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

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
@@ -606,6 +610,7 @@ 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>)
@@ -682,6 +687,7 @@ 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

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/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;
@@ -170,14 +171,22 @@ bool BufferCache::DownloadBufferMemory(Buffer& buffer, uint64_t vaddr, uint64_t
 	                       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), owner = std::move(temporary)] {
+	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;
 }
 
@@ -256,10 +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) {
@@ -614,7 +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/textureCache.cpp

@@ -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 {

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/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.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/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/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/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);

tests/AudioOutTimingTests.cpp

@@ -0,0 +1,401 @@
+// Exercise production output scheduling with a deterministic clock and consuming SDL queues.
+#include "common/threads.h"
+#include "libs/dualSenseHaptics.h"
+
+#include <SDL3/SDL.h>
+#include <algorithm>
+#include <array>
+#include <cstdio>
+#include <cstdlib>
+#include <memory>
+#include <vector>
+
+namespace {
+void Check(bool condition, const char* message) {
+	if (!condition) {
+		std::fprintf(stderr, "AudioOutTimingTests: %s\n", message);
+		std::abort();
+	}
+}
+
+struct Stream {
+	SDL_AudioSpec spec;
+	uint64_t      frames = 0; // Millionths of a frame retain exact simulated consumption.
+	uint64_t      updated;
+};
+
+uint64_t                             now       = 1000000;
+uint32_t                             oversleep = 0, processing = 0;
+std::vector<uint32_t>                sleeps;
+std::vector<std::unique_ptr<Stream>> streams;
+bool                                 fail_open = false, fail_put = false, stalled = false;
+bool                                 pad_connected = false, pad_bluetooth = false;
+uint64_t                             pad_queue_us = 0;
+int                                  clears = 0;
+
+Stream& GetStream(SDL_AudioStream* stream) {
+	return *reinterpret_cast<Stream*>(stream);
+}
+
+void Drain(Stream& stream) {
+	if (!stalled) {
+		const auto consumed = (now - stream.updated) * stream.spec.freq;
+		stream.frames -= std::min(stream.frames, consumed);
+	}
+	stream.updated = now;
+}
+} // namespace
+
+namespace Fake {
+bool InitSubSystem(SDL_InitFlags) {
+	return true;
+}
+void QuitSubSystem(SDL_InitFlags) {}
+bool ResumeAudioStreamDevice(SDL_AudioStream*) {
+	return true;
+}
+
+SDL_AudioStream* OpenAudioDeviceStream(SDL_AudioDeviceID, const SDL_AudioSpec* spec,
+                                       SDL_AudioStreamCallback, void*) {
+	if (fail_open) {
+		return nullptr;
+	}
+
+	streams.push_back(std::make_unique<Stream>(Stream {*spec, 0, now}));
+	return reinterpret_cast<SDL_AudioStream*>(streams.back().get());
+}
+
+void DestroyAudioStream(SDL_AudioStream* stream) {
+	const auto it = std::find_if(streams.begin(), streams.end(), [stream](const auto& value) {
+		return value.get() == &GetStream(stream);
+	});
+	Check(it != streams.end(), "unknown stream destroyed");
+	streams.erase(it);
+}
+
+int GetAudioStreamQueued(SDL_AudioStream* stream) {
+	auto& state = GetStream(stream);
+	Drain(state);
+	return static_cast<int>(state.frames / 1000000 * SDL_AUDIO_BYTESIZE(state.spec.format) *
+	                        state.spec.channels);
+}
+
+bool PutAudioStreamData(SDL_AudioStream* stream, const void*, int bytes) {
+	now += processing;
+	if (fail_put) {
+		return false;
+	}
+	auto& state = GetStream(stream);
+	Drain(state);
+	state.frames += static_cast<uint64_t>(bytes) * 1000000 /
+	                (SDL_AUDIO_BYTESIZE(state.spec.format) * state.spec.channels);
+	return true;
+}
+
+bool ClearAudioStream(SDL_AudioStream* stream) {
+	auto& state   = GetStream(stream);
+	state.frames  = 0;
+	state.updated = now;
+	clears++;
+	return true;
+}
+} // namespace Fake
+
+namespace Common {
+class TimingThread: public Thread {
+public:
+	static void SleepMicro(uint32_t micros) {
+		sleeps.push_back(micros);
+		now += micros + oversleep;
+	}
+};
+} // namespace Common
+
+#define Thread                      TimingThread
+#define SDL_InitSubSystem           Fake::InitSubSystem
+#define SDL_QuitSubSystem           Fake::QuitSubSystem
+#define SDL_OpenAudioDeviceStream   Fake::OpenAudioDeviceStream
+#define SDL_ResumeAudioStreamDevice Fake::ResumeAudioStreamDevice
+#define SDL_DestroyAudioStream      Fake::DestroyAudioStream
+#define SDL_GetAudioStreamQueued    Fake::GetAudioStreamQueued
+#define SDL_PutAudioStreamData      Fake::PutAudioStreamData
+#define SDL_ClearAudioStream        Fake::ClearAudioStream
+#include "libs/audio.cpp"
+#undef Thread
+#undef SDL_InitSubSystem
+#undef SDL_QuitSubSystem
+#undef SDL_OpenAudioDeviceStream
+#undef SDL_ResumeAudioStreamDevice
+#undef SDL_DestroyAudioStream
+#undef SDL_GetAudioStreamQueued
+#undef SDL_PutAudioStreamData
+#undef SDL_ClearAudioStream
+
+namespace Libs::Controller {
+int GetActiveControllerId() {
+	return 0;
+}
+
+namespace DualSenseHaptics {
+Stream* Open(uint32_t, bool) {
+	static int pad_stream;
+	return pad_connected ? reinterpret_cast<Stream*>(&pad_stream) : nullptr;
+}
+void     Close(Stream*) {}
+bool UsesBluetooth(const Stream* stream) {
+	return stream != nullptr && pad_bluetooth;
+}
+uint64_t Queue(Stream* stream, int, const void*, uint32_t, uint32_t, bool, const int*) {
+	return stream != nullptr ? pad_queue_us : 0;
+}
+} // namespace DualSenseHaptics
+} // namespace Libs::Controller
+
+namespace Libs::LibKernel {
+uint64_t KYTY_SYSV_ABI KernelGetProcessTime() {
+	return now;
+}
+} // namespace Libs::LibKernel
+
+namespace Loader::Timer {
+double GetTimeMs() {
+	return static_cast<double>(now) / 1000;
+}
+} // namespace Loader::Timer
+
+namespace {
+using Audio = Libs::Audio::Audio;
+
+struct Fixture {
+	Audio                   audio;
+	std::array<float, 2048> pcm {};
+
+	Fixture() {
+		Check(streams.empty(), "stream leaked between tests");
+		now        = 1000000;
+		processing = oversleep = 0;
+		clears                 = 0;
+		fail_open = fail_put = stalled = false;
+		pad_connected = pad_bluetooth = false;
+		pad_queue_us = 0;
+		sleeps.clear();
+	}
+
+	Audio::Id Open(uint32_t frames = 256, int type = 0) {
+		const auto id = audio.AudioOutOpen(type, frames, 48000, Audio::Format::FloatStereo);
+		Check(id.IsValid(), "port open failed");
+		return id;
+	}
+
+	void Output(Audio::Id port, bool blocking = true) {
+		Audio::OutputParam param {port, pcm.data()};
+		audio.AudioOutOutputs(&param, 1, blocking);
+	}
+
+	void Prime(Audio::Id port) {
+		const auto start = now;
+		for (int i = 0; i < 8; i++) {
+			Output(port);
+		}
+		Check(now == start, "priming was paced before the 40 ms cushion filled");
+		Output(port);
+		Check(now - start >= 5333 && now - start <= 5334,
+		      "priming accumulated a multi-block deadline");
+		Drain(*streams.front());
+		Check(streams.front()->frames >= uint64_t {7 * 256} * 1000000,
+		      "first paced output consumed the priming cushion");
+	}
+};
+
+void TestPrimingAndSteadyCadence() {
+	Fixture    f;
+	const auto port = f.Open();
+	f.Prime(port);
+	processing = 125;
+	oversleep  = 175;
+	for (int i = 0; i < 10; i++) {
+		f.Output(port);
+	}
+	const auto start = now;
+	for (int i = 0; i < 3000; i++) {
+		f.Output(port);
+	}
+	Check(now - start >= 15999999 && now - start <= 16000001,
+	      "processing, sleep overshoot, or fractional periods accumulated timing drift");
+}
+
+void TestUnderrunAndStalledQueueRecovery() {
+	Fixture    f;
+	const auto port = f.Open();
+	f.Prime(port);
+	now += 1000000;
+	f.Prime(port);
+
+	// A stopped device must clear its stale queue and restart priming, not catch up old deadlines.
+	stalled = true;
+	for (int i = 0; i < 3 && clears == 0; i++) {
+		f.Output(port);
+	}
+	Check(clears == 1, "stalled playback queue was not cleared");
+	stalled          = false;
+	const auto start = now;
+	for (int i = 0; i < 7; i++) {
+		f.Output(port);
+	}
+	Check(now == start, "queue clear did not restart unpaced priming");
+	f.Output(port);
+	Check(now - start >= 5333 && now - start <= 5334,
+	      "queue clear retained an old pacing deadline");
+}
+
+void TestFallbackClockAndLongGap() {
+	Fixture f;
+	fail_open       = true;
+	const auto port = f.Open();
+	f.Output(port);
+	const auto start = now;
+	oversleep        = 175;
+	for (int i = 0; i < 3000; i++) {
+		now += 125;
+		f.Output(port);
+	}
+	Check(now - start >= 16000174 && now - start <= 16000176,
+	      "device-free output accumulated timing drift");
+	now += 1000000;
+	f.Output(port);
+	const auto resumed = now;
+	f.Output(port);
+	Check(now - resumed >= 5333 && now - resumed <= 5509,
+	      "long interruption caused a burst of device-free catch-up submissions");
+}
+
+void TestAsyncDoesNotAccumulateDeadlines() {
+	Fixture    f;
+	const auto port  = f.Open();
+	const auto start = now;
+	for (int i = 0; i < 8; i++) {
+		f.Output(port, false);
+	}
+	Check(now == start, "asynchronous output slept");
+	f.Output(port);
+	Check(now - start <= 5334, "asynchronous submissions accumulated a blocking deadline");
+	for (int i = 0; i < 3; i++) {
+		f.Output(port);
+	}
+	now += 1000000;
+	f.Output(port, false);
+	const auto resumed = now;
+	for (int i = 0; i < 7; i++) {
+		f.Output(port);
+	}
+	Check(now == resumed, "async-to-blocking transition did not reset an emptied queue");
+}
+
+void TestSynchronizedBatchAndInactivePorts() {
+	Fixture            f;
+	const auto         vibration  = f.Open(256, 10);
+	const auto         main       = f.Open();
+	const auto         background = f.Open();
+	Audio::OutputParam batch[] {
+	    {vibration, f.pcm.data()}, {main, f.pcm.data()}, {background, f.pcm.data()}};
+	const auto start = now;
+	for (int i = 0; i < 8; i++) {
+		f.audio.AudioOutOutputs(batch, 3);
+	}
+	Check(now == start, "vibration fallback slowed a device-backed batch's priming");
+	for (int i = 0; i < 3; i++) {
+		sleeps.clear();
+		f.audio.AudioOutOutputs(batch, 3);
+		Check(sleeps.size() == 1 && sleeps.front() >= 5333 && sleeps.front() <= 5334,
+		      "synchronized streams accumulated separate pacing waits");
+	}
+
+	Audio::OutputParam inactive[] {{main, nullptr}, {vibration, f.pcm.data()}};
+	now += 1000000;
+	f.audio.AudioOutOutputs(inactive, 2);
+	const auto resumed = now;
+	f.audio.AudioOutOutputs(inactive, 2);
+	Check(now - resumed >= 5333 && now - resumed <= 5334,
+	      "a device without PCM disabled fallback pacing");
+}
+
+void TestFallbackUsesEachPortsPeriod() {
+	Fixture f;
+	fail_open                     = true;
+	const auto         short_port = f.Open(128);
+	const auto         long_port  = f.Open(256);
+	Audio::OutputParam batch[] {{short_port, f.pcm.data()}, {long_port, f.pcm.data()}};
+	f.audio.AudioOutOutputs(batch, 2);
+	const auto start = now;
+	f.audio.AudioOutOutputs(batch, 2);
+	Check(now - start >= 5333 && now - start <= 5334,
+	      "fallback batch used the first port's period for every port");
+}
+
+void TestFailedQueueUsesFallbackClock() {
+	Fixture    f;
+	const auto port = f.Open();
+	fail_put        = true;
+	f.Output(port);
+	const auto start = now;
+	for (int i = 0; i < 3; i++) {
+		f.Output(port);
+	}
+	Check(now - start == 16000, "repeated queue failures reset fallback pacing");
+}
+
+void TestControllerSpeakerPacing() {
+	Fixture f;
+	fail_open     = true; // No PC audio device paces the pad speaker.
+	pad_connected = true;
+	pad_queue_us  = 80000;
+	const auto speaker = f.Open(256, 4);
+
+	pad_bluetooth = true;
+	f.Output(speaker);
+	const auto bluetooth_start = now;
+	f.Output(speaker);
+	Check(now - bluetooth_start >= 5333 && now - bluetooth_start <= 5334,
+	      "Bluetooth speaker waited on its HID queue instead of the sample clock");
+
+	pad_bluetooth = false;
+	const auto usb_start = now;
+	f.Output(speaker);
+	Check(now - usb_start >= 5333 && now - usb_start <= 5334,
+	      "stalled USB speaker waited longer than one audio block");
+}
+
+void TestInvalidBatchSize() {
+	Libs::Audio::AudioOut::AudioOutOutputParam param {};
+	for (const auto count: {0u, 33u}) {
+		Check(Libs::Audio::AudioOut::AudioOutOutputs(&param, count) ==
+		          Libs::Audio::AUDIO_OUT_ERROR_INVALID_SIZE,
+		      "invalid batch size was not rejected before accessing ports");
+	}
+}
+
+void TestZeroOutputFrequency() {
+	Fixture f;
+	Check(Libs::Audio::AudioOut::AudioOutOpen(0, 0, 0, 256, 0, 4) ==
+	          Libs::Audio::AUDIO_OUT_ERROR_INVALID_SAMPLE_FREQ,
+	      "public output open accepted zero frequency");
+	Check(!f.audio.AudioOutOpen(10, 256, 0, Audio::Format::FloatStereo).IsValid(),
+	      "internal output open accepted zero frequency");
+	Check(f.Open().ToInt() == 1, "rejected output port occupied a handle");
+}
+} // namespace
+
+int main() {
+	TestPrimingAndSteadyCadence();
+	TestUnderrunAndStalledQueueRecovery();
+	TestFallbackClockAndLongGap();
+	TestAsyncDoesNotAccumulateDeadlines();
+	TestSynchronizedBatchAndInactivePorts();
+	TestFallbackUsesEachPortsPeriod();
+	TestFailedQueueUsesFallbackClock();
+	TestControllerSpeakerPacing();
+	TestInvalidBatchSize();
+	TestZeroOutputFrequency();
+	Check(streams.empty(), "output stream leaked");
+	std::puts("AudioOutTimingTests: all cases passed");
+}

tests/MemoryTrackerTests.cpp

@@ -238,14 +238,31 @@ struct TrackerHarness {
   MemoryTracker tracker;
 };
 
+uint8_t *AllocateFixedGuestRange(uint64_t size, uintptr_t offset) {
+  // Keep the mapping inside the tracker's guest address space and preserve the
+  // requested region alignment. A fixed address can be occupied by the host
+  // process (notably by the macOS runner's ASLR layout).
+  constexpr uintptr_t first_base = 0x0000000200000000ull;
+  constexpr uintptr_t stride = 0x0000000100000000ull;
+  for (uintptr_t attempt = 0; attempt < 256; attempt++) {
+    auto *wanted = reinterpret_cast<void *>(first_base + offset + attempt * stride);
+    auto *memory = static_cast<uint8_t *>(
+        VirtualAlloc(wanted, size, MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE));
+    if (memory == wanted) {
+      return memory;
+    }
+    if (memory != nullptr) {
+      Check(VirtualFree(memory, 0, MEM_RELEASE) != 0,
+            "release unexpected fixed allocation failed");
+    }
+  }
+  Check(false, "no free fixed guest address found");
+  return nullptr;
+}
+
 uint8_t *Allocate(PageManager &manager, uint64_t pages) {
-  constexpr uintptr_t base = 0x0000000200010000ull;
   const auto size = manager.GetPageSize() * pages;
-  auto *memory = static_cast<uint8_t *>(
-      VirtualAlloc(reinterpret_cast<void *>(base), size,
-                   MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE));
-  Check(memory == reinterpret_cast<void *>(base), "fixed VirtualAlloc failed");
-  return memory;
+  return AllocateFixedGuestRange(size, 0x10000);
 }
 
 void Release(uint8_t *memory) {
@@ -386,16 +403,10 @@ void TestCpuDirtyUpload() {
 }
 
 void TestRangeInvalidation() {
-  constexpr uintptr_t base = 0x0000000201000000ull;
   TrackerHarness harness;
   auto &tracker = harness.tracker;
-  auto &page_manager = harness.page_manager;
   constexpr uint64_t size = Libs::Graphics::TRACKER_REGION_SIZE * 2;
-  auto *memory = static_cast<uint8_t *>(
-      VirtualAlloc(reinterpret_cast<void *>(base), size,
-                   MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE));
-  Check(memory == reinterpret_cast<void *>(base),
-        "range invalidation allocation failed");
+  auto *memory = AllocateFixedGuestRange(size, 0x1000000);
   const auto address = reinterpret_cast<uint64_t>(memory);
 
   tracker.ForEachUploadRange(
@@ -614,16 +625,12 @@ void TestGpuDownloadProtectionMirrors() {
 }
 
 void TestCrossRegionUpload() {
-  constexpr uintptr_t base = 0x0000000200010000ull;
   constexpr uint64_t region_size = 4ull * 1024ull * 1024ull;
   TrackerHarness harness;
   auto &tracker = harness.tracker;
   auto &page_manager = harness.page_manager;
   const auto page_size = page_manager.GetPageSize();
-  auto *memory = static_cast<uint8_t *>(
-      VirtualAlloc(reinterpret_cast<void *>(base), region_size * 2,
-                   MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE));
-  Check(memory == reinterpret_cast<void *>(base), "fixed VirtualAlloc failed");
+  auto *memory = AllocateFixedGuestRange(region_size * 2, 0x10000);
   const auto address = reinterpret_cast<uint64_t>(memory);
   const auto boundary = (address + region_size - 1) & ~(region_size - 1);
   uint32_t ranges = 0;
@@ -871,6 +878,15 @@ void TestFullRegionGpuUnmarkBatching() {
         [&]() noexcept {
           (void)tracker.IsRegionCpuModified(address, page_size);
         });
+  } else if (std::strcmp(name, "recursive-tracking-lock") == 0) {
+    Libs::Graphics::TrackingSpinLock lock;
+    lock.lock();
+    lock.lock();
+  } else if (std::strcmp(name, "non-owner-tracking-unlock") == 0) {
+    Libs::Graphics::TrackingSpinLock lock;
+    lock.lock();
+    std::thread worker([&] { lock.unlock(); });
+    worker.join();
   }
   std::_Exit(0x7f);
 }
@@ -928,7 +944,8 @@ void CheckDeathCase(const char *name) {
 }
 
 void TestFatalPaths() {
-  for (const char *name : {"gpu-dirty-explicit-cpu", "reentrant-upload"}) {
+  for (const char *name : {"gpu-dirty-explicit-cpu", "reentrant-upload",
+                           "recursive-tracking-lock", "non-owner-tracking-unlock"}) {
     CheckDeathCase(name);
   }
 }

tests/PadHapticsTests.cpp

@@ -1,11 +1,17 @@
 // Exercise the production module with real SDL streams and fake USB endpoints.
 #include <SDL3/SDL.h>
+#include <opus.h>
 #include <algorithm>
 #include <array>
+#include <atomic>
+#include <chrono>
 #include <cmath>
 #include <cstdio>
 #include <cstdlib>
 #include <memory>
+#include <numeric>
+#include <semaphore>
+#include <thread>
 #include <vector>
 
 namespace {
@@ -24,23 +30,32 @@ struct Rumble {
 	Uint16 large, small;
 	Uint32 duration;
 };
-std::vector<Device>                 devices;
-std::vector<SDL_AudioStream*>       streams;
-Rumble                              rumble {};
-Uint64                              now        = 1000;
-int                                 audio_refs = 0, opens = 0, rumble_calls = 0;
-bool                                fail_open = false, fail_resume = false;
-int                                 actual_channels = 4;
-SDL_AudioDeviceID                   opened_device   = 0;
-SDL_AudioSpec                       opened_spec {};
-std::array<Rumble, 4>               pad_rumble {};
-std::array<std::vector<uint8_t>, 4> effects;
-std::vector<SDL_JoystickID>         connected_pads {1};
-std::array<Rumble, 4>               cached_rumble {};
-bool                                fail_effect = false, wireless = false;
-SDL_AudioStream*                   default_stream = nullptr;
-bool                               fail_default_resume = false;
-int                                active_controller = 1;
+std::vector<Device>                   devices;
+std::vector<SDL_AudioStream*>         streams;
+Rumble                                rumble {};
+Uint64                                now        = 1000;
+int                                   audio_refs = 0, opens = 0, rumble_calls = 0;
+bool                                  fail_open = false, fail_resume = false;
+int                                   actual_channels = 4;
+SDL_AudioDeviceID                     opened_device   = 0;
+SDL_AudioSpec                         opened_spec {};
+std::array<Rumble, 4>                 pad_rumble {};
+std::array<std::vector<uint8_t>, 4>   effects;
+std::vector<SDL_JoystickID>           connected_pads {1};
+std::array<Rumble, 4>                 cached_rumble {};
+bool                                  fail_effect = false, wireless = false, wireless_pad3 = false;
+bool                                  hid_available = false, fail_hid_write = false;
+bool                                  duplicate_hid  = false;
+int                                   hid_opens      = 0;
+int                                   opened_hid_pad = 0;
+std::vector<std::array<uint8_t, 547>> hid_reports;
+std::atomic_bool                      block_opus_encode {false}, block_hid_write {false};
+std::atomic_int                       hid_closes {0};
+std::binary_semaphore                 encode_started {0}, allow_encode {0};
+std::binary_semaphore                 write_started {0}, allow_write {0};
+SDL_AudioStream*                      default_stream      = nullptr;
+bool                                  fail_default_resume = false;
+int                                   active_controller   = 1;
 } // namespace
 
 namespace Fake {
@@ -116,8 +131,9 @@ void DestroyAudioStream(SDL_AudioStream* stream) {
 		return;
 	}
 	const auto it = std::find(streams.begin(), streams.end(), stream);
-	Check(it != streams.end(), "stream destroyed twice");
-	streams.erase(it);
+	if (it != streams.end()) {
+		streams.erase(it);
+	}
 	if (stream == default_stream) {
 		default_stream = nullptr;
 	}
@@ -134,6 +150,73 @@ SDL_Gamepad* GetGamepadFromID(SDL_JoystickID id) {
 	return id == 1 || id == 3 ? reinterpret_cast<SDL_Gamepad*>(static_cast<uintptr_t>(id))
 	                          : nullptr;
 }
+SDL_GamepadType GetGamepadType(SDL_Gamepad* pad) {
+	return GetGamepadTypeForID(static_cast<SDL_JoystickID>(reinterpret_cast<uintptr_t>(pad)));
+}
+Uint16 GetGamepadVendor(SDL_Gamepad*) {
+	return 0x054c;
+}
+Uint16 GetGamepadProduct(SDL_Gamepad*) {
+	return 0x0ce6;
+}
+const char* GetGamepadSerial(SDL_Gamepad* pad) {
+	return reinterpret_cast<uintptr_t>(pad) == 3 ? "aa-bb-cc-dd-ee-03" : "aa-bb-cc-dd-ee-01";
+}
+Uint64 GetTicksNS() {
+	return now * 1000000;
+}
+SDL_Thread* CreateThread(SDL_ThreadFunction, const char*, void*) {
+	return reinterpret_cast<SDL_Thread*>(1);
+}
+void                 WaitThread(SDL_Thread*, int*) {}
+SDL_hid_device_info* HidEnumerate(unsigned short, unsigned short) {
+	if (!hid_available) {
+		return nullptr;
+	}
+	static wchar_t             serial1[] = L"AA:BB:CC:DD:EE:01";
+	static wchar_t             serial3[] = L"AA:BB:CC:DD:EE:03";
+	static char                path1[]   = "pad-1";
+	static char                path3[]   = "pad-3";
+	static SDL_hid_device_info pad3 {}, pad1 {};
+	pad3.path          = path3;
+	pad3.serial_number = duplicate_hid ? serial1 : serial3;
+	pad3.bus_type      = SDL_HID_API_BUS_BLUETOOTH;
+	pad1.path          = path1;
+	pad1.serial_number = serial1;
+	pad1.bus_type      = SDL_HID_API_BUS_BLUETOOTH;
+	pad1.next          = &pad3;
+	return &pad1;
+}
+void            HidFreeEnumeration(SDL_hid_device_info*) {}
+SDL_hid_device* HidOpenPath(const char* path) {
+	hid_opens++;
+	opened_hid_pad = path[4] == '1' ? 1 : 3;
+	return reinterpret_cast<SDL_hid_device*>(static_cast<uintptr_t>(opened_hid_pad));
+}
+void HidClose(SDL_hid_device*) {
+	hid_closes++;
+}
+int HidWrite(SDL_hid_device*, const unsigned char* data, size_t size) {
+	if (block_hid_write.exchange(false)) {
+		write_started.release();
+		allow_write.acquire();
+	}
+	if (fail_hid_write || size != 547) {
+		return -1;
+	}
+	std::array<uint8_t, 547> report {};
+	std::copy_n(data, report.size(), report.begin());
+	hid_reports.push_back(report);
+	return static_cast<int>(size);
+}
+opus_int32 OpusEncodeFloat(OpusEncoder* encoder, const float* pcm, int frames,
+                           unsigned char* packet, opus_int32 bytes) {
+	if (block_opus_encode.exchange(false)) {
+		encode_started.release();
+		allow_encode.acquire();
+	}
+	return ::opus_encode_float(encoder, pcm, frames, packet, bytes);
+}
 SDL_JoystickID* GetGamepads(int* count) {
 	*count    = static_cast<int>(connected_pads.size());
 	auto* ids = static_cast<SDL_JoystickID*>(
@@ -142,8 +225,10 @@ SDL_JoystickID* GetGamepads(int* count) {
 	ids[connected_pads.size()] = 0;
 	return ids;
 }
-SDL_JoystickConnectionState GetGamepadConnectionState(SDL_Gamepad*) {
-	return wireless ? SDL_JOYSTICK_CONNECTION_WIRELESS : SDL_JOYSTICK_CONNECTION_WIRED;
+SDL_JoystickConnectionState GetGamepadConnectionState(SDL_Gamepad* pad) {
+	return wireless || (wireless_pad3 && reinterpret_cast<uintptr_t>(pad) == 3)
+	           ? SDL_JOYSTICK_CONNECTION_WIRELESS
+	           : SDL_JOYSTICK_CONNECTION_WIRED;
 }
 bool RumbleGamepad(SDL_Gamepad* pad, Uint16 large, Uint16 small, Uint32 duration) {
 	const auto id = reinterpret_cast<uintptr_t>(pad);
@@ -173,22 +258,37 @@ bool SendGamepadEffect(SDL_Gamepad* pad, const void* data, int size) {
 }
 } // namespace Fake
 
-#define SDL_GetTicks                  Fake::GetTicks
-#define SDL_InitSubSystem             Fake::InitSubSystem
-#define SDL_QuitSubSystem             Fake::QuitSubSystem
-#define SDL_GetAudioPlaybackDevices   Fake::GetAudioPlaybackDevices
-#define SDL_GetAudioDeviceName        Fake::GetAudioDeviceName
-#define SDL_GetAudioDeviceFormat      Fake::GetAudioDeviceFormat
-#define SDL_OpenAudioDeviceStream     Fake::OpenAudioDeviceStream
-#define SDL_ResumeAudioStreamDevice   Fake::ResumeAudioStreamDevice
-#define SDL_DestroyAudioStream        Fake::DestroyAudioStream
-#define SDL_GetAudioStreamDevice      Fake::GetAudioStreamDevice
-#define SDL_GetGamepadTypeForID       Fake::GetGamepadTypeForID
+#define SDL_GetTicks                Fake::GetTicks
+#define SDL_InitSubSystem           Fake::InitSubSystem
+#define SDL_QuitSubSystem           Fake::QuitSubSystem
+#define SDL_GetAudioPlaybackDevices Fake::GetAudioPlaybackDevices
+#define SDL_GetAudioDeviceName      Fake::GetAudioDeviceName
+#define SDL_GetAudioDeviceFormat    Fake::GetAudioDeviceFormat
+#define SDL_OpenAudioDeviceStream   Fake::OpenAudioDeviceStream
+#define SDL_ResumeAudioStreamDevice Fake::ResumeAudioStreamDevice
+#define SDL_DestroyAudioStream      Fake::DestroyAudioStream
+#define SDL_GetAudioStreamDevice    Fake::GetAudioStreamDevice
+#define SDL_GetGamepadTypeForID     Fake::GetGamepadTypeForID
+#define SDL_GetGamepadType          Fake::GetGamepadType
+#define SDL_GetGamepadVendor        Fake::GetGamepadVendor
+#define SDL_GetGamepadProduct       Fake::GetGamepadProduct
+#define SDL_GetGamepadSerial        Fake::GetGamepadSerial
+#define SDL_GetTicksNS              Fake::GetTicksNS
+#undef SDL_CreateThread
+#define SDL_CreateThread              Fake::CreateThread
+#define SDL_WaitThread                Fake::WaitThread
+#define SDL_hid_enumerate             Fake::HidEnumerate
+#define SDL_hid_free_enumeration      Fake::HidFreeEnumeration
+#define SDL_hid_open_path             Fake::HidOpenPath
+#define SDL_hid_close                 Fake::HidClose
+#define SDL_hid_write                 Fake::HidWrite
+#define opus_encode_float             Fake::OpusEncodeFloat
 #define SDL_GetGamepadFromID          Fake::GetGamepadFromID
 #define SDL_RumbleGamepad             Fake::RumbleGamepad
 #define SDL_SendGamepadEffect         Fake::SendGamepadEffect
 #define SDL_GetGamepads               Fake::GetGamepads
 #define SDL_GetGamepadConnectionState Fake::GetGamepadConnectionState
+#include "libs/dualSenseBluetooth.cpp"
 #include "libs/dualSenseHaptics.cpp"
 #include "libs/audio.cpp"
 #undef SDL_GetTicks
@@ -202,6 +302,19 @@ bool SendGamepadEffect(SDL_Gamepad* pad, const void* data, int size) {
 #undef SDL_DestroyAudioStream
 #undef SDL_GetAudioStreamDevice
 #undef SDL_GetGamepadTypeForID
+#undef SDL_GetGamepadType
+#undef SDL_GetGamepadVendor
+#undef SDL_GetGamepadProduct
+#undef SDL_GetGamepadSerial
+#undef SDL_GetTicksNS
+#undef SDL_CreateThread
+#undef SDL_WaitThread
+#undef SDL_hid_enumerate
+#undef SDL_hid_free_enumeration
+#undef SDL_hid_open_path
+#undef SDL_hid_close
+#undef SDL_hid_write
+#undef opus_encode_float
 #undef SDL_GetGamepadFromID
 #undef SDL_RumbleGamepad
 #undef SDL_SendGamepadEffect
@@ -228,6 +341,7 @@ double GetTimeMs() {
 
 namespace {
 namespace Haptics = Libs::Controller::DualSenseHaptics;
+namespace Bluetooth = Libs::Controller::DualSenseBluetooth;
 using Port        = std::unique_ptr<Haptics::Stream, decltype(&Haptics::Close)>;
 constexpr std::array<int, 2>   unity {32768, 32768};
 constexpr std::array<float, 4> pcm {0.5f, 0.5f, 0.5f, 0.5f};
@@ -240,11 +354,17 @@ struct Fixture {
 		cached_rumble  = {};
 		effects        = {};
 		connected_pads = {1};
-		fail_effect = wireless = false;
-		default_stream = nullptr;
+		fail_effect = wireless = wireless_pad3 = false;
+		hid_available = fail_hid_write = false;
+		duplicate_hid                  = false;
+		hid_opens                      = 0;
+		hid_closes                     = 0;
+		opened_hid_pad                 = 0;
+		hid_reports.clear();
+		default_stream      = nullptr;
 		fail_default_resume = false;
-		active_controller = 1;
-		now                    = 1000;
+		active_controller   = 1;
+		now                 = 1000;
 		opens = rumble_calls = 0;
 		fail_open = fail_resume = false;
 		actual_channels         = 4;
@@ -255,8 +375,8 @@ struct Fixture {
 	}
 };
 
-Port Open(bool speaker = false) {
-	auto* stream = Haptics::Open(48000, speaker);
+Port Open(bool speaker = false, uint32_t freq = 48000) {
+	auto* stream = Haptics::Open(freq, speaker);
 	Check(stream != nullptr, "port open failed");
 	return Port(stream, Haptics::Close);
 }
@@ -516,6 +636,446 @@ void TestSpeakerRequiresUnambiguousUsbDevice() {
 	      "USB to Bluetooth change retained the USB speaker");
 }
 
+void TestBluetoothAudioAndIdentity() {
+	Fixture f;
+	wireless = hid_available = true;
+	connected_pads           = {1, 3};
+	devices.clear(); // Bluetooth does not enumerate a quad USB audio endpoint.
+	auto               vibration = Open();
+	auto               speaker   = Open(true);
+	std::vector<float> haptic(2048, 0.5f);
+	std::vector<float> sound(2048, 0.25f);
+	Check(Haptics::SetVibration(1, 100, 80) && pad_rumble[1].large == 100 * 0x101,
+	      "Bluetooth controller did not receive normal rumble");
+	Check(Haptics::Queue(vibration.get(), 1, haptic.data(), 1024, 2, true, unity.data()) != 0,
+	      "Bluetooth haptics did not open");
+	Check(Haptics::Queue(speaker.get(), 1, sound.data(), 1024, 2, true, unity.data()) != 0,
+	      "Bluetooth speaker did not open");
+	Check(hid_opens == 1 && opened_hid_pad == 1 && effects[1][7] == 0x30,
+	      "Bluetooth audio was not matched to the active controller");
+	Check(pad_rumble[1].large == 0, "Bluetooth haptics did not suppress motor rumble");
+	auto pending = Libs::Controller::DualSenseBluetooth::Prepare(now * 1000000);
+	Check(pending.size() == 1, "Bluetooth streams were not combined in one report");
+	Libs::Controller::DualSenseBluetooth::Send(pending);
+	Check(hid_reports.size() == 1, "Bluetooth audio report was not sent");
+	const auto& report = hid_reports.back();
+	Check(report[0] == 0x39 && report[2] == 0x91 && report[3] == 6 && report[4] == 0x7e &&
+	          report[10] == 0xd2 && report[11] == 64 && report[140] == 0xd3 && report[141] == 200,
+	      "Bluetooth audio sub-packets are malformed");
+	Check(report[12] > 0 && report[13] > 0,
+	      "haptic waveform was not downsampled into the actuator channels");
+	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++) {
+		Check(report[report.size() - 4 + i] == static_cast<uint8_t>(crc >> (8 * i)),
+		      "Bluetooth audio CRC is invalid");
+	}
+	int   opus_error = OPUS_OK;
+	auto* decoder    = opus_decoder_create(48000, 2, &opus_error);
+	Check(decoder != nullptr && opus_error == OPUS_OK, "Opus decoder unavailable in test");
+	std::array<float, 960> decoded {};
+	Check(opus_decode_float(decoder, report.data() + 142, 200, decoded.data(), 480, 0) == 480,
+	      "Bluetooth speaker packet could not be decoded");
+	opus_decoder_destroy(decoder);
+	const auto mean = std::accumulate(decoded.begin(), decoded.end(), 0.0f) / decoded.size();
+	Check(mean > 0.05f && mean < 0.35f, "Bluetooth speaker audio was not preserved");
+	Check(Libs::Controller::DualSenseBluetooth::Prepare(now * 1000000).empty(),
+	      "Bluetooth sender ignored report pacing");
+	now += 300;
+	Haptics::UpdateBluetoothRumble();
+	Check(pad_rumble[1].large == 100 * 0x101,
+	      "Bluetooth haptics did not restore normal rumble afterward");
+
+	// A different active DualSense has a different HID address, even if both are
+	// paired.
+	Check(Haptics::Queue(vibration.get(), 3, haptic.data(), 1024, 2, true, unity.data()) != 0 &&
+	          opened_hid_pad == 3,
+	      "Bluetooth stream followed the wrong controller");
+	Check(Haptics::Queue(speaker.get(), 3, sound.data(), 1024, 2, true, unity.data()) != 0,
+	      "Bluetooth speaker did not follow the selected controller");
+	Check(effects[1][7] == 0, "old Bluetooth controller remained speaker-routed");
+	wireless       = false;
+	connected_pads = {3};
+	devices        = {{20, "Speakers (DualSense Wireless Controller)", 4}};
+	Check(Haptics::Queue(speaker.get(), 3, sound.data(), 1024, 2, true, unity.data()) != 0 &&
+	          opened_device == 20 && effects[3][7] == 0x30,
+	      "Bluetooth to USB change did not reopen the USB speaker immediately");
+	speaker.reset();
+	vibration.reset();
+}
+
+void TestBluetoothFailureFallsBack() {
+	Fixture f;
+	wireless = hid_available   = true;
+	auto               speaker = Open(true);
+	std::vector<float> sound(2048, 0.25f);
+	Check(Haptics::Queue(speaker.get(), 1, sound.data(), 1024, 2, true, unity.data()) != 0,
+	      "Bluetooth port did not initially open");
+	fail_hid_write = true;
+	auto pending   = Libs::Controller::DualSenseBluetooth::Prepare(now * 1000000);
+	Libs::Controller::DualSenseBluetooth::Send(pending);
+	Check(Haptics::Queue(speaker.get(), 1, sound.data(), 1024, 2, true, unity.data()) == 0 &&
+	          effects[1][7] == 0,
+	      "failed HID write did not restore the normal audio fallback");
+	speaker.reset();
+}
+
+void TestBluetoothFormatsAndRoutingFailure() {
+	Fixture f;
+	wireless = hid_available = fail_effect = true;
+	auto speaker                           = Open(true);
+	Check(Haptics::Queue(speaker.get(), 1, pcm.data(), 2, 2, true, unity.data()) == 0 &&
+	          !Haptics::UsesBluetooth(speaker.get()) && hid_closes == 1 && effects[1].empty(),
+	      "failed Bluetooth speaker routing leaked a session or suppressed fallback");
+	fail_effect = false;
+	now += 2001;
+	std::vector<int16_t> mono(1024, 16384);
+	const int            volume = 16384;
+	Queue(speaker, mono.data(), 1024, 1, false, &volume);
+	auto pending = Bluetooth::Prepare(now * 1000000);
+	Check(pending.size() == 1 && hid_opens == 2 && effects[1][7] == 0x30 &&
+	          std::abs(pending[0].audio[200] - 0.25f) < 1e-4f &&
+	          std::abs(pending[0].audio[201] - 0.25f) < 1e-4f,
+	      "Bluetooth mono int16 volume mapping or speaker routing recovery failed");
+}
+
+void TestBluetoothEncodingDoesNotBlockAudioQueue() {
+	Fixture f;
+	wireless = hid_available = true;
+	auto speaker = Open(true);
+	std::vector<float> sound(2048, 0.25f);
+	Check(Haptics::Queue(speaker.get(), 1, sound.data(), 1024, 2, true, unity.data()) != 0,
+	      "Bluetooth speaker did not open for concurrent encoding test");
+	block_opus_encode = true;
+	std::thread encoder([&] {
+		auto pending = Libs::Controller::DualSenseBluetooth::Prepare(now * 1000000);
+		Libs::Controller::DualSenseBluetooth::Send(pending);
+	});
+	encode_started.acquire();
+	std::binary_semaphore queue_finished {0};
+	std::atomic_bool      queue_ok {false};
+	std::thread audio([&] {
+		queue_ok = Haptics::Queue(speaker.get(), 1, sound.data(), 1024, 2, true,
+		                          unity.data()) != 0;
+		queue_finished.release();
+	});
+	const bool queued_without_encoder =
+	    queue_finished.try_acquire_for(std::chrono::seconds(2));
+	allow_encode.release();
+	audio.join();
+	encoder.join();
+	Check(queued_without_encoder && queue_ok,
+	      "Bluetooth audio queue waited for Opus encoding");
+	speaker.reset();
+}
+
+void TestBluetoothCloseCancelsPreparedReport() {
+	Fixture f;
+	wireless = hid_available     = true;
+	auto               speaker   = Open(true);
+	auto               vibration = Open();
+	std::vector<float> sound(2048, 0.25f);
+	Queue(speaker, sound.data(), 1024);
+	Queue(vibration, sound.data(), 1024);
+	const auto base    = now * 1000000;
+	auto       pending = Bluetooth::Prepare(base);
+	Check(pending.size() == 1, "Bluetooth report was not prepared before stream close");
+	speaker.reset();
+	Bluetooth::Send(pending);
+	Check(hid_reports.empty(), "closed speaker's prepared report was sent");
+	Queue(vibration, sound.data(), 1024);
+	Bluetooth::Send(Bluetooth::Prepare(base + Bluetooth::PERIOD_NS));
+	Check(hid_reports.size() == 1 && hid_reports.back()[140] == 0,
+	      "closing the speaker prevented the remaining haptic stream from sending");
+}
+
+void TestBluetoothCloseDuringEncoding() {
+	Fixture f;
+	wireless = hid_available   = true;
+	auto               speaker = Open(true);
+	std::vector<float> sound(2048, 0.25f);
+	Queue(speaker, sound.data(), 1024);
+	block_opus_encode = true;
+	std::thread encoder([&] { Bluetooth::Send(Bluetooth::Prepare(now * 1000000)); });
+	encode_started.acquire();
+	std::binary_semaphore close_finished {0};
+	std::thread           closer([&] {
+		speaker.reset();
+		close_finished.release();
+	});
+	const bool closed_without_encoder  = close_finished.try_acquire_for(std::chrono::seconds(2));
+	const bool hid_alive_during_encode = hid_closes == 0;
+	allow_encode.release();
+	closer.join();
+	encoder.join();
+	Check(closed_without_encoder && hid_alive_during_encode && hid_closes == 1 &&
+	          hid_reports.empty(),
+	      "closing during encoding blocked, destroyed a live session, or sent stale audio");
+}
+
+void TestBluetoothCloseWaitsForWriteWithoutBlockingQueue() {
+	Fixture f;
+	wireless = hid_available     = true;
+	auto               speaker   = Open(true);
+	auto               vibration = Open();
+	std::vector<float> sound(2048, 0.25f);
+	Queue(speaker, sound.data(), 1024);
+	Queue(vibration, sound.data(), 1024);
+	auto pending = Bluetooth::Prepare(now * 1000000);
+	Check(pending.size() == 1, "Bluetooth report was not prepared before blocked write");
+	block_hid_write = true;
+	std::thread sender([&] { Bluetooth::Send(pending); });
+	write_started.acquire();
+	std::binary_semaphore close_started {0}, close_finished {0}, queue_finished {0};
+	std::thread           closer([&] {
+		close_started.release();
+		speaker.reset();
+		close_finished.release();
+	});
+	close_started.acquire();
+	const bool closed_before_write = close_finished.try_acquire_for(std::chrono::milliseconds(50));
+	std::atomic_bool queue_ok {false};
+	std::thread      audio([&] {
+		queue_ok =
+		    Haptics::Queue(vibration.get(), 1, sound.data(), 1024, 2, true, unity.data()) != 0;
+		queue_finished.release();
+	});
+	const bool       queued_without_write = queue_finished.try_acquire_for(std::chrono::seconds(2));
+	allow_write.release();
+	audio.join();
+	closer.join();
+	sender.join();
+	Check(!closed_before_write && queued_without_write && queue_ok && hid_reports.size() == 1 &&
+	          effects[1][7] == 0,
+	      "close failed to drain the HID write independently of the audio queue");
+}
+
+void TestBluetoothSpeakerWaitsForItsAudioBlock() {
+	Fixture f;
+	wireless = hid_available = true;
+	auto vibration = Open();
+	auto speaker = Open(true);
+	std::vector<float> haptic(2048, 0.5f);
+	std::vector<float> sound(2048, 0.25f);
+	Check(Haptics::Queue(vibration.get(), 1, haptic.data(), 1024, 2, true, unity.data()) != 0,
+	      "Bluetooth vibration stream did not open");
+	Check(Libs::Controller::DualSenseBluetooth::Prepare(now * 1000000).empty(),
+	      "haptics sent a silent speaker packet before the audio batch arrived");
+	Check(Haptics::Queue(speaker.get(), 1, sound.data(), 256, 2, true, unity.data()) != 0,
+	      "Bluetooth speaker did not accept a short block");
+	Check(Libs::Controller::DualSenseBluetooth::Prepare(now * 1000000).empty(),
+	      "short Bluetooth speaker block was padded with silence immediately");
+	Check(Haptics::Queue(speaker.get(), 1, sound.data(), 768, 2, true, unity.data()) != 0,
+	      "Bluetooth speaker did not accept the remainder of its block");
+	auto reports = Libs::Controller::DualSenseBluetooth::Prepare(now * 1000000);
+	Check(reports.size() == 1 && reports[0].audio[100] > 0.2f && reports[0].data[12] > 0,
+	      "Bluetooth speaker and haptics were not combined into a complete report");
+	speaker.reset();
+	vibration.reset();
+}
+
+void TestBluetoothShortHapticBlock() {
+	Fixture f;
+	wireless = hid_available     = true;
+	auto               vibration = Open();
+	std::vector<float> sound(2048, 0.25f);
+	Queue(vibration, sound.data(), 256);
+	const auto base = now * 1000000;
+	Check(Bluetooth::Prepare(base).empty() && Bluetooth::Prepare(base + 1000000).empty(),
+	      "short haptic block was padded with silence before its remainder arrived");
+	Queue(vibration, sound.data(), 768);
+	Check(Bluetooth::Prepare(base + 1000000).empty(),
+	      "completed haptic block skipped its speaker coalescing window");
+	auto pending = Bluetooth::Prepare(base + 2000000);
+	Check(pending.size() == 1 && pending[0].data[12 + 100] > 0,
+	      "complete haptic block did not play after its remainder arrived");
+}
+
+void TestBluetoothIncrementalBatchCoalesces() {
+	Fixture f;
+	wireless = hid_available = true;
+	auto                   vibration = Open();
+	auto speaker = Open(true);
+	std::array<float, 512> sound;
+	sound.fill(0.25f);
+	for (int block = 0; block < 3; ++block) {
+		Queue(vibration, sound.data(), 256);
+		Queue(speaker, sound.data(), 256);
+		Check(Bluetooth::Prepare(now * 1000000).empty(),
+		      "incremental Bluetooth batch was padded before it was complete");
+		now += 5;
+	}
+	Queue(vibration, sound.data(), 256);
+	Check(Bluetooth::Prepare(now * 1000000).empty(),
+	      "completed haptics sent before the same batch's final speaker block arrived");
+	Queue(speaker, sound.data(), 256);
+	auto pending = Bluetooth::Prepare(now * 1000000);
+	Check(pending.size() == 1 && pending[0].audio[1800] > 0.2f && pending[0].data[12 + 100] > 0,
+	      "completed incremental speaker and haptics batch did not play immediately");
+}
+
+void TestBluetoothLowRateCompleteBlock() {
+	Fixture f;
+	wireless = hid_available           = true;
+	auto                       speaker = Open(true, 8000);
+	std::array<float, 171 * 2> sound;
+	sound.fill(0.25f);
+	Queue(speaker, sound.data(), 171);
+	auto pending = Bluetooth::Prepare(now * 1000000);
+	Check(pending.size() == 1 && pending[0].audio[1000] > 0.2f,
+	      "resampler lookahead delayed a complete low-rate speaker block");
+}
+
+void TestBluetoothCompleteStreamDoesNotWaitForPartialPeer() {
+	Fixture f;
+	wireless = hid_available    = true;
+	auto               complete = Open(true);
+	auto               partial  = Open(true);
+	std::vector<float> sound(2048, 0.25f);
+	Queue(complete, sound.data(), 1024);
+	Queue(partial, sound.data(), 256);
+	auto pending = Bluetooth::Prepare(now * 1000000);
+	Check(pending.size() == 1 && pending[0].audio[100] > 0.4f && pending[0].audio[1000] > 0.2f,
+	      "complete speaker stream waited for a partial peer instead of mixing available audio");
+}
+
+void TestBluetoothShortFinalBlock() {
+	for (bool speaker: {false, true}) {
+		Fixture f;
+		wireless = hid_available = true;
+		auto               port  = Open(speaker);
+		std::vector<float> sound(2048, 0.25f);
+		Queue(port, sound.data(), 256);
+		const auto base = now * 1000000;
+		Check(Bluetooth::Prepare(base).empty(),
+		      "short final Bluetooth block sent before more audio could arrive");
+		Check(Bluetooth::Prepare(base + Bluetooth::PERIOD_NS).size() == 1,
+		      "short final Bluetooth block was never played");
+	}
+}
+
+void TestBluetoothOverflowKeepsRecentAudio() {
+	Fixture f;
+	wireless = hid_available = true;
+	auto speaker = Open(true);
+	std::vector<float> sound(2048);
+	for (int block = 1; block <= 4; ++block) {
+		std::fill(sound.begin(), sound.end(), block * 0.1f);
+		Check(Haptics::Queue(speaker.get(), 1, sound.data(), 1024, 2, true,
+		                     unity.data()) != 0,
+		      "Bluetooth speaker overflow queue failed");
+	}
+	auto first = Libs::Controller::DualSenseBluetooth::Prepare(now * 1000000);
+	Check(first.size() == 1 && first[0].audio[200] > 0.15f && first[0].audio[200] < 0.25f,
+	      "Bluetooth queue overflow lost all buffered audio instead of the oldest block");
+	auto second = Libs::Controller::DualSenseBluetooth::Prepare(now * 1000000 +
+	                                                        Libs::Controller::DualSenseBluetooth::PERIOD_NS);
+	Check(second.size() == 1 && second[0].audio[200] > 0.25f && second[0].audio[200] < 0.35f,
+	      "Bluetooth queue overflow did not preserve the next audio block");
+	speaker.reset();
+}
+
+void TestBluetoothHapticOverflowKeepsRecentAudio() {
+	Fixture f;
+	wireless = hid_available = true;
+	auto vibration = Open();
+	std::vector<float> haptic(2048);
+	for (int block = 1; block <= 4; ++block) {
+		std::fill(haptic.begin(), haptic.end(), block * 0.1f);
+		Check(Haptics::Queue(vibration.get(), 1, haptic.data(), 1024, 2, true,
+		                     unity.data()) != 0,
+		      "Bluetooth haptic overflow queue failed");
+	}
+	const auto base = now * 1000000;
+	Check(Libs::Controller::DualSenseBluetooth::Prepare(base).empty(),
+	      "haptic-only report skipped its coalescing window");
+	auto first = Libs::Controller::DualSenseBluetooth::Prepare(base + 1000000);
+	Check(first.size() == 1 && first[0].data[12 + 32] > 18 && first[0].data[12 + 32] < 32,
+	      "Bluetooth haptic overflow discarded buffered waveform instead of one report");
+	vibration.reset();
+}
+
+void TestBluetoothRecoversAfterLateWrite() {
+	Fixture f;
+	wireless = hid_available = true;
+	auto speaker = Open(true);
+	std::vector<float> sound(2048, 0.25f);
+	for (int i = 0; i < 3; ++i) {
+		Check(Haptics::Queue(speaker.get(), 1, sound.data(), 1024, 2, true,
+		                     unity.data()) != 0,
+		      "Bluetooth speaker could not buffer three reports");
+	}
+	const auto base = now * 1000000;
+	const auto period = Libs::Controller::DualSenseBluetooth::PERIOD_NS;
+	Check(Libs::Controller::DualSenseBluetooth::Prepare(base).size() == 1,
+	      "first Bluetooth report was not prepared");
+	Check(Libs::Controller::DualSenseBluetooth::Prepare(base + period + 10000000).size() == 1,
+	      "late Bluetooth report was not prepared");
+	Check(Libs::Controller::DualSenseBluetooth::Prepare(base + 2 * period).size() == 1,
+	      "Bluetooth sender did not recover its schedule after a late write");
+	speaker.reset();
+}
+
+void TestUsbOverflowKeepsRecentAudio() {
+	Fixture f;
+	auto speaker = Open(true);
+	std::vector<float> sound(2048);
+	for (int block = 1; block <= 4; ++block) {
+		std::fill(sound.begin(), sound.end(), block * 0.1f);
+		Check(Haptics::Queue(speaker.get(), 1, sound.data(), 1024, 2, true,
+		                     unity.data()) != 0,
+		      "USB speaker overflow queue failed");
+	}
+	std::array<float, 4> first {};
+	Check(!streams.empty() &&
+	          SDL_GetAudioStreamData(streams.back(), first.data(), sizeof(first)) ==
+	              sizeof(first) &&
+	          first[0] > 0.15f && first[0] < 0.25f,
+	      "USB queue overflow did not preserve the newest buffered audio");
+	// A single oversized input must also preserve only its newest 80 ms.
+	sound.assign(4096 * 2, 0.1f);
+	std::fill(sound.begin() + 256 * 2, sound.end(), 0.5f);
+	const auto queued = Haptics::Queue(speaker.get(), 1, sound.data(), 4096, 2, true, unity.data());
+	Check(queued == 80000 && SDL_GetAudioStreamQueued(streams.back()) == 3840 * 4 * sizeof(float) &&
+	          SDL_GetAudioStreamData(streams.back(), first.data(), sizeof(first)) ==
+	              sizeof(first) &&
+	          first[0] == 0.5f && first[1] == 0.5f,
+	      "oversized USB block did not retain only the newest 80 ms");
+	speaker.reset();
+}
+
+void TestBluetoothAmbiguousDevice() {
+	Fixture f;
+	wireless = hid_available = duplicate_hid = true;
+	auto speaker                             = Open(true);
+	Check(Haptics::Queue(speaker.get(), 1, pcm.data(), 2, 2, true, unity.data()) == 0 &&
+	          hid_opens == 0 && effects[1].empty(),
+	      "ambiguous Bluetooth devices were routed arbitrarily");
+	speaker.reset();
+}
+
+void TestUsbAndBluetoothFeatureParity() {
+	Fixture f;
+	connected_pads = {1, 3};
+	wireless_pad3 = hid_available       = true;
+	auto               wired_speaker    = Open(true);
+	auto               wireless_speaker = Open(true);
+	std::vector<float> sound(2048, 0.25f);
+	Check(Haptics::Queue(wired_speaker.get(), 1, sound.data(), 1024, 2, true, unity.data()) != 0 &&
+	          opened_device == 10,
+	      "Bluetooth pad disabled the wired pad speaker");
+	Check(Haptics::Queue(wireless_speaker.get(), 3, sound.data(), 1024, 2, true, unity.data()) !=
+	              0 &&
+	          opened_hid_pad == 3,
+	      "wired pad disabled the Bluetooth speaker");
+	Check(Haptics::Queue(wired_speaker.get(), 1, sound.data(), 1024, 2, true, unity.data()) != 0 &&
+	          effects[1][7] == 0x30 && effects[3][7] == 0,
+	      "returning to the wired speaker left Bluetooth routing active");
+	wireless_speaker.reset();
+	wired_speaker.reset();
+}
+
 void TestAudioSpeakerFallback() {
 	Fixture f;
 	Libs::Audio::Audio audio;
@@ -545,6 +1105,26 @@ void TestAudioSpeakerRouting() {
 	ExpectPcm({0.5f, 0.5f, 0, 0, 0.5f, 0.5f, 0, 0});
 }
 
+void TestBluetoothAudioSpeakerRouting() {
+	Fixture f;
+	wireless = hid_available = true;
+	devices.clear();
+	Libs::Audio::Audio audio;
+	const auto port = audio.AudioOutOpen(4, 1024, 48000, Libs::Audio::Audio::Format::FloatStereo);
+	Check(port.IsValid() && default_stream != nullptr, "Bluetooth pad speaker port did not open");
+	std::vector<float> sound(2048, 0.25f);
+	Libs::Audio::Audio::OutputParam output {port, sound.data()};
+	Check(audio.AudioOutOutputs(&output, 1, false) == 1024 && hid_opens == 1 &&
+	          SDL_GetAudioStreamQueued(default_stream) == 0,
+	      "Bluetooth pad speaker also played on the host output");
+	fail_hid_write = true;
+	auto pending = Libs::Controller::DualSenseBluetooth::Prepare(now * 1000000);
+	Libs::Controller::DualSenseBluetooth::Send(pending);
+	Check(audio.AudioOutOutputs(&output, 1, false) == 1024 &&
+	          SDL_GetAudioStreamQueued(default_stream) > 0,
+	      "Bluetooth write failure did not use the normal speaker output");
+}
+
 void TestAudioDefaultResumeFailure() {
 	Fixture f;
 	Libs::Audio::Audio audio;
@@ -564,6 +1144,7 @@ int main() {
 	SDL_SetHint(SDL_HINT_AUDIO_DRIVER, "dummy");
 	TestAudioSpeakerFallback();
 	TestAudioSpeakerRouting();
+	TestBluetoothAudioSpeakerRouting();
 	TestAudioDefaultResumeFailure();
 	TestFormatsAndVolume();
 	TestDiscoveryAndHotplug();
@@ -571,6 +1152,25 @@ int main() {
 	TestSpeakerUnplug();
 	TestSpeakerSwitchAndFailures();
 	TestSpeakerRequiresUnambiguousUsbDevice();
+	TestBluetoothAudioAndIdentity();
+	TestBluetoothFailureFallsBack();
+	TestBluetoothFormatsAndRoutingFailure();
+	TestBluetoothEncodingDoesNotBlockAudioQueue();
+	TestBluetoothCloseCancelsPreparedReport();
+	TestBluetoothCloseDuringEncoding();
+	TestBluetoothCloseWaitsForWriteWithoutBlockingQueue();
+	TestBluetoothSpeakerWaitsForItsAudioBlock();
+	TestBluetoothShortHapticBlock();
+	TestBluetoothIncrementalBatchCoalesces();
+	TestBluetoothLowRateCompleteBlock();
+	TestBluetoothCompleteStreamDoesNotWaitForPartialPeer();
+	TestBluetoothShortFinalBlock();
+	TestBluetoothOverflowKeepsRecentAudio();
+	TestBluetoothHapticOverflowKeepsRecentAudio();
+	TestBluetoothRecoversAfterLateWrite();
+	TestUsbOverflowKeepsRecentAudio();
+	TestBluetoothAmbiguousDevice();
+	TestUsbAndBluetoothFeatureParity();
 	TestFailuresAndBoundedQueue();
 	TestRumbleLeaseAndDuration();
 	TestSwitchStopsOldRumble();

tests/PageManagerTests.cpp

@@ -192,13 +192,25 @@ uint8_t *Allocate(uint64_t size, uint32_t protection = PAGE_READWRITE,
   Check(memory == reinterpret_cast<void *>(test_address),
         "fixed low VirtualAlloc failed");
 #elif defined(__APPLE__)
-  mach_vm_address_t raw = test_address;
-  Check(mach_vm_allocate(mach_task_self(), &raw, size, VM_FLAGS_FIXED) ==
-            KERN_SUCCESS &&
-            mach_vm_protect(mach_task_self(), raw, size, false,
-                            static_cast<vm_prot_t>(ToHostProt(protection))) ==
-                KERN_SUCCESS,
-        "fixed low mach_vm_allocate failed");
+  mach_vm_address_t raw = 0;
+  bool allocated = false;
+  constexpr uintptr_t stride = 0x0000000100000000ull;
+  for (uintptr_t candidate = test_address;
+       Libs::Graphics::GuestRange{candidate, size}.Valid(); candidate += stride) {
+    raw = candidate;
+    if (mach_vm_allocate(mach_task_self(), &raw, size, VM_FLAGS_FIXED) !=
+        KERN_SUCCESS) {
+      continue;
+    }
+    if (mach_vm_protect(mach_task_self(), raw, size, false,
+                        static_cast<vm_prot_t>(ToHostProt(protection))) ==
+        KERN_SUCCESS) {
+      allocated = true;
+      break;
+    }
+    mach_vm_deallocate(mach_task_self(), raw, size);
+  }
+  Check(allocated, "no free fixed guest address found");
   auto *memory = reinterpret_cast<uint8_t *>(raw);
   AllocationSizes()[memory] = static_cast<size_t>(size);
 #else

tests/ScalarProvenanceTests.cpp

@@ -281,6 +281,61 @@ void TestInvariantAndDivergentPhi() {
         "divergent phi was accepted");
 }
 
+void TestOperandMutationLifecycle() {
+  Fixture fixture(6);
+  const auto definition = fixture.Emit(ValueOpcode::IAdd32, {Value(0x55u), Value(0u)});
+  auto *source = definition.TryInstruction();
+  const auto inserted = fixture.Emit(ValueOpcode::BitFieldInsert,
+      {definition, definition, Value(4u), Value(4u)});
+  Check(source->UseCount() == 2, "inline operands lost distinct reverse-use indices");
+  inserted.TryInstruction()->SetArg(1, Value(3u));
+  Check(source->UseCount() == 1, "inline operand mutation did not detach its old use");
+  source->ReplaceUsesWith(Value(0xaau));
+  uint32_t result = 0;
+  Check(source->UseCount() == 0 &&
+            SrtWalker(fixture.program, {}).Evaluate(inserted, result) && result == 0x3au,
+        "inline operand replacement changed the evaluated value");
+
+  for (const auto opcode : {ValueOpcode::GetImageResource, ValueOpcode::MakeImageAddress}) {
+    auto &large = fixture.program.value_storage.emplace_back(opcode);
+    auto &replacement = fixture.program.value_storage.emplace_back(opcode);
+    const auto count = large.NumArgs();
+    for (size_t index = 0; index < count; ++index) {
+      large.SetArg(index, definition);
+      replacement.SetArg(index, Value(static_cast<uint32_t>(index)));
+    }
+    const auto user = fixture.Emit(ValueOpcode::Identity, {Value(&large)}, 0, 1);
+    Check(source->UseCount() == count, "large operands lost reverse-use indices");
+    large.ReplaceUsesWith(Value(&replacement));
+    Check(source->UseCount() == 0 && large.GetOpcode() == ValueOpcode::Identity &&
+              large.NumArgs() == 1 && large.Arg(0) == Value(&replacement) &&
+              user.Resolve() == Value(&replacement) && replacement.UseCount() == 2,
+          "large operand replacement did not preserve users or detach old operands");
+    large.Invalidate();
+    large.Invalidate();
+    Check(large.NumArgs() == 0 && replacement.UseCount() == 1,
+          "repeated invalidation detached a surviving large-value user");
+  }
+
+  auto &phi = fixture.BlockAt(5).AppendNewInst(ValueOpcode::Phi);
+  phi.SetFlags(Type::U32);
+  for (size_t index = 0; index < 5; ++index) {
+    phi.AddPhiOperand(&fixture.BlockAt(index), definition);
+  }
+  phi.SetArg(3, Value(7u));
+  Check(phi.NumArgs() == 5 && phi.NumPhiBlocks() == 5 && source->UseCount() == 4,
+        "Phi mutation lost its incoming values or reverse uses");
+  for (size_t index = 0; index < 5; ++index) {
+    Check(phi.PhiBlock(index) == &fixture.BlockAt(index),
+          "Phi mutation changed an incoming predecessor");
+  }
+  const auto sum = fixture.Emit(ValueOpcode::IAdd32, {Value(&phi), Value(5u)}, 0, 5);
+  phi.ReplaceUsesWith(Value(6u));
+  Check(source->UseCount() == 0 && phi.NumPhiBlocks() == 0 && phi.NumArgs() == 1 &&
+            SrtWalker(fixture.program, {}).Evaluate(sum, result) && result == 11,
+        "Phi replacement retained incoming uses or changed the consumer value");
+}
+
 void TestControlDependentStandaloneLoadStaysTyped() {
   Fixture fixture(3);
   const auto memory = fixture.AddMemory(ResourceKind::ScalarAddress);
@@ -634,6 +689,7 @@ int main() {
     TestShaderBaseAndUserData();
     TestCarryAndBitFields();
     TestInvariantAndDivergentPhi();
+    TestOperandMutationLifecycle();
     TestControlDependentStandaloneLoadStaysTyped();
     TestRuntime64BitDescriptorOps();
     TestUniformFirstLaneSamplerLod();

tests/ShaderRecompilerComputeTests.cpp

@@ -4208,7 +4208,30 @@ public:
         cache.FillBuffer(base + offset, sizeof(window_value), window_value,
                          false);
       }
+      const auto window_publication_tick = scheduler.CurrentTick();
+      std::binary_semaphore window_publication_entered{0};
+      std::binary_semaphore release_window_publication{0};
+      std::atomic<bool> window_readback_returned{false};
+      scheduler.DeferPriorityOperation([&] {
+        window_publication_entered.release();
+        release_window_publication.acquire();
+        Libs::LibKernel::Memory::WriteBacking(
+            base + window_inside_offset, &window_stale, sizeof(window_stale));
+      });
+      std::jthread release_window_callback([&] {
+        window_publication_entered.acquire();
+        const auto deadline = std::chrono::steady_clock::now() +
+                              std::chrono::milliseconds(100);
+        while (!window_readback_returned.load() &&
+               std::chrono::steady_clock::now() < deadline) {
+          std::this_thread::yield();
+        }
+        release_window_publication.release();
+      });
       cache.ReadMemory(base + window_fault_offset, sizeof(window_value));
+      window_readback_returned = true;
+      release_window_callback.join();
+      scheduler.WaitPriorityOperations(window_publication_tick);
       uint32_t window_inside_backing = 0;
       uint32_t window_outside_backing = 0;
       Libs::LibKernel::Memory::TryReadBacking(base + window_inside_offset,
@@ -4217,7 +4240,7 @@ public:
       Libs::LibKernel::Memory::TryReadBacking(base + window_outside_offset,
                                               &window_outside_backing,
                                               sizeof(window_outside_backing));
-      Require(name, "widened-window boundary",
+      Require(name, "widened-window publication order and boundary",
               window_inside_backing == window_value &&
                   window_outside_backing == window_stale &&
                   !cache.HasGpuDirtyBytes(base + window_inside_offset,
@@ -4228,7 +4251,8 @@ public:
                                              sizeof(window_value)) &&
                   cache.IsRegionGpuModified(base + window_outside_offset,
                                             sizeof(window_value)),
-              "readback did not honor the clamped half-open 512 KiB window");
+              "readback did not publish after the older overlapping callback "
+              "or honor the clamped half-open 512 KiB window");
       cache.ReadMemory(base + window_outside_offset, sizeof(window_value));
 
       Libs::LibKernel::Memory::WriteBacking(base + first_offset, &first_stale,