget::edit_input_mapping);
 	connect(m_ui->edit_button, &QToolButton::clicked, this,
 	        &ConfigurationListWidget::edit_configuration);
 	connect(m_ui->delete_button, &QToolButton::clicked, this,
@@ -717,6 +723,14 @@ void ConfigurationListWidget::edit_global_settings() {
 	}
 }
 
+void ConfigurationListWidget::edit_input_mapping() {
+	InputMappingDialog dialog(m_global_info.host_input_mapping, this);
+	if (dialog.exec() == QDialog::Accepted) {
+		m_global_info.host_input_mapping = dialog.Mapping();
+		WriteSettings();
+	}
+}
+
 void ConfigurationListWidget::ClearCustomSettings(ConfigurationItem* item) {
 	delete m_custom_infos.take(item->GetInfo().game_path);
 }

src/launcher/src/inputMappingDialog.cpp

@@ -0,0 +1,311 @@
+#include "inputMappingDialog.h"
+
+#include <QAbstractItemView>
+#include <QDialogButtonBox>
+#include <QHBoxLayout>
+#include <QHash>
+#include <QHeaderView>
+#include <QKeyEvent>
+#include <QKeySequence>
+#include <QLabel>
+#include <QMouseEvent>
+#include <QPushButton>
+#include <QStringList>
+#include <QTreeWidget>
+#include <QVBoxLayout>
+
+namespace {
+
+constexpr int BINDING_COLUMN = 1;
+
+struct PadButton {
+	const char* id;
+	const char* label;
+	const char* default_binding;
+};
+
+constexpr PadButton PAD_BUTTONS[] = {
+    {"Up", "D-pad Up", "Up"},
+    {"Down", "D-pad Down", "Down"},
+    {"Left", "D-pad Left", "Left"},
+    {"Right", "D-pad Right", "Right"},
+    {"Triangle", "Triangle", "I"},
+    {"Circle", "Circle", "L"},
+    {"Cross", "Cross", "J"},
+    {"Square", "Square", "K"},
+    {"L1", "L1", "Q"},
+    {"R1", "R1", "E"},
+    {"L2", "L2", ""},
+    {"R2", "R2", ""},
+    {"L3", "L3", "Left Shift"},
+    {"R3", "R3", "Left Ctrl"},
+    {"Options", "Options", "Return"},
+    {"TouchPad", "Touch pad", "Backspace"},
+};
+
+QString KeypadName(int key) {
+	if (key >= Qt::Key_0 && key <= Qt::Key_9) {
+		return QStringLiteral("Keypad %1").arg(key - Qt::Key_0);
+	}
+
+	switch (key) {
+		case Qt::Key_Return:
+		case Qt::Key_Enter: return QStringLiteral("Keypad Enter");
+		case Qt::Key_Slash: return QStringLiteral("Keypad /");
+		case Qt::Key_Asterisk: return QStringLiteral("Keypad *");
+		case Qt::Key_Minus: return QStringLiteral("Keypad -");
+		case Qt::Key_Plus: return QStringLiteral("Keypad +");
+		case Qt::Key_Period: return QStringLiteral("Keypad .");
+		case Qt::Key_Equal: return QStringLiteral("Keypad =");
+		case Qt::Key_Comma: return QStringLiteral("Keypad ,");
+		default: return {};
+	}
+}
+
+QString KeyName(const QKeyEvent& event) {
+	const int key = event.key();
+	if (event.modifiers().testFlag(Qt::KeypadModifier)) {
+		return KeypadName(key);
+	}
+	if (key >= Qt::Key_A && key <= Qt::Key_Z) {
+		return QChar(key);
+	}
+	if (key >= Qt::Key_0 && key <= Qt::Key_9) {
+		return QChar(key);
+	}
+	if (key >= Qt::Key_F1 && key <= Qt::Key_F24) {
+		return QStringLiteral("F%1").arg(key - Qt::Key_F1 + 1);
+	}
+
+	switch (key) {
+		case Qt::Key_Return:
+		case Qt::Key_Enter: return QStringLiteral("Return");
+		case Qt::Key_Backspace: return QStringLiteral("Backspace");
+		case Qt::Key_Tab: return QStringLiteral("Tab");
+		case Qt::Key_Shift: return QStringLiteral("Left Shift");
+		case Qt::Key_Control: return QStringLiteral("Left Ctrl");
+		case Qt::Key_Alt: return QStringLiteral("Left Alt");
+		case Qt::Key_Meta: return QStringLiteral("Left GUI");
+		case Qt::Key_Insert: return QStringLiteral("Insert");
+		case Qt::Key_Delete: return QStringLiteral("Delete");
+		case Qt::Key_Home: return QStringLiteral("Home");
+		case Qt::Key_End: return QStringLiteral("End");
+		case Qt::Key_PageUp: return QStringLiteral("PageUp");
+		case Qt::Key_PageDown: return QStringLiteral("PageDown");
+		case Qt::Key_Left: return QStringLiteral("Left");
+		case Qt::Key_Right: return QStringLiteral("Right");
+		case Qt::Key_Up: return QStringLiteral("Up");
+		case Qt::Key_Down: return QStringLiteral("Down");
+		case Qt::Key_CapsLock: return QStringLiteral("CapsLock");
+		case Qt::Key_NumLock: return QStringLiteral("Numlock");
+		case Qt::Key_ScrollLock: return QStringLiteral("ScrollLock");
+		case Qt::Key_Pause: return QStringLiteral("Pause");
+		case Qt::Key_Print: return QStringLiteral("PrintScreen");
+		default: break;
+	}
+
+	if (event.modifiers() != Qt::NoModifier) {
+		return {};
+	}
+	const auto name = QKeySequence(key).toString(QKeySequence::PortableText);
+	return name.size() == 1 ? name : QString();
+}
+
+class InputCaptureDialog final: public QDialog {
+public:
+	explicit InputCaptureDialog(QWidget* parent): QDialog(parent) {
+		setWindowTitle(tr("Set Binding"));
+		setModal(true);
+		setMinimumWidth(360);
+
+		auto* layout = new QVBoxLayout(this);
+		m_label      = new QLabel(
+		    tr("Press a key or mouse button.\nSpace and F1 are reserved; Esc cancels."), this);
+		m_label->setAlignment(Qt::AlignCenter);
+		layout->addWidget(m_label);
+	}
+
+	[[nodiscard]] const QString& Binding() const { return m_binding; }
+
+protected:
+	void keyPressEvent(QKeyEvent* event) override {
+		if (event->isAutoRepeat()) {
+			return;
+		}
+		if (event->key() == Qt::Key_Escape) {
+			reject();
+			return;
+		}
+		if (event->key() == Qt::Key_Space || event->key() == Qt::Key_F1) {
+			m_label->setText(tr("That key is reserved by the emulator."));
+			return;
+		}
+
+		m_binding = KeyName(*event);
+		if (!m_binding.isEmpty()) {
+			accept();
+		} else {
+			m_label->setText(tr("That key is not supported."));
+		}
+	}
+
+	void mousePressEvent(QMouseEvent* event) override {
+		switch (event->button()) {
+			case Qt::LeftButton: m_binding = QStringLiteral("Mouse:Left"); break;
+			case Qt::RightButton: m_binding = QStringLiteral("Mouse:Right"); break;
+			case Qt::MiddleButton: m_binding = QStringLiteral("Mouse:Middle"); break;
+			case Qt::BackButton: m_binding = QStringLiteral("Mouse:X1"); break;
+			case Qt::ForwardButton: m_binding = QStringLiteral("Mouse:X2"); break;
+			default: return;
+		}
+		accept();
+	}
+
+private:
+	QLabel* m_label = nullptr;
+	QString m_binding;
+};
+
+QHash<QString, QString> ParseMapping(const QStringList& mapping) {
+	QHash<QString, QString> result;
+	for (const auto& entry: mapping) {
+		const auto separator = entry.indexOf(QLatin1Char('='));
+		if (separator > 0 && separator + 1 < entry.size()) {
+			const auto binding = entry.mid(separator + 1);
+			for (auto item = result.begin(); item != result.end();) {
+				if (item.value().compare(binding, Qt::CaseInsensitive) == 0) {
+					item = result.erase(item);
+				} else {
+					++item;
+				}
+			}
+			result.insert(entry.left(separator), binding);
+		}
+	}
+	return result;
+}
+
+} // namespace
+
+InputMappingDialog::InputMappingDialog(const QStringList& mapping, QWidget* parent)
+    : QDialog(parent), m_customized(!mapping.isEmpty()) {
+	setWindowTitle(tr("Input Mapping"));
+	resize(440, 520);
+
+	auto* layout = new QVBoxLayout(this);
+	layout->addWidget(new QLabel(tr("Map keyboard or mouse input to DualSense buttons.\n"
+	                                "Defaults: left stick W/A/S/D, right stick T/F/G/H."),
+	                             this));
+
+	m_bindings = new QTreeWidget(this);
+	m_bindings->setColumnCount(2);
+	m_bindings->setHeaderLabels({tr("DualSense button"), tr("Host input")});
+	m_bindings->setRootIsDecorated(false);
+	m_bindings->setSelectionMode(QAbstractItemView::SingleSelection);
+	m_bindings->header()->setSectionResizeMode(0, QHeaderView::Stretch);
+	m_bindings->header()->setSectionResizeMode(1, QHeaderView::Stretch);
+	layout->addWidget(m_bindings);
+
+	const auto parsed = ParseMapping(mapping);
+	for (const auto& button: PAD_BUTTONS) {
+		auto* item = new QTreeWidgetItem(m_bindings);
+		item->setText(0, tr(button.label));
+		item->setData(0, Qt::UserRole, QString::fromLatin1(button.id));
+		SetBinding(item, m_customized ? parsed.value(QString::fromLatin1(button.id))
+		                              : QString::fromLatin1(button.default_binding));
+	}
+	m_bindings->setCurrentItem(m_bindings->topLevelItem(0));
+
+	auto* controls  = new QHBoxLayout;
+	m_change_button = new QPushButton(tr("Change..."), this);
+	m_clear_button  = new QPushButton(tr("Clear"), this);
+	auto* defaults  = new QPushButton(tr("Defaults"), this);
+	controls->addWidget(m_change_button);
+	controls->addWidget(m_clear_button);
+	controls->addWidget(defaults);
+	controls->addStretch();
+	layout->addLayout(controls);
+
+	auto* buttons = new QDialogButtonBox(QDialogButtonBox::Ok | QDialogButtonBox::Cancel, this);
+	layout->addWidget(buttons);
+
+	connect(m_bindings, &QTreeWidget::itemDoubleClicked, this,
+	        [this](QTreeWidgetItem*, int) { ChangeBinding(); });
+	connect(m_bindings, &QTreeWidget::itemSelectionChanged, this, [this]() { UpdateButtons(); });
+	connect(m_change_button, &QPushButton::clicked, this, [this]() { ChangeBinding(); });
+	connect(m_clear_button, &QPushButton::clicked, this, [this]() { ClearBinding(); });
+	connect(defaults, &QPushButton::clicked, this, [this]() { RestoreDefaults(); });
+	connect(buttons, &QDialogButtonBox::accepted, this, &QDialog::accept);
+	connect(buttons, &QDialogButtonBox::rejected, this, &QDialog::reject);
+	UpdateButtons();
+}
+
+QStringList InputMappingDialog::Mapping() const {
+	QStringList result;
+	if (!m_customized) {
+		return result;
+	}
+	for (int index = 0; index < m_bindings->topLevelItemCount(); index++) {
+		const auto* item    = m_bindings->topLevelItem(index);
+		const auto  binding = item->data(BINDING_COLUMN, Qt::UserRole).toString();
+		if (!binding.isEmpty()) {
+			result.append(item->data(0, Qt::UserRole).toString() + QLatin1Char('=') + binding);
+		}
+	}
+	return result;
+}
+
+void InputMappingDialog::ChangeBinding() {
+	auto* item = m_bindings->currentItem();
+	if (item == nullptr) {
+		return;
+	}
+
+	InputCaptureDialog dialog(this);
+	if (dialog.exec() != QDialog::Accepted) {
+		return;
+	}
+
+	for (int index = 0; index < m_bindings->topLevelItemCount(); index++) {
+		auto* other = m_bindings->topLevelItem(index);
+		if (other != item && other->data(BINDING_COLUMN, Qt::UserRole)
+		                             .toString()
+		                             .compare(dialog.Binding(), Qt::CaseInsensitive) == 0) {
+			SetBinding(other, {});
+		}
+	}
+	SetBinding(item, dialog.Binding());
+	m_customized = true;
+}
+
+void InputMappingDialog::ClearBinding() {
+	SetBinding(m_bindings->currentItem(), {});
+	m_customized = true;
+}
+
+void InputMappingDialog::RestoreDefaults() {
+	for (int index = 0; index < m_bindings->topLevelItemCount(); index++) {
+		SetBinding(m_bindings->topLevelItem(index),
+		           QString::fromLatin1(PAD_BUTTONS[index].default_binding));
+	}
+	m_customized = false;
+}
+
+void InputMappingDialog::SetBinding(QTreeWidgetItem* item, const QString& binding) {
+	if (item == nullptr) {
+		return;
+	}
+	item->setData(BINDING_COLUMN, Qt::UserRole, binding);
+	item->setText(BINDING_COLUMN, binding.isEmpty() ? tr("None") : binding);
+	UpdateButtons();
+}
+
+void InputMappingDialog::UpdateButtons() {
+	if (m_change_button == nullptr) {
+		return;
+	}
+	const auto* item = m_bindings->currentItem();
+	m_change_button->setEnabled(item != nullptr);
+	m_clear_button->setEnabled(item != nullptr &&
+	                           !item->data(BINDING_COLUMN, Qt::UserRole).toString().isEmpty());
+}

src/launcher/src/mainDialog.cpp

@@ -220,6 +220,9 @@ static QStringList CreateEmulatorArgs(const Configuration& info) {
 	args << "--printf-output-file" << info.printf_output_file;
 	args << "--profiler-direction" << EnumToText(info.profiler_direction);
 	args << "--spirv-debug-printf" << "false";
+	for (const auto& binding: info.host_input_mapping) {
+		args << "--keymap" << binding;
+	}
 	if (info.renderdoc_enabled) {
 		args << "--rd";
 	}
@@ -438,7 +441,10 @@ void MainDialogPrivate::Run() {
 
 	m_running_item->SetRunning(true);
 
-	m_main_dialog->RunInterpreter(&m_process, m_running_item->GetInfo());
+	Configuration info;
+	info.CopyFrom(m_running_item->GetInfo());
+	info.host_input_mapping = m_ui->widget->GetHostInputMapping();
+	m_main_dialog->RunInterpreter(&m_process, info);
 
 	Update();
 }

src/libs/agc.cpp

@@ -6,6 +6,7 @@
 #include "common/logging/log.h"
 #include "common/stringUtils.h"
 #include "common/virtualMemory.h"
+#include "graphics/guest_gpu/command_processor/commandProcessor.h"
 #include "graphics/guest_gpu/gpu_defs.h"
 #include "graphics/guest_gpu/graphicsRun.h"
 #include "graphics/guest_gpu/hardwareContext.h"
@@ -1444,29 +1445,72 @@ int KYTY_SYSV_ABI GraphicsSuspendPoint() {
 	return OK;
 }
 
-uint32_t* KYTY_SYSV_ABI GraphicsUnknownQj7QZpgr9Uw(CommandBuffer* buf, uint32_t mode,
-                                                   uint32_t value) {
-	PRINT_NAME();
-	// ResetContextState?
-	LOGF("\t mode  = 0x%08" PRIx32 "\n"
-	     "\t value = 0x%08" PRIx32 "\n",
-	     mode, value);
+[[nodiscard]] static constexpr uint32_t context_state_op_size_dw(uint32_t operation) {
+	switch (static_cast<ContextStateOperation>(operation)) {
+		case ContextStateOperation::Clear: return 5;
+		case ContextStateOperation::Push:
+		case ContextStateOperation::Pop: return 27;
+		case ContextStateOperation::PushClear: return 32;
+		default: return 0;
+	}
+}
 
-	if (buf == nullptr) {
+uint32_t* KYTY_SYSV_ABI GraphicsDcbContextStateOp(CommandBuffer* buf, uint32_t operation) {
+	PRINT_NAME();
+	LOGF("\t operation = 0x%08" PRIx32 "\n", operation);
+
+	const auto size_dw = context_state_op_size_dw(operation);
+	if (buf == nullptr || size_dw == 0) {
 		return nullptr;
 	}
 
 	buf->DbgDump();
 
-	auto* cmd = buf->AllocateDW(1);
+	uint32_t* first = nullptr;
+	const auto append = [&](uint32_t segment_size_dw) {
+		auto* segment = buf->AllocateDW(segment_size_dw);
+		if (segment == nullptr) {
+			return false;
+		}
+		std::fill_n(segment, segment_size_dw, 0u);
+		if (first == nullptr) {
+			first      = segment;
+			segment[0] = KYTY_PM4(segment_size_dw, Pm4::IT_NOP, Pm4::R_CONTEXT_STATE);
+			segment[1] = operation;
+		} else {
+			segment[0] = KYTY_PM4(segment_size_dw, Pm4::IT_NOP, Pm4::R_ZERO);
+		}
+		return true;
+	};
+	const auto append_save_restore = [&] {
+		// The native helper reserves 22 dwords, then emits 5-, 8-, and 9-dword packets.
+		return buf->ReserveDW(22) && append(5) && append(8) && append(9);
+	};
 
-	if (cmd == nullptr) {
-		return nullptr;
+	// Preserve native packet sizes, allocation boundaries, Push/Pop ordering, and
+	// returned first-packet address. The native packets reference process-private state buffers,
+	// so their net Cx operation is represented by the first HLE packet and type-3 NOPs.
+	bool complete = false;
+	switch (static_cast<ContextStateOperation>(operation)) {
+		case ContextStateOperation::Clear: complete = append(5); break;
+		case ContextStateOperation::Push:
+			complete = append_save_restore() && append(3) && append(2);
+			break;
+		case ContextStateOperation::Pop:
+			complete = append(3) && append_save_restore() && append(2);
+			break;
+		case ContextStateOperation::PushClear:
+			complete = append_save_restore() && append(3) && append(2) && append(5);
+			break;
+		default: break;
 	}
 
-	cmd[0] = 0x80000000u;
+	return complete ? first : nullptr;
+}
 
-	return cmd;
+uint64_t KYTY_SYSV_ABI GraphicsDcbContextStateOpGetSize(uint32_t operation) {
+	PRINT_NAME();
+	return static_cast<uint64_t>(context_state_op_size_dw(operation)) * sizeof(uint32_t);
 }
 
 uint64_t KYTY_SYSV_ABI GraphicsGetIsTrinityMode() {

src/libs/agc.h

@@ -67,8 +67,8 @@ int KYTY_SYSV_ABI GraphicsUnknownJumpPatchSetTarget(uint32_t* cmd, const volatil
 int KYTY_SYSV_ABI GraphicsJumpPatchSetTarget(uint32_t* cmd, const volatile uint32_t* target,
                                              uint32_t size_in_dwords);
 int KYTY_SYSV_ABI GraphicsSuspendPoint();
-uint32_t* KYTY_SYSV_ABI GraphicsUnknownQj7QZpgr9Uw(CommandBuffer* buf, uint32_t mode,
-                                                   uint32_t value);
+uint32_t* KYTY_SYSV_ABI GraphicsDcbContextStateOp(CommandBuffer* buf, uint32_t operation);
+uint64_t KYTY_SYSV_ABI  GraphicsDcbContextStateOpGetSize(uint32_t operation);
 uint64_t KYTY_SYSV_ABI  GraphicsGetIsTrinityMode();
 
 uint32_t KYTY_SYSV_ABI GraphicsDriverGetDefaultOwner();

src/libs/controller.cpp

@@ -85,8 +85,6 @@ private:
 
 static GameController* g_controller = nullptr;
 
-static constexpr int KEYBOARD_CONTROLLER_ID = -1000;
-
 static uint8_t pad_connected_count_to_u8(int connected_count) {
 	return static_cast<uint8_t>(connected_count > 255 ? 255 : connected_count);
 }
@@ -117,7 +115,7 @@ KYTY_SUBSYSTEM_INIT(Controller) {
 	EXIT_IF(g_controller != nullptr);
 
 	g_controller = new GameController;
-	g_controller->Connect(KEYBOARD_CONTROLLER_ID);
+	g_controller->Connect(HOST_INPUT_CONTROLLER_ID);
 }
 
 KYTY_SUBSYSTEM_UNEXPECTED_SHUTDOWN(Controller) {}
@@ -155,7 +153,7 @@ void GameController::CheckActive() {
 	if (!m_connected_ids.empty()) {
 		new_active_id = m_connected_ids[0];
 		for (const auto id: m_connected_ids) {
-			if (id != KEYBOARD_CONTROLLER_ID) {
+			if (id != HOST_INPUT_CONTROLLER_ID) {
 				new_active_id = id;
 				break;
 			}

src/libs/controller.h

@@ -9,6 +9,8 @@ namespace Libs::Controller {
 
 KYTY_SUBSYSTEM_DEFINE(Controller);
 
+constexpr int HOST_INPUT_CONTROLLER_ID = -1000;
+
 constexpr uint32_t PAD_BUTTON_L3        = 0x00000002;
 constexpr uint32_t PAD_BUTTON_R3        = 0x00000004;
 constexpr uint32_t PAD_BUTTON_OPTIONS   = 0x00000008;

src/libs/libGraphicsDriver.cpp

@@ -178,7 +178,8 @@ LIB_DEFINE(InitGraphicsDriver_1) {
 	LIB_FUNC("bxca0BK4FNg", Gen5::GraphicsUnknownJumpPatchSetTarget);
 	LIB_FUNC("2BS4EtAaF28", Gen5::GraphicsJumpPatchSetTarget);
 	LIB_FUNC("h9z6+0hEydk", Gen5::GraphicsSuspendPoint);
-	LIB_FUNC("qj7QZpgr9Uw", Gen5::GraphicsUnknownQj7QZpgr9Uw);
+	LIB_FUNC("qj7QZpgr9Uw", Gen5::GraphicsDcbContextStateOp);
+	LIB_FUNC("H6vHS5cidSA", Gen5::GraphicsDcbContextStateOpGetSize);
 	LIB_FUNC("BfBDZGbti7A", Gen5::GraphicsGetIsTrinityMode);
 	LIB_FUNC("dbOlWdppb4o", LibGen5::Graphics5UnknownDb);
 

src/libs/libKernel.cpp

@@ -13,6 +13,7 @@
 #include "kernel/memory.h"
 #include "kernel/pthread.h"
 #include "kernel/semaphore.h"
+#include "kernel/syncOnAddress.h"
 #include "libs/errno.h"
 #include "libs/libs.h"
 #include "libs/network.h"
@@ -2134,26 +2135,37 @@ uint64_t KYTY_SYSV_ABI cfwBSQyr5Ys(uint64_t a1, uint64_t a2, uint64_t a3, uint64
 	return 0;
 }
 
-uint64_t KYTY_SYSV_ABI KernelSyncOnAddressV1(uint64_t op, uint64_t address, uint64_t value,
-                                             uint64_t size, uint64_t timeout, uint64_t flags) {
-	static std::atomic_uint32_t log_count = 0;
-	const auto                  index     = log_count.fetch_add(1, std::memory_order_relaxed);
-
-	if (index < 16) {
-		LOGF("\t libkernel_sync_on_address_v1: op=0x%016" PRIx64 ", address=0x%016" PRIx64
-		     ", value=0x%016" PRIx64 ", size=0x%016" PRIx64 ", timeout=0x%016" PRIx64
-		     ", flags=0x%016" PRIx64 "\n",
-		     op, address, value, size, timeout, flags);
+static void LogExperimentalSyncOnAddress(std::atomic_bool& logged, const char* function_name) {
+	if (!logged.exchange(true, std::memory_order_relaxed)) {
+		::printf("WARNING: %s is experimental\n", function_name);
+		LOGF("WARNING: %s is experimental\n", function_name);
 	}
+}
 
-	if (op != 0 && address == 0 && value == 0 && size == 0) {
-		// This unsupported form is used as a yield/wait by some Unity jobs.
-		// SleepMicro() uses a sub-millisecond busy wait on Windows, which can
-		// pin every worker thread when the guest polls this path.
-		Common::Thread::Sleep(timeout == 0 ? 1 : 2);
-	}
+int KYTY_SYSV_ABI KernelSyncOnAddressWait(volatile uint32_t* address, uint32_t expected,
+                                          const uint32_t* timeout_micros) {
+	return LibKernel::SyncOnAddress::Wait32(address, expected, timeout_micros,
+	                                        LibKernel::KernelDispatchPendingSignalForCurrentThread);
+}
 
-	return 0;
+int KYTY_SYSV_ABI KernelSyncOnAddressWait32(volatile uint32_t* address, uint32_t expected,
+                                            const uint32_t* timeout_micros) {
+	static std::atomic_bool logged {false};
+	LogExperimentalSyncOnAddress(logged, "sceKernelSyncOnAddressWait32");
+	return LibKernel::SyncOnAddress::Wait32(address, expected, timeout_micros,
+	                                        LibKernel::KernelDispatchPendingSignalForCurrentThread);
+}
+
+int KYTY_SYSV_ABI KernelSyncOnAddressWait64(volatile uint64_t* address, uint64_t expected,
+                                            const uint32_t* timeout_micros) {
+	static std::atomic_bool logged {false};
+	LogExperimentalSyncOnAddress(logged, "sceKernelSyncOnAddressWait64");
+	return LibKernel::SyncOnAddress::Wait64(address, expected, timeout_micros,
+	                                        LibKernel::KernelDispatchPendingSignalForCurrentThread);
+}
+
+int KYTY_SYSV_ABI KernelSyncOnAddressWake(volatile void* address, int32_t count) {
+	return LibKernel::SyncOnAddress::Wake(address, count);
 }
 
 LIB_DEFINE(InitLibKernel_1_Posix) {
@@ -3405,8 +3417,10 @@ LIB_DEFINE(InitLibKernel_1) {
 	LIB_FUNC("Xjoosiw+XPI", LibKernel::KernelUuidCreate);
 	LIB_FUNC("DLORcroUqbc", LibKernel::KernelGetOpenPsId);
 	LIB_FUNC("zE-wXIZjLoM", LibKernel::KernelDebugRaiseExceptionOnReleaseMode);
-	LIB_FUNC("Hc4CaR6JBL0", Posix::KernelSyncOnAddressV1);
-	LIB_FUNC("q2y-wDIVWZA", Posix::KernelSyncOnAddressV1);
+	LIB_FUNC("Hc4CaR6JBL0", Posix::KernelSyncOnAddressWait);
+	LIB_FUNC("B2n8aDorSH4", Posix::KernelSyncOnAddressWait32);
+	LIB_FUNC("PZQhiiLXRFs", Posix::KernelSyncOnAddressWait64);
+	LIB_FUNC("q2y-wDIVWZA", Posix::KernelSyncOnAddressWake);
 
 	AddLibkernelUnityFunc(s, "Qhv5ARAoOEc",
 	                      reinterpret_cast<uint64_t>(LibKernel::KernelRemoveExceptionHandler),

src/main.cpp

@@ -64,6 +64,7 @@ static void PrintUsage() {
 	::printf("  --spirv-debug-printf <true|false>    Enable SPIR-V debug printf.\n");
 	::printf(
 	    "  --readback-linear-images <true|false> Read back writable linear images on submit.\n");
+	::printf("  --keymap <Pad=Key>                   DualSense mapping; may be repeated.\n");
 	::printf("  --rd                                 Enable RenderDoc capture.\n");
 }
 
@@ -247,6 +248,13 @@ static bool ParseArgs(int argc, char* argv[], RunOptions& options, bool& show_he
 				::printf("invalid boolean for %s: %s\n", arg.c_str(), value.c_str());
 				return false;
 			}
+		} else if (arg == "--keymap") {
+			const auto split = value.find('=');
+			if (split == std::string::npos || split == 0 || split + 1 == value.size()) {
+				::printf("invalid keymap: %s\n", value.c_str());
+				return false;
+			}
+			options.config.keymap.push_back(value);
 		} else {
 			::printf("unknown option: %s\n", arg.c_str());
 			return false;

src/utils.cmake

@@ -101,8 +101,6 @@ if(KYTY_CLANG_CL)
 	string(REGEX REPLACE "/MD" "/MT" CMAKE_C_FLAGS "${CMAKE_C_FLAGS}")
 			
 	set(KYTY_CPP_FLAGS "${KYTY_CPP_FLAGS} /utf-8 /Oy- /wd4244 /wd4305 /wd4800 /wd4345")
-	
-	add_link_options("$<$<CONFIG:RELEASE>:/OPT:NOREF>")
   
 	if(KYTY_WARNINGS_ARE_ERRORS)
 		#set(KYTY_CPP_FLAGS "${KYTY_CPP_FLAGS} /WX")

tests/ShaderRecompilerComputeTests.cpp

@@ -42,6 +42,7 @@
 #include "graphics/shader/rectListShader.h"
 #include "graphics/shader/shader.h"
 #include "kernel/memory.h"
+#include "libs/agc.h"
 #include "spirv-tools/libspirv.hpp"
 
 #if __has_include("graphics/host_gpu/renderer/renderTargetBarriers.h")
@@ -17653,6 +17654,136 @@ void CheckPm4StencilInfoValueLane(RenderContext& renderer) {
 	std::printf("[host]    %-32s ok\n", "Pm4StencilInfoValueLane");
 }
 
+void CheckPm4ContextStateOperations(RenderContext& renderer) {
+	GraphicsInitJmpTables();
+	CommandProcessor processor(renderer);
+	struct AgcCommandBufferLayout {
+		using Callback = KYTY_SYSV_ABI bool (*)(Gen5::CommandBuffer*, uint32_t, void*);
+
+		uint32_t* bottom;
+		uint32_t* top;
+		uint32_t* cursor_up;
+		uint32_t* cursor_down;
+		Callback  callback;
+		void*     user_data;
+		uint32_t  reserved_dw;
+	};
+	static_assert(offsetof(AgcCommandBufferLayout, cursor_up) == 0x10);
+	static_assert(offsetof(AgcCommandBufferLayout, reserved_dw) == 0x30);
+
+	HW::RenderControl original_control {};
+	original_control.depth_clear_enable   = true;
+	original_control.stencil_clear_enable = true;
+	processor.GetCtx().SetRenderControl(original_control);
+	processor.GetCtx().SetRenderTargetMask(0x1234abcd);
+	processor.GetUcfg().SetPrimitiveType(0x35);
+	g_hw_sh_indirect_func[Pm4::SPI_SHADER_PGM_LO_PS](processor, Pm4::SPI_SHADER_PGM_LO_PS,
+	                                                 0x00123456);
+	const auto shader_address = processor.GetShCtx().GetPs().ps_regs.data_addr;
+	const auto invoke = [&](ContextStateOperation operation) {
+		std::array<uint32_t, 32> packet {};
+		std::array<uint32_t, 6>  segment_sizes {};
+		AgcCommandBufferLayout   dcb {packet.data(), packet.data() + packet.size(), packet.data(),
+		                              packet.data() + packet.size(), nullptr, nullptr, 0};
+		const auto operation_value = static_cast<uint32_t>(operation);
+		size_t     segment_count  = 0;
+		switch (operation) {
+			case ContextStateOperation::Clear: segment_sizes = {5}; segment_count = 1; break;
+			case ContextStateOperation::Push:
+				segment_sizes = {5, 8, 9, 3, 2};
+				segment_count = 5;
+				break;
+			case ContextStateOperation::Pop:
+				segment_sizes = {3, 5, 8, 9, 2};
+				segment_count = 5;
+				break;
+			case ContextStateOperation::PushClear:
+				segment_sizes = {5, 8, 9, 3, 2, 5};
+				segment_count = 6;
+				break;
+		}
+		const auto size_dw =
+		    static_cast<size_t>(Gen5::GraphicsDcbContextStateOpGetSize(operation_value) / 4);
+		auto* emitted = Gen5::GraphicsDcbContextStateOp(
+		    reinterpret_cast<Gen5::CommandBuffer*>(&dcb), operation_value);
+		bool   packet_matches = true;
+		size_t segment_offset = 0;
+		for (size_t i = 0; i < segment_count; i++) {
+			const auto segment_size = segment_sizes[i];
+			const auto selector = (i == 0 ? Pm4::R_CONTEXT_STATE : Pm4::R_ZERO);
+			packet_matches &=
+			    packet[segment_offset] == KYTY_PM4(segment_size, Pm4::IT_NOP, selector);
+			for (size_t j = 1; j < segment_size; j++) {
+				packet_matches &= packet[segment_offset + j] ==
+				                  (i == 0 && j == 1 ? operation_value : 0u);
+			}
+			segment_offset += segment_size;
+		}
+		Require("Pm4ContextState", "HLE packet",
+		        emitted == packet.data() && dcb.cursor_up == packet.data() + size_dw &&
+		            segment_offset == size_dw && packet_matches,
+		        "context-state HLE packet does not match its real allocation size");
+		Pm4Execution execution;
+		return processor.Process(execution, packet.data(), size_dw);
+	};
+
+	Require("Pm4ContextState", "push-clear",
+	        invoke(ContextStateOperation::PushClear) == Pm4ProcessResult::Complete &&
+	            !processor.GetCtx().GetRenderControl().depth_clear_enable &&
+	            !processor.GetCtx().GetRenderControl().stencil_clear_enable &&
+	            processor.GetCtx().GetRenderTargetMask() == HW::Context {}.GetRenderTargetMask() &&
+	            processor.GetUcfg().GetPrimType() == 0x35 &&
+	            processor.GetShCtx().GetPs().ps_regs.data_addr == shader_address,
+	        "push-clear did not save and clear only Cx state");
+
+	processor.GetCtx().SetRenderTargetMask(0x55aa55aa);
+	Require("Pm4ContextState", "pop",
+	        invoke(ContextStateOperation::Pop) == Pm4ProcessResult::Complete &&
+	            processor.GetCtx().GetRenderControl().depth_clear_enable &&
+	            processor.GetCtx().GetRenderControl().stencil_clear_enable &&
+	            processor.GetCtx().GetRenderTargetMask() == 0x1234abcd,
+	        "pop did not restore the complete saved Cx state");
+
+	Require("Pm4ContextState", "push",
+	        invoke(ContextStateOperation::Push) == Pm4ProcessResult::Complete &&
+	            processor.GetCtx().GetRenderTargetMask() == 0x1234abcd,
+	        "push changed live Cx state");
+	processor.GetCtx().SetRenderTargetMask(0x01020304);
+	Require("Pm4ContextState", "push-pop restore",
+	        invoke(ContextStateOperation::Pop) == Pm4ProcessResult::Complete &&
+	            processor.GetCtx().GetRenderTargetMask() == 0x1234abcd,
+	        "ordinary push/pop did not restore Cx state");
+
+	Require("Pm4ContextState", "clear",
+	        invoke(ContextStateOperation::Clear) == Pm4ProcessResult::Complete &&
+	            !processor.GetCtx().GetRenderControl().depth_clear_enable &&
+	            processor.GetCtx().GetRenderTargetMask() == HW::Context {}.GetRenderTargetMask() &&
+	            processor.GetUcfg().GetPrimType() == 0x35 &&
+	            processor.GetShCtx().GetPs().ps_regs.data_addr == shader_address,
+	        "clear reset state outside the Cx register domain");
+
+	processor.GetCtx().SetRenderTargetMask(0xabcdef01);
+	Require("Pm4ContextState", "push before processor reset",
+	        invoke(ContextStateOperation::Push) == Pm4ProcessResult::Complete,
+	        "push before reset failed");
+	processor.Reset();
+	processor.GetCtx().SetRenderTargetMask(0x76543210);
+	Require("Pm4ContextState", "processor reset discards push",
+	        invoke(ContextStateOperation::Push) == Pm4ProcessResult::Complete &&
+	            invoke(ContextStateOperation::Pop) == Pm4ProcessResult::Complete &&
+	            processor.GetCtx().GetRenderTargetMask() == 0x76543210,
+	        "processor reset retained an invalid saved Cx state");
+
+	Require("Pm4ContextState", "HLE packet size",
+	        Gen5::GraphicsDcbContextStateOpGetSize(0) == 20 &&
+	            Gen5::GraphicsDcbContextStateOpGetSize(1) == 108 &&
+	            Gen5::GraphicsDcbContextStateOpGetSize(2) == 108 &&
+	            Gen5::GraphicsDcbContextStateOpGetSize(3) == 128 &&
+	            Gen5::GraphicsDcbContextStateOpGetSize(4) == 0,
+	        "context-state HLE sizes do not match libSceAgc");
+	std::printf("[host]    %-32s ok\n", "Pm4ContextState");
+}
+
 void CheckPm4WaitResume(RenderContext& renderer) {
 	GraphicsInitJmpTables();
 	CommandProcessor processor(renderer);
@@ -17789,6 +17920,11 @@ int main(int argc, char** argv) {
 		vulkan.CheckGpuCommandLane();
 		return 0;
 	}
+	if (argc == 2 && std::strcmp(argv[1], "--context-state-only") == 0) {
+		VulkanHarness vulkan;
+		CheckPm4ContextStateOperations(vulkan.RuntimeRenderer());
+		return 0;
+	}
 	if (argc == 2 && std::strcmp(argv[1], "--image-overlap-only") == 0) {
 		CheckDepthAttachmentWrites();
 		CheckDynamicRenderingState();
@@ -17923,6 +18059,7 @@ int main(int argc, char** argv) {
 	CheckVulkan13FeatureRequirements();
 	CheckPm4AcquireMemNoOp(vulkan.RuntimeRenderer());
 	CheckPm4StencilInfoValueLane(vulkan.RuntimeRenderer());
+	CheckPm4ContextStateOperations(vulkan.RuntimeRenderer());
 	CheckPm4WaitResume(vulkan.RuntimeRenderer());
 	CheckPm4CeCompletion(vulkan.RuntimeRenderer());
 	CheckEmbeddedFetchVertexOffset();

tests/SyncOnAddressTests.cpp

@@ -0,0 +1,269 @@
+#include "kernel/syncOnAddress.h"
+
+#include "libs/errno.h"
+
+#include <atomic>
+#include <chrono>
+#include <climits>
+#include <cstdio>
+#include <cstdlib>
+#include <thread>
+#include <vector>
+
+namespace {
+
+using Libs::LibKernel::SyncOnAddress::Wait32;
+using Libs::LibKernel::SyncOnAddress::Wait64;
+using Libs::LibKernel::SyncOnAddress::Wake;
+
+std::atomic<int> g_signal_poll_count{0};
+
+void Check(bool value, const char *text) {
+  if (!value) {
+    std::fprintf(stderr, "SyncOnAddressTests: failed: %s\n", text);
+    std::abort();
+  }
+}
+
+template <typename T> void Store(T *address, T value) {
+  std::atomic_ref<T>(*address).store(value, std::memory_order_release);
+}
+
+void CountSignalPoll() {
+  g_signal_poll_count.fetch_add(1, std::memory_order_relaxed);
+}
+
+void TestInvalidAddress() {
+  uint32_t timeout = 1;
+  Check(Wait32(nullptr, 0, &timeout) == Libs::LibKernel::KERNEL_ERROR_EINVAL,
+        "wait32 rejects a null address");
+  Check(Wait64(nullptr, 0, &timeout) == Libs::LibKernel::KERNEL_ERROR_EINVAL,
+        "wait64 rejects a null address");
+  Check(Wake(nullptr, 1) == Libs::LibKernel::KERNEL_ERROR_EINVAL,
+        "wake rejects a null address");
+
+  alignas(uint64_t) uint8_t bytes[16] = {};
+  auto *misaligned = reinterpret_cast<uint32_t *>(bytes + 1);
+  Check(Wait32(misaligned, 0, &timeout) == Libs::LibKernel::KERNEL_ERROR_EINVAL,
+        "wait32 rejects a misaligned address");
+  Check(Wait64(reinterpret_cast<uint64_t *>(bytes + 4), 0, &timeout) ==
+            Libs::LibKernel::KERNEL_ERROR_EINVAL,
+        "wait64 rejects a misaligned address");
+  Check(Wake(misaligned, 1) == Libs::LibKernel::KERNEL_ERROR_EINVAL,
+        "wake rejects a misaligned address");
+  uint64_t aligned = 0;
+  Check(Wake(&aligned, -1) == Libs::LibKernel::KERNEL_ERROR_EINVAL,
+        "wake rejects a negative count");
+}
+
+void TestMismatchReturnsImmediately() {
+  uint32_t word = 7;
+  uint64_t word64 = UINT64_C(0x100000000);
+  uint32_t timeout = 500000;
+  const auto start = std::chrono::steady_clock::now();
+  Check(Wait32(&word, 6, &timeout, CountSignalPoll) == OK,
+        "mismatch succeeds without parking");
+  Check(Wait64(&word64, 0, &timeout, CountSignalPoll) == OK,
+        "wait64 compares all 64 bits");
+  Check(std::chrono::steady_clock::now() - start <
+            std::chrono::milliseconds(100),
+        "mismatched value is a fast path");
+  Check(g_signal_poll_count.load(std::memory_order_relaxed) >= 2,
+        "mismatch remains a guest signal safe-point");
+}
+
+void TestTimeout() {
+  uint32_t word = 0;
+  uint32_t timeout = 20000;
+  const auto start = std::chrono::steady_clock::now();
+  Check(Wait32(&word, 0, &timeout) == Libs::LibKernel::KERNEL_ERROR_ETIMEDOUT,
+        "matching value times out");
+  const auto elapsed = std::chrono::steady_clock::now() - start;
+  Check(elapsed >= std::chrono::milliseconds(10),
+        "timeout does not return too early");
+  Check(elapsed < std::chrono::milliseconds(500), "timeout remains bounded");
+
+  timeout = 0;
+  Check(Wait32(&word, 0, &timeout) == Libs::LibKernel::KERNEL_ERROR_ETIMEDOUT,
+        "zero timeout polls a matching value");
+  word = 1;
+  Check(Wait32(&word, 0, &timeout) == OK,
+        "zero timeout succeeds for a mismatched value");
+}
+
+void TestValueChangeAndWake() {
+  uint64_t word = 0;
+  uint32_t timeout = 1000000;
+  std::atomic<bool> ready{false};
+  int result = Libs::LibKernel::KERNEL_ERROR_ETIMEDOUT;
+
+  std::thread waiter([&] {
+    ready.store(true, std::memory_order_release);
+    result = Wait64(&word, 0, &timeout);
+  });
+  while (!ready.load(std::memory_order_acquire)) {
+    std::this_thread::yield();
+  }
+  Store(&word, UINT64_C(0x100000000));
+  Check(Wake(&word, 1) == OK, "wake-one succeeds");
+  waiter.join();
+  Check(result == OK, "a value change plus wake releases the waiter");
+}
+
+void TestWakeOneThenAll() {
+  constexpr int WAITER_COUNT = 4;
+  uint32_t word = 0;
+  uint32_t timeout = 1000000;
+  std::atomic<int> ready{0};
+  std::atomic<int> returned{0};
+  int results[WAITER_COUNT] = {};
+  std::vector<std::thread> waiters;
+
+  for (int i = 0; i < WAITER_COUNT; i++) {
+    waiters.emplace_back([&, i] {
+      ready.fetch_add(1, std::memory_order_release);
+      results[i] = Wait32(&word, 0, &timeout);
+      returned.fetch_add(1, std::memory_order_release);
+    });
+  }
+  while (ready.load(std::memory_order_acquire) != WAITER_COUNT) {
+    std::this_thread::yield();
+  }
+  std::this_thread::sleep_for(std::chrono::milliseconds(50));
+  Check(Wake(&word, 1) == OK, "wake-one succeeds with multiple waiters");
+  const auto one_deadline =
+      std::chrono::steady_clock::now() + std::chrono::milliseconds(500);
+  while (returned.load(std::memory_order_acquire) == 0 &&
+         std::chrono::steady_clock::now() < one_deadline) {
+    std::this_thread::yield();
+  }
+  Check(returned.load(std::memory_order_acquire) == 1,
+        "wake-one releases exactly one waiter");
+
+  Check(Wake(&word, 2) == OK, "wake-two succeeds");
+  const auto two_deadline =
+      std::chrono::steady_clock::now() + std::chrono::milliseconds(500);
+  while (returned.load(std::memory_order_acquire) < 3 &&
+         std::chrono::steady_clock::now() < two_deadline) {
+    std::this_thread::yield();
+  }
+  Check(returned.load(std::memory_order_acquire) == 3,
+        "wake-two releases exactly two more waiters");
+
+  Check(Wake(&word, INT_MAX) == OK, "wake-all succeeds");
+  for (auto &waiter : waiters) {
+    waiter.join();
+  }
+  Check(returned.load(std::memory_order_acquire) == WAITER_COUNT,
+        "wake-all releases the remaining waiters");
+  for (int result : results) {
+    Check(result == OK, "explicitly woken waiters return success");
+  }
+}
+
+void TestAddressesAreIsolated() {
+  uint32_t first = 0;
+  uint32_t second = 0;
+  uint32_t first_timeout = 1000000;
+  uint32_t second_timeout = 1000000;
+  std::atomic<int> ready{0};
+  std::atomic<bool> first_returned{false};
+  std::atomic<bool> second_returned{false};
+  int first_result = 0;
+  int second_result = 0;
+
+  std::thread first_waiter([&] {
+    ready.fetch_add(1, std::memory_order_release);
+    first_result = Wait32(&first, 0, &first_timeout);
+    first_returned.store(true, std::memory_order_release);
+  });
+  std::thread second_waiter([&] {
+    ready.fetch_add(1, std::memory_order_release);
+    second_result = Wait32(&second, 0, &second_timeout);
+    second_returned.store(true, std::memory_order_release);
+  });
+  while (ready.load(std::memory_order_acquire) != 2) {
+    std::this_thread::yield();
+  }
+  std::this_thread::sleep_for(std::chrono::milliseconds(50));
+  Check(Wake(&first, 1) == OK, "first address wakes");
+  std::this_thread::sleep_for(std::chrono::milliseconds(50));
+  Check(first_returned.load(std::memory_order_acquire),
+        "first address waiter returned");
+  Check(!second_returned.load(std::memory_order_acquire),
+        "waking one address does not release another address");
+  Check(Wake(&second, 1) == OK, "second address wakes");
+
+  first_waiter.join();
+  second_waiter.join();
+  Check(first_result == OK && second_result == OK,
+        "isolated waiters return success");
+}
+
+void TestCompareRegisterWakeRace() {
+  for (int i = 0; i < 100; i++) {
+    uint32_t word = 0;
+    uint32_t timeout = 500000;
+    std::atomic<bool> ready{false};
+    int result = Libs::LibKernel::KERNEL_ERROR_ETIMEDOUT;
+    std::thread waiter([&] {
+      ready.store(true, std::memory_order_release);
+      result = Wait32(&word, 0, &timeout);
+    });
+    while (!ready.load(std::memory_order_acquire)) {
+      std::this_thread::yield();
+    }
+    Store(&word, uint32_t{1});
+    (void)Wake(&word, 1);
+    waiter.join();
+    Check(result == OK, "compare/register/wake race never loses progress");
+  }
+}
+
+void TestWakeZeroIsNoOp() {
+  constexpr int WAITER_COUNT = 2;
+  uint32_t word = 0;
+  uint32_t timeout = 1000000;
+  std::atomic<int> ready{0};
+  std::atomic<int> returned{0};
+  int results[WAITER_COUNT] = {};
+  std::vector<std::thread> waiters;
+
+  for (int i = 0; i < WAITER_COUNT; i++) {
+    waiters.emplace_back([&, i] {
+      ready.fetch_add(1, std::memory_order_release);
+      results[i] = Wait32(&word, 0, &timeout);
+      returned.fetch_add(1, std::memory_order_release);
+    });
+  }
+  while (ready.load(std::memory_order_acquire) != WAITER_COUNT) {
+    std::this_thread::yield();
+  }
+  std::this_thread::sleep_for(std::chrono::milliseconds(50));
+  Check(Wake(&word, 0) == OK, "zero-count wake succeeds");
+  std::this_thread::sleep_for(std::chrono::milliseconds(20));
+  Check(returned.load(std::memory_order_acquire) == 0,
+        "zero-count wake releases no waiters");
+  Check(Wake(&word, INT_MAX) == OK, "wake-all succeeds after zero-count wake");
+  for (auto &waiter : waiters) {
+    waiter.join();
+  }
+  for (int result : results) {
+    Check(result == OK, "wake-all releases waiters after zero-count no-op");
+  }
+}
+
+} // namespace
+
+int main() {
+  TestInvalidAddress();
+  TestMismatchReturnsImmediately();
+  TestTimeout();
+  TestValueChangeAndWake();
+  TestWakeOneThenAll();
+  TestAddressesAreIsolated();
+  TestCompareRegisterWakeRace();
+  TestWakeZeroIsNoOp();
+  std::printf("SyncOnAddressTests: all passed\n");
+  return 0;
+}