DualSense haptics and controller speaker sound bluethooth (#969)

* pad: send DualSense speaker and haptic audio over Bluetooth

* ci: fix Opus clang-cl build and macOS memory tracker test

* docs: remove Bluetooth notes from project README

* pad: keep Bluetooth audio work off emulation timing path

* pad: smooth DualSense speaker audio on USB and Bluetooth

* tests: find available macOS guest range for page manager

* pad: keep haptic waveform when Bluetooth queue overflows

* audio: reject zero output sample frequency

* Simplify DualSense audio routing and fix Bluetooth batching

---------

Co-authored-by: Codex <codex@openai.com>
Co-authored-by: nmzik <Nmzik@mail.ru>

3rdparty/CMakeLists.txt

@@ -2,10 +2,31 @@ set(BUILD_SHARED_LIBS OFF CACHE BOOL "" FORCE)
 
 add_subdirectory(cpuinfo)
 
-# The Windows guest red-zone patcher needs a decoder/encoder and a small
-# trampoline assembler. Pin the same revisions used by shadPS4's implementation.
 include(FetchContent)
 
+# The Bluetooth DualSense transports speaker audio as fixed-size Opus packets.
+set(kyty_build_testing_saved "${BUILD_TESTING}")
+set(BUILD_TESTING OFF)
+set(OPUS_BUILD_PROGRAMS OFF CACHE BOOL "" FORCE)
+set(OPUS_INSTALL_PKG_CONFIG_MODULE OFF CACHE BOOL "" FORCE)
+set(OPUS_INSTALL_CMAKE_CONFIG_MODULE OFF CACHE BOOL "" FORCE)
+if(MSVC AND CMAKE_C_COMPILER_ID STREQUAL "Clang")
+	# Opus assumes MSVC accepts SSE4.1 intrinsics without per-file flags, but
+	# clang-cl rejects them in its baseline x64 files. Its AVX2 dispatch also
+	# depends on the SSE4.1 declarations, so retain the x64 SSE2 baseline only.
+	set(OPUS_X86_MAY_HAVE_SSE4_1 OFF CACHE BOOL "" FORCE)
+	set(OPUS_X86_MAY_HAVE_AVX2 OFF CACHE BOOL "" FORCE)
+endif()
+FetchContent_Declare(opus
+	GIT_REPOSITORY https://github.com/xiph/opus.git
+	GIT_TAG ddbe48383984d56acd9e1ab6a090c54ca6b735a6
+	GIT_SHALLOW FALSE)
+FetchContent_MakeAvailable(opus)
+set(BUILD_TESTING "${kyty_build_testing_saved}")
+install(FILES "${opus_SOURCE_DIR}/COPYING" DESTINATION "licenses/opus")
+
+# The Windows guest red-zone patcher needs a decoder/encoder and a small
+# trampoline assembler. Pin the same revisions used by shadPS4's implementation.
 set(XBYAK_BUILD_TESTS OFF CACHE BOOL "" FORCE)
 FetchContent_Declare(
 	xbyak
@@ -167,6 +188,7 @@ endif()
 
 set(TRACY_STATIC ON CACHE BOOL "" FORCE)
 set(TRACY_ENABLE ON CACHE BOOL "" FORCE)
+set(TRACY_NO_CRASH_HANDLER ON CACHE BOOL "" FORCE)
 set(TRACY_ON_DEMAND ON CACHE BOOL "" FORCE)
 set(TRACY_DELAYED_INIT ON CACHE BOOL "" FORCE)
 set(TRACY_MANUAL_LIFETIME ON CACHE BOOL "" FORCE)

CMakeLists.txt

@@ -332,7 +332,7 @@ if (KYTY_CLANG_CL)
 	set_source_files_properties(${kyty_emulator_src} PROPERTIES COMPILE_FLAGS "-Wno-pragma-pack -Wno-deprecated-declarations -D_TIMESPEC_DEFINED")
 endif()
 
-set(kyty_emulator_link_libraries common Vulkan::Headers spirv-tools-opt spirv-tools SDL3::SDL3-static imgui xxhash FFmpeg::ffmpeg fmt::fmt nlohmann_json::nlohmann_json LibAtrac9 kyty_git_version Zydis::Zydis xbyak::xbyak)
+set(kyty_emulator_link_libraries common Vulkan::Headers spirv-tools-opt spirv-tools SDL3::SDL3-static imgui xxhash FFmpeg::ffmpeg fmt::fmt nlohmann_json::nlohmann_json LibAtrac9 kyty_git_version Zydis::Zydis xbyak::xbyak opus)
 
 # Linux system libraries.
 if(LINUX)
@@ -481,6 +481,10 @@ add_executable(audio_out2_port_tests EXCLUDE_FROM_ALL
 target_link_libraries(audio_out2_port_tests common fmt::fmt)
 target_include_directories(audio_out2_port_tests PRIVATE ${inc_headers})
 
+add_executable(audio_out_timing_tests EXCLUDE_FROM_ALL "${KYTY_TESTS_DIR}/AudioOutTimingTests.cpp")
+target_link_libraries(audio_out_timing_tests common fmt::fmt SDL3::SDL3-static)
+target_include_directories(audio_out_timing_tests PRIVATE ${inc_headers})
+
 add_executable(ngs2_sampler_tests EXCLUDE_FROM_ALL
 	"${KYTY_TESTS_DIR}/Ngs2SamplerTests.cpp"
 	"${KYTY_SOURCE_DIR}/loader/timer.cpp"
@@ -491,7 +495,7 @@ target_include_directories(ngs2_sampler_tests PRIVATE ${inc_headers})
 add_executable(pad_haptics_tests EXCLUDE_FROM_ALL
 	"${KYTY_TESTS_DIR}/PadHapticsTests.cpp"
 )
-target_link_libraries(pad_haptics_tests common fmt::fmt SDL3::SDL3-static)
+target_link_libraries(pad_haptics_tests common fmt::fmt SDL3::SDL3-static opus)
 target_include_directories(pad_haptics_tests PRIVATE ${inc_headers})
 
 add_executable(save_data_memory_tests EXCLUDE_FROM_ALL
@@ -500,6 +504,10 @@ 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})
@@ -602,9 +610,11 @@ if(BUILD_TESTING)
 	add_test(NAME kernel_file_system COMMAND $<TARGET_FILE:kernel_file_system_tests>)
 	add_test(NAME sync_on_address COMMAND $<TARGET_FILE:sync_on_address_tests>)
 	add_test(NAME audio_out2_port COMMAND $<TARGET_FILE:audio_out2_port_tests>)
+	add_test(NAME audio_out_timing COMMAND $<TARGET_FILE:audio_out_timing_tests>)
 	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 $<TARGET_FILE:http_uri_parse_tests>)
@@ -677,9 +687,11 @@ if(BUILD_TESTING)
 		shader_recompiler_compute_tests
 		virtual_memory_allocation_tests
 		audio_out2_port_tests
+		audio_out_timing_tests
 		ngs2_sampler_tests
 		pad_haptics_tests
 		save_data_memory_tests
+		png_enc_tests
 		ces_tests
 		save_data_dialog_tests
 		http_uri_parse_tests
@@ -767,6 +779,10 @@ endif()
 install(TARGETS kyty_emulator DESTINATION .)
 if(KYTY_BUILD_LAUNCHER)
 	add_subdirectory("${KYTY_SOURCE_DIR}/launcher" "${CMAKE_CURRENT_BINARY_DIR}/launcher")
+
+	if(LINUX)
+		target_link_options(launcher PRIVATE $<$<CONFIG:Release>:-Wl,--strip-all>)
+	endif()
 endif()
 if(KYTY_CLANG_CL)
 	install(FILES "${KYTY_THIRD_PARTY_DIR}/winpthread/bin/libwinpthread-1.dll" DESTINATION .)

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/common/threads.cpp

@@ -2,7 +2,6 @@
 
 #include "common/assert.h"
 
-#include <algorithm>
 #include <atomic>
 #include <cerrno>
 #include <chrono>             // IWYU pragma: keep
@@ -26,8 +25,7 @@
 #ifdef KYTY_WIN_CS
 #include <windows.h> // IWYU pragma: keep
 // IWYU pragma: no_include <winbase.h>
-constexpr DWORD    KYTY_CS_SPIN_COUNT          = 4000;
-constexpr uint64_t KYTY_SLEEP_SPIN_LIMIT_100NS = 500; // 50 us
+constexpr DWORD KYTY_CS_SPIN_COUNT = 4000;
 
 #ifndef CREATE_WAITABLE_TIMER_HIGH_RESOLUTION
 #define CREATE_WAITABLE_TIMER_HIGH_RESOLUTION 0x00000002
@@ -38,30 +36,6 @@ static void SleepHighResolution100ns(uint64_t units_100ns) {
 		return;
 	}
 
-	// Keep spinning only where a kernel transition is
-	// likely to cost more than the requested delay; ordinary millisecond sleeps use the
-	// per-thread high-resolution waitable timer below.
-	if (units_100ns <= KYTY_SLEEP_SPIN_LIMIT_100NS) {
-		LARGE_INTEGER frequency {};
-		LARGE_INTEGER start {};
-		if (QueryPerformanceFrequency(&frequency) != 0 && QueryPerformanceCounter(&start) != 0 &&
-		    frequency.QuadPart > 0) {
-			const auto wait_ticks =
-			    static_cast<LONGLONG>((static_cast<long double>(units_100ns) *
-			                           static_cast<long double>(frequency.QuadPart)) /
-			                          10000000.0L);
-			const auto    deadline = start.QuadPart + std::max<LONGLONG>(wait_ticks, 1);
-			LARGE_INTEGER now {};
-			do {
-				if (QueryPerformanceCounter(&now) == 0) {
-					break;
-				}
-				YieldProcessor();
-			} while (now.QuadPart < deadline);
-			return;
-		}
-	}
-
 	thread_local HANDLE timer = CreateWaitableTimerExW(
 	    nullptr, nullptr, CREATE_WAITABLE_TIMER_HIGH_RESOLUTION, TIMER_ALL_ACCESS);
 	if (timer == nullptr) {
@@ -131,14 +105,13 @@ static SleepConditionVariableCS_func_t ResolveSleepConditionVariableCS() {
 #endif
 
 #ifdef KYTY_POSIX_HIGH_RES_SLEEP
-// Spin for very short waits; use an absolute deadline for longer waits.
+// An absolute deadline preserves the requested sleep across signal interruptions.
 static void SleepHighResolutionNanos(uint64_t nanos) {
 	if (nanos == 0) {
 		return;
 	}
 
 	constexpr uint64_t NANOS_PER_SEC = 1000000000;
-	constexpr uint64_t SPIN_LIMIT_NS = 50000; // below this a context switch dominates
 
 	timespec deadline {};
 	if (clock_gettime(CLOCK_MONOTONIC, &deadline) != 0) {
@@ -150,17 +123,6 @@ static void SleepHighResolutionNanos(uint64_t nanos) {
 	deadline.tv_sec += static_cast<time_t>(target_nsec / NANOS_PER_SEC);
 	deadline.tv_nsec = static_cast<long>(target_nsec % NANOS_PER_SEC);
 
-	if (nanos <= SPIN_LIMIT_NS) {
-		timespec now {};
-		do {
-			if (clock_gettime(CLOCK_MONOTONIC, &now) != 0) {
-				return;
-			}
-		} while (now.tv_sec < deadline.tv_sec ||
-		         (now.tv_sec == deadline.tv_sec && now.tv_nsec < deadline.tv_nsec));
-		return;
-	}
-
 	while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME, &deadline, nullptr) == EINTR) {
 	}
 }

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/command_processor/pm4Dispatch.cpp

@@ -140,6 +140,8 @@ constexpr auto MakeShaderDispatchTable() {
 	g_hw_sh_func[Pm4::COMPUTE_PGM_RSRC3]               = HwShSetCsRegisters;
 	g_hw_sh_func[Pm4::COMPUTE_PACE_ID]                 = HwShSetCsRegisters;
 	g_hw_sh_func[Pm4::SPI_GRAPHICS_SHADER_CONTROL_PS]  = HwShSetRegisters;
+	g_hw_sh_func[Pm4::SPI_SHADER_USER_DATA_ADDR_LO_PS] = HwShSetRegisters;
+	g_hw_sh_func[Pm4::SPI_SHADER_USER_DATA_ADDR_HI_PS] = HwShSetRegisters;
 	g_hw_sh_func[Pm4::SPI_GRAPHICS_SHADER_CONTROL_GS]  = HwShSetRegisters;
 	g_hw_sh_func[Pm4::SPI_SHADER_USER_DATA_ADDR_LO_GS] = HwShSetRegisters;
 	g_hw_sh_func[Pm4::SPI_SHADER_USER_DATA_ADDR_HI_GS] = HwShSetRegisters;

src/graphics/guest_gpu/command_processor/pm4Handlers.cpp

@@ -3622,6 +3622,18 @@ void GraphicsInitJmpTablesShIndirect() {
 	g_hw_sh_indirect_func[Pm4::SPI_GRAPHICS_SHADER_CONTROL_PS] = [](KYTY_HW_SH_INDIRECT_ARGS) {
 		HwShIgnoreShaderRegister(cmd_offset, value);
 	};
+	for (uint32_t offset = Pm4::SPI_SHADER_USER_DATA_ADDR_LO_PS;
+	     offset <= Pm4::SPI_SHADER_USER_DATA_ADDR_HI_PS; offset++) {
+		g_hw_sh_indirect_func[offset] = [](KYTY_HW_SH_INDIRECT_ARGS) {
+			cp.GetShCtx().SetPsUserDataAddress(cmd_offset - Pm4::SPI_SHADER_USER_DATA_ADDR_LO_PS,
+			                                   value);
+		};
+	}
+	for (uint32_t offset = 4; offset <= 5; offset++) {
+		g_hw_sh_indirect_func[offset] = [](KYTY_HW_SH_INDIRECT_ARGS) {
+			cp.GetShCtx().SetPsAuxiliaryTableAddress(cmd_offset - 4, value);
+		};
+	}
 	g_hw_sh_indirect_func[Pm4::SPI_SHADER_PACE_ID_GS] = [](KYTY_HW_SH_INDIRECT_ARGS) {
 		HwShIgnoreShaderRegister(cmd_offset, value);
 	};

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/guest_gpu/hardwareContext.h

@@ -458,6 +458,9 @@ struct PsShaderResource2 {
 
 struct PsStageRegisters {
 	uint64_t          data_addr = 0;
+	uint64_t          user_data_addr = 0;
+	// Hades II uses PS offsets 4:5 for a byte-addressed descriptor table.
+	uint64_t          auxiliary_table_addr = 0;
 	PsShaderResource1 rsrc1;
 	PsShaderResource2 rsrc2;
 };
@@ -1017,6 +1020,12 @@ public:
 	void SetGsShaderResource2(const GsShaderResource2& rsrc2) { m_vs.gs_regs.rsrc2 = rsrc2; }
 
 	void SetPsShaderBase(uint64_t addr) { m_ps.ps_regs.data_addr = addr; }
+	void SetPsUserDataAddress(uint32_t word, uint32_t value) {
+		SetUserDataAddressWord(m_ps.ps_regs.user_data_addr, word, value);
+	}
+	void SetPsAuxiliaryTableAddress(uint32_t word, uint32_t value) {
+		SetUserDataAddressWord(m_ps.ps_regs.auxiliary_table_addr, word, value);
+	}
 	void SetPsShaderResource1(const PsShaderResource1& rsrc1) { m_ps.ps_regs.rsrc1 = rsrc1; }
 	void SetPsShaderResource2(const PsShaderResource2& rsrc2) { m_ps.ps_regs.rsrc2 = rsrc2; }
 

src/graphics/guest_gpu/pm4.h

@@ -697,6 +697,8 @@ constexpr uint32_t CX_NUM = 0x3FF + 1;
 
 /* Shader registers */
 
+constexpr uint32_t SPI_SHADER_USER_DATA_ADDR_LO_PS = 0x2;
+constexpr uint32_t SPI_SHADER_USER_DATA_ADDR_HI_PS = 0x3;
 constexpr uint32_t SPI_SHADER_PACE_ID_PS            = 0x6;
 constexpr uint32_t SPI_GRAPHICS_SHADER_CONTROL_PS    = 0x7;
 constexpr uint32_t SPI_SHADER_PGM_LO_PS              = 0x8;

src/graphics/host_gpu/regionManager.h

@@ -30,21 +30,14 @@ class TrackingSpinLock final {
 public:
 	void lock() noexcept {
 		const auto thread = CurrentThread();
-		if (m_owner.load(std::memory_order_relaxed) == thread) {
-			EXIT("recursive region tracking lock\n");
-		}
 		while (m_lock.test_and_set(std::memory_order_acquire)) {
-			if (m_owner.load(std::memory_order_relaxed) == thread) {
-				EXIT("recursive region tracking lock while contended\n");
-			}
+			EXIT_NOT_IMPLEMENTED(m_owner.load(std::memory_order_relaxed) == thread);
 			std::atomic_signal_fence(std::memory_order_seq_cst);
 		}
 		m_owner.store(thread, std::memory_order_relaxed);
 	}
 	void unlock() noexcept {
-		if (m_owner.load(std::memory_order_relaxed) != CurrentThread()) {
-			EXIT("region tracking lock released by non-owner\n");
-		}
+		EXIT_NOT_IMPLEMENTED(m_owner.load(std::memory_order_relaxed) != CurrentThread());
 		m_owner.store(0, std::memory_order_relaxed);
 		m_lock.clear(std::memory_order_release);
 	}
@@ -88,8 +81,7 @@ public:
 	template <DirtySource source>
 	[[nodiscard]] bool IsModified(uint64_t offset, uint64_t size) const {
 		const auto [start, end] = GetPageRange(m_cpu_addr + offset, size);
-		const auto& bits        = GetBits<source>();
-		return RegionBits(bits, start, end).Any();
+		return GetBits<source>().FirstRangeFrom(start).first < end;
 	}
 
 	template <DirtySource source, bool enable>

src/graphics/host_gpu/renderer/cache/bufferCache.cpp

@@ -110,6 +110,7 @@ void BufferCache::DeleteBuffer(BufferId id) {
 	}
 }
 
+template <bool async>
 bool BufferCache::DownloadBufferMemory(Buffer& buffer, uint64_t vaddr, uint64_t size) {
 	std::vector<vk::BufferCopy> copies;
 	uint64_t                    total_size     = 0;
@@ -129,11 +130,16 @@ bool BufferCache::DownloadBufferMemory(Buffer& buffer, uint64_t vaddr, uint64_t
 		return false;
 	}
 
-	const auto [mapped, offset] = m_download_buffer.Map(total_size, 64);
+	auto [mapped, offset] = m_download_buffer.Map(total_size, 64);
+	std::unique_ptr<Buffer> temporary;
 	if (mapped == nullptr) {
-		EXIT("BufferCache: download exceeds 64 MiB staging buffer capacity\n");
+		temporary = std::make_unique<Buffer>(m_graphics, m_scheduler, MemoryUsage::Download, 0,
+		                                     vk::BufferUsageFlagBits::eTransferDst, total_size);
+		mapped = temporary->Mapped().data();
+	} else {
+		m_download_buffer.Commit();
 	}
-	m_download_buffer.Commit();
+	const auto& download = temporary ? *temporary : m_download_buffer;
 	for (auto& copy: copies) {
 		copy.dstOffset += offset;
 	}
@@ -152,27 +158,35 @@ bool BufferCache::DownloadBufferMemory(Buffer& buffer, uint64_t vaddr, uint64_t
 	native.pipelineBarrier(vk::PipelineStageFlagBits::eAllCommands,
 	                       vk::PipelineStageFlagBits::eTransfer, {}, 0, nullptr, 1, &before, 0,
 	                       nullptr);
-	native.copyBuffer(buffer.Handle(), m_download_buffer.Handle(),
+	native.copyBuffer(buffer.Handle(), download.Handle(),
 	                  static_cast<uint32_t>(copies.size()), copies.data());
 
 	auto after          = before;
 	after.srcAccessMask = vk::AccessFlagBits::eTransferWrite;
 	after.dstAccessMask = vk::AccessFlagBits::eHostRead;
-	after.buffer        = m_download_buffer.Handle();
+	after.buffer        = download.Handle();
 	after.offset        = offset;
 	after.size          = total_size;
 	native.pipelineBarrier(vk::PipelineStageFlagBits::eTransfer,
 	                       vk::PipelineStageFlagBits::eAllCommands |
 	                           vk::PipelineStageFlagBits::eHost,
 	                       {}, 0, nullptr, 1, &after, 0, nullptr);
-	m_scheduler.DeferPriorityOperation([this, mapped, offset, total_size, buffer_address,
-	                                    copies = std::move(copies)] {
-		m_download_buffer.Invalidate(offset, total_size);
+	auto publish = [this, mapped, offset, total_size, buffer_address,
+	                copies = std::move(copies), owner = std::move(temporary)] {
+		(owner ? *owner : m_download_buffer).Invalidate(offset, total_size);
 		for (const auto& copy: copies) {
 			Libs::LibKernel::Memory::WriteBacking(buffer_address + copy.srcOffset,
 			                                      mapped + (copy.dstOffset - offset), copy.size);
 		}
-	});
+	};
+	if constexpr (async) {
+		m_scheduler.DeferPriorityOperation(std::move(publish));
+	} else {
+		const auto tick = m_scheduler.CurrentTick();
+		m_scheduler.Wait(tick);
+		m_scheduler.WaitPriorityOperations(tick);
+		publish();
+	}
 	return true;
 }
 
@@ -251,10 +265,7 @@ void BufferCache::ReadMemory(uint64_t vaddr, uint64_t size, bool is_write) {
 		const auto window_begin = std::max(Common::AlignDown(vaddr, WindowSize), buffer_begin);
 		const auto window_end = std::min(std::max(window_begin + WindowSize, vaddr + size), buffer_end);
 
-		if (DownloadBufferMemory(buffer, window_begin, window_end - window_begin)) {
-			const auto tick = m_scheduler.CurrentTick();
-			m_scheduler.Wait(tick);
-			m_scheduler.WaitPriorityOperations(tick);
+		if (DownloadBufferMemory<false>(buffer, window_begin, window_end - window_begin)) {
 			m_memory_tracker.UnmarkRegionAsGpuModified(window_begin, window_end - window_begin);
 		}
 		if (is_write) {
@@ -491,8 +502,7 @@ std::pair<Buffer*, uint64_t> BufferCache::ObtainBufferForImage(uint64_t vaddr, u
 	}
 
 	auto [staging, stage_offset] = m_staging_buffer.Map(size, 16);
-	if (staging == nullptr || (!Libs::LibKernel::Memory::TryReadBacking(vaddr, staging, size) &&
-	                           !Libs::LibKernel::Memory::TryReadPrtBacking(vaddr, staging, size))) {
+	if (staging == nullptr || !Libs::LibKernel::Memory::TryReadSparseBacking(vaddr, staging, size)) {
 		EXIT("BufferCache: failed to read mapped guest image backing\n");
 	}
 	m_staging_buffer.Commit();
@@ -610,7 +620,7 @@ void BufferCache::RunGarbageCollector() {
 			return false;
 		}
 		if (dirty) {
-			EXIT_IF(!DownloadBufferMemory(buffer, buffer.CpuAddress(), buffer.Size()));
+			EXIT_NOT_IMPLEMENTED(!DownloadBufferMemory<true>(buffer, buffer.CpuAddress(), buffer.Size()));
 			dirty_buffers.push_back(id);
 		} else {
 			m_memory_tracker.UntrackMemory(buffer.CpuAddress(), buffer.Size());

src/graphics/host_gpu/renderer/cache/bufferCache.h

@@ -117,7 +117,8 @@ private:
 	[[nodiscard]] vk::Buffer UploadCopies(Buffer& buffer, std::span<vk::BufferCopy> copies,
 	                                      uint64_t total_size);
 	[[nodiscard]] bool SynchronizeBufferFromImage(Buffer& buffer, uint64_t vaddr, uint64_t size);
-	// Queues backing publication; callers wait before clearing dirty pages or reusing their data.
+	// Synchronous downloads publish before returning; asynchronous callers wait before reuse.
+	template <bool async>
 	[[nodiscard]] bool DownloadBufferMemory(Buffer& buffer, uint64_t vaddr, uint64_t size);
 
 	GraphicContext&                                   m_graphics;

src/graphics/host_gpu/renderer/cache/samplerCache.cpp

@@ -123,8 +123,11 @@ vk::Sampler SamplerCache::GetSampler(const ShaderSamplerResource& r, bool intege
 	sampler_info.addressModeU = to_vk_address_mode(r.ClampX());
 	sampler_info.addressModeV = to_vk_address_mode(r.ClampY());
 	sampler_info.addressModeW = to_vk_address_mode(r.ClampZ());
-	sampler_info.mipLodBias =
-	    static_cast<float>(static_cast<int16_t>((r.LodBias() ^ 0x2000u) - 0x2000u)) / 256.0f;
+    const auto max_lod_bias = m_graphics.GetPhysicalDeviceProperties().limits.maxSamplerLodBias;
+    sampler_info.mipLodBias = std::clamp(
+        static_cast<float>(static_cast<int16_t>((r.LodBias() ^ 0x2000u) - 0x2000u)) / 256.0f,
+        -max_lod_bias, max_lod_bias
+    );
 	sampler_info.anisotropyEnable        = (aniso ? VK_TRUE : VK_FALSE);
 	sampler_info.maxAnisotropy           = aniso_ratio;
 	sampler_info.compareEnable           = (r.DepthCompareFunc() != 0 ? VK_TRUE : VK_FALSE);

src/graphics/host_gpu/renderer/cache/textureCache.cpp

@@ -583,7 +583,7 @@ bool TextureCache::CopyD16(Image& destination, Image& source) {
 	}
 
 	auto                         depth_buffer = m_tiler.GetScratchBuffer(depth_size);
-	auto                         color_buffer = m_tiler.GetScratchBuffer(color_size);
+	auto                         color_buffer = m_tiler.GetScratchBuffer(color_size, depth_buffer.buffer);
 	const TileManager::D16Layout promote_layout {
 	    .width               = depth.info.extent.width,
 	    .height              = depth.info.extent.height,
@@ -1071,7 +1071,7 @@ void TextureCache::UploadImage(Image& image, Buffer& source, uint64_t source_off
 		                     texels_per_slice > UINT64_MAX / transfer_bytes);
 		const uint64_t transfer_slice = texels_per_slice * transfer_bytes;
 		EXIT_NOT_IMPLEMENTED(transfer_slice > UINT64_MAX / layers);
-		auto promoted = m_tiler.GetScratchBuffer(transfer_slice * layers);
+		auto promoted = m_tiler.GetScratchBuffer(transfer_slice * layers, linear.buffer);
 		m_tiler.ConvertD16(
 		    linear, promoted, TileManager::D16Direction::Promote,
 		    info.pixel_format == vk::Format::eD32SfloatS8Uint,
@@ -1533,7 +1533,7 @@ bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address
 	auto&          image = m_slot_images[selected];
 	vk::ClearValue clear {};
 	if (aspect == vk::ImageAspectFlagBits::eColor) {
-		if (!DecodePackedColorClear(image.info.pixel_format, packed_clear, clear.color)) {
+		if (!DecodeColorDwordFill(image.info.pixel_format, packed_clear, clear.color)) {
 			return false;
 		}
 	} else {
@@ -1679,7 +1679,7 @@ void TextureCache::DownloadDepth(Image& image, Buffer& destination, uint64_t des
 	auto host_linear = m_tiler.GetScratchBuffer(transfer_size);
 	image.Download(copies, host_linear.buffer, 0, host_linear.size);
 	const bool tiled        = info.IsTiled();
-	auto       guest_linear = tiled ? m_tiler.GetScratchBuffer(info.data.size)
+	auto       guest_linear = tiled ? m_tiler.GetScratchBuffer(info.data.size, host_linear.buffer)
 	                                : TileManager::Result {destination.Handle(), destination_offset,
 	                                                       destination.Size() - destination_offset};
 	m_tiler.ConvertD16(host_linear, guest_linear, TileManager::D16Direction::Demote,

src/graphics/host_gpu/renderer/image/imageInfo.h

@@ -428,6 +428,20 @@ inline constexpr std::array<VideoOutFormatPolicy, 7> VIDEO_OUT_FORMAT_POLICIES {
 	return true;
 }
 
+// Unlike a register clear, a DWORD fill repeats the same word across the entire pixel.
+// Keep this separate: the first register word alone cannot describe a 64-bit color.
+[[nodiscard]] inline bool DecodeColorDwordFill(vk::Format format, uint32_t packed,
+                                               vk::ClearColorValue& clear) {
+	if (format == vk::Format::eR16G16B16A16Sfloat) {
+		if (packed != 0) {
+			return false;
+		}
+		clear = vk::ClearColorValue {};
+		return true;
+	}
+	return DecodePackedColorClear(format, packed, clear);
+}
+
 [[nodiscard]] inline bool DecodePackedStencilClear(uint32_t packed, uint8_t& clear) {
 	const auto value = static_cast<uint8_t>(packed);
 	if (packed != static_cast<uint32_t>(value) * 0x01010101u) {

src/graphics/host_gpu/renderer/image/imageView.h

@@ -31,20 +31,14 @@ namespace ImageViewOps {
 	       ImageViewOps::IsFormatDepthCompatible(view_format);
 }
 
-[[nodiscard]] inline bool IsValidImageSwizzle(uint32_t swizzle) noexcept {
+[[nodiscard]] inline bool IsValidImageSwizzle(uint32_t swizzle,
+                                             uint32_t component_count = 4) noexcept {
 	if ((swizzle & ~0xfffu) != 0) {
 		return false;
 	}
 	for (uint32_t channel = 0; channel < 4; channel++) {
-		switch (GetDstSel(swizzle, channel)) {
-			case 0:
-			case 1:
-			case 4:
-			case 5:
-			case 6:
-			case 7: break;
-			default: return false;
-		}
+		const auto selector = GetDstSel(swizzle, channel);
+		if (selector > 1 && (selector < 4 || selector >= 4 + component_count)) return false;
 	}
 	return true;
 }
@@ -81,15 +75,8 @@ SelectSampledColorView(vk::Format image_format, vk::Format view_format, uint32_t
 [[nodiscard]] inline bool IsSupportedSampledDepthView(vk::Format image_format,
                                                       vk::Format view_format,
                                                       uint32_t   swizzle) noexcept {
-	if (!IsSupportedSampledDepthFormat(image_format, view_format)) {
-		return false;
-	}
-	switch (swizzle) {
-		case DstSel(4, 4, 4, 4):
-		case DstSel(4, 0, 0, 0):
-		case DstSel(4, 0, 0, 1): return true;
-		default: return false;
-	}
+	return IsSupportedSampledDepthFormat(image_format, view_format) &&
+	       IsValidImageSwizzle(swizzle, 1);
 }
 
 [[nodiscard]] inline bool

src/graphics/host_gpu/renderer/image/tiler.cpp

@@ -88,30 +88,6 @@ TileManager::~TileManager() {
 	}
 }
 
-TileManager::Scratch TileManager::AllocateScratch(uint64_t size) {
-	EXIT_IF(size == 0);
-	vk::BufferCreateInfo create {};
-	create.size  = size;
-	create.usage = vk::BufferUsageFlagBits::eStorageBuffer | vk::BufferUsageFlagBits::eTransferSrc |
-	               vk::BufferUsageFlagBits::eTransferDst;
-
-	VmaAllocationCreateInfo allocate {};
-	allocate.usage       = VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE;
-	VkBuffer      buffer = VK_NULL_HANDLE;
-	VmaAllocation memory = nullptr;
-	const auto    raw    = static_cast<VkBufferCreateInfo>(create);
-	RequireVulkanSuccess(static_cast<vk::Result>(vmaCreateBuffer(
-	                         m_graphics.allocator, &raw, &allocate, &buffer, &memory, nullptr)),
-	                     "allocate TileManager scratch buffer");
-	return {buffer, memory, size};
-}
-
-void TileManager::DeferDestroy(Scratch scratch) {
-	auto allocator = m_graphics.allocator;
-	m_scheduler.DeferOperation(
-	    [allocator, scratch] { vmaDestroyBuffer(allocator, scratch.buffer, scratch.allocation); });
-}
-
 void TileManager::Prepare(bool tile, uint64_t tiled_capacity, uint64_t linear_capacity,
                           std::span<const GpuTileInfo> infos, uint64_t source_base,
                           uint64_t target_base, std::vector<Dispatch>& dispatches) {
@@ -369,8 +345,7 @@ TileManager::Result TileManager::Detile(vk::Buffer tiled, uint64_t tiled_offset,
 	const uint64_t        source_base = tiled_offset & (descriptor_alignment - 1);
 	std::vector<Dispatch> dispatches;
 	Prepare(false, tiled_capacity, linear_capacity, infos, source_base, 0, dispatches);
-	auto scratch = AllocateScratch(Common::AlignUp(linear_capacity, 4));
-	DeferDestroy(scratch);
+	auto scratch = GetScratchBuffer(linear_capacity, tiled);
 	Record(tiled, tiled_offset, tiled_capacity, scratch.buffer, 0, scratch.size, dispatches,
 	       true);
 	return {scratch.buffer, 0, linear_capacity};
@@ -400,11 +375,9 @@ void TileManager::TileImage(Image& image, std::span<const vk::BufferImageCopy> r
 	    std::max<uint64_t>(limits.minStorageBufferOffsetAlignment, 4);
 	const uint64_t        target_base = tiled_offset & (descriptor_alignment - 1);
 	std::vector<Dispatch> dispatches;
-	// Reserve all stream parameters before creating a scheduler-lived scratch dependency:
-	// StreamBuffer::Map is allowed to submit the current tick when it wraps.
+	// Reserve stream parameters before recording the image download and conversion.
 	Prepare(true, tiled_capacity, linear_capacity, infos, 0, target_base, dispatches);
-	auto linear = AllocateScratch(Common::AlignUp(linear_capacity, 4));
-	DeferDestroy(linear);
+	auto linear = GetScratchBuffer(linear_capacity, tiled);
 	image.Download(regions, linear.buffer, 0, linear.size);
 	Result source {linear.buffer, 0, linear.size};
 	if (transform == ColorTransform::SwapBgra16) {
@@ -414,10 +387,19 @@ void TileManager::TileImage(Image& image, std::span<const vk::BufferImageCopy> r
 	       dispatches, false);
 }
 
-TileManager::Result TileManager::GetScratchBuffer(uint64_t size) {
-	auto scratch = AllocateScratch(Common::AlignUp(size, 4));
-	DeferDestroy(scratch);
-	return {scratch.buffer, 0, scratch.size};
+TileManager::Result TileManager::GetScratchBuffer(uint64_t size, vk::Buffer input) {
+	EXIT_IF(size == 0);
+	size = Common::AlignUp(size, 4);
+	auto& buffer = m_scratch[m_scratch[0] && m_scratch[0]->Handle() == input ? 1 : 0];
+	if (!buffer || buffer->Size() < size) {
+		if (buffer) {
+			m_scheduler.DeferOperation([old = std::move(buffer)]() mutable { old.reset(); });
+		}
+		buffer = std::make_unique<Buffer>(m_graphics, m_scheduler, MemoryUsage::DeviceLocal, 0,
+		    vk::BufferUsageFlagBits::eStorageBuffer | vk::BufferUsageFlagBits::eTransferSrc |
+		        vk::BufferUsageFlagBits::eTransferDst, size);
+	}
+	return {buffer->Handle(), 0, size};
 }
 
 TileManager::StorageBinding TileManager::BindStorage(Result buffer, uint64_t size) const {
@@ -668,9 +650,7 @@ void TileManager::SwapBgra16(Result input, Result output, uint32_t pixels) {
 
 TileManager::Result TileManager::SwapBgra16(Result input) {
 	EXIT_NOT_IMPLEMENTED(input.size == 0 || input.size % 8u != 0 || input.size / 8u > UINT32_MAX);
-	auto output = AllocateScratch(input.size);
-	DeferDestroy(output);
-	Result result {output.buffer, 0, output.size};
+	auto result = GetScratchBuffer(input.size, input.buffer);
 	SwapBgra16(input, result, static_cast<uint32_t>(input.size / 8u));
 	return result;
 }

src/graphics/host_gpu/renderer/image/tiler.h

@@ -6,13 +6,14 @@
 #include "graphics/host_gpu/vulkanCommon.h"
 
 #include <array>
+#include <memory>
 #include <span>
 #include <vector>
-#include <vk_mem_alloc.h>
 
 namespace Libs::Graphics {
 
 class CommandScheduler;
+class Buffer;
 class Image;
 class StreamBuffer;
 struct GraphicContext;
@@ -62,7 +63,7 @@ public:
 	~TileManager();
 	KYTY_CLASS_NO_COPY(TileManager);
 
-	// The returned device-local buffer remains alive through the current scheduler tick.
+	// Consume scratch results before the next acquisition, or pass their buffer as input.
 	[[nodiscard]] Result Detile(vk::Buffer tiled, uint64_t tiled_offset, uint64_t tiled_capacity,
 	                            uint64_t linear_capacity, std::span<const GpuTileInfo> infos);
 	void Tile(vk::Buffer linear, uint64_t linear_offset, uint64_t linear_capacity, vk::Buffer tiled,
@@ -71,7 +72,7 @@ public:
 	               uint64_t tiled_offset, uint64_t tiled_capacity, uint64_t linear_capacity,
 	               std::span<const GpuTileInfo> infos,
 	               ColorTransform               transform = ColorTransform::None);
-	[[nodiscard]] Result GetScratchBuffer(uint64_t size);
+	[[nodiscard]] Result GetScratchBuffer(uint64_t size, vk::Buffer input = nullptr);
 	void                 ConvertD16(Result source, Result target, D16Direction direction, bool d32,
 	                                const D16Layout& layout);
 	[[nodiscard]] Result SwapBgra16(Result input);
@@ -105,23 +106,16 @@ private:
 		uint32_t pipeline_slot = 0;
 		uint64_t params_offset = 0;
 	};
-	struct Scratch {
-		vk::Buffer    buffer     = nullptr;
-		VmaAllocation allocation = nullptr;
-		uint64_t      size       = 0;
-	};
 	struct StorageBinding {
 		vk::DescriptorBufferInfo info;
 		uint32_t                 base = 0;
 	};
 
-	[[nodiscard]] Scratch         AllocateScratch(uint64_t size);
 	[[nodiscard]] StorageBinding  BindStorage(Result buffer, uint64_t size) const;
 	[[nodiscard]] static uint32_t ConversionRows(uint64_t offset, uint64_t row_stride,
 	                                             uint64_t active, uint32_t remaining,
 	                                             uint64_t alignment, uint64_t max_range,
 	                                             uint32_t max_groups) noexcept;
-	void                          DeferDestroy(Scratch scratch);
 	void Prepare(bool tile, uint64_t tiled_capacity, uint64_t linear_capacity,
 	             std::span<const GpuTileInfo> infos, uint64_t source_base, uint64_t target_base,
 	             std::vector<Dispatch>& dispatches);
@@ -134,6 +128,7 @@ private:
 	GraphicContext&                         m_graphics;
 	CommandScheduler&                       m_scheduler;
 	StreamBuffer&                           m_stream_buffer;
+	std::array<std::unique_ptr<Buffer>, 2>   m_scratch;
 	vk::DescriptorSetLayout                 m_descriptor_layout = nullptr;
 	vk::PipelineLayout                      m_pipeline_layout   = nullptr;
 	std::array<vk::Pipeline, PipelineCount> m_pipelines {};

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

@@ -0,0 +1,58 @@
+#include "graphics/host_gpu/renderer/pipeline/blendMapping.h"
+
+#include "graphics/guest_gpu/hardwareContext.h"
+
+#include <initializer_list>
+
+namespace Libs::Graphics {
+
+bool BlendFactorIsDualSource(uint8_t factor) {
+	return factor >= static_cast<uint8_t>(Prospero::BlendFactor::kSrc1Color) &&
+	       factor <= static_cast<uint8_t>(Prospero::BlendFactor::kOneMinusSrc1Alpha);
+}
+
+namespace {
+
+bool BlendFactorIsConstantColor(uint8_t factor) {
+	return factor == static_cast<uint8_t>(Prospero::BlendFactor::kConstantColor) ||
+	       factor == static_cast<uint8_t>(Prospero::BlendFactor::kOneMinusConstantColor);
+}
+
+} // namespace
+
+BlendMappingSupport ClassifyBlendMapping(const HW::BlendControl&                blend,
+                                         const Prospero::ColorComponentMapping& mapping) {
+	// Color constants are not swizzled with the exports; scalar constant alpha is unaffected.
+	if (!mapping.IsIdentity() && (BlendFactorIsConstantColor(blend.color_srcblend) ||
+	                              BlendFactorIsConstantColor(blend.color_destblend))) {
+		return BlendMappingSupport::Unsupported;
+	}
+	if (mapping.Map(3) == 3u) {
+		return BlendMappingSupport::Direct;
+	}
+	// Moving alpha requires the same equation for all channels.
+	if (blend.separate_alpha_blend && (blend.alpha_srcblend != blend.color_srcblend ||
+	                                   blend.alpha_destblend != blend.color_destblend ||
+	                                   blend.alpha_comb_fcn != blend.color_comb_fcn)) {
+		return BlendMappingSupport::Unsupported;
+	}
+	auto support = BlendMappingSupport::Direct;
+	for (const auto factor: {blend.color_srcblend, blend.color_destblend}) {
+		if (BlendFactorIsDualSource(factor)) {
+			return BlendMappingSupport::Unsupported;
+		}
+		switch (static_cast<Prospero::BlendFactor>(factor)) {
+			case Prospero::BlendFactor::kSrcAlpha:
+			case Prospero::BlendFactor::kOneMinusSrcAlpha:
+				support = BlendMappingSupport::SourceAlpha;
+				break;
+			case Prospero::BlendFactor::kDstAlpha:
+			case Prospero::BlendFactor::kOneMinusDstAlpha:
+			case Prospero::BlendFactor::kSrcAlphaSaturate: return BlendMappingSupport::Unsupported;
+			default: break;
+		}
+	}
+	return support;
+}
+
+} // namespace Libs::Graphics

src/graphics/host_gpu/renderer/pipeline/blendMapping.h

@@ -0,0 +1,27 @@
+#ifndef EMULATOR_SRC_GRAPHICS_HOST_GPU_RENDERER_PIPELINE_BLENDMAPPING_H_
+#define EMULATOR_SRC_GRAPHICS_HOST_GPU_RENDERER_PIPELINE_BLENDMAPPING_H_
+
+#include <cstdint>
+
+namespace Libs::Graphics {
+
+namespace HW {
+struct BlendControl;
+}
+namespace Prospero {
+struct ColorComponentMapping;
+}
+
+enum class BlendMappingSupport {
+	Direct,
+	SourceAlpha, // Requires logical alpha in the second blend source.
+	Unsupported,
+};
+
+bool                BlendFactorIsDualSource(uint8_t factor);
+BlendMappingSupport ClassifyBlendMapping(const HW::BlendControl&                blend,
+                                         const Prospero::ColorComponentMapping& mapping);
+
+} // namespace Libs::Graphics
+
+#endif // EMULATOR_SRC_GRAPHICS_HOST_GPU_RENDERER_PIPELINE_BLENDMAPPING_H_

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

@@ -974,6 +974,41 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 		        (pipeline_bind_point == vk::PipelineBindPoint::eCompute &&
 		         shader_stage != vk::ShaderStageFlagBits::eCompute));
 	}
+	for (const auto* reader: prepared_bindings) {
+		const auto& reads = reader->runtime->resources->specialization_reads;
+		if (reads.empty()) continue;
+		for (const auto* writer: prepared_bindings) {
+			if (writer->runtime->program->has_address_writes) {
+				EXIT("scalar resource reads cannot be proven disjoint from shader address writes\n");
+			}
+		}
+		for (const auto [address, size]: reads) {
+			for (const auto id: m_bound_images) {
+				const auto* image = m_context.GetTextureCache().m_slot_images.try_get(id);
+				if (image == nullptr ||
+				    (!image->binding.shader_write && !image->binding.is_target)) continue;
+				for (const auto written: {image->info.data, image->info.stencil,
+				                          image->info.metadata.range}) {
+					if (written.size != 0 && ImageRangeOverlaps(address, size,
+					                                          written.address, written.size)) {
+						EXIT("scalar resource reads overlap an image or attachment write\n");
+					}
+				}
+			}
+			for (const auto* writer: prepared_bindings) {
+				const auto& program = *writer->runtime->program;
+				for (uint32_t i = 0; i < writer->buffer_sources.size(); ++i) {
+					const auto resource = program.bindings.descriptors.front().resources[i];
+					if (!program.info.buffers[resource].written) continue;
+					const auto& written = writer->buffer_sources[i];
+					if (written.size != 0 && ImageRangeOverlaps(address, size,
+					                                          written.address, written.size)) {
+						EXIT("scalar resource reads overlap a shader buffer write\n");
+					}
+				}
+			}
+		}
+	}
 	m_descriptor_buffers.clear();
 	m_descriptor_images.clear();
 	m_descriptor_writes.clear();

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

@@ -11,6 +11,7 @@
 #include "graphics/host_gpu/renderer/debug.h"
 #include "graphics/host_gpu/renderer/depthRenderTarget.h"
 #include "graphics/host_gpu/renderer/image/imageView.h"
+#include "graphics/host_gpu/renderer/pipeline/blendMapping.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
 #include "graphics/shader/recompiler/ShaderRecompiler.h"
@@ -219,7 +220,6 @@ struct PipelineCache::ProgramCache {
 		ShaderRecompiler::IR::ResourceSnapshot       resources;
 		ShaderRecompiler::IR::ResourceSpecialization specialization;
 		std::vector<Permutation>                    permutations;
-		bool                                        skip_dispatch = false;
 	};
 
 	struct ProgramKeyHash {
@@ -287,9 +287,6 @@ struct PipelineCache::ProgramCache {
 		lookup_key.code_size       = static_cast<uint32_t>(params.code.size());
 		BuildStageStaticKey(input_info, lookup_key.static_state);
 		auto                                         entry = programs.find(lookup_key);
-		if (entry != programs.end() && entry->second.skip_dispatch) {
-			return {};
-		}
 		const ShaderRecompiler::IR::SrtRuntime       runtime {
 		    .user_data                  = user_data,
 		    .shader_base                = params.Base(),
@@ -357,11 +354,6 @@ struct PipelineCache::ProgramCache {
 		}
 		DumpShaderOriginal(stage_name, options.shader_hash, params.code);
 		auto translated = ShaderRecompiler::TranslateProgram(params.code, options);
-		if (translated.skip_dispatch) {
-			entry = programs.try_emplace(lookup_key, ShaderRecompiler::IR::ResourcePlan {}).first;
-			entry->second.skip_dispatch = true;
-			return {};
-		}
 		if (entry == programs.end()) {
 			entry = programs.try_emplace(lookup_key,
 			    ShaderRecompiler::IR::ExtractResourcePlan(translated.program)).first;
@@ -609,21 +601,30 @@ PipelineCache::GraphicsPrograms PipelineCache::GetGraphicsPrograms(
 	}
 	ShaderParams pixel_params;
 	if (pixel_active) {
-		pixel_params = PrepareProgram(pixel_regs, sh, target_export_mapping, pixel_info);
-		const auto& blend          = context.GetBlendControl(0);
-		const auto  is_dual_source = [](uint8_t factor) {
-			return factor >= static_cast<uint8_t>(Prospero::BlendFactor::kSrc1Color) &&
-			       factor <= static_cast<uint8_t>(Prospero::BlendFactor::kOneMinusSrc1Alpha);
-		};
+		pixel_params      = PrepareProgram(pixel_regs, sh, target_export_mapping, pixel_info);
+		const auto& blend = context.GetBlendControl(0);
 		pixel_info.dual_source_blending =
 		    blend.enable && !context.GetRenderTarget(0).info.blend_bypass &&
-		    (is_dual_source(blend.color_srcblend) || is_dual_source(blend.color_destblend) ||
-		     (blend.separate_alpha_blend &&
-		      (is_dual_source(blend.alpha_srcblend) || is_dual_source(blend.alpha_destblend))));
+		    (BlendFactorIsDualSource(blend.color_srcblend) ||
+		     BlendFactorIsDualSource(blend.color_destblend) ||
+		     (blend.separate_alpha_blend && (BlendFactorIsDualSource(blend.alpha_srcblend) ||
+		                                     BlendFactorIsDualSource(blend.alpha_destblend))));
 		if (pixel_info.dual_source_blending) {
-			// MRT1 supplies a second blend source for the same render target as MRT0.
+			// MRT1 supplies the second blend source for target 0.
 			pixel_info.target_output_mode[1]    = pixel_info.target_output_mode[0];
 			pixel_info.target_export_mapping[1] = pixel_info.target_export_mapping[0];
+		} else if (blend.enable && !context.GetRenderTarget(0).info.blend_bypass &&
+		           pixel_info.target_output_mode[0] != 0 && pixel_info.target_output_mode[0] != 7 &&
+		           std::all_of(std::begin(pixel_info.target_output_mode) + 1,
+		                       std::end(pixel_info.target_output_mode),
+		                       [](uint8_t mode) { return mode == 0; }) &&
+		           ClassifyBlendMapping(blend, pixel_info.target_export_mapping[0]) ==
+		               BlendMappingSupport::SourceAlpha) {
+			// Preserve logical alpha when the export mapping moves it.
+			pixel_info.alpha_blend_source_remap = true;
+			pixel_info.dual_source_blending     = true;
+			pixel_info.target_output_mode[1]    = pixel_info.target_output_mode[0];
+			pixel_info.target_export_mapping[1] = {};
 		}
 	}
 	if (context.GetClipControl().clip_disable) {
@@ -713,14 +714,35 @@ PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(
 		}
 		const auto& rt                        = ctx.GetRenderTarget(colors[i].target_slot);
 		const auto& bc                        = ctx.GetBlendControl(colors[i].target_slot);
-		static_params.color_srcblend[slot]       = bc.color_srcblend;
-		static_params.color_comb_fcn[slot]       = bc.color_comb_fcn;
-		static_params.color_destblend[slot]      = bc.color_destblend;
-		static_params.alpha_srcblend[slot]       = bc.alpha_srcblend;
-		static_params.alpha_comb_fcn[slot]       = bc.alpha_comb_fcn;
-		static_params.alpha_destblend[slot]      = bc.alpha_destblend;
-		static_params.separate_alpha_blend[slot] = bc.separate_alpha_blend;
-		static_params.blend_enable[slot]         = bc.enable && !rt.info.blend_bypass;
+		const bool alpha_remap =
+		    slot == 0 && ps_input_info != nullptr && ps_input_info->alpha_blend_source_remap;
+		static_params.blend_enable[slot] = bc.enable && !rt.info.blend_bypass;
+		if (static_params.blend_enable[slot] && !alpha_remap &&
+		    ClassifyBlendMapping(bc, colors[i].export_mapping) != BlendMappingSupport::Direct) {
+			static_params.blend_enable[slot] = false;
+			static std::atomic_bool warned = false;
+			if (!warned.exchange(true, std::memory_order_relaxed)) {
+				Log::WriteToConsoleAndLog(fmt::format(
+				    "Warning: blending disabled for unsupported color mapping "
+				    "(slot={} mapping=0x{:02x} color={}/{} alpha={}/{} separate={}).\n",
+				    slot, colors[i].export_mapping.packed, bc.color_srcblend, bc.color_destblend,
+				    bc.alpha_srcblend, bc.alpha_destblend, bc.separate_alpha_blend ? 1 : 0));
+			}
+		}
+		if (alpha_remap) {
+			static_params.blend_alpha_source_remap = true;
+		}
+		if (static_params.blend_enable[slot]) {
+			static_params.color_srcblend[slot]       = bc.color_srcblend;
+			static_params.color_comb_fcn[slot]       = bc.color_comb_fcn;
+			static_params.color_destblend[slot]      = bc.color_destblend;
+			static_params.separate_alpha_blend[slot] = bc.separate_alpha_blend;
+			if (bc.separate_alpha_blend) {
+				static_params.alpha_srcblend[slot]  = bc.alpha_srcblend;
+				static_params.alpha_comb_fcn[slot]  = bc.alpha_comb_fcn;
+				static_params.alpha_destblend[slot] = bc.alpha_destblend;
+			}
+		}
 	}
 	const bool with_depth =
 	    depth.desc.view_info.format != vk::Format::eUndefined && static_cast<bool>(depth.image_id);

src/graphics/host_gpu/renderer/pipeline/pipelineCache.h

@@ -55,6 +55,7 @@ struct PipelineStaticParameters {
 	uint8_t                    alpha_destblend[RENDER_COLOR_ATTACHMENTS_MAX]      = {};
 	bool                       separate_alpha_blend[RENDER_COLOR_ATTACHMENTS_MAX] = {};
 	bool                       blend_enable[RENDER_COLOR_ATTACHMENTS_MAX]         = {};
+	bool                       blend_alpha_source_remap                           = false;
 
 	bool operator==(const PipelineStaticParameters& other) const noexcept;
 };
@@ -64,7 +65,7 @@ struct PipelineStaticParameters {
 static_assert(std::is_trivially_copyable_v<PipelineStaticParameters>);
 static_assert(std::is_standard_layout_v<PipelineStaticParameters>);
 static_assert(alignof(PipelineStaticParameters) == 1);
-static_assert(sizeof(PipelineStaticParameters) == 125);
+static_assert(sizeof(PipelineStaticParameters) == 126);
 
 struct PipelineRenderingState {
 	std::array<vk::Format, RENDER_COLOR_ATTACHMENTS_MAX> color_formats {};

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

@@ -137,14 +137,17 @@ static void GetInputFormat(const ShaderBufferResource& res, vk::Format& format,
 	}
 }
 
-static vk::BlendFactor GetBlendFactor(uint32_t factor) {
+static vk::BlendFactor GetBlendFactor(uint32_t factor, bool remap_source_alpha) {
 	switch (static_cast<Prospero::BlendFactor>(factor)) {
 		case Prospero::BlendFactor::kZero: return vk::BlendFactor::eZero;
 		case Prospero::BlendFactor::kOne: return vk::BlendFactor::eOne;
 		case Prospero::BlendFactor::kSrcColor: return vk::BlendFactor::eSrcColor;
 		case Prospero::BlendFactor::kOneMinusSrcColor: return vk::BlendFactor::eOneMinusSrcColor;
-		case Prospero::BlendFactor::kSrcAlpha: return vk::BlendFactor::eSrcAlpha;
-		case Prospero::BlendFactor::kOneMinusSrcAlpha: return vk::BlendFactor::eOneMinusSrcAlpha;
+		case Prospero::BlendFactor::kSrcAlpha:
+			return remap_source_alpha ? vk::BlendFactor::eSrc1Color : vk::BlendFactor::eSrcAlpha;
+		case Prospero::BlendFactor::kOneMinusSrcAlpha:
+			return remap_source_alpha ? vk::BlendFactor::eOneMinusSrc1Color
+			                          : vk::BlendFactor::eOneMinusSrcAlpha;
 		case Prospero::BlendFactor::kDstAlpha: return vk::BlendFactor::eDstAlpha;
 		case Prospero::BlendFactor::kOneMinusDstAlpha: return vk::BlendFactor::eOneMinusDstAlpha;
 		case Prospero::BlendFactor::kDstColor: return vk::BlendFactor::eDstColor;
@@ -327,9 +330,6 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 			}
 		}
 
-		EXIT_NOT_IMPLEMENTED(vs_input_info.resources[index].AddTid());
-		EXIT_NOT_IMPLEMENTED(vs_input_info.resources[index].SwizzleEnabled());
-
 		EXIT_IF(registers_num < 1 || registers_num > 4);
 	}
 
@@ -397,17 +397,20 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 		color_blend_attachment[i].colorWriteMask =
 		    vk::ColorComponentFlags {static_params.color_mask[i]};
 		color_blend_attachment[i].blendEnable = static_params.blend_enable[i] ? VK_TRUE : VK_FALSE;
+		// Only target 0 can use the second blend source carrying logical alpha.
+		const bool remap_source_alpha         = i == 0 && static_params.blend_alpha_source_remap;
 		color_blend_attachment[i].srcColorBlendFactor =
-		    GetBlendFactor(static_params.color_srcblend[i]);
+		    GetBlendFactor(static_params.color_srcblend[i], remap_source_alpha);
 		color_blend_attachment[i].dstColorBlendFactor =
-		    GetBlendFactor(static_params.color_destblend[i]);
+		    GetBlendFactor(static_params.color_destblend[i], remap_source_alpha);
 		color_blend_attachment[i].colorBlendOp = GetBlendOp(static_params.color_comb_fcn[i]);
 		color_blend_attachment[i].srcAlphaBlendFactor =
-		    (static_params.separate_alpha_blend[i] ? GetBlendFactor(static_params.alpha_srcblend[i])
-		                                           : color_blend_attachment[i].srcColorBlendFactor);
+		    (static_params.separate_alpha_blend[i]
+		         ? GetBlendFactor(static_params.alpha_srcblend[i], remap_source_alpha)
+		         : color_blend_attachment[i].srcColorBlendFactor);
 		color_blend_attachment[i].dstAlphaBlendFactor =
 		    (static_params.separate_alpha_blend[i]
-		         ? GetBlendFactor(static_params.alpha_destblend[i])
+		         ? GetBlendFactor(static_params.alpha_destblend[i], remap_source_alpha)
 		         : color_blend_attachment[i].dstColorBlendFactor);
 		color_blend_attachment[i].alphaBlendOp =
 		    (static_params.separate_alpha_blend[i] ? GetBlendOp(static_params.alpha_comb_fcn[i])

src/graphics/host_gpu/renderer/renderCompute.cpp

@@ -256,10 +256,6 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 	input_info.dispatch_thread_dimensions = use_thread_dimensions;
 	const auto compute_program =
 	    m_context.GetPipelineCache().GetComputeProgram(cs_regs, sh_regs, input_info);
-	if (!compute_program) {
-		// Temporary until RT is implemented.
-		return;
-	}
 	if (use_thread_dimensions) {
 		input_info.dispatch_threads_num[0]    = thread_group_x;
 		input_info.dispatch_threads_num[1]    = thread_group_y;
@@ -268,15 +264,14 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 
 	const auto& program   = *input_info.stage.program;
 	const auto& resources = *input_info.stage.resources;
-	if (TryConsumeComputeMetaClear(input_info, buffer)) {
-		ResetBindings();
-		return;
-	}
-	if (TryConsumeComputeImageClear(input_info, buffer, thread_group_x, thread_group_y,
-	                                thread_group_z, mode)) {
+	if (resources.specialization_reads.empty() &&
+	    (TryConsumeComputeMetaClear(input_info, buffer) ||
+	     TryConsumeComputeImageClear(input_info, buffer, thread_group_x, thread_group_y,
+	                                 thread_group_z, mode))) {
 		ResetBindings();
 		return;
 	}
+
 	const bool large_workgroup =
 	    (input_info.threads_num[0] * input_info.threads_num[1] * input_info.threads_num[2] >= 512);
 	const bool                   has_sampler = !program.info.samplers.empty();
@@ -417,10 +412,6 @@ void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
 	ShaderComputeInputInfo input_info {};
 	const auto compute_program = m_context.GetPipelineCache().GetComputeProgram(
 	    cs_regs, buffer.GetRegisters().GetShaderRegisters(), input_info);
-	if (!compute_program) {
-		// Temporary until RT is implemented.
-		return;
-	}
 	buffer.EndRendering();
 	auto& pipeline = m_context.GetPipelineCache().GetComputePipeline(input_info, compute_program);
 	auto& bindings = m_compute_bindings;

src/graphics/host_gpu/renderer/renderContext.cpp

@@ -119,6 +119,10 @@ void RenderContext::UnmapMemory(uint64_t vaddr, uint64_t size) {
 }
 
 void RenderContext::PrepareBda() {
+	if (!m_bda_logged) {
+		Log::WriteToConsoleAndLog("GPU: using buffer device address (BDA) shader memory access.\n");
+		m_bda_logged = true;
+	}
 	std::shared_lock lock(m_mapped_ranges_mutex);
 	m_mapped_ranges.ForEach([this](uint64_t start, uint64_t end) {
 		m_buffer_cache.SynchronizeBuffersInRange(start, end - start);

src/graphics/host_gpu/renderer/renderContext.h

@@ -81,6 +81,7 @@ private:
 	std::unique_ptr<GuestGpu> m_gpu;
 	VideoOut::VideoOutDriver* m_video_out = nullptr;
 	bool                      m_fault_process_pending = false;
+	bool                      m_bda_logged = false;
 
 	Common::Mutex                        m_interrupt_mutex;
 	std::vector<InterruptEqRegistration> m_interrupt_eqs;

src/graphics/host_gpu/renderer/renderDraw.cpp

@@ -1032,6 +1032,7 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 	uint32_t   mesh_groups = 0;
 	if (mesh_active) {
 		const auto& mesh = state.vertex_info[0].mesh;
+		EXIT_NOT_IMPLEMENTED(mesh.fast_launch && (draw.IsIndexed() || primitive_restart_enable));
 		static std::atomic_bool restart_warned = false;
 		if (primitive_restart_enable && !restart_warned.exchange(true, std::memory_order_relaxed)) {
 			std::printf("Warning: primitive restart is not implemented for mesh shaders; "

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/presentation/window/vulkanWindow.cpp

@@ -137,11 +137,13 @@ static uint32_t VulkanFindQueueFamily(vk::PhysicalDevice device, vk::SurfaceKHR
 	return static_cast<uint32_t>(-1);
 }
 
+// On failure out_device is null and out_rejections says why each device was skipped.
 // NOLINTNEXTLINE(readability-function-cognitive-complexity)
 static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surface,
                                      const std::vector<const char*>& device_extensions,
                                      SurfaceCapabilities&            out_capabilities,
-                                     vk::PhysicalDevice& out_device, uint32_t& out_queue_family) {
+                                     vk::PhysicalDevice& out_device, uint32_t& out_queue_family,
+                                     std::string& out_rejections) {
 	EXIT_IF(instance == nullptr);
 	EXIT_IF(surface == nullptr);
 
@@ -170,14 +172,22 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 		vk::PhysicalDeviceProperties device_properties {};
 		device.getProperties(&device_properties);
 
+		const auto reject = [&](const std::string& reason) {
+			LOGF("%s\n", reason.c_str());
+			out_rejections +=
+			    skip_device ? "; " : fmt::format("\n  {}: ", device_properties.deviceName.data());
+			out_rejections += reason;
+			skip_device = true;
+		};
+
 		LOGF("Vulkan device: %s\n", device_properties.deviceName.data());
 		if (device_properties.apiVersion < VULKAN_TARGET_API_VERSION) {
-			LOGF("Vulkan %u.%u is required, but device supports only %u.%u.%u\n",
-			     VK_VERSION_MAJOR(VULKAN_TARGET_API_VERSION),
-			     VK_VERSION_MINOR(VULKAN_TARGET_API_VERSION),
-			     VK_VERSION_MAJOR(device_properties.apiVersion),
-			     VK_VERSION_MINOR(device_properties.apiVersion),
-			     VK_VERSION_PATCH(device_properties.apiVersion));
+			reject(fmt::format("Vulkan {}.{} is required, but device supports only {}.{}.{}",
+			                   VK_VERSION_MAJOR(VULKAN_TARGET_API_VERSION),
+			                   VK_VERSION_MINOR(VULKAN_TARGET_API_VERSION),
+			                   VK_VERSION_MAJOR(device_properties.apiVersion),
+			                   VK_VERSION_MINOR(device_properties.apiVersion),
+			                   VK_VERSION_PATCH(device_properties.apiVersion)));
 			continue;
 		}
 
@@ -203,7 +213,7 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 		fragment_barycentric.pNext = &depth_clip_control;
 		features12.pNext           = &fragment_barycentric;
 #endif
-		features13.pNext = &features12;
+		features13.pNext       = &features12;
 		device_features2.pNext = &features13;
 
 		device.getFeatures2(&device_features2);
@@ -212,151 +222,82 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 
 		const auto queue_family = VulkanFindQueueFamily(device, surface);
 		if (queue_family == static_cast<uint32_t>(-1)) {
-			LOGF("No universal graphics, compute, and presentation queue\n");
-			skip_device = true;
+			reject("No universal graphics, compute, and presentation queue");
 		}
 
+		const auto check_feature = [&](vk::Bool32 supported, const char* name,
+		                               vk::Bool32 required = VK_TRUE) {
+			if (required == VK_TRUE && supported != VK_TRUE) {
+				reject(fmt::format("{} is not supported", name));
+			}
+		};
+
+#if defined(__APPLE__)
 		if (color_write_ext.colorWriteEnable != VK_TRUE) {
 			LOGF("colorWriteEnable is not supported\n");
-#if !defined(__APPLE__)
-			skip_device = true;
+		}
+#else
+		check_feature(color_write_ext.colorWriteEnable, "colorWriteEnable");
 #endif
-		}
-		if (image_view_min_lod.minLod != VK_TRUE) {
-			LOGF("image view minLod is not supported\n");
-			skip_device = true;
-		}
+		check_feature(image_view_min_lod.minLod, "image view minLod");
 
-		if (depth_clip_control.depthClipControl != VK_TRUE) {
-			LOGF("depthClipControl is not supported\n");
-			skip_device = true;
-		}
+		check_feature(depth_clip_control.depthClipControl, "depthClipControl");
+#if defined(__APPLE__)
 		if (depth_clip_enable.depthClipEnable != VK_TRUE) {
 			LOGF("depthClipEnable is not supported\n");
-#if !defined(__APPLE__)
-			skip_device = true;
+		}
+#else
+		check_feature(depth_clip_enable.depthClipEnable, "depthClipEnable");
 #endif
-		}
 #if !defined(__APPLE__)
-		if (device_features2.features.depthClamp != VK_TRUE) {
-			LOGF("depthClamp is not supported\n");
-			skip_device = true;
-		}
-		if (fragment_barycentric.fragmentShaderBarycentric != VK_TRUE) {
-			LOGF("fragmentShaderBarycentric is not supported\n");
-			skip_device = true;
-		}
+		check_feature(device_features2.features.depthClamp, "depthClamp");
+		check_feature(fragment_barycentric.fragmentShaderBarycentric, "fragmentShaderBarycentric");
 #endif
 
-		if (required_features12.samplerMirrorClampToEdge == VK_TRUE &&
-		    features12.samplerMirrorClampToEdge != VK_TRUE) {
-			LOGF("samplerMirrorClampToEdge is not supported\n");
-			skip_device = true;
-		}
-		if (required_features12.timelineSemaphore == VK_TRUE &&
-		    features12.timelineSemaphore != VK_TRUE) {
-			LOGF("timelineSemaphore is not supported\n");
-			skip_device = true;
-		}
-		if (required_features12.shaderOutputLayer == VK_TRUE &&
-		    features12.shaderOutputLayer != VK_TRUE) {
-			LOGF("shaderOutputLayer is not supported\n");
-			skip_device = true;
-		}
-		if (required_features12.shaderOutputViewportIndex == VK_TRUE &&
-		    features12.shaderOutputViewportIndex != VK_TRUE) {
-			LOGF("shaderOutputViewportIndex is not supported\n");
-			skip_device = true;
-		}
-		if (required_features12.bufferDeviceAddress == VK_TRUE &&
-		    features12.bufferDeviceAddress != VK_TRUE) {
-			LOGF("bufferDeviceAddress is not supported\n");
-			skip_device = true;
-		}
-		if (required_features12.shaderBufferInt64Atomics == VK_TRUE &&
-		    features12.shaderBufferInt64Atomics != VK_TRUE) {
-			LOGF("shaderBufferInt64Atomics is not supported\n");
-			skip_device = true;
-		}
-		if (features13.robustImageAccess != VK_TRUE) {
-			LOGF("robustImageAccess is not supported\n");
-			skip_device = true;
-		}
-		if (required_features13.dynamicRendering == VK_TRUE &&
-		    features13.dynamicRendering != VK_TRUE) {
-			LOGF("dynamicRendering is not supported\n");
-			skip_device = true;
-		}
-		if (required_features13.synchronization2 == VK_TRUE &&
-		    features13.synchronization2 != VK_TRUE) {
-			LOGF("synchronization2 is not supported\n");
-			skip_device = true;
-		}
-		if (device_features2.features.sampleRateShading != VK_TRUE) {
-			LOGF("sampleRateShading is not supported\n");
-			skip_device = true;
-		}
-		if (device_features2.features.depthBiasClamp != VK_TRUE) {
-			LOGF("depthBiasClamp is not supported\n");
-			skip_device = true;
-		}
-		if (device_features2.features.shaderClipDistance != VK_TRUE) {
-			LOGF("shaderClipDistance is not supported\n");
-			skip_device = true;
-		}
-		if (device_features2.features.shaderCullDistance != VK_TRUE) {
-			LOGF("shaderCullDistance is not supported\n");
-			skip_device = true;
-		}
-		if (device_features2.features.largePoints != VK_TRUE) {
-			LOGF("largePoints is not supported\n");
-			skip_device = true;
-		}
-		if (device_features2.features.multiViewport != VK_TRUE) {
-			LOGF("multiViewport is not supported\n");
-			skip_device = true;
-		}
-		if (device_features2.features.fillModeNonSolid != VK_TRUE) {
-			LOGF("fillModeNonSolid is not supported\n");
-			skip_device = true;
-		}
-		if (device_features2.features.fragmentStoresAndAtomics != VK_TRUE) {
-			LOGF("fragmentStoresAndAtomics is not supported\n");
-			skip_device = true;
-		}
+		check_feature(features12.samplerMirrorClampToEdge, "samplerMirrorClampToEdge",
+		              required_features12.samplerMirrorClampToEdge);
+		check_feature(features12.timelineSemaphore, "timelineSemaphore",
+		              required_features12.timelineSemaphore);
+		check_feature(features12.shaderOutputLayer, "shaderOutputLayer",
+		              required_features12.shaderOutputLayer);
+		check_feature(features12.shaderOutputViewportIndex, "shaderOutputViewportIndex",
+		              required_features12.shaderOutputViewportIndex);
+		check_feature(features12.bufferDeviceAddress, "bufferDeviceAddress",
+		              required_features12.bufferDeviceAddress);
+		check_feature(features12.shaderBufferInt64Atomics, "shaderBufferInt64Atomics",
+		              required_features12.shaderBufferInt64Atomics);
+		check_feature(features13.robustImageAccess, "robustImageAccess");
+		check_feature(features13.dynamicRendering, "dynamicRendering",
+		              required_features13.dynamicRendering);
+		check_feature(features13.synchronization2, "synchronization2",
+		              required_features13.synchronization2);
+		check_feature(device_features2.features.sampleRateShading, "sampleRateShading");
+		check_feature(device_features2.features.depthBiasClamp, "depthBiasClamp");
+		check_feature(device_features2.features.shaderClipDistance, "shaderClipDistance");
+		check_feature(device_features2.features.shaderCullDistance, "shaderCullDistance");
+		check_feature(device_features2.features.largePoints, "largePoints");
+		check_feature(device_features2.features.multiViewport, "multiViewport");
+		check_feature(device_features2.features.fillModeNonSolid, "fillModeNonSolid");
+		check_feature(device_features2.features.fragmentStoresAndAtomics,
+		              "fragmentStoresAndAtomics");
 
-		if (device_features2.features.samplerAnisotropy != VK_TRUE) {
-			LOGF("samplerAnisotropy is not supported\n");
-			skip_device = true;
-		}
-		if (device_features2.features.robustBufferAccess != VK_TRUE) {
-			LOGF("robustBufferAccess is not supported\n");
-			skip_device = true;
-		}
+		check_feature(device_features2.features.samplerAnisotropy, "samplerAnisotropy");
+		check_feature(device_features2.features.robustBufferAccess, "robustBufferAccess");
+#if defined(__APPLE__)
 		if (device_features2.features.depthBounds != VK_TRUE) {
 			LOGF("depthBounds is not supported\n");
-#if !defined(__APPLE__)
-			skip_device = true;
+		}
+#else
+		check_feature(device_features2.features.depthBounds, "depthBounds");
 #endif
-		}
-		if (device_features2.features.shaderStorageImageWriteWithoutFormat != VK_TRUE) {
-			LOGF("shaderStorageImageWriteWithoutFormat is not supported\n");
-			skip_device = true;
-		}
+		check_feature(device_features2.features.shaderStorageImageWriteWithoutFormat,
+		              "shaderStorageImageWriteWithoutFormat");
 
-		if (device_features2.features.shaderImageGatherExtended != VK_TRUE) {
-			LOGF("shaderImageGatherExtended is not supported\n");
-			skip_device = true;
-		}
+		check_feature(device_features2.features.shaderImageGatherExtended,
+		              "shaderImageGatherExtended");
 
-		if (device_features2.features.independentBlend != VK_TRUE) {
-			LOGF("independentBlend is not supported\n");
-			skip_device = true;
-		}
-		if (device_features2.features.tessellationShader != VK_TRUE) {
-			LOGF("tessellationShader is not supported\n");
-			skip_device = true;
-		}
+		check_feature(device_features2.features.independentBlend, "independentBlend");
+		check_feature(device_features2.features.tessellationShader, "tessellationShader");
 
 		if (!skip_device) {
 			auto available_extensions = EnumerateVulkan<vk::ExtensionProperties>(
@@ -368,8 +309,7 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 
 			for (const char* ext: device_extensions) {
 				if (!HasExtension(available_extensions, ext)) {
-					skip_device = true;
-					break;
+					reject(fmt::format("{} is not supported", ext));
 				}
 			}
 
@@ -387,86 +327,46 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 
 			if (!(candidate_capabilities.capabilities.supportedUsageFlags &
 			      vk::ImageUsageFlagBits::eTransferDst)) {
-				LOGF("Surface cannot be destination of blit\n");
-				skip_device = true;
+				reject("Surface cannot be destination of blit");
 			}
 		}
 
-		if (!skip_device && !CheckFormat(device, vk::Format::eD32Sfloat, true,
-		                                 vk::FormatFeatureFlagBits::eDepthStencilAttachment)) {
-			LOGF("Format vk::Format::eD32Sfloat cannot be used as depth buffer\n");
-			skip_device = true;
-		}
-
-		if (!skip_device && !CheckFormat(device, vk::Format::eD32SfloatS8Uint, true,
-		                                 vk::FormatFeatureFlagBits::eDepthStencilAttachment)) {
-			LOGF("Format vk::Format::eD32SfloatS8Uint cannot be used as depth buffer\n");
-			skip_device = true;
-		}
-
-		if (!skip_device && !CheckFormat(device, vk::Format::eD16Unorm, true,
-		                                 vk::FormatFeatureFlagBits::eDepthStencilAttachment)) {
-			LOGF("Format vk::Format::eD16Unorm cannot be used as depth buffer\n");
-			skip_device = true;
-		}
-
-		if (!skip_device && !CheckFormat(device, vk::Format::eBc3SrgbBlock, true,
-		                                 vk::FormatFeatureFlagBits::eSampledImage |
-		                                     vk::FormatFeatureFlagBits::eTransferDst)) {
-			LOGF("Format vk::Format::eBc3SrgbBlock cannot be used as texture\n");
-			skip_device = true;
-		}
-
-		if (!skip_device && !CheckFormat(device, vk::Format::eR8G8B8A8Srgb, true,
-		                                 vk::FormatFeatureFlagBits::eSampledImage |
-		                                     vk::FormatFeatureFlagBits::eTransferDst)) {
-			LOGF("Format vk::Format::eR8G8B8A8Srgb cannot be used as texture\n");
-			skip_device = true;
-		}
-
-		if (!skip_device && !CheckFormat(device, vk::Format::eR8Unorm, true,
-		                                 vk::FormatFeatureFlagBits::eSampledImage |
-		                                     vk::FormatFeatureFlagBits::eTransferDst)) {
-			LOGF("Format vk::Format::eR8Unorm cannot be used as texture\n");
-			skip_device = true;
-		}
-
-		if (!skip_device && !CheckFormat(device, vk::Format::eR8G8Unorm, true,
-		                                 vk::FormatFeatureFlagBits::eSampledImage |
-		                                     vk::FormatFeatureFlagBits::eTransferDst)) {
-			LOGF("Format vk::Format::eR8G8Unorm cannot be used as texture\n");
-			skip_device = true;
-		}
-
-		if (!skip_device && !CheckFormat(device, vk::Format::eR8G8B8A8Srgb, true,
-		                                 vk::FormatFeatureFlagBits::eStorageImage |
-		                                     vk::FormatFeatureFlagBits::eTransferDst)) {
-			LOGF("Format vk::Format::eR8G8B8A8Srgb cannot be used as texture\n");
-
-			if (!skip_device && !CheckFormat(device, vk::Format::eR8G8B8A8Unorm, true,
-			                                 vk::FormatFeatureFlagBits::eStorageImage |
-			                                     vk::FormatFeatureFlagBits::eTransferDst)) {
-				LOGF("Format vk::Format::eR8G8B8A8Unorm cannot be used as texture\n");
-				skip_device = true;
+		const auto check_format = [&](vk::Format format, vk::FormatFeatureFlags features,
+		                              const char* usage) {
+			if (!skip_device && !CheckFormat(device, format, true, features)) {
+				reject(fmt::format("Format vk::Format::e{} cannot be used as {}",
+				                   vk::to_string(format), usage));
 			}
+		};
+
+		const auto depth_features = vk::FormatFeatureFlagBits::eDepthStencilAttachment;
+		const auto texture_features =
+		    vk::FormatFeatureFlagBits::eSampledImage | vk::FormatFeatureFlagBits::eTransferDst;
+		const auto storage_features =
+		    vk::FormatFeatureFlagBits::eStorageImage | vk::FormatFeatureFlagBits::eTransferDst;
+
+		check_format(vk::Format::eD32Sfloat, depth_features, "depth buffer");
+		check_format(vk::Format::eD32SfloatS8Uint, depth_features, "depth buffer");
+		check_format(vk::Format::eD16Unorm, depth_features, "depth buffer");
+		check_format(vk::Format::eBc3SrgbBlock, texture_features, "texture");
+		check_format(vk::Format::eR8G8B8A8Srgb, texture_features, "texture");
+		check_format(vk::Format::eR8Unorm, texture_features, "texture");
+		check_format(vk::Format::eR8G8Unorm, texture_features, "texture");
+
+		if (!skip_device &&
+		    !CheckFormat(device, vk::Format::eR8G8B8A8Srgb, true, storage_features)) {
+			LOGF("Format vk::Format::eR8G8B8A8Srgb cannot be used as storage image\n");
+			check_format(vk::Format::eR8G8B8A8Unorm, storage_features, "storage image");
 		}
 
-		if (!skip_device && !CheckFormat(device, vk::Format::eB8G8R8A8Srgb, true,
-		                                 vk::FormatFeatureFlagBits::eStorageImage |
-		                                     vk::FormatFeatureFlagBits::eTransferDst)) {
-			LOGF("Format vk::Format::eB8G8R8A8Srgb cannot be used as texture\n");
-
-			if (!skip_device && !CheckFormat(device, vk::Format::eB8G8R8A8Unorm, true,
-			                                 vk::FormatFeatureFlagBits::eStorageImage |
-			                                     vk::FormatFeatureFlagBits::eTransferDst)) {
-				LOGF("Format vk::Format::eB8G8R8A8Unorm cannot be used as texture\n");
-				skip_device = true;
-			}
+		if (!skip_device &&
+		    !CheckFormat(device, vk::Format::eB8G8R8A8Srgb, true, storage_features)) {
+			LOGF("Format vk::Format::eB8G8R8A8Srgb cannot be used as storage image\n");
+			check_format(vk::Format::eB8G8R8A8Unorm, storage_features, "storage image");
 		}
 
 		if (!skip_device && device_properties.limits.maxSamplerAnisotropy < 16.0f) {
-			LOGF("maxSamplerAnisotropy < 16.0f");
-			skip_device = true;
+			reject("maxSamplerAnisotropy < 16.0f");
 		}
 
 		if (skip_device) {
@@ -520,16 +420,20 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 #endif
 	depth_clip_control.depthClipControl = VK_TRUE;
 
-	auto features12  = WindowContext::RequiredVulkan12Features();
-	features12.pNext = &depth_clip_control;
-
+	const bool workgroup_layout_extension =
+	    HasExtension(device_extensions, VK_KHR_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_EXTENSION_NAME);
+	vk::PhysicalDeviceWorkgroupMemoryExplicitLayoutFeaturesKHR supported_workgroup_layout {};
+	vk::PhysicalDeviceVulkan12Features supported_features12 {};
+	supported_features12.pNext = workgroup_layout_extension ? &supported_workgroup_layout : nullptr;
 	vk::PhysicalDeviceVulkan13Features supported_features13 {};
+	supported_features13.pNext = &supported_features12;
 
 	const auto robustness2_ext_enabled =
 	    HasExtension(device_extensions, VK_EXT_ROBUSTNESS_2_EXTENSION_NAME);
 
 	vk::PhysicalDeviceRobustness2FeaturesEXT supported_robustness2 {};
 	if (robustness2_ext_enabled) {
+		supported_robustness2.pNext = supported_features13.pNext;
 		supported_features13.pNext = &supported_robustness2;
 	}
 
@@ -557,6 +461,22 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 		supported_features2.pNext = &provoking_vertex;
 	}
 	physical_device.getFeatures2(&supported_features2);
+
+	auto features12 = WindowContext::RequiredVulkan12Features();
+	features12.shaderSharedInt64Atomics = supported_features12.shaderSharedInt64Atomics;
+	vk::PhysicalDeviceWorkgroupMemoryExplicitLayoutFeaturesKHR workgroup_layout {};
+	workgroup_layout.workgroupMemoryExplicitLayout =
+	    supported_workgroup_layout.workgroupMemoryExplicitLayout;
+	workgroup_layout.pNext = &depth_clip_control;
+	features12.pNext = workgroup_layout_extension ? static_cast<void*>(&workgroup_layout)
+	                                             : static_cast<void*>(&depth_clip_control);
+	if (!features12.shaderSharedInt64Atomics || !workgroup_layout.workgroupMemoryExplicitLayout) {
+		Log::WriteToConsoleAndLog(fmt::format(
+		    "WARNING: Native 64-bit LDS atomics are unavailable: shaderSharedInt64Atomics={}, "
+		    "workgroupMemoryExplicitLayout={}. Continuing with the selected Vulkan device.\n",
+		    features12.shaderSharedInt64Atomics != VK_FALSE,
+		    workgroup_layout.workgroupMemoryExplicitLayout != VK_FALSE));
+	}
 	graphics.mesh_shader_enabled = mesh_extension && supported_mesh.meshShader;
 
 	vk::PhysicalDeviceSubgroupSizeControlProperties subgroup_size_control {};
@@ -1004,11 +924,13 @@ void WindowContext::CreateVulkan() {
 	}
 #endif
 
+	std::string rejected_devices;
 	VulkanFindPhysicalDevice(graphic_ctx.instance, surface, device_extensions, surface_capabilities,
-	                         graphic_ctx.physical_device, graphic_ctx.queue_family);
+	                         graphic_ctx.physical_device, graphic_ctx.queue_family,
+	                         rejected_devices);
 
 	if (graphic_ctx.physical_device == nullptr) {
-		EXIT("Could not find suitable device");
+		EXIT("Could not find suitable device:%s", rejected_devices.c_str());
 	}
 
 	vk::PhysicalDevicePushDescriptorProperties push_descriptor_properties {};
@@ -1049,6 +971,7 @@ void WindowContext::CreateVulkan() {
 		for (const auto* extension: {VK_EXT_ROBUSTNESS_2_EXTENSION_NAME,
 		                             VK_EXT_PROVOKING_VERTEX_EXTENSION_NAME,
 		                             VK_EXT_MESH_SHADER_EXTENSION_NAME,
+		                             VK_KHR_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_EXTENSION_NAME,
 		                             VK_EXT_DEPTH_RANGE_UNRESTRICTED_EXTENSION_NAME}) {
 			if (HasExtension(available_extensions, extension)) {
 				device_extensions.push_back(extension);

src/graphics/shader/recompiler/ShaderRecompiler.cpp

@@ -19,7 +19,6 @@
 
 #include <algorithm>
 #include <array>
-#include <atomic>
 #include <chrono>
 #include <fmt/format.h>
 #include <map>
@@ -523,20 +522,6 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 	     GetDumpLabel(options), StageName(options.stage), options.shader_hash,
 	     static_cast<uint64_t>(decoded.instructions.size()), phase_ms());
 
-	// Temporary workaround for games that compile ray-tracing shaders before
-	// the player can select a mode without ray tracing.
-	if (options.stage == ShaderType::Compute && decoded.has_bvh) {
-		static std::atomic_flag warned = ATOMIC_FLAG_INIT;
-		if (!warned.test_and_set(std::memory_order_relaxed)) {
-			const auto& bvh = decoded.instructions.back();
-			Log::WriteToConsoleAndLog(fmt::format(
-			    "Warning: ray tracing is not implemented; skipping compute dispatches containing "
-			    "BVH intersection instructions (shader=0x{:016x}, pc=0x{:08x}, opcode=0x{:02x}).\n",
-			    options.shader_hash, bvh.pc, bvh.opcode_id));
-		}
-		return {.skip_dispatch = true};
-	}
-
 	std::string decoded_dump;
 	if (options.dump_ir) {
 		decoded_dump = Decoder::ProgramToString(decoded);
@@ -609,7 +594,7 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 	     GetDumpLabel(options), StageName(options.stage), options.shader_hash,
 	     static_cast<uint64_t>(ir.blocks.size()), phase_ms());
 	IR::RewriteToSsa(ir.blocks);
-	IR::ConstantPropagationPass(ir.blocks);
+	IR::ConstantPropagationPass(ir.blocks, ir.wave_size);
 	IR::ResolveControlFlowIdentities(ir);
 	IR::RemoveIdentities(ir.blocks);
 	IR::EliminateDeadCode(ir.blocks);
@@ -617,7 +602,7 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 	if (read_lane_stats.rewritten_reads != 0) {
 		LOGF("%s read-lane elimination: reads=%" PRIu32 "\n", GetDumpLabel(options),
 		     read_lane_stats.rewritten_reads);
-		IR::ConstantPropagationPass(ir.blocks);
+		IR::ConstantPropagationPass(ir.blocks, ir.wave_size);
 		IR::ResolveControlFlowIdentities(ir);
 		IR::RemoveIdentities(ir.blocks);
 		IR::EliminateDeadCode(ir.blocks);
@@ -645,7 +630,6 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 CompileResult CompileProgram(TranslateResult translated, const CompileOptions& options,
                              const IR::ResourceSpecialization& specialization,
                              uint32_t push_data_start_dword) {
-	EXIT_IF(translated.skip_dispatch);
 	const auto emit_begin = std::chrono::steady_clock::now();
 	auto& ir = translated.program;
 	IR::ApplyResourceSpecialization(ir, specialization);

src/graphics/shader/recompiler/ShaderRecompiler.h

@@ -28,7 +28,6 @@ struct TranslateResult {
 	IR::Program program;
 	std::string decoded_dump;
 	std::string cfg_dump;
-	bool        skip_dispatch = false;
 };
 
 struct CompileResult {

src/graphics/shader/recompiler/Tessellation.cpp

@@ -217,7 +217,7 @@ void LowerTessellationMemory(IR::Program& program, const CompileOptions& options
 			    2, ActiveAddress(*block, it, it->Arg(2), it->Arg(it->NumArgs() - 1u), resolved));
 		}
 	}
-	ConstantPropagationPass(program.blocks);
+	ConstantPropagationPass(program.blocks, program.wave_size);
 	uint32_t reads = 0, writes = 0, factors = 0;
 	for (auto* block: program.blocks) {
 		for (auto it = block->begin(); it != block->end(); ++it) {
@@ -317,7 +317,7 @@ void LowerTessellationMemory(IR::Program& program, const CompileOptions& options
 			inst.Invalidate();
 		}
 	}
-	ConstantPropagationPass(program.blocks);
+	ConstantPropagationPass(program.blocks, program.wave_size);
 	RemoveIdentities(program.blocks);
 	EliminateDeadCode(program.blocks);
 	LOGF("%s tessellation lowering: reads=%u writes=%u factors=%u\n",

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

@@ -122,6 +122,16 @@ void ValidateNativeProgram(const IR::Program& program) {
 			              program.memory_info[index].kind == IR::ResourceKind::IndirectBuffer;
 		       });
 	};
+	const auto local_flat_handle = [&](const IR::Inst& handle) {
+		return !handle.Uses().empty() &&
+		       std::ranges::all_of(handle.Uses(), [&](const IR::Use& use) {
+			       if (IR::AddressOpcodeInfoOf(use.user->GetOpcode()).access == IR::AddressAccess::None)
+				       return false;
+			       const auto index = use.user->Flags<IR::MemoryFlags>().index;
+			       return index < program.memory_info.size() &&
+			              program.memory_info[index].kind == IR::ResourceKind::FlatLocal;
+		       });
+	};
 	for (const auto* block: program.blocks) {
 		for (const auto& inst: *block) {
 			const auto dense = inst.Flags<uint32_t>();
@@ -138,7 +148,7 @@ void ValidateNativeProgram(const IR::Program& program) {
 					if (planning_only_handle(inst)) {
 						break;
 					}
-					if (inst.NumArgs() != 2 || !program.info.uses_dma) {
+					if (inst.NumArgs() != 2 || (!program.info.uses_dma && !local_flat_handle(inst))) {
 						Fail(program, "typed address handle has invalid DMA metadata");
 					}
 					break;
@@ -188,11 +198,16 @@ Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program
 				if (memory_index >= program.memory_info.size()) {
 					Fail(program, "address operation has invalid memory metadata");
 				}
-				if (program.memory_info[memory_index].kind == IR::ResourceKind::Scratch) {
+				const auto kind = program.memory_info[memory_index].kind;
+				if (kind == IR::ResourceKind::Scratch || kind == IR::ResourceKind::FlatLocal) {
 					if (program.scratch_dwords == 0) {
 						Fail(program, "scratch operation has no per-thread storage");
 					}
 					requirements.function_scratch = true;
+					if (kind == IR::ResourceKind::FlatLocal && program.stage != ShaderType::Compute &&
+					    program.stage != ShaderType::Mesh) {
+						requirements.function_lds = true;
+					}
 				} else if (address_access == IR::AddressAccess::Write) {
 					Fail(program, "writable FLAT/GLOBAL addresses require GPU ownership tracking");
 				}
@@ -229,6 +244,13 @@ Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program
 				if (kind != IR::ResourceKind::Lds && kind != IR::ResourceKind::Gds) {
 					Fail(program, "shared operation has invalid resource kind");
 				}
+				if (shared_access == IR::SharedAccess::Atomic &&
+				    IR::SharedComponentCount(inst.GetOpcode()) == 2u) {
+					if (kind != IR::ResourceKind::Lds || program.stage != ShaderType::Compute) {
+						Fail(program, "64-bit shared atomics require compute LDS");
+					}
+					requirements.shared_int64_atomics = true;
+				}
 				if (program.stage != ShaderType::Compute && program.stage != ShaderType::Mesh &&
 				    kind == IR::ResourceKind::Lds) {
 					requirements.function_lds = true;
@@ -241,6 +263,7 @@ Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program
 				}
 			}
 			switch (inst.GetOpcode()) {
+				case IR::ValueOpcode::BvhIntersect: requirements.bvh = true; break;
 				case IR::ValueOpcode::Ballot: requirements.subgroup_ballot = true; break;
 				case IR::ValueOpcode::DppMoveU32:
 				case IR::ValueOpcode::ReadFirstLane:

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

@@ -444,6 +444,18 @@ uint32_t EmitUGreaterThan64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
 	return CompareOrdered64(state, arg0, arg1, spv::OpUGreaterThan, spv::OpUGreaterThan);
 }
 
+uint32_t EmitSLessThanEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
+	return CompareOrdered64(state, arg0, arg1, spv::OpSLessThan, spv::OpULessThanEqual);
+}
+
+uint32_t EmitULessThanEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
+	return CompareOrdered64(state, arg0, arg1, spv::OpULessThan, spv::OpULessThanEqual);
+}
+
+uint32_t EmitUGreaterThanEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
+	return CompareOrdered64(state, arg0, arg1, spv::OpUGreaterThan, spv::OpUGreaterThanEqual);
+}
+
 uint32_t EmitFPIsNan32(EmitterState& state, uint32_t arg0) {
 	return EmitNative<spv::OpFUnordNotEqual, IR::Type::U1>(state, arg0, arg0);
 }

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

@@ -0,0 +1,348 @@
+#include "graphics/shader/recompiler/backend/spirv/spirvEmitterInstructions.h"
+
+namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter {
+namespace {
+
+using Vec3 = std::array<uint32_t, 3>;
+
+uint32_t Extract(EmitterState& s, uint32_t type, uint32_t vector, uint32_t index) {
+	const auto result = s.builder.AllocateId();
+	s.builder.AddFunction(spv::OpCompositeExtract, type, result, vector, index);
+	return result;
+}
+
+uint32_t Vector(EmitterState& s, uint32_t type, std::span<const uint32_t> components) {
+	const auto result = s.builder.AllocateId();
+	s.builder.AddFunction(spv::OpCompositeConstruct, type, result, components);
+	return result;
+}
+
+uint32_t NodeWord(EmitterState& s, uint32_t block, uint32_t word) {
+	const auto offset = Binary(s, spv::OpShiftLeftLogical, TypeU32(s), word, ConstantU32(s, 2));
+	const auto address = Binary(s, spv::OpIAdd, TypeScalarU64(s), block,
+	                            Unary(s, spv::OpUConvert, TypeScalarU64(s), offset));
+	const auto pointer = Unary(s, spv::OpConvertUToPtr, TypePhysicalU32Pointer(s), address);
+	const auto result = s.builder.AllocateId();
+	s.builder.AddFunction(spv::OpLoad, TypeU32(s), result, pointer, spv::MemoryAccessAlignedMask, 4u);
+	return result;
+}
+
+uint32_t Triangle(EmitterState& s, uint32_t block, uint32_t kind, uint32_t bary,
+                  Vec3 origin, Vec3 direction) {
+	const auto f = TypeF32(s), u = TypeU32(s), b = TypeBool(s);
+	const auto zero = ConstantF32Value(s, 0.0f);
+	const auto op = [&](spv::Op code, uint32_t x, uint32_t y) {
+		const auto value = Binary(s, code, f, x, y);
+		s.builder.AddAnnotation(spv::OpDecorate, value, spv::DecorationNoContraction);
+		return value;
+	};
+	const auto cmp = [&](spv::Op code, uint32_t x, uint32_t y) { return Binary(s, code, b, x, y); };
+	const auto either = [&](uint32_t x, uint32_t y) { return Binary(s, spv::OpLogicalOr, b, x, y); };
+	const auto both = [&](uint32_t x, uint32_t y) { return Binary(s, spv::OpLogicalAnd, b, x, y); };
+	const auto abs = [&](uint32_t x) { return EmitGlsl<GLSLstd450FAbs, IR::Type::F32>(s, x); };
+	const auto is_kind = [&](uint32_t n) { return cmp(spv::OpIEqual, kind, ConstantU32(s, n)); };
+	// The four triangles are (0,1,2), (1,3,2), (2,3,4), (2,4,0).
+	const std::array vertices {
+		EmitGlsl<GLSLstd450UMin, IR::Type::U32>(s, kind, ConstantU32(s, 2)),
+		Select(s, u, is_kind(0), ConstantU32(s, 1),
+		       Select(s, u, is_kind(3), ConstantU32(s, 4), ConstantU32(s, 3))),
+		Select(s, u, is_kind(2), ConstantU32(s, 4),
+		       Select(s, u, is_kind(3), ConstantU32(s, 0), ConstantU32(s, 2))) };
+	std::array<Vec3, 3> positions;
+	for (uint32_t i = 0; i < 3; ++i) {
+		const auto first = Binary(s, spv::OpIMul, u, vertices[i], ConstantU32(s, 3));
+		for (uint32_t axis = 0; axis < 3; ++axis)
+			positions[i][axis] = Unary(s, spv::OpBitcast, f,
+			    NodeWord(s, block, Binary(s, spv::OpIAdd, u, first, ConstantU32(s, axis))));
+	}
+	const auto flag = NodeWord(s, block, ConstantU32(s, 15));
+	const auto y_largest = cmp(spv::OpFOrdLessThan, abs(direction[0]), abs(direction[1]));
+	const auto z_largest = cmp(spv::OpFOrdLessThan,
+	    Select(s, f, y_largest, abs(direction[1]), abs(direction[0])), abs(direction[2]));
+	const auto rotate = [&](Vec3 value) {
+		Vec3 result;
+		for (uint32_t i = 0; i < 3; ++i)
+			result[i] = Select(s, f, z_largest, value[i],
+			    Select(s, f, y_largest, value[(i + 2) % 3], value[(i + 1) % 3]));
+		return result;
+	};
+	origin = rotate(origin);
+	direction = rotate(direction);
+	std::array<Vec3, 3> projected;
+	for (uint32_t i = 0; i < 3; ++i) {
+		const auto position = rotate(positions[i]);
+		Vec3 relative;
+		for (uint32_t axis = 0; axis < 3; ++axis)
+			relative[axis] = op(spv::OpFSub, position[axis], origin[axis]);
+		projected[i] = {op(spv::OpFSub, op(spv::OpFMul, relative[0], direction[2]),
+		                              op(spv::OpFMul, direction[0], relative[2])),
+		                op(spv::OpFSub, op(spv::OpFMul, relative[1], direction[2]),
+		                              op(spv::OpFMul, direction[1], relative[2])), relative[2]};
+	}
+	Vec3 edge, weighted;
+	for (uint32_t i = 0; i < 3; ++i) {
+		const auto& first = projected[(i + 1) % 3];
+		const auto& second = projected[(i + 2) % 3];
+		edge[i] = op(spv::OpFSub, op(spv::OpFMul, second[0], first[1]),
+		                        op(spv::OpFMul, second[1], first[0]));
+		weighted[i] = op(spv::OpFMul, edge[i], direction[2]);
+	}
+	auto numerator = op(spv::OpFAdd, op(spv::OpFAdd,
+	    op(spv::OpFMul, edge[0], projected[0][2]), op(spv::OpFMul, edge[1], projected[1][2])),
+	    op(spv::OpFMul, edge[2], projected[2][2]));
+	auto denominator = op(spv::OpFAdd, op(spv::OpFAdd, weighted[0], weighted[1]), weighted[2]);
+	uint32_t negative = ConstantBool(s, false), positive = negative;
+	for (auto value: edge) {
+		negative = either(negative, cmp(spv::OpFOrdLessThan, value, zero));
+		positive = either(positive, cmp(spv::OpFOrdGreaterThan, value, zero));
+	}
+	const auto winding = cmp(spv::OpFOrdGreaterThan, denominator, zero);
+	auto missed = either(both(negative, positive), cmp(spv::OpFOrdEqual, denominator, zero));
+	missed = either(missed, Unary(s, spv::OpIsNan, b, numerator));
+	missed = either(missed, cmp(spv::OpFOrdLessThan,
+	    Select(s, f, winding, numerator, Unary(s, spv::OpFNegate, f, numerator)), zero));
+	for (uint32_t i = 0; i < 3; ++i) {
+		const auto first = projected[(i + 1) % 3][1], second = projected[(i + 2) % 3][1];
+		const auto first_zero = cmp(spv::OpFOrdEqual, first, zero);
+		const auto horizontal = both(first_zero, cmp(spv::OpFOrdEqual, second, zero));
+		const auto below = either(cmp(spv::OpFOrdLessThan, first, zero),
+		    both(first_zero, cmp(spv::OpFOrdGreaterThan, second, zero)));
+		const auto right = Binary(s, spv::OpLogicalNotEqual, b, below, winding);
+		const auto excluded = Select(s, b, horizontal,
+		    cmp(spv::OpFOrdGreaterThan, projected[i][1], zero), right);
+		missed = either(missed, both(cmp(spv::OpFOrdEqual, edge[i], zero), excluded));
+	}
+	const auto remap = Binary(s, spv::OpShiftRightLogical, u, flag,
+	    Binary(s, spv::OpIMul, u, kind, ConstantU32(s, 8)));
+	const auto coordinate = [&](uint32_t shift) {
+		const auto index = Binary(s, spv::OpBitwiseAnd, u,
+		    Binary(s, spv::OpShiftRightLogical, u, remap, ConstantU32(s, shift)), ConstantU32(s, 3));
+		return Unary(s, spv::OpBitcast, u, Select(s, f,
+		    cmp(spv::OpIEqual, index, ConstantU32(s, 1)), weighted[1],
+		    Select(s, f, cmp(spv::OpIEqual, index, ConstantU32(s, 2)), weighted[2], weighted[0])));
+	};
+	const std::array result {
+		Unary(s, spv::OpBitcast, u, Select(s, f, missed, ConstantF32(s, 0x7f800000), numerator)),
+		Unary(s, spv::OpBitcast, u, Select(s, f, missed, ConstantF32Value(s, 1.0f), denominator)),
+		Select(s, u, bary, coordinate(0), Binary(s, spv::OpIAdd, u, flag, kind)),
+		Select(s, u, bary, coordinate(2), Select(s, u, missed, ConstantU32(s, 0), ConstantU32(s, 1))) };
+	return Vector(s, TypeU32Vector(s, 4), result);
+}
+
+std::array<uint32_t, 6> Box(EmitterState& s, uint32_t first, uint32_t second,
+                           uint32_t half, uint32_t child) {
+	const auto f = TypeF32(s);
+	const auto half_label = s.builder.AllocateId(), full_label = s.builder.AllocateId();
+	const auto merge = s.builder.AllocateId();
+	s.builder.AddFunction(spv::OpSelectionMerge, merge, spv::SelectionControlMaskNone);
+	s.builder.AddFunction(spv::OpBranchConditional, half, half_label, full_label);
+	EmitLabel(s, half_label);
+	std::array<uint32_t, 6> packed, full, result;
+	for (uint32_t pair = 0; pair < 3; ++pair) {
+		const auto word = NodeWord(s, first, ConstantU32(s, 4 + child * 3 + pair));
+		const auto unpacked = EmitGlsl<GLSLstd450UnpackHalf2x16, IR::Type::F32x2>(s, word);
+		packed[pair * 2] = Extract(s, f, unpacked, 0);
+		packed[pair * 2 + 1] = Extract(s, f, unpacked, 1);
+	}
+	s.builder.AddFunction(spv::OpBranch, merge);
+	EmitLabel(s, full_label);
+	for (uint32_t component = 0; component < 6; ++component) {
+		const auto word = 4 + child * 6 + component;
+		full[component] = Unary(s, spv::OpBitcast, f,
+		    NodeWord(s, word < 16 ? first : second, ConstantU32(s, word % 16)));
+	}
+	s.builder.AddFunction(spv::OpBranch, merge);
+	EmitLabel(s, merge);
+	for (uint32_t component = 0; component < 6; ++component) {
+		result[component] = s.builder.AllocateId();
+		s.builder.AddFunction(spv::OpPhi, f, result[component], packed[component], half_label,
+		                      full[component], full_label);
+	}
+	return result;
+}
+
+uint32_t Boxes(EmitterState& s, uint32_t first, uint32_t second, uint32_t half,
+               uint32_t grow, uint32_t sort, uint32_t extent, const Vec3& origin, const Vec3& inverse) {
+	const auto f = TypeF32(s), u = TypeU32(s), b = TypeBool(s);
+	const auto zero = ConstantF32Value(s, 0.0f);
+	std::array<uint32_t, 4> depths, indices;
+	Vec3 forward;
+	for (uint32_t axis = 0; axis < 3; ++axis)
+		forward[axis] = Binary(s, spv::OpFOrdGreaterThanEqual, b, inverse[axis], zero);
+	for (uint32_t child = 0; child < 4; ++child) {
+		const auto box = Box(s, first, second, half, child);
+		Vec3 near, far;
+		for (uint32_t axis = 0; axis < 3; ++axis) {
+			const auto lo = Binary(s, spv::OpFMul, f,
+			    Binary(s, spv::OpFSub, f, box[axis], origin[axis]), inverse[axis]);
+			const auto hi = Binary(s, spv::OpFMul, f,
+			    Binary(s, spv::OpFSub, f, box[axis + 3], origin[axis]), inverse[axis]);
+			near[axis] = Select(s, f, forward[axis], lo, hi);
+			far[axis] = Select(s, f, forward[axis], hi, lo);
+		}
+		const auto entry = EmitGlsl<GLSLstd450FMax, IR::Type::F32>(s,
+		    EmitGlsl<GLSLstd450FMax, IR::Type::F32>(s, near[0], near[1]), near[2]);
+		const auto exit = EmitGlsl<GLSLstd450FMin, IR::Type::F32>(s,
+		    EmitGlsl<GLSLstd450FMin, IR::Type::F32>(s, far[0], far[1]), far[2]);
+		// Valid hits have nonnegative exit distance. Saturate ULP growth at +infinity.
+		const auto bits = Unary(s, spv::OpBitcast, u, Select(s, f,
+		    Binary(s, spv::OpFOrdEqual, b, exit, zero), zero, exit));
+		const auto grown_bits = EmitGlsl<GLSLstd450UMin, IR::Type::U32>(s,
+		    Binary(s, spv::OpIAdd, u, bits, grow), ConstantU32(s, 0x7f800000));
+		const auto grown = Unary(s, spv::OpBitcast, f, grown_bits);
+		auto hit = Binary(s, spv::OpLogicalAnd, b,
+		    Binary(s, spv::OpFOrdGreaterThanEqual, b, exit, zero),
+		    Binary(s, spv::OpFOrdLessThan, b, entry, extent));
+		hit = Binary(s, spv::OpLogicalAnd, b, hit,
+		    Binary(s, spv::OpFOrdLessThanEqual, b, entry, grown));
+		depths[child] = Unary(s, spv::OpBitcast, u, Select(s, f, hit,
+		    Select(s, f, Binary(s, spv::OpFOrdGreaterThan, b, entry, zero), entry, zero),
+		    ConstantF32(s, 0x7f800000)));
+		indices[child] = Select(s, u, hit, NodeWord(s, first, ConstantU32(s, child)), ConstantU32(s, ~0u));
+	}
+	// A select network keeps the original child order when sorting is disabled.
+	constexpr std::array<std::pair<uint32_t, uint32_t>, 5> network {{{0,1}, {2,3}, {0,2}, {1,3}, {1,2}}};
+	for (size_t step = 0; step < network.size(); ++step) {
+		const auto [a, c] = network[step];
+		const auto swap = Binary(s, spv::OpLogicalAnd, b, sort,
+		    Binary(s, spv::OpUGreaterThan, b, depths[a], depths[c]));
+		const auto depth = depths[a], index = indices[a];
+		const auto later = [&](uint32_t position) {
+			return std::ranges::any_of(std::span(network).subspan(step + 1), [=](auto pair) {
+				return pair.first == position || pair.second == position;
+			});
+		};
+		if (later(a)) depths[a] = Select(s, u, swap, depths[c], depth);
+		if (later(c)) depths[c] = Select(s, u, swap, depth, depths[c]);
+		indices[a] = Select(s, u, swap, indices[c], index);
+		indices[c] = Select(s, u, swap, index, indices[c]);
+	}
+	return Vector(s, TypeU32Vector(s, 4), indices);
+}
+
+} // namespace
+
+void DefineBvhIntersect(EmitterState& s) {
+	if (!s.requirements.bvh) return;
+	const auto u = TypeU32(s), f = TypeF32(s), b = TypeBool(s), wide = TypeScalarU64(s);
+	const auto vec4 = TypeU32Vector(s, 4), vec3 = TypeF32Vector(s, 3);
+	const auto signature = s.builder.Type(spv::OpTypeFunction, vec4, vec4, u, f, vec3, vec3, vec3);
+	s.bvh_intersect_function = s.builder.AllocateId();
+	s.builder.AddName(s.bvh_intersect_function, "bvh_intersect");
+	s.builder.AddFunction(spv::OpFunction, vec4, s.bvh_intersect_function,
+	                      spv::FunctionControlMaskNone, signature);
+	std::array<uint32_t, 6> args;
+	const std::array types {vec4, u, f, vec3, vec3, vec3};
+	for (uint32_t i = 0; i < args.size(); ++i) {
+		args[i] = s.builder.AllocateId();
+		s.builder.AddFunction(spv::OpFunctionParameter, types[i], args[i]);
+	}
+	EmitLabel(s, s.builder.AllocateId());
+	const auto descriptor = args[0], node = args[1], extent = args[2];
+	Vec3 origin, direction, inverse;
+	for (uint32_t i = 0; i < 3; ++i) {
+		origin[i] = Extract(s, f, args[3], i);
+		direction[i] = Extract(s, f, args[4], i);
+		inverse[i] = Extract(s, f, args[5], i);
+	}
+	std::array<uint32_t, 4> words;
+	for (uint32_t i = 0; i < 4; ++i) words[i] = Extract(s, u, descriptor, i);
+	const auto and_bits = [&](uint32_t value, uint32_t mask) {
+		return Binary(s, spv::OpBitwiseAnd, u, value, ConstantU32(s, mask));
+	};
+	const auto shr = [&](uint32_t value, uint32_t shift) {
+		return Binary(s, spv::OpShiftRightLogical, u, value, ConstantU32(s, shift));
+	};
+	const auto kind = and_bits(node, 7);
+	const auto triangle = Binary(s, spv::OpULessThan, b, kind, ConstantU32(s, 4));
+	const auto full = Binary(s, spv::OpIEqual, b, kind, ConstantU32(s, 5));
+	const auto half = Binary(s, spv::OpIEqual, b, kind, ConstantU32(s, 4));
+	const auto bary = Binary(s, spv::OpINotEqual, b, and_bits(words[3], 0x01000000), ConstantU32(s, 0));
+	const auto sort = Binary(s, spv::OpINotEqual, b, and_bits(words[1], 0x80000000), ConstantU32(s, 0));
+	const auto grow = and_bits(shr(words[1], 23), 0xff);
+	const auto base = Binary(s, spv::OpShiftLeftLogical, wide,
+	    DeviceAddressFromWords(s, words[0], and_bits(words[1], 0xff)), ConstantDeviceAddress(s, 8));
+	const auto index = Unary(s, spv::OpUConvert, wide, shr(node, 3));
+	const auto last_index = Binary(s, spv::OpIAdd, wide, index,
+	    Select(s, wide, full, ConstantDeviceAddress(s, 1), ConstantDeviceAddress(s, 0)));
+	const auto address = Binary(s, spv::OpIAdd, wide, base,
+	    Binary(s, spv::OpShiftLeftLogical, wide, index, ConstantDeviceAddress(s, 6)));
+	const auto last_address = Binary(s, spv::OpIAdd, wide, base,
+	    Binary(s, spv::OpShiftLeftLogical, wide, last_index, ConstantDeviceAddress(s, 6)));
+	auto valid = Binary(s, spv::OpIEqual, b, shr(words[3], 28), ConstantU32(s, 8));
+	const auto require = [&](uint32_t condition) {
+		valid = Binary(s, spv::OpLogicalAnd, b, valid, condition);
+	};
+	require(Binary(s, spv::OpULessThanEqual, b, kind, ConstantU32(s, 5)));
+	require(Binary(s, spv::OpULessThanEqual, b, last_index, DeviceAddressFromWords(s, words[2], and_bits(words[3], 0x3ff))));
+	require(Binary(s, spv::OpULessThan, b, last_address, ConstantDeviceAddress(s, uint64_t{1} << 48)));
+	require(Unary(s, spv::OpLogicalNot, b, Unary(s, spv::OpIsNan, b, extent)));
+	for (uint32_t i = 0; i < 3; ++i) {
+		for (const auto value: {origin[i], direction[i], inverse[i]})
+			require(Unary(s, spv::OpLogicalNot, b, Unary(s, spv::OpIsNan, b, value)));
+		require(Unary(s, spv::OpLogicalNot, b, Unary(s, spv::OpIsInf, b, origin[i])));
+	}
+	const std::array triangle_invalid {ConstantU32(s, 0x7f800000), ConstantU32(s, 0x3f800000),
+	                                  ConstantU32(s, ~0u), ConstantU32(s, 0)};
+	const std::array box_invalid {ConstantU32(s, ~0u), ConstantU32(s, ~0u),
+	                             ConstantU32(s, ~0u), ConstantU32(s, ~0u)};
+	std::array<uint32_t, 4> invalid_words;
+	for (uint32_t i = 0; i < 4; ++i)
+		invalid_words[i] = Select(s, u, triangle, triangle_invalid[i], box_invalid[i]);
+	const auto invalid = Vector(s, vec4, invalid_words);
+	const auto result = EmitValueOrDefaultIfCondition(s, valid, vec4, invalid, [&] {
+		const auto first = GetBdaPointer(s, address);
+		const auto second = EmitValueOrDefaultIfCondition(s, full, wide, ConstantDeviceAddress(s, 0), [&] {
+			return GetBdaPointer(s, Binary(s, spv::OpIAdd, wide, address, ConstantDeviceAddress(s, 64)));
+		});
+		const auto present = Binary(s, spv::OpLogicalAnd, b,
+		    Binary(s, spv::OpINotEqual, b, first, ConstantDeviceAddress(s, 0)),
+		    Binary(s, spv::OpLogicalOr, b, Unary(s, spv::OpLogicalNot, b, full),
+		        Binary(s, spv::OpINotEqual, b, second, ConstantDeviceAddress(s, 0))));
+		return EmitValueOrDefaultIfCondition(s, present, vec4, invalid, [&] {
+			const auto tri_label = s.builder.AllocateId(), box_label = s.builder.AllocateId();
+			const auto merge = s.builder.AllocateId();
+			s.builder.AddFunction(spv::OpSelectionMerge, merge, spv::SelectionControlMaskNone);
+			s.builder.AddFunction(spv::OpBranchConditional, triangle, tri_label, box_label);
+			EmitLabel(s, tri_label);
+			const auto tri_value = Triangle(s, first, kind, bary, origin, direction);
+			const auto tri_exit = s.current_label;
+			s.builder.AddFunction(spv::OpBranch, merge);
+			EmitLabel(s, box_label);
+			const auto box_value = Boxes(s, first, second, half, grow, sort, extent, origin, inverse);
+			const auto box_exit = s.current_label;
+			s.builder.AddFunction(spv::OpBranch, merge);
+			EmitLabel(s, merge);
+			const auto value = s.builder.AllocateId();
+			s.builder.AddFunction(spv::OpPhi, vec4, value, tri_value, tri_exit, box_value, box_exit);
+			return value;
+		});
+	});
+	s.builder.AddFunction(spv::OpReturnValue, result);
+	s.builder.AddFunction(spv::OpFunctionEnd);
+}
+
+uint32_t EmitBvhIntersect(ValueEmitContext& ctx, const IR::Inst& inst) {
+	auto& s = ctx.state;
+	const auto* ray = ctx.ImageAddress(inst.Arg(1));
+	if (ray == nullptr || ray->NumArgs() < 11 || s.bvh_intersect_function == 0)
+		ctx.Fail(inst, "BVH intersection requires eleven address components and its shared function");
+	const auto vector = [&](uint32_t first) {
+		Vec3 values;
+		for (uint32_t i = 0; i < 3; ++i)
+			values[i] = Unary(s, spv::OpBitcast, TypeF32(s), ctx.Arg(*ray, first + i));
+		return Vector(s, TypeF32Vector(s, 3), values);
+	};
+	const auto descriptor = ctx.Arg(inst, 0), node = ctx.Arg(*ray, 0);
+	const auto extent = Unary(s, spv::OpBitcast, TypeF32(s), ctx.Arg(*ray, 1));
+	const auto origin = vector(2), direction = vector(5), inverse = vector(8);
+	return EmitValueOrDefaultIfCondition(s, ctx.Arg(inst, 2), TypeU32Vector(s, 4),
+	    ConstantU32CompositeZero(s, 4), [&] {
+		const auto result = s.builder.AllocateId();
+		s.builder.AddFunction(spv::OpFunctionCall, TypeU32Vector(s, 4), result,
+		    s.bvh_intersect_function, descriptor, node, extent, origin, direction, inverse);
+		return result;
+	});
+}
+
+} // namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter

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

@@ -452,6 +452,11 @@ void EmitSetAttribute(ValueEmitContext& ctx, const IR::Inst& inst) {
 	if (exp.kind == IR::ExportTargetKind::Null || exp.en == 0u) {
 		return;
 	}
+	// Skip dormant color exports after their valid mask; MRT1 is reserved for logical alpha.
+	if (state.program.stage == ShaderType::Pixel && exp.kind == IR::ExportTargetKind::Mrt &&
+	    exp.index != 0 && state.input_info.pixel->alpha_blend_source_remap) {
+		return;
+	}
 	EmitIfCondition(state, exec, [&]() {
 		const auto data = ctx.Arg(inst, 0);
 		if (exp.kind == IR::ExportTargetKind::Primitive) {
@@ -492,6 +497,18 @@ void EmitSetAttribute(ValueEmitContext& ctx, const IR::Inst& inst) {
 		const bool uint_output = MrtOutputMode(state, exp) == 7u;
 		const auto vector_type = uint_output ? TypeU32Vector(state, 4) : TypeF32Vector(state, 4);
 		auto       value       = ExportVector(ctx, data, exp, uint_output);
+		if (state.program.stage == ShaderType::Pixel && exp.kind == IR::ExportTargetKind::Mrt &&
+		    exp.index == 0 && !uint_output && state.input_info.pixel->alpha_blend_source_remap) {
+			// Broadcast logical alpha before swizzling the primary output.
+			const auto blend_output =
+			    OutputVariableForExport(state, {.kind = IR::ExportTargetKind::Mrt, .index = 1});
+			if (blend_output != 0) {
+				const auto alpha = state.builder.AllocateId();
+				state.builder.AddFunction(spv::OpVectorShuffle, vector_type, alpha, value, value,
+				                          3u, 3u, 3u, 3u);
+				state.builder.AddFunction(spv::OpStore, blend_output, alpha);
+			}
+		}
 		if (state.program.stage == ShaderType::Pixel && exp.kind == IR::ExportTargetKind::Mrt &&
 		    exp.index < state.input_info.pixel->target_export_mapping.size()) {
 			const auto mapping = state.input_info.pixel->target_export_mapping[exp.index];

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

@@ -539,19 +539,6 @@ uint32_t StoreTexel(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uint32_t d
 	return PackImageTexel(ctx, mem, texel);
 }
 
-spv::Op ImageAtomicOpcode(IR::ValueOpcode opcode) {
-	switch (opcode) {
-		case IR::ValueOpcode::ImageAtomicSwap32: return spv::OpAtomicExchange;
-		case IR::ValueOpcode::ImageAtomicIAdd32: return spv::OpAtomicIAdd;
-		case IR::ValueOpcode::ImageAtomicUMin32: return spv::OpAtomicUMin;
-		case IR::ValueOpcode::ImageAtomicUMax32: return spv::OpAtomicUMax;
-		case IR::ValueOpcode::ImageAtomicAnd32: return spv::OpAtomicAnd;
-		case IR::ValueOpcode::ImageAtomicOr32: return spv::OpAtomicOr;
-		case IR::ValueOpcode::ImageAtomicXor32: return spv::OpAtomicXor;
-		default: return spv::OpNop;
-	}
-}
-
 } // namespace
 
 void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
@@ -876,10 +863,9 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 		ctx.Define(inst, ResultVector(ctx, result, numeric_class, dref, mem));
 		return;
 	}
-	const auto atomic_opcode = ImageAtomicOpcode(op);
 	if (image_info.access == IR::ImageAccess::Atomic) {
 		const auto dimension = image.dimension;
-		ctx.Define(inst, EmitValueOrZeroIfCondition(state, ctx.Arg(inst, 3), [&]() {
+		ctx.Define(inst, EmitValueOrZeroIfCondition(state, ctx.Arg(inst, inst.NumArgs() - 1), [&]() {
 			           const auto pointer      = state.builder.AllocateId();
 			           const auto pointer_type = state.builder.Type(
 			               spv::OpTypePointer, spv::StorageClassImage, TypeU32(state));
@@ -896,12 +882,8 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 					                                   op == IR::ValueOpcode::ImageAtomicFMax32);
 				                               });
 			           }
-			           const auto old = state.builder.AllocateId();
-			           state.builder.AddFunction(atomic_opcode, TypeU32(state), old, pointer,
-			                                     ConstantU32(state, spv::ScopeDevice),
-			                                     ConstantU32(state, spv::MemorySemanticsMaskNone),
-			                                     ctx.Arg(inst, 2));
-			           EmitDeviceAtomicMemoryBarrier(state);
+			           const auto old = EmitAtomicOperation(ctx, inst, pointer, spv::ScopeDevice);
+			           EmitAtomicMemoryBarrier(state, IR::ResourceKind::Image);
 			           return old;
 		           }));
 		return;

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

@@ -111,6 +111,9 @@ uint32_t EmitIEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1);
 uint32_t EmitINotEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1);
 uint32_t EmitULessThan64(EmitterState& state, uint32_t arg0, uint32_t arg1);
 uint32_t EmitSLessThan64(EmitterState& state, uint32_t arg0, uint32_t arg1);
+uint32_t EmitSLessThanEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1);
+uint32_t EmitULessThanEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1);
+uint32_t EmitUGreaterThanEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1);
 uint32_t EmitUGreaterThan64(EmitterState& state, uint32_t arg0, uint32_t arg1);
 EMIT_NATIVE(LogicalOr, OpLogicalOr, U1, uint32_t, uint32_t)
 EMIT_NATIVE(LogicalAnd, OpLogicalAnd, U1, uint32_t, uint32_t)
@@ -133,6 +136,9 @@ uint32_t EmitFloatCompare32(ValueEmitContext& ctx, const IR::Inst& inst) {
 	return EmitNative<opcode, IR::Type::U1>(ctx.state, operand(0), operand(1));
 }
 inline constexpr auto EmitFPOrdEqual32 = EmitFloatCompare32<spv::OpFOrdEqual>;
+EMIT_NATIVE(FPOrdEqual64, OpFOrdEqual, U1, uint32_t, uint32_t)
+EMIT_NATIVE(FPOrdLessThanEqual64, OpFOrdLessThanEqual, U1, uint32_t, uint32_t)
+EMIT_NATIVE(FPOrdGreaterThanEqual64, OpFOrdGreaterThanEqual, U1, uint32_t, uint32_t)
 inline constexpr auto EmitFPUnordEqual32 = EmitFloatCompare32<spv::OpFUnordEqual>;
 inline constexpr auto EmitFPOrdNotEqual32 = EmitFloatCompare32<spv::OpFOrdNotEqual>;
 inline constexpr auto EmitFPUnordNotEqual32 = EmitFloatCompare32<spv::OpFUnordNotEqual>;
@@ -149,6 +155,7 @@ inline constexpr auto EmitFPCmpClass32 = EmitClassMaskF32;
 EMIT_NATIVE(FPAdd32, OpFAdd, F32, uint32_t, uint32_t)
 EMIT_NATIVE(FPSub32, OpFSub, F32, uint32_t, uint32_t)
 EMIT_NATIVE(FPMul32, OpFMul, F32, uint32_t, uint32_t)
+EMIT_NATIVE(FPAdd64, OpFAdd, F64, uint32_t, uint32_t)
 EMIT_NATIVE(FPMul64, OpFMul, F64, uint32_t, uint32_t)
 inline constexpr auto EmitFPFma64 =
     EmitGlsl<GLSLstd450Fma, IR::Type::F64, uint32_t, uint32_t, uint32_t>;
@@ -222,6 +229,7 @@ void                  EmitLoadMemory(ValueEmitContext& ctx, const IR::Inst& inst
 void                  EmitStoreMemory(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitAtomic32(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitBufferAtomic64(ValueEmitContext& ctx, const IR::Inst& inst);
+void                  EmitSharedAtomic64(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitBufferFloatAtomic(ValueEmitContext& ctx, const IR::Inst& inst);
 void                  EmitSharedFloatAtomic(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitSharedIncDec(ValueEmitContext& ctx, const IR::Inst& inst);
@@ -257,6 +265,7 @@ inline constexpr auto EmitBufferAtomicUMin32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicSMax32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicUMax32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicAnd32     = EmitAtomic32;
+inline constexpr auto EmitBufferAtomicAnd64     = EmitBufferAtomic64;
 inline constexpr auto EmitBufferAtomicOr32      = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicOr64      = EmitBufferAtomic64;
 inline constexpr auto EmitBufferAtomicXor32     = EmitAtomic32;
@@ -278,6 +287,7 @@ inline constexpr auto EmitSharedAtomicFMin32    = EmitSharedFloatAtomic;
 inline constexpr auto EmitSharedAtomicFMax32    = EmitSharedFloatAtomic;
 inline constexpr auto EmitSharedAtomicSwap32    = EmitAtomic32;
 inline constexpr auto EmitSharedAtomicIAdd32    = EmitAtomic32;
+inline constexpr auto EmitSharedAtomicIAdd64    = EmitSharedAtomic64;
 inline constexpr auto EmitSharedAtomicISub32    = EmitAtomic32;
 inline constexpr auto EmitSharedAtomicInc32     = EmitSharedIncDec;
 inline constexpr auto EmitSharedAtomicDec32     = EmitSharedIncDec;
@@ -287,20 +297,25 @@ inline constexpr auto EmitSharedAtomicSMax32    = EmitAtomic32;
 inline constexpr auto EmitSharedAtomicUMax32    = EmitAtomic32;
 inline constexpr auto EmitSharedAtomicAnd32     = EmitAtomic32;
 inline constexpr auto EmitSharedAtomicOr32      = EmitAtomic32;
+inline constexpr auto EmitSharedAtomicOr64      = EmitSharedAtomic64;
 inline constexpr auto EmitSharedAtomicXor32     = EmitAtomic32;
 inline constexpr auto EmitDataAppend            = EmitAppendConsume;
 inline constexpr auto EmitDataConsume           = EmitAppendConsume;
 uint32_t              EmitSwizzleU32(ValueEmitContext& ctx, const IR::Inst& inst);
 void                  EmitImage(ValueEmitContext& ctx, const IR::Inst& inst);
+uint32_t              EmitBvhIntersect(ValueEmitContext& ctx, const IR::Inst& inst);
 inline constexpr auto EmitImageQueryDimensions = EmitImage;
 inline constexpr auto EmitImageQueryLod        = EmitImage;
 inline constexpr auto EmitImageRead            = EmitImage;
 inline constexpr auto EmitImageWrite           = EmitImage;
 inline constexpr auto EmitImageSampleRaw       = EmitImage;
 inline constexpr auto EmitImageGatherRaw       = EmitImage;
+inline constexpr auto EmitImageAtomicCompareSwap32 = EmitImage;
 inline constexpr auto EmitImageAtomicSwap32    = EmitImage;
 inline constexpr auto EmitImageAtomicIAdd32    = EmitImage;
+inline constexpr auto EmitImageAtomicSMin32    = EmitImage;
 inline constexpr auto EmitImageAtomicUMin32    = EmitImage;
+inline constexpr auto EmitImageAtomicSMax32    = EmitImage;
 inline constexpr auto EmitImageAtomicUMax32    = EmitImage;
 inline constexpr auto EmitImageAtomicAnd32     = EmitImage;
 inline constexpr auto EmitImageAtomicOr32      = EmitImage;

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

@@ -58,6 +58,7 @@ constexpr std::array<ImageDimensionInfo, 7> ImageDimensions {{
 const ImageDimensionInfo& ImageDimensionInfoFor(ImageDimension dimension);
 
 struct SpirvRequirements {
+	bool bvh                          = false;
 	bool subgroup_ballot              = false;
 	bool subgroup_shuffle             = false;
 	bool subgroup_local_invocation_id = false;
@@ -67,6 +68,7 @@ struct SpirvRequirements {
 	bool function_scratch             = false;
 	bool pixel_valid_mask             = false;
 	bool buffer_int64_atomics         = false;
+	bool shared_int64_atomics         = false;
 	bool coherent_buffers             = false;
 	bool float64                      = false;
 };
@@ -82,6 +84,20 @@ struct EmitterState {
 	Builder                                          builder;
 	const IR::Program&                               program;
 	ShaderStageInputInfo                             input_info;
+	uint32_t                                        void_type = 0;
+	uint32_t                                        bool_type = 0;
+	uint32_t                                        u32_type = 0;
+	uint32_t                                        native_u64_type = 0;
+	uint32_t                                        i32_type = 0;
+	uint32_t                                        f32_type = 0;
+	uint32_t                                        f64_type = 0;
+	uint32_t                                        u32_pair_type = 0;
+	uint32_t                                        i32_pair_type = 0;
+	uint32_t                                        function_type = 0;
+	std::array<uint32_t, 3>                          bool_vector_types {};
+	std::array<uint32_t, 3>                          u32_vector_types {};
+	std::array<uint32_t, 3>                          i32_vector_types {};
+	std::array<uint32_t, 3>                          f32_vector_types {};
 	std::array<uint32_t, 6>                          tess_variables {};
 	uint32_t                                         tess_inner_variable = 0;
 	uint32_t                                         tess_patch_base     = 0;
@@ -95,12 +111,14 @@ struct EmitterState {
 	uint32_t                                         bda_pagetable_variable  = 0;
 	uint32_t                                         fault_buffer_variable   = 0;
 	uint32_t                                         bda_pointer_function    = 0;
+	uint32_t                                         bvh_intersect_function  = 0;
 	uint32_t                                         gds_variable            = 0;
 	uint32_t                                         gds_length              = 0;
 	uint32_t                                         push_constant_variable  = 0;
 	uint32_t                                         shader_data_storage_variable = 0;
 	uint32_t                                         flattened_srt_variable  = 0;
 	uint32_t                                         lds_variable            = 0;
+	uint32_t                                         lds_u64_variable        = 0;
 	std::array<uint32_t, 2>                          scratch_variable {};
 	std::array<uint32_t, IR::ImageBindingCount>      image_variables {};
 	uint32_t                   sampler_variable                      = 0;
@@ -376,6 +394,7 @@ Prospero::BufferFormat StorageBufferFormat(const EmitterState& state, const IR::
 void EmitMemoryOffsets(EmitterState& state);
 
 uint32_t LdsDwordCount(const EmitterState& state);
+void EnsureLdsStorage(EmitterState& state);
 
 struct MemoryResourceAccess {
 	IR::ResourceKind      kind             = IR::ResourceKind::None;
@@ -417,7 +436,12 @@ uint32_t EmitUFloatToF32Bits(EmitterState& state, uint32_t raw, uint32_t bits);
 uint32_t NormalizeFormatComponent(EmitterState& state, const Format::BufferFormatInfo& info,
                                   uint32_t component, uint32_t raw);
 
-void EmitDeviceAtomicMemoryBarrier(EmitterState& state);
+spv::Op SpirvAtomicOpcode(IR::ValueOpcode opcode);
+
+uint32_t EmitAtomicOperation(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t pointer,
+                             uint32_t scope);
+
+void EmitAtomicMemoryBarrier(EmitterState& state, IR::ResourceKind kind);
 
 uint32_t EmitFloatAtomicReplacement(EmitterState& state, uint32_t old, uint32_t source,
                                     bool max_value);
@@ -484,6 +508,10 @@ uint32_t EmitF16BitsToF32(EmitterState& state, uint32_t bits);
 void EmitProgram(EmitterState& state);
 
 void DefineGetBdaPointer(EmitterState& state);
+void DefineBvhIntersect(EmitterState& state);
+uint32_t GetBdaPointer(EmitterState& state, uint32_t address);
+uint32_t ConstantDeviceAddress(EmitterState& state, uint64_t value);
+uint32_t DeviceAddressFromWords(EmitterState& state, uint32_t low, uint32_t high);
 
 // These templates accept local lambdas from several emitter translation units.
 template <typename Fn>
@@ -525,14 +553,7 @@ uint32_t EmitValueOrZeroIfCondition(EmitterState& state, uint32_t condition, Fn&
 
 template <typename Fn>
 uint32_t AtomicUpdate(EmitterState& state, uint32_t pointer, IR::ResourceKind kind, Fn&& desired) {
-	const auto scope  = kind == IR::ResourceKind::Lds ? spv::ScopeWorkgroup : spv::ScopeDevice;
-	const auto memory = [&] {
-		switch (kind) {
-			case IR::ResourceKind::Lds: return spv::MemorySemanticsWorkgroupMemoryMask;
-			case IR::ResourceKind::Image: return spv::MemorySemanticsImageMemoryMask;
-			default: return spv::MemorySemanticsUniformMemoryMask;
-		}
-	}();
+	const auto scope     = kind == IR::ResourceKind::Lds ? spv::ScopeWorkgroup : spv::ScopeDevice;
 	const auto preheader = state.builder.AllocateId();
 	const auto header    = state.builder.AllocateId();
 	const auto cont      = state.builder.AllocateId();
@@ -561,8 +582,7 @@ uint32_t AtomicUpdate(EmitterState& state, uint32_t pointer, IR::ResourceKind ki
 	EmitLabel(state, cont);
 	state.builder.AddFunction(spv::OpBranch, header);
 	EmitLabel(state, merge);
-	state.builder.AddFunction(spv::OpMemoryBarrier, ConstantU32(state, scope),
-	                          ConstantU32(state, spv::MemorySemanticsAcquireReleaseMask | memory));
+	EmitAtomicMemoryBarrier(state, kind);
 	return observed;
 }
 

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

@@ -1,6 +1,7 @@
 #include "graphics/shader/recompiler/backend/spirv/spirvEmitterInstructions.h"
 
 #include "graphics/host_gpu/renderer/cache/bufferCache.h"
+#include "graphics/shader/recompiler/frontend/decode/ShaderDecoder.h"
 
 #include <algorithm>
 
@@ -105,31 +106,23 @@ uint32_t AddU64Low(EmitterState& state, uint32_t low, uint32_t high, uint32_t ad
 	return result;
 }
 
-uint32_t ScratchByteAddress(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uint32_t low,
-                            uint32_t high) {
-	auto& state     = ctx.state;
-	auto  immediate = static_cast<int32_t>(mem.offset);
+uint32_t AddAddressOffset(EmitterState& state, const IR::MemoryInfo& mem, uint32_t low,
+                          uint32_t& high) {
+	const auto immediate = static_cast<int32_t>(mem.offset);
+	if (immediate == 0) return low;
 	const auto immediate_low  = ConstantU32(state, static_cast<uint32_t>(immediate));
 	const auto immediate_high = ConstantU32(state, immediate < 0 ? UINT32_MAX : 0u);
-	low                      = AddU64Low(state, low, high, immediate_low, immediate_high, high);
+	return AddU64Low(state, low, high, immediate_low, immediate_high, high);
+}
+
+uint32_t ScratchByteAddress(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uint32_t low,
+                            uint32_t high) {
+	auto& state = ctx.state;
+	low = AddAddressOffset(state, mem, low, high);
 	const auto valid = Binary(state, spv::OpIEqual, TypeBool(state), high, ConstantU32(state, 0));
 	return Select(state, TypeU32(state), valid, low, ConstantU32(state, UINT32_MAX));
 }
 
-uint32_t ConstantDeviceAddress(EmitterState& state, uint64_t value) {
-	return state.builder.Constant(spv::OpConstant, TypeScalarU64(state),
-	                              static_cast<uint32_t>(value),
-	                              static_cast<uint32_t>(value >> 32u));
-}
-
-uint32_t DeviceAddressFromWords(EmitterState& state, uint32_t low, uint32_t high) {
-	const auto low64  = Unary(state, spv::OpUConvert, TypeScalarU64(state), low);
-	const auto high64 = Binary(state, spv::OpShiftLeftLogical, TypeScalarU64(state),
-	                           Unary(state, spv::OpUConvert, TypeScalarU64(state), high),
-	                           ConstantDeviceAddress(state, 32));
-	return Binary(state, spv::OpBitwiseOr, TypeScalarU64(state), low64, high64);
-}
-
 uint32_t GuestAddress(ValueEmitContext& ctx, const IR::Inst& inst, const IR::MemoryInfo& mem) {
 	auto& state = ctx.state;
 	auto  low   = ctx.Arg(inst, 1);
@@ -186,17 +179,9 @@ void RecordBdaFault(EmitterState& state, uint32_t page) {
 	                          Binary(state, spv::OpBitwiseOr, TypeU32(state), value, bit));
 }
 
-uint32_t GetBdaPointer(ValueEmitContext& ctx, uint32_t address) {
-	auto&      state  = ctx.state;
-	const auto result = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpFunctionCall, TypeScalarU64(state), result,
-	                          state.bda_pointer_function, address);
-	return result;
-}
-
 uint32_t LoadBdaDword(ValueEmitContext& ctx, uint32_t address) {
 	auto&      state   = ctx.state;
-	const auto bda     = GetBdaPointer(ctx, address);
+	const auto bda     = GetBdaPointer(state, address);
 	const auto present =
 	    Binary(state, spv::OpINotEqual, TypeBool(state), bda, ConstantDeviceAddress(state, 0));
 	return EmitValueOrZeroIfCondition(state, present, [&]() {
@@ -511,6 +496,50 @@ void StoreWordInBounds(ValueEmitContext& ctx, const MemoryResourceAccess& resour
 	                              resource.memory_access);
 }
 
+template <typename Fn>
+void ForEachLocalFlatAccess(ValueEmitContext& ctx, const IR::Inst& inst, Fn&& emit) {
+	auto& state = ctx.state;
+	auto mem = ctx.Memory(inst);
+	auto high = ctx.Arg(inst, 2);
+	const auto low = AddAddressOffset(state, mem, ctx.Arg(inst, 1), high);
+	const auto index = Binary(state, spv::OpShiftRightLogical, TypeU32(state), low, ConstantU32(state, 2));
+	const auto aligned = Binary(state, spv::OpIEqual, TypeBool(state), ConstantU32(state, 0),
+	    Binary(state, spv::OpBitwiseAnd, TypeU32(state), low, ConstantU32(state, 3)));
+	for (const auto kind: {IR::ResourceKind::Scratch, IR::ResourceKind::Lds}) {
+		mem.kind = kind;
+		const auto resource = PrepareMemoryResourceAccess(state, mem);
+		const auto aperture = kind == IR::ResourceKind::Scratch ? Decoder::PrivateApertureHigh
+		                                                       : Decoder::SharedApertureHigh;
+		const auto selected = Binary(state, spv::OpIEqual, TypeBool(state), high, ConstantU32(state, aperture));
+		const auto valid = AndCondition(state, aligned,
+		    AndCondition(state, selected, EmitMemoryElementInBounds(state, resource, index)));
+		emit(resource, index, valid);
+	}
+}
+
+uint32_t LoadLocalFlat(ValueEmitContext& ctx, const IR::Inst& inst) {
+	return EmitValueOrZeroIfCondition(ctx.state, ctx.Arg(inst, inst.NumArgs() - 1), [&]() {
+		uint32_t result = 0;
+		ForEachLocalFlatAccess(ctx, inst, [&](const auto& resource, uint32_t index, uint32_t valid) {
+			const auto value = EmitValueOrZeroIfCondition(ctx.state, valid, [&]() {
+				return LoadWordInBounds(ctx, resource, index);
+			});
+			result = result == 0 ? value : Binary(ctx.state, spv::OpBitwiseOr, TypeU32(ctx.state), result, value);
+		});
+		return result;
+	});
+}
+
+void StoreLocalFlat(ValueEmitContext& ctx, const IR::Inst& inst) {
+	EmitIfCondition(ctx.state, ctx.Arg(inst, inst.NumArgs() - 1), [&]() {
+		ForEachLocalFlatAccess(ctx, inst, [&](const auto& resource, uint32_t index, uint32_t valid) {
+			EmitIfCondition(ctx.state, valid, [&]() {
+				StoreWordInBounds(ctx, resource, index, ctx.Arg(inst, inst.NumArgs() - 2));
+			});
+		});
+	});
+}
+
 void StoreWordPrepared(ValueEmitContext& ctx, const IR::Inst& inst, const IR::MemoryInfo& mem,
                        const MemoryResourceAccess& resource, uint32_t data) {
 	const auto index = EmitMemoryElementIndex(ctx.state, resource, DwordIndex(ctx, inst, mem));
@@ -526,54 +555,6 @@ void StoreWord(ValueEmitContext& ctx, const IR::Inst& inst, IR::MemoryInfo mem)
 	});
 }
 
-spv::Op SpirvAtomicOpcode(IR::ValueOpcode opcode) {
-	switch (opcode) {
-		case IR::ValueOpcode::BufferAtomicCmpSwap32: return spv::OpAtomicCompareExchange;
-		case IR::ValueOpcode::BufferAtomicSwap32:
-		case IR::ValueOpcode::BufferAtomicSwap64:
-		case IR::ValueOpcode::SharedAtomicSwap32: return spv::OpAtomicExchange;
-		case IR::ValueOpcode::BufferAtomicIAdd32:
-		case IR::ValueOpcode::SharedAtomicIAdd32: return spv::OpAtomicIAdd;
-		case IR::ValueOpcode::BufferAtomicISub32:
-		case IR::ValueOpcode::SharedAtomicISub32: return spv::OpAtomicISub;
-		case IR::ValueOpcode::BufferAtomicSMin32:
-		case IR::ValueOpcode::SharedAtomicSMin32: return spv::OpAtomicSMin;
-		case IR::ValueOpcode::BufferAtomicUMin32:
-		case IR::ValueOpcode::SharedAtomicUMin32: return spv::OpAtomicUMin;
-		case IR::ValueOpcode::BufferAtomicSMax32:
-		case IR::ValueOpcode::SharedAtomicSMax32: return spv::OpAtomicSMax;
-		case IR::ValueOpcode::BufferAtomicUMax32:
-		case IR::ValueOpcode::SharedAtomicUMax32: return spv::OpAtomicUMax;
-		case IR::ValueOpcode::BufferAtomicAnd32:
-		case IR::ValueOpcode::SharedAtomicAnd32: return spv::OpAtomicAnd;
-		case IR::ValueOpcode::BufferAtomicOr32:
-		case IR::ValueOpcode::BufferAtomicOr64:
-		case IR::ValueOpcode::SharedAtomicOr32: return spv::OpAtomicOr;
-		case IR::ValueOpcode::BufferAtomicXor32:
-		case IR::ValueOpcode::SharedAtomicXor32: return spv::OpAtomicXor;
-		default: return spv::OpNop;
-	}
-}
-
-uint32_t EmitAtomicOperation(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t pointer,
-                             uint32_t scope) {
-	const auto old = ctx.state.builder.AllocateId();
-	if (inst.GetOpcode() == IR::ValueOpcode::BufferAtomicCmpSwap32) {
-		const auto desired    = ctx.Arg(inst, inst.NumArgs() - 3);
-		const auto comparator = ctx.Arg(inst, inst.NumArgs() - 2);
-		ctx.state.builder.AddFunction(
-		    spv::OpAtomicCompareExchange, TypeU32(ctx.state), old, pointer,
-		    ConstantU32(ctx.state, scope), ConstantU32(ctx.state, spv::MemorySemanticsMaskNone),
-		    ConstantU32(ctx.state, spv::MemorySemanticsMaskNone), desired, comparator);
-	} else {
-		const auto value = ctx.Arg(inst, inst.NumArgs() - 2);
-		ctx.state.builder.AddFunction(SpirvAtomicOpcode(inst.GetOpcode()), TypeU32(ctx.state), old,
-		                              pointer, ConstantU32(ctx.state, scope),
-		                              ConstantU32(ctx.state, spv::MemorySemanticsMaskNone), value);
-	}
-	return old;
-}
-
 template <typename Fn>
 uint32_t EmitAtomicAccess(ValueEmitContext& ctx, const IR::Inst& inst,
                           const IR::MemoryInfo& mem, Fn&& operation) {
@@ -960,6 +941,27 @@ void StoreWideShared(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t compo
 
 } // namespace
 
+uint32_t ConstantDeviceAddress(EmitterState& state, uint64_t value) {
+	return state.builder.Constant(spv::OpConstant, TypeScalarU64(state),
+	                              static_cast<uint32_t>(value),
+	                              static_cast<uint32_t>(value >> 32u));
+}
+
+uint32_t DeviceAddressFromWords(EmitterState& state, uint32_t low, uint32_t high) {
+	const auto low64  = Unary(state, spv::OpUConvert, TypeScalarU64(state), low);
+	const auto high64 = Binary(state, spv::OpShiftLeftLogical, TypeScalarU64(state),
+	                           Unary(state, spv::OpUConvert, TypeScalarU64(state), high),
+	                           ConstantDeviceAddress(state, 32));
+	return Binary(state, spv::OpBitwiseOr, TypeScalarU64(state), low64, high64);
+}
+
+uint32_t GetBdaPointer(EmitterState& state, uint32_t address) {
+	const auto result = state.builder.AllocateId();
+	state.builder.AddFunction(spv::OpFunctionCall, TypeScalarU64(state), result,
+	                          state.bda_pointer_function, address);
+	return result;
+}
+
 void DefineGetBdaPointer(EmitterState& state) {
 	if (!state.program.info.uses_dma) {
 		return;
@@ -1023,14 +1025,7 @@ uint32_t EmitAtomic32(ValueEmitContext& ctx, const IR::Inst& inst) {
 		const auto scope =
 		    mem.kind == IR::ResourceKind::Lds ? spv::ScopeWorkgroup : spv::ScopeDevice;
 		const auto old = EmitAtomicOperation(ctx, inst, pointer, scope);
-		if (mem.kind == IR::ResourceKind::Lds) {
-			const auto semantics =
-			    spv::MemorySemanticsAcquireReleaseMask | spv::MemorySemanticsWorkgroupMemoryMask;
-			ctx.state.builder.AddFunction(spv::OpMemoryBarrier, ConstantU32(ctx.state, scope),
-			                              ConstantU32(ctx.state, semantics));
-		} else {
-			EmitDeviceAtomicMemoryBarrier(ctx.state);
-		}
+		EmitAtomicMemoryBarrier(ctx.state, mem.kind);
 		return old;
 	});
 }
@@ -1058,12 +1053,38 @@ uint32_t EmitBufferAtomic64(ValueEmitContext& ctx, const IR::Inst& inst) {
 			                                            TypeStorageBufferU64ElementPointer(state)),
 			            ConstantU32(state, spv::ScopeDevice),
 			            ConstantU32(state, spv::MemorySemanticsMaskNone), value);
-			        EmitDeviceAtomicMemoryBarrier(state);
+			        EmitAtomicMemoryBarrier(state, mem.kind);
 			        return Unary(state, spv::OpBitcast, TypeU64(state), old);
 		        });
 	    });
 }
 
+void EmitSharedAtomic64(ValueEmitContext& ctx, const IR::Inst& inst) {
+	auto& state = ctx.state;
+	const auto& mem = ctx.Memory(inst);
+	EnsureLdsStorage(state);
+	EmitIfCondition(state, ctx.Arg(inst, 2), [&]() {
+		const auto address = Binary(state, spv::OpBitwiseAnd, TypeU32(state),
+		                            ByteAddress(ctx, inst, mem), ConstantU32(state, 0xfff8u));
+		const auto index = Binary(state, spv::OpShiftRightLogical, TypeU32(state), address,
+		                          ConstantU32(state, 3u));
+		const auto in_bounds = Binary(state, spv::OpULessThan, TypeBool(state), index,
+		                              ConstantU32(state, LdsDwordCount(state) / 2u));
+		EmitIfCondition(state, in_bounds, [&]() {
+			const auto pointer = state.builder.AllocateId();
+			state.builder.AddFunction(spv::OpAccessChain,
+			                          TypePointer(state, spv::StorageClassWorkgroup, TypeScalarU64(state)),
+			                          pointer, state.lds_u64_variable, ConstantU32(state, 0), index);
+			const auto value = Unary(state, spv::OpBitcast, TypeScalarU64(state), ctx.Arg(inst, 1));
+			state.builder.AddFunction(
+			    SpirvAtomicOpcode(inst.GetOpcode()), TypeScalarU64(state), state.builder.AllocateId(),
+			    pointer, ConstantU32(state, spv::ScopeWorkgroup),
+			    ConstantU32(state, spv::MemorySemanticsAcquireReleaseMask |
+			                           spv::MemorySemanticsWorkgroupMemoryMask), value);
+		});
+	});
+}
+
 uint32_t EmitBufferFloatAtomic(ValueEmitContext& ctx, const IR::Inst& inst) {
 	const auto& mem       = ctx.Memory(inst);
 	const bool  max_value = inst.GetOpcode() == IR::ValueOpcode::BufferAtomicFMax32;
@@ -1216,7 +1237,9 @@ void EmitLoadMemory(ValueEmitContext& ctx, const IR::Inst& inst) {
 	const auto shared_components = IR::SharedComponentCount(op);
 	const auto address_info      = IR::AddressOpcodeInfoOf(op);
 	uint32_t   value;
-	if (buffer_components > 1u)
+	if (mem.kind == IR::ResourceKind::FlatLocal)
+		value = LoadLocalFlat(ctx, inst);
+	else if (buffer_components > 1u)
 		value = LoadWideBuffer(ctx, inst, buffer_components);
 	else if (shared_components > 1u)
 		value = LoadWideShared(ctx, inst, shared_components);
@@ -1243,7 +1266,9 @@ void EmitStoreMemory(ValueEmitContext& ctx, const IR::Inst& inst) {
 	const auto  buffer_components = IR::BufferComponentCount(op);
 	const auto  shared_components = IR::SharedComponentCount(op);
 	const auto  type              = inst.Arg(inst.NumArgs() - 2).GetType();
-	if (buffer_components > 1u)
+	if (mem.kind == IR::ResourceKind::FlatLocal)
+		StoreLocalFlat(ctx, inst);
+	else if (buffer_components > 1u)
 		StoreWideBuffer(ctx, inst, buffer_components);
 	else if (shared_components > 1u)
 		StoreWideShared(ctx, inst, shared_components);

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

@@ -63,16 +63,34 @@ uint32_t LdsDwordCount(const EmitterState& state) {
 	return workgroup != nullptr ? workgroup->lds_size_dwords : 8192u;
 }
 
-static void EnsureLdsStorage(EmitterState& state) {
+void EnsureLdsStorage(EmitterState& state) {
 	if (state.lds_variable != 0) {
 		return;
 	}
 	if (ShaderWorkgroupInput(state.program.stage, state.input_info) == nullptr) {
 		EXIT("function LDS was not prepared before SPIR-V function emission\n");
 	}
-	state.lds_variable = state.builder.DefineGlobalVariable(
-	    TypeU32ArrayPointer(state, spv::StorageClassWorkgroup, LdsDwordCount(state)),
-	    spv::StorageClassWorkgroup);
+	const auto define = [&](uint32_t type, uint32_t bytes) {
+		const auto array = state.builder.DecoratedType(
+		    spv::OpTypeArray, {{spv::OpDecorate, {spv::DecorationArrayStride, bytes}}}, type,
+		    ConstantU32(state, std::max(LdsDwordCount(state) * 4u / bytes, 1u)));
+		const auto block = state.builder.DecoratedType(
+		    spv::OpTypeStruct, {{spv::OpMemberDecorate, {0, spv::DecorationOffset, 0}},
+		                        {spv::OpDecorate, {spv::DecorationBlock}}}, array);
+		const auto variable = state.builder.DefineGlobalVariable(
+		    TypePointer(state, spv::StorageClassWorkgroup, block), spv::StorageClassWorkgroup);
+		state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationAliased);
+		return variable;
+	};
+	if (state.requirements.shared_int64_atomics) {
+		state.lds_variable = define(TypeU32(state), 4u);
+		state.lds_u64_variable = define(TypeScalarU64(state), 8u);
+		state.builder.AddName(state.lds_u64_variable, "lds_qwords");
+	} else {
+		state.lds_variable = state.builder.DefineGlobalVariable(
+		    TypeU32ArrayPointer(state, spv::StorageClassWorkgroup, LdsDwordCount(state)),
+		    spv::StorageClassWorkgroup);
+	}
 	state.builder.AddName(state.lds_variable, "lds_dwords");
 }
 
@@ -167,6 +185,11 @@ uint32_t EmitMemoryElementPointer(EmitterState& state, const MemoryResourceAcces
 		    : ShaderWorkgroupInput(state.program.stage, state.input_info) != nullptr
 		        ? spv::StorageClassWorkgroup
 		        : spv::StorageClassFunction;
+		if (access.kind == IR::ResourceKind::Lds && state.requirements.shared_int64_atomics) {
+			state.builder.AddFunction(spv::OpAccessChain, TypeU32ElementPointer(state, storage_class),
+			                          pointer, access.object_pointer, ConstantU32(state, 0), index);
+			return pointer;
+		}
 		state.builder.AddFunction(spv::OpAccessChain, TypeU32ElementPointer(state, storage_class),
 		                          pointer, access.object_pointer, index);
 		return pointer;
@@ -283,11 +306,79 @@ uint32_t NormalizeFormatComponent(EmitterState& state, const Format::BufferForma
 	}
 }
 
-void EmitDeviceAtomicMemoryBarrier(EmitterState& state) {
-	const auto semantics =
-	    spv::MemorySemanticsAcquireReleaseMask | spv::MemorySemanticsUniformMemoryMask;
-	state.builder.AddFunction(spv::OpMemoryBarrier, ConstantU32(state, spv::ScopeDevice),
-	                          ConstantU32(state, semantics));
+spv::Op SpirvAtomicOpcode(IR::ValueOpcode opcode) {
+	switch (opcode) {
+		case IR::ValueOpcode::ImageAtomicCompareSwap32:
+		case IR::ValueOpcode::BufferAtomicCmpSwap32: return spv::OpAtomicCompareExchange;
+		case IR::ValueOpcode::ImageAtomicSwap32:
+		case IR::ValueOpcode::BufferAtomicSwap32:
+		case IR::ValueOpcode::BufferAtomicSwap64:
+		case IR::ValueOpcode::SharedAtomicSwap32: return spv::OpAtomicExchange;
+		case IR::ValueOpcode::ImageAtomicIAdd32:
+		case IR::ValueOpcode::BufferAtomicIAdd32:
+		case IR::ValueOpcode::SharedAtomicIAdd64:
+		case IR::ValueOpcode::SharedAtomicIAdd32: return spv::OpAtomicIAdd;
+		case IR::ValueOpcode::BufferAtomicISub32:
+		case IR::ValueOpcode::SharedAtomicISub32: return spv::OpAtomicISub;
+		case IR::ValueOpcode::ImageAtomicSMin32:
+		case IR::ValueOpcode::BufferAtomicSMin32:
+		case IR::ValueOpcode::SharedAtomicSMin32: return spv::OpAtomicSMin;
+		case IR::ValueOpcode::ImageAtomicUMin32:
+		case IR::ValueOpcode::BufferAtomicUMin32:
+		case IR::ValueOpcode::SharedAtomicUMin32: return spv::OpAtomicUMin;
+		case IR::ValueOpcode::ImageAtomicSMax32:
+		case IR::ValueOpcode::BufferAtomicSMax32:
+		case IR::ValueOpcode::SharedAtomicSMax32: return spv::OpAtomicSMax;
+		case IR::ValueOpcode::ImageAtomicUMax32:
+		case IR::ValueOpcode::BufferAtomicUMax32:
+		case IR::ValueOpcode::SharedAtomicUMax32: return spv::OpAtomicUMax;
+		case IR::ValueOpcode::ImageAtomicAnd32:
+		case IR::ValueOpcode::BufferAtomicAnd32:
+		case IR::ValueOpcode::BufferAtomicAnd64:
+		case IR::ValueOpcode::SharedAtomicAnd32: return spv::OpAtomicAnd;
+		case IR::ValueOpcode::ImageAtomicOr32:
+		case IR::ValueOpcode::BufferAtomicOr32:
+		case IR::ValueOpcode::BufferAtomicOr64:
+		case IR::ValueOpcode::SharedAtomicOr64:
+		case IR::ValueOpcode::SharedAtomicOr32: return spv::OpAtomicOr;
+		case IR::ValueOpcode::ImageAtomicXor32:
+		case IR::ValueOpcode::BufferAtomicXor32:
+		case IR::ValueOpcode::SharedAtomicXor32: return spv::OpAtomicXor;
+		default: return spv::OpNop;
+	}
+}
+
+uint32_t EmitAtomicOperation(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t pointer,
+                             uint32_t scope) {
+	const auto opcode = SpirvAtomicOpcode(inst.GetOpcode());
+	const auto old    = ctx.state.builder.AllocateId();
+	if (opcode == spv::OpAtomicCompareExchange) {
+		const auto desired    = ctx.Arg(inst, inst.NumArgs() - 3);
+		const auto comparator = ctx.Arg(inst, inst.NumArgs() - 2);
+		ctx.state.builder.AddFunction(
+		    spv::OpAtomicCompareExchange, TypeU32(ctx.state), old, pointer,
+		    ConstantU32(ctx.state, scope), ConstantU32(ctx.state, spv::MemorySemanticsMaskNone),
+		    ConstantU32(ctx.state, spv::MemorySemanticsMaskNone), desired, comparator);
+	} else {
+		const auto value = ctx.Arg(inst, inst.NumArgs() - 2);
+		ctx.state.builder.AddFunction(opcode, TypeU32(ctx.state), old, pointer,
+		                              ConstantU32(ctx.state, scope),
+		                              ConstantU32(ctx.state, spv::MemorySemanticsMaskNone), value);
+	}
+	return old;
+}
+
+void EmitAtomicMemoryBarrier(EmitterState& state, IR::ResourceKind kind) {
+	const auto scope  = kind == IR::ResourceKind::Lds ? spv::ScopeWorkgroup : spv::ScopeDevice;
+	const auto memory = [&] {
+		switch (kind) {
+			case IR::ResourceKind::Lds: return spv::MemorySemanticsWorkgroupMemoryMask;
+			case IR::ResourceKind::Image: return spv::MemorySemanticsImageMemoryMask;
+			default: return spv::MemorySemanticsUniformMemoryMask;
+		}
+	}();
+	state.builder.AddFunction(spv::OpMemoryBarrier, ConstantU32(state, scope),
+	                          ConstantU32(state, spv::MemorySemanticsAcquireReleaseMask | memory));
 }
 
 uint32_t EmitFloatAtomicReplacement(EmitterState& state, uint32_t old, uint32_t source,

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

@@ -6,19 +6,32 @@
 namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter {
 
 uint32_t TypeVoid(EmitterState& state) {
-	return state.builder.Type(spv::OpTypeVoid);
+	if (state.void_type == 0) {
+		state.void_type = state.builder.Type(spv::OpTypeVoid);
+	}
+	return state.void_type;
 }
 
 uint32_t TypeBool(EmitterState& state) {
-	return state.builder.Type(spv::OpTypeBool);
+	if (state.bool_type == 0) {
+		state.bool_type = state.builder.Type(spv::OpTypeBool);
+	}
+	return state.bool_type;
 }
 
 uint32_t TypeBoolVector(EmitterState& state, uint32_t components) {
-	return state.builder.Type(spv::OpTypeVector, TypeBool(state), components);
+	auto& type = state.bool_vector_types[components - 2u];
+	if (type == 0) {
+		type = state.builder.Type(spv::OpTypeVector, TypeBool(state), components);
+	}
+	return type;
 }
 
 uint32_t TypeU32(EmitterState& state) {
-	return state.builder.Type(spv::OpTypeInt, 32, 0);
+	if (state.u32_type == 0) {
+		state.u32_type = state.builder.Type(spv::OpTypeInt, 32, 0);
+	}
+	return state.u32_type;
 }
 
 uint32_t TypeU64(EmitterState& state) {
@@ -26,33 +39,55 @@ uint32_t TypeU64(EmitterState& state) {
 }
 
 uint32_t TypeScalarU64(EmitterState& state) {
-	return state.builder.Type(spv::OpTypeInt, 64, 0);
+	if (state.native_u64_type == 0) {
+		state.native_u64_type = state.builder.Type(spv::OpTypeInt, 64, 0);
+	}
+	return state.native_u64_type;
 }
 
 uint32_t TypeU32Pair(EmitterState& state) {
-	const auto element = TypeU32(state);
-	return state.builder.Type(spv::OpTypeStruct, element, element);
+	if (state.u32_pair_type == 0) {
+		const auto element = TypeU32(state);
+		state.u32_pair_type = state.builder.Type(spv::OpTypeStruct, element, element);
+	}
+	return state.u32_pair_type;
 }
 
 uint32_t TypeI32(EmitterState& state) {
-	return state.builder.Type(spv::OpTypeInt, 32, 1);
+	if (state.i32_type == 0) {
+		state.i32_type = state.builder.Type(spv::OpTypeInt, 32, 1);
+	}
+	return state.i32_type;
 }
 
 uint32_t TypeI32Pair(EmitterState& state) {
-	const auto element = TypeI32(state);
-	return state.builder.Type(spv::OpTypeStruct, element, element);
+	if (state.i32_pair_type == 0) {
+		const auto element = TypeI32(state);
+		state.i32_pair_type = state.builder.Type(spv::OpTypeStruct, element, element);
+	}
+	return state.i32_pair_type;
 }
 
 uint32_t TypeF32(EmitterState& state) {
-	return state.builder.Type(spv::OpTypeFloat, 32);
+	if (state.f32_type == 0) {
+		state.f32_type = state.builder.Type(spv::OpTypeFloat, 32);
+	}
+	return state.f32_type;
 }
 
 uint32_t TypeF64(EmitterState& state) {
-	return state.builder.Type(spv::OpTypeFloat, 64);
+	if (state.f64_type == 0) {
+		state.f64_type = state.builder.Type(spv::OpTypeFloat, 64);
+	}
+	return state.f64_type;
 }
 
 uint32_t TypeU32Vector(EmitterState& state, uint32_t components) {
-	return state.builder.Type(spv::OpTypeVector, TypeU32(state), components);
+	auto& type = state.u32_vector_types[components - 2u];
+	if (type == 0) {
+		type = state.builder.Type(spv::OpTypeVector, TypeU32(state), components);
+	}
+	return type;
 }
 
 uint32_t TypeU32Composite(EmitterState& state, uint32_t components) {
@@ -61,11 +96,19 @@ uint32_t TypeU32Composite(EmitterState& state, uint32_t components) {
 }
 
 uint32_t TypeI32Vector(EmitterState& state, uint32_t components) {
-	return state.builder.Type(spv::OpTypeVector, TypeI32(state), components);
+	auto& type = state.i32_vector_types[components - 2u];
+	if (type == 0) {
+		type = state.builder.Type(spv::OpTypeVector, TypeI32(state), components);
+	}
+	return type;
 }
 
 uint32_t TypeF32Vector(EmitterState& state, uint32_t components) {
-	return state.builder.Type(spv::OpTypeVector, TypeF32(state), components);
+	auto& type = state.f32_vector_types[components - 2u];
+	if (type == 0) {
+		type = state.builder.Type(spv::OpTypeVector, TypeF32(state), components);
+	}
+	return type;
 }
 
 uint32_t TypePointer(EmitterState& state, spv::StorageClass storage_class, uint32_t pointee) {
@@ -73,7 +116,10 @@ uint32_t TypePointer(EmitterState& state, spv::StorageClass storage_class, uint3
 }
 
 uint32_t TypeFunction(EmitterState& state) {
-	return state.builder.Type(spv::OpTypeFunction, TypeVoid(state));
+	if (state.function_type == 0) {
+		state.function_type = state.builder.Type(spv::OpTypeFunction, TypeVoid(state));
+	}
+	return state.function_type;
 }
 
 uint32_t StorageRuntimeArrayType(EmitterState& state) {
@@ -626,10 +672,15 @@ void DefineModule(EmitterState& state) {
 		state.builder.RequireCapability(spv::CapabilityPhysicalStorageBufferAddresses);
 		state.builder.RequireExtension("SPV_KHR_physical_storage_buffer");
 	}
-	if (state.requirements.buffer_int64_atomics) {
+	if (state.requirements.buffer_int64_atomics || state.requirements.shared_int64_atomics) {
 		state.builder.RequireCapability(spv::CapabilityInt64);
 		state.builder.RequireCapability(spv::CapabilityInt64Atomics);
 	}
+	if (state.requirements.shared_int64_atomics) {
+		state.builder.RequireVersion(0x00010400u);
+		state.builder.RequireExtension("SPV_KHR_workgroup_memory_explicit_layout");
+		state.builder.RequireCapability(spv::CapabilityWorkgroupMemoryExplicitLayoutKHR);
+	}
 	if (state.clip_distance_variable != 0) {
 		state.builder.RequireCapability(spv::CapabilityClipDistance);
 	}

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

@@ -243,6 +243,12 @@ void Invoke(Return (*emit)(Context&, Args...), ValueEmitContext& ctx, const IR::
 }
 
 void EmitDirectInstruction(ValueEmitContext& ctx, const IR::Inst& inst) {
+	if (ctx.half != 0 && (inst.GetOpcode() == IR::ValueOpcode::Ballot ||
+	                     inst.GetOpcode() == IR::ValueOpcode::ReadFirstLane)) {
+		// Both operations already combine both emulated halves into one whole-wave result.
+		ctx.Define(inst, ctx.other_half->Result(inst));
+		return;
+	}
 	switch (inst.GetOpcode()) {
 #define VALUE_OPCODE(name, ...)                                                                    \
 	case IR::ValueOpcode::name: return Invoke(Emit##name, ctx, inst);
@@ -679,6 +685,7 @@ void EmitProgram(EmitterState& state) {
 		}
 	}
 	DefineGetBdaPointer(state);
+	DefineBvhIntersect(state);
 	for (const auto* block: program.blocks) {
 		if (std::ranges::any_of(*block, [](const IR::Inst& inst) {
 			    return inst.GetOpcode() == IR::ValueOpcode::SwizzleU32 ||

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

@@ -192,8 +192,11 @@ constexpr MimgGatherInfo MIMG_GATHER_OPCODE_LIST[] = {
 
 constexpr Detail::OpcodeMap MIMG_ATOMIC_OPCODE_LIST[] = {
     {0x0fu, Opcode::IMAGE_ATOMIC_SWAP},
+    {0x10u, Opcode::IMAGE_ATOMIC_CMPSWAP},
     {0x11u, Opcode::IMAGE_ATOMIC_ADD},
+    {0x14u, Opcode::IMAGE_ATOMIC_SMIN},
     {0x15u, Opcode::IMAGE_ATOMIC_UMIN},
+    {0x16u, Opcode::IMAGE_ATOMIC_SMAX},
     {0x17u, Opcode::IMAGE_ATOMIC_UMAX},
     {0x18u, Opcode::IMAGE_ATOMIC_AND},
     {0x19u, Opcode::IMAGE_ATOMIC_OR},
@@ -225,6 +228,7 @@ Opcode DecodeMimgOpcode(uint32_t opcode, const MimgSampleInfo* sample, const Mim
 		case 0x09u: return Opcode::IMAGE_STORE_MIP;
 		case 0x0eu: return Opcode::IMAGE_GET_RESINFO;
 		case 0x60u: return Opcode::IMAGE_GET_LOD;
+		case 0xe6u: return Opcode::IMAGE_BVH_INTERSECT_RAY;
 		default: return Opcode::UNSUPPORTED;
 	}
 }
@@ -253,6 +257,7 @@ uint32_t DecodeMimgAddressComponents(uint32_t opcode, ImageDimension dimension,
 	}
 
 	switch (opcode) {
+		case 0xe6u: return 11u;
 		case 0x0eu: return 1u;
 		case 0x01u:
 		case 0x09u: return ImageCoordComponents(dimension) + 1u;
@@ -355,11 +360,20 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 		SetUnsupported(inst, Family::MIMG, opcode,
 		               "MIMG image gather requires exactly one dmask bit");
 	}
+	if (inst.opcode == Opcode::IMAGE_ATOMIC_CMPSWAP && inst.dmask != 0x3u) {
+		SetUnsupported(inst, Family::MIMG, opcode,
+		               "MIMG image compare-and-swap requires 32-bit DMASK 0x3");
+	}
 	const bool supports_d16 = sample != nullptr || gather != nullptr || opcode == 0x00u ||
 	                          opcode == 0x01u || opcode == 0x08u || opcode == 0x09u;
 	if (d16 && !supports_d16) {
 		SetUnsupported(inst, Family::MIMG, opcode, "MIMG opcode does not support D16 data");
 	}
+	if (opcode == 0xe6u &&
+	    (a16 || !r128 || inst.dmask != 0xfu || (nsa_dwords != 0u && nsa_dwords != 3u))) {
+		SetUnsupported(inst, Family::MIMG, opcode,
+		               "BVH intersection requires eleven full-float ray DWORDs and R128/dmask:0xf");
+	}
 
 	DecodeVectorGpr(vdata, inst.dst);
 	DecodeVectorGpr(vaddr, inst.src0);

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

@@ -61,6 +61,7 @@ constexpr MemoryOpcodeInfo MUBUF_OPCODE_LIST[] = {
     {0x3fu, Opcode::BUFFER_ATOMIC_FMIN, 1, 32},
     {0x40u, Opcode::BUFFER_ATOMIC_FMAX, 1, 32},
     {0x50u, Opcode::BUFFER_ATOMIC_SWAP_X2, 2, 32},
+    {0x59u, Opcode::BUFFER_ATOMIC_AND_X2, 2, 32},
     {0x5au, Opcode::BUFFER_ATOMIC_OR_X2, 2, 32},
 };
 
@@ -106,6 +107,7 @@ constexpr MemoryOpcodeInfo DS_OPCODE_LIST[] = {
     {0x39u, Opcode::DS_READ_I8, 1, 8, true},     {0x3au, Opcode::DS_READ_U8, 1, 8},
     {0x3bu, Opcode::DS_READ_I16, 1, 16, true},   {0x3cu, Opcode::DS_READ_U16, 1, 16},
     {0x3du, Opcode::DS_CONSUME, 1, 32},          {0x3eu, Opcode::DS_APPEND, 1, 32},
+    {0x40u, Opcode::DS_ADD_U64, 2, 32},         {0x4au, Opcode::DS_OR_B64, 2, 32},
     {0x4du, Opcode::DS_WRITE_B64, 2, 32},        {0x4eu, Opcode::DS_WRITE2_B64, 4, 32},
     {0x4fu, Opcode::DS_WRITE2ST64_B64, 4, 32},   {0x76u, Opcode::DS_READ_B64, 2, 32},
     {0x77u, Opcode::DS_READ2_B64, 4, 32},        {0x78u, Opcode::DS_READ2ST64_B64, 4, 32},
@@ -167,6 +169,7 @@ bool IsDsWriteOpcode(Opcode opcode) {
 bool IsDsAtomicOpcode(Opcode opcode) {
 	switch (opcode) {
 		case Opcode::DS_ADD_U32:
+		case Opcode::DS_ADD_U64:
 		case Opcode::DS_ADD_RTN_U32:
 		case Opcode::DS_SUB_U32:
 		case Opcode::DS_SUB_RTN_U32:
@@ -185,6 +188,7 @@ bool IsDsAtomicOpcode(Opcode opcode) {
 		case Opcode::DS_AND_B32:
 		case Opcode::DS_AND_RTN_B32:
 		case Opcode::DS_OR_B32:
+		case Opcode::DS_OR_B64:
 		case Opcode::DS_OR_RTN_B32:
 		case Opcode::DS_XOR_B32:
 		case Opcode::DS_XOR_RTN_B32:

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

@@ -46,6 +46,7 @@ constexpr OpcodeMap SOP1_OPCODE_LIST[] = {
     {0x09u, Opcode::S_WQM_B32},
     {0x0au, Opcode::S_WQM_B64},
     {0x0bu, Opcode::S_BREV_B32},
+    {0x0cu, Opcode::S_BREV_B64},
     {0x0fu, Opcode::S_BCNT1_I32_B32},
     {0x10u, Opcode::S_BCNT1_I32_B64},
     {0x13u, Opcode::S_FF1_I32_B32},
@@ -108,6 +109,7 @@ constexpr OpcodeMap SOPP_OPCODE_LIST[] = {
     {0x12u, Opcode::S_TRAP},
     {0x16u, Opcode::S_TTRACEDATA},
     {0x17u, Opcode::S_CBRANCH_CDBGSYS},
+    {0x19u, Opcode::S_CBRANCH_CDBGSYS_OR_USER},
     {0x20u, Opcode::S_INST_PREFETCH},
     {0x23u, Opcode::S_WAITCNT_DEPCTR},
 };

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

@@ -184,8 +184,9 @@ std::string FormatExp(const Instruction& inst) {
 
 bool IsConditionalBranch(Opcode opcode) {
 	switch (opcode) {
-		// DevKit NGG should be disabled.
-		case Opcode::S_CBRANCH_CDBGSYS: return false;
+		// Conditional shader debugging is disabled.
+		case Opcode::S_CBRANCH_CDBGSYS:
+		case Opcode::S_CBRANCH_CDBGSYS_OR_USER: return false;
 		case Opcode::S_CBRANCH_SCC0:
 		case Opcode::S_CBRANCH_SCC1:
 		case Opcode::S_CBRANCH_VCCZ:
@@ -253,6 +254,8 @@ void DecodeScalarSource(uint32_t code, uint32_t pc, Operand& operand) {
 		case 125u: operand.kind = OperandKind::Null; return;
 		case 126u: operand.kind = OperandKind::ExecLo; return;
 		case 127u: operand.kind = OperandKind::ExecHi; return;
+		case 235u: operand.kind = OperandKind::SharedBase; return;
+		case 237u: operand.kind = OperandKind::PrivateBase; return;
 		case 239u: operand.kind = OperandKind::PopsExitingWaveId; return;
 		case 248u:
 			operand.kind      = OperandKind::FloatInlineConstant;
@@ -403,7 +406,6 @@ void DecodeProgram(std::span<const uint32_t> code, Program& program) {
 	program.instructions.clear();
 	program.instructions.reserve(code.size());
 	program.code = code;
-	program.has_bvh = false;
 
 	std::vector<bool> branch_targets;
 	for (uint32_t word_index = 0; word_index < code.size();) {
@@ -412,10 +414,6 @@ void DecodeProgram(std::span<const uint32_t> code, Program& program) {
 
 		const auto& inst = program.instructions.back();
 		word_index += inst.word_count;
-		if (inst.family == Family::MIMG && (inst.opcode_id == 0xe6u || inst.opcode_id == 0xe7u)) {
-			program.has_bvh = true;
-			return;
-		}
 
 		if (IsDirectBranch(inst.opcode)) {
 			const auto target_index = inst.branch_target / sizeof(uint32_t);
@@ -454,6 +452,8 @@ std::string OperandToString(const Operand& operand) {
 		case OperandKind::Scc: text = "scc"; break;
 		case OperandKind::M0: text = "m0"; break;
 		case OperandKind::PopsExitingWaveId: text = "pops_exiting_wave_id"; break;
+		case OperandKind::SharedBase: text = "shared_base"; break;
+		case OperandKind::PrivateBase: text = "private_base"; break;
 		case OperandKind::Null: text = "null"; break;
 		default: text = "unknown"; break;
 	}
@@ -504,6 +504,7 @@ std::string InstructionToString(const Instruction& inst) {
 			                                               OperandToString(inst.src0).c_str()));
 		case Opcode::S_ABS_I32:
 		case Opcode::S_BREV_B32:
+		case Opcode::S_BREV_B64:
 		case Opcode::S_BCNT1_I32_B32:
 		case Opcode::S_FLBIT_I32_B32:
 		case Opcode::S_FF1_I32_B32:
@@ -560,6 +561,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::S_CBRANCH_EXECZ:
 		case Opcode::S_CBRANCH_EXECNZ:
 		case Opcode::S_CBRANCH_CDBGSYS:
+		case Opcode::S_CBRANCH_CDBGSYS_OR_USER:
 			return WithUnsupportedReason(inst, fmt::format("0x{:08x}: {} 0x{:08x}", inst.pc,
 			                                               magic_enum::enum_name(inst.opcode),
 			                                               inst.branch_target));
@@ -572,9 +574,12 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::IMAGE_SAMPLE:
 		case Opcode::IMAGE_STORE:
 		case Opcode::IMAGE_STORE_MIP:
+		case Opcode::IMAGE_ATOMIC_CMPSWAP:
 		case Opcode::IMAGE_ATOMIC_SWAP:
 		case Opcode::IMAGE_ATOMIC_ADD:
+		case Opcode::IMAGE_ATOMIC_SMIN:
 		case Opcode::IMAGE_ATOMIC_UMIN:
+		case Opcode::IMAGE_ATOMIC_SMAX:
 		case Opcode::IMAGE_ATOMIC_UMAX:
 		case Opcode::IMAGE_ATOMIC_AND:
 		case Opcode::IMAGE_ATOMIC_OR:
@@ -641,6 +646,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::BUFFER_ATOMIC_SMAX:
 		case Opcode::BUFFER_ATOMIC_UMAX:
 		case Opcode::BUFFER_ATOMIC_AND:
+		case Opcode::BUFFER_ATOMIC_AND_X2:
 		case Opcode::BUFFER_ATOMIC_OR:
 		case Opcode::BUFFER_ATOMIC_OR_X2:
 		case Opcode::BUFFER_ATOMIC_XOR:
@@ -663,6 +669,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::FLAT_STORE_DWORDX3:
 		case Opcode::FLAT_STORE_DWORDX4:
 		case Opcode::DS_ADD_U32:
+		case Opcode::DS_ADD_U64:
 		case Opcode::DS_ADD_RTN_U32:
 		case Opcode::DS_SUB_U32:
 		case Opcode::DS_SUB_RTN_U32:
@@ -681,6 +688,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::DS_AND_B32:
 		case Opcode::DS_AND_RTN_B32:
 		case Opcode::DS_OR_B32:
+		case Opcode::DS_OR_B64:
 		case Opcode::DS_OR_RTN_B32:
 		case Opcode::DS_XOR_B32:
 		case Opcode::DS_XOR_RTN_B32:

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

@@ -10,6 +10,10 @@
 
 namespace Libs::Graphics::ShaderRecompiler::Decoder {
 
+// PS5 flat stack addresses use these SH_MEM_BASES aperture tags in VA[63:32].
+constexpr uint32_t PrivateApertureHigh = 0x70000000u;
+constexpr uint32_t SharedApertureHigh = 0x80000000u;
+
 enum class Family {
 	Unknown,
 	SOP1,
@@ -43,6 +47,7 @@ enum class Opcode {
 	S_ABS_I32,
 	S_ABSDIFF_I32,
 	S_BREV_B32,
+	S_BREV_B64,
 	S_BCNT1_I32_B32,
 	S_BCNT1_I32_B64,
 	S_FF1_I32_B32,
@@ -240,6 +245,7 @@ enum class Opcode {
 	V_MAD_U32_U24,
 	V_MAD_U64_U32,
 	V_FMA_F64,
+	V_ADD_F64,
 	V_MUL_F64,
 	V_FMA_F32,
 	V_FMA_F16,
@@ -355,6 +361,10 @@ enum class Opcode {
 	V_CMP_NEQ_F32,
 	V_CMP_NLT_F32,
 	V_CMP_TRU_F32,
+	V_CMP_EQ_F64,
+	V_CMP_LE_F64,
+	V_CMPX_LE_F64,
+	V_CMPX_GE_F64,
 	V_CMPX_LT_F32,
 	V_CMPX_EQ_F32,
 	V_CMPX_LE_F32,
@@ -384,12 +394,19 @@ enum class Opcode {
 	V_CMP_GT_I16,
 	V_CMP_NE_I16,
 	V_CMP_GE_I16,
+	V_CMPX_LT_I16,
+	V_CMPX_EQ_I16,
+	V_CMPX_LE_I16,
+	V_CMPX_GT_I16,
+	V_CMPX_NE_I16,
+	V_CMPX_GE_I16,
 	V_CMP_LT_F16,
 	V_CMP_EQ_F16,
 	V_CMP_LE_F16,
 	V_CMP_GT_F16,
 	V_CMP_LG_F16,
 	V_CMP_GE_F16,
+	V_CMP_NGE_F16,
 	V_CMP_NGT_F16,
 	V_CMP_NEQ_F16,
 	V_CMP_NLT_F16,
@@ -399,6 +416,7 @@ enum class Opcode {
 	V_CMPX_GT_F16,
 	V_CMPX_GE_F16,
 	V_CMPX_NGT_F16,
+	V_CMPX_NLE_F16,
 	V_CMPX_NEQ_F16,
 	V_CMPX_NLT_F16,
 	V_CMPX_LT_I32,
@@ -425,11 +443,16 @@ enum class Opcode {
 	V_CMP_GE_U32,
 	V_CMP_T_U32,
 	V_CMP_EQ_I64,
+	V_CMP_LT_I64,
+	V_CMP_LE_I64,
 	V_CMP_LT_U64,
 	V_CMP_EQ_U64,
+	V_CMP_LE_U64,
 	V_CMP_GT_U64,
 	V_CMP_NE_U64,
+	V_CMP_GE_U64,
 	V_CMPX_NE_I64,
+	V_CMPX_LE_U64,
 	V_CMPX_NE_U64,
 	V_CMPX_LT_U32,
 	V_CMPX_EQ_U32,
@@ -489,6 +512,7 @@ enum class Opcode {
 	BUFFER_ATOMIC_SMAX,
 	BUFFER_ATOMIC_UMAX,
 	BUFFER_ATOMIC_AND,
+	BUFFER_ATOMIC_AND_X2,
 	BUFFER_ATOMIC_OR,
 	BUFFER_ATOMIC_OR_X2,
 	BUFFER_ATOMIC_XOR,
@@ -509,6 +533,7 @@ enum class Opcode {
 	FLAT_STORE_DWORDX3,
 	FLAT_STORE_DWORDX4,
 	DS_ADD_U32,
+	DS_ADD_U64,
 	DS_ADD_RTN_U32,
 	DS_SUB_U32,
 	DS_SUB_RTN_U32,
@@ -527,6 +552,7 @@ enum class Opcode {
 	DS_AND_B32,
 	DS_AND_RTN_B32,
 	DS_OR_B32,
+	DS_OR_B64,
 	DS_OR_RTN_B32,
 	DS_XOR_B32,
 	DS_XOR_RTN_B32,
@@ -565,6 +591,7 @@ enum class Opcode {
 	DS_WRITE_B128,
 	DS_WRITE_ADDTID_B32,
 	DS_READ_ADDTID_B32,
+	IMAGE_BVH_INTERSECT_RAY,
 	IMAGE_GET_RESINFO,
 	IMAGE_GET_LOD,
 	IMAGE_LOAD,
@@ -572,8 +599,11 @@ enum class Opcode {
 	IMAGE_STORE,
 	IMAGE_STORE_MIP,
 	IMAGE_ATOMIC_SWAP,
+	IMAGE_ATOMIC_CMPSWAP,
 	IMAGE_ATOMIC_ADD,
+	IMAGE_ATOMIC_SMIN,
 	IMAGE_ATOMIC_UMIN,
+	IMAGE_ATOMIC_SMAX,
 	IMAGE_ATOMIC_UMAX,
 	IMAGE_ATOMIC_AND,
 	IMAGE_ATOMIC_OR,
@@ -605,6 +635,7 @@ enum class Opcode {
 	S_CBRANCH_EXECZ,
 	S_CBRANCH_EXECNZ,
 	S_CBRANCH_CDBGSYS,
+	S_CBRANCH_CDBGSYS_OR_USER,
 	S_SENDMSG,
 	S_SETREG_B32,
 	S_SLEEP,
@@ -631,6 +662,8 @@ enum class OperandKind {
 	Scc,
 	M0,
 	PopsExitingWaveId,
+	SharedBase,
+	PrivateBase,
 	Null,
 	Vgpr,
 };
@@ -742,7 +775,6 @@ struct Instruction {
 struct Program {
 	std::span<const uint32_t> code;
 	std::vector<Instruction>  instructions;
-	bool                     has_bvh = false;
 };
 
 // Code spans are trusted to contain complete instructions, valid branch targets, and 32-bit PCs.

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

@@ -228,6 +228,8 @@ constexpr VopcOpcodeInfo VOPC_OPCODE_LIST[] = {
     {0x19u, Opcode::V_CMPX_NGE_F32},       {0x1au, Opcode::V_CMPX_NLG_F32},
     {0x1bu, Opcode::V_CMPX_NGT_F32},       {0x1cu, Opcode::V_CMPX_NLE_F32},
     {0x1du, Opcode::V_CMPX_NEQ_F32},       {0x1eu, Opcode::V_CMPX_NLT_F32},
+    {0x22u, Opcode::V_CMP_EQ_F64, false},  {0x23u, Opcode::V_CMP_LE_F64, false},
+    {0x33u, Opcode::V_CMPX_LE_F64, false}, {0x36u, Opcode::V_CMPX_GE_F64, false},
     {0x80u, Opcode::V_CMP_F_I32},          {0x81u, Opcode::V_CMP_LT_I32},
     {0x82u, Opcode::V_CMP_EQ_I32},         {0x83u, Opcode::V_CMP_LE_I32},
     {0x84u, Opcode::V_CMP_GT_I32},         {0x85u, Opcode::V_CMP_NE_I32},
@@ -239,6 +241,9 @@ 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}, {0x9au, Opcode::V_CMPX_EQ_I16, false},
+    {0x9bu, Opcode::V_CMPX_LE_I16, false}, {0x9cu, Opcode::V_CMPX_GT_I16, false},
+    {0x9du, Opcode::V_CMPX_NE_I16, false}, {0x9eu, Opcode::V_CMPX_GE_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},
@@ -248,21 +253,25 @@ constexpr VopcOpcodeInfo VOPC_OPCODE_LIST[] = {
     {0xc3u, Opcode::V_CMP_LE_U32},         {0xc4u, Opcode::V_CMP_GT_U32},
     {0xc5u, Opcode::V_CMP_NE_U32},         {0xc6u, Opcode::V_CMP_GE_U32},
     {0xc7u, Opcode::V_CMP_T_U32},          {0xa2u, Opcode::V_CMP_EQ_I64, false},
+    {0xa1u, Opcode::V_CMP_LT_I64, false},  {0xa3u, Opcode::V_CMP_LE_I64, false},
     {0xb5u, Opcode::V_CMPX_NE_I64, false}, {0xd1u, Opcode::V_CMPX_LT_U32},
     {0xd2u, Opcode::V_CMPX_EQ_U32},        {0xd3u, Opcode::V_CMPX_LE_U32},
     {0xd4u, Opcode::V_CMPX_GT_U32},        {0xd5u, Opcode::V_CMPX_NE_U32},
     {0xd6u, Opcode::V_CMPX_GE_U32},        {0xe1u, Opcode::V_CMP_LT_U64, false},
-    {0xe2u, Opcode::V_CMP_EQ_U64, false},  {0xe4u, Opcode::V_CMP_GT_U64, false},
-    {0xe5u, Opcode::V_CMP_NE_U64, false},
+    {0xe2u, Opcode::V_CMP_EQ_U64, false},  {0xe3u, Opcode::V_CMP_LE_U64, false},
+    {0xe4u, Opcode::V_CMP_GT_U64, false},  {0xe5u, Opcode::V_CMP_NE_U64, false},
+    {0xe6u, Opcode::V_CMP_GE_U64, false},  {0xf3u, Opcode::V_CMPX_LE_U64, false},
     {0xf5u, Opcode::V_CMPX_NE_U64, false}, {0xc9u, Opcode::V_CMP_LT_F16},
     {0xcau, Opcode::V_CMP_EQ_F16},         {0xcbu, Opcode::V_CMP_LE_F16},
     {0xccu, Opcode::V_CMP_GT_F16},         {0xcdu, Opcode::V_CMP_LG_F16},
-    {0xceu, Opcode::V_CMP_GE_F16},         {0xebu, Opcode::V_CMP_NGT_F16},
+    {0xceu, Opcode::V_CMP_GE_F16},         {0xe9u, Opcode::V_CMP_NGE_F16},
+    {0xebu, Opcode::V_CMP_NGT_F16},
     {0xedu, Opcode::V_CMP_NEQ_F16},        {0xeeu, Opcode::V_CMP_NLT_F16},
     {0xd9u, Opcode::V_CMPX_LT_F16},        {0xdau, Opcode::V_CMPX_EQ_F16},
     {0xdbu, Opcode::V_CMPX_LE_F16},        {0xdcu, Opcode::V_CMPX_GT_F16},
     {0xdeu, Opcode::V_CMPX_GE_F16},        {0xfbu, Opcode::V_CMPX_NGT_F16},
-    {0xfdu, Opcode::V_CMPX_NEQ_F16},       {0xfeu, Opcode::V_CMPX_NLT_F16},
+    {0xfcu, Opcode::V_CMPX_NLE_F16},       {0xfdu, Opcode::V_CMPX_NEQ_F16},
+    {0xfeu, Opcode::V_CMPX_NLT_F16},
 };
 
 constexpr auto VOPC_OPS = Detail::MakeOpcodeTable<0x100>(VOPC_OPCODE_LIST);
@@ -278,6 +287,7 @@ constexpr OpcodeMap VOP3_OPCODE_LIST[] = {
     {0x147u, Opcode::V_CUBEMA_F32},
     {0x14bu, Opcode::V_FMA_F32},
     {0x14cu, Opcode::V_FMA_F64},
+    {0x164u, Opcode::V_ADD_F64},
     {0x165u, Opcode::V_MUL_F64},
     {0x148u, Opcode::V_BFE_U32},
     {0x149u, Opcode::V_BFE_I32},
@@ -827,6 +837,10 @@ bool IsVopcFloatCompareOpcode(Opcode opcode) {
 		case Opcode::V_CMP_F_F32:
 		case Opcode::V_CMP_LT_F32:
 		case Opcode::V_CMP_EQ_F32:
+		case Opcode::V_CMP_EQ_F64:
+		case Opcode::V_CMP_LE_F64:
+		case Opcode::V_CMPX_LE_F64:
+		case Opcode::V_CMPX_GE_F64:
 		case Opcode::V_CMP_LE_F32:
 		case Opcode::V_CMP_GT_F32:
 		case Opcode::V_CMP_LG_F32:
@@ -859,6 +873,7 @@ bool IsVopcFloatCompareOpcode(Opcode opcode) {
 		case Opcode::V_CMP_GT_F16:
 		case Opcode::V_CMP_LG_F16:
 		case Opcode::V_CMP_GE_F16:
+		case Opcode::V_CMP_NGE_F16:
 		case Opcode::V_CMP_NGT_F16:
 		case Opcode::V_CMP_NEQ_F16:
 		case Opcode::V_CMP_NLT_F16:
@@ -868,6 +883,7 @@ bool IsVopcFloatCompareOpcode(Opcode opcode) {
 		case Opcode::V_CMPX_GT_F16:
 		case Opcode::V_CMPX_GE_F16:
 		case Opcode::V_CMPX_NGT_F16:
+		case Opcode::V_CMPX_NLE_F16:
 		case Opcode::V_CMPX_NEQ_F16:
 		case Opcode::V_CMPX_NLT_F16:
 		case Opcode::V_CMP_CLASS_F32:
@@ -1136,7 +1152,9 @@ VopcSdwaFields DecodeVopcSdwaFields(uint32_t modifier) {
 }
 
 bool SupportsVopcSdwa(Opcode opcode) {
-	return opcode != Opcode::UNSUPPORTED;
+	return opcode != Opcode::UNSUPPORTED && opcode != Opcode::V_CMP_EQ_F64 &&
+	       opcode != Opcode::V_CMP_LE_F64 && opcode != Opcode::V_CMPX_LE_F64 &&
+	       opcode != Opcode::V_CMPX_GE_F64;
 }
 
 void DecodeVopcSdwa(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index,
@@ -1201,6 +1219,7 @@ void DecodeVopcDpp(uint32_t pc, std::span<const uint32_t> code, uint32_t word_in
 
 uint32_t NativeVop3SourceCount(Opcode opcode) {
 	switch (opcode) {
+		case Opcode::V_ADD_F64:
 		case Opcode::V_MUL_F64:
 		case Opcode::V_MUL_LO_U32:
 		case Opcode::V_MUL_HI_U32:
@@ -1352,6 +1371,7 @@ bool SupportsNativeVop3SourceModifiers(Opcode opcode) {
 		case Opcode::V_MAX_F32:
 		case Opcode::V_MAC_F32:
 		case Opcode::V_MAD_F32:
+		case Opcode::V_ADD_F64:
 		case Opcode::V_MUL_F64:
 		case Opcode::V_FMA_F64:
 		case Opcode::V_FMA_F32:
@@ -1457,6 +1477,8 @@ void ApplyNativeVop3SourceModifiers(Instruction& inst, uint32_t abs, uint32_t ne
 
 bool IsVopcCompareExec(Opcode opcode) {
 	switch (opcode) {
+		case Opcode::V_CMPX_LE_F64:
+		case Opcode::V_CMPX_GE_F64:
 		case Opcode::V_CMPX_LT_F32:
 		case Opcode::V_CMPX_EQ_F32:
 		case Opcode::V_CMPX_LE_F32:
@@ -1484,7 +1506,14 @@ bool IsVopcCompareExec(Opcode opcode) {
 		case Opcode::V_CMPX_NE_U32:
 		case Opcode::V_CMPX_GE_U32:
 		case Opcode::V_CMPX_NE_I64:
+		case Opcode::V_CMPX_LE_U64:
 		case Opcode::V_CMPX_NE_U64:
+		case Opcode::V_CMPX_LT_I16:
+		case Opcode::V_CMPX_EQ_I16:
+		case Opcode::V_CMPX_LE_I16:
+		case Opcode::V_CMPX_GT_I16:
+		case Opcode::V_CMPX_NE_I16:
+		case Opcode::V_CMPX_GE_I16:
 		case Opcode::V_CMPX_LT_U16:
 		case Opcode::V_CMPX_EQ_U16:
 		case Opcode::V_CMPX_GT_U16:
@@ -1494,6 +1523,7 @@ bool IsVopcCompareExec(Opcode opcode) {
 		case Opcode::V_CMPX_GT_F16:
 		case Opcode::V_CMPX_GE_F16:
 		case Opcode::V_CMPX_NGT_F16:
+		case Opcode::V_CMPX_NLE_F16:
 		case Opcode::V_CMPX_NEQ_F16:
 		case Opcode::V_CMPX_NLT_F16: return true;
 		default: return false;

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

@@ -1,5 +1,7 @@
 #include "graphics/shader/recompiler/frontend/translate/Translator.h"
 
+#include <bit>
+
 namespace Libs::Graphics::ShaderRecompiler::Frontend {
 
 void Translator::EmitCompareResult(const Decoder::Instruction& inst, IR::U1 value, bool scalar,
@@ -26,9 +28,20 @@ void Translator::EmitCompareConstant(const Decoder::Instruction& inst, bool valu
 
 void Translator::EmitIntegerCompare(const Decoder::Instruction& inst, IR::ValueOpcode opcode,
                                     IR::Type type, bool scalar, bool cmpx) {
-	const auto lhs = ReadOperand(inst.src0, type);
-	const auto rhs = ReadOperand(inst.src1, type);
-	EmitCompareResult(inst, IR::U1(ir.Emit(opcode, {lhs, rhs})), scalar, cmpx);
+	const bool signed_64 = opcode == IR::ValueOpcode::SLessThan64 ||
+	                       opcode == IR::ValueOpcode::SLessThanEqual64 ||
+	                       inst.opcode == Decoder::Opcode::V_CMP_EQ_I64 ||
+	                       inst.opcode == Decoder::Opcode::V_CMPX_NE_I64;
+	const auto read = [&](const Decoder::Operand& operand) {
+		// RDNA2 expands signed 64-bit integer literals by sign extension.
+		if (signed_64 && operand.kind == Decoder::OperandKind::LiteralConstant) {
+			return IR::Value(static_cast<uint64_t>(
+			    static_cast<int64_t>(std::bit_cast<int32_t>(operand.value))));
+		}
+		return ReadOperand(operand, type);
+	};
+	EmitCompareResult(inst, IR::U1(ir.Emit(opcode, {read(inst.src0), read(inst.src1)})), scalar,
+	                  cmpx);
 }
 
 void Translator::EmitInteger16Compare(const Decoder::Instruction& inst, IR::ValueOpcode opcode,
@@ -40,11 +53,13 @@ void Translator::EmitInteger16Compare(const Decoder::Instruction& inst, IR::Valu
 
 void Translator::EmitFloatCompare(const Decoder::Instruction& inst, IR::ValueOpcode opcode,
                                   bool half, bool cmpx) {
+	const auto type = IR::ArgTypeOf(opcode, 0);
 	const auto lhs =
-	    half ? IR::Value(ReadF16AsF32(inst.src0)) : ReadOperand(inst.src0, IR::Type::F32);
+	    half ? IR::Value(ReadF16AsF32(inst.src0)) : ReadOperand(inst.src0, type);
 	const auto rhs =
-	    half ? IR::Value(ReadF16AsF32(inst.src1)) : ReadOperand(inst.src1, IR::Type::F32);
-	const IR::FPCompareFlags flags{.flush_input_denorms = !half && flush_f32_inputs};
+	    half ? IR::Value(ReadF16AsF32(inst.src1)) : ReadOperand(inst.src1, type);
+	const IR::FPCompareFlags flags{
+	    .flush_input_denorms = !half && type == IR::Type::F32 && flush_f32_inputs};
 	EmitCompareResult(inst, IR::U1(ir.Emit(opcode, {lhs, rhs}, flags)), false, cmpx);
 }
 

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

@@ -558,9 +558,18 @@ void Translator::IMAGE_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode
 	const auto memory   = MemoryInfoFromDecoded(inst);
 	const auto resource = GetImageResource(memory);
 	const auto address  = MakeImageAddress(inst, MemorySourceAt(inst, 1));
-	const auto result =
-	    ir.Emit(opcode, {resource, address, ReadU32(MemorySourceAt(inst, 0)), ir.GetExec()},
-	            AddMemoryInfo(memory, inst.pc));
+	const auto data     = MemorySourceAt(inst, 0);
+	const auto flags    = AddMemoryInfo(memory, inst.pc);
+	IR::Value  result;
+	if (opcode == IR::ValueOpcode::ImageAtomicCompareSwap32) {
+		// VDATA supplies the replacement; VDATA+1 supplies the comparison.
+		result = ir.Emit(opcode,
+		                 {resource, address, ReadU32(data), ReadU32(OffsetOperand(data, 1u)),
+		                  ir.GetExec()},
+		                 flags);
+	} else {
+		result = ir.Emit(opcode, {resource, address, ReadU32(data), ir.GetExec()}, flags);
+	}
 	if (inst.glc) {
 		WriteOperand(inst.dst, result);
 	}
@@ -570,7 +579,10 @@ void Translator::DS_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode opc
                            bool returns_value) {
 	const auto memory  = MemoryInfoFromDecoded(inst);
 	const auto address = ReadU32(MemorySourceAt(inst, 1));
-	const auto result  = ir.Emit(opcode, {address, ReadU32(MemorySourceAt(inst, 0)), ir.GetExec()},
+	const auto data_src = MemorySourceAt(inst, 0);
+	const IR::Value value =
+	    memory.data_dwords == 2u ? IR::Value(ReadU64(data_src)) : IR::Value(ReadU32(data_src));
+	const auto result  = ir.Emit(opcode, {address, value, ir.GetExec()},
 	                             AddMemoryInfo(memory, inst.pc));
 	if (returns_value) {
 		WriteOperand(inst.dst, result);
@@ -880,6 +892,15 @@ void Translator::DS_BPERMUTE_B32(const Decoder::Instruction& inst) {
 
 void Translator::EmitMemory(const Decoder::Instruction& inst) {
 	switch (inst.opcode) {
+		case Decoder::Opcode::IMAGE_BVH_INTERSECT_RAY: {
+			const auto result = ir.Emit(IR::ValueOpcode::BvhIntersect,
+			    {ConstructU32x4(inst.src1, 4), MakeImageAddress(inst, inst.src0), ir.GetExec()});
+			for (uint32_t component = 0; component < 4; ++component) {
+				WriteOperand(OffsetOperand(inst.dst, component),
+				    ir.Emit(IR::ValueOpcode::CompositeExtractU32x4, {result, IR::Value(component)}));
+			}
+			return;
+		}
 		case Decoder::Opcode::S_LOAD_DWORD:
 		case Decoder::Opcode::S_LOAD_DWORDX2:
 		case Decoder::Opcode::S_LOAD_DWORDX4:
@@ -954,6 +975,8 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 			return BUFFER_ATOMIC(inst, IR::ValueOpcode::BufferAtomicUMax32);
 		case Decoder::Opcode::BUFFER_ATOMIC_AND:
 			return BUFFER_ATOMIC(inst, IR::ValueOpcode::BufferAtomicAnd32);
+		case Decoder::Opcode::BUFFER_ATOMIC_AND_X2:
+			return BUFFER_ATOMIC(inst, IR::ValueOpcode::BufferAtomicAnd64);
 		case Decoder::Opcode::BUFFER_ATOMIC_OR:
 			return BUFFER_ATOMIC(inst, IR::ValueOpcode::BufferAtomicOr32);
 		case Decoder::Opcode::BUFFER_ATOMIC_OR_X2:
@@ -967,6 +990,8 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 
 		case Decoder::Opcode::DS_ADD_U32:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicIAdd32, false);
+		case Decoder::Opcode::DS_ADD_U64:
+			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicIAdd64, false);
 		case Decoder::Opcode::DS_ADD_RTN_U32:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicIAdd32, true);
 		case Decoder::Opcode::DS_SUB_U32:
@@ -1003,6 +1028,8 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicAnd32, true);
 		case Decoder::Opcode::DS_OR_B32:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicOr32, false);
+		case Decoder::Opcode::DS_OR_B64:
+			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicOr64, false);
 		case Decoder::Opcode::DS_OR_RTN_B32:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicOr32, true);
 		case Decoder::Opcode::DS_XOR_B32:
@@ -1012,12 +1039,18 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::DS_WRXCHG_RTN_B32:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicSwap32, true);
 
+		case Decoder::Opcode::IMAGE_ATOMIC_CMPSWAP:
+			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicCompareSwap32);
 		case Decoder::Opcode::IMAGE_ATOMIC_SWAP:
 			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicSwap32);
 		case Decoder::Opcode::IMAGE_ATOMIC_ADD:
 			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicIAdd32);
+		case Decoder::Opcode::IMAGE_ATOMIC_SMIN:
+			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicSMin32);
 		case Decoder::Opcode::IMAGE_ATOMIC_UMIN:
 			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicUMin32);
+		case Decoder::Opcode::IMAGE_ATOMIC_SMAX:
+			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicSMax32);
 		case Decoder::Opcode::IMAGE_ATOMIC_UMAX:
 			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicUMax32);
 		case Decoder::Opcode::IMAGE_ATOMIC_AND:

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

@@ -155,6 +155,12 @@ void Translator::EmitScalar(const Decoder::Instruction& inst) {
 		case O::S_BREV_B32:
 			return SimpleInteger(inst, IR::ValueOpcode::BitReverse32, IR::Type::U32, false, false,
 			                     false);
+		case O::S_BREV_B64: {
+			const auto value = ReadU32Pair(inst.src0);
+			WriteU32Pair(inst.dst, {IR::U32(ir.Emit(IR::ValueOpcode::BitReverse32, {value[1]})),
+			                        IR::U32(ir.Emit(IR::ValueOpcode::BitReverse32, {value[0]}))});
+			return;
+		}
 		case O::S_BCNT1_I32_B32:
 			return SimpleInteger(inst, IR::ValueOpcode::BitCount32, IR::Type::U32, false, false,
 			                     true);
@@ -231,6 +237,7 @@ void Translator::EmitScalar(const Decoder::Instruction& inst) {
 		case O::S_CBRANCH_EXECZ:
 		case O::S_CBRANCH_EXECNZ:
 		case O::S_CBRANCH_CDBGSYS:
+		case O::S_CBRANCH_CDBGSYS_OR_USER:
 		case O::S_ENDPGM: return;
 		default: return FailMissingTranslation(inst);
 	}

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

@@ -135,6 +135,8 @@ IR::U32 Translator::ReadRawU32(const Decoder::Operand& operand) {
 		case Decoder::OperandKind::IntegerInlineConstant:
 		case Decoder::OperandKind::FloatInlineConstant: return IR::U32(IR::Value(operand.value));
 		case Decoder::OperandKind::Null:
+		case Decoder::OperandKind::SharedBase:
+		case Decoder::OperandKind::PrivateBase:
 		case Decoder::OperandKind::PopsExitingWaveId: return IR::U32(IR::Value(0u));
 		case Decoder::OperandKind::Sgpr:
 			return ir.GetScalarReg(static_cast<IR::ScalarReg>(operand.reg));
@@ -474,6 +476,12 @@ IR::U32 Translator::ReadU32(const Decoder::Operand& operand) {
 }
 
 std::array<IR::U32, 2> Translator::ReadU32Pair(const Decoder::Operand& operand) {
+	if (operand.kind == Decoder::OperandKind::PrivateBase ||
+	    operand.kind == Decoder::OperandKind::SharedBase) {
+		return {IR::U32(IR::Value(0u)), IR::U32(IR::Value(
+		    operand.kind == Decoder::OperandKind::PrivateBase ? Decoder::PrivateApertureHigh
+		                                                      : Decoder::SharedApertureHigh))};
+	}
 	if (operand.kind == Decoder::OperandKind::ExecLo) {
 		return {ir.GetExecLo(), ir.GetExecHi()};
 	}
@@ -560,6 +568,12 @@ 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 +581,7 @@ 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)}));
 	}
@@ -880,13 +894,35 @@ void IncludeInstructionVectorRegisters(const Decoder::Instruction& inst, uint32_
 		case Decoder::Opcode::V_CVT_F64_F32:
 		case Decoder::Opcode::V_CVT_F64_U32: include_vector(inst.dst, 2u); break;
 		case Decoder::Opcode::V_FMA_F64: include_vector(inst.src2, 2u); [[fallthrough]];
+		case Decoder::Opcode::V_ADD_F64:
 		case Decoder::Opcode::V_MUL_F64: include_vector(inst.src1, 2u); [[fallthrough]];
 		case Decoder::Opcode::V_RCP_F64: include_vector(inst.dst, 2u); [[fallthrough]];
 		case Decoder::Opcode::V_CVT_F32_F64: include_vector(inst.src0, 2u); break;
+		case Decoder::Opcode::V_CMP_EQ_F64:
+		case Decoder::Opcode::V_CMP_LE_F64:
+		case Decoder::Opcode::V_CMPX_LE_F64:
+		case Decoder::Opcode::V_CMPX_GE_F64:
+		case Decoder::Opcode::V_CMP_EQ_I64:
+		case Decoder::Opcode::V_CMP_LT_I64:
+		case Decoder::Opcode::V_CMP_LE_I64:
+		case Decoder::Opcode::V_CMP_LT_U64:
+		case Decoder::Opcode::V_CMP_EQ_U64:
+		case Decoder::Opcode::V_CMP_LE_U64:
+		case Decoder::Opcode::V_CMP_GT_U64:
+		case Decoder::Opcode::V_CMP_NE_U64:
+		case Decoder::Opcode::V_CMP_GE_U64:
+		case Decoder::Opcode::V_CMPX_NE_I64:
+		case Decoder::Opcode::V_CMPX_LE_U64:
+		case Decoder::Opcode::V_CMPX_NE_U64:
+			include_vector(inst.src0, 2u);
+			include_vector(inst.src1, 2u);
+			break;
 		default: break;
 	}
 	if (inst.family == Decoder::Family::DS) {
 		switch (inst.opcode) {
+			case Decoder::Opcode::DS_ADD_U64:
+			case Decoder::Opcode::DS_OR_B64:
 			case Decoder::Opcode::DS_WRITE_B64:
 			case Decoder::Opcode::DS_WRITE_B96:
 			case Decoder::Opcode::DS_WRITE_B128: include_vector(inst.src1, inst.data_dwords); break;
@@ -1120,12 +1156,13 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 				return entry_ir.ISub(lhs, minimum(lhs, rhs));
 			};
 			const auto local = builtin(IR::StageInputKind::LocalInvocationIndex);
-			const auto primitive_chunk = entry_ir.IMul(builtin(IR::StageInputKind::WorkgroupId, 0),
-			                                           u32(mesh.primitives_per_group));
+			const auto group = builtin(IR::StageInputKind::WorkgroupId, 0);
+			const auto primitive_chunk = mesh.fast_launch ? group :
+			    entry_ir.IMul(group, u32(mesh.primitives_per_group));
 			const auto step  = u32(mesh.InputPrimitiveStep());
 			const auto size  = u32(mesh.InputPrimitiveSize());
-			const auto chunk = entry_ir.IMul(primitive_chunk, step);
-			const auto vertices =
+			const auto chunk = mesh.fast_launch ? group : entry_ir.IMul(primitive_chunk, step);
+			const auto vertices = mesh.fast_launch ? u32(mesh.vertices_per_group) :
 			    minimum(subtract_saturate(draw(0), chunk), u32(mesh.vertices_per_group));
 			const auto primitives = entry_ir.Select(
 			    entry_ir.ULessThan(vertices, size), u32(0),
@@ -1146,56 +1183,64 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 			    entry_ir.BitwiseOr(wave_info, entry_ir.BitwiseOr(entry_ir.ShiftLeftLogical(
 			                                                         primitive_count, u32(8)),
 			                                                     vertex_count)));
-			// GS adjacency addresses local ES records in LDS. Fans retain the draw's
-			// center in every subgroup; strip winding follows the global primitive.
-			const auto vertex = entry_ir.IMul(local, step);
-			auto       first  = vertex;
-			auto       second = u32(0);
-			auto       third  = u32(0);
-			if (mesh.InputPrimitiveSize() >= 2u) {
-				second = entry_ir.IAdd(vertex, u32(1));
+			if (mesh.fast_launch) {
+				// Fast launch broadcasts the subgroup's base vertex and instance to every lane.
+				entry_ir.SetVectorReg(static_cast<IR::VectorReg>(5), entry_ir.IAdd(draw(1), group));
+				entry_ir.SetVectorReg(
+				    static_cast<IR::VectorReg>(6),
+				    entry_ir.IAdd(draw(2), builtin(IR::StageInputKind::WorkgroupId, 1)));
+			} else {
+				// GS adjacency addresses local ES records in LDS. Fans retain the draw's
+				// center in every subgroup; strip winding follows the global primitive.
+				const auto vertex = entry_ir.IMul(local, step);
+				auto       first  = vertex;
+				auto       second = u32(0);
+				auto       third  = u32(0);
+				if (mesh.InputPrimitiveSize() >= 2u) {
+					second = entry_ir.IAdd(vertex, u32(1));
+				}
+				if (mesh.InputPrimitiveSize() == 3u) {
+					third = entry_ir.IAdd(vertex, u32(2));
+				}
+				auto input_vertex = entry_ir.IAdd(chunk, local);
+				if (mesh.input_primitive == static_cast<uint32_t>(Prospero::PrimitiveType::kTriFan)) {
+					first = u32(0);
+					input_vertex = entry_ir.Select(entry_ir.IEqual(local, u32(0)), u32(0), input_vertex);
+				} else if (mesh.input_primitive == static_cast<uint32_t>(Prospero::PrimitiveType::kTriStrip)) {
+					const auto parity = entry_ir.BitwiseAnd(entry_ir.IAdd(primitive_chunk, local), u32(1));
+					first = entry_ir.IAdd(first, parity);
+					second = entry_ir.ISub(second, parity);
+				}
+				entry_ir.SetVectorReg(static_cast<IR::VectorReg>(0),
+				                      entry_ir.BitwiseOr(entry_ir.ShiftLeftLogical(first, u32(2)),
+				                                         entry_ir.ShiftLeftLogical(second, u32(18))));
+				entry_ir.SetVectorReg(static_cast<IR::VectorReg>(1),
+				                      entry_ir.ShiftLeftLogical(third, u32(2)));
+				const auto index_bytes  = draw(3);
+				const auto indexed      = entry_ir.INotEqual(index_bytes, u32(0));
+				const auto index_low    = draw(4);
+				const auto byte_offset  = entry_ir.IAdd(entry_ir.BitwiseAnd(index_low, u32(3)),
+				                                           entry_ir.IMul(input_vertex, index_bytes));
+				const auto index_resource = entry_ir.Emit(
+				    IR::ValueOpcode::GetAddressResource,
+				    {entry_ir.BitwiseAnd(index_low, u32(~3u)), draw(5)});
+				const auto memory_index = static_cast<uint32_t>(result.memory_info.size());
+				result.memory_info.push_back({.kind = IR::ResourceKind::Global});
+				const auto packed_index = entry_ir.Emit(
+				    IR::ValueOpcode::LoadAddressU32,
+				    {index_resource, entry_ir.BitwiseAnd(byte_offset, u32(~3u)), u32(0),
+				     entry_ir.LogicalAnd(indexed, entry_ir.ULessThan(local, vertices))},
+				    IR::MemoryFlags {.index = memory_index});
+				const auto index = IR::U32(entry_ir.Emit(
+				    IR::ValueOpcode::BitFieldUExtract,
+				    {packed_index, entry_ir.IMul(entry_ir.BitwiseAnd(byte_offset, u32(3)), u32(8)),
+				     entry_ir.IMul(index_bytes, u32(8))}));
+				entry_ir.SetVectorReg(static_cast<IR::VectorReg>(5),
+				                      entry_ir.IAdd(draw(1), entry_ir.Select(indexed, index, input_vertex)));
+				entry_ir.SetVectorReg(
+				    static_cast<IR::VectorReg>(8),
+				    entry_ir.IAdd(draw(2), builtin(IR::StageInputKind::WorkgroupId, 1)));
 			}
-			if (mesh.InputPrimitiveSize() == 3u) {
-				third = entry_ir.IAdd(vertex, u32(2));
-			}
-			auto input_vertex = entry_ir.IAdd(chunk, local);
-			if (mesh.input_primitive == static_cast<uint32_t>(Prospero::PrimitiveType::kTriFan)) {
-				first = u32(0);
-				input_vertex = entry_ir.Select(entry_ir.IEqual(local, u32(0)), u32(0), input_vertex);
-			} else if (mesh.input_primitive == static_cast<uint32_t>(Prospero::PrimitiveType::kTriStrip)) {
-				const auto parity = entry_ir.BitwiseAnd(entry_ir.IAdd(primitive_chunk, local), u32(1));
-				first = entry_ir.IAdd(first, parity);
-				second = entry_ir.ISub(second, parity);
-			}
-			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(0),
-			                      entry_ir.BitwiseOr(entry_ir.ShiftLeftLogical(first, u32(2)),
-			                                         entry_ir.ShiftLeftLogical(second, u32(18))));
-			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(1),
-			                      entry_ir.ShiftLeftLogical(third, u32(2)));
-			const auto index_bytes  = draw(3);
-			const auto indexed      = entry_ir.INotEqual(index_bytes, u32(0));
-			const auto index_low    = draw(4);
-			const auto byte_offset  = entry_ir.IAdd(entry_ir.BitwiseAnd(index_low, u32(3)),
-			                                           entry_ir.IMul(input_vertex, index_bytes));
-			const auto index_resource = entry_ir.Emit(
-			    IR::ValueOpcode::GetAddressResource,
-			    {entry_ir.BitwiseAnd(index_low, u32(~3u)), draw(5)});
-			const auto memory_index = static_cast<uint32_t>(result.memory_info.size());
-			result.memory_info.push_back({.kind = IR::ResourceKind::Global});
-			const auto packed_index = entry_ir.Emit(
-			    IR::ValueOpcode::LoadAddressU32,
-			    {index_resource, entry_ir.BitwiseAnd(byte_offset, u32(~3u)), u32(0),
-			     entry_ir.LogicalAnd(indexed, entry_ir.ULessThan(local, vertices))},
-			    IR::MemoryFlags {.index = memory_index});
-			const auto index = IR::U32(entry_ir.Emit(
-			    IR::ValueOpcode::BitFieldUExtract,
-			    {packed_index, entry_ir.IMul(entry_ir.BitwiseAnd(byte_offset, u32(3)), u32(8)),
-			     entry_ir.IMul(index_bytes, u32(8))}));
-			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(5),
-			                      entry_ir.IAdd(draw(1), entry_ir.Select(indexed, index, input_vertex)));
-			entry_ir.SetVectorReg(
-			    static_cast<IR::VectorReg>(8),
-			    entry_ir.IAdd(draw(2), builtin(IR::StageInputKind::WorkgroupId, 1)));
 		} else if (options.stage == ShaderType::Local) {
 			entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(3), IR::U32(IR::Value(64u)));
 			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(2),

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

@@ -114,18 +114,40 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 			EmitIntegerCompare(inst, IR::ValueOpcode::INotEqual64, IR::Type::U64, false, true);
 			return;
 
+		case O::V_CMP_LT_I64:
+			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThan64, IR::Type::U64, false, false);
+			return;
+		case O::V_CMP_LE_I64:
+			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThanEqual64, IR::Type::U64, false, false);
+			return;
+		case O::V_CMP_LE_U64:
+			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThanEqual64, IR::Type::U64, false, false);
+			return;
+		case O::V_CMPX_LE_U64:
+			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThanEqual64, IR::Type::U64, false, true);
+			return;
+		case O::V_CMP_GE_U64:
+			EmitIntegerCompare(inst, IR::ValueOpcode::UGreaterThanEqual64, IR::Type::U64, false, false);
+			return;
+
 		case O::V_CMP_EQ_U16:
 			EmitInteger16Compare(inst, IR::ValueOpcode::IEqual32, false, false);
 			return;
 		case O::V_CMP_EQ_I16:
 			EmitInteger16Compare(inst, IR::ValueOpcode::IEqual32, true, false);
 			return;
+		case O::V_CMPX_EQ_I16:
+			EmitInteger16Compare(inst, IR::ValueOpcode::IEqual32, true, true);
+			return;
 		case O::V_CMP_NE_U16:
 			EmitInteger16Compare(inst, IR::ValueOpcode::INotEqual32, false, false);
 			return;
 		case O::V_CMP_NE_I16:
 			EmitInteger16Compare(inst, IR::ValueOpcode::INotEqual32, true, false);
 			return;
+		case O::V_CMPX_NE_I16:
+			EmitInteger16Compare(inst, IR::ValueOpcode::INotEqual32, true, true);
+			return;
 		case O::V_CMP_GT_U16:
 			EmitInteger16Compare(inst, IR::ValueOpcode::UGreaterThan32, false, false);
 			return;
@@ -150,21 +172,45 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_CMP_GT_I16:
 			EmitInteger16Compare(inst, IR::ValueOpcode::SGreaterThan32, true, false);
 			return;
+		case O::V_CMPX_GT_I16:
+			EmitInteger16Compare(inst, IR::ValueOpcode::SGreaterThan32, true, true);
+			return;
 		case O::V_CMP_GE_I16:
 			EmitInteger16Compare(inst, IR::ValueOpcode::SGreaterThanEqual32, true, false);
 			return;
+		case O::V_CMPX_GE_I16:
+			EmitInteger16Compare(inst, IR::ValueOpcode::SGreaterThanEqual32, true, true);
+			return;
 		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;
+		case O::V_CMPX_LE_I16:
+			EmitInteger16Compare(inst, IR::ValueOpcode::SLessThanEqual32, true, true);
+			return;
 
 		case O::V_CMP_F_F32: EmitCompareConstant(inst, false, false, false); return;
 		case O::V_CMP_TRU_F32: EmitCompareConstant(inst, true, false, false); return;
 		case O::V_CMP_EQ_F32:
 			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdEqual32, false, false);
 			return;
+		case O::V_CMP_EQ_F64:
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdEqual64, false, false);
+			return;
+		case O::V_CMP_LE_F64:
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdLessThanEqual64, false, false);
+			return;
+		case O::V_CMPX_LE_F64:
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdLessThanEqual64, false, true);
+			return;
+		case O::V_CMPX_GE_F64:
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdGreaterThanEqual64, false, true);
+			return;
 		case O::V_CMPX_EQ_F32:
 			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdEqual32, false, true);
 			return;
@@ -267,12 +313,18 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_CMPX_LE_F16:
 			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdLessThanEqual32, true, true);
 			return;
+		case O::V_CMP_NGE_F16:
+			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordLessThan32, true, false);
+			return;
 		case O::V_CMP_NGT_F16:
 			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordLessThanEqual32, true, false);
 			return;
 		case O::V_CMPX_NGT_F16:
 			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordLessThanEqual32, true, true);
 			return;
+		case O::V_CMPX_NLE_F16:
+			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordGreaterThan32, true, true);
+			return;
 		case O::V_CMP_NEQ_F16:
 			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordNotEqual32, true, false);
 			return;
@@ -415,6 +467,7 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_SUB_F32: return FloatBinary(inst, IR::ValueOpcode::FPSub32, false);
 		case O::V_SUBREV_F32: return FloatBinary(inst, IR::ValueOpcode::FPSub32, true);
 		case O::V_MUL_F32: return FloatBinary(inst, IR::ValueOpcode::FPMul32, false);
+		case O::V_ADD_F64: return FloatBinary(inst, IR::ValueOpcode::FPAdd64, false);
 		case O::V_MUL_F64: return FloatBinary(inst, IR::ValueOpcode::FPMul64, false);
 		case O::V_MIN_F32: return FloatBinary(inst, IR::ValueOpcode::FPMin32, false);
 		case O::V_MAX_F32: return FloatBinary(inst, IR::ValueOpcode::FPMax32, false);

src/graphics/shader/recompiler/ir/Program.cpp

@@ -142,6 +142,7 @@ CompiledShaderInfo Program::TakeCompiledInfo() && {
 	    .user_data_base  = user_data_base,
 	    .user_data_count = user_data_count,
 	    .scratch_dwords  = scratch_dwords,
+	    .has_address_writes = has_address_writes,
 	    .info            = std::move(info),
 	    .bindings        = std::move(bindings),
 	};

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

@@ -3,6 +3,7 @@
 
 #include <array>
 #include <cstdint>
+#include <utility>
 #include <vector>
 
 namespace Libs::Graphics::ShaderRecompiler::IR {
@@ -34,6 +35,7 @@ struct ResourceSnapshot {
 	std::vector<DescriptorValue> samplers;
 	std::vector<uint32_t>        flattened_srt;
 	std::vector<uint32_t>        user_data;
+	std::vector<std::pair<uint64_t, uint64_t>> specialization_reads;
 	UniformFill                 uniform_fill;
 };
 

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

@@ -30,6 +30,7 @@ enum class ResourceKind {
 	Buffer,
 	IndirectBuffer,
 	Flat,
+	FlatLocal,
 	Global,
 	Scratch,
 	Lds,
@@ -40,7 +41,8 @@ enum class ResourceKind {
 
 [[nodiscard]] constexpr bool IsAddressResourceKind(ResourceKind kind) {
 	return kind == ResourceKind::ScalarAddress || kind == ResourceKind::Flat ||
-	       kind == ResourceKind::Global || kind == ResourceKind::Scratch;
+	       kind == ResourceKind::FlatLocal || kind == ResourceKind::Global ||
+	       kind == ResourceKind::Scratch;
 }
 
 struct MemoryInfo {
@@ -497,6 +499,7 @@ struct ResourceBlock {
 	Value                 condition;
 	std::vector<uint32_t> successors;
 	std::vector<uint32_t> sources;
+	std::vector<uint32_t> srt_reads;
 };
 
 // Stable shader metadata consumed by the renderer after native IR has been discarded.
@@ -508,6 +511,7 @@ struct CompiledShaderInfo {
 	uint32_t                      user_data_count     = 64;
 	uint32_t                      scratch_dwords      = 0;
 	uint32_t                      param_export_mask   = 0;
+	bool                          has_address_writes  = false;
 	ShaderInfo                    info;
 	BindingLayout                 bindings;
 };
@@ -562,7 +566,6 @@ struct ResourcePlan {
 	mutable std::vector<uint8_t>            visited_blocks;
 	mutable std::vector<uint32_t>           pending_blocks;
 	mutable std::vector<uint32_t>           material_keys;
-	mutable std::vector<std::pair<uint64_t, uint64_t>> specialization_reads;
 };
 
 struct Program: ResourcePlan {

src/graphics/shader/recompiler/ir/Value.cpp

@@ -2,7 +2,7 @@
 
 #include <algorithm>
 #include <cstring>
-#include <limits>
+#include <memory>
 
 namespace Libs::Graphics::ShaderRecompiler::IR {
 
@@ -135,10 +135,16 @@ bool Value::operator==(const Value& other) const {
 }
 
 Inst::Inst(ValueOpcode value_opcode, uint64_t value_flags)
-    : opcode(value_opcode), flags(value_flags) {
-	const auto count = NumArgsOf(opcode);
-	if (count != std::numeric_limits<size_t>::max()) {
-		args.resize(count);
+    : opcode(value_opcode),
+      num_args(value_opcode == ValueOpcode::Phi ? PhiArity
+                                               : static_cast<uint8_t>(NumArgsOf(value_opcode))),
+      flags(value_flags) {
+	if (num_args == PhiArity) {
+		std::destroy_at(&fixed_args);
+		std::construct_at(&phi_args);
+	} else if (num_args > InlineArity) {
+		std::destroy_at(&fixed_args);
+		std::construct_at(&large_args, num_args);
 	}
 }
 
@@ -154,8 +160,8 @@ Type Inst::GetType() const {
 	if (opcode == ValueOpcode::Phi) {
 		return static_cast<Type>(flags);
 	}
-	if (opcode == ValueOpcode::Identity && !args.empty()) {
-		return args.front().GetType();
+	if (opcode == ValueOpcode::Identity && num_args != 0) {
+		return Arg(0).GetType();
 	}
 	return TypeOf(opcode);
 }
@@ -173,21 +179,24 @@ size_t Inst::UseCount() const {
 }
 
 size_t Inst::NumArgs() const {
-	return args.size();
+	return num_args == PhiArity ? phi_args.size() : num_args;
 }
 
 size_t Inst::NumPhiBlocks() const {
-	return phi_blocks.size();
+	return num_args == PhiArity ? phi_args.size() : 0;
 }
 
 Value Inst::Arg(size_t index) const {
-	EXIT_IF(index >= args.size());
-	return args[index];
+	EXIT_IF(index >= NumArgs());
+	if (num_args <= InlineArity) {
+		return fixed_args[index];
+	}
+	return num_args == PhiArity ? phi_args[index].second : large_args[index];
 }
 
 Block* Inst::PhiBlock(size_t index) const {
-	EXIT_IF(opcode != ValueOpcode::Phi || index >= phi_blocks.size());
-	return phi_blocks[index];
+	EXIT_IF(opcode != ValueOpcode::Phi || index >= phi_args.size());
+	return phi_args[index].first;
 }
 
 Block* Inst::Parent() const {
@@ -203,15 +212,17 @@ void Inst::SetParent(Block* block) {
 }
 
 void Inst::SetArg(size_t index, Value value) {
-	if (index >= args.size()) {
-		EXIT_IF(NumArgsOf(opcode) != std::numeric_limits<size_t>::max());
-		args.resize(index + 1);
-	}
-	const auto old = args[index];
+	const auto old = Arg(index);
 	if (auto* old_inst = old.TryInstruction(); old_inst != nullptr) {
 		RemoveUse(old_inst, index);
 	}
-	args[index] = value;
+	if (num_args <= InlineArity) {
+		fixed_args[index] = value;
+	} else if (num_args == PhiArity) {
+		phi_args[index].second = value;
+	} else {
+		large_args[index] = value;
+	}
 	if (auto* new_inst = value.TryInstruction(); new_inst != nullptr) {
 		AddUse(new_inst, index);
 	}
@@ -219,9 +230,8 @@ void Inst::SetArg(size_t index, Value value) {
 
 void Inst::AddPhiOperand(Block* predecessor, Value value) {
 	EXIT_IF(opcode != ValueOpcode::Phi);
-	const auto index = args.size();
-	args.push_back(value);
-	phi_blocks.push_back(predecessor);
+	const auto index = phi_args.size();
+	phi_args.emplace_back(predecessor, value);
 	if (auto* value_inst = value.TryInstruction(); value_inst != nullptr) {
 		AddUse(value_inst, index);
 	}
@@ -234,21 +244,12 @@ void Inst::ReplaceUsesWith(Value replacement, bool preserve) {
 	}
 	Invalidate();
 	if (preserve) {
-		ReplaceOpcode(ValueOpcode::Identity);
-		args.resize(1);
+		opcode = ValueOpcode::Identity;
+		num_args = 1;
 		SetArg(0, replacement);
 	}
 }
 
-void Inst::ReplaceOpcode(ValueOpcode value_opcode) {
-	opcode           = value_opcode;
-	const auto count = NumArgsOf(opcode);
-	if (count != std::numeric_limits<size_t>::max()) {
-		EXIT_IF(!args.empty() && args.size() != count);
-		args.resize(count);
-	}
-}
-
 void Inst::Invalidate() {
 	ClearArgs();
 	opcode = ValueOpcode::Void;
@@ -269,13 +270,21 @@ void Inst::RemoveUse(Inst* used, size_t operand) {
 }
 
 void Inst::ClearArgs() {
-	for (size_t index = 0; index < args.size(); index++) {
-		if (auto* value_inst = args[index].TryInstruction(); value_inst != nullptr) {
+	for (size_t index = 0; index < NumArgs(); index++) {
+		if (auto* value_inst = Arg(index).TryInstruction(); value_inst != nullptr) {
 			RemoveUse(value_inst, index);
 		}
 	}
-	args.clear();
-	phi_blocks.clear();
+	if (num_args == PhiArity) {
+		std::destroy_at(&phi_args);
+		std::construct_at(&fixed_args);
+	} else if (num_args > InlineArity) {
+		std::destroy_at(&large_args);
+		std::construct_at(&fixed_args);
+	} else {
+		fixed_args.fill(Value {});
+	}
+	num_args = 0;
 }
 
 } // namespace Libs::Graphics::ShaderRecompiler::IR

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

@@ -4,6 +4,7 @@
 #include "graphics/shader/recompiler/ir/Reg.h"
 #include "graphics/shader/recompiler/ir/opcodes/ValueOpcodes.h"
 
+#include <array>
 #include <bit>
 #include <cstdint>
 #include <cstring>
@@ -133,7 +134,6 @@ public:
 	void SetArg(size_t index, Value value);
 	void AddPhiOperand(Block* predecessor, Value value);
 	void ReplaceUsesWith(Value replacement, bool preserve = true);
-	void ReplaceOpcode(ValueOpcode opcode);
 	void Invalidate();
 
 	template <typename T>
@@ -156,13 +156,22 @@ private:
 	void RemoveUse(Inst* used, size_t operand);
 	void ClearArgs();
 
+	static constexpr uint8_t InlineArity = 4;
+	static constexpr uint8_t PhiArity = UINT8_MAX;
+
 	ValueOpcode         opcode;
+	uint8_t             num_args;
+	mutable uint32_t    evaluation_index = UINT32_MAX;
 	uint64_t            flags;
 	Block*              parent = nullptr;
-	std::vector<Value>  args;
-	std::vector<Block*> phi_blocks;
+	union {
+		std::array<Value, InlineArity> fixed_args {};
+		std::vector<Value> large_args;
+		std::vector<std::pair<Block*, Value>> phi_args;
+	};
 	std::vector<Use>    uses;
-	mutable uint32_t    evaluation_index = UINT32_MAX;
 };
 
+static_assert(sizeof(Inst) <= 112, "Inst operand storage unintentionally increased");
+
 } // namespace Libs::Graphics::ShaderRecompiler::IR

src/graphics/shader/recompiler/ir/opcodes/ValueOpcodes.cpp

@@ -123,6 +123,7 @@ BufferAccess BufferAccessOf(ValueOpcode opcode) {
 		case ValueOpcode::BufferAtomicSMax32:
 		case ValueOpcode::BufferAtomicUMax32:
 		case ValueOpcode::BufferAtomicAnd32:
+		case ValueOpcode::BufferAtomicAnd64:
 		case ValueOpcode::BufferAtomicOr32:
 		case ValueOpcode::BufferAtomicOr64:
 		case ValueOpcode::BufferAtomicXor32:
@@ -135,6 +136,7 @@ BufferAccess BufferAccessOf(ValueOpcode opcode) {
 uint32_t BufferComponentCount(ValueOpcode opcode) {
 	switch (opcode) {
 		case ValueOpcode::BufferAtomicSwap64:
+		case ValueOpcode::BufferAtomicAnd64:
 		case ValueOpcode::BufferAtomicOr64:
 		case ValueOpcode::LoadBufferU32x2:
 		case ValueOpcode::StoreBufferU32x2: return 2u;
@@ -164,6 +166,7 @@ SharedAccess SharedAccessOf(ValueOpcode opcode) {
 		case ValueOpcode::SharedAtomicFMax32:
 		case ValueOpcode::SharedAtomicSwap32:
 		case ValueOpcode::SharedAtomicIAdd32:
+		case ValueOpcode::SharedAtomicIAdd64:
 		case ValueOpcode::SharedAtomicISub32:
 		case ValueOpcode::SharedAtomicInc32:
 		case ValueOpcode::SharedAtomicDec32:
@@ -173,6 +176,7 @@ SharedAccess SharedAccessOf(ValueOpcode opcode) {
 		case ValueOpcode::SharedAtomicUMax32:
 		case ValueOpcode::SharedAtomicAnd32:
 		case ValueOpcode::SharedAtomicOr32:
+		case ValueOpcode::SharedAtomicOr64:
 		case ValueOpcode::SharedAtomicXor32: return SharedAccess::Atomic;
 		case ValueOpcode::DataAppend: return SharedAccess::Append;
 		case ValueOpcode::DataConsume: return SharedAccess::Consume;
@@ -182,6 +186,8 @@ SharedAccess SharedAccessOf(ValueOpcode opcode) {
 
 uint32_t SharedComponentCount(ValueOpcode opcode) {
 	switch (opcode) {
+		case ValueOpcode::SharedAtomicIAdd64:
+		case ValueOpcode::SharedAtomicOr64:
 		case ValueOpcode::LoadSharedU32x2:
 		case ValueOpcode::WriteSharedU32x2: return 2u;
 		case ValueOpcode::LoadSharedU32x3:
@@ -214,9 +220,12 @@ ImageOpcodeInfo ImageOpcodeInfoOf(ValueOpcode opcode) {
 			return {ImageAccess::Read, ImageResourceClass::Sampled, true};
 		case ValueOpcode::ImageWrite:
 			return {ImageAccess::Write, ImageResourceClass::Storage, false};
+		case ValueOpcode::ImageAtomicCompareSwap32:
 		case ValueOpcode::ImageAtomicSwap32:
 		case ValueOpcode::ImageAtomicIAdd32:
+		case ValueOpcode::ImageAtomicSMin32:
 		case ValueOpcode::ImageAtomicUMin32:
+		case ValueOpcode::ImageAtomicSMax32:
 		case ValueOpcode::ImageAtomicUMax32:
 		case ValueOpcode::ImageAtomicAnd32:
 		case ValueOpcode::ImageAtomicOr32:

src/graphics/shader/recompiler/ir/opcodes/ValueOpcodes.h

@@ -9,7 +9,7 @@
 
 namespace Libs::Graphics::ShaderRecompiler::IR {
 
-enum class ValueOpcode {
+enum class ValueOpcode : uint16_t {
 #define VALUE_OPCODE(name, ...) name,
 #include "graphics/shader/recompiler/ir/opcodes/ValueOpcodes.inc"
 #undef VALUE_OPCODE

src/graphics/shader/recompiler/ir/opcodes/ValueOpcodes.inc

@@ -137,6 +137,9 @@ VALUE_OPCODE(SMedTri32, U32, U32, U32, U32)
 VALUE_OPCODE(UMedTri32, U32, U32, U32, U32)
 VALUE_OPCODE(SLessThan32, U1, U32, U32)
 VALUE_OPCODE(SLessThan64, U1, U64, U64)
+VALUE_OPCODE(SLessThanEqual64, U1, U64, U64)
+VALUE_OPCODE(ULessThanEqual64, U1, U64, U64)
+VALUE_OPCODE(UGreaterThanEqual64, U1, U64, U64)
 VALUE_OPCODE(ULessThan32, U1, U32, U32)
 VALUE_OPCODE(ULessThan64, U1, U64, U64)
 VALUE_OPCODE(IEqual32, U1, U32, U32)
@@ -155,6 +158,9 @@ VALUE_OPCODE(LogicalAnd, U1, U1, U1)
 VALUE_OPCODE(LogicalXor, U1, U1, U1)
 VALUE_OPCODE(LogicalNot, U1, U1)
 VALUE_OPCODE(FPOrdEqual32, U1, F32, F32)
+VALUE_OPCODE(FPOrdEqual64, U1, F64, F64)
+VALUE_OPCODE(FPOrdLessThanEqual64, U1, F64, F64)
+VALUE_OPCODE(FPOrdGreaterThanEqual64, U1, F64, F64)
 VALUE_OPCODE(FPUnordEqual32, U1, F32, F32)
 VALUE_OPCODE(FPOrdNotEqual32, U1, F32, F32)
 VALUE_OPCODE(FPUnordNotEqual32, U1, F32, F32)
@@ -173,6 +179,7 @@ VALUE_OPCODE(FPAdd32, F32, F32, F32)
 VALUE_OPCODE(FPSub32, F32, F32, F32)
 VALUE_OPCODE(FPFma32, F32, F32, F32, F32)
 VALUE_OPCODE(FPMul32, F32, F32, F32)
+VALUE_OPCODE(FPAdd64, F64, F64, F64)
 VALUE_OPCODE(FPMul64, F64, F64, F64)
 VALUE_OPCODE(FPFma64, F64, F64, F64, F64)
 VALUE_OPCODE(FPRecip64, F64, F64)
@@ -245,6 +252,7 @@ VALUE_OPCODE(BufferAtomicUMin32, U32, BufferResource, U32, U32, U32, U32, U1)
 VALUE_OPCODE(BufferAtomicSMax32, U32, BufferResource, U32, U32, U32, U32, U1)
 VALUE_OPCODE(BufferAtomicUMax32, U32, BufferResource, U32, U32, U32, U32, U1)
 VALUE_OPCODE(BufferAtomicAnd32, U32, BufferResource, U32, U32, U32, U32, U1)
+VALUE_OPCODE(BufferAtomicAnd64, U64, BufferResource, U32, U32, U32, U64, U1)
 VALUE_OPCODE(BufferAtomicOr32, U32, BufferResource, U32, U32, U32, U32, U1)
 VALUE_OPCODE(BufferAtomicOr64, U64, BufferResource, U32, U32, U32, U64, U1)
 VALUE_OPCODE(BufferAtomicXor32, U32, BufferResource, U32, U32, U32, U32, U1)
@@ -267,6 +275,7 @@ VALUE_OPCODE(SharedAtomicFMin32, Void, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicFMax32, Void, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicSwap32, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicIAdd32, U32, U32, U32, U1)
+VALUE_OPCODE(SharedAtomicIAdd64, Void, U32, U64, U1)
 VALUE_OPCODE(SharedAtomicISub32, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicInc32, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicDec32, U32, U32, U32, U1)
@@ -276,20 +285,25 @@ VALUE_OPCODE(SharedAtomicSMax32, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicUMax32, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicAnd32, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicOr32, U32, U32, U32, U1)
+VALUE_OPCODE(SharedAtomicOr64, Void, U32, U64, U1)
 VALUE_OPCODE(SharedAtomicXor32, U32, U32, U32, U1)
 VALUE_OPCODE(DataAppend, U32, U32, U1)
 VALUE_OPCODE(DataConsume, U32, U32, U1)
 VALUE_OPCODE(SwizzleU32, U32, U32, U32, U1)
 
 VALUE_OPCODE(ImageQueryDimensions, U32x4, ImageResource, ImageAddress)
+VALUE_OPCODE(BvhIntersect, U32x4, U32x4, ImageAddress, U1)
 VALUE_OPCODE(ImageQueryLod, U32x4, ImageResource, SamplerResource, ImageAddress)
 VALUE_OPCODE(ImageRead, U32x4, ImageResource, ImageAddress, U1)
 VALUE_OPCODE(ImageWrite, Void, ImageResource, ImageAddress, U32x4, U1)
 VALUE_OPCODE(ImageSampleRaw, U32x4, ImageResource, SamplerResource, ImageAddress)
 VALUE_OPCODE(ImageGatherRaw, U32x4, ImageResource, SamplerResource, ImageAddress)
+VALUE_OPCODE(ImageAtomicCompareSwap32, U32, ImageResource, ImageAddress, U32, U32, U1)
 VALUE_OPCODE(ImageAtomicSwap32, U32, ImageResource, ImageAddress, U32, U1)
 VALUE_OPCODE(ImageAtomicIAdd32, U32, ImageResource, ImageAddress, U32, U1)
+VALUE_OPCODE(ImageAtomicSMin32, U32, ImageResource, ImageAddress, U32, U1)
 VALUE_OPCODE(ImageAtomicUMin32, U32, ImageResource, ImageAddress, U32, U1)
+VALUE_OPCODE(ImageAtomicSMax32, U32, ImageResource, ImageAddress, U32, U1)
 VALUE_OPCODE(ImageAtomicUMax32, U32, ImageResource, ImageAddress, U32, U1)
 VALUE_OPCODE(ImageAtomicAnd32, U32, ImageResource, ImageAddress, U32, U1)
 VALUE_OPCODE(ImageAtomicOr32, U32, ImageResource, ImageAddress, U32, U1)

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

@@ -4,6 +4,7 @@
 #include <algorithm>
 #include <bit>
 #include <cstdint>
+#include <unordered_map>
 #include <unordered_set>
 
 namespace Libs::Graphics::ShaderRecompiler::IR {
@@ -575,6 +576,17 @@ void FoldInstruction(Block& block, Block::iterator instruction,
 		case ValueOpcode::UGreaterThan64:
 			FoldU64Compare(inst, [](uint64_t a, uint64_t b) { return a > b; });
 			return;
+		case ValueOpcode::ULessThanEqual64:
+			FoldU64Compare(inst, [](uint64_t a, uint64_t b) { return a <= b; });
+			return;
+		case ValueOpcode::UGreaterThanEqual64:
+			FoldU64Compare(inst, [](uint64_t a, uint64_t b) { return a >= b; });
+			return;
+		case ValueOpcode::SLessThanEqual64:
+			FoldU64Compare(inst, [](uint64_t a, uint64_t b) {
+				return std::bit_cast<int64_t>(a) <= std::bit_cast<int64_t>(b);
+			});
+			return;
 		case ValueOpcode::SLessThan64:
 			FoldU64Compare(inst, [](uint64_t a, uint64_t b) {
 				return std::bit_cast<int64_t>(a) < std::bit_cast<int64_t>(b);
@@ -649,12 +661,110 @@ void FoldInstruction(Block& block, Block::iterator instruction,
 	}
 }
 
+// In wave32, extracting this invocation's bit from a ballot recovers its predicate.
+// Keep that identity through scalar EXEC operations and their loop-carried mask Phis.
+class LaneMaskProjection {
+public:
+	explicit LaneMaskProjection(std::unordered_set<Inst*>& lowered_ancillary)
+	    : m_lowered_ancillary(lowered_ancillary) {}
+
+	void Fold(Inst& inst) {
+		if (inst.GetOpcode() != ValueOpcode::INotEqual32 || !Immediate(Arg(inst, 1), 0u)) return;
+		const auto* bit = Arg(inst, 0).TryInstruction();
+		if (bit == nullptr || bit->GetOpcode() != ValueOpcode::BitwiseAnd32 ||
+		    !Immediate(Arg(*bit, 1), 1u)) return;
+		const auto* shift = Arg(*bit, 0).TryInstruction();
+		if (shift == nullptr || shift->GetOpcode() != ValueOpcode::ShiftRightLogical32) return;
+		const auto* index = Arg(*shift, 1).TryInstruction();
+		if (index == nullptr || index->GetOpcode() != ValueOpcode::BitwiseAnd32 ||
+		    !Immediate(Arg(*index, 1), 31u)) return;
+		const auto* lane = Arg(*index, 0).TryInstruction();
+		if (lane == nullptr || lane->GetOpcode() != ValueOpcode::LaneId) return;
+		m_visited.clear();
+		m_grounded = false;
+		if (CanProject(Arg(*shift, 0)) && m_grounded) Replace(inst, Project(Arg(*shift, 0)));
+	}
+
+private:
+	static bool Immediate(Value value, uint32_t expected) {
+		return IsImmediate(value, Type::U32) && value.U32() == expected;
+	}
+
+	static Value BallotPredicate(const Inst& inst) {
+		if (inst.GetOpcode() != ValueOpcode::CompositeExtractU32x4 ||
+		    !Immediate(Arg(inst, 1), 0u)) return {};
+		const auto* source = Arg(inst, 0).TryInstruction();
+		return source != nullptr && source->GetOpcode() == ValueOpcode::Ballot
+		           ? Arg(*source, 0) : Value {};
+	}
+
+	bool CanProject(Value value) {
+		if (value.IsImmediate()) {
+			const bool supported = Immediate(value, 0u) || Immediate(value, UINT32_MAX);
+			m_grounded |= supported;
+			return supported;
+		}
+		const auto* inst = value.TryInstruction();
+		if (inst == nullptr) return false;
+		if (m_values.contains(inst) || !BallotPredicate(*inst).IsEmpty()) {
+			m_grounded = true;
+			return true;
+		}
+		if (!m_visited.insert(inst).second) return true;
+		const auto op = inst->GetOpcode();
+		if (op != ValueOpcode::BitwiseAnd32 && op != ValueOpcode::BitwiseOr32 &&
+		    op != ValueOpcode::BitwiseNot32 && op != ValueOpcode::SelectU32 && op != ValueOpcode::Phi)
+			return false;
+		for (size_t arg = op == ValueOpcode::SelectU32 ? 1u : 0u; arg < inst->NumArgs(); ++arg) {
+			if (!CanProject(Arg(*inst, arg))) return false;
+		}
+		return inst->NumArgs() != 0;
+	}
+
+	Value Project(Value value) {
+		if (value.IsImmediate()) return Value(value.U32() != 0u);
+		auto* source = value.TryInstruction();
+		if (const auto found = m_values.find(source); found != m_values.end()) return found->second.Resolve();
+		if (const auto predicate = BallotPredicate(*source); !predicate.IsEmpty()) return predicate;
+		auto* block = source->Parent();
+		const auto where = std::find_if(block->begin(), block->end(),
+		                               [source](const Inst& candidate) { return &candidate == source; });
+		const auto op = source->GetOpcode() == ValueOpcode::BitwiseAnd32 ? ValueOpcode::LogicalAnd
+		              : source->GetOpcode() == ValueOpcode::BitwiseOr32 ? ValueOpcode::LogicalOr
+		              : source->GetOpcode() == ValueOpcode::BitwiseNot32 ? ValueOpcode::LogicalNot
+		              : source->GetOpcode() == ValueOpcode::SelectU32 ? ValueOpcode::SelectU1
+		                                                               : ValueOpcode::Phi;
+		auto result = op == ValueOpcode::Phi ? block->PrependNewInst(where, op)
+		            : op == ValueOpcode::LogicalNot
+		                ? block->PrependNewInst(where, op, {Value(false)})
+		            : op == ValueOpcode::SelectU1
+		                ? block->PrependNewInst(where, op, {Arg(*source, 0), Value(false), Value(false)})
+		                : block->PrependNewInst(where, op, {Value(false), Value(false)});
+		m_values.emplace(source, Value(&*result));
+		if (op == ValueOpcode::Phi) result->SetFlags(Type::U1);
+		for (size_t arg = op == ValueOpcode::SelectU1 ? 1u : 0u; arg < source->NumArgs(); ++arg) {
+			const auto projected = Project(Arg(*source, arg));
+			if (op == ValueOpcode::Phi) result->AddPhiOperand(source->PhiBlock(arg), projected);
+			else result->SetArg(arg, projected);
+		}
+		FoldInstruction(*block, result, m_lowered_ancillary);
+		return Value(&*result).Resolve();
+	}
+
+	std::unordered_set<Inst*>& m_lowered_ancillary;
+	std::unordered_set<const Inst*> m_visited;
+	std::unordered_map<const Inst*, Value> m_values;
+	bool m_grounded = false;
+};
+
 } // namespace
 
-void ConstantPropagationPass(const BlockList& blocks) {
+void ConstantPropagationPass(const BlockList& blocks, uint32_t wave_size) {
 	std::unordered_set<Inst*> lowered_ancillary;
+	LaneMaskProjection mask_projection(lowered_ancillary);
 	for (auto* block: blocks) {
 		for (auto inst = block->begin(); inst != block->end(); ++inst) {
+			if (wave_size == 32u) mask_projection.Fold(*inst);
 			FoldInstruction(*block, inst, lowered_ancillary);
 		}
 	}

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

@@ -4,6 +4,6 @@
 
 namespace Libs::Graphics::ShaderRecompiler::IR {
 
-void ConstantPropagationPass(const BlockList& blocks);
+void ConstantPropagationPass(const BlockList& blocks, uint32_t wave_size = 64);
 
 } // namespace Libs::Graphics::ShaderRecompiler::IR

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

@@ -157,34 +157,15 @@ bool CaptureStrictRead(void* userdata, uint64_t address, std::span<uint32_t> val
 
 bool CaptureOrdinaryRead(void* userdata, uint64_t address, std::span<uint32_t> values) {
 	auto& capture = *static_cast<ReadCapture*>(userdata);
-	if (!capture.source.read_memory(capture.source.userdata, address, values)) {
-		return false;
+	if (capture.source.read_memory != nullptr) {
+		if (!capture.source.read_memory(capture.source.userdata, address, values)) return false;
+	} else {
+		std::memcpy(values.data(), reinterpret_cast<const void*>(address), values.size_bytes());
 	}
 	capture.ranges.emplace_back(address, values.size_bytes());
 	return true;
 }
 
-bool WrittenBuffersDisjoint(const ResourcePlan& program, const ResourceSnapshot& snapshot,
-                            std::span<const std::pair<uint64_t, uint64_t>> reads) {
-	for (uint32_t i = 0; i < program.info.buffers.size(); ++i) {
-		if (!program.info.buffers[i].written) continue;
-		ShaderBufferResource buffer;
-		if (!DecodeBufferDescriptor(snapshot.buffers[i], buffer)) return false;
-		const auto base = buffer.Base48();
-		const auto size = buffer.GetSize();
-		if (size == 0u) continue;
-		if (size - 1u > AddressMask - base) return false;
-		for (const auto [address, bytes]: reads) {
-			if (bytes == 0u) continue;
-			if (address > AddressMask || bytes - 1u > AddressMask - address ||
-			    (address <= base ? base - address < bytes : address - base < size)) {
-				return false;
-			}
-		}
-	}
-	return true;
-}
-
 const DescriptorSource* Source(const ResourcePlan& program, uint32_t source) {
 	if (source >= program.descriptor_sources.size()) {
 		return nullptr;
@@ -638,9 +619,10 @@ bool BuildSamplerPlan(const ShaderInfo& base, const Images& images, SamplerPlan&
 	return true;
 }
 
-static std::vector<ResourceBlock> ResourceControlFlow(const Program& program) {
-	// Any shader write may alias a scalar predicate read, including on a later loop visit.
-	if (program.blocks.size() != program.block_info.size() || HasShaderMemoryWrites(program)) {
+template <typename Predicate>
+static std::vector<ResourceBlock> ResourceControlFlow(const Program& program, const Predicate& predicate) {
+	// The renderer checks captured scalar reads against final buffer and image write ranges.
+	if (program.blocks.size() != program.block_info.size() || program.has_address_writes) {
 		return {};
 	}
 	std::unordered_map<uint32_t, uint32_t> indices;
@@ -659,9 +641,7 @@ static std::vector<ResourceBlock> ResourceControlFlow(const Program& program) {
 			case CFG::TerminatorKind::Branch: successors.push_back(terminator.true_block); break;
 			case CFG::TerminatorKind::ConditionalBranch:
 				successors = {terminator.true_block, terminator.false_block};
-				if (ValidateRuntimeValue(program, info.condition, RuntimeValueType::Integer)) {
-					block.condition = info.condition;
-				}
+				block.condition = info.condition;
 				break;
 			case CFG::TerminatorKind::IndirectBranch:
 				successors = terminator.indirect_targets;
@@ -678,13 +658,17 @@ static std::vector<ResourceBlock> ResourceControlFlow(const Program& program) {
 		}
 		for (const auto& inst: *program.blocks[i]) {
 			const auto op     = inst.GetOpcode();
+			if (op == ValueOpcode::ReadConst) {
+				block.srt_reads.push_back(inst.Arg(1).Resolve().U32());
+				continue;
+			}
 			const auto buffer = BufferAccessOf(op);
 			const auto image  = ImageOpcodeInfoOf(op);
 			if (buffer == BufferAccess::None && image.access == ImageAccess::None) {
 				continue;
 			}
 			const auto& memory = program.memory_info.at(inst.Flags<MemoryFlags>().index);
-			if (memory.planning_only) {
+			if (memory.planning_only || memory.kind == ResourceKind::IndirectBuffer) {
 				continue;
 			}
 			if (buffer != BufferAccess::None) {
@@ -700,6 +684,7 @@ static std::vector<ResourceBlock> ResourceControlFlow(const Program& program) {
 		block.sources.erase(std::unique(block.sources.begin(), block.sources.end()),
 		                    block.sources.end());
 	}
+	for (auto& block: blocks) block.condition = predicate(block.condition);
 	if (std::ranges::none_of(
 	        blocks, [](const ResourceBlock& block) { return !block.condition.IsEmpty(); })) {
 		return {};
@@ -914,10 +899,38 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 	for (const auto& read: program.srt_reads) {
 		plan.srt_reads.push_back({Clone(read.value), read.flat_offset});
 	}
-	plan.control_flow = ResourceControlFlow(program);
-	for (auto& block: plan.control_flow) {
-		block.condition = Clone(block.condition);
-	}
+	// A proven uniform factor can decide a branch even when its other lanes are unknown.
+	// Keep only that Boolean structure, never the varying shader dependency graph.
+	Value unknown;
+	std::function<Value(Value)> ClonePredicate = [&](Value value) -> Value {
+		value = value.Resolve();
+		if (value.IsEmpty()) return {};
+		if (ValidateRuntimeValue(program, value, RuntimeValueType::Integer)) return Clone(value);
+		const auto* inst = value.TryInstruction();
+		if (inst == nullptr) return {};
+		const auto op = inst->GetOpcode();
+		if (op == ValueOpcode::ConditionRef || op == ValueOpcode::LogicalNot) {
+			const auto operand = ClonePredicate(inst->Arg(0));
+			if (operand.IsEmpty() || op == ValueOpcode::ConditionRef) return operand;
+			auto& node = plan.value_storage.emplace_back(op);
+			node.SetArg(0, operand);
+			return Value(&node);
+		}
+		if (op != ValueOpcode::LogicalAnd && op != ValueOpcode::LogicalOr) return {};
+		auto left = ClonePredicate(inst->Arg(0));
+		auto right = ClonePredicate(inst->Arg(1));
+		if (left.IsEmpty() && right.IsEmpty()) return {};
+		if (left.IsEmpty() || right.IsEmpty()) {
+			if (unknown.IsEmpty()) unknown = Value(&plan.value_storage.emplace_back(ValueOpcode::UndefU1));
+			if (left.IsEmpty()) left = unknown;
+			if (right.IsEmpty()) right = unknown;
+		}
+		auto& node = plan.value_storage.emplace_back(op);
+		node.SetArg(0, left);
+		node.SetArg(1, right);
+		return Value(&node);
+	};
+	plan.control_flow = ResourceControlFlow(program, ClonePredicate);
 	plan.uniform_fill = AnalyzeUniformFill(program);
 	for (uint32_t i = 0; i < plan.uniform_fill.fill.words; ++i) {
 		plan.uniform_fill.values[i] = Clone(plan.uniform_fill.values[i]);
@@ -936,12 +949,15 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 		MarkCleanFlatSlots(plan, Source(plan, source->indirect_image->table_source),
 		                   plan.clean_flat_slots);
 	}
-	if (masked_image) {
-		plan.resource_tracking_complete &= !program.has_address_writes &&
-		    !std::ranges::any_of(plan.info.images, &ImageResource::written);
-		plan.capture_specialization_reads =
-		    std::ranges::any_of(plan.info.buffers, &BufferResource::written);
+	plan.capture_specialization_reads = masked_image || !plan.control_flow.empty();
+	// Writable descriptor addresses must come from clean backing for the alias proof.
+	for (const auto& buffer: plan.info.buffers) {
+		if (buffer.written) MarkCleanFlatSlots(plan, Source(plan, buffer.source), plan.clean_flat_slots);
 	}
+	for (const auto& image: plan.info.images) {
+		if (image.written) MarkCleanFlatSlots(plan, Source(plan, image.source), plan.clean_flat_slots);
+	}
+	if (masked_image) plan.resource_tracking_complete &= !program.has_address_writes;
 	return plan;
 }
 
@@ -952,21 +968,22 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 		return false;
 	}
 	const bool capture_reads = program.capture_specialization_reads;
-	auto& reads = program.specialization_reads;
+	auto& reads = snapshot.specialization_reads;
+	reads.clear();
 	ReadCapture capture {runtime, reads};
 	SrtRuntime observed = runtime;
 	if (capture_reads) {
-		reads.clear();
 		observed.userdata = &capture;
-		observed.read_specialization_memory = CaptureStrictRead;
-		if (observed.read_memory != nullptr) observed.read_memory = CaptureOrdinaryRead;
+		observed.read_specialization_memory = runtime.read_specialization_memory != nullptr
+		                                         ? CaptureStrictRead : nullptr;
+		observed.read_memory = CaptureOrdinaryRead;
 	}
 	SrtWalker clean(program, CleanRuntime(observed));
 	SrtWalker walker(program, observed, program.clean_flat_slots, &clean);
-	const auto active = clean.FindActiveSources();
 	if (!walker.RefreshFlatBuffer(snapshot.flattened_srt)) {
 		return false;
 	}
+	const auto active = std::span<const uint8_t>(program.active_sources);
 	snapshot.uniform_fill = {};
 	const auto& fill = program.uniform_fill;
 	const auto words = fill.fill.words;
@@ -979,12 +996,12 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 		snapshot.uniform_fill = fill.fill;
 		snapshot.uniform_fill.value = stored[0];
 	}
-	const auto evaluate = [&](uint32_t source, DescriptorValue& value) {
+	const auto evaluate = [&](uint32_t source, DescriptorValue& value, bool written = false) {
 		if (source >= program.descriptor_sources.size()) {
 			return false;
 		}
 		if (active.empty() || active[source]) {
-			return walker.EvaluateDescriptor(source, value);
+			return (written ? clean : walker).EvaluateDescriptor(source, value);
 		}
 		value = {};
 		value.dword_count = program.descriptor_sources[source].dword_count;
@@ -992,19 +1009,11 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 	};
 	snapshot.buffers.resize(program.info.buffers.size());
 	for (uint32_t i = 0; i < program.info.buffers.size(); ++i) {
-		if (!evaluate(program.info.buffers[i].source, snapshot.buffers[i])) {
+		if (!evaluate(program.info.buffers[i].source, snapshot.buffers[i],
+		              program.info.buffers[i].written)) {
 			return false;
 		}
 	}
-	if (capture_reads) {
-		for (uint32_t i = 0; i < program.info.buffers.size(); ++i) {
-			const auto& buffer = program.info.buffers[i];
-			if (!buffer.written || (!active.empty() && !active[buffer.source])) continue;
-			DescriptorValue strict;
-			if (!clean.EvaluateDescriptor(buffer.source, strict) ||
-			    strict != snapshot.buffers[i]) return false;
-		}
-	}
 	snapshot.images.resize(program.info.images.size());
 	specialization.images.resize(program.info.images.size());
 	for (uint32_t i = 0; i < program.info.images.size(); ++i) {
@@ -1040,7 +1049,7 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 				return false;
 			}
 		} else {
-			if (!evaluate(image.source, snapshot.images[i])) {
+			if (!evaluate(image.source, snapshot.images[i], image.written)) {
 				return false;
 			}
 			if (!ValidImageDescriptor(snapshot.images[i], image.r128)) {
@@ -1054,7 +1063,6 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 			return false;
 		}
 	}
-	if (capture_reads && !WrittenBuffersDisjoint(program, snapshot, reads)) return false;
 	snapshot.user_data.assign(runtime.user_data.begin(), runtime.user_data.end());
 	return BuildResourceSpecialization(program, snapshot, specialization);
 }

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

@@ -44,6 +44,34 @@ uint32_t PossibleU32Bits(Value value) {
 	}
 }
 
+bool IsNoncanonicalFlatAddress(Value high, Value active) {
+	high = high.Resolve();
+	active = active.Resolve();
+	// A predicated VGPR write supplies its new value on exactly the lanes which
+	// execute this access. Do not discard a selection under a different mask.
+	while (const auto* select = high.TryInstruction()) {
+		if (select->GetOpcode() != ValueOpcode::SelectU32 || select->Arg(0).Resolve() != active) break;
+		high = select->Arg(1).Resolve();
+	}
+	uint32_t lower = 0, upper = UINT32_MAX;
+	if (high.IsImmediate() && high.GetType() == Type::U32) {
+		lower = upper = high.U32();
+	} else if (const auto* inst = high.TryInstruction();
+	           inst != nullptr && inst->GetOpcode() == ValueOpcode::UMedTri32) {
+		// The median lies between any two constant inputs, regardless of the third.
+		uint32_t count = 0;
+		for (size_t index = 0; index < inst->NumArgs(); ++index) {
+			const auto value = inst->Arg(index).Resolve();
+			if (!value.IsImmediate() || value.GetType() != Type::U32) continue;
+			if (count++ == 0) lower = upper = value.U32();
+			else { lower = std::min(lower, value.U32()); upper = std::max(upper, value.U32()); }
+		}
+		if (count < 2) return false;
+	}
+	// Leave one high DWORD on each side for the signed instruction offset/carry.
+	return lower > 0x00008000u && upper < 0xffff7fffu;
+}
+
 Value CanonicalizeSampleAdjustDword3(Value value) {
 	for (;;) {
 		value            = value.Resolve();
@@ -1649,6 +1677,10 @@ private:
 
 	void Collect(Inst& inst) {
 		const auto op           = inst.GetOpcode();
+		if (op == ValueOpcode::BvhIntersect) {
+			m_info.uses_dma = true;
+			return;
+		}
 		const auto buffer       = BufferAccessOf(op);
 		const auto address_info = AddressOpcodeInfoOf(op);
 		const auto image_info   = ImageOpcodeInfoOf(op);
@@ -1695,6 +1727,15 @@ private:
 			if (!IsAddressResourceKind(memory.kind)) {
 				Fail(flags.pc, "address operation has invalid resource kind");
 			}
+			if (memory.kind == ResourceKind::Flat && memory.address_is_full &&
+			    IsNoncanonicalFlatAddress(inst.Arg(2), inst.Arg(inst.NumArgs() - 1))) {
+				ValidateAddressHandle(inst.Arg(0), flags.pc);
+				if (memory.data_bits != 32u || m_program.scratch_dwords == 0) {
+					Fail(flags.pc, "local FLAT access requires DWORD data and per-thread scratch storage");
+				}
+				m_program.memory_info[flags.index].kind = ResourceKind::FlatLocal;
+				return;
+			}
 			if (memory.kind == ResourceKind::Scratch) {
 				handle = inst.Arg(0).Resolve().TryInstruction();
 				if (handle == nullptr || handle->GetOpcode() != ValueOpcode::GetScratchResource ||

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

@@ -326,6 +326,8 @@ void CollectBuiltinInputs(const Program& program, ShaderInfo& info) {
 }
 
 void CollectOutputs(const Program& program, ShaderStageInputInfo input_info, ShaderInfo& info) {
+	const bool alpha_remap = program.stage == ShaderType::Pixel && input_info.pixel != nullptr &&
+	                         input_info.pixel->alpha_blend_source_remap;
 	for (const auto* block: program.blocks) {
 		for (const auto& inst: *block) {
 			if (inst.GetOpcode() != ValueOpcode::SetAttribute) {
@@ -385,8 +387,14 @@ void CollectOutputs(const Program& program, ShaderStageInputInfo input_info, Sha
 					          export_info.index, fmt::format("out_param_{}", export_info.index));
 					break;
 				case ExportTargetKind::Mrt:
+					if (alpha_remap && export_info.index != 0) {
+						break;
+					}
 					AddOutput(info, StageOutputKind::Mrt, export_info.index, export_info.index,
 					          fmt::format("out_mrt_{}", export_info.index));
+					if (alpha_remap) {
+						AddOutput(info, StageOutputKind::Mrt, 1, 1, "out_mrt_1");
+					}
 					break;
 				default: break;
 			}

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

@@ -795,18 +795,26 @@ bool SrtWalker::EvaluateInst(const Inst& inst, uint64_t& result) {
 				return true;
 			}
 			return false;
-		case ValueOpcode::LogicalAnd:
-			if (binary()) {
-				result = (a != 0u) && (b != 0u);
+		case ValueOpcode::LogicalAnd: {
+			const bool left = Arg(inst, 0, a);
+			if (left && a == 0u) {
+				result = 0u;
 				return true;
 			}
-			return false;
-		case ValueOpcode::LogicalOr:
-			if (binary()) {
-				result = (a != 0u) || (b != 0u);
+			if (!Arg(inst, 1, b) || (b != 0u && !left)) return false;
+			result = b != 0u;
+			return true;
+		}
+		case ValueOpcode::LogicalOr: {
+			const bool left = Arg(inst, 0, a);
+			if (left && a != 0u) {
+				result = 1u;
 				return true;
 			}
-			return false;
+			if (!Arg(inst, 1, b) || (b == 0u && !left)) return false;
+			result = b != 0u;
+			return true;
+		}
 		case ValueOpcode::LogicalXor:
 			if (binary()) {
 				result = (a != 0u) != (b != 0u);
@@ -844,16 +852,31 @@ bool SrtWalker::EvaluateDescriptor(uint32_t source, DescriptorValue& result) {
 	return true;
 }
 
-std::span<const uint8_t> SrtWalker::FindActiveSources() {
-	if (m_program.control_flow.empty()) {
-		return {};
-	}
+bool SrtWalker::RefreshFlatBuffer(std::vector<uint32_t>& flat) {
+	if (!m_program.srt_plan_complete) return false;
+	const auto refresh = [&](uint32_t slot) {
+		if (slot >= m_program.srt_reads.size()) return false;
+		const auto& read = m_program.srt_reads[slot];
+		const bool clean = read.flat_offset < m_clean_flat_slots.size() &&
+		                   m_clean_flat_slots[read.flat_offset] != 0u;
+		if (clean && (m_clean_evaluator == nullptr || m_runtime.read_specialization_memory == nullptr))
+			return false;
+		auto& evaluator = clean ? *m_clean_evaluator : *this;
+		return read.flat_offset < flat.size() && evaluator.Evaluate(read.value, flat[read.flat_offset]);
+	};
 	auto& active = m_program.active_sources;
+	if (m_program.control_flow.empty()) {
+		active.clear();
+		flat.resize(m_program.srt_reads.size());
+		for (uint32_t slot = 0; slot < m_program.srt_reads.size(); ++slot) {
+			if (!refresh(slot)) return false;
+		}
+		return true;
+	}
+	flat.assign(m_program.srt_reads.size(), 0u);
 	active.assign(m_program.descriptor_sources.size(), 1u);
 	for (const auto& block: m_program.control_flow) {
-		for (const auto source: block.sources) {
-			active.at(source) = 0u;
-		}
+		for (const auto source: block.sources) active.at(source) = 0u;
 	}
 	auto& visited = m_program.visited_blocks;
 	auto& pending = m_program.pending_blocks;
@@ -863,41 +886,22 @@ std::span<const uint8_t> SrtWalker::FindActiveSources() {
 	while (!pending.empty()) {
 		const auto index = pending.back();
 		pending.pop_back();
-		if (visited.at(index)) {
-			continue;
-		}
+		if (visited.at(index)) continue;
 		visited[index] = 1u;
 		const auto& block = m_program.control_flow[index];
-		for (const auto source: block.sources) {
-			active[source] = 1u;
+		for (const auto source: block.sources) active[source] = 1u;
+		for (const auto slot: block.srt_reads) {
+			if (!refresh(slot)) return false;
 		}
 		uint32_t condition = 0;
+		auto& predicate = m_clean_evaluator != nullptr ? *m_clean_evaluator : *this;
 		if (!block.condition.IsEmpty() && m_runtime.read_specialization_memory != nullptr &&
-		    Evaluate(block.condition, condition)) {
+		    predicate.Evaluate(block.condition, condition)) {
 			pending.push_back(block.successors[condition != 0u ? 0u : 1u]);
 		} else {
 			pending.insert(pending.end(), block.successors.begin(), block.successors.end());
 		}
 	}
-	return active;
-}
-
-bool SrtWalker::RefreshFlatBuffer(std::vector<uint32_t>& flat) {
-	if (!m_program.srt_plan_complete) {
-		return false;
-	}
-	flat.resize(m_program.srt_reads.size());
-	for (const auto& read: m_program.srt_reads) {
-		const bool clean = read.flat_offset < m_clean_flat_slots.size() &&
-		                   m_clean_flat_slots[read.flat_offset] != 0u;
-		if (clean && (m_clean_evaluator == nullptr || m_runtime.read_specialization_memory == nullptr)) {
-			return false;
-		}
-		auto& evaluator = clean ? *m_clean_evaluator : *this;
-		if (read.flat_offset >= flat.size() || !evaluator.Evaluate(read.value, flat[read.flat_offset])) {
-			return false;
-		}
-	}
 	return true;
 }
 

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

@@ -38,8 +38,7 @@ public:
 
 	bool Evaluate(Value value, uint32_t& result);
 	bool EvaluateDescriptor(uint32_t source, DescriptorValue& result);
-	// An empty span means that all sources are active.
-	std::span<const uint8_t> FindActiveSources();
+	// Refreshes reachable scalar reads and active descriptor sources in one walk.
 	bool RefreshFlatBuffer(std::vector<uint32_t>& flat);
 
 private:

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

@@ -165,7 +165,6 @@ Value InitialValue(M0Tag) {
 
 enum class ReadStep { Start, SetValue, PushPhiArgument };
 
-template <typename T>
 struct ReadState {
 	Block*   block = nullptr;
 	Value    result;
@@ -183,8 +182,8 @@ public:
 
 	template <typename T>
 	Value Read(T variable, Block* root) {
-		std::vector<ReadState<T>> stack {{}, {.block = root}};
-		const auto                prepare_phi = [&]() {
+		stack.assign({{}, {.block = root}});
+		const auto prepare_phi = [&]() {
 			auto&      state        = stack.back();
 			const auto predecessors = state.block->ImmPredecessors();
 			if (state.pred == predecessors.size()) {
@@ -290,6 +289,7 @@ private:
 
 	std::unordered_map<Block*, std::map<Variable, Inst*>> incomplete_phis;
 	DefTable                                              definitions;
+	std::vector<ReadState>                               stack;
 };
 
 void VisitInstruction(Pass& pass, Block* block, Inst& inst) {

src/graphics/shader/shader.cpp

@@ -41,44 +41,33 @@
 
 namespace Libs::Graphics {
 
-struct ShaderBinaryInfo {
-	uint8_t  signature[7];
-	uint8_t  version;
-	uint32_t pssl_or_cg  : 1;
-	uint32_t cached      : 1;
-	uint32_t type        : 4;
-	uint32_t source_type : 2;
-	uint32_t length      : 24;
-	uint8_t  chunk_usage_base_offset_dw;
-	uint8_t  num_input_usage_slots;
-	uint8_t  is_srt                 : 1;
-	uint8_t  is_srt_used_info_valid : 1;
-	uint8_t  is_extended_usage_info : 1;
-	uint8_t  reserved2              : 5;
-	uint8_t  reserved3;
-	uint32_t hash0;
-	uint32_t hash1;
-	uint32_t crc32;
+struct ShaderMapEntry {
+	ShaderMappedData data;
+	uint64_t hash;
 };
 
-static std::unique_ptr<std::unordered_map<uint64_t, ShaderMappedData>> g_shader_map;
+static std::unique_ptr<std::unordered_map<uint64_t, ShaderMapEntry>> g_shader_map;
 static std::mutex                                                      g_shader_map_mutex;
 
 void ShaderInit() {
 	EXIT_IF(g_shader_map != nullptr);
 
-	g_shader_map = std::make_unique<std::unordered_map<uint64_t, ShaderMappedData>>();
+	g_shader_map = std::make_unique<std::unordered_map<uint64_t, ShaderMapEntry>>();
 }
 
 void ShaderMapUserData(uint64_t addr, const ShaderMappedData& data) {
 	EXIT_IF(g_shader_map == nullptr);
 
+	if (addr == 0 || data.code_size_bytes == 0 || data.code_size_bytes % sizeof(uint32_t) != 0) {
+		EXIT("ShaderMapUserData shader=0x%016" PRIx64 " has invalid AGC shader_size=0x%08" PRIx32 "\n",
+		     addr, data.code_size_bytes);
+	}
+	const auto hash = XXH3_64bits(reinterpret_cast<const void*>(addr), data.code_size_bytes);
 	std::scoped_lock lock(g_shader_map_mutex);
-
-	(*g_shader_map)[addr] = data;
+	(*g_shader_map)[addr] = {data, hash};
 }
 
-static ShaderMappedData ShaderGetMappedData(uint64_t addr, const char* label) {
+static ShaderMapEntry ShaderGetMappedData(uint64_t addr, const char* label) {
 	EXIT_IF(g_shader_map == nullptr);
 
 	std::scoped_lock lock(g_shader_map_mutex);
@@ -90,37 +79,15 @@ static ShaderMappedData ShaderGetMappedData(uint64_t addr, const char* label) {
 	EXIT("%s shader=0x%016" PRIx64 " is missing from ShaderMap\n", label, addr);
 }
 
-static const ShaderBinaryInfo* GetBinaryInfo(const uint32_t* code) {
-	EXIT_IF(code == nullptr);
-
-	if (code[0] == 0xBEEB03FF) {
-		return reinterpret_cast<const ShaderBinaryInfo*>(code +
-		                                                 static_cast<size_t>(code[1] + 1) * 2);
-	}
-
-	return nullptr;
-}
-
-static uint64_t GetDeclaredShaderHash(uint64_t shader_addr) {
-	const auto* header = GetBinaryInfo(reinterpret_cast<const uint32_t*>(shader_addr));
-	return header != nullptr ? (static_cast<uint64_t>(header->hash1) << 32u) | header->hash0 : 0;
-}
-
-static ShaderParams GetShaderParams(uint64_t shader_addr, const char* label, uint64_t declared_hash,
+static ShaderParams GetShaderParams(uint64_t shader_addr, uint64_t hash,
 	                                std::span<const uint32_t> user_data,
 	                                const ShaderMappedData& data, uint32_t user_data_base = 0) {
-	if (data.code_size_bytes == 0 || data.code_size_bytes % sizeof(uint32_t) != 0) {
-		EXIT("%s hash=0x%016" PRIx64 " shader=0x%016" PRIx64
-		     " has invalid AGC shader_size=0x%08" PRIx32 "\n",
-		     label, declared_hash, shader_addr, data.code_size_bytes);
-	}
 	const auto code_words = data.code_size_bytes / sizeof(uint32_t);
 	const auto code = std::span {reinterpret_cast<const uint32_t*>(shader_addr), code_words};
 	ShaderParams params {
 	    .code            = code,
 	    .user_data_count = static_cast<uint32_t>(user_data.size()) + user_data_base,
-	    .hash            = declared_hash != 0 ? declared_hash
-	                                          : XXH3_64bits(code.data(), code.size_bytes()),
+	    .hash            = hash,
 	};
 	EXIT_IF(user_data.size() > HW::UserSgprInfo::SGPRS_MAX ||
 	        params.user_data_count > params.user_data.size());
@@ -414,6 +381,8 @@ static void ShaderApplyAttribSemantics(ShaderVertexInputInfo& info,
 		r.fields[1]       = sharp[1];
 		r.fields[2]       = sharp[2];
 		r.fields[3]       = sharp[3];
+		EXIT_NOT_IMPLEMENTED(r.AddTid());
+		EXIT_NOT_IMPLEMENTED(r.SwizzleEnabled());
 		if (format != Prospero::VertexAttribFormat::kInvalid) {
 			const auto                   format_raw    = static_cast<uint32_t>(format);
 			const auto                   buffer_format = format_raw >> 2u;
@@ -650,7 +619,8 @@ void BuildStageStaticKey(const ShaderVertexInputInfo& info, std::vector<uint32_t
 		key.insert(key.end(), {mesh.wave_size, mesh.host_subgroup_size, mesh.lds_size_dwords,
 		                       mesh.scratch_size_dwords, mesh.input_primitive,
 		                       mesh.primitives_per_group, mesh.vertices_per_group,
-		                       mesh.max_vertices, mesh.max_primitives, mesh.provoking_vertex});
+		                       mesh.max_vertices, mesh.max_primitives, mesh.provoking_vertex,
+		                       static_cast<uint32_t>(mesh.fast_launch)});
 	}
 	key.push_back(info.tess.input_control_points);
 	if (info.tess.input_control_points != 0) {
@@ -662,12 +632,9 @@ void BuildStageStaticKey(const ShaderVertexInputInfo& info, std::vector<uint32_t
 	for (int i = 0; i < info.resources_num; i++) {
 		const auto& resource    = info.resources[i];
 		const auto& destination = info.resources_dst[i];
-		key.push_back(destination.register_start);
 		key.push_back(destination.registers_num);
-		key.push_back(destination.fetch_index);
 		key.push_back(static_cast<uint32_t>(destination.attr_id));
-		key.push_back(resource.fields[1] & 0xbfff0000u); // Stride and swizzle enable.
-		key.push_back(resource.fields[3] & 0x3087ffffu); // Channels, format, OOB, and add TID.
+		key.push_back(resource.fields[3] & 0x7ffffu); // Channels and format used by embedded fetch.
 	}
 }
 
@@ -693,6 +660,7 @@ void BuildStageStaticKey(const ShaderPixelInputInfo& info, std::vector<uint32_t>
 	key.push_back(static_cast<uint32_t>(info.ps_sample_mask_export_enable));
 	key.push_back(static_cast<uint32_t>(info.ps_early_z));
 	key.push_back(static_cast<uint32_t>(info.dual_source_blending));
+	key.push_back(static_cast<uint32_t>(info.alpha_blend_source_remap));
 	key.insert(key.end(), std::begin(info.target_output_mode), std::end(info.target_output_mode));
 	for (uint32_t base = 0; base < info.target_export_mapping.size(); base += 4u) {
 		uint32_t packed = 0;
@@ -725,11 +693,10 @@ void BuildStageStaticKey(const ShaderComputeInputInfo& info, std::vector<uint32_
 ShaderParams PrepareProgram(const HW::VertexShaderInfo& regs, const HW::Context& context,
                             const HW::UserConfig& user_config, ShaderVertexInputInfo& info) {
 	const auto& sh     = context.GetShaderRegisters();
-	const auto data = ShaderGetMappedData(regs.es_regs.data_addr, "ShaderGetInputInfoVS():");
+	const auto [data, hash] = ShaderGetMappedData(regs.es_regs.data_addr, "ShaderGetInputInfoVS():");
 	const bool merged = (context.GetShaderStages() & 0x20u) != 0;
 	auto        params = GetShaderParams(
-	    regs.es_regs.data_addr, "ShaderRecompiler VS",
-	    GetDeclaredShaderHash(regs.es_regs.data_addr),
+	    regs.es_regs.data_addr, hash,
 	    std::span<const uint32_t>(regs.gs_user_sgpr.value, regs.gs_regs.rsrc2.user_sgpr), data,
 	    merged ? 8u : 0u);
 	if (!merged) {
@@ -752,12 +719,14 @@ ShaderParams PrepareProgram(const HW::VertexShaderInfo& regs, const HW::Context&
 	mesh.provoking_vertex    = context.GetModeControl().provoking_vtx_last ? 2u : 0u;
 	mesh.lds_size_dwords     = static_cast<uint32_t>(regs.gs_regs.rsrc2.lds_size) * 128u;
 	mesh.scratch_size_dwords = data.scratch_size_dwords;
+	const auto fast_launch = (context.GetShaderStages() >> 19u) & 3u;
+	EXIT_NOT_IMPLEMENTED(fast_launch > 1u);
+	mesh.fast_launch = fast_launch != 0;
 	if (data.type == Prospero::ShaderBinaryType::kGsFront) {
 		EXIT_IF(regs.gs_regs.data_addr == 0);
-		const auto back = ShaderGetMappedData(regs.gs_regs.data_addr, "ShaderGetInputInfoGS():");
+		const auto [back, back_hash] = ShaderGetMappedData(regs.gs_regs.data_addr, "ShaderGetInputInfoGS():");
 		const auto back_params =
-		    GetShaderParams(regs.gs_regs.data_addr, "ShaderRecompiler GS",
-		                    GetDeclaredShaderHash(regs.gs_regs.data_addr), {}, back);
+		    GetShaderParams(regs.gs_regs.data_addr, back_hash, {}, back);
 		params.back_code = back_params.code;
 		params.user_data[0] = static_cast<uint32_t>(regs.gs_regs.user_data_addr);
 		params.user_data[1] = static_cast<uint32_t>(regs.gs_regs.user_data_addr >> 32u);
@@ -765,9 +734,20 @@ ShaderParams PrepareProgram(const HW::VertexShaderInfo& regs, const HW::Context&
 		params.hash = XXH3_64bits(hashes, sizeof(hashes));
 		mesh.scratch_size_dwords = std::max(mesh.scratch_size_dwords, back.scratch_size_dwords);
 	}
-	EXIT_NOT_IMPLEMENTED(regs.gs_regs.rsrc1.gs_vgpr_component_count != 3u ||
-	                     regs.gs_regs.rsrc2.es_vgpr_component_count != 3u);
 	const auto& group = user_config.GetGeControl();
+	if (mesh.fast_launch) {
+		const auto& user_vgpr = user_config.GetGeUserVgprEn();
+		EXIT_NOT_IMPLEMENTED(data.type != Prospero::ShaderBinaryType::kGs ||
+		                     regs.gs_regs.rsrc1.gs_vgpr_component_count != 0u ||
+		                     regs.gs_regs.rsrc2.es_vgpr_component_count != 1u ||
+		                     user_config.GetPrimType() != Prospero::PrimitiveType::kPointList ||
+		                     group.primitive_group_size != 1u || group.vertex_group_size != 1u ||
+		                     user_vgpr.vgpr1 || user_vgpr.vgpr2 || user_vgpr.vgpr3 ||
+		                     mesh.max_vertices != sh.m_vgtGsMaxVertOut);
+	} else {
+		EXIT_NOT_IMPLEMENTED(regs.gs_regs.rsrc1.gs_vgpr_component_count != 3u ||
+		                     regs.gs_regs.rsrc2.es_vgpr_component_count != 3u);
+	}
 	if ((user_config.GetPrimType() != Prospero::PrimitiveType::kPointList &&
 	     user_config.GetPrimType() != Prospero::PrimitiveType::kLineList &&
 	     user_config.GetPrimType() != Prospero::PrimitiveType::kTriFan &&
@@ -780,7 +760,8 @@ ShaderParams PrepareProgram(const HW::VertexShaderInfo& regs, const HW::Context&
 		     mesh.input_primitive, sh.m_vgtGsOutPrimType, sh.m_vgtGsMaxVertOut,
 		     group.primitive_group_size, group.vertex_group_size, mesh.max_vertices);
 	}
-	mesh.max_primitives       = group.primitive_group_size * (sh.m_vgtGsMaxVertOut - 2u);
+	mesh.max_primitives = mesh.fast_launch ? mesh.max_vertices :
+	                      group.primitive_group_size * (sh.m_vgtGsMaxVertOut - 2u);
 	mesh.primitives_per_group = std::min({static_cast<uint32_t>(group.primitive_group_size),
 	                                      mesh.InputPrimitiveCount(group.vertex_group_size),
 	                                      mesh.max_vertices / sh.m_vgtGsMaxVertOut});
@@ -796,9 +777,9 @@ std::array<ShaderParams, 3>
 PrepareTessellationPrograms(const HW::VertexShaderInfo& regs, const HW::Context& context,
                             std::array<ShaderVertexInputInfo, 3>& input_info) {
 	const auto& sh        = context.GetShaderRegisters();
-	const auto  local     = ShaderGetMappedData(regs.ls_regs.data_addr, "ShaderGetInputInfoLS():");
-	const auto  control   = ShaderGetMappedData(regs.hs_regs.data_addr, "ShaderGetInputInfoHS():");
-	const auto evaluation = ShaderGetMappedData(regs.es_regs.data_addr, "ShaderGetInputInfoTES():");
+	const auto [local, local_hash] = ShaderGetMappedData(regs.ls_regs.data_addr, "ShaderGetInputInfoLS():");
+	const auto [control, control_hash] = ShaderGetMappedData(regs.hs_regs.data_addr, "ShaderGetInputInfoHS():");
+	const auto [evaluation, evaluation_hash] = ShaderGetMappedData(regs.es_regs.data_addr, "ShaderGetInputInfoTES():");
 	EXIT_NOT_IMPLEMENTED(local.type != Prospero::ShaderBinaryType::kHsFront ||
 	                     control.type != Prospero::ShaderBinaryType::kHsBack ||
 	                     evaluation.type != Prospero::ShaderBinaryType::kGs);
@@ -809,12 +790,9 @@ PrepareTessellationPrograms(const HW::VertexShaderInfo& regs, const HW::Context&
 	const auto local_users      = std::span(regs.hs_user_sgpr.value, regs.hs_regs.rsrc2.user_sgpr);
 	const auto evaluation_users = std::span(regs.gs_user_sgpr.value, regs.gs_regs.rsrc2.user_sgpr);
 	std::array<ShaderParams, 3> params {
-	    GetShaderParams(regs.ls_regs.data_addr, "ShaderRecompiler LS",
-	                    GetDeclaredShaderHash(regs.ls_regs.data_addr), local_users, local),
-	    GetShaderParams(regs.hs_regs.data_addr, "ShaderRecompiler HS",
-	                    GetDeclaredShaderHash(regs.hs_regs.data_addr), local_users, control, 8u),
-	    GetShaderParams(regs.es_regs.data_addr, "ShaderRecompiler TES",
-	                    GetDeclaredShaderHash(regs.es_regs.data_addr), evaluation_users,
+	    GetShaderParams(regs.ls_regs.data_addr, local_hash, local_users, local),
+	    GetShaderParams(regs.hs_regs.data_addr, control_hash, local_users, control, 8u),
+	    GetShaderParams(regs.es_regs.data_addr, evaluation_hash, evaluation_users,
 	                    evaluation),
 	};
 	// The fused HS back half receives its separate user-data address in s0:s1.
@@ -854,19 +832,19 @@ ShaderParams PrepareProgram(
     const HW::PixelShaderInfo& regs, const HW::ShaderRegisters& sh,
     std::span<const Prospero::ColorComponentMapping, 8> target_export_mapping,
     ShaderPixelInputInfo&                               ps_info) {
-	const auto data = ShaderGetMappedData(regs.ps_regs.data_addr, "ShaderGetInputInfoPS():");
+	const auto [data, hash] = ShaderGetMappedData(regs.ps_regs.data_addr, "ShaderGetInputInfoPS():");
 	ShaderGetStaticInputInfoPS(regs, sh, target_export_mapping, data, ps_info);
 	return GetShaderParams(
-	    regs.ps_regs.data_addr, "ShaderRecompiler PS", GetDeclaredShaderHash(regs.ps_regs.data_addr),
+	    regs.ps_regs.data_addr, hash,
 	    std::span<const uint32_t>(regs.ps_user_sgpr.value, regs.ps_regs.rsrc2.user_sgpr), data);
 }
 
 ShaderParams PrepareProgram(const HW::ComputeShaderInfo& regs, const HW::ShaderRegisters& sh,
                             ShaderComputeInputInfo& info) {
-	const auto data = ShaderGetMappedData(regs.cs_regs.data_addr, "ShaderGetInputInfoCS():");
+	const auto [data, hash] = ShaderGetMappedData(regs.cs_regs.data_addr, "ShaderGetInputInfoCS():");
 	ShaderGetStaticInputInfoCS(regs, sh, data, info);
 	return GetShaderParams(
-	    regs.cs_regs.data_addr, "ShaderRecompiler CS", GetDeclaredShaderHash(regs.cs_regs.data_addr),
+	    regs.cs_regs.data_addr, hash,
 	    std::span<const uint32_t>(regs.cs_user_sgpr.value, regs.cs_regs.user_sgpr), data);
 }
 

src/graphics/shader/shader.h

@@ -88,6 +88,7 @@ struct ShaderMeshInputInfo: ShaderWorkgroupInputInfo {
 	uint32_t max_vertices         = 0;
 	uint32_t max_primitives       = 0;
 	uint32_t provoking_vertex     = 0;
+	bool     fast_launch          = false;
 
 	[[nodiscard]] constexpr uint32_t InputPrimitiveSize() const {
 		switch (static_cast<Prospero::PrimitiveType>(input_primitive)) {
@@ -178,6 +179,8 @@ struct ShaderPixelInputInfo {
 	bool                                           ps_sample_mask_export_enable = false;
 	bool                                           ps_sample_shading            = false;
 	bool                                           dual_source_blending         = false;
+	// Export logical alpha through MRT1 for blending after channel swizzling.
+	bool                                           alpha_blend_source_remap     = false;
 	bool                                           ps_early_z                   = false;
 	bool                                           ps_execute_on_noop           = false;
 	ShaderStageRuntime                             stage;

src/kernel/memory.cpp

@@ -963,18 +963,7 @@ static bool IsInPrtAperture(uint64_t address, uint64_t size = 1) {
 	return false;
 }
 
-bool TryReadPrtBacking(uint64_t vaddr, void* data, uint64_t size) {
-	std::vector<VirtualRanges::Range> ranges;
-	if (g_guest_address_space == nullptr || g_virtual_ranges == nullptr ||
-	    !IsInPrtAperture(vaddr, size) || !g_virtual_ranges->QuerySpan(vaddr, size, &ranges)) {
-		return false;
-	}
-	if (std::any_of(ranges.begin(), ranges.end(), [](const auto& range) {
-		    return !IsReservedRangeType(range.type) &&
-		           !g_guest_address_space->BackingContains(range.start, range.size);
-	    })) {
-		return false;
-	}
+bool TryReadSparseBacking(uint64_t vaddr, void* data, uint64_t size) {
 	return g_guest_address_space->TryReadSparseBacking(vaddr, data, size);
 }
 

src/kernel/memory.h

@@ -115,7 +115,7 @@ int MapAutomaticMemory(uint64_t vaddr, size_t size, int type, int prot);
 bool                   TryWriteBacking(uint64_t vaddr, const void* data, uint64_t size);
 bool                   TryReadBacking(uint64_t vaddr, void* data, uint64_t size);
 bool                   TryReadGpuCleanBacking(uint64_t vaddr, void* data, uint64_t size);
-bool                   TryReadPrtBacking(uint64_t vaddr, void* data, uint64_t size);
+bool                   TryReadSparseBacking(uint64_t vaddr, void* data, uint64_t size);
 [[nodiscard]] uint64_t ClampRangeSize(uint64_t vaddr, uint64_t size);
 void                   WriteBacking(uint64_t vaddr, const void* data, uint64_t size) noexcept;
 void                   InvalidateMemory(uint64_t vaddr, uint64_t size);

src/kernel/memoryAddressSpace.inc

@@ -212,14 +212,21 @@ public:
 		}
 
 		auto* destination = static_cast<uint8_t*>(data);
-		std::memset(destination, 0, size);
+		auto  cursor      = vaddr;
 		for (auto mapping = first; mapping != m_maps.end() && mapping->second.vaddr < end;
 		     ++mapping) {
 			const auto& entry  = mapping->second;
 			const auto  start  = std::max(vaddr, entry.vaddr);
 			const auto  bytes  = std::min(end, entry.vaddr + entry.size) - start;
 			const auto  offset = entry.backing_offset + start - entry.vaddr;
-			std::memcpy(destination + start - vaddr, m_backing_base + offset, bytes);
+			if (cursor < start) {
+				std::memset(destination + (cursor - vaddr), 0, start - cursor);
+			}
+			std::memcpy(destination + (start - vaddr), m_backing_base + offset, bytes);
+			cursor = start + bytes;
+		}
+		if (cursor < end) {
+			std::memset(destination + (cursor - vaddr), 0, end - cursor);
 		}
 		return true;
 	}
@@ -731,6 +738,20 @@ public:
 	}
 
 	bool TryReadSparseBacking(uint64_t vaddr, void* data, uint64_t size) {
+		if (!Owns(vaddr, size)) {
+			return false;
+		}
+		std::lock_guard lock(m_mutex);
+		auto            mapping = m_mapped.upper_bound(vaddr);
+		if (mapping != m_mapped.begin()) {
+			--mapping;
+		}
+		for (; mapping != m_mapped.end() && mapping->second.vaddr < vaddr + size; ++mapping) {
+			const auto& region = mapping->second;
+			if (region.type == MappingType::Private && region.vaddr + region.size > vaddr) {
+				return false;
+			}
+		}
 		return m_backing->TryReadSparseBacking(vaddr, data, size);
 	}
 

src/kernel/pthread.cpp

@@ -512,15 +512,6 @@ static void SchedulerBackoffOnce() {
 #endif
 }
 
-static bool SleepMicroSchedulerBackoff(uint64_t microseconds) {
-	if (microseconds > 1) {
-		return false;
-	}
-
-	SchedulerBackoffOnce();
-	return true;
-}
-
 static void SleepMicroWithSignalPoll(uint64_t microseconds) {
 	if (microseconds == 0) {
 		KernelDispatchPendingSignalForCurrentThread();
@@ -529,9 +520,7 @@ static void SleepMicroWithSignalPoll(uint64_t microseconds) {
 
 	while (microseconds > 0) {
 		const auto step = std::min<uint64_t>(microseconds, SIGNAL_APC_POLL_MICROS);
-		if (!SleepMicroSchedulerBackoff(step)) {
-			Common::Thread::SleepMicro(step);
-		}
+		Common::Thread::SleepMicro(step);
 		microseconds -= step;
 		KernelDispatchPendingSignalForCurrentThread();
 	}
@@ -3842,28 +3831,16 @@ void KYTY_SYSV_ABI KernelSetThreadDtors(thread_dtors_func_t dtors) {
 }
 
 int KYTY_SYSV_ABI KernelUsleep(KernelUseconds microseconds) {
-	Common::Timer t;
-	t.Start();
 	SleepMicroWithSignalPoll(microseconds);
-	// double ts = t.GetTimeS();
-	// LOGF("\tactual: %g microseconds\n", ts * 1000000.0);
 	return OK;
 }
 
 unsigned int KYTY_SYSV_ABI KernelSleep(unsigned int seconds) {
-	PRINT_NAME();
-	LOGF("\tsleep: %u\n", seconds);
-	Common::Timer t;
-	t.Start();
 	SleepMicroWithSignalPoll(static_cast<uint64_t>(seconds) * 1000000ull);
-	double ts = t.GetTimeS();
-	LOGF("\tactual: %g seconds\n", ts);
 	return OK;
 }
 
 int KYTY_SYSV_ABI KernelNanosleep(const KernelTimespec* rqtp, KernelTimespec* rmtp) {
-	PRINT_NAME();
-
 	if (rqtp == nullptr) {
 		return KERNEL_ERROR_EFAULT;
 	}
@@ -3880,13 +3857,7 @@ int KYTY_SYSV_ABI KernelNanosleep(const KernelTimespec* rqtp, KernelTimespec* rm
 	uint64_t nanos =
 	    static_cast<uint64_t>(rqtp->tv_sec) * 1000000000ull + static_cast<uint64_t>(rqtp->tv_nsec);
 
-	LOGF("\tnanosleep: %" PRIu64 "\n", nanos);
-
-	Common::Timer t;
-	t.Start();
 	SleepNanoWithSignalPoll(nanos);
-	double ts = t.GetTimeS();
-	LOGF("\tactual: %g nanoseconds\n", ts * 1000000000.0);
 
 	if (rmtp != nullptr) {
 		rmtp->tv_sec  = 0;

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/audio.cpp

@@ -1,6 +1,5 @@
 #include "libs/audio.h"
 
-#include <SDL3/SDL.h>
 #include "common/assert.h"
 #include "common/emulatorConfig.h"
 #include "common/logging/log.h"
@@ -12,7 +11,9 @@
 #include "libs/errno.h"
 #include "libs/libs.h"
 
+#include <SDL3/SDL.h>
 #include <algorithm>
+#include <array>
 #include <cstring>
 #include <limits>
 #include <magic_enum.hpp>
@@ -37,6 +38,24 @@ constexpr uint64_t AUDIO_OUT_TARGET_LATENCY_US = 40000;
 constexpr int      AUDIO_IN_SILENT_STATE_DEVICE_NONE = 0x1;
 constexpr uint32_t AUDIO_IN_GRAIN_MAX_ASYNC          = 384;
 
+struct OutputClock {
+	uint64_t next_time = 0;
+	uint64_t remainder = 0;
+
+	uint64_t Advance(uint64_t now, uint32_t frames, uint32_t rate) {
+		// Preserve small scheduling errors; discard timing debt after a discontinuity.
+		if (next_time == 0 || (now > next_time && now - next_time > 100000)) {
+			next_time = now;
+			remainder = 0;
+		}
+		const auto deadline = next_time;
+		const auto duration = 1000000ULL * frames + remainder;
+		next_time += duration / rate;
+		remainder = duration % rate;
+		return deadline;
+	}
+};
+
 static bool audio_out_port_type_is_valid(int type) {
 	return (type >= AUDIO_OUT_PORT_TYPE_MAIN && type <= AUDIO_OUT_PORT_TYPE_PADSPK) ||
 	       type == AUDIO_OUT_PORT_TYPE_VIBRATION || type == AUDIO_OUT_PORT_TYPE_AUDIO3D ||
@@ -98,17 +117,17 @@ public:
 
 private:
 	struct PortOut {
-		bool     used             = false;
-		int      type             = 0;
-		uint32_t samples_num      = 0;
-		uint32_t freq             = 0;
-		Format   format           = Format::Unknown;
-		uint64_t last_output_time = 0;
-		bool     queue_primed     = false;
-		int      channels_num     = 0;
-		int      volume[12]       = {};
+		bool        used        = false;
+		int         type        = 0;
+		uint32_t    samples_num = 0;
+		uint32_t    freq        = 0;
+		Format      format      = Format::Unknown;
+		OutputClock clock;
+		bool        queue_primed = false;
+		int         channels_num = 0;
+		int         volume[12]   = {};
 
-		SDL_AudioStream*                     stream  = nullptr;
+		SDL_AudioStream*                      stream  = nullptr;
 		Controller::DualSenseHaptics::Stream* haptics = nullptr;
 	};
 
@@ -376,10 +395,7 @@ bool Audio::QueueSdlAudio(PortOut* port, const void* data, bool blocking) {
 		        : 2u;
 		min_queued_size           = prepared_size * std::clamp(buffers, 2u, 16u);
 		const auto wait_start      = LibKernel::KernelGetProcessTime();
-		auto queued                = SDL_GetAudioStreamQueued(port->stream);
-		if (queued < static_cast<int>(prepared_size)) {
-			port->queue_primed = false;
-		}
+		auto       queued          = SDL_GetAudioStreamQueued(port->stream);
 		while (queued > static_cast<int>(min_queued_size)) {
 			if (LibKernel::KernelGetProcessTime() - wait_start > 200000) {
 				SDL_ClearAudioStream(port->stream);
@@ -389,19 +405,20 @@ bool Audio::QueueSdlAudio(PortOut* port, const void* data, bool blocking) {
 			Common::Thread::SleepMicro(1000);
 			queued = SDL_GetAudioStreamQueued(port->stream);
 		}
-		if (port->queue_primed) {
-			const auto next_time = port->last_output_time + buffer_us;
-			const auto now       = LibKernel::KernelGetProcessTime();
-			if (next_time > now) {
-				Common::Thread::SleepMicro(next_time - now);
-			}
+		if (queued < static_cast<int>(prepared_size)) {
+			port->queue_primed = false;
 		}
 	}
 
 	if (!SDL_PutAudioStreamData(port->stream, prepared_data, static_cast<int>(prepared_size))) {
 		LOGF("AudioOut: SDL_PutAudioStreamData failed: %s\n", SDL_GetError());
+		port->queue_primed = false;
 		return false;
 	}
+	// Unpaced priming must not advance the playback deadline into the future.
+	if (!port->queue_primed) {
+		port->clock = {};
+	}
 	if (blocking && !port->queue_primed &&
 	    SDL_GetAudioStreamQueued(port->stream) >= static_cast<int>(min_queued_size)) {
 		port->queue_primed = true;
@@ -411,18 +428,21 @@ bool Audio::QueueSdlAudio(PortOut* port, const void* data, bool blocking) {
 }
 
 Audio::Id Audio::AudioOutOpen(int type, uint32_t samples_num, uint32_t freq, Format format) {
+	if (freq == 0) {
+		return Id::Invalid();
+	}
+
 	Common::LockGuard lock(m_mutex);
 
 	for (int id = 0; id < OUT_PORTS_MAX; id++) {
 		if (!m_out_ports[id].used) {
 			auto& port = m_out_ports[id];
 
-			port.used             = true;
-			port.type             = type;
-			port.samples_num      = samples_num;
-			port.freq             = freq;
-			port.format           = format;
-			port.last_output_time = 0;
+			port.used        = true;
+			port.type        = type;
+			port.samples_num = samples_num;
+			port.freq        = freq;
+			port.format      = format;
 
 			switch (format) {
 				case Format::Signed16bitMono:
@@ -524,61 +544,68 @@ bool Audio::AudioOutSetVolume(Id handle, uint32_t bitflag, const int* volume) {
 }
 
 uint32_t Audio::AudioOutOutputs(OutputParam* params, uint32_t num, bool blocking) {
-	EXIT_NOT_IMPLEMENTED(num == 0);
+	EXIT_NOT_IMPLEMENTED(num == 0 || num > OUT_PORTS_MAX);
 	EXIT_NOT_IMPLEMENTED(!AudioOutValid(params[0].handle));
 
-	const auto& first_port = m_out_ports[params[0].handle.GetId()];
-
-	uint64_t block_time   = (1000000 * first_port.samples_num) / first_port.freq;
-	uint64_t current_time = LibKernel::KernelGetProcessTime();
-
-	uint64_t max_wait_time = 0;
-
-	for (uint32_t i = 0; i < num; i++) {
-		uint64_t next_time = m_out_ports[params[i].handle.GetId()].last_output_time + block_time;
-		uint64_t wait_time = (next_time > current_time ? next_time - current_time : 0);
-		max_wait_time      = (wait_time > max_wait_time ? wait_time : max_wait_time);
-	}
-
-	bool any_port_has_device = false;
-	for (uint32_t i = 0; i < num; i++) {
-		if (m_out_ports[params[i].handle.GetId()].stream != nullptr) {
-			any_port_has_device = true;
-			break;
-		}
-	}
-
-	// One real output device is enough to pace the whole synchronized batch. Applying the fallback
-	// when a vibration port is present would rate-limit the device-backed ports to exactly 1x and
-	// prevent their SDL queues from building an underrun cushion.
-	if (blocking && max_wait_time != 0 && !any_port_has_device) {
-		Common::Thread::SleepMicro(max_wait_time);
-	}
+	const auto&                     first_port = m_out_ports[params[0].handle.GetId()];
+	std::array<bool, OUT_PORTS_MAX> paced {};
+	bool                            any_device = false;
+	uint64_t                        deadline   = 0;
 
 	for (uint32_t i = 0; i < num; i++) {
 		auto& port = m_out_ports[params[i].handle.GetId()];
 
 		uint64_t controller_queued_us = 0;
+		bool controller_uses_bluetooth = false;
 		if (port.haptics != nullptr) {
 			// Keep the stream alive against close and volume changes.
 			Common::LockGuard lock(m_mutex);
 			controller_queued_us = Controller::DualSenseHaptics::Queue(
 			    port.haptics, Controller::GetActiveControllerId(), params[i].data, port.samples_num,
 			    static_cast<uint32_t>(port.channels_num), FormatIsFloat(port.format), port.volume);
+			controller_uses_bluetooth =
+			    Controller::DualSenseHaptics::UsesBluetooth(port.haptics);
 		}
 		if (controller_queued_us == 0) {
 			// No DualSense took it (e.g. it was unplugged); a pad speaker port plays on the main
 			// output instead, and a vibration port has none.
-			QueueSdlAudio(&port, params[i].data, blocking);
-		} else if (blocking && port.type == AUDIO_OUT_PORT_TYPE_PADSPK &&
-		           controller_queued_us > AUDIO_OUT_TARGET_LATENCY_US) {
-			// Vibration never paces output; the speaker is audible, so pace it like other ports.
-			Common::Thread::SleepMicro(controller_queued_us - AUDIO_OUT_TARGET_LATENCY_US);
+			if (QueueSdlAudio(&port, params[i].data, blocking)) {
+				any_device = true;
+				paced[i]   = port.queue_primed;
+			}
+		} else {
+			port.queue_primed = false;
+			// Bluetooth HID has its own sender. Let the sample clock pace a batch
+			// without another audio device, instead of waiting on the HID queue.
+			if (port.type == AUDIO_OUT_PORT_TYPE_PADSPK && !controller_uses_bluetooth) {
+				any_device = true;
+				if (blocking && controller_queued_us > AUDIO_OUT_TARGET_LATENCY_US) {
+					// Limit a stalled USB endpoint to one audio block of waiting.
+					const uint64_t block_us = 1000000ULL * port.samples_num / port.freq;
+					deadline = std::max(deadline, LibKernel::KernelGetProcessTime() +
+					                                  std::min(controller_queued_us - AUDIO_OUT_TARGET_LATENCY_US,
+					                                           block_us));
+				}
+			}
 		}
 	}
 
+	// Submit the whole batch before pacing it. Audible devices own its cadence while
+	// priming; silent/vibration-only batches use the same sample clock as SDL output.
+	const auto now = LibKernel::KernelGetProcessTime();
 	for (uint32_t i = 0; i < num; i++) {
-		m_out_ports[params[i].handle.GetId()].last_output_time = LibKernel::KernelGetProcessTime();
+		auto& port = m_out_ports[params[i].handle.GetId()];
+		if (blocking && (paced[i] || !any_device)) {
+			deadline = std::max(deadline, port.clock.Advance(now, port.samples_num, port.freq));
+		} else {
+			port.clock = {};
+		}
+		if (!blocking) {
+			port.queue_primed = false;
+		}
+	}
+	if (deadline > now) {
+		Common::Thread::SleepMicro(deadline - now);
 	}
 
 	return first_port.samples_num;
@@ -826,6 +853,9 @@ int KYTY_SYSV_ABI AudioOutOpen(int user_id, int type, int index, uint32_t len, u
 	if (!audio_out_port_type_is_valid(type)) {
 		return AUDIO_OUT_ERROR_INVALID_PORT_TYPE;
 	}
+	if (freq == 0) {
+		return AUDIO_OUT_ERROR_INVALID_SAMPLE_FREQ;
+	}
 	EXIT_NOT_IMPLEMENTED(index != 0);
 
 	Audio::Format format       = Audio::Format::Unknown;
@@ -935,6 +965,9 @@ int KYTY_SYSV_ABI AudioOutOutputs(AudioOutOutputParam* param, uint32_t num) {
 	PRINT_NAME();
 
 	EXIT_NOT_IMPLEMENTED(param == nullptr);
+	if (num == 0 || num > Audio::OUT_PORTS_MAX) {
+		return AUDIO_OUT_ERROR_INVALID_SIZE;
+	}
 
 	Audio::OutputParam params[Audio::OUT_PORTS_MAX];
 

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/dualSenseBluetooth.cpp

@@ -0,0 +1,467 @@
+#include "libs/dualSenseBluetooth.h"
+
+#include "common/logging/log.h"
+
+#include <SDL3/SDL.h>
+#include <SDL3/SDL_hidapi.h>
+#include <algorithm>
+#include <array>
+#include <atomic>
+#include <chrono>
+#include <cmath>
+#include <condition_variable>
+#include <map>
+#include <memory>
+#include <mutex>
+#include <opus.h>
+#include <string>
+#include <vector>
+
+namespace Libs::Controller::DualSenseBluetooth {
+
+namespace {
+// Report 0x39 and its sub-packets are described by awalol/DS5Dongle (MIT):
+// https://github.com/awalol/DS5Dongle
+// Each report plays two 32-frame stereo haptic blocks at 3 kHz (21.333 ms).
+constexpr uint64_t PERIOD_NS   = 64000000000ULL / 3000;
+constexpr int      HAPTIC_RATE = 3000;
+constexpr int      HAPTIC_REPORT_FLOATS = 2 * 32 * 2;
+// Two 480-frame Opus packets occupy the same 21.333 ms slot. Resampling
+// 48 kHz input to 45 kHz before encoding at 48 kHz compensates for playback.
+constexpr int    SPEAKER_RATE = 45000;
+constexpr int    OPUS_FRAMES  = 480;
+constexpr int    OPUS_BYTES   = 200;
+constexpr size_t REPORT_BYTES = 547;
+constexpr int    MAX_QUEUE_MS = 80;
+constexpr uint64_t COALESCE_NS = 1000000;
+
+struct Session;
+std::mutex                              g_mutex;
+std::condition_variable                 g_wake;
+std::map<int, std::shared_ptr<Session>> g_sessions;
+SDL_Thread*                             g_thread   = nullptr;
+bool                                    g_stopping = false;
+void (*g_tick)()                                   = nullptr;
+
+template <typename Char>
+std::string Address(const Char* serial) {
+	std::string address;
+	if (serial == nullptr) {
+		return address;
+	}
+	for (; *serial != 0; ++serial) {
+		if (*serial >= '0' && *serial <= '9') {
+			address.push_back(static_cast<char>(*serial));
+		} else if (*serial >= 'a' && *serial <= 'f') {
+			address.push_back(static_cast<char>(*serial));
+		} else if (*serial >= 'A' && *serial <= 'F') {
+			address.push_back(static_cast<char>(*serial - 'A' + 'a'));
+		} else if (*serial != ':' && *serial != '-') {
+			return {};
+		}
+	}
+	return address.size() == 12 ? address : std::string {};
+}
+
+SDL_hid_device* OpenDevice(int controller) {
+	SDL_LockJoysticks();
+	auto*      pad      = SDL_GetGamepadFromID(static_cast<SDL_JoystickID>(controller));
+	const bool wireless = pad != nullptr && SDL_GetGamepadType(pad) == SDL_GAMEPAD_TYPE_PS5 &&
+	                      SDL_GetGamepadConnectionState(pad) == SDL_JOYSTICK_CONNECTION_WIRELESS;
+	const auto vendor  = wireless ? SDL_GetGamepadVendor(pad) : 0;
+	const auto product = wireless ? SDL_GetGamepadProduct(pad) : 0;
+	const auto address = wireless ? Address(SDL_GetGamepadSerial(pad)) : std::string {};
+	SDL_UnlockJoysticks();
+	if (vendor != 0x054c || (product != 0x0ce6 && product != 0x0df2) || address.empty()) {
+		return nullptr;
+	}
+	auto*       devices = SDL_hid_enumerate(vendor, product);
+	std::string path;
+	bool        ambiguous = false;
+	for (auto* device = devices; device != nullptr; device = device->next) {
+		if (device->bus_type == SDL_HID_API_BUS_BLUETOOTH && device->path != nullptr &&
+		    Address(device->serial_number) == address &&
+		    (device->usage_page == 0 || (device->usage_page == 1 && device->usage == 5))) {
+			if (!path.empty()) {
+				ambiguous = true;
+				break;
+			}
+			path = device->path;
+		}
+	}
+	SDL_hid_free_enumeration(devices);
+	return !ambiguous && !path.empty() ? SDL_hid_open_path(path.c_str()) : nullptr;
+}
+
+void FinishReport(std::array<uint8_t, REPORT_BYTES>& report) {
+	const uint8_t prefix = 0xa2;
+	auto          crc    = SDL_crc32(0, &prefix, 1);
+	crc                  = SDL_crc32(crc, report.data(), report.size() - 4);
+	for (size_t i = 0; i < 4; i++) {
+		report[report.size() - 4 + i] = static_cast<uint8_t>(crc >> (8 * i));
+	}
+}
+
+struct Session {
+	SDL_hid_device*       hid     = nullptr;
+	OpusEncoder*          encoder = nullptr;
+	std::vector<Stream*>  streams;
+	std::mutex            io_mutex;
+	std::atomic<bool>     active {true};
+	std::atomic<bool>     failed {false};
+	std::atomic<uint64_t> generation {0};
+	uint64_t              next_send        = 0;
+	uint64_t              pending_deadline = 0;
+	uint8_t               sequence  = 0;
+	uint8_t               counter   = 0;
+
+	~Session() {
+		if (encoder != nullptr) {
+			opus_encoder_destroy(encoder);
+		}
+		if (hid != nullptr) {
+			SDL_hid_close(hid);
+		}
+	}
+
+	bool CreateEncoder() {
+		if (encoder != nullptr) {
+			return true;
+		}
+		int error = OPUS_OK;
+		encoder   = opus_encoder_create(48000, 2, OPUS_APPLICATION_AUDIO, &error);
+		if (encoder == nullptr || error != OPUS_OK) {
+			return false;
+		}
+		if (opus_encoder_ctl(encoder, OPUS_SET_BITRATE(OPUS_BYTES * 8 * 100)) == OPUS_OK &&
+		    opus_encoder_ctl(encoder, OPUS_SET_VBR(0)) == OPUS_OK &&
+		    opus_encoder_ctl(encoder, OPUS_SET_COMPLEXITY(0)) == OPUS_OK) {
+			return true;
+		}
+		opus_encoder_destroy(encoder);
+		encoder = nullptr;
+		return false;
+	}
+};
+} // namespace
+
+struct Stream {
+	std::shared_ptr<Session> session;
+	SDL_AudioStream*         audio   = nullptr;
+	uint32_t                 freq    = 0;
+	bool                     speaker = false;
+};
+
+namespace {
+template <size_t N>
+void Mix(SDL_AudioStream* stream, std::array<float, N>& mixed) {
+	std::array<float, N> pcm {};
+	const int            bytes = SDL_GetAudioStreamData(stream, pcm.data(), sizeof(pcm));
+	for (int i = 0; i < std::max(0, bytes) / static_cast<int>(sizeof(float)); i++) {
+		mixed[i] += pcm[i];
+	}
+}
+
+struct PendingReport {
+	std::shared_ptr<Session>          session;
+	uint64_t                          generation;
+	std::array<uint8_t, REPORT_BYTES> data {};
+	std::array<float, 2 * OPUS_FRAMES * 2> audio {};
+	bool                              speaker = false;
+};
+
+bool PrepareReport(Session& session, PendingReport& pending, uint64_t now) {
+	bool has_data      = false;
+	bool ready         = false;
+	bool speaker_ready = false;
+	for (const auto* stream: session.streams) {
+		const bool available = SDL_GetAudioStreamAvailable(stream->audio) > 0;
+		// Both payloads cover 64 / 3000 seconds. Measure input duration so SDL's
+		// resampler lookahead does not make a complete block appear partial.
+		const auto frames =
+		    (static_cast<uint64_t>(stream->freq) * (HAPTIC_REPORT_FLOATS / 2) + HAPTIC_RATE - 1) /
+		    HAPTIC_RATE;
+		const bool full = available && SDL_GetAudioStreamQueued(stream->audio) >=
+		                                   static_cast<int>(frames * 2 * sizeof(float));
+		has_data |= available;
+		ready |= full;
+		speaker_ready |= stream->speaker && full;
+		pending.speaker |= stream->speaker;
+	}
+	if (!has_data) {
+		session.pending_deadline = 0;
+		return false;
+	}
+	if (!ready || (!speaker_ready && (pending.speaker || session.next_send == 0))) {
+		// Accumulate short blocks for at most one report period. A full stream
+		// must not wait on a partial peer; only briefly coalesce haptics with
+		// speaker audio arriving in the same emulated batch.
+		const auto deadline = now + (ready ? COALESCE_NS : PERIOD_NS);
+		session.pending_deadline =
+		    session.pending_deadline == 0 ? deadline : std::min(session.pending_deadline, deadline);
+		if (now < session.pending_deadline) {
+			return false;
+		}
+	}
+	session.pending_deadline = 0;
+	std::array<float, HAPTIC_REPORT_FLOATS> haptics {};
+	for (auto* stream: session.streams) {
+		if (stream->speaker) {
+			Mix(stream->audio, pending.audio);
+		} else {
+			Mix(stream->audio, haptics);
+		}
+	}
+	auto& report     = pending.data;
+	report[0]        = 0x39;
+	report[1]        = static_cast<uint8_t>(session.sequence << 4);
+	session.sequence = (session.sequence + 1) & 0x0f;
+	report[2]        = 0x91;
+	report[3]        = 6;
+	report[4]        = 0x7e;
+	std::fill_n(report.begin() + 5, 4, 48);
+	session.counter += 2;
+	report[9]  = session.counter;
+	report[10] = 0xd2;
+	report[11] = 64;
+	for (size_t i = 0; i < haptics.size(); i++) {
+		const auto value = std::isfinite(haptics[i]) ? haptics[i] : 0.0f;
+		report[12 + i]   = static_cast<uint8_t>(
+            static_cast<int>(std::lround(std::clamp(value, -1.0f, 1.0f) * 127)));
+	}
+	if (pending.speaker) {
+		report[140] = 0xd3;
+		report[141] = OPUS_BYTES;
+	}
+	return true;
+}
+
+bool EncodeReport(PendingReport& pending) {
+	if (!pending.session->active || pending.session->failed) {
+		return false;
+	}
+	if (pending.speaker) {
+		for (auto& sample: pending.audio) {
+			sample = std::isfinite(sample) ? std::clamp(sample, -1.0f, 1.0f) : 0.0f;
+		}
+		for (int packet = 0; packet < 2; packet++) {
+			if (opus_encode_float(pending.session->encoder,
+			                      pending.audio.data() + packet * OPUS_FRAMES * 2, OPUS_FRAMES,
+			                      pending.data.data() + 142 + packet * OPUS_BYTES,
+			                      OPUS_BYTES) != OPUS_BYTES) {
+				pending.session->failed = true;
+				return false;
+			}
+		}
+	}
+	FinishReport(pending.data);
+	return true;
+}
+
+std::vector<PendingReport> Prepare(uint64_t now) {
+	std::vector<PendingReport> pending;
+	{
+		std::lock_guard lock(g_mutex);
+		for (auto& [id, session]: g_sessions) {
+			if (session->failed || now < session->next_send) {
+				continue;
+			}
+			PendingReport report {session, session->generation.load()};
+			if (PrepareReport(*session, report, now)) {
+				pending.push_back(std::move(report));
+				// Regain timing after a late HID write without sending reports in a
+				// burst. Reset the cadence only on the first report.
+				session->next_send = session->next_send == 0
+				                         ? now + PERIOD_NS
+				                         : std::max(session->next_send + PERIOD_NS,
+				                                    now + PERIOD_NS / 2);
+			}
+		}
+	}
+	// Opus encoding can take milliseconds; the emulation audio thread must be able
+	// to queue its next block while this work runs.
+	pending.erase(std::remove_if(pending.begin(), pending.end(),
+	                             [](PendingReport& report) { return !EncodeReport(report); }),
+	              pending.end());
+	return pending;
+}
+
+void Send(const std::vector<PendingReport>& pending) {
+	for (const auto& report: pending) {
+		auto&           session = *report.session;
+		std::lock_guard io_lock(session.io_mutex);
+		if (!session.active || session.failed || report.generation != session.generation) {
+			continue;
+		}
+		if (SDL_hid_write(session.hid, report.data.data(), report.data.size()) !=
+		    static_cast<int>(report.data.size())) {
+			session.failed = true;
+			LOGF("DualSenseBluetooth: audio HID write failed: %s\n", SDL_GetError());
+		}
+	}
+}
+
+int SDLCALL Sender(void*) {
+	std::unique_lock lock(g_mutex);
+	while (!g_stopping) {
+		auto tick = g_tick;
+		lock.unlock();
+		auto pending = Prepare(SDL_GetTicksNS());
+		Send(pending);
+		if (tick != nullptr) {
+			tick();
+		}
+		lock.lock();
+		if (g_sessions.empty()) {
+			g_wake.wait(lock, [] { return g_stopping || !g_sessions.empty(); });
+		} else {
+			const auto now = SDL_GetTicksNS();
+			uint64_t   wait_ns = 10000000; // Check rumble expiry at most 10 ms late.
+			for (const auto& [id, session]: g_sessions) {
+				if (session->next_send > now) {
+					wait_ns = std::min(wait_ns, session->next_send - now);
+				}
+				if (session->pending_deadline > now) {
+					wait_ns = std::min(wait_ns, session->pending_deadline - now);
+				}
+			}
+			g_wake.wait_for(lock, std::chrono::nanoseconds(wait_ns));
+		}
+	}
+	return 0;
+}
+} // namespace
+
+Stream* Open(uint32_t freq, bool speaker, int controller, void (*tick)()) {
+	if (freq == 0 || freq > 192000) {
+		return nullptr;
+	}
+	std::lock_guard lock(g_mutex);
+	if (g_stopping) {
+		return nullptr;
+	}
+	auto                     it = g_sessions.find(controller);
+	std::shared_ptr<Session> session;
+	if (it != g_sessions.end() && !it->second->failed) {
+		session = it->second;
+	} else {
+		session      = std::make_shared<Session>();
+		session->hid = OpenDevice(controller);
+		if (session->hid == nullptr) {
+			return nullptr;
+		}
+	}
+	if (speaker && !session->CreateEncoder()) {
+		return nullptr;
+	}
+	const SDL_AudioSpec source {SDL_AUDIO_F32, 2, static_cast<int>(freq)};
+	const SDL_AudioSpec target {SDL_AUDIO_F32, 2, speaker ? SPEAKER_RATE : HAPTIC_RATE};
+	auto*               audio = SDL_CreateAudioStream(&source, &target);
+	if (audio == nullptr) {
+		return nullptr;
+	}
+	if (g_thread == nullptr) {
+		g_thread = SDL_CreateThread(Sender, "DualSense Bluetooth audio", nullptr);
+		if (g_thread == nullptr) {
+			SDL_DestroyAudioStream(audio);
+			return nullptr;
+		}
+	}
+	g_tick       = tick;
+	auto* stream = new Stream {session, audio, freq, speaker};
+	session->streams.push_back(stream);
+	session->pending_deadline = 0;
+	g_sessions[controller] = session;
+	g_wake.notify_one();
+	return stream;
+}
+
+uint64_t Queue(Stream* stream, const float* stereo, uint32_t frames) {
+	if (stream == nullptr || stereo == nullptr || frames == 0) {
+		return 0;
+	}
+	std::lock_guard lock(g_mutex);
+	if (!stream->session->active || stream->session->failed) {
+		return 0;
+	}
+	const uint32_t max_frames = stream->freq * MAX_QUEUE_MS / 1000;
+	if (frames > max_frames) {
+		stereo += static_cast<size_t>(frames - max_frames) * 2;
+		frames = max_frames;
+	}
+	int       queued = std::max(0, SDL_GetAudioStreamQueued(stream->audio));
+	const int bytes  = static_cast<int>(frames * 2 * sizeof(float));
+	if (queued + bytes > static_cast<int>(max_frames * 2 * sizeof(float))) {
+		// Discard complete old report slots, preserving the newest queued audio.
+		// Clearing the entire stream here creates an audible gap after a brief
+		// Windows Bluetooth write stall.
+		std::array<float, 2 * OPUS_FRAMES * 2> discarded {};
+		const int discard_bytes = stream->speaker
+		                              ? static_cast<int>(sizeof(discarded))
+		                              : static_cast<int>(HAPTIC_REPORT_FLOATS * sizeof(float));
+		const int target = static_cast<int>(max_frames * 2 * sizeof(float)) - bytes;
+		while (queued > target) {
+			if (SDL_GetAudioStreamData(stream->audio, discarded.data(), discard_bytes) <= 0) {
+				SDL_ClearAudioStream(stream->audio);
+				break;
+			}
+			queued = std::max(0, SDL_GetAudioStreamQueued(stream->audio));
+		}
+	}
+	queued = std::max(0, SDL_GetAudioStreamQueued(stream->audio));
+	if (!SDL_PutAudioStreamData(stream->audio, stereo, bytes)) {
+		return 0;
+	}
+	g_wake.notify_one();
+	return static_cast<uint64_t>(queued + bytes) * 1000000 / (stream->freq * 2 * sizeof(float));
+}
+
+void Close(Stream* stream) {
+	if (stream == nullptr) {
+		return;
+	}
+	auto session = stream->session;
+	{
+		std::lock_guard lock(g_mutex);
+		auto&           streams = session->streams;
+		streams.erase(std::remove(streams.begin(), streams.end(), stream), streams.end());
+		session->generation++;
+		session->pending_deadline = 0;
+		if (streams.empty()) {
+			session->active = false;
+			for (auto it = g_sessions.begin(); it != g_sessions.end(); ++it) {
+				if (it->second == session) {
+					g_sessions.erase(it);
+					break;
+				}
+			}
+		}
+		SDL_DestroyAudioStream(stream->audio);
+		delete stream;
+	}
+	// Drain a write already in flight before the caller changes speaker routing.
+	std::lock_guard io_lock(session->io_mutex);
+}
+
+void Shutdown() {
+	SDL_Thread* thread;
+	{
+		std::lock_guard lock(g_mutex);
+		g_stopping = true;
+		for (auto& [id, session]: g_sessions) {
+			session->active = false;
+		}
+		thread = g_thread;
+		g_wake.notify_one();
+	}
+	if (thread != nullptr) {
+		SDL_WaitThread(thread, nullptr);
+	}
+	std::lock_guard lock(g_mutex);
+	g_sessions.clear();
+	g_thread   = nullptr;
+	g_tick     = nullptr;
+	g_stopping = false;
+}
+
+} // namespace Libs::Controller::DualSenseBluetooth

src/libs/dualSenseBluetooth.h

@@ -0,0 +1,14 @@
+#ifndef KYTY_LIBS_DUALSENSE_BLUETOOTH_H_
+#define KYTY_LIBS_DUALSENSE_BLUETOOTH_H_
+
+#include <cstdint>
+
+namespace Libs::Controller::DualSenseBluetooth {
+struct Stream;
+Stream*  Open(uint32_t frequency, bool speaker, int controller, void (*tick)());
+uint64_t Queue(Stream* stream, const float* samples, uint32_t frames);
+void     Close(Stream* stream);
+void     Shutdown();
+} // namespace Libs::Controller::DualSenseBluetooth
+
+#endif

src/libs/dualSenseHaptics.cpp

@@ -2,6 +2,7 @@
 
 #include "common/logging/log.h"
 #include "common/threads.h"
+#include "libs/dualSenseBluetooth.h"
 
 #include <SDL3/SDL.h>
 #include <algorithm>
@@ -13,18 +14,20 @@
 namespace Libs::Controller::DualSenseHaptics {
 
 struct Stream {
-	uint32_t           freq       = 0;
-	bool               speaker    = false;
-	int                controller = -1;
-	SDL_AudioStream*   sdl        = nullptr;
-	SDL_AudioDeviceID  device     = 0;
-	uint64_t           next_check = 0;
-	std::vector<float> buffer;
+	uint32_t                    freq       = 0;
+	bool                        speaker    = false;
+	int                         controller = -1;
+	SDL_AudioStream*            sdl        = nullptr;
+	DualSenseBluetooth::Stream* bluetooth  = nullptr;
+	SDL_AudioDeviceID           device     = 0;
+	uint64_t                    next_check = 0;
+	std::vector<float>          buffer;
 };
 
 namespace {
 
-constexpr int      FRAME_BYTES = 4 * sizeof(float);
+constexpr int      FRAME_BYTES  = 4 * sizeof(float);
+constexpr uint32_t MAX_QUEUE_MS = 80;
 Common::Mutex      g_mutex;
 int                g_controller      = -1;
 int                g_haptics_streams = 0;
@@ -107,11 +110,22 @@ void SDLCALL UpdateRumble(void*, SDL_AudioStream*, int, int) {
 	}
 }
 
+void UpdateBluetoothRumble() {
+	UpdateRumble(nullptr, nullptr, 0, 0);
+}
+
 void CloseDevice(Stream* stream) {
+	const bool had_device = stream->sdl != nullptr || stream->bluetooth != nullptr;
+	if (stream->bluetooth != nullptr) {
+		DualSenseBluetooth::Close(stream->bluetooth);
+		stream->bluetooth = nullptr;
+	}
 	if (stream->sdl != nullptr) {
 		// Destroy outside g_mutex: SDL waits for any running get callback.
 		SDL_DestroyAudioStream(stream->sdl);
 		stream->sdl = nullptr;
+	}
+	if (had_device) {
 		Common::LockGuard lock(g_mutex);
 		if (!stream->speaker && --g_haptics_streams == 0 && g_haptics_until != 0) {
 			g_haptics_until = 0;
@@ -127,26 +141,34 @@ void CloseDevice(Stream* stream) {
 	}
 }
 
+bool IsWireless(int controller) {
+	SDL_LockJoysticks();
+	auto*      pad = SDL_GetGamepadFromID(static_cast<SDL_JoystickID>(controller));
+	const bool wireless =
+	    pad != nullptr && SDL_GetGamepadConnectionState(pad) == SDL_JOYSTICK_CONNECTION_WIRELESS;
+	SDL_UnlockJoysticks();
+	return wireless;
+}
+
 bool CanUseDevice(int controller) {
 	// SDL exposes no portable association between a gamepad and its USB audio endpoint.
 	// Require one wired DualSense; guessing with multiple pads could play on the wrong one.
 	SDL_LockJoysticks();
-	auto* pad = SDL_GetGamepadFromID(static_cast<SDL_JoystickID>(controller));
-	bool  usable =
-	    pad != nullptr && SDL_GetGamepadConnectionState(pad) == SDL_JOYSTICK_CONNECTION_WIRED;
-	int   count           = 0;
-	auto* pads            = SDL_GetGamepads(&count);
-	int   dualsense_count = 0;
-	bool  found           = false;
+	int   count                 = 0;
+	auto* pads                  = SDL_GetGamepads(&count);
+	int   wired_dualsense_count = 0;
+	bool  found                 = false;
 	for (int i = 0; i < count; i++) {
-		if (SDL_GetGamepadTypeForID(pads[i]) == SDL_GAMEPAD_TYPE_PS5) {
-			dualsense_count++;
+		auto* candidate = SDL_GetGamepadFromID(pads[i]);
+		if (candidate != nullptr && SDL_GetGamepadTypeForID(pads[i]) == SDL_GAMEPAD_TYPE_PS5 &&
+		    SDL_GetGamepadConnectionState(candidate) == SDL_JOYSTICK_CONNECTION_WIRED) {
+			wired_dualsense_count++;
 			found |= pads[i] == static_cast<SDL_JoystickID>(controller);
 		}
 	}
 	SDL_free(pads);
 	SDL_UnlockJoysticks();
-	return usable && found && dualsense_count == 1;
+	return found && wired_dualsense_count == 1;
 }
 
 SDL_AudioDeviceID FindDevice() {
@@ -169,32 +191,42 @@ SDL_AudioDeviceID FindDevice() {
 	return device;
 }
 
-void RefreshDevice(Stream* stream, int controller) {
-	if (stream->controller != controller) {
+void RefreshDevice(Stream* stream, int controller, bool wireless) {
+	if (stream->controller != controller || (wireless && stream->sdl != nullptr) ||
+	    (!wireless && stream->bluetooth != nullptr)) {
 		CloseDevice(stream);
 		stream->controller = controller;
 		stream->next_check = 0;
 	}
 	const auto now = SDL_GetTicks();
-	if (now < stream->next_check) {
+	if (stream->bluetooth != nullptr || now < stream->next_check) {
 		return;
 	}
 	stream->next_check = now + 2000;
-	const auto device  = FindDevice();
+	const auto device  = wireless ? 0 : FindDevice();
 	if (stream->device == device && stream->sdl != nullptr) {
 		return;
 	}
 	CloseDevice(stream);
 	stream->device = device;
-	if (device == 0) {
+	if (!wireless && device == 0) {
 		return;
 	}
-	const SDL_AudioSpec desired {SDL_AUDIO_F32, 4, static_cast<int>(stream->freq)};
-	auto*               sdl = SDL_OpenAudioDeviceStream(device, &desired, UpdateRumble, nullptr);
-	SDL_AudioSpec       actual {};
-	bool ready = sdl != nullptr &&
-	             SDL_GetAudioDeviceFormat(SDL_GetAudioStreamDevice(sdl), &actual, nullptr) &&
-	             actual.channels == 4 && SDL_ResumeAudioStreamDevice(sdl);
+	SDL_AudioStream*            sdl       = nullptr;
+	DualSenseBluetooth::Stream* bluetooth = nullptr;
+	bool                        ready;
+	if (wireless) {
+		bluetooth = DualSenseBluetooth::Open(stream->freq, stream->speaker, controller,
+		                                     UpdateBluetoothRumble);
+		ready     = bluetooth != nullptr;
+	} else {
+		const SDL_AudioSpec desired {SDL_AUDIO_F32, 4, static_cast<int>(stream->freq)};
+		SDL_AudioSpec       actual {};
+		sdl   = SDL_OpenAudioDeviceStream(device, &desired, UpdateRumble, nullptr);
+		ready = sdl != nullptr &&
+		        SDL_GetAudioDeviceFormat(SDL_GetAudioStreamDevice(sdl), &actual, nullptr) &&
+		        actual.channels == 4 && SDL_ResumeAudioStreamDevice(sdl);
+	}
 	if (ready) {
 		Common::LockGuard lock(g_mutex);
 		if (stream->speaker) {
@@ -207,14 +239,56 @@ void RefreshDevice(Stream* stream, int controller) {
 		}
 	}
 	if (ready) {
-		stream->sdl = sdl;
-		LOGF("DualSenseHaptics: playing on '%s'\n", SDL_GetAudioDeviceName(device));
+		stream->sdl       = sdl;
+		stream->bluetooth = bluetooth;
+		if (!wireless) {
+			LOGF("DualSenseHaptics: playing on '%s'\n", SDL_GetAudioDeviceName(device));
+		}
 	} else {
+		DualSenseBluetooth::Close(bluetooth);
 		SDL_DestroyAudioStream(sdl);
-		LOGF("DualSenseHaptics: cannot open quad output: %s\n", SDL_GetError());
+		if (!wireless) {
+			LOGF("DualSenseHaptics: cannot open quad output: %s\n", SDL_GetError());
+		}
 	}
 }
 
+uint64_t QueueUsbAudio(Stream* stream, uint32_t frames) {
+	auto      queued = std::max(0, SDL_GetAudioStreamQueued(stream->sdl));
+	const int bytes  = static_cast<int>(frames * FRAME_BYTES);
+	const int max_bytes =
+	    static_cast<int>(static_cast<uint64_t>(stream->freq) * FRAME_BYTES * MAX_QUEUE_MS / 1000);
+	if (queued + bytes > max_bytes) {
+		// Keep the newest sound after a brief device stall. Clearing the entire
+		// stream produces a gap that is much longer than one audio block.
+		SDL_AudioSpec              target {};
+		std::array<uint8_t, 16384> discarded {};
+		if (SDL_GetAudioStreamFormat(stream->sdl, nullptr, &target)) {
+			const int frame_bytes = SDL_AUDIO_FRAMESIZE(target);
+			if (frame_bytes > 0) {
+				const int chunk_frames =
+				    std::min<int>(static_cast<int>(discarded.size()) / frame_bytes,
+				                  static_cast<int>((static_cast<uint64_t>(frames) * target.freq +
+				                                    stream->freq - 1) /
+				                                   stream->freq));
+				while (queued + bytes > max_bytes && chunk_frames > 0 &&
+				       SDL_GetAudioStreamData(stream->sdl, discarded.data(),
+				                              chunk_frames * frame_bytes) > 0) {
+					queued = std::max(0, SDL_GetAudioStreamQueued(stream->sdl));
+				}
+			}
+		}
+		if (queued + bytes > max_bytes) {
+			SDL_ClearAudioStream(stream->sdl);
+			queued = 0;
+		}
+	}
+	if (!SDL_PutAudioStreamData(stream->sdl, stream->buffer.data(), bytes)) {
+		return 0;
+	}
+	return static_cast<uint64_t>(queued + bytes) * 1000000 / (stream->freq * FRAME_BYTES);
+}
+
 } // namespace
 
 Stream* Open(uint32_t freq, bool speaker) {
@@ -232,18 +306,30 @@ void Close(Stream* stream) {
 	}
 }
 
+bool UsesBluetooth(const Stream* stream) {
+	return stream != nullptr && stream->bluetooth != nullptr;
+}
+
 uint64_t Queue(Stream* stream, int controller, const void* data, uint32_t frames, uint32_t channels,
                bool is_float, const int* volume) {
 	if (stream == nullptr || data == nullptr || frames == 0 || channels == 0 || volume == nullptr) {
 		return 0;
 	}
-	if (!CanUseDevice(controller)) {
+	const auto max_frames =
+	    static_cast<uint32_t>(static_cast<uint64_t>(stream->freq) * MAX_QUEUE_MS / 1000);
+	const auto first_frame = frames > max_frames ? frames - max_frames : 0;
+	frames -= first_frame;
+	if (frames == 0) {
+		return 0;
+	}
+	const bool wireless = IsWireless(controller);
+	if (!wireless && !CanUseDevice(controller)) {
 		CloseDevice(stream);
 		stream->next_check = 0; // Look for the device as soon as a DualSense is active again.
 		return 0;
 	}
-	RefreshDevice(stream, controller);
-	if (stream->sdl == nullptr) {
+	RefreshDevice(stream, controller, wireless);
+	if (stream->sdl == nullptr && stream->bluetooth == nullptr) {
 		return 0;
 	}
 	if (stream->speaker) {
@@ -257,47 +343,42 @@ uint64_t Queue(Stream* stream, int controller, const void* data, uint32_t frames
 			return 0;
 		}
 	}
-	stream->buffer.assign(static_cast<size_t>(frames) * 4, 0.0f);
+	const uint32_t output_channels = wireless ? 2 : 4;
+	const uint32_t channel_offset  = wireless || stream->speaker ? 0 : 2;
+	stream->buffer.assign(static_cast<size_t>(frames) * output_channels, 0.0f);
 	bool audible = false;
 	for (uint32_t frame = 0; frame < frames; frame++) {
 		for (uint32_t ch = 0; ch < 2; ch++) {
 			const auto src_ch = channels == 1 ? 0 : ch;
-			const auto index  = static_cast<size_t>(frame) * channels + src_ch;
+			const auto index  = (static_cast<size_t>(first_frame) + frame) * channels + src_ch;
 			float      value  = is_float ? static_cast<const float*>(data)[index]
 			                             : static_cast<const int16_t*>(data)[index] / 32768.0f;
 			value *= volume[src_ch] / 32768.0f;
-			// The speaker plays the right front channel; the actuators play the back ones.
-			stream->buffer[static_cast<size_t>(frame) * 4 + ch + (stream->speaker ? 0 : 2)] = value;
-			audible |= std::fabs(value) > 1.0f / 1024;
+			// USB places the speaker in front and the actuators in back; Bluetooth takes stereo.
+			stream->buffer[static_cast<size_t>(frame) * output_channels + ch + channel_offset] =
+			    value;
+			audible |= std::isfinite(value) && std::fabs(value) > 1.0f / 1024;
 		}
 	}
-	auto queued = std::max(0, SDL_GetAudioStreamQueued(stream->sdl));
-	if (queued > static_cast<int>(stream->freq * FRAME_BYTES * 80 / 1000)) {
-		SDL_ClearAudioStream(stream->sdl);
-		queued = 0;
-	}
-	const int  bytes = static_cast<int>(frames * FRAME_BYTES);
 	const auto queued_us =
-	    static_cast<uint64_t>(queued + bytes) * 1000000 / (stream->freq * FRAME_BYTES);
+	    wireless ? DualSenseBluetooth::Queue(stream->bluetooth, stream->buffer.data(), frames)
+	             : QueueUsbAudio(stream, frames);
+	if (queued_us == 0) {
+		CloseDevice(stream);
+		stream->next_check = SDL_GetTicks() + 2000;
+		return 0;
+	}
 	if (audible && !stream->speaker) {
 		Common::LockGuard lock(g_mutex);
-		if (SDL_GetGamepadTypeForID(static_cast<SDL_JoystickID>(controller)) !=
-		    SDL_GAMEPAD_TYPE_PS5) {
-			return 0;
-		}
 		SelectController(controller);
-		const auto playback_now = SDL_GetTicks();
-		const bool starting     = g_haptics_until <= playback_now;
+		const auto now      = SDL_GetTicks();
+		const bool starting = g_haptics_until <= now;
 		g_haptics_until =
-		    std::max(g_haptics_until, playback_now + std::max<uint64_t>(queued_us / 1000, 250));
+		    std::max(g_haptics_until, now + std::max<uint64_t>(queued_us / 1000, 250));
 		if (starting) {
 			ApplyRumble();
 		}
 	}
-	if (!SDL_PutAudioStreamData(stream->sdl, stream->buffer.data(), bytes)) {
-		CloseDevice(stream);
-		return 0;
-	}
 	return queued_us;
 }
 
@@ -315,6 +396,7 @@ bool SetVibration(int controller, uint8_t large_motor, uint8_t small_motor) {
 }
 
 void Shutdown() {
+	DualSenseBluetooth::Shutdown();
 	Common::LockGuard lock(g_mutex);
 	SelectController(-1);
 	SelectSpeaker(-1);

src/libs/dualSenseHaptics.h

@@ -11,9 +11,11 @@ struct Stream;
 // ports on the actuators.
 Stream* Open(uint32_t freq, bool speaker);
 void    Close(Stream* stream);
+// True while this stream is routed through the asynchronous Bluetooth HID sender.
+bool UsesBluetooth(const Stream* stream);
 // Returns how long the queued audio still plays, in microseconds, or 0 when no DualSense took it
-// (the active pad has no usable USB audio device or its speaker cannot be routed).
-// Multiple DualSenses or matching audio endpoints are ambiguous and are not selected.
+// (the active pad has no usable controller audio device or its speaker cannot be routed).
+// Ambiguous USB or Bluetooth endpoints are not selected.
 uint64_t Queue(Stream* stream, int controller, const void* data, uint32_t frames, uint32_t channels,
                bool is_float, const int* volume);
 // Returns false for other gamepad types, which retain the normal rumble path.

src/libs/libAmpr.cpp

@@ -12,12 +12,18 @@
 
 #include <algorithm>
 #include <array>
+#include <atomic>
+#include <chrono>
+#include <condition_variable>
 #include <cstdint>
 #include <cstring>
+#include <deque>
 #include <limits>
 #include <mutex>
 #include <string>
+#include <thread>
 #include <unordered_map>
+#include <variant>
 #include <vector>
 
 #ifdef min
@@ -36,13 +42,6 @@ struct ResultBuffer {
 	uint32_t errorOffset = 0;
 };
 
-struct SubmissionState {
-	uint64_t command_buffer   = 0;
-	uint64_t result           = 0;
-	int32_t  execution_result = 0;
-	uint32_t error_offset     = 0;
-};
-
 struct ResolvedPathInfo {
 	int         result    = OK;
 	uint32_t    file_id   = 0xffffffffu;
@@ -53,7 +52,8 @@ struct ResolvedPathInfo {
 
 static std::mutex                                        g_mutex;
 static uint32_t                                          g_next_submission_id = 1;
-static std::unordered_map<uint32_t, SubmissionState>     g_submissions;
+static std::unordered_map<uint32_t, bool>                g_submissions;
+static std::condition_variable                           g_submission_cv;
 static std::unordered_map<uint32_t, std::string>         g_files;
 static std::unordered_map<uint32_t, uint64_t>            g_file_sizes;
 static std::unordered_map<std::string, ResolvedPathInfo> g_resolved_paths;
@@ -339,38 +339,35 @@ static int ResolvePathsCommon(const char* const* path_list, uint32_t count, uint
 	return results != nullptr ? static_cast<int>(success_count) : OK;
 }
 
-static uint32_t AllocateSubmissionId(uint64_t command_buffer, uint64_t result) {
+static uint32_t AllocateSubmissionId() {
 	std::scoped_lock lock(g_mutex);
-
 	auto id = g_next_submission_id++;
 	if (id == 0) {
 		id = g_next_submission_id++;
 	}
-	g_submissions[id] = SubmissionState {command_buffer, result};
+	g_submissions.emplace(id, false);
 	return id;
 }
 
-static void SetSubmissionResult(uint32_t submission_id, int32_t execution_result,
-                                uint32_t error_offset) {
-	std::scoped_lock lock(g_mutex);
-	auto             it = g_submissions.find(submission_id);
-	if (it != g_submissions.end()) {
-		it->second.execution_result = execution_result;
-		it->second.error_offset     = error_offset;
+static void SetSubmissionComplete(uint32_t submission_id) {
+	if (submission_id == 0) {
+		return;
 	}
+	std::scoped_lock lock(g_mutex);
+	g_submissions.at(submission_id) = true;
+	g_submission_cv.notify_all();
 }
 
-static bool CompleteSubmission(uint32_t submission_id, SubmissionState* state) {
-	std::scoped_lock lock(g_mutex);
-	auto             it = g_submissions.find(submission_id);
-	if (it == g_submissions.end()) {
+static bool CompleteSubmission(uint32_t submission_id) {
+	std::unique_lock lock(g_mutex);
+	if (!g_submissions.contains(submission_id)) {
 		return false;
 	}
-	if (state != nullptr) {
-		*state = it->second;
-	}
-	g_submissions.erase(it);
-	return true;
+	g_submission_cv.wait(lock, [submission_id] {
+		const auto it = g_submissions.find(submission_id);
+		return it == g_submissions.end() || it->second;
+	});
+	return g_submissions.erase(submission_id) != 0;
 }
 
 static int WriteResult(void* result, int32_t execution_result = 0, uint32_t error_offset = 0) {
@@ -389,8 +386,9 @@ static int WriteResult(void* result, int32_t execution_result = 0, uint32_t erro
 } // namespace AprShared
 
 namespace LibAmpr::Ampr {
-static int ExecuteAprCommandBuffer(uint64_t command_buffer, int32_t* execution_result,
-                                   uint32_t* error_offset);
+enum class Engine { Apr, Amm };
+static int QueueCommandBuffer(Engine engine, uint64_t command_buffer, uint32_t bytes,
+                              uint32_t priority, void* result, uint32_t* out_submission_id);
 }
 
 namespace LibKernelApr {
@@ -536,89 +534,37 @@ static int KYTY_SYSV_ABI ResolveFilepathsWithPrefixToIdsAndFileSizesForEach(
 	    AprShared::ResolvePathsCommon(path_list, count, ids, sizes, nullptr, prefix, results));
 }
 
-static int KYTY_SYSV_ABI SubmitCommandBufferAndGetResult(void*     command_buffer, uint64_t,
+static int KYTY_SYSV_ABI SubmitCommandBufferAndGetResult(void* command_buffer, uint32_t priority,
                                                          void*     result,
                                                          uint32_t* out_submission_id) {
 	PRINT_NAME();
-
-	if (command_buffer == nullptr) {
-		return KernelSyscallResult(LibKernel::KERNEL_ERROR_EINVAL);
-	}
-
-	const auto command_buffer_addr = reinterpret_cast<uint64_t>(command_buffer);
-	const auto id =
-	    AprShared::AllocateSubmissionId(command_buffer_addr, reinterpret_cast<uint64_t>(result));
-
-	int32_t  execution_result = OK;
-	uint32_t error_offset     = 0;
-	auto submit_result = LibAmpr::Ampr::ExecuteAprCommandBuffer(command_buffer_addr,
-	                                                            &execution_result, &error_offset);
-	if (submit_result != OK) {
-		return KernelSyscallResult(submit_result);
-	}
-	AprShared::SetSubmissionResult(id, execution_result, error_offset);
-
-	if (out_submission_id != nullptr) {
-		if (!AprShared::WriteGuest(reinterpret_cast<uint64_t>(out_submission_id), id)) {
-			return KernelSyscallResult(LibKernel::KERNEL_ERROR_EFAULT);
-		}
-	}
-
-	return KernelSyscallResult(AprShared::WriteResult(result, execution_result, error_offset));
+	return KernelSyscallResult(LibAmpr::Ampr::QueueCommandBuffer(
+	    LibAmpr::Ampr::Engine::Apr, reinterpret_cast<uint64_t>(command_buffer), 0, priority, result,
+	    out_submission_id));
 }
 
-static int KYTY_SYSV_ABI SubmitCommandBuffer(void* command_buffer, uint64_t) {
+static int KYTY_SYSV_ABI SubmitCommandBuffer(void* command_buffer, uint32_t priority) {
 	PRINT_NAME();
-
-	if (command_buffer == nullptr) {
-		return KernelSyscallResult(LibKernel::KERNEL_ERROR_EINVAL);
-	}
-
-	int32_t  execution_result = OK;
-	uint32_t error_offset     = 0;
-	return KernelSyscallResult(LibAmpr::Ampr::ExecuteAprCommandBuffer(
-	    reinterpret_cast<uint64_t>(command_buffer), &execution_result, &error_offset));
+	return KernelSyscallResult(LibAmpr::Ampr::QueueCommandBuffer(
+	    LibAmpr::Ampr::Engine::Apr, reinterpret_cast<uint64_t>(command_buffer), 0, priority,
+	    nullptr, nullptr));
 }
 
-static int KYTY_SYSV_ABI SubmitCommandBufferAndGetId(void*     command_buffer, uint64_t,
+static int KYTY_SYSV_ABI SubmitCommandBufferAndGetId(void* command_buffer, uint32_t priority,
                                                      uint32_t* out_submission_id) {
 	PRINT_NAME();
-
-	if (command_buffer == nullptr || out_submission_id == nullptr) {
+	if (out_submission_id == nullptr) {
 		return KernelSyscallResult(LibKernel::KERNEL_ERROR_EINVAL);
 	}
-
-	const auto command_buffer_addr = reinterpret_cast<uint64_t>(command_buffer);
-	const auto id                  = AprShared::AllocateSubmissionId(command_buffer_addr, 0);
-
-	int32_t  execution_result = OK;
-	uint32_t error_offset     = 0;
-	auto submit_result = LibAmpr::Ampr::ExecuteAprCommandBuffer(command_buffer_addr,
-	                                                            &execution_result, &error_offset);
-	if (submit_result != OK) {
-		return KernelSyscallResult(submit_result);
-	}
-	AprShared::SetSubmissionResult(id, execution_result, error_offset);
-
-	if (!AprShared::WriteGuest(reinterpret_cast<uint64_t>(out_submission_id), id)) {
-		return KernelSyscallResult(LibKernel::KERNEL_ERROR_EFAULT);
-	}
-
-	return OK;
+	return KernelSyscallResult(LibAmpr::Ampr::QueueCommandBuffer(
+	    LibAmpr::Ampr::Engine::Apr, reinterpret_cast<uint64_t>(command_buffer), 0, priority,
+	    nullptr, out_submission_id));
 }
 
 static int KYTY_SYSV_ABI WaitCommandBuffer(uint32_t submission_id) {
 	PRINT_NAME();
-
-	AprShared::SubmissionState state {};
-	if (!AprShared::CompleteSubmission(submission_id, &state)) {
-		return KernelSyscallResult(LibKernel::KERNEL_ERROR_ESRCH);
-	}
-	if (state.result != 0) {
-		return KernelSyscallResult(AprShared::WriteResult(
-		    reinterpret_cast<void*>(state.result), state.execution_result, state.error_offset));
-	}
-	return OK;
+	return KernelSyscallResult(
+	    AprShared::CompleteSubmission(submission_id) ? OK : LibKernel::KERNEL_ERROR_ESRCH);
 }
 
 } // namespace Apr
@@ -705,41 +651,45 @@ struct CommandBufferState {
 	bool     header_validated = false;
 	bool     buffer_validated = false;
 	struct ReadFileCommand {
-		uint64_t record_offset = 0;
-		uint32_t file_id       = 0;
-		uint64_t destination   = 0;
-		uint64_t size          = 0;
-		uint64_t file_offset   = 0;
+		uint32_t file_id     = 0;
+		uint64_t destination = 0;
+		uint64_t size        = 0;
+		uint64_t file_offset = 0;
 	};
 	struct KernelEventCommand {
-		uint64_t record_offset = 0;
-		uint64_t eq            = 0;
-		int32_t  id            = 0;
-		uint64_t data          = 0;
+		uint64_t eq   = 0;
+		int32_t  id   = 0;
+		uint64_t data = 0;
 	};
 	struct WriteAddressCommand {
-		uint64_t record_offset = 0;
-		uint64_t address       = 0;
-		uint64_t value         = 0;
+		uint64_t address = 0;
+		uint64_t value   = 0;
+	};
+	struct WaitAddressCommand {
+		uint64_t address;
+		uint64_t value;
+		uint8_t  compare;
 	};
 	struct AmmMapCommand {
-		uint64_t       record_offset = 0;
-		AmmCommandKind kind          = AmmCommandKind::MapAuto;
-		uint64_t       va            = 0;
-		uint64_t       dmem_offset   = 0;
-		uint64_t       size          = 0;
-		int32_t        type          = 0;
-		int32_t        prot          = 0;
-		uint8_t        gpu_mask_id   = 0;
+		AmmCommandKind kind        = AmmCommandKind::MapAuto;
+		uint64_t       va          = 0;
+		uint64_t       dmem_offset = 0;
+		uint64_t       size        = 0;
+		int32_t        type        = 0;
+		int32_t        prot        = 0;
+		uint8_t        gpu_mask_id = 0;
 	};
-	std::vector<ReadFileCommand>     read_file_commands;
-	std::vector<KernelEventCommand>  kernel_event_commands;
-	std::vector<WriteAddressCommand> write_address_commands;
-	std::vector<AmmMapCommand>       amm_map_commands;
-	bool                             gather_scatter_valid       = false;
-	uint32_t                         gather_scatter_file_id     = 0;
-	uint64_t                         gather_scatter_destination = 0;
-	uint64_t                         gather_scatter_file_offset = 0;
+	using CommandData = std::variant<ReadFileCommand, KernelEventCommand, WriteAddressCommand,
+	                                 AmmMapCommand, WaitAddressCommand>;
+	struct Command {
+		uint64_t record_offset;
+		CommandData data;
+	};
+	std::vector<Command> commands;
+	bool                 gather_scatter_valid       = false;
+	uint32_t             gather_scatter_file_id     = 0;
+	uint64_t             gather_scatter_destination = 0;
+	uint64_t             gather_scatter_file_offset = 0;
 };
 
 struct AmmUsageStatsData {
@@ -757,11 +707,6 @@ static std::unordered_map<uint64_t, CommandBufferState> g_command_buffers;
 static std::unordered_map<uint64_t, uint64_t>           g_command_buffer_aliases;
 using CommandBufferIterator = std::unordered_map<uint64_t, CommandBufferState>::iterator;
 
-static bool HasQueuedCommands(const CommandBufferState& state) {
-	return !state.read_file_commands.empty() || !state.kernel_event_commands.empty() ||
-	       !state.write_address_commands.empty() || !state.amm_map_commands.empty();
-}
-
 static void RegisterCommandBufferAliasLocked(uint64_t command_buffer, uint64_t buffer) {
 	for (auto it = g_command_buffer_aliases.begin(); it != g_command_buffer_aliases.end();) {
 		if (it->second == command_buffer) {
@@ -950,7 +895,7 @@ static bool UpdateCommandBufferTypeFlags(uint64_t command_buffer, uint32_t set_b
 static std::unordered_map<uint64_t, CommandBufferState>::iterator
 ResolveCommandBufferStateLocked(uint64_t command_buffer) {
 	auto it = g_command_buffers.find(command_buffer);
-	if (it != g_command_buffers.end() && HasQueuedCommands(it->second)) {
+	if (it != g_command_buffers.end() && !it->second.commands.empty()) {
 		return it;
 	}
 
@@ -999,10 +944,7 @@ static bool WriteCommandBufferPointers(uint64_t command_buffer, uint64_t buffer,
 	state.write_offset     = write_offset;
 	state.header_validated = true;
 	state.buffer_validated = false;
-	state.read_file_commands.clear();
-	state.kernel_event_commands.clear();
-	state.write_address_commands.clear();
-	state.amm_map_commands.clear();
+	state.commands.clear();
 	state.gather_scatter_valid       = false;
 	state.gather_scatter_file_id     = 0;
 	state.gather_scatter_destination = 0;
@@ -1107,7 +1049,8 @@ static bool AppendReadFileRecord(uint64_t command_buffer, uint8_t opcode, uint32
 	std::memset(reinterpret_cast<void*>(state.buffer + record_offset), 0,
 	            static_cast<size_t>(record_size));
 	std::memcpy(reinterpret_cast<void*>(state.buffer + record_offset), &opcode, sizeof(opcode));
-	state.read_file_commands.push_back({record_offset, file_id, destination, size, file_offset});
+	state.commands.push_back({record_offset, CommandBufferState::ReadFileCommand {
+	                                             file_id, destination, size, file_offset}});
 	if (!CommitCommandBufferRecord(command_buffer, &state, record_size)) {
 		return false;
 	}
@@ -1145,30 +1088,11 @@ static bool AppendKernelEventRecord(uint64_t command_buffer, uint64_t eq, int32_
 	const auto record_offset = state.write_offset;
 	std::memcpy(reinterpret_cast<void*>(state.buffer + record_offset), record.data(),
 	            record.size());
-	state.kernel_event_commands.push_back({record_offset, eq, id, data});
+	state.commands.push_back(
+	    {record_offset, CommandBufferState::KernelEventCommand {eq, id, data}});
 	return CommitCommandBufferRecord(command_buffer, &state, KERNEL_EVENT_RECORD_SIZE);
 }
 
-static bool AppendWriteAddressRecord(uint64_t command_buffer, uint64_t address, uint64_t value,
-                                     uint64_t record_size = 0x20) {
-	std::scoped_lock      lock(g_command_buffer_mutex);
-	CommandBufferIterator it;
-	if (!GetOrCreateCommandBufferStateLocked(command_buffer, &it)) {
-		return false;
-	}
-
-	auto& state = it->second;
-	if (!EnsureCommandBufferRecordSpace(command_buffer, &state, record_size)) {
-		return false;
-	}
-
-	const auto record_offset = state.write_offset;
-	std::memset(reinterpret_cast<void*>(state.buffer + record_offset), 0,
-	            static_cast<size_t>(record_size));
-	state.write_address_commands.push_back({record_offset, address, value});
-	return CommitCommandBufferRecord(command_buffer, &state, record_size);
-}
-
 static bool ValidateAmmMapArgs(uint64_t va, uint64_t size) {
 	return va != 0 && size != 0 && (va & (AMM_PAGE_SIZE - 1u)) == 0 &&
 	       (size & (AMM_PAGE_SIZE - 1u)) == 0 && va + size >= va;
@@ -1226,8 +1150,8 @@ static int ExecuteAmmMapCommand(const CommandBufferState::AmmMapCommand& command
 	    static_cast<int64_t>(command.dmem_offset), AMM_PAGE_SIZE);
 }
 
-static bool AppendAmmMapRecord(uint64_t                                 command_buffer,
-                               const CommandBufferState::AmmMapCommand& cmd, uint64_t record_size) {
+static bool AppendCommandRecord(uint64_t command_buffer, CommandBufferState::CommandData command,
+                                uint64_t record_size) {
 	std::scoped_lock      lock(g_command_buffer_mutex);
 	CommandBufferIterator it;
 	if (!GetOrCreateCommandBufferStateLocked(command_buffer, &it)) {
@@ -1239,149 +1163,204 @@ static bool AppendAmmMapRecord(uint64_t                                 command_
 		return false;
 	}
 
-	auto record          = cmd;
-	record.record_offset = state.write_offset;
 	auto* record_addr    = reinterpret_cast<void*>(state.buffer + state.write_offset);
 	std::memset(record_addr, 0, static_cast<size_t>(record_size));
-	state.amm_map_commands.push_back(record);
+	state.commands.push_back({state.write_offset, std::move(command)});
 	return CommitCommandBufferRecord(command_buffer, &state, record_size);
 }
 
 static int ReadHostFileToGuest(const std::string& host_path, uint64_t file_offset,
                                uint64_t destination, uint64_t size, uint64_t* bytes_read);
 
-static int ExecuteAprCommandBuffer(uint64_t command_buffer, int32_t* execution_result,
-                                   uint32_t* error_offset) {
-	if (execution_result == nullptr || error_offset == nullptr) {
-		return LibKernel::KERNEL_ERROR_EINVAL;
+static int ExecuteCommand(const CommandBufferState::Command& entry) {
+	if (const auto* payload = std::get_if<CommandBufferState::ReadFileCommand>(&entry.data)) {
+		const auto& command = *payload;
+		std::string host_path;
+		if (!AprShared::TryGetHostPath(command.file_id, &host_path)) {
+			LOGF("\tAPR submit failed for unknown file id: 0x%08" PRIx32 "\n", command.file_id);
+			return LibKernel::KERNEL_ERROR_ENOENT;
+		}
+
+		uint64_t bytes_read = 0;
+		auto     result = ReadHostFileToGuest(host_path, command.file_offset, command.destination,
+		                                      command.size, &bytes_read);
+		if (result != OK) {
+			LOGF("\tAPR submit read failed: id=0x%08" PRIx32 ", result=0x%08" PRIx32 ", path=%s\n",
+			     command.file_id, static_cast<uint32_t>(result), host_path.c_str());
+			return result;
+		}
+	} else if (const auto* payload =
+	               std::get_if<CommandBufferState::KernelEventCommand>(&entry.data)) {
+		const auto& command = *payload;
+		const auto  eq      = static_cast<LibKernel::EventQueue::KernelEqueue>(command.eq);
+		auto        result  = LibKernel::EventQueue::KernelTriggerUserEvent(
+		    eq, command.id, reinterpret_cast<void*>(command.data));
+		if (result != OK) {
+			LOGF("\tAPR submit event failed: eq=0x%016" PRIx64 ", id=%" PRId32
+			     ", result=0x%08" PRIx32 "\n",
+			     command.eq, command.id, static_cast<uint32_t>(result));
+			return result;
+		}
+	} else if (const auto* payload =
+	               std::get_if<CommandBufferState::WriteAddressCommand>(&entry.data)) {
+		const auto& command = *payload;
+		std::atomic_ref(*reinterpret_cast<uint64_t*>(command.address))
+		    .store(command.value, std::memory_order_release);
+	} else if (const auto* payload = std::get_if<CommandBufferState::AmmMapCommand>(&entry.data)) {
+		const auto& command = *payload;
+		int         result  = OK;
+		if (command.kind == AmmCommandKind::Unmap) {
+			result = LibKernel::Memory::KernelMunmap(command.va, command.size);
+		} else {
+			result = ExecuteAmmMapCommand(command);
+		}
+
+		if (result != OK) {
+			LOGF("\tAMM submit command failed: kind=%u va=0x%016" PRIx64 " dmem=0x%016" PRIx64
+			     " size=0x%016" PRIx64 " type=%" PRId32 " prot=0x%08" PRIx32 " result=0x%08" PRIx32
+			     "\n",
+			     static_cast<uint32_t>(command.kind), command.va, command.dmem_offset, command.size,
+			     command.type, static_cast<uint32_t>(command.prot), static_cast<uint32_t>(result));
+			return result;
+		}
+	}
+	return OK;
+}
+
+struct Submission {
+	std::vector<CommandBufferState::Command> commands;
+	size_t                                   cursor = 0;
+	void*                                    result = nullptr;
+	uint32_t                                 id     = 0;
+};
+
+static bool IsWaitSatisfied(const CommandBufferState::WaitAddressCommand& wait) {
+	const uint64_t value =
+	    std::atomic_ref(*reinterpret_cast<uint64_t*>(wait.address)).load(std::memory_order_acquire);
+	switch (wait.compare) {
+		case 0: return value == wait.value;
+		case 1: return value > wait.value;
+		case 2: return value < wait.value;
+		case 3: return value != wait.value;
+		case 4: return static_cast<int64_t>(value - wait.value) >= 0;
+		case 5: return static_cast<int64_t>(value) > static_cast<int64_t>(wait.value);
+		default: return static_cast<int64_t>(value) < static_cast<int64_t>(wait.value);
+	}
+}
+
+class CommandEngine {
+public:
+	explicit CommandEngine(size_t priorities)
+	    : m_queues(priorities), m_worker([this](std::stop_token stop) { Run(stop); }) {}
+
+	~CommandEngine() {
+		std::scoped_lock lock(m_mutex);
+		m_worker.request_stop();
+		m_cv.notify_all();
 	}
 
-	*execution_result = OK;
-	*error_offset     = 0;
+	void Submit(uint32_t priority, Submission submission) {
+		std::scoped_lock lock(m_mutex);
+		m_queues[priority].push_back(std::move(submission));
+		m_cv.notify_one();
+	}
 
+private:
+	void Run(std::stop_token stop) {
+		std::unique_lock lock(m_mutex);
+		while (!stop.stop_requested()) {
+			std::deque<Submission>* ready   = nullptr;
+			bool                    pending = false;
+			for (auto& queue: m_queues) {
+				if (queue.empty()) {
+					continue;
+				}
+				pending          = true;
+				auto& submission = queue.front();
+				if (submission.cursor < submission.commands.size()) {
+					const auto* wait = std::get_if<CommandBufferState::WaitAddressCommand>(
+					    &submission.commands[submission.cursor].data);
+					if (wait != nullptr && !IsWaitSatisfied(*wait)) {
+						continue;
+					}
+				}
+				ready = &queue;
+				break;
+			}
+			if (ready == nullptr) {
+				if (pending) {
+					// CPU and GPU fence writes do not notify the submission queue.
+					m_cv.wait_for(lock, std::chrono::microseconds(100));
+				} else {
+					m_cv.wait(lock);
+				}
+				continue;
+			}
+
+			auto& submission = ready->front();
+			lock.unlock();
+			int      result       = OK;
+			uint32_t error_offset = 0;
+			if (submission.cursor < submission.commands.size()) {
+				const auto& command = submission.commands[submission.cursor++];
+				result              = ExecuteCommand(command);
+				if (result != OK) {
+					error_offset = static_cast<uint32_t>(command.record_offset);
+				}
+			}
+			const bool complete = result != OK || submission.cursor == submission.commands.size();
+			if (complete) {
+				AprShared::WriteResult(submission.result, result, error_offset);
+				AprShared::SetSubmissionComplete(submission.id);
+			}
+			lock.lock();
+			if (complete) {
+				ready->pop_front();
+			}
+		}
+	}
+
+	std::mutex                          m_mutex;
+	std::condition_variable             m_cv;
+	std::vector<std::deque<Submission>> m_queues;
+	std::jthread                        m_worker;
+};
+
+static int QueueCommandBuffer(Engine engine, uint64_t command_buffer, uint32_t bytes,
+                              uint32_t priority, void* result, uint32_t* out_submission_id) {
+	const bool amm = engine == Engine::Amm;
+	if (command_buffer == 0 || priority > (amm ? 2u : 6u)) {
+		return LibKernel::KERNEL_ERROR_EINVAL;
+	}
+	if ((result != nullptr &&
+	     !AprShared::IsValidGuestRange(reinterpret_cast<uint64_t>(result),
+	                                   sizeof(AprShared::ResultBuffer), true)) ||
+	    (out_submission_id != nullptr &&
+	     !AprShared::IsValidGuestRange(reinterpret_cast<uint64_t>(out_submission_id),
+	                                   sizeof(uint32_t), true))) {
+		return LibKernel::KERNEL_ERROR_EFAULT;
+	}
 	CommandBufferState state {};
 	if (!TryGetCommandBufferState(command_buffer, &state)) {
 		return LibKernel::KERNEL_ERROR_EFAULT;
 	}
-
-	enum class CommandKind {
-		ReadFile,
-		KernelEvent,
-		WriteAddress,
-		AmmMap,
-	};
-
-	struct OrderedCommand {
-		uint64_t    record_offset = 0;
-		CommandKind kind          = CommandKind::ReadFile;
-		size_t      index         = 0;
-	};
-
-	std::vector<OrderedCommand> ordered;
-	ordered.reserve(state.read_file_commands.size() + state.kernel_event_commands.size() +
-	                state.write_address_commands.size() + state.amm_map_commands.size());
-	for (size_t i = 0; i < state.read_file_commands.size(); i++) {
-		if (state.read_file_commands[i].record_offset < state.write_offset) {
-			ordered.push_back(
-			    {state.read_file_commands[i].record_offset, CommandKind::ReadFile, i});
-		}
+	const uint64_t end = amm ? bytes : state.write_offset;
+	if (end == 0 || end > state.write_offset) {
+		return LibKernel::KERNEL_ERROR_EINVAL;
 	}
-	for (size_t i = 0; i < state.kernel_event_commands.size(); i++) {
-		if (state.kernel_event_commands[i].record_offset < state.write_offset) {
-			ordered.push_back(
-			    {state.kernel_event_commands[i].record_offset, CommandKind::KernelEvent, i});
-		}
+	state.commands.erase(std::lower_bound(state.commands.begin(), state.commands.end(), end,
+	                                      [](const auto& command, uint64_t offset) {
+		                                      return command.record_offset < offset;
+	                                      }),
+	                     state.commands.end());
+	const uint32_t id = out_submission_id != nullptr ? AprShared::AllocateSubmissionId() : 0;
+	if (out_submission_id != nullptr) {
+		AprShared::WriteGuest(reinterpret_cast<uint64_t>(out_submission_id), id);
 	}
-	for (size_t i = 0; i < state.write_address_commands.size(); i++) {
-		if (state.write_address_commands[i].record_offset < state.write_offset) {
-			ordered.push_back(
-			    {state.write_address_commands[i].record_offset, CommandKind::WriteAddress, i});
-		}
-	}
-	for (size_t i = 0; i < state.amm_map_commands.size(); i++) {
-		if (state.amm_map_commands[i].record_offset < state.write_offset) {
-			ordered.push_back({state.amm_map_commands[i].record_offset, CommandKind::AmmMap, i});
-		}
-	}
-
-	std::sort(ordered.begin(), ordered.end(), [](const OrderedCommand& a, const OrderedCommand& b) {
-		return a.record_offset < b.record_offset;
-	});
-
-	for (const auto& entry: ordered) {
-		switch (entry.kind) {
-			case CommandKind::ReadFile: {
-				const auto& command = state.read_file_commands[entry.index];
-				std::string host_path;
-				if (!AprShared::TryGetHostPath(command.file_id, &host_path)) {
-					LOGF("\tAPR submit failed for unknown file id: 0x%08" PRIx32 "\n",
-					     command.file_id);
-					*execution_result = LibKernel::KERNEL_ERROR_ENOENT;
-					*error_offset     = static_cast<uint32_t>(command.record_offset);
-					return OK;
-				}
-
-				uint64_t bytes_read = 0;
-				auto result = ReadHostFileToGuest(host_path, command.file_offset,
-				                                  command.destination, command.size, &bytes_read);
-				if (result != OK) {
-					LOGF("\tAPR submit read failed: id=0x%08" PRIx32 ", result=0x%08" PRIx32
-					     ", path=%s\n",
-					     command.file_id, static_cast<uint32_t>(result), host_path.c_str());
-					*execution_result = result;
-					*error_offset     = static_cast<uint32_t>(command.record_offset);
-					return OK;
-				}
-			} break;
-			case CommandKind::KernelEvent: {
-				const auto& command = state.kernel_event_commands[entry.index];
-				const auto  eq      = static_cast<LibKernel::EventQueue::KernelEqueue>(command.eq);
-				auto        result  = LibKernel::EventQueue::KernelTriggerUserEvent(
-				    eq, command.id, reinterpret_cast<void*>(command.data));
-				if (result != OK) {
-					LOGF("\tAPR submit event failed: eq=0x%016" PRIx64 ", id=%" PRId32
-					     ", result=0x%08" PRIx32 "\n",
-					     command.eq, command.id, static_cast<uint32_t>(result));
-					*execution_result = result;
-					*error_offset     = static_cast<uint32_t>(command.record_offset);
-					return OK;
-				}
-			} break;
-			case CommandKind::WriteAddress: {
-				const auto& command = state.write_address_commands[entry.index];
-				if (!AprShared::WriteGuest(command.address, command.value)) {
-					LOGF("\tAMPR submit write-address failed: address=0x%016" PRIx64
-					     " value=0x%016" PRIx64 "\n",
-					     command.address, command.value);
-					*execution_result = LibKernel::KERNEL_ERROR_EFAULT;
-					*error_offset     = static_cast<uint32_t>(command.record_offset);
-					return OK;
-				}
-			} break;
-			case CommandKind::AmmMap: {
-				const auto& command = state.amm_map_commands[entry.index];
-				int         result  = OK;
-				if (command.kind == AmmCommandKind::Unmap) {
-					result = LibKernel::Memory::KernelMunmap(command.va, command.size);
-				} else {
-					result = ExecuteAmmMapCommand(command);
-				}
-
-				if (result != OK) {
-					LOGF("\tAMM submit command failed: kind=%u va=0x%016" PRIx64
-					     " dmem=0x%016" PRIx64 " size=0x%016" PRIx64 " type=%" PRId32
-					     " prot=0x%08" PRIx32 " result=0x%08" PRIx32 "\n",
-					     static_cast<uint32_t>(command.kind), command.va, command.dmem_offset,
-					     command.size, command.type, static_cast<uint32_t>(command.prot),
-					     static_cast<uint32_t>(result));
-					*execution_result = result;
-					*error_offset     = static_cast<uint32_t>(command.record_offset);
-					return OK;
-				}
-			} break;
-		}
-	}
-
+	static CommandEngine apr_engine(7);
+	static CommandEngine amm_engine(3);
+	(amm ? amm_engine : apr_engine)
+	    .Submit(priority, {std::move(state.commands), 0, result, id});
 	return OK;
 }
 
@@ -1437,8 +1416,11 @@ static int ReadHostFileToGuest(const std::string& host_path, uint64_t file_offse
 		return OK;
 	}
 
-	std::vector<uint8_t> buffer(
-	    static_cast<size_t>(std::min<uint64_t>(APR_HOST_READ_CHUNK_SIZE, readable)));
+	static thread_local std::vector<uint8_t> buffer;
+	const auto buffer_size = static_cast<size_t>(std::min<uint64_t>(APR_HOST_READ_CHUNK_SIZE, readable));
+	if (buffer.size() < buffer_size) {
+		buffer.resize(buffer_size);
+	}
 	while (*bytes_read < readable) {
 		const auto request =
 		    static_cast<uint32_t>(std::min<uint64_t>(buffer.size(), readable - *bytes_read));
@@ -1935,11 +1917,20 @@ static int KYTY_SYSV_ABI CommandBufferPopMarker(void* command_buffer) {
 	return AppendNoOpCommand(command_buffer, sizeof(uint32_t));
 }
 
-static int KYTY_SYSV_ABI CommandBufferWaitOnAddress(void* command_buffer, volatile uint64_t*,
-                                                    uint64_t, uint8_t, uint8_t) {
+static int KYTY_SYSV_ABI CommandBufferWaitOnAddress(void*              command_buffer,
+                                                    volatile uint64_t* address, uint64_t value,
+                                                    uint8_t compare, uint8_t) {
 	PRINT_NAME();
-
-	return AppendNoOpCommand(command_buffer, 0x20);
+	if (command_buffer == nullptr || address == nullptr ||
+	    (reinterpret_cast<uint64_t>(address) & 7) != 0 || compare > 6) {
+		return LibKernel::KERNEL_ERROR_EINVAL;
+	}
+	return AppendCommandRecord(reinterpret_cast<uint64_t>(command_buffer),
+	                           CommandBufferState::WaitAddressCommand {
+	                               reinterpret_cast<uint64_t>(address), value, compare},
+	                           0x20)
+	           ? OK
+	           : LibKernel::KERNEL_ERROR_EBUSY;
 }
 
 static int KYTY_SYSV_ABI CommandBufferWaitOnCounter(void* command_buffer, uint8_t, uint8_t,
@@ -1951,11 +1942,14 @@ static int KYTY_SYSV_ABI CommandBufferWaitOnCounter(void* command_buffer, uint8_
 
 static int AppendWriteAddressCommand(void* command_buffer, volatile uint64_t* address,
                                      uint64_t value) {
-	if (command_buffer == nullptr || address == nullptr) {
+	if (command_buffer == nullptr || address == nullptr ||
+	    (reinterpret_cast<uint64_t>(address) & 7) != 0) {
 		return LibKernel::KERNEL_ERROR_EINVAL;
 	}
-	return AppendWriteAddressRecord(reinterpret_cast<uint64_t>(command_buffer),
-	                                reinterpret_cast<uint64_t>(address), value)
+	return AppendCommandRecord(
+	           reinterpret_cast<uint64_t>(command_buffer),
+	           CommandBufferState::WriteAddressCommand {reinterpret_cast<uint64_t>(address), value},
+	           0x20)
 	           ? OK
 	           : LibKernel::KERNEL_ERROR_EBUSY;
 }
@@ -2213,8 +2207,8 @@ static int KYTY_SYSV_ABI AmmCommandBufferMap(void* command_buffer, uint64_t va,
 	command.type = type;
 	command.prot = prot;
 
-	return AppendAmmMapRecord(reinterpret_cast<uint64_t>(command_buffer), command,
-	                          AMM_MAP_RECORD_SIZE)
+	return AppendCommandRecord(reinterpret_cast<uint64_t>(command_buffer), command,
+	                           AMM_MAP_RECORD_SIZE)
 	           ? OK
 	           : LibKernel::KERNEL_ERROR_EBUSY;
 }
@@ -2236,8 +2230,8 @@ static int KYTY_SYSV_ABI AmmCommandBufferMapWithGpuMaskId(void* command_buffer,
 	command.prot        = prot;
 	command.gpu_mask_id = gpu_mask_id;
 
-	return AppendAmmMapRecord(reinterpret_cast<uint64_t>(command_buffer), command,
-	                          AMM_MAP_RECORD_SIZE)
+	return AppendCommandRecord(reinterpret_cast<uint64_t>(command_buffer), command,
+	                           AMM_MAP_RECORD_SIZE)
 	           ? OK
 	           : LibKernel::KERNEL_ERROR_EBUSY;
 }
@@ -2260,8 +2254,8 @@ static int KYTY_SYSV_ABI AmmCommandBufferMapDirect(void* command_buffer, uint64_
 	command.type        = type;
 	command.prot        = prot;
 
-	return AppendAmmMapRecord(reinterpret_cast<uint64_t>(command_buffer), command,
-	                          AMM_MAP_DIRECT_RECORD_SIZE)
+	return AppendCommandRecord(reinterpret_cast<uint64_t>(command_buffer), command,
+	                           AMM_MAP_DIRECT_RECORD_SIZE)
 	           ? OK
 	           : LibKernel::KERNEL_ERROR_EBUSY;
 }
@@ -2286,8 +2280,8 @@ static int KYTY_SYSV_ABI AmmCommandBufferMapDirectWithGpuMaskId(void* command_bu
 	command.prot        = prot;
 	command.gpu_mask_id = gpu_mask_id;
 
-	return AppendAmmMapRecord(reinterpret_cast<uint64_t>(command_buffer), command,
-	                          AMM_MAP_DIRECT_RECORD_SIZE)
+	return AppendCommandRecord(reinterpret_cast<uint64_t>(command_buffer), command,
+	                           AMM_MAP_DIRECT_RECORD_SIZE)
 	           ? OK
 	           : LibKernel::KERNEL_ERROR_EBUSY;
 }
@@ -2304,8 +2298,8 @@ static int KYTY_SYSV_ABI AmmCommandBufferUnmap(void* command_buffer, uint64_t va
 	command.va   = va;
 	command.size = size;
 
-	return AppendAmmMapRecord(reinterpret_cast<uint64_t>(command_buffer), command,
-	                          AMM_UNMAP_RECORD_SIZE)
+	return AppendCommandRecord(reinterpret_cast<uint64_t>(command_buffer), command,
+	                           AMM_UNMAP_RECORD_SIZE)
 	           ? OK
 	           : LibKernel::KERNEL_ERROR_EBUSY;
 }
@@ -2369,90 +2363,36 @@ static int KYTY_SYSV_ABI AmmSetPageTablePoolOccupancyNotificationThreshold(uint3
 	return OK;
 }
 
-static int KYTY_SYSV_ABI AmmSubmitCommandBuffer(void* command_buffer_base, uint32_t, uint32_t) {
+static int KYTY_SYSV_ABI AmmSubmitCommandBuffer(void* command_buffer_base, uint32_t bytes,
+                                                uint32_t priority) {
 	PRINT_NAME();
-
-	if (command_buffer_base == nullptr) {
-		return LibKernel::KERNEL_ERROR_EINVAL;
-	}
-
-	int32_t  execution_result = OK;
-	uint32_t error_offset     = 0;
-	return ExecuteAprCommandBuffer(reinterpret_cast<uint64_t>(command_buffer_base),
-	                               &execution_result, &error_offset);
+	return QueueCommandBuffer(Engine::Amm, reinterpret_cast<uint64_t>(command_buffer_base), bytes,
+	                          priority, nullptr, nullptr);
 }
 
-static int KYTY_SYSV_ABI AmmSubmitCommandBufferAndGetId(void* command_buffer_base, uint32_t,
-                                                        uint32_t, uint32_t* out_submission_id) {
+static int KYTY_SYSV_ABI AmmSubmitCommandBufferAndGetId(void* command_buffer_base, uint32_t bytes,
+                                                        uint32_t  priority,
+                                                        uint32_t* out_submission_id) {
 	PRINT_NAME();
-
-	if (command_buffer_base == nullptr || out_submission_id == nullptr) {
+	if (out_submission_id == nullptr) {
 		return LibKernel::KERNEL_ERROR_EINVAL;
 	}
-
-	const auto command_buffer_addr = reinterpret_cast<uint64_t>(command_buffer_base);
-	const auto id                  = AprShared::AllocateSubmissionId(command_buffer_addr, 0);
-
-	int32_t  execution_result = OK;
-	uint32_t error_offset     = 0;
-	auto     submit_result =
-	    ExecuteAprCommandBuffer(command_buffer_addr, &execution_result, &error_offset);
-	if (submit_result != OK) {
-		return submit_result;
-	}
-	AprShared::SetSubmissionResult(id, execution_result, error_offset);
-
-	if (!AprShared::WriteGuest(reinterpret_cast<uint64_t>(out_submission_id), id)) {
-		return LibKernel::KERNEL_ERROR_EFAULT;
-	}
-
-	return OK;
+	return QueueCommandBuffer(Engine::Amm, reinterpret_cast<uint64_t>(command_buffer_base), bytes,
+	                          priority, nullptr, out_submission_id);
 }
 
-static int KYTY_SYSV_ABI AmmSubmitCommandBufferAndGetResult(void* command_buffer_base, uint32_t,
-                                                            uint32_t, void* result,
+static int KYTY_SYSV_ABI AmmSubmitCommandBufferAndGetResult(void*    command_buffer_base,
+                                                            uint32_t bytes, uint32_t priority,
+                                                            void*     result,
                                                             uint32_t* out_submission_id) {
 	PRINT_NAME();
-
-	if (command_buffer_base == nullptr) {
-		return LibKernel::KERNEL_ERROR_EINVAL;
-	}
-
-	const auto command_buffer_addr = reinterpret_cast<uint64_t>(command_buffer_base);
-	const auto id =
-	    AprShared::AllocateSubmissionId(command_buffer_addr, reinterpret_cast<uint64_t>(result));
-
-	int32_t  execution_result = OK;
-	uint32_t error_offset     = 0;
-	auto     submit_result =
-	    ExecuteAprCommandBuffer(command_buffer_addr, &execution_result, &error_offset);
-	if (submit_result != OK) {
-		return submit_result;
-	}
-	AprShared::SetSubmissionResult(id, execution_result, error_offset);
-
-	if (out_submission_id != nullptr &&
-	    !AprShared::WriteGuest(reinterpret_cast<uint64_t>(out_submission_id), id)) {
-		return LibKernel::KERNEL_ERROR_EFAULT;
-	}
-
-	return AprShared::WriteResult(result, execution_result, error_offset);
+	return QueueCommandBuffer(Engine::Amm, reinterpret_cast<uint64_t>(command_buffer_base), bytes,
+	                          priority, result, out_submission_id);
 }
 
 static int KYTY_SYSV_ABI AmmWaitCommandBufferCompletion(uint32_t submission_id) {
 	PRINT_NAME();
-
-	AprShared::SubmissionState state {};
-	if (!AprShared::CompleteSubmission(submission_id, &state)) {
-		return LibKernel::KERNEL_ERROR_ESRCH;
-	}
-
-	if (state.result != 0) {
-		return AprShared::WriteResult(reinterpret_cast<void*>(state.result), state.execution_result,
-		                              state.error_offset);
-	}
-
-	return OK;
+	return AprShared::CompleteSubmission(submission_id) ? OK : LibKernel::KERNEL_ERROR_ESRCH;
 }
 
 } // namespace Ampr

src/libs/libKernel.cpp

@@ -1797,6 +1797,12 @@ int KYTY_SYSV_ABI KernelRtldThreadAtexitDecrement(uint64_t* /*c*/) {
 	return 0;
 }
 
+static uint64_t KYTY_SYSV_ABI KernelGetAvailableCpumask() {
+	PRINT_NAME();
+
+	return 0x1fff;
+}
+
 static KYTY_SYSV_ABI int KernelGetCurrentCpu() {
 	PRINT_NAME();
 
@@ -3340,6 +3346,7 @@ LIB_DEFINE(InitLibKernel_1) {
 	LIB_FUNC("vYU8P9Td2Zo", KernelAioInitializeImpl);
 	LIB_FUNC("WhCc1w3EhSI", LibKernel::KernelSetThreadAtexitReport);
 	LIB_FUNC("WkwEd3N7w0Y", LibKernel::KernelInstallExceptionHandler);
+	LIB_FUNC("La9uyZv4Kvw", LibKernel::KernelGetAvailableCpumask);
 	LIB_FUNC("g0VTBxfJyu0", LibKernel::KernelGetCurrentCpu);
 	LIB_FUNC("wzvqT4UqKX8", LibKernel::KernelLoadStartModule);
 	LIB_FUNC("Xjoosiw+XPI", LibKernel::KernelUuidCreate);