input: correct gyro orientation drift with the accelerometer (#899)

* input: correct gyro orientation drift with the accelerometer

* input: simplify gyro drift correction and reset gravity on reconnect

---------

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

CMakeLists.txt

@@ -500,6 +500,10 @@ add_executable(save_data_memory_tests EXCLUDE_FROM_ALL
 target_link_libraries(save_data_memory_tests common fmt::fmt)
 target_include_directories(save_data_memory_tests PRIVATE ${inc_headers})
 
+add_executable(png_enc_tests EXCLUDE_FROM_ALL "${KYTY_TESTS_DIR}/PngEncTests.cpp")
+target_link_libraries(png_enc_tests common fmt::fmt)
+target_include_directories(png_enc_tests PRIVATE ${inc_headers})
+
 add_executable(ces_tests EXCLUDE_FROM_ALL "${KYTY_TESTS_DIR}/CesTests.cpp")
 target_link_libraries(ces_tests common fmt::fmt SDL3::SDL3-static)
 target_include_directories(ces_tests PRIVATE ${inc_headers})
@@ -605,6 +609,7 @@ if(BUILD_TESTING)
 	add_test(NAME ngs2_sampler COMMAND $<TARGET_FILE:ngs2_sampler_tests>)
 	add_test(NAME pad_haptics COMMAND $<TARGET_FILE:pad_haptics_tests>)
 	add_test(NAME save_data_memory COMMAND $<TARGET_FILE:save_data_memory_tests>)
+	add_test(NAME png_enc COMMAND $<TARGET_FILE:png_enc_tests>)
 	add_test(NAME ces COMMAND $<TARGET_FILE:ces_tests>)
 	add_test(NAME save_data_dialog COMMAND $<TARGET_FILE:save_data_dialog_tests>)
 	add_test(NAME http_uri_parse COMMAND $<TARGET_FILE:http_uri_parse_tests>)
@@ -680,6 +685,7 @@ if(BUILD_TESTING)
 		ngs2_sampler_tests
 		pad_haptics_tests
 		save_data_memory_tests
+		png_enc_tests
 		ces_tests
 		save_data_dialog_tests
 		http_uri_parse_tests

src/common/emulatorConfig.cpp

@@ -47,6 +47,10 @@ const std::string& GetAudioInputDevice() {
 	return g_config->audio_input_device;
 }
 
+const std::optional<ControllerColor>& GetControllerColor() {
+	return g_config->controller_color;
+}
+
 PresentMode GetPresentMode() {
 	return g_config->present_mode;
 }

src/common/emulatorConfig.h

@@ -4,7 +4,9 @@
 #include "common/common.h"
 
 #include <cstddef>
+#include <array>
 #include <filesystem>
+#include <optional>
 #include <string>
 #include <vector>
 
@@ -26,6 +28,7 @@ enum class LogDirection { Silent, Console, File };
 enum class PresentMode { Fifo, Mailbox, Immediate };
 
 using Keymap = std::vector<std::string>;
+using ControllerColor = std::array<uint8_t, 3>;
 
 constexpr uint32_t DEFAULT_CONSOLE_LANGUAGE = 1;
 constexpr uint32_t MAX_CONSOLE_LANGUAGE     = 29;
@@ -44,6 +47,7 @@ struct ConfigOptions {
 	std::string            user_name                   = "Kyty";
 	int32_t                user_id                     = DEFAULT_USER_ID;
 	std::string            audio_input_device;
+	std::optional<ControllerColor> controller_color;
 	PresentMode            present_mode                = PresentMode::Mailbox;
 	int32_t                gpu_index                   = -1;
 	bool                   fullscreen_enabled          = false;
@@ -81,6 +85,7 @@ uint32_t GetScreenHeight();
 const std::string& GetUserName();
 int32_t  GetUserId();
 const std::string& GetAudioInputDevice();
+const std::optional<ControllerColor>& GetControllerColor();
 PresentMode GetPresentMode();
 int32_t GetGpuIndex();
 bool     FullscreenEnabled();

src/graphics/guest_gpu/command_processor/commandProcessor.h

@@ -62,7 +62,6 @@ public:
 
 	void            BufferInit();
 	void            BufferFlush();
-	void            BufferFlushAndWait();
 	void            BufferWait();
 	HW::Context&    GetCtx() { return m_ctx; }
 	HW::UserConfig& GetUcfg() { return m_ucfg; }

src/graphics/guest_gpu/graphicsRun.cpp

@@ -35,7 +35,6 @@ namespace Libs::Graphics {
 
 static thread_local CommandProcessor* g_current_processor = nullptr;
 static thread_local Pm4Execution*     g_current_execution = nullptr;
-static thread_local bool              g_gpu_mutex_owned   = false;
 static thread_local bool              g_gpu_thread        = false;
 static thread_local GuestGpu*         g_gpu_state         = nullptr;
 
@@ -54,27 +53,6 @@ struct DrawIndexedIndirectArgs {
 	uint32_t start_instance_location;
 };
 
-class GpuMutexLock final {
-public:
-	explicit GpuMutexLock(Common::Mutex& mutex): m_mutex(mutex) {
-		if (g_gpu_mutex_owned) {
-			EXIT("recursive GPU mutex acquisition\n");
-		}
-		g_gpu_mutex_owned = true;
-		m_mutex.Lock();
-	}
-	~GpuMutexLock() {
-		if (!g_gpu_mutex_owned) {
-			EXIT("invalid GPU mutex release\n");
-		}
-		m_mutex.Unlock();
-		g_gpu_mutex_owned = false;
-	}
-
-private:
-	Common::Mutex& m_mutex;
-};
-
 static bool GraphicsRunDebugDumpEnabled() {
 	return Config::GraphicsDebugDumpEnabled() &&
 	       Config::GetPrintfDirection() != Config::LogDirection::Silent;
@@ -84,6 +62,7 @@ GuestGpu::GuestGpu(RenderContext& renderer): m_renderer(renderer) {
 	EXIT_NOT_IMPLEMENTED(!Common::Thread::IsMainThread());
 	GraphicsInitJmpTables();
 	m_gfx_cp = std::make_unique<CommandProcessor>(renderer, 0);
+	m_gfx_cp->Reset();
 	m_thread = std::jthread(ThreadRun, this);
 }
 
@@ -160,20 +139,16 @@ void GuestGpu::Submit(std::span<const uint32_t> draw_commands,
 	if (draw_commands.empty()) {
 		return;
 	}
-	GpuMutexLock lock(m_submission_mutex);
-	Submission   submission;
+	Submission submission;
 	submission.type              = SubmissionType::Graphics;
 	submission.queue_id          = 0;
 	submission.commands          = draw_commands;
 	submission.constant_commands = constant_commands;
-	submission.reset_processor   = m_graphics_done;
-	m_graphics_done              = false;
 	Enqueue(std::move(submission));
 }
 
 void GuestGpu::SubmitCompute(uint32_t queue, std::span<const uint32_t> commands) {
 	EXIT_IF(commands.empty());
-	GpuMutexLock lock(m_submission_mutex);
 
 	EXIT_NOT_IMPLEMENTED(queue < ComputeQueueBase || queue >= ComputeQueueBase + ComputeQueueCount);
 
@@ -186,22 +161,21 @@ void GuestGpu::SubmitCompute(uint32_t queue, std::span<const uint32_t> commands)
 }
 
 void GuestGpu::SubmitFlipPreparation(uint64_t request_id) {
-	GpuMutexLock lock(m_submission_mutex);
-	Submission   submission;
+	Submission submission;
 	submission.type            = SubmissionType::FlipPreparation;
 	submission.queue_id        = 0;
-	submission.reset_processor = m_graphics_done;
 	submission.flip_request_id = request_id;
-	m_graphics_done            = false;
 	Enqueue(std::move(submission));
 }
 
-void GuestGpu::Done() {
-	GpuMutexLock lock(m_submission_mutex);
-	if (!IsGpuThread()) {
-		WaitForIdle();
-	}
-	m_graphics_done = true;
+void GuestGpu::SuspendPoint() {
+	EXIT_IF(IsGpuThread() || CommandScheduler::InDeferredOperation());
+	// Do not hold a queue lock while waiting: asynchronous work may be needed to
+	// finish the preceding graphics frame. The first point returns immediately.
+	m_suspend_point_ready->acquire();
+	Submission submission;
+	submission.type = SubmissionType::SuspendPoint;
+	Enqueue(std::move(submission));
 	m_done_num++;
 }
 
@@ -229,6 +203,9 @@ void CommandProcessor::Reset() {
 	m_context_state_pushed             = false;
 	m_index_type_and_size              = 0;
 	m_index_buffer_size                = 0;
+	m_index_base_addr                  = 0;
+	m_num_instances                    = 1;
+	m_predicate_skip                   = false;
 	m_user_data_marker                 = HW::UserSgprType::Unknown;
 	m_draw_indirect_args_base_addr     = 0;
 	m_dispatch_indirect_args_base_addr = 0;
@@ -268,10 +245,6 @@ void CommandProcessor::BufferFlush() {
 	GetScheduler().Flush();
 }
 
-void CommandProcessor::BufferFlushAndWait() {
-	GetScheduler().FlushAndWait();
-}
-
 void CommandProcessor::BufferWait() {
 	BufferInit();
 	GetScheduler().Finish();
@@ -459,6 +432,7 @@ void GuestGpu::Enqueue(Submission submission) {
 }
 
 void GuestGpu::WaitForIdle() {
+	EXIT_IF(IsGpuThread() || CommandScheduler::InDeferredOperation());
 	Common::LockGuard lock(m_queue_mutex);
 	while (m_processing || !m_commands.empty() || m_submission_count != 0) {
 		m_idle.Wait(&m_queue_mutex);
@@ -564,11 +538,7 @@ void GuestGpu::ThreadRun(void* data) {
 
 bool GuestGpu::Process(Submission& submission) {
 	const bool first_slice = !submission.started;
-	auto& cp = GetProcessor(submission.queue_id);
-
-	if (first_slice && submission.reset_processor) {
-		cp.Reset();
-	}
+	auto&      cp          = GetProcessor(submission.queue_id);
 
 	if (first_slice) {
 		submission.started = true;
@@ -645,6 +615,13 @@ bool GuestGpu::Process(Submission& submission) {
 			m_renderer.RunGarbageCollector();
 			cp.PrepareCpuFlip(submission.flip_request_id);
 			break;
+		case SubmissionType::SuspendPoint:
+			cp.EmitGlobalBarrier();
+			m_renderer.GetCommandScheduler().DeferPriorityOperation(
+			    [ready = m_suspend_point_ready] { ready->release(); });
+			cp.BufferFlush();
+			cp.Reset();
+			break;
 	}
 
 	return complete;
@@ -849,9 +826,7 @@ void CommandProcessor::SetPredication(uint32_t condition, uint32_t op, uint32_t
 			}
 		} break;
 		case 0x03:
-			if (wait_op != 0) {
-				BufferFlushAndWait();
-			}
+			// The wait selector applies only to Z-pass query readiness.
 			EXIT_NOT_IMPLEMENTED(address == nullptr);
 			value = *reinterpret_cast<const volatile uint64_t*>(address);
 			break;
@@ -1034,9 +1009,6 @@ void CommandProcessor::DispatchDirect(uint32_t thread_group_x, uint32_t thread_g
 	m_sh_ctx.SetCsWaveSize(Pm4::ComputeWaveSize(mode));
 
 	uint32_t frame_num = 0;
-	// uint32_t local_x   = 1;
-	// uint32_t local_y   = 1;
-	// uint32_t local_z   = 1;
 
 	{
 		frame_num = m_renderer.GetGpu().GetFrameNum();
@@ -1058,35 +1030,9 @@ void CommandProcessor::DispatchDirect(uint32_t thread_group_x, uint32_t thread_g
 			}
 		}
 
-		const auto& cs = m_sh_ctx.GetCs().cs_regs;
-		// local_x        = std::max(cs.num_thread_x, 1u);
-		// local_y        = std::max(cs.num_thread_y, 1u);
-		// local_z        = std::max(cs.num_thread_z, 1u);
 		m_renderer.GetRenderExecutor().DispatchDirect(m_submit_id, CurrentBuffer(), thread_group_x,
 		                                              thread_group_y, thread_group_z, mode);
 	}
-
-	/*constexpr uint32_t DispatchInitiatorUseThreadDimensions = 1u << 5u;
-	auto               group_count = [](uint32_t threads, uint32_t group_size) {
-	    return (threads == 0
-	                ? 0u
-	                : (threads + std::max(group_size, 1u) - 1u) / std::max(group_size, 1u));
-	};
-
-	auto groups_x = thread_group_x;
-	auto groups_y = thread_group_y;
-	auto groups_z = thread_group_z;
-	if ((mode & DispatchInitiatorUseThreadDimensions) != 0) {
-	    groups_x = group_count(thread_group_x, local_x);
-	    groups_y = group_count(thread_group_y, local_y);
-	    groups_z = group_count(thread_group_z, local_z);
-	}
-
-	const uint64_t invocations =
-	    static_cast<uint64_t>(groups_x) * groups_y * groups_z * local_x * local_y * local_z;
-	if (invocations != 0) {
-	    BufferFlushAndWait();
-	}*/
 }
 
 void CommandProcessor::DispatchIndirect(uint64_t args_addr, uint32_t mode) {

src/graphics/guest_gpu/graphicsRun.h

@@ -13,6 +13,7 @@
 #include <deque>
 #include <memory>
 #include <mutex>
+#include <semaphore>
 #include <span>
 #include <thread>
 
@@ -32,11 +33,14 @@ public:
 	void               SendCommandSync(Common::UniqueFunction<void>&& command);
 
 	// Submitted command memory is borrowed and must remain valid until GPU execution completes.
-	void              Submit(std::span<const uint32_t> draw_commands,
-	                         std::span<const uint32_t> constant_commands);
-	void              SubmitCompute(uint32_t queue, std::span<const uint32_t> commands);
-	void              SubmitFlipPreparation(uint64_t request_id);
-	void              Done();
+	void Submit(std::span<const uint32_t> draw_commands,
+	            std::span<const uint32_t> constant_commands);
+	void SubmitCompute(uint32_t queue, std::span<const uint32_t> commands);
+	void SubmitFlipPreparation(uint64_t request_id);
+	// Insert an ordered graphics drain; only a previous suspend point can block the caller.
+	void SuspendPoint();
+	// Wait for guest command processing, including all compute queues (not native GPU completion).
+	void              WaitForIdle();
 	[[nodiscard]] int GetFrameNum() const;
 
 	[[nodiscard]] static bool IsGpuThread() noexcept;
@@ -48,7 +52,7 @@ private:
 	static constexpr uint32_t ComputeQueueBase     = 0x20;
 	static constexpr uint32_t QueueCount           = 1 + ComputeQueueCount;
 
-	enum class SubmissionType { Graphics, Compute, FlipPreparation };
+	enum class SubmissionType { Graphics, Compute, FlipPreparation, SuspendPoint };
 
 	struct Submission {
 		SubmissionType            type     = SubmissionType::Graphics;
@@ -57,7 +61,6 @@ private:
 		std::span<const uint32_t> constant_commands;
 		Pm4Execution              command_execution;
 		Pm4Execution              constant_execution;
-		bool                      reset_processor   = false;
 		bool                      started           = false;
 		bool                      command_complete  = false;
 		bool                      constant_complete = false;
@@ -66,14 +69,12 @@ private:
 	};
 
 	void              Enqueue(Submission submission);
-	void              WaitForIdle();
 	void              ProcessCommands();
 	bool              Process(Submission& submission);
 	static void       ThreadRun(void* data);
 	CommandProcessor& GetProcessor(uint32_t queue_id);
 
 	RenderContext&                                 m_renderer;
-	Common::Mutex                                  m_submission_mutex;
 	Common::Mutex                                  m_queue_mutex;
 	std::mutex                                     m_shutdown_mutex;
 	Common::CondVar                                m_work_available;
@@ -84,10 +85,12 @@ private:
 	uint32_t                                       m_next_queue        = 0;
 	uint32_t                                       m_submission_count  = 0;
 	bool                                           m_processing        = false;
-	bool                                           m_graphics_done     = true;
 	bool                                           m_accepting         = true;
 	bool                                           m_stopping          = false;
 	bool                                           m_shutdown_complete = false;
+	// Completion callbacks can outlive GuestGpu during renderer shutdown.
+	std::shared_ptr<std::binary_semaphore> m_suspend_point_ready =
+	    std::make_shared<std::binary_semaphore>(1);
 
 	std::unique_ptr<CommandProcessor>                                m_gfx_cp;
 	std::array<std::unique_ptr<CommandProcessor>, ComputeQueueCount> m_compute_cp;

src/graphics/presentation/videoOut.cpp

@@ -1814,6 +1814,38 @@ KYTY_SYSV_ABI int VideoOutGetOutputStatus(int handle, VideoOutOutputStatus* stat
 	return OK;
 }
 
+KYTY_SYSV_ABI int VideoOutAllowOutputResolutionWqhdDetection(int handle) {
+	if (!DriverState().IsOpened(handle)) {
+		return VIDEO_OUT_ERROR_INVALID_HANDLE;
+	}
+	return OK;
+}
+
+KYTY_SYSV_ABI int VideoOutVrrPegToFixedRate(int handle, uint64_t arg1, uint64_t arg2) {
+	PRINT_NAME();
+
+	static std::atomic_bool logged {false};
+	if (!logged.exchange(true, std::memory_order_relaxed)) {
+		LOGF("\t handle = %d\n"
+		     "\t arg1   = 0x%016" PRIx64 "\n"
+		     "\t arg2   = 0x%016" PRIx64 "\n",
+		     handle, arg1, arg2);
+	}
+
+	return OK;
+}
+
+KYTY_SYSV_ABI int VideoOutVrrUnpegFromFixedRate(int handle) {
+	PRINT_NAME();
+
+	static std::atomic_bool logged {false};
+	if (!logged.exchange(true, std::memory_order_relaxed)) {
+		LOGF("\t handle = %d\n", handle);
+	}
+
+	return OK;
+}
+
 static int ValidateOutputConfig(int handle, uint64_t mode, const VideoOutOutputOptions* options,
                                 void* reserved_ptr, uint64_t reserved) {
 	if (!DriverState().IsOpened(handle)) {

src/graphics/presentation/videoOut.h

@@ -90,6 +90,9 @@ KYTY_SYSV_ABI int VideoOutGetEventData(const LibKernel::EventQueue::KernelEvent*
 KYTY_SYSV_ABI int VideoOutGetEventCount(const LibKernel::EventQueue::KernelEvent* ev);
 KYTY_SYSV_ABI int VideoOutWaitVblank(int handle);
 KYTY_SYSV_ABI int VideoOutGetOutputStatus(int handle, VideoOutOutputStatus* status);
+KYTY_SYSV_ABI int VideoOutAllowOutputResolutionWqhdDetection(int handle);
+KYTY_SYSV_ABI int VideoOutVrrPegToFixedRate(int handle, uint64_t arg1, uint64_t arg2);
+KYTY_SYSV_ABI int VideoOutVrrUnpegFromFixedRate(int handle);
 KYTY_SYSV_ABI int VideoOutInitializeOutputOptions(VideoOutOutputOptions* options);
 KYTY_SYSV_ABI int VideoOutIsOutputSupported(int handle, uint64_t mode,
                                             const VideoOutOutputOptions* options,

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

@@ -384,6 +384,7 @@ enum class Opcode {
 	V_CMP_GT_I16,
 	V_CMP_NE_I16,
 	V_CMP_GE_I16,
+	V_CMPX_LT_I16,
 	V_CMP_LT_F16,
 	V_CMP_EQ_F16,
 	V_CMP_LE_F16,

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

@@ -239,6 +239,7 @@ constexpr VopcOpcodeInfo VOPC_OPCODE_LIST[] = {
     {0x92u, Opcode::V_CMPX_EQ_I32},        {0x93u, Opcode::V_CMPX_LE_I32},
     {0x94u, Opcode::V_CMPX_GT_I32},        {0x95u, Opcode::V_CMPX_NE_I32},
     {0x96u, Opcode::V_CMPX_GE_I32},        {0x98u, Opcode::V_CMPX_CLASS_F32},
+    {0x99u, Opcode::V_CMPX_LT_I16, false},
     {0xa9u, Opcode::V_CMP_LT_U16},         {0xaau, Opcode::V_CMP_EQ_U16},
     {0xabu, Opcode::V_CMP_LE_U16},         {0xacu, Opcode::V_CMP_GT_U16},
     {0xadu, Opcode::V_CMP_NE_U16},         {0xaeu, Opcode::V_CMP_GE_U16},
@@ -1487,6 +1488,7 @@ bool IsVopcCompareExec(Opcode opcode) {
 		case Opcode::V_CMPX_GE_U32:
 		case Opcode::V_CMPX_NE_I64:
 		case Opcode::V_CMPX_NE_U64:
+		case Opcode::V_CMPX_LT_I16:
 		case Opcode::V_CMPX_LT_U16:
 		case Opcode::V_CMPX_EQ_U16:
 		case Opcode::V_CMPX_GT_U16:

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

@@ -560,6 +560,12 @@ IR::U32 Translator::ReadU16LaneRaw(const Decoder::Operand& operand, bool high_la
 	} else if (operand.sdwa_sel == 4u || operand.sdwa_sel == 5u) {
 		offset = operand.sdwa_sel == 5u ? 16u : 0u;
 	}
+	if (width == 8u && operand.sdwa_sext) {
+		// Extend the selected byte into the 16-bit operand before halfword modifiers.
+		const auto value = IR::U32(
+		    ir.Emit(IR::ValueOpcode::BitFieldSExtract, {bits, IR::Value(offset), IR::Value(width)}));
+		return ir.BitwiseAnd(value, IR::U32(IR::Value(0xffffu)));
+	}
 	return IR::U32(
 	    ir.Emit(IR::ValueOpcode::BitFieldUExtract, {bits, IR::Value(offset), IR::Value(width)}));
 }
@@ -567,7 +573,7 @@ IR::U32 Translator::ReadU16LaneRaw(const Decoder::Operand& operand, bool high_la
 IR::U32 Translator::ReadU16LaneAsU32(const Decoder::Operand& operand, bool high_lane,
                                      bool sign_extend) {
 	auto value = Read16LaneBits(operand, high_lane);
-	if (sign_extend || operand.sdwa_sext) {
+	if (sign_extend) {
 		value = IR::U32(
 		    ir.Emit(IR::ValueOpcode::BitFieldSExtract, {value, IR::Value(0u), IR::Value(16u)}));
 	}

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

@@ -156,6 +156,9 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_CMP_LT_I16:
 			EmitInteger16Compare(inst, IR::ValueOpcode::SLessThan32, true, false);
 			return;
+		case O::V_CMPX_LT_I16:
+			EmitInteger16Compare(inst, IR::ValueOpcode::SLessThan32, true, true);
+			return;
 		case O::V_CMP_LE_I16:
 			EmitInteger16Compare(inst, IR::ValueOpcode::SLessThanEqual32, true, false);
 			return;

src/launcher/forms/configuration_edit_dialog.ui

@@ -72,6 +72,27 @@
           <property name="toolTip"><string>Language</string></property>
          </widget>
         </item>
+        <item row="3" column="0">
+         <widget class="QLabel" name="label_controller_color">
+          <property name="text"><string>DualSense lightbar:</string></property>
+         </widget>
+        </item>
+        <item row="3" column="1">
+         <layout class="QHBoxLayout" name="controllerColorLayout">
+          <item>
+           <widget class="QPushButton" name="button_controller_color">
+            <property name="toolTip"><string>Choose and preview a fixed LED color. Bluetooth DualSense may take several seconds to update.</string></property>
+            <property name="text"><string>Custom</string></property>
+           </widget>
+          </item>
+          <item>
+           <widget class="QPushButton" name="button_controller_color_reset">
+            <property name="toolTip"><string>Let the game control the lightbar</string></property>
+            <property name="text"><string>Default</string></property>
+           </widget>
+          </item>
+         </layout>
+        </item>
        </layout>
       </widget>
      </item>

src/launcher/include/configuration.h

@@ -5,6 +5,7 @@
 
 #include <QByteArray>
 #include <QChar>
+#include <QColor>
 #include <QMetaEnum>
 #include <QMetaType>
 #include <QObject>
@@ -88,6 +89,7 @@ public:
 	QString                user_name                   = "Kyty";
 	int                    user_id                     = Config::DEFAULT_USER_ID;
 	QString                audio_input_device;
+	QString                controller_color;
 	PresentMode            present_mode                = PresentMode::Mailbox;
 	int                    gpu_index                   = -1;
 	bool                   fullscreen_enabled          = false;
@@ -119,6 +121,7 @@ public:
 		user_name                   = other.user_name;
 		user_id                     = other.user_id;
 		audio_input_device          = other.audio_input_device;
+		controller_color            = other.controller_color;
 		present_mode                = other.present_mode;
 		gpu_index                   = other.gpu_index;
 		fullscreen_enabled          = other.fullscreen_enabled;
@@ -165,6 +168,7 @@ public:
 		KYTY_CFG_SET(user_name);
 		KYTY_CFG_SET(user_id);
 		KYTY_CFG_SET(audio_input_device);
+		KYTY_CFG_SET(controller_color);
 		KYTY_CFG_SET(present_mode);
 		KYTY_CFG_SET(gpu_index);
 		KYTY_CFG_SET(fullscreen_enabled);
@@ -204,6 +208,8 @@ public:
 		                         ? saved_user_id
 		                         : Config::DEFAULT_USER_ID;
 		audio_input_device = s->value("audio_input_device", audio_input_device).toString();
+		const QColor color(s->value("controller_color", controller_color).toString());
+		controller_color = color.isValid() ? color.name(QColor::HexRgb) : QString {};
 		KYTY_CFG_GET(present_mode);
 		gpu_index = s->value("gpu_index", -1).toInt();
 		if (EnumToText(present_mode).isEmpty()) {

src/launcher/include/configurationEditDialog.h

@@ -31,6 +31,9 @@ public:
 	void                      SetGameDirectories(const QStringList& dirs);
 	[[nodiscard]] QStringList GetGameDirectories() const;
 
+signals:
+	void PreviewControllerColor(const QString& color);
+
 private:
 	Ui::ConfigurationEditDialog* m_ui = nullptr;
 	Configuration&               m_info;

src/launcher/include/configurationListWidget.h

@@ -33,6 +33,9 @@ public:
 	ConfigurationItem*                     GetSelectedItem() { return m_selected_item; }
 
 	[[nodiscard]] const QString& GetSettingsFile() const { return m_settings_file; }
+	[[nodiscard]] const QString& GetGlobalControllerColor() const {
+		return m_global_info.controller_color;
+	}
 	[[nodiscard]] std::unique_ptr<Configuration>
 	CreateConfiguration(const ConfigurationItem& item) const;
 
@@ -44,6 +47,7 @@ signals:
 
 	void Run();
 	void Select();
+	void PreviewControllerColor(const QString& color);
 
 protected:
 	void changeEvent(QEvent* event) override;

src/launcher/include/controllerLightbar.h

@@ -0,0 +1,29 @@
+#ifndef CONTROLLER_LIGHTBAR_H
+#define CONTROLLER_LIGHTBAR_H
+
+#include <QColor>
+#include <QHash>
+#include <QObject>
+#include <QString>
+#include <QTimer>
+
+#include <SDL3/SDL_gamepad.h>
+
+class ControllerLightbar final: public QObject {
+public:
+	explicit ControllerLightbar(QObject* parent = nullptr);
+	~ControllerLightbar() override;
+
+	void SetColor(const QString& color);
+	void Stop();
+
+private:
+	void Refresh();
+	void ApplyColor(SDL_Gamepad* gamepad) const;
+
+	QTimer                              m_timer;
+	QHash<SDL_JoystickID, SDL_Gamepad*> m_gamepads;
+	QColor                              m_color;
+};
+
+#endif // CONTROLLER_LIGHTBAR_H

src/launcher/src/configurationEditDialog.cpp

@@ -7,6 +7,7 @@
 
 #include <QAbstractItemView>
 #include <QCheckBox>
+#include <QColorDialog>
 #include <QComboBox>
 #include <QCoreApplication>
 #include <QDir>
@@ -40,6 +41,22 @@ constexpr char SETTINGS_CFG_DIALOG[]               = "ConfigurationEditDialog";
 constexpr char SETTINGS_CFG_LAST_GEOMETRY[]        = "geometry";
 constexpr int  GLOBAL_SETTINGS_GAME_DIRS_MIN_WIDTH = 560;
 
+static void UpdateControllerColorButton(QPushButton* button, const QString& hex) {
+	const QColor color(hex);
+	if (!color.isValid()) {
+		button->setProperty("controllerColor", QString {});
+		button->setText(QObject::tr("Custom"));
+		button->setStyleSheet({});
+		return;
+	}
+	const auto normalized = color.name(QColor::HexRgb);
+	button->setProperty("controllerColor", normalized);
+	button->setText(normalized.toUpper());
+	button->setStyleSheet(QStringLiteral("background-color: %1; color: %2;")
+	                          .arg(normalized, color.lightness() < 128 ? QStringLiteral("white")
+	                                                                   : QStringLiteral("black")));
+}
+
 const QStringList CONSOLE_LANGUAGE_NAMES = {
     "Japanese",
     "English (United States)",
@@ -110,6 +127,21 @@ ConfigurationEditDialog::ConfigurationEditDialog(Configuration& info, QWidget* p
 	connect(m_ui->ok_button, &QPushButton::clicked, this, &ConfigurationEditDialog::save);
 	connect(m_ui->cancel_button, &QPushButton::clicked, this, &QDialog::reject);
 	connect(m_ui->clear_button, &QPushButton::clicked, this, &ConfigurationEditDialog::clear);
+	connect(m_ui->button_controller_color, &QPushButton::clicked, this, [this]() {
+		const auto current = m_ui->button_controller_color->property("controllerColor").toString();
+		const auto initial =
+		    current.isEmpty() ? QColor(QStringLiteral("#0070d1")) : QColor(current);
+		const auto color = QColorDialog::getColor(initial, this, tr("DualSense lightbar color"));
+		if (color.isValid()) {
+			const auto hex = color.name(QColor::HexRgb);
+			UpdateControllerColorButton(m_ui->button_controller_color, hex);
+			emit PreviewControllerColor(hex);
+		}
+	});
+	connect(m_ui->button_controller_color_reset, &QPushButton::clicked, this, [this]() {
+		UpdateControllerColorButton(m_ui->button_controller_color, {});
+		emit PreviewControllerColor({});
+	});
 	connect(m_ui->comboBox_shader_log_direction, &QComboBox::currentTextChanged, this,
 	        [this](const QString& text) {
 		        auto log = TextToEnum<Configuration::LogDirection>(text);
@@ -169,6 +201,7 @@ void ConfigurationEditDialog::Init(const Configuration& info) {
 	m_ui->lineEdit_user_name->setMaxLength(static_cast<int>(Config::MAX_USER_NAME_LENGTH));
 	m_ui->lineEdit_user_name->setText(info.user_name);
 	m_ui->spinBox_user_id->setValue(info.user_id);
+	UpdateControllerColorButton(m_ui->button_controller_color, info.controller_color);
 	auto* microphone = m_ui->comboBox_audio_input_device;
 	microphone->clear();
 	microphone->addItem(tr("None"), QString {});
@@ -370,6 +403,7 @@ static void UpdateInfo(Configuration& info, Ui::ConfigurationEditDialog& ui) {
 	info.user_name = ui.lineEdit_user_name->text().trimmed();
 	info.user_id   = ui.spinBox_user_id->value();
 	info.audio_input_device = ui.comboBox_audio_input_device->currentData().toString();
+	info.controller_color   = ui.button_controller_color->property("controllerColor").toString();
 	info.screen_resolution =
 	    TextToEnum<Configuration::Resolution>(ui.comboBox_screen_resolution->currentText());
 	info.present_mode =
@@ -427,6 +461,7 @@ void ConfigurationEditDialog::save() {
 void ConfigurationEditDialog::clear() {
 	Configuration default_info;
 	Init(default_info);
+	emit PreviewControllerColor({});
 
 	if (m_show_game_dirs) {
 		m_game_dirs_list->clear();

src/launcher/src/configurationListWidget.cpp

@@ -751,6 +751,8 @@ void ConfigurationListWidget::edit_configuration() {
 	auto                    info = CreateConfiguration(*item);
 	ConfigurationEditDialog dlg(*info, this);
 	dlg.setWindowTitle(tr("Edit game settings"));
+	connect(&dlg, &ConfigurationEditDialog::PreviewControllerColor, this,
+	        &ConfigurationListWidget::PreviewControllerColor);
 
 	if (dlg.exec() == QDialog::Accepted) {
 		info->custom_settings           = true;
@@ -761,6 +763,8 @@ void ConfigurationListWidget::edit_configuration() {
 		WriteSettings();
 		item->Update();
 		SelectItem(item);
+	} else {
+		emit Select();
 	}
 }
 
@@ -788,6 +792,8 @@ void ConfigurationListWidget::edit_global_settings() {
 	ConfigurationEditDialog dlg(info, this);
 	dlg.setWindowTitle(tr("Global settings"));
 	dlg.SetGameDirectories(m_game_dirs);
+	connect(&dlg, &ConfigurationEditDialog::PreviewControllerColor, this,
+	        &ConfigurationListWidget::PreviewControllerColor);
 
 	if (dlg.exec() == QDialog::Accepted) {
 		m_global_info.CopyEmulatorSettingsFrom(info);
@@ -799,6 +805,7 @@ void ConfigurationListWidget::edit_global_settings() {
 			ScanGameDirectory();
 		}
 	}
+	emit Select();
 }
 
 void ConfigurationListWidget::edit_input_mapping() {

src/launcher/src/controllerLightbar.cpp

@@ -0,0 +1,84 @@
+#include "controllerLightbar.h"
+
+#include <QDebug>
+
+#include <SDL3/SDL.h>
+
+ControllerLightbar::ControllerLightbar(QObject* parent): QObject(parent) {
+	m_timer.setInterval(100);
+	connect(&m_timer, &QTimer::timeout, this, &ControllerLightbar::Refresh);
+}
+
+ControllerLightbar::~ControllerLightbar() {
+	Stop();
+}
+
+void ControllerLightbar::SetColor(const QString& color) {
+	const QColor parsed(color);
+	if (!parsed.isValid()) {
+		Stop();
+		return;
+	}
+	if (m_color == parsed) {
+		return;
+	}
+	if (!m_timer.isActive()) {
+		if (!SDL_InitSubSystem(SDL_INIT_GAMEPAD)) {
+			qWarning() << "Cannot initialize controller color preview:" << SDL_GetError();
+			return;
+		}
+		m_timer.start();
+	}
+	m_color = parsed;
+	for (auto* gamepad: m_gamepads) {
+		ApplyColor(gamepad);
+	}
+	Refresh();
+}
+
+void ControllerLightbar::Stop() {
+	if (!m_timer.isActive()) {
+		return;
+	}
+	m_timer.stop();
+	for (auto* gamepad: m_gamepads) {
+		SDL_CloseGamepad(gamepad);
+	}
+	m_gamepads.clear();
+	SDL_QuitSubSystem(SDL_INIT_GAMEPAD);
+	m_color = {};
+}
+
+void ControllerLightbar::Refresh() {
+	// Bluetooth DualSense may defer an accepted LED command until its connection
+	// animation ends. Polling events also delivers those deferred commands.
+	SDL_Event event;
+	while (SDL_PollEvent(&event)) {
+		if (event.type == SDL_EVENT_GAMEPAD_ADDED && !m_gamepads.contains(event.gdevice.which)) {
+			auto* gamepad = SDL_OpenGamepad(event.gdevice.which);
+			if (gamepad == nullptr) {
+				qWarning() << "Cannot open controller color preview:" << SDL_GetError();
+				continue;
+			}
+			if (SDL_GetBooleanProperty(SDL_GetGamepadProperties(gamepad),
+			                           SDL_PROP_GAMEPAD_CAP_RGB_LED_BOOLEAN, false)) {
+				m_gamepads.insert(event.gdevice.which, gamepad);
+				ApplyColor(gamepad);
+			} else {
+				SDL_CloseGamepad(gamepad);
+			}
+		} else if (event.type == SDL_EVENT_GAMEPAD_REMOVED) {
+			if (auto* gamepad = m_gamepads.take(event.gdevice.which); gamepad != nullptr) {
+				SDL_CloseGamepad(gamepad);
+			}
+		}
+	}
+}
+
+void ControllerLightbar::ApplyColor(SDL_Gamepad* gamepad) const {
+	if (!SDL_SetGamepadLED(gamepad, static_cast<Uint8>(m_color.red()),
+	                       static_cast<Uint8>(m_color.green()),
+	                       static_cast<Uint8>(m_color.blue()))) {
+		qWarning() << "Cannot set controller LED:" << SDL_GetError();
+	}
+}

src/launcher/src/mainDialog.cpp

@@ -3,6 +3,7 @@
 #include "configuration.h"
 #include "configurationItem.h"
 #include "configurationListWidget.h"
+#include "controllerLightbar.h"
 #include "gameContent.h"
 #include "patchesDialog.h"
 #include "updateChecker.h"
@@ -88,6 +89,7 @@ private:
 	Ui::MainDialog* m_ui             = {nullptr};
 	MainDialog*     m_main_dialog    = nullptr;
 	UpdateChecker*  m_update_checker = nullptr;
+	ControllerLightbar m_lightbar;
 	QString         m_interpreter;
 
 	QProcess m_process;
@@ -103,6 +105,7 @@ MainDialog::MainDialog(QWidget* parent): QDialog(parent), m_p(new MainDialogPriv
 }
 
 MainDialogPrivate::~MainDialogPrivate() {
+	m_process.disconnect(this);
 	delete m_ui;
 }
 
@@ -121,6 +124,12 @@ void MainDialogPrivate::Setup(MainDialog* main_dialog) {
 	connect(main_dialog, &MainDialog::Start, this, &MainDialogPrivate::FindInterpreter,
 	        Qt::QueuedConnection);
 	connect(m_ui->widget, &ConfigurationListWidget::Select, this, &MainDialogPrivate::Update);
+	connect(m_ui->widget, &ConfigurationListWidget::PreviewControllerColor, this,
+	        [this](const QString& color) {
+		        if (m_process.state() == QProcess::NotRunning) {
+			        m_lightbar.SetColor(color);
+		        }
+	        });
 	connect(m_ui->widget, &ConfigurationListWidget::Run, this, &MainDialogPrivate::Run);
 	connect(m_ui->check_updates_link, &QLabel::linkActivated, this,
 	        [this](const QString&) { m_update_checker->Check(true); });
@@ -135,14 +144,14 @@ void MainDialogPrivate::Setup(MainDialog* main_dialog) {
 		m_ui->widget->WriteSettings();
 	});
 
-	connect(&m_process,
-	        static_cast<void (QProcess::*)(int, QProcess::ExitStatus)>(&QProcess::finished),
-	        [this](int /*exitCode*/, QProcess::ExitStatus /*exitStatus*/) {
-		        if (m_running_item != nullptr) {
-			        m_running_item->SetRunning(false);
-		        }
-		        Update();
-	        });
+	connect(&m_process, &QProcess::stateChanged, this, [this](QProcess::ProcessState state) {
+		if (state == QProcess::NotRunning) {
+			if (m_running_item != nullptr) {
+				m_running_item->SetRunning(false);
+			}
+			Update();
+		}
+	});
 
 	m_ui->label_settings_file->setText(tr("Settings file: ") + m_ui->widget->GetSettingsFile());
 
@@ -219,6 +228,9 @@ static QStringList CreateEmulatorArgs(const Configuration& info) {
 	if (!info.audio_input_device.isEmpty()) {
 		args << "--mic" << info.audio_input_device;
 	}
+	if (!info.controller_color.isEmpty()) {
+		args << "--controller-color" << info.controller_color;
+	}
 	args << "--present-mode" << EnumToText(info.present_mode);
 	if (info.gpu_index >= 0) {
 		args << "--gpu" << QString::number(info.gpu_index);
@@ -493,6 +505,7 @@ void MainDialogPrivate::Run() {
 	}
 
 	m_running_item->SetRunning(true);
+	m_lightbar.Stop();
 
 	auto info = m_ui->widget->CreateConfiguration(*m_running_item);
 	m_main_dialog->RunInterpreter(&m_process, *info);
@@ -512,6 +525,12 @@ void MainDialogPrivate::Update() {
 	}
 
 	m_ui->widget->SetRunEnabled(run_enabled);
+	if (m_process.state() != QProcess::NotRunning) {
+		m_lightbar.Stop();
+		return;
+	}
+	m_lightbar.SetColor(item != nullptr ? m_ui->widget->CreateConfiguration(*item)->controller_color
+	                                    : m_ui->widget->GetGlobalControllerColor());
 }
 
 #include "mainDialog.moc"

src/libs/agc.cpp

@@ -1552,7 +1552,7 @@ int KYTY_SYSV_ABI AgcSuspendPoint() {
 	PRINT_NAME();
 
 	EXIT_IF(g_renderer == nullptr);
-	g_renderer->GetGpu().Done();
+	g_renderer->GetGpu().SuspendPoint();
 
 	return OK;
 }

src/libs/controller.cpp

@@ -133,6 +133,9 @@ private:
 	int              m_connected_count = 0;
 	bool             m_motion_enabled  = true;
 	uint64_t         m_gyro_time       = 0;
+	// Accelerometer gravity direction in the reset frame.
+	std::array<float, 3> m_up_reference {0.0f, 1.0f, 0.0f};
+	bool                 m_up_reference_valid = false;
 	ControllerState  m_state;
 	ControllerState  m_states[STATES_MAX];
 	bool             m_obtained[STATES_MAX] {};
@@ -297,6 +300,9 @@ void GameController::Connect(int id) {
 	if (id != HOST_INPUT_CONTROLLER_ID) {
 		if (auto* pad = SDL_GetGamepadFromID(static_cast<SDL_JoystickID>(id));
 		    pad != nullptr) {
+			if (const auto& color = Config::GetControllerColor()) {
+				(void)SDL_SetGamepadLED(pad, (*color)[0], (*color)[1], (*color)[2]);
+			}
 			for (auto sensor: {SDL_SENSOR_ACCEL, SDL_SENSOR_GYRO}) {
 				if (SDL_GamepadHasSensor(pad, sensor) &&
 				    !SDL_SetGamepadSensorEnabled(pad, sensor, true)) {
@@ -345,6 +351,7 @@ void GameController::CheckActive() {
 	m_connected     = new_connected;
 	m_state         = {};
 	m_gyro_time     = 0;
+	m_up_reference_valid = false;
 	m_states_num    = 0;
 	m_first_state   = 0;
 	m_next_touch_id = 1;
@@ -436,6 +443,29 @@ void GameController::TouchPad(int id, int finger, bool down, float x, float y) {
 	}
 }
 
+namespace {
+
+using Vec3 = std::array<float, 3>;
+using Quat = std::array<float, 4>; // x, y, z, w
+
+Vec3 QuatRotate(const Quat& q, const Vec3& v) {
+	const Vec3 t {2.0f * (q[1] * v[2] - q[2] * v[1]), 2.0f * (q[2] * v[0] - q[0] * v[2]),
+	              2.0f * (q[0] * v[1] - q[1] * v[0])};
+	return {v[0] + q[3] * t[0] + (q[1] * t[2] - q[2] * t[1]),
+	        v[1] + q[3] * t[1] + (q[2] * t[0] - q[0] * t[2]),
+	        v[2] + q[3] * t[2] + (q[0] * t[1] - q[1] * t[0])};
+}
+
+Quat QuatConjugate(const Quat& q) {
+	return {-q[0], -q[1], -q[2], q[3]};
+}
+
+float Length(const Vec3& v) {
+	return std::sqrt(v[0] * v[0] + v[1] * v[1] + v[2] * v[2]);
+}
+
+} // namespace
+
 void GameController::Motion(int id, Sensor sensor, const float* data, uint64_t time_us) {
 	Common::LockGuard lock(m_mutex);
 	if (id != m_active_id || !m_motion_enabled) {
@@ -447,20 +477,38 @@ void GameController::Motion(int id, Sensor sensor, const float* data, uint64_t t
 		for (int i = 0; i < 3; i++) {
 			m_state.accel[i] = data[i] / SDL_STANDARD_GRAVITY;
 		}
+		// Capture gravity near 1 G to limit interference from linear acceleration.
+		const float magnitude = Length(m_state.accel);
+		if (!m_up_reference_valid && magnitude > 0.9f && magnitude < 1.1f) {
+			const Vec3 up {m_state.accel[0] / magnitude, m_state.accel[1] / magnitude,
+			               m_state.accel[2] / magnitude};
+			m_up_reference       = QuatRotate(m_state.orientation, up);
+			m_up_reference_valid = true;
+		}
 	} else {
 		std::copy_n(data, 3, m_state.gyro.begin());
+		// Mahony tilt correction; heading is unobservable and reported gyro values stay raw.
+		auto        rate      = m_state.gyro;
+		const float magnitude = Length(m_state.accel);
+		if (m_up_reference_valid && magnitude > 0.8f && magnitude < 1.2f) {
+			const Vec3 measured {m_state.accel[0] / magnitude, m_state.accel[1] / magnitude,
+			                     m_state.accel[2] / magnitude};
+			const Vec3 predicted = QuatRotate(QuatConjugate(m_state.orientation), m_up_reference);
+			rate[0] += measured[1] * predicted[2] - measured[2] * predicted[1];
+			rate[1] += measured[2] * predicted[0] - measured[0] * predicted[2];
+			rate[2] += measured[0] * predicted[1] - measured[1] * predicted[0];
+		}
 		// Do not extrapolate a single sample across lost reports (e.g. loss of window focus).
 		constexpr uint64_t max_gyro_interval_us = 100000;
 		if (m_gyro_time != 0 && time_us > m_gyro_time &&
 		    time_us - m_gyro_time <= max_gyro_interval_us) {
-			const float dt = static_cast<float>(time_us - m_gyro_time) * 0.000001f;
-			const float speed =
-			    std::sqrt(data[0] * data[0] + data[1] * data[1] + data[2] * data[2]);
+			const float dt         = static_cast<float>(time_us - m_gyro_time) * 0.000001f;
+			const float speed      = Length(rate);
 			const float half_angle = speed * dt * 0.5f;
 			const float scale      = speed > 0.0f ? std::sin(half_angle) / speed : 0.0f;
-			const float x          = data[0] * scale;
-			const float y          = data[1] * scale;
-			const float z          = data[2] * scale;
+			const float x          = rate[0] * scale;
+			const float y          = rate[1] * scale;
+			const float z          = rate[2] * scale;
 			const float w          = std::cos(half_angle);
 			const auto  q          = m_state.orientation;
 			// Accumulate body-local rotation relative to connection / orientation reset.
@@ -486,19 +534,21 @@ void GameController::Motion(int id, Sensor sensor, const float* data, uint64_t t
 void GameController::SetMotionSensorState(bool enable) {
 	Common::LockGuard lock(m_mutex);
 	if (m_motion_enabled != enable) {
-		m_motion_enabled = enable;
-		m_gyro_time      = 0;
-		m_state.accel    = {0.0f, 1.0f, 0.0f};
-		m_state.gyro     = {};
-		m_state.time     = LibKernel::KernelGetProcessTime();
+		m_motion_enabled     = enable;
+		m_gyro_time          = 0;
+		m_state.accel        = {0.0f, 1.0f, 0.0f};
+		m_state.gyro         = {};
+		m_up_reference_valid = false;
+		m_state.time         = LibKernel::KernelGetProcessTime();
 		AddState();
 	}
 }
 
 void GameController::ResetOrientation() {
 	Common::LockGuard lock(m_mutex);
-	m_state.orientation = {0.0f, 0.0f, 0.0f, 1.0f};
-	m_gyro_time         = 0;
+	m_state.orientation  = {0.0f, 0.0f, 0.0f, 1.0f};
+	m_gyro_time          = 0;
+	m_up_reference_valid = false;
 	m_state.time        = LibKernel::KernelGetProcessTime();
 	AddState();
 }
@@ -581,6 +631,11 @@ int GameController::GetActiveControllerId() {
 
 void GameController::SetLightBar(uint8_t r, uint8_t g, uint8_t b) {
 	Common::LockGuard lock(m_mutex);
+	if (const auto& color = Config::GetControllerColor()) {
+		r = (*color)[0];
+		g = (*color)[1];
+		b = (*color)[2];
+	}
 	if (auto* pad = SDL_GetGamepadFromID(static_cast<SDL_JoystickID>(m_active_id));
 	    pad != nullptr) {
 		(void)SDL_SetGamepadLED(pad, r, g, b);

src/libs/libNet.cpp

@@ -3834,6 +3834,27 @@ namespace LibSharePlay {
 
 LIB_VERSION("SharePlay", 1, "SharePlay", 1, 1);
 
+constexpr int SHARE_PLAY_ERROR_INVALID_ARGS        = -2129788927;
+constexpr int SHARE_PLAY_ERROR_ALREADY_INITIALIZED = -2129788925;
+constexpr int SHARE_PLAY_ERROR_NOT_INITIALIZED     = -2129788924;
+
+struct SharePlayConnectionInfoA {
+	int32_t  status;
+	int32_t  mode;
+	char     host_online_id[20];
+	char     visitor_online_id[20];
+	uint64_t host_account_id;
+	uint64_t visitor_account_id;
+	int32_t  host_user_id;
+	int32_t  visitor_user_id;
+};
+
+static_assert(sizeof(SharePlayConnectionInfoA) == 72);
+static_assert(offsetof(SharePlayConnectionInfoA, host_account_id) == 48);
+static_assert(offsetof(SharePlayConnectionInfoA, host_user_id) == 64);
+
+static bool g_share_play_initialized = false;
+
 static int KYTY_SYSV_ABI SharePlayInitialize(void* heap, size_t heap_size) {
 	PRINT_NAME();
 
@@ -3841,18 +3862,41 @@ static int KYTY_SYSV_ABI SharePlayInitialize(void* heap, size_t heap_size) {
 	     "\t heap_size = %" PRIu64 "\n",
 	     reinterpret_cast<uint64_t>(heap), static_cast<uint64_t>(heap_size));
 
+	if (g_share_play_initialized) {
+		return SHARE_PLAY_ERROR_ALREADY_INITIALIZED;
+	}
+	if (heap != nullptr && heap_size < 6u * 1024u) {
+		return SHARE_PLAY_ERROR_INVALID_ARGS;
+	}
+	g_share_play_initialized = true;
 	return 0;
 }
 
 static int KYTY_SYSV_ABI SharePlayTerminate() {
 	PRINT_NAME();
 
+	if (!g_share_play_initialized) {
+		return SHARE_PLAY_ERROR_NOT_INITIALIZED;
+	}
+	g_share_play_initialized = false;
+	return 0;
+}
+
+static int KYTY_SYSV_ABI SharePlayGetCurrentConnectionInfoA(SharePlayConnectionInfoA* info) {
+	if (!g_share_play_initialized) {
+		return SHARE_PLAY_ERROR_NOT_INITIALIZED;
+	}
+	if (info == nullptr) {
+		return SHARE_PLAY_ERROR_INVALID_ARGS;
+	}
+	*info = {.status = 0, .host_user_id = -1, .visitor_user_id = -1};
 	return 0;
 }
 
 LIB_DEFINE(InitPlatform_1_SharePlay) {
 	LIB_FUNC("isruqthpYcw", LibSharePlay::SharePlayInitialize);
 	LIB_FUNC("UaLjloJinow", LibSharePlay::SharePlayTerminate);
+	LIB_FUNC("+MCXJlWdi+s", LibSharePlay::SharePlayGetCurrentConnectionInfoA);
 }
 
 } // namespace LibSharePlay

src/libs/libPngEnc.cpp

@@ -0,0 +1,349 @@
+#include "common/abi.h"
+#include "common/logging/log.h"
+#include "libs/libs.h"
+#include "loader/symbolDatabase.h"
+
+#include <bit>
+#include <cstddef>
+#include <cstring>
+#include <limits>
+#include <memory>
+#include <mutex>
+#include <new>
+
+#define STB_IMAGE_WRITE_IMPLEMENTATION
+#define STB_IMAGE_WRITE_STATIC
+#define STBIWDEF static inline
+#define STBI_WRITE_NO_STDIO
+#include "stb_image_write.h"
+
+namespace Libs {
+
+LIB_VERSION("PngEnc", 1, "PngEnc", 1, 1);
+
+namespace PngEnc {
+
+constexpr int32_t PNG_ENC_ERROR_INVALID_ADDR   = -2140602111; // 0x80690101
+constexpr int32_t PNG_ENC_ERROR_INVALID_SIZE   = -2140602110; // 0x80690102
+constexpr int32_t PNG_ENC_ERROR_INVALID_PARAM  = -2140602109; // 0x80690103
+constexpr int32_t PNG_ENC_ERROR_INVALID_HANDLE = -2140602108; // 0x80690104
+constexpr int32_t PNG_ENC_ERROR_DATA_OVERFLOW  = -2140602096; // 0x80690110
+constexpr int32_t PNG_ENC_ERROR_FATAL          = -2140602080; // 0x80690120
+
+constexpr uint32_t PNG_ENC_MAX_IMAGE_WIDTH       = 1000000;
+constexpr uint32_t PNG_ENC_MAX_IMAGE_HEIGHT      = 1000000;
+constexpr uint32_t PNG_ENC_MAX_FILTER_NUMBER     = 4;
+constexpr uint16_t PNG_ENC_COLOR_SPACE_RGB       = 3;
+constexpr uint16_t PNG_ENC_COLOR_SPACE_RGBA      = 19;
+constexpr uint16_t PNG_ENC_PIXEL_FORMAT_R8G8B8A8 = 0;
+constexpr uint16_t PNG_ENC_PIXEL_FORMAT_B8G8R8A8 = 1;
+constexpr uint16_t PNG_ENC_FILTER_TYPE_SUB       = 1;
+constexpr uint16_t PNG_ENC_FILTER_TYPE_UP        = 2;
+constexpr uint16_t PNG_ENC_FILTER_TYPE_AVERAGE   = 4;
+constexpr uint16_t PNG_ENC_FILTER_TYPE_PAETH     = 8;
+
+struct PngEncCreateParam {
+	uint32_t this_size;
+	uint32_t attribute;
+	uint32_t max_image_width;
+	uint32_t max_filter_number;
+};
+
+struct PngEncEncodeParam {
+	const uint8_t* image_mem_addr;
+	uint8_t*       png_mem_addr;
+	uint32_t       image_mem_size;
+	uint32_t       png_mem_size;
+	uint32_t       image_width;
+	uint32_t       image_height;
+	uint32_t       image_pitch;
+	uint16_t       pixel_format;
+	uint16_t       color_space;
+	uint16_t       bit_depth;
+	uint16_t       clut_number;
+	uint16_t       filter_type;
+	uint16_t       compression_level;
+};
+
+struct PngEncOutputInfo {
+	uint32_t data_size;
+	uint32_t processed_height;
+};
+
+static_assert(sizeof(PngEncCreateParam) == 16);
+static_assert(sizeof(PngEncEncodeParam) == 48);
+
+struct PngEncContext {
+	uint64_t magic;
+	uint32_t max_image_width;
+	uint32_t max_filter_number;
+};
+
+constexpr uint64_t PNG_ENC_CONTEXT_MAGIC = 0x4b595459504e4745ull; // KYTYPNGE
+
+struct PngEncWriter {
+	uint8_t* dst;
+	uint32_t capacity;
+	uint32_t size;
+	bool     overflow;
+};
+
+// stb_image_write keeps its compression level and filter choice in globals
+static std::mutex g_png_enc_mutex;
+
+static void PngEncWrite(void* context, void* data, int size) {
+	auto* writer = static_cast<PngEncWriter*>(context);
+
+	if (writer->overflow || size < 0 ||
+	    static_cast<uint64_t>(writer->size) + static_cast<uint32_t>(size) > writer->capacity) {
+		writer->overflow = true;
+		return;
+	}
+
+	std::memcpy(writer->dst + writer->size, data, static_cast<size_t>(size));
+	writer->size += static_cast<uint32_t>(size);
+}
+
+static int PngEncForcedFilter(uint16_t filter_type) {
+	if (filter_type == (PNG_ENC_FILTER_TYPE_SUB | PNG_ENC_FILTER_TYPE_UP |
+	                    PNG_ENC_FILTER_TYPE_AVERAGE | PNG_ENC_FILTER_TYPE_PAETH)) {
+		return -1;
+	}
+	// stb's adaptive mode cannot restrict itself to the guest's filter mask.
+	if ((filter_type & PNG_ENC_FILTER_TYPE_SUB) != 0) {
+		return 1;
+	}
+	if ((filter_type & PNG_ENC_FILTER_TYPE_UP) != 0) {
+		return 2;
+	}
+	if ((filter_type & PNG_ENC_FILTER_TYPE_AVERAGE) != 0) {
+		return 3;
+	}
+	if ((filter_type & PNG_ENC_FILTER_TYPE_PAETH) != 0) {
+		return 4;
+	}
+	return 0;
+}
+
+static int32_t ValidateCreateParam(const PngEncCreateParam* param) {
+	if (param == nullptr) {
+		return PNG_ENC_ERROR_INVALID_ADDR;
+	}
+
+	LOGF("\t this_size         = %" PRIu32 "\n", param->this_size);
+	LOGF("\t attribute         = %" PRIu32 "\n", param->attribute);
+	LOGF("\t max_image_width   = %" PRIu32 "\n", param->max_image_width);
+	LOGF("\t max_filter_number = %" PRIu32 "\n", param->max_filter_number);
+
+	if (param->this_size != sizeof(PngEncCreateParam) || param->attribute != 0 ||
+	    param->max_filter_number > PNG_ENC_MAX_FILTER_NUMBER) {
+		return PNG_ENC_ERROR_INVALID_PARAM;
+	}
+
+	if (param->max_image_width == 0 || param->max_image_width > PNG_ENC_MAX_IMAGE_WIDTH) {
+		return PNG_ENC_ERROR_INVALID_SIZE;
+	}
+
+	return 0;
+}
+
+static int32_t KYTY_SYSV_ABI PngEncQueryMemorySize(const PngEncCreateParam* param) {
+	PRINT_NAME();
+
+	if (int32_t result = ValidateCreateParam(param); result != 0) {
+		return result;
+	}
+
+	return sizeof(PngEncContext);
+}
+
+static int32_t KYTY_SYSV_ABI PngEncCreate(const PngEncCreateParam* param, void* memory_address,
+                                          uint32_t memory_size, void** handle) {
+	PRINT_NAME();
+
+	LOGF("\t memory_address    = %p\n", memory_address);
+	LOGF("\t memory_size       = %" PRIu32 "\n", memory_size);
+
+	if (int32_t result = ValidateCreateParam(param); result != 0) {
+		return result;
+	}
+
+	if (memory_address == nullptr || handle == nullptr) {
+		return PNG_ENC_ERROR_INVALID_ADDR;
+	}
+
+	if (memory_size < sizeof(PngEncContext)) {
+		return PNG_ENC_ERROR_INVALID_SIZE;
+	}
+
+	auto* ctx              = static_cast<PngEncContext*>(memory_address);
+	ctx->magic             = PNG_ENC_CONTEXT_MAGIC;
+	ctx->max_image_width   = param->max_image_width;
+	ctx->max_filter_number = param->max_filter_number;
+	*handle                = ctx;
+
+	return 0;
+}
+
+static int32_t KYTY_SYSV_ABI PngEncEncode(void* handle, const PngEncEncodeParam* param,
+                                          PngEncOutputInfo* output_info) {
+	PRINT_NAME();
+
+	if (handle == nullptr) {
+		return PNG_ENC_ERROR_INVALID_HANDLE;
+	}
+
+	const auto* ctx = static_cast<const PngEncContext*>(handle);
+	if (ctx->magic != PNG_ENC_CONTEXT_MAGIC) {
+		return PNG_ENC_ERROR_INVALID_HANDLE;
+	}
+
+	if (param == nullptr) {
+		return PNG_ENC_ERROR_INVALID_PARAM;
+	}
+
+	LOGF("\t image_mem_addr    = %p\n", static_cast<const void*>(param->image_mem_addr));
+	LOGF("\t png_mem_addr      = %p\n", static_cast<const void*>(param->png_mem_addr));
+	LOGF("\t image_mem_size    = %" PRIu32 "\n", param->image_mem_size);
+	LOGF("\t png_mem_size      = %" PRIu32 "\n", param->png_mem_size);
+	LOGF("\t image_width       = %" PRIu32 "\n", param->image_width);
+	LOGF("\t image_height      = %" PRIu32 "\n", param->image_height);
+	LOGF("\t image_pitch       = %" PRIu32 "\n", param->image_pitch);
+	LOGF("\t pixel_format      = %" PRIu16 "\n", param->pixel_format);
+	LOGF("\t color_space       = %" PRIu16 "\n", param->color_space);
+	LOGF("\t bit_depth         = %" PRIu16 "\n", param->bit_depth);
+	LOGF("\t filter_type       = %" PRIu16 "\n", param->filter_type);
+	LOGF("\t compression_level = %" PRIu16 "\n", param->compression_level);
+
+	if (param->image_mem_addr == nullptr || param->png_mem_addr == nullptr ||
+	    (reinterpret_cast<uintptr_t>(param->image_mem_addr) & 3u) != 0) {
+		return PNG_ENC_ERROR_INVALID_ADDR;
+	}
+
+	if (param->pixel_format != PNG_ENC_PIXEL_FORMAT_R8G8B8A8 &&
+	    param->pixel_format != PNG_ENC_PIXEL_FORMAT_B8G8R8A8) {
+		return PNG_ENC_ERROR_INVALID_PARAM;
+	}
+
+	if (param->color_space != PNG_ENC_COLOR_SPACE_RGB &&
+	    param->color_space != PNG_ENC_COLOR_SPACE_RGBA) {
+		return PNG_ENC_ERROR_INVALID_PARAM;
+	}
+
+	if (param->bit_depth != 8 || param->clut_number != 0 || param->compression_level > 9) {
+		return PNG_ENC_ERROR_INVALID_PARAM;
+	}
+
+	constexpr uint16_t filter_mask = PNG_ENC_FILTER_TYPE_SUB | PNG_ENC_FILTER_TYPE_UP |
+	                                 PNG_ENC_FILTER_TYPE_AVERAGE | PNG_ENC_FILTER_TYPE_PAETH;
+	if ((param->filter_type & ~filter_mask) != 0 ||
+	    static_cast<uint32_t>(std::popcount(param->filter_type)) > ctx->max_filter_number) {
+		return PNG_ENC_ERROR_INVALID_PARAM;
+	}
+
+	const uint32_t width  = param->image_width;
+	const uint32_t height = param->image_height;
+	if (width == 0 || height == 0 || width > ctx->max_image_width ||
+	    width > PNG_ENC_MAX_IMAGE_WIDTH || height > PNG_ENC_MAX_IMAGE_HEIGHT) {
+		return PNG_ENC_ERROR_INVALID_SIZE;
+	}
+
+	const int      components = (param->color_space == PNG_ENC_COLOR_SPACE_RGBA ? 4 : 3);
+	const uint32_t row_size   = width * static_cast<uint32_t>(components);
+	const uint32_t min_pitch  = width * 4u;
+	const uint32_t pitch      = param->image_pitch;
+
+	// Leave room for deflate expansion and stb's signed-int buffer-capacity doubling.
+	constexpr uint64_t max_filtered_size = std::numeric_limits<int>::max() / 2u;
+	if (static_cast<uint64_t>(row_size + 1u) * height > max_filtered_size ||
+	    param->png_mem_size == 0 || pitch < min_pitch || (pitch & 3u) != 0 ||
+	    static_cast<uint64_t>(pitch) * height > param->image_mem_size) {
+		return PNG_ENC_ERROR_INVALID_SIZE;
+	}
+
+	const bool                 bgr = (param->pixel_format == PNG_ENC_PIXEL_FORMAT_B8G8R8A8);
+	std::unique_ptr<uint8_t[]> pixels(new (std::nothrow)
+	                                      uint8_t[static_cast<size_t>(row_size) * height]);
+	if (pixels == nullptr) {
+		return PNG_ENC_ERROR_FATAL;
+	}
+
+	for (uint32_t y = 0; y < height; y++) {
+		const auto* src = param->image_mem_addr + static_cast<size_t>(y) * pitch;
+		auto*       dst = pixels.get() + static_cast<size_t>(y) * row_size;
+
+		for (uint32_t x = 0; x < width; x++) {
+			dst[0] = src[bgr ? 2 : 0];
+			dst[1] = src[1];
+			dst[2] = src[bgr ? 0 : 2];
+			if (components == 4) {
+				dst[3] = src[3];
+			}
+
+			src += 4;
+			dst += components;
+		}
+	}
+
+	PngEncWriter writer {param->png_mem_addr, param->png_mem_size, 0, false};
+
+	int ok = 0;
+	{
+		std::scoped_lock lock(g_png_enc_mutex);
+
+		stbi_write_png_compression_level = param->compression_level;
+		stbi_write_force_png_filter      = PngEncForcedFilter(param->filter_type);
+
+		ok = stbi_write_png_to_func(PngEncWrite, &writer, static_cast<int>(width),
+		                            static_cast<int>(height), components, pixels.get(),
+		                            static_cast<int>(row_size));
+	}
+
+	if (output_info != nullptr) {
+		output_info->data_size        = (writer.overflow ? 0 : writer.size);
+		output_info->processed_height = (ok != 0 && !writer.overflow ? height : 0);
+	}
+
+	if (ok == 0) {
+		LOGF("\t encode failed\n");
+		return PNG_ENC_ERROR_FATAL;
+	}
+
+	if (writer.overflow) {
+		LOGF("\t output buffer too small\n");
+		return PNG_ENC_ERROR_DATA_OVERFLOW;
+	}
+
+	LOGF("\t data_size = %" PRIu32 "\n", writer.size);
+
+	return static_cast<int32_t>(writer.size);
+}
+
+static int32_t KYTY_SYSV_ABI PngEncDelete(void* handle) {
+	PRINT_NAME();
+
+	if (handle == nullptr) {
+		return PNG_ENC_ERROR_INVALID_HANDLE;
+	}
+
+	auto* ctx = static_cast<PngEncContext*>(handle);
+	if (ctx->magic != PNG_ENC_CONTEXT_MAGIC) {
+		return PNG_ENC_ERROR_INVALID_HANDLE;
+	}
+
+	ctx->magic = 0;
+	return 0;
+}
+
+} // namespace PngEnc
+
+LIB_DEFINE(InitPngEnc_1) {
+	PRINT_NAME_ENABLE(true);
+
+	LIB_FUNC("9030RnBDoh4", PngEnc::PngEncQueryMemorySize);
+	LIB_FUNC("7aGTPfrqT9s", PngEnc::PngEncCreate);
+	LIB_FUNC("xgDjJKpcyHo", PngEnc::PngEncEncode);
+	LIB_FUNC("RUrWdwTWZy8", PngEnc::PngEncDelete);
+}
+
+} // namespace Libs

src/libs/libVideoOut.cpp

@@ -39,6 +39,8 @@ LIB_DEFINE(InitVideoOut_1) {
 	LIB_FUNC("Mt4QHHkxkOc", VideoOut::VideoOutGetEventCount);
 	LIB_FUNC("j6RaAUlaLv0", VideoOut::VideoOutWaitVblank);
 	LIB_FUNC("utPrVdxio-8", VideoOut::VideoOutGetOutputStatus);
+	LIB_FUNC("5tRaBjtdTzY", VideoOut::VideoOutVrrPegToFixedRate);
+	LIB_FUNC("T4ucGB8CsnM", VideoOut::VideoOutVrrUnpegFromFixedRate);
 	LIB_FUNC("+I4K03i3EL0", VideoOut::VideoOutInitializeOutputOptions);
 	LIB_FUNC("Nv8c-Kb+DUM", VideoOut::VideoOutIsOutputSupported);
 	LIB_FUNC("w0hLuNarQxY", VideoOut::VideoOutConfigureOutput);
@@ -46,6 +48,7 @@ LIB_DEFINE(InitVideoOut_1) {
 	LIB_FUNC("MCJ8SkzsQxY", VideoOut::VideoOutLatencyMeasureSetStartPoint);
 	LIB_FUNC("DYhhWbJSeRg", VideoOut::VideoOutColorSettingsSetGamma);
 	LIB_FUNC("pv9CI5VC+R0", VideoOut::VideoOutAdjustColor);
+	LIB_FUNC("w7Ipp9Xl7hg", VideoOut::VideoOutAllowOutputResolutionWqhdDetection);
 }
 
 } // namespace LibGen5

src/libs/libs.cpp

@@ -91,6 +91,7 @@ LIB_DEFINE(InitNet_1);
 LIB_DEFINE(InitPad_1);
 LIB_DEFINE(InitPlayGo_1);
 LIB_DEFINE(InitPngDec_1);
+LIB_DEFINE(InitPngEnc_1);
 LIB_DEFINE(InitPlatform_1);
 LIB_DEFINE(InitRudp_1);
 LIB_DEFINE(InitSaveData_1);
@@ -129,6 +130,7 @@ void InitAll(Loader::SymbolDatabase* s) {
 	InitPlayGo_1(s);
 	LibPsml::InitPsml_1(s);
 	InitPngDec_1(s);
+	InitPngEnc_1(s);
 	InitPlatform_1(s);
 	InitRudp_1(s);
 	LibRtc::InitRtc_1(s);

src/main.cpp

@@ -53,6 +53,7 @@ static void PrintUsage() {
 	::printf("  --user-id <num>                      Local user ID. Default: %d.\n",
 	         Config::DEFAULT_USER_ID);
 	::printf("  --mic <name>                        Capture from this microphone; omit for silence.\n");
+	::printf("  --controller-color <#RRGGBB>        Override the controller lightbar color.\n");
 	::printf(
 	    "  --present-mode <value>               Fifo, Mailbox, or Immediate. Default: Mailbox.\n");
 	::printf(
@@ -145,6 +146,20 @@ static bool ParseUint32(const std::string& value, uint32_t& out) {
 	return true;
 }
 
+static bool ParseControllerColor(const std::string& value, Config::ControllerColor& out) {
+	if (value.size() != 7 || value[0] != '#') {
+		return false;
+	}
+	uint32_t rgb = 0;
+	auto [end, error] = std::from_chars(value.data() + 1, value.data() + value.size(), rgb, 16);
+	if (error != std::errc {} || end != value.data() + value.size()) {
+		return false;
+	}
+	out = {static_cast<uint8_t>(rgb >> 16), static_cast<uint8_t>(rgb >> 8),
+	       static_cast<uint8_t>(rgb)};
+	return true;
+}
+
 static bool ParseInt32(const std::string& value, int32_t& out) {
 	int32_t number    = 0;
 	auto [end, error] = std::from_chars(value.data(), value.data() + value.size(), number);
@@ -302,6 +317,13 @@ static bool ParseArgs(int argc, char* argv[], RunOptions& options, bool& show_he
 			}
 		} else if (arg == "--mic") {
 			options.config.audio_input_device = value;
+		} else if (arg == "--controller-color") {
+			Config::ControllerColor color {};
+			if (!ParseControllerColor(value, color)) {
+				::printf("invalid controller color (expected #RRGGBB): %s\n", value.c_str());
+				return false;
+			}
+			options.config.controller_color = color;
 		} else if (arg == "--present-mode") {
 			if (!ParseEnum(value, options.config.present_mode)) {
 				::printf("invalid present mode: %s\n", value.c_str());

tests/PngEncTests.cpp

@@ -0,0 +1,247 @@
+#include "libs/libPngEnc.cpp"
+
+#define STB_IMAGE_IMPLEMENTATION
+#define STBI_ONLY_PNG
+#define STBI_NO_STDIO
+#include "stb_image.h"
+
+#include <algorithm>
+#include <array>
+#include <cstdio>
+#include <cstdlib>
+#include <vector>
+
+namespace Loader {
+void SymbolDatabase::Add(const SymbolResolve&, uint64_t, const std::string&) {}
+namespace Timer {
+double GetTimeMs() {
+	return 0.0;
+}
+} // namespace Timer
+} // namespace Loader
+
+namespace {
+using namespace Libs::PngEnc;
+
+#define CHECK(condition)                                                                           \
+	do {                                                                                           \
+		if (!(condition)) {                                                                        \
+			std::fprintf(stderr, "PngEncTests:%d: %s\n", __LINE__, #condition);                    \
+			std::abort();                                                                          \
+		}                                                                                          \
+	} while (false)
+
+struct Encoder {
+	PngEncContext context {};
+	void*         handle = nullptr;
+
+	explicit Encoder(uint32_t max_filters = 4) {
+		const PngEncCreateParam create {sizeof(PngEncCreateParam), 0, PNG_ENC_MAX_IMAGE_WIDTH,
+		                                max_filters};
+		CHECK(PngEncQueryMemorySize(&create) == sizeof(context));
+		CHECK(PngEncCreate(&create, &context, sizeof(context), &handle) == 0);
+	}
+
+	~Encoder() {
+		CHECK(PngEncDelete(handle) == 0);
+		CHECK(PngEncDelete(handle) == PNG_ENC_ERROR_INVALID_HANDLE);
+	}
+};
+
+PngEncEncodeParam EncodeParam(const uint8_t* source, uint32_t source_size, uint8_t* destination,
+                              uint32_t destination_size) {
+	PngEncEncodeParam param {};
+	param.image_mem_addr    = source;
+	param.png_mem_addr      = destination;
+	param.image_mem_size    = source_size;
+	param.png_mem_size      = destination_size;
+	param.image_width       = 2;
+	param.image_height      = 2;
+	param.image_pitch       = 8;
+	param.color_space       = PNG_ENC_COLOR_SPACE_RGBA;
+	param.bit_depth         = 8;
+	param.compression_level = 6;
+	return param;
+}
+
+uint32_t ReadBe32(const uint8_t* bytes) {
+	return (uint32_t {bytes[0]} << 24u) | (uint32_t {bytes[1]} << 16u) |
+	       (uint32_t {bytes[2]} << 8u) | bytes[3];
+}
+
+void TestPixels() {
+	Encoder                           encoder;
+	constexpr std::array<uint8_t, 16> rgba {255, 0, 13,  0,  4,   210, 6,   255,
+	                                        7,   8, 220, 64, 111, 12,  113, 128};
+	for (const bool bgr: {false, true}) {
+		for (const bool alpha: {false, true}) {
+			for (const bool padded: {false, true}) {
+				const uint32_t       pitch = padded ? 12 : 8;
+				std::vector<uint8_t> source(pitch * 2, 0xee);
+				std::vector<uint8_t> expected;
+				for (uint32_t y = 0; y < 2; ++y) {
+					for (uint32_t x = 0; x < 2; ++x) {
+						const auto* pixel = rgba.data() + (y * 2 + x) * 4;
+						auto*       input = source.data() + y * pitch + x * 4;
+						input[0]          = pixel[bgr ? 2 : 0];
+						input[1]          = pixel[1];
+						input[2]          = pixel[bgr ? 0 : 2];
+						input[3]          = pixel[3];
+						expected.insert(expected.end(), pixel, pixel + (alpha ? 4 : 3));
+					}
+				}
+				std::array<uint8_t, 1024> png {};
+				auto param = EncodeParam(source.data(), source.size(), png.data(), png.size());
+				param.pixel_format =
+				    bgr ? PNG_ENC_PIXEL_FORMAT_B8G8R8A8 : PNG_ENC_PIXEL_FORMAT_R8G8B8A8;
+				param.color_space = alpha ? PNG_ENC_COLOR_SPACE_RGBA : PNG_ENC_COLOR_SPACE_RGB;
+				param.image_pitch = pitch;
+				PngEncOutputInfo info {};
+				const auto       size = PngEncEncode(encoder.handle, &param, &info);
+				CHECK(size > 0 && info.data_size == static_cast<uint32_t>(size));
+				CHECK(info.processed_height == 2);
+				int   width = 0, height = 0, components = 0;
+				auto* decoded =
+				    stbi_load_from_memory(png.data(), size, &width, &height, &components, 0);
+				CHECK(decoded != nullptr);
+				CHECK(width == 2 && height == 2 && components == (alpha ? 4 : 3));
+				CHECK(std::memcmp(decoded, expected.data(), expected.size()) == 0);
+				stbi_image_free(decoded);
+			}
+		}
+	}
+}
+
+void TestFilterMasks() {
+	Encoder                            encoder;
+	alignas(4) std::array<uint8_t, 16> pixels {};
+	std::array<uint8_t, 1024>          png {};
+	auto param = EncodeParam(pixels.data(), pixels.size(), png.data(), png.size());
+	for (uint16_t mask = 0; mask <= 15; ++mask) {
+		param.filter_type = mask;
+		const auto size   = PngEncEncode(encoder.handle, &param, nullptr);
+		CHECK(size > 0);
+		CHECK(std::memcmp(png.data() + 37, "IDAT", 4) == 0);
+		std::array<char, 18> filtered {};
+		CHECK(stbi_zlib_decode_buffer(filtered.data(), filtered.size(),
+		                              reinterpret_cast<const char*>(png.data() + 41),
+		                              ReadBe32(png.data() + 33)) == filtered.size());
+		for (size_t row = 0; row < 2; ++row) {
+			const auto filter = static_cast<uint8_t>(filtered[row * 9]);
+			CHECK(filter <= 4);
+			CHECK(filter == 0 ? (mask == 0 || mask == 15) : (mask & (1u << (filter - 1u))) != 0);
+		}
+	}
+	param.filter_type = 16;
+	CHECK(PngEncEncode(encoder.handle, &param, nullptr) == PNG_ENC_ERROR_INVALID_PARAM);
+	for (uint32_t count = 0; count <= 4; ++count) {
+		Encoder limited(count);
+		for (uint16_t mask = 0; mask <= 15; ++mask) {
+			param.filter_type = mask;
+			const auto result = PngEncEncode(limited.handle, &param, nullptr);
+			CHECK(std::popcount(mask) <= count ? result > 0
+			                                   : result == PNG_ENC_ERROR_INVALID_PARAM);
+		}
+	}
+}
+
+void TestOutputCapacity() {
+	Encoder                            encoder;
+	alignas(4) std::array<uint8_t, 16> pixels {};
+	std::array<uint8_t, 1024>          png {};
+	auto       param = EncodeParam(pixels.data(), pixels.size(), png.data(), png.size());
+	const auto size  = PngEncEncode(encoder.handle, &param, nullptr);
+	CHECK(size > 0);
+	std::vector<uint8_t> guarded(size + 2, 0xa5);
+	param.png_mem_addr = guarded.data() + 1;
+	param.png_mem_size = size;
+	PngEncOutputInfo info {};
+	CHECK(PngEncEncode(encoder.handle, &param, &info) == size);
+	CHECK(guarded.front() == 0xa5 && guarded.back() == 0xa5);
+	CHECK(std::memcmp(guarded.data() + 1, png.data(), size) == 0);
+	std::fill(guarded.begin(), guarded.end(), 0xa5);
+	param.png_mem_size = size - 1;
+	CHECK(PngEncEncode(encoder.handle, &param, &info) == PNG_ENC_ERROR_DATA_OVERFLOW);
+	CHECK(info.data_size == 0 && info.processed_height == 0);
+	CHECK(std::all_of(guarded.begin(), guarded.end(), [](uint8_t byte) { return byte == 0xa5; }));
+}
+
+void TestValidation() {
+	Encoder           encoder;
+	PngEncCreateParam create {sizeof(PngEncCreateParam), 0, 2, 4};
+	CHECK(PngEncQueryMemorySize(nullptr) == PNG_ENC_ERROR_INVALID_ADDR);
+	CHECK(PngEncCreate(&create, nullptr, sizeof(PngEncContext), &encoder.handle) ==
+	      PNG_ENC_ERROR_INVALID_ADDR);
+	CHECK(PngEncCreate(&create, &encoder.context, 0, &encoder.handle) ==
+	      PNG_ENC_ERROR_INVALID_SIZE);
+	create.max_image_width = 0;
+	CHECK(PngEncQueryMemorySize(&create) == PNG_ENC_ERROR_INVALID_SIZE);
+	create.max_image_width = PNG_ENC_MAX_IMAGE_WIDTH + 1;
+	CHECK(PngEncQueryMemorySize(&create) == PNG_ENC_ERROR_INVALID_SIZE);
+	create.max_image_width = 2;
+	create.this_size       = 0;
+	CHECK(PngEncQueryMemorySize(&create) == PNG_ENC_ERROR_INVALID_PARAM);
+	create.this_size         = sizeof(create);
+	create.max_filter_number = 5;
+	CHECK(PngEncQueryMemorySize(&create) == PNG_ENC_ERROR_INVALID_PARAM);
+	alignas(4) std::array<uint8_t, 16> pixels {};
+	std::array<uint8_t, 1024>          png {};
+	const auto valid = EncodeParam(pixels.data(), pixels.size(), png.data(), png.size());
+	CHECK(PngEncEncode(nullptr, &valid, nullptr) == PNG_ENC_ERROR_INVALID_HANDLE);
+	CHECK(PngEncEncode(encoder.handle, nullptr, nullptr) == PNG_ENC_ERROR_INVALID_PARAM);
+	for (const auto width: {0u, PNG_ENC_MAX_IMAGE_WIDTH + 1, 0x40000000u, 0xffffffffu}) {
+		auto param        = valid;
+		param.image_width = width;
+		CHECK(PngEncEncode(encoder.handle, &param, nullptr) == PNG_ENC_ERROR_INVALID_SIZE);
+	}
+	for (const auto height: {0u, PNG_ENC_MAX_IMAGE_HEIGHT + 1, 0xffffffffu}) {
+		auto param           = valid;
+		param.image_width    = 1;
+		param.image_height   = height;
+		param.image_mem_size = 0xffffffffu;
+		CHECK(PngEncEncode(encoder.handle, &param, nullptr) == PNG_ENC_ERROR_INVALID_SIZE);
+	}
+	for (const auto height: {262144u, 400000u, 0x7ffffu}) {
+		auto param           = valid;
+		param.image_width    = 1024;
+		param.image_height   = height;
+		param.image_pitch    = 4096;
+		param.image_mem_size = 0xffffffffu;
+		CHECK(PngEncEncode(encoder.handle, &param, nullptr) == PNG_ENC_ERROR_INVALID_SIZE);
+	}
+	for (const auto pitch: {0u, 7u, 9u, 0xfffffffcu}) {
+		auto param        = valid;
+		param.image_pitch = pitch;
+		CHECK(PngEncEncode(encoder.handle, &param, nullptr) == PNG_ENC_ERROR_INVALID_SIZE);
+	}
+	auto param           = valid;
+	param.image_pitch    = 12;
+	param.image_mem_size = 20;
+	CHECK(PngEncEncode(encoder.handle, &param, nullptr) == PNG_ENC_ERROR_INVALID_SIZE);
+	param = valid;
+	param.image_mem_size--;
+	CHECK(PngEncEncode(encoder.handle, &param, nullptr) == PNG_ENC_ERROR_INVALID_SIZE);
+	param                = valid;
+	param.image_mem_addr = nullptr;
+	CHECK(PngEncEncode(encoder.handle, &param, nullptr) == PNG_ENC_ERROR_INVALID_ADDR);
+	param.image_mem_addr = pixels.data() + 1;
+	CHECK(PngEncEncode(encoder.handle, &param, nullptr) == PNG_ENC_ERROR_INVALID_ADDR);
+	param           = valid;
+	param.bit_depth = 16;
+	CHECK(PngEncEncode(encoder.handle, &param, nullptr) == PNG_ENC_ERROR_INVALID_PARAM);
+	param             = valid;
+	param.clut_number = 1;
+	CHECK(PngEncEncode(encoder.handle, &param, nullptr) == PNG_ENC_ERROR_INVALID_PARAM);
+	param                   = valid;
+	param.compression_level = 10;
+	CHECK(PngEncEncode(encoder.handle, &param, nullptr) == PNG_ENC_ERROR_INVALID_PARAM);
+}
+} // namespace
+
+int main() {
+	TestPixels();
+	TestFilterMasks();
+	TestOutputCapacity();
+	TestValidation();
+	std::puts("PNG encoder tests passed");
+}

tests/ShaderRecompilerComputeTests.cpp

@@ -2402,6 +2402,232 @@ public:
     std::printf("[host]    %-32s ok\n", "StreamBufferRing");
   }
 
+  void CheckGpuSuspendPoint() {
+    constexpr const char *name = "GpuSuspendPoint";
+    EnsureRuntimeContext();
+    auto &context = Renderer();
+    context.InitializeGpu(nullptr);
+    auto &gpu = context.GetGpu();
+    auto &scheduler = context.GetCommandScheduler();
+    CommandProcessor setup(context, 0);
+
+    vk::SemaphoreTypeCreateInfo timeline_type{};
+    timeline_type.sType = vk::StructureType::eSemaphoreTypeCreateInfo;
+    timeline_type.semaphoreType = vk::SemaphoreType::eTimeline;
+    vk::SemaphoreCreateInfo timeline_create{};
+    timeline_create.sType = vk::StructureType::eSemaphoreCreateInfo;
+    timeline_create.pNext = &timeline_type;
+    vk::Semaphore gate = nullptr;
+    Require(name, "timeline create",
+            m_runtime_context.device.createSemaphore(
+                &timeline_create, nullptr, &gate) == vk::Result::eSuccess,
+            "failed to create the native completion gate");
+    const auto signal = [&](uint64_t value) {
+      vk::SemaphoreSignalInfo info{};
+      info.sType = vk::StructureType::eSemaphoreSignalInfo;
+      info.semaphore = gate;
+      info.value = value;
+      Require(name, "timeline signal",
+              m_runtime_context.device.signalSemaphore(&info) ==
+                  vk::Result::eSuccess,
+              "failed to release native GPU work");
+    };
+    const auto finish = [&] {
+      gpu.WaitForIdle();
+      gpu.SendCommandSync([&] {
+        scheduler.Finish();
+        scheduler.WaitPriorityOperations(scheduler.CurrentTick() - 1);
+      });
+    };
+    const auto write = [](uint32_t *packet, uint32_t *destination,
+                          uint32_t value, bool predicated = false) {
+      const auto address = reinterpret_cast<uint64_t>(destination);
+      packet[0] = KYTY_PM4(5, Pm4::IT_WRITE_DATA, 0) | uint32_t{predicated};
+      packet[1] = 0;
+      packet[2] = static_cast<uint32_t>(address);
+      packet[3] = static_cast<uint32_t>(address >> 32u);
+      packet[4] = value;
+    };
+
+    gpu.SendCommandSync([&] {
+      setup.BufferInit();
+      SubmitInfo blocked;
+      blocked.AddWait(gate, 1);
+      scheduler.Flush(blocked);
+    });
+    std::binary_semaphore first_returned{0};
+    std::jthread first([&] {
+      gpu.SuspendPoint();
+      first_returned.release();
+    });
+    const bool first_nonblocking =
+        first_returned.try_acquire_for(std::chrono::seconds(2));
+    if (!first_nonblocking) {
+      signal(1);
+    }
+    first.join();
+    Require(name, "first call", first_nonblocking,
+            "the first suspend point waited for native GPU completion");
+
+    // Let the marker reach Vulkan while its completion remains gated.
+    std::binary_semaphore marker_parsed{0};
+    std::jthread parser([&] {
+      gpu.WaitForIdle();
+      marker_parsed.release();
+    });
+    const bool parser_nonblocking =
+        marker_parsed.try_acquire_for(std::chrono::seconds(2));
+    if (!parser_nonblocking) {
+      signal(1);
+    }
+    parser.join();
+    Require(name, "marker insertion", parser_nonblocking,
+            "inserting a suspend point blocked the GPU command thread");
+
+    std::binary_semaphore second_entered{0};
+    std::binary_semaphore second_returned{0};
+    std::jthread second([&] {
+      second_entered.release();
+      gpu.SuspendPoint();
+      second_returned.release();
+    });
+    second_entered.acquire();
+    const bool second_overtook_completion =
+        second_returned.try_acquire_for(std::chrono::milliseconds(100));
+
+    uint32_t compute_value = 0;
+    std::array<uint32_t, 5> compute_commands{};
+    write(compute_commands.data(), &compute_value, 33);
+    bool host_lane_responded = false;
+    std::binary_semaphore compute_progress{0};
+    std::jthread compute([&] {
+      gpu.SubmitCompute(0x20, compute_commands);
+      gpu.SendCommandSync([&] { host_lane_responded = true; });
+      gpu.WaitForIdle();
+      compute_progress.release();
+    });
+    const bool compute_did_not_wait =
+        compute_progress.try_acquire_for(std::chrono::seconds(2));
+    signal(1);
+    second.join();
+    compute.join();
+    finish();
+    Require(name, "one outstanding point", !second_overtook_completion,
+            "a second suspend point returned before the first native completion");
+    Require(name, "independent queue progress",
+            compute_did_not_wait && host_lane_responded && compute_value == 33,
+            "waiting for a previous suspend point blocked compute submissions "
+            "or the host command lane");
+    Require(name, "frame count", gpu.GetFrameNum() == 2,
+            "suspend-point insertion did not advance the frame counter once");
+
+    // A marker must stay behind a stalled graphics stream, then clear its
+    // predicate before the next stream. Async compute releases that stream.
+    alignas(16) uint64_t predicate = 0;
+    uint32_t label = 0;
+    uint32_t prefix = 0;
+    uint32_t skipped = 0;
+    uint32_t after_reset = 0;
+    const auto predicate_address = reinterpret_cast<uint64_t>(&predicate);
+    const auto label_address = reinterpret_cast<uint64_t>(&label);
+    std::array<uint32_t, 21> before{};
+    before[0] = KYTY_PM4(4, Pm4::IT_SET_PREDICATION, 0);
+    before[1] = (3u << 16u) | (1u << 8u);
+    before[2] = static_cast<uint32_t>(predicate_address);
+    before[3] = static_cast<uint32_t>(predicate_address >> 32u);
+    write(before.data() + 4, &skipped, 11, true);
+    write(before.data() + 9, &prefix, 22);
+    before[14] = KYTY_PM4(7, Pm4::IT_WAIT_REG_MEM, 0);
+    before[15] = 0x10u | 3u;
+    before[16] = static_cast<uint32_t>(label_address);
+    before[17] = static_cast<uint32_t>(label_address >> 32u);
+    before[18] = 1;
+    before[19] = UINT32_MAX;
+    std::array<uint32_t, 5> after{};
+    write(after.data(), &after_reset, 44, true);
+    std::binary_semaphore lane_entered{0};
+    std::binary_semaphore release_lane{0};
+    gpu.SendCommand([&] {
+      lane_entered.release();
+      release_lane.acquire();
+    });
+    lane_entered.acquire();
+    gpu.Submit(before, {});
+    std::binary_semaphore queued_returned{0};
+    std::jthread queued([&] {
+      gpu.SuspendPoint();
+      queued_returned.release();
+    });
+    const bool queued_nonblocking =
+        queued_returned.try_acquire_for(std::chrono::seconds(2));
+    if (!queued_nonblocking) {
+      release_lane.release();
+      gpu.SendCommandSync([&] { label = 1; });
+      queued.join();
+      finish();
+      Require(name, "queued insertion", false,
+              "suspend-point insertion drained pending guest graphics");
+    }
+    queued.join();
+    gpu.Submit(after, {});
+    release_lane.release();
+    bool prefix_reached = false;
+    bool graphics_ordered = false;
+    const auto deadline = std::chrono::steady_clock::now() +
+                          std::chrono::seconds(2);
+    do {
+      gpu.SendCommandSync([&] {
+        prefix_reached = prefix == 22;
+        graphics_ordered = skipped == 0 && after_reset == 0;
+      });
+      std::this_thread::yield();
+    } while (!prefix_reached && std::chrono::steady_clock::now() < deadline);
+    std::array<uint32_t, 5> release_graphics{};
+    write(release_graphics.data(), &label, 1);
+    gpu.SubmitCompute(0x20, release_graphics);
+    finish();
+    Require(name, "graphics FIFO",
+            prefix_reached && graphics_ordered && skipped == 0,
+            "suspend point or later graphics overtook a stalled graphics stream");
+    Require(name, "graphics reset", after_reset == 44,
+            "suspend point preserved the previous frame's predicate");
+
+    // Delay the completion callback beyond the owner's shutdown when possible.
+    // Its credit must remain alive independently of the GuestGpu object.
+    std::binary_semaphore priority_entered{0};
+    std::binary_semaphore release_priority{0};
+    gpu.SendCommandSync([&] {
+      scheduler.DeferPriorityOperation([&] {
+        priority_entered.release();
+        release_priority.acquire();
+      });
+      SubmitInfo blocked;
+      blocked.AddWait(gate, 2);
+      scheduler.Flush(blocked);
+    });
+    gpu.SuspendPoint();
+    gpu.WaitForIdle();
+    std::binary_semaphore shutdown_returned{0};
+    std::jthread shutdown([&] {
+      context.ShutdownGpu();
+      shutdown_returned.release();
+    });
+    const bool shutdown_overtook_native =
+        shutdown_returned.try_acquire_for(std::chrono::milliseconds(100));
+    signal(2);
+    priority_entered.acquire();
+    if (!shutdown_overtook_native) {
+      (void)shutdown_returned.try_acquire_for(std::chrono::seconds(2));
+    }
+    release_priority.release();
+    shutdown.join();
+    scheduler.DrainPriorityOperations();
+    m_runtime_context.device.destroySemaphore(gate, nullptr);
+    Require(name, "shutdown native completion", !shutdown_overtook_native,
+            "shutdown returned while the suspend point was still executing");
+    std::printf("[host]    %-32s ok\n", name);
+  }
+
   void CheckGpuCommandLane() {
     EnsureRuntimeContext();
     auto &context = Renderer();
@@ -2647,7 +2873,7 @@ public:
     graphics_gate_entered.acquire();
     live_graphics_commands[4] = 22;
     graphics_gate_release.release();
-    gpu.Done();
+    gpu.WaitForIdle();
     Require("GpuCommandLane", "borrowed graphics commands",
             live_graphics_value == 22,
             "graphics submission executed a copied PM4 stream");
@@ -2665,7 +2891,7 @@ public:
     compute_gate_entered.acquire();
     live_compute_commands[4] = 44;
     compute_gate_release.release();
-    gpu.Done();
+    gpu.WaitForIdle();
     Require("GpuCommandLane", "borrowed compute commands",
             live_compute_value == 44,
             "compute submission executed a copied PM4 stream");
@@ -2721,7 +2947,7 @@ public:
     write_packet(no_interrupt_graphics_commands.data(),
                  &no_interrupt_graphics_value, 55);
     gpu.Submit(no_interrupt_graphics_commands, {});
-    gpu.Done();
+    gpu.WaitForIdle();
     finish_gpu();
 
     LibKernel::EventQueue::KernelEvent interrupt_event{};
@@ -2745,7 +2971,7 @@ public:
     interrupt_gate_entered.acquire();
     live_interrupt_commands[2] = 0;
     interrupt_gate_release.release();
-    gpu.Done();
+    gpu.WaitForIdle();
     finish_gpu();
 
     interrupt_count = 0;
@@ -2761,7 +2987,7 @@ public:
     auto graphics_interrupt_commands = make_interrupt_packet(
         1, 1, &graphics_interrupt_label, 0x11223344u, 0);
     gpu.Submit(graphics_interrupt_commands, {});
-    gpu.Done();
+    gpu.WaitForIdle();
     finish_gpu();
 
     interrupt_count = 0;
@@ -2787,7 +3013,7 @@ public:
     compute_interrupt_gate_entered.acquire();
     compute_interrupt_commands[2] |= 1u << 24u;
     compute_interrupt_gate_release.release();
-    gpu.Done();
+    gpu.WaitForIdle();
     finish_gpu();
 
     interrupt_count = 0;
@@ -2803,7 +3029,7 @@ public:
     auto compute_clock_commands = make_interrupt_packet(
         3, 1, &compute_clock_label, 0, 0x567u);
     gpu.SubmitCompute(0x20, compute_clock_commands);
-    gpu.Done();
+    gpu.WaitForIdle();
     finish_gpu();
 
     interrupt_count = 0;
@@ -2821,7 +3047,7 @@ public:
         2, 2, &compute_done_label, compute_done_value, 0x678u);
     compute_done_commands[1] = 0x62fu; // CS_DONE, shader-done index, no GCR action.
     gpu.SubmitCompute(0x20, compute_done_commands);
-    gpu.Done();
+    gpu.WaitForIdle();
     finish_gpu();
 
     interrupt_count = 0;
@@ -2896,7 +3122,7 @@ public:
       });
       std::this_thread::yield();
     }
-    gpu.Done();
+    gpu.WaitForIdle();
     Require("GpuCommandLane", "packet-boundary command polling",
             packet_marker_a_at_callback == 11 &&
                 packet_marker_b_at_callback == 0 && packet_marker_a == 11 &&
@@ -2934,21 +3160,21 @@ public:
     uint32_t ordered_suffix = 0;
     std::jthread ordered([&] {
       ordered_started.release();
-      gpu.Done();
+      gpu.WaitForIdle();
       ordered_suffix = suffix;
       ordered_finished = true;
     });
     ordered_started.acquire();
     gpu.SendCommandSync([&] {
-      Require("GpuCommandLane", "submit done barrier", !ordered_finished.load(),
-              "submit done returned before a queued submission");
+      Require("GpuCommandLane", "guest queue idle fence", !ordered_finished.load(),
+              "idle fence returned before a queued submission");
       label = 1;
     });
     ordered.join();
     Require("GpuCommandLane", "ordered completion",
             prefix == 11 && suffix == 22 && ordered_suffix == 22 &&
                 ordered_finished.load(),
-            "submit done did not drain prior PM4 work");
+            "idle fence did not drain prior PM4 work");
 
     auto &resources = context;
     constexpr uint64_t empty_unmap_base = 0x0000000200400000ull;
@@ -2971,7 +3197,7 @@ public:
       unmap_complete.acquire();
     }
     unmap_thread.join();
-    gpu.Done();
+    gpu.WaitForIdle();
     Require("GpuCommandLane", "unmap queue progress",
             unmap_returned &&
                 !resources.IsMapped(empty_unmap_base, empty_unmap_size) &&
@@ -3098,7 +3324,7 @@ public:
             dma_cursor == dma_commands.size(),
             "DMA_DATA packet stream has the wrong size");
     gpu.Submit(dma_commands, {});
-    gpu.Done();
+    gpu.WaitForIdle();
     constexpr uint32_t clean_fill_value = 0xdecafbad;
     gpu.SendCommandSync([&] {
       auto &buffer_cache = resources.GetBufferCache();
@@ -3336,15 +3562,12 @@ public:
     std::binary_semaphore command_entered{0};
     std::binary_semaphore release_command{0};
     std::atomic<bool> shutdown_complete{false};
-    std::atomic<bool> submission_lane_entered{false};
     bool shutdown_owner_visible = false;
     resources.MapMemory(empty_unmap_base, empty_unmap_size);
     gpu.SendCommand([&] {
       command_entered.release();
       release_command.acquire();
       shutdown_owner_visible = &context.GetGpu() == &gpu;
-      gpu.Done();
-      submission_lane_entered = true;
     });
     command_entered.acquire();
     std::jthread shutdown_thread([&] {
@@ -3360,8 +3583,7 @@ public:
     shutdown_thread.join();
     Require(
         "GpuCommandLane", "owned shutdown completion",
-        shutdown_complete.load() && submission_lane_entered.load() &&
-            shutdown_owner_visible,
+        shutdown_complete.load() && shutdown_owner_visible,
         "GPU owner did not remain available while draining queued work");
     resources.UnmapMemory(empty_unmap_base, empty_unmap_size);
     Require("GpuCommandLane", "unmap after shutdown",
@@ -23539,6 +23761,158 @@ TestCase VectorVopcCmpxLtU16CapturedSdwaExecMask() {
   return test;
 }
 
+TestCase VectorVopcCmpxLtI16CapturedSdwaExecMask() {
+  using O = ShaderOpcode;
+  enum Encoding { Captured, Compact, Vop3, HighLhs, HighRhs, SdwaOperands };
+  struct CompareCase {
+    u32 lhs;
+    u32 rhs;
+    u32 expected_exec;
+    Encoding encoding = Captured;
+    u32 incoming_exec = 1;
+    u32 src0_sel = 6;
+    u32 src1_sel = 6;
+    u32 src0_sext = 0;
+    u32 src1_sext = 0;
+  };
+  const std::array<CompareCase, 30> cases{{
+      {0xffff000eu, 15u, 1}, // Ignore the high half: 14 < 15.
+      {0x8000000fu, 15u, 0}, // Equality is false despite a negative high half.
+      {0x00000010u, 15u, 0},
+      {0x7fff0000u, 15u, 1},
+      {0x0000ffffu, 15u, 1}, // Signed -1, not unsigned 65535.
+      {0x00008000u, 15u, 1}, // Minimum signed halfword.
+      {0x00007fffu, 15u, 0}, // Maximum signed halfword.
+      {0x0000fff0u, 15u, 1},
+      {0x0000ffffu, 15u, 0, Captured, 0}, // Cannot reactivate an inactive lane.
+      {0xffff000fu, 15u, 0, Captured, 0},
+      {0x1234ffffu, 0xabcd0000u, 1, Compact}, // -1 < 0.
+      {0xffff0000u, 0x0000ffffu, 0, Compact}, // 0 < -1 is false.
+      {0xabcd8000u, 0x12347fffu, 1, Vop3},
+      {0x12347fffu, 0xabcd8000u, 0, Vop3},
+      {0xffff0010u, 15u, 1, HighLhs}, // Select -1 in WORD_1.
+      {0x0010ffffu, 15u, 0, HighLhs}, // Select 16, not the negative low half.
+      {0x1234fffeu, 0xffff8000u, 1, HighRhs}, // -2 < selected -1.
+      {0x1234ffffu, 0xabcd0000u, 1, SdwaOperands, 1, 4, 4, 1, 1},
+      {0x12345680u, 0u, 1, SdwaOperands, 1, 0, 6, 1, 0}, // BYTE_0 SEXT: -128.
+      {0x1234567fu, 0u, 0, SdwaOperands, 1, 0, 6, 1, 0}, // +127.
+      {0x12345680u, 0u, 0, SdwaOperands, 1, 0, 6, 0, 0}, // Without SEXT: +128.
+      {0x12348000u, 0u, 1, SdwaOperands, 1, 1, 6, 1, 0},
+      {0x12803456u, 0u, 1, SdwaOperands, 1, 2, 6, 1, 0},
+      {0x80345678u, 0u, 1, SdwaOperands, 1, 3, 6, 1, 0},
+      {0x0000ffffu, 0x12345680u, 0, SdwaOperands, 1, 6, 0, 0, 1}, // -1 < -128 is false.
+      {0x0000ff7fu, 0x12345680u, 1, SdwaOperands, 1, 6, 0, 0, 1}, // -129 < -128.
+      {0x0000ffffu, 0x12347f00u, 1, SdwaOperands, 1, 6, 1, 0, 1},
+      {0u, 0x12803456u, 0, SdwaOperands, 1, 6, 2, 0, 1},
+      {0u, 0x80345678u, 0, SdwaOperands, 1, 6, 3, 0, 1},
+      {0xff000001u, 0x000100ffu, 1, SdwaOperands, 1, 3, 2, 1, 1},
+  }};
+  constexpr u32 vcc_lo = 0x76543210u;
+  constexpr u32 vcc_hi = 0x89abcdefu;
+  TestCase test;
+  test.name = "VectorVopcCmpxLtI16CapturedSdwaExecMask";
+  for (const auto &entry : cases) {
+    test.initial.push_back(entry.lhs);
+  }
+  test.expected = test.initial;
+  auto &code = test.code;
+  for (u32 i = 0; i < cases.size(); ++i) {
+    const auto &entry = cases[i];
+    AppendVMovU32(&code, 30, i * 4u);
+    AppendBufferLoadDword(&code, 1, 30); // Runtime input prevents constant folding.
+    AppendVMovU32(&code, 2, entry.rhs);
+    AppendSMovLiteral(&code, 106, vcc_lo);
+    AppendSMovLiteral(&code, 107, vcc_hi);
+    code.push_back(EncodeSMovB32(126, InlineU32(entry.incoming_exec)));
+    switch (entry.encoding) {
+    case Compact: code.push_back(0x7d320501u); break;
+    case Vop3: code.insert(code.end(), {0xd499007eu, 0x00020501u}); break;
+    case HighLhs: code.insert(code.end(), {0x7d331ef9u, 0x86050001u}); break;
+    case HighRhs:
+      code.push_back(EncodeVopc(0x99u, 249u, 2u));
+      code.push_back(EncodeVopcSdwa(1u, 0u, 0u, 6u, 5u));
+      break;
+    case SdwaOperands:
+      code.push_back(EncodeVopc(0x99u, 249u, 2u));
+      code.push_back(EncodeVopcSdwa(1u, 0u, 0u, entry.src0_sel, entry.src1_sel,
+                                  entry.src0_sext, entry.src1_sext));
+      break;
+    default: code.insert(code.end(), {0x7d331ef9u, 0x86060001u}); break;
+    }
+    code.push_back(EncodeSMovB32(20, 126)); // Snapshot EXEC before restoring it.
+    code.push_back(EncodeSMovB32(21, 127));
+    code.push_back(EncodeSMovB32(126, InlineU32(1)));
+    const u32 out = static_cast<u32>(cases.size()) + i * 4u;
+    AppendStoreSgprPair(&code, 20, out);
+    AppendStoreSgprPair(&code, 106, out + 2u); // CMPX preserves both VCC halves.
+    test.expected.insert(test.expected.end(),
+                         {entry.expected_exec, 0u, vcc_lo, vcc_hi});
+  }
+  AppendEnd(&code);
+  test.opcodes = {O::V_MOV_B32, O::S_MOV_B32, O::BUFFER_LOAD_DWORD,
+                  O::V_CMPX_LT_I16, O::BUFFER_STORE_DWORD, O::S_ENDPGM};
+  test.decoded_counts = {{"V_CMPX_LT_I16", cases.size()}};
+  test.required_spirv = {"OpSLessThan", "OpBitFieldSExtract"};
+  return test;
+}
+
+TestCase VectorVopcCmpxLtI16WaveMasks(u32 wave_size) {
+  using O = ShaderOpcode;
+  constexpr u32 marker = 0xfeedabcdu;
+  constexpr u32 vcc_lo = 0x01234567u;
+  constexpr u32 vcc_hi = 0x89abcdefu;
+  constexpr u32 expected_exec_lo = 0xaaaaaaaau;
+  const u32 expected_exec_hi = wave_size == 64u ? 0x33333333u : 0u;
+  TestCase test;
+  test.name = wave_size == 32u ? "VectorVopcCmpxLtI16Wave32Masks"
+                              : "VectorVopcCmpxLtI16Wave64Masks";
+  test.initial.assign(6u * wave_size, 0xdeadbeefu);
+  constexpr u32 inputs[] = {0xffff000eu, 0x0000ffffu, 0x8000000fu, 0x12348000u};
+  for (u32 lane = 0; lane < wave_size; ++lane) {
+    test.initial[lane] = inputs[lane % 4u];
+  }
+  test.expected = test.initial;
+  for (u32 lane = 0; lane < wave_size; ++lane) {
+    test.expected[wave_size + lane] = expected_exec_lo;
+    test.expected[2u * wave_size + lane] = expected_exec_hi;
+    test.expected[3u * wave_size + lane] = vcc_lo;
+    test.expected[4u * wave_size + lane] = vcc_hi;
+    const u32 mask = lane < 32u ? expected_exec_lo : expected_exec_hi;
+    if ((mask & (1u << (lane % 32u))) != 0u) {
+      test.expected[5u * wave_size + lane] = marker;
+    }
+  }
+  auto &code = test.code;
+  code.push_back(EncodeVop2(0x1a, 30, InlineU32(2), 0));
+  AppendBufferLoadDword(&code, 1, 30);
+  AppendVMovLiteral(&code, 4, marker);
+  code.push_back(EncodeSop1(0x04, 10, 126)); // Save the full incoming EXEC.
+  AppendSMovLiteral(&code, 106, vcc_lo);
+  AppendSMovLiteral(&code, 107, vcc_hi);
+  AppendSMovLiteral(&code, 126, 0xeeeeeeeeu);
+  AppendSMovLiteral(&code, 127, 0x77777777u);
+  code.insert(code.end(), {0x7d331ef9u, 0x86060001u});
+  code.push_back(EncodeSMovB32(20, 126));
+  code.push_back(EncodeSMovB32(21, 127));
+  AppendStoreVgprAtLaneDwordOffset(&code, 4, 0, 5u * wave_size);
+  code.push_back(EncodeSop1(0x04, 126, 10)); // Restore lanes to inspect the raw masks.
+  AppendStoreSgprAtLaneDwordOffset(&code, 20, 0, wave_size);
+  AppendStoreSgprAtLaneDwordOffset(&code, 21, 0, 2u * wave_size);
+  AppendStoreSgprAtLaneDwordOffset(&code, 106, 0, 3u * wave_size);
+  AppendStoreSgprAtLaneDwordOffset(&code, 107, 0, 4u * wave_size);
+  AppendEnd(&code);
+  test.compute_info.threads_num[0] = wave_size;
+  test.compute_info.threads_num[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::V_MOV_B32, O::S_MOV_B32, O::S_MOV_B64, O::V_LSHLREV_B32,
+                  O::V_ADD_NC_U32, O::BUFFER_LOAD_DWORD, O::V_CMPX_LT_I16,
+                  O::BUFFER_STORE_DWORD, O::S_ENDPGM};
+  test.required_spirv = {"OpSLessThan", "OpBitFieldSExtract"};
+  return test;
+}
+
 TestCase VectorVopcCmpxEqU16SdwaCompactVop3ExecMask() {
   using O = ShaderOpcode;
   struct CompareCase {
@@ -23546,9 +23920,9 @@ TestCase VectorVopcCmpxEqU16SdwaCompactVop3ExecMask() {
     u32 rhs;
     u32 incoming_exec;
     u32 expected_exec;
-    u32 encoding = 0; // 0: SDWA, 1: compact, 2: VOP3, 3: SDWA sign-extend.
+    u32 encoding = 0; // SDWA, compact, VOP3, sign-extended words/bytes.
   };
-  const std::array<CompareCase, 12> cases{{
+  const std::array<CompareCase, 16> cases{{
       {0x12340000u, 0x56780000u, 1, 1}, // Equal low half; ignore high half.
       {0xffff0001u, 0xabcd0001u, 1, 1},
       {0x00000000u, 0x12340000u, 1, 1},
@@ -23561,6 +23935,10 @@ TestCase VectorVopcCmpxEqU16SdwaCompactVop3ExecMask() {
       {0x00008000u, 0x0000ffffu, 1, 0, 2},
       {0x0001ffffu, 0x12340001u, 1, 0, 3}, // Sign-extended low halves differ.
       {0x00000001u, 0x00010001u, 0, 0}, // CMPX cannot reactivate an inactive lane.
+      {0x12345680u, 0x0000ff80u, 1, 1, 4}, // SEXT BYTE_0 is unsigned U16 0xff80.
+      {0x123456ffu, 0x0000ffffu, 1, 1, 4},
+      {0x0000ff80u, 0x12348000u, 1, 1, 5}, // SEXT BYTE_1 on src1.
+      {0x1234ffffu, 0x5678ffffu, 1, 1, 6}, // SEXT WORD_0 versus DWORD low half.
   }};
   constexpr u32 vcc_hi = 0x89abcdefu;
   TestCase test;
@@ -23585,6 +23963,18 @@ TestCase VectorVopcCmpxEqU16SdwaCompactVop3ExecMask() {
       code.push_back(EncodeVopc(0xba, 249u, 1u));
       code.push_back(EncodeVopcSdwa(0u, 0u, 0u, 6u, 6u, 1u, 1u));
       break;
+    case 4:
+      code.push_back(EncodeVopc(0xba, 249u, 1u));
+      code.push_back(EncodeVopcSdwa(0u, 0u, 0u, 0u, 6u, 1u, 0u));
+      break;
+    case 5:
+      code.push_back(EncodeVopc(0xba, 249u, 1u));
+      code.push_back(EncodeVopcSdwa(0u, 0u, 0u, 6u, 1u, 0u, 1u));
+      break;
+    case 6:
+      code.push_back(EncodeVopc(0xba, 249u, 1u));
+      code.push_back(EncodeVopcSdwa(0u, 0u, 0u, 4u, 6u, 1u, 0u));
+      break;
     default:
       code.push_back(EncodeVopc(0xba, 249u, 1u));
       code.push_back(EncodeVopcSdwa(0u));
@@ -30372,6 +30762,10 @@ std::vector<TestCase> MakeCases() {
   AddCase(VectorVopcSdwaCmpxWritesExecMask);
   AddCase(VectorVopcCmpxGtU16CapturedSdwaExecMask);
   AddCase(VectorVopcCmpxLtU16CapturedSdwaExecMask);
+  AddCase(VectorVopcCmpxLtI16CapturedSdwaExecMask);
+  for (u32 wave_size : {32u, 64u}) {
+    cases.push_back(VectorVopcCmpxLtI16WaveMasks(wave_size));
+  }
   AddCase(VectorVopcCmpxEqU16SdwaCompactVop3ExecMask);
   AddCase(VectorVopcCmpNgtF16CapturedSdwaAndEdges);
   AddCase(VectorVopcCmpNltF16CapturedSdwaAndEdges);
@@ -33981,12 +34375,15 @@ void CheckPm4Predication(RenderContext &renderer) {
                          uint32_t wait, const void *source, bool skip) {
     predicated = unconditional = 0;
     const auto packet = commands(op, condition, wait, source);
+    const auto tick = renderer.GetCommandScheduler().CurrentTick();
     Pm4Execution execution;
     Require(name, stage,
             processor.Process(execution, packet) == Pm4ProcessResult::Complete &&
                 processor.ShouldSkipPredicatedPackets() == skip &&
                 predicated == (skip ? 0u : 11u) && unconditional == 22,
             "predicate polarity, tagged packet execution, or packet consumption is wrong");
+    Require(name, stage, renderer.GetCommandScheduler().CurrentTick() == tick,
+            "predication unnecessarily submitted and drained the host queue");
   };
   struct Case {
     uint64_t delta;
@@ -35290,6 +35687,13 @@ int main(int argc, char **argv) {
     RunCase(&vulkan, VectorVopcCmpxEqU16SdwaCompactVop3ExecMask());
     return 0;
   }
+  if (argc == 2 && std::strcmp(argv[1], "--cmpx-lt-i16-only") == 0) {
+    VulkanHarness vulkan;
+    RunCase(&vulkan, VectorVopcCmpxLtI16CapturedSdwaExecMask());
+    RunCase(&vulkan, VectorVopcCmpxLtI16WaveMasks(32));
+    RunCase(&vulkan, VectorVopcCmpxLtI16WaveMasks(64));
+    return 0;
+  }
   if (argc == 2 && std::strcmp(argv[1], "--cmpx-lt-u16-only") == 0) {
     VulkanHarness vulkan;
     RunCase(&vulkan, VectorVopcCmpxLtU16CapturedSdwaExecMask());
@@ -35506,6 +35910,11 @@ int main(int argc, char **argv) {
     CheckPm4CeCompletion(vulkan.RuntimeRenderer());
     return 0;
   }
+  if (argc == 2 && std::strcmp(argv[1], "--suspend-point-only") == 0) {
+    VulkanHarness vulkan;
+    vulkan.CheckGpuSuspendPoint();
+    return 0;
+  }
   if (argc == 2 && std::strcmp(argv[1], "--alignbyte-only") == 0) {
     VulkanHarness vulkan;
     RunCase(&vulkan, VectorAlignByteUsesFiveBitByteOffset());
@@ -35915,6 +36324,7 @@ int main(int argc, char **argv) {
   for (const auto &test : graphics_tests) {
     RunGraphicsCase(&vulkan, test);
   }
+  vulkan.CheckGpuSuspendPoint();
   vulkan.CheckGpuCommandLane();
   if (skipped_device_checks) {
     std::printf(