minor cleanup

audio: preserve NGS2 streaming buffer ownership

CMakeLists.txt

@@ -40,6 +40,11 @@ if (NOT (WIN32 OR LINUX OR APPLE))
     message(FATAL_ERROR "only Windows, Linux, and macOS builds are supported")
 endif()
 
+if(WIN32)
+	# Prevent Windows headers from defining min/max macros in all targets.
+	add_compile_definitions(NOMINMAX)
+endif()
+
 set(CMAKE_CXX_STANDARD 20)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_CXX_EXTENSIONS OFF)
@@ -298,28 +303,27 @@ list(APPEND kyty_emulator_src "${fault_buffer_shader_header}")
 list(APPEND kyty_emulator_src ${gpu_tiler_shader_headers})
 
 set(gpu_blit_generated_dir "${PROJECT_BINARY_DIR}/gpu_blit_shaders")
-set(gpu_blit_shader_sources
-	"${gpu_tiler_shader_dir}/gpu_blit_fs_triangle.vert"
-	"${gpu_tiler_shader_dir}/gpu_blit_color_to_ms_depth.frag"
-	"${gpu_tiler_shader_dir}/gpu_video_out_overlay.frag"
-)
-foreach(shader_source IN LISTS gpu_blit_shader_sources)
-	get_filename_component(shader_name "${shader_source}" NAME_WE)
+function(add_gpu_blit_shader shader_name shader_file)
+	set(shader_source "${gpu_tiler_shader_dir}/${shader_file}")
 	set(shader_spv "${gpu_blit_generated_dir}/${shader_name}.spv")
 	set(shader_header "${gpu_blit_generated_dir}/${shader_name}_spv.h")
 	string(TOUPPER "${shader_name}_SPV" shader_symbol)
 	add_custom_command(
 		OUTPUT "${shader_header}"
 		COMMAND ${CMAKE_COMMAND} -E make_directory "${gpu_blit_generated_dir}"
-		COMMAND "${KYTY_GLSLANG_VALIDATOR}" -V --target-env vulkan1.3 -Os
+		COMMAND "${KYTY_GLSLANG_VALIDATOR}" -V --target-env vulkan1.3 -Os ${ARGN}
 		        "-I${gpu_tiler_shader_dir}" -o "${shader_spv}" "${shader_source}"
 		COMMAND ${CMAKE_COMMAND} -DINPUT=${shader_spv} -DOUTPUT=${shader_header}
 		        -DSYMBOL=${shader_symbol} -P "${KYTY_SOURCE_DIR}/embed_spirv.cmake"
 		DEPENDS "${shader_source}"
 		VERBATIM
 	)
-	list(APPEND gpu_blit_shader_headers "${shader_header}")
-endforeach()
+	set(gpu_blit_shader_headers ${gpu_blit_shader_headers} "${shader_header}" PARENT_SCOPE)
+endfunction()
+add_gpu_blit_shader(gpu_blit_fs_triangle gpu_blit_fs_triangle.vert)
+add_gpu_blit_shader(gpu_blit_color_to_ms_depth gpu_blit_color_to_ms_depth.frag)
+add_gpu_blit_shader(gpu_video_out_overlay gpu_video_out_overlay.frag)
+add_gpu_blit_shader(gpu_imgui_hdr gpu_video_out_overlay.frag -DIMGUI_HDR=1)
 list(APPEND kyty_emulator_src ${gpu_blit_shader_headers})
 
 list(APPEND kyty_emulator_src
@@ -399,7 +403,7 @@ add_executable(scalar_provenance_tests EXCLUDE_FROM_ALL
 	"${KYTY_SOURCE_DIR}/graphics/shader/recompiler/ir/passes/ResourceMaterialization.cpp"
 	"${KYTY_SOURCE_DIR}/graphics/shader/recompiler/ir/passes/SrtWalker.cpp"
 )
-target_link_libraries(scalar_provenance_tests fmt::fmt)
+target_link_libraries(scalar_provenance_tests fmt::fmt xbyak::xbyak)
 target_include_directories(scalar_provenance_tests PRIVATE ${inc_headers})
 
 add_executable(page_manager_tests EXCLUDE_FROM_ALL
@@ -467,7 +471,7 @@ add_executable(resource_materialization_tests EXCLUDE_FROM_ALL
 	"${KYTY_SOURCE_DIR}/graphics/shader/recompiler/ir/passes/ResourceMaterialization.cpp"
 	"${KYTY_SOURCE_DIR}/graphics/shader/recompiler/ir/passes/SrtWalker.cpp"
 )
-target_link_libraries(resource_materialization_tests fmt::fmt)
+target_link_libraries(resource_materialization_tests fmt::fmt xbyak::xbyak)
 target_include_directories(resource_materialization_tests PRIVATE ${inc_headers})
 
 add_executable(resource_tracking_tests EXCLUDE_FROM_ALL
@@ -480,7 +484,7 @@ add_executable(resource_tracking_tests EXCLUDE_FROM_ALL
 	"${KYTY_SOURCE_DIR}/graphics/shader/recompiler/ir/passes/ResourceMaterialization.cpp"
 	"${KYTY_SOURCE_DIR}/graphics/shader/recompiler/ir/passes/ShaderInfoCollection.cpp"
 )
-target_link_libraries(resource_tracking_tests fmt::fmt)
+target_link_libraries(resource_tracking_tests fmt::fmt xbyak::xbyak)
 target_include_directories(resource_tracking_tests PRIVATE ${inc_headers})
 
 add_executable(audio_out2_port_tests EXCLUDE_FROM_ALL

README.md

@@ -248,6 +248,43 @@ sources, so it needs network access; a fully sandboxed `nix build` would require
 inputs. A Vulkan 1.3 driver must be available at runtime (on NixOS,
 `hardware.graphics.enable = true`).
 
+### Building on Fedora and Fedora Atomic (Bazzite, Bluefin, Silverblue)
+
+On regular Fedora, install dependencies on the host. On Fedora Atomic variants such as Bazzite,
+Bluefin and Silverblue, use [Distrobox](https://distrobox.it/) and run the install and build steps
+inside the container. Keep the checkout in your home directory so it is shared with the container.
+Add `--nvidia` to the create command for hosts using the proprietary NVIDIA driver:
+
+```bash
+distrobox create --name kyty-build --image registry.fedoraproject.org/fedora:latest
+distrobox enter kyty-build
+```
+
+Install the toolchain, Qt 6 and SDL3 dependencies:
+
+```bash
+sudo dnf install -y \
+  clang lld ninja-build cmake git glslang python3 pkgconf-pkg-config \
+  mesa-libGL-devel libX11-devel libXcursor-devel libXext-devel libXfixes-devel \
+  libXi-devel libXrandr-devel libXScrnSaver-devel libXtst-devel libxkbcommon-devel \
+  alsa-lib-devel pulseaudio-libs-devel systemd-devel dbus-devel \
+  wayland-devel wayland-protocols-devel qt6-qtbase-devel
+```
+
+From the repository root, follow the launcher configure, build and install steps in
+[Building on Linux](#building-on-linux), omitting `-DCMAKE_PREFIX_PATH="$Qt6_DIR"`; CMake finds
+Fedora's packaged Qt 6 automatically.
+
+Run `./_Build/linux/install/launcher` on regular Fedora or inside the container. To launch a
+container build from the host, leave the container with `exit`, then run from the repository root:
+
+```bash
+distrobox enter kyty-build -- ./_Build/linux/install/launcher
+```
+
+Qt libraries and plugins are bundled. You can also launch directly on a Fedora Atomic host
+if its other system libraries are compatible with the container build.
+
 ### Building on macOS
 
 macOS builds target x86-64 and run under Rosetta 2 on Apple Silicon, so the PS5's x86-64 game

src/common/emulatorConfig.cpp

@@ -72,6 +72,10 @@ bool FullscreenEnabled() {
 	return g_config->fullscreen_enabled;
 }
 
+bool HdrEnabled() {
+	return g_config->hdr_enabled;
+}
+
 bool HideCursorEnabled() {
 	return g_config->hide_cursor_enabled;
 }
@@ -96,6 +100,10 @@ bool VulkanValidationEnabled() {
 	return g_config->vulkan_validation_enabled;
 }
 
+bool IsNetworkConnected() {
+	return g_config->network_connected;
+}
+
 bool ShaderValidationEnabled() {
 	return g_config->shader_validation_enabled;
 }
@@ -182,4 +190,8 @@ const Keymap& GetKeymap() {
 	return g_config->keymap;
 }
 
+const std::vector<std::string>& GetGuestArguments() {
+	return g_config->guest_arguments;
+}
+
 } // namespace Config

src/common/emulatorConfig.h

@@ -53,10 +53,12 @@ struct ConfigOptions {
 	PresentMode            present_mode                = PresentMode::Mailbox;
 	int32_t                gpu_index                   = -1;
 	bool                   fullscreen_enabled          = false;
+	bool                   hdr_enabled                 = false;
 	bool                   hide_cursor_enabled         = false;
 	bool                   splash_enabled              = true;
 	bool                   vr_enabled                  = false;
 	bool                   amd_cpu_enabled             = false;
+	bool                   network_connected           = false;
 	uint32_t               vblank_frequency            = 60;
 	uint32_t               console_language            = DEFAULT_CONSOLE_LANGUAGE;
 	bool                   vulkan_validation_enabled   = false;
@@ -83,6 +85,8 @@ struct ConfigOptions {
 	bool red_zone_protection_enabled = false;
 #endif
 	Keymap keymap;
+	// Ordinary boots need a non-null argv[0]; --guest-exec supplies the caller's arguments.
+	std::vector<std::string> guest_arguments {""};
 };
 
 void Load(const ConfigOptions& cfg);
@@ -98,9 +102,11 @@ uint32_t GetControllerVibrationIntensity();
 PresentMode GetPresentMode();
 int32_t GetGpuIndex();
 bool     FullscreenEnabled();
+bool     HdrEnabled();
 bool     HideCursorEnabled();
 bool     VrEnabled();
 bool     AmdCpuEnabled();
+bool     IsNetworkConnected();
 uint32_t GetVblankFrequency();
 uint32_t GetConsoleLanguage();
 bool     VulkanValidationEnabled();
@@ -136,6 +142,7 @@ bool RedZoneProtectionEnabled();
 #endif
 
 const Keymap& GetKeymap();
+const std::vector<std::string>& GetGuestArguments();
 
 } // namespace Config
 

src/common/threads.cpp

@@ -8,10 +8,6 @@
 #include <condition_variable> // IWYU pragma: keep
 #include <mutex>
 
-#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS && KYTY_COMPILER == KYTY_COMPILER_CLANG
-#define KYTY_WIN_CS
-#endif
-
 // macOS has no clock_nanosleep.
 #if KYTY_PLATFORM != KYTY_PLATFORM_WINDOWS && !defined(__APPLE__)
 #define KYTY_POSIX_HIGH_RES_SLEEP
@@ -22,7 +18,7 @@
 #include <string>
 #include <thread>
 
-#ifdef KYTY_WIN_CS
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 #include <windows.h> // IWYU pragma: keep
 // IWYU pragma: no_include <winbase.h>
 constexpr DWORD KYTY_CS_SPIN_COUNT = 4000;
@@ -133,7 +129,7 @@ namespace Common {
 using thread_id_t = std::thread::id;
 
 struct MutexPrivate {
-#ifdef KYTY_WIN_CS
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 	MutexPrivate() { InitializeCriticalSectionAndSpinCount(&m_cs, KYTY_CS_SPIN_COUNT); }
 	~MutexPrivate() { DeleteCriticalSection(&m_cs); }
 	KYTY_CLASS_NO_COPY(MutexPrivate);
@@ -144,7 +140,7 @@ struct MutexPrivate {
 };
 
 struct CondVarPrivate {
-#ifdef KYTY_WIN_CS
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 	CondVarPrivate() {
 		static auto func = ResolveInitializeConditionVariable();
 		EXIT_NOT_IMPLEMENTED(func == nullptr);
@@ -216,7 +212,7 @@ void Thread::Detach() {
 }
 
 void Thread::SleepMicro(uint32_t micros) {
-#ifdef KYTY_WIN_CS
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 	SleepHighResolution100ns(static_cast<uint64_t>(micros) * 10);
 #elif defined(KYTY_POSIX_HIGH_RES_SLEEP)
 	SleepHighResolutionNanos(static_cast<uint64_t>(micros) * 1000);
@@ -226,7 +222,7 @@ void Thread::SleepMicro(uint32_t micros) {
 }
 
 void Thread::SleepNano(uint64_t nanos) {
-#ifdef KYTY_WIN_CS
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 	SleepHighResolution100ns((nanos + 99) / 100);
 #elif defined(KYTY_POSIX_HIGH_RES_SLEEP)
 	SleepHighResolutionNanos(nanos);
@@ -262,7 +258,7 @@ Mutex::~Mutex() {
 }
 
 void Mutex::Lock() {
-#ifdef KYTY_WIN_CS
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 	EnterCriticalSection(&m_mutex->m_cs);
 #else
 	m_mutex->m_mutex.lock();
@@ -270,7 +266,7 @@ void Mutex::Lock() {
 }
 
 void Mutex::Unlock() {
-#ifdef KYTY_WIN_CS
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 	LeaveCriticalSection(&m_mutex->m_cs);
 #else
 	m_mutex->m_mutex.unlock();
@@ -278,7 +274,7 @@ void Mutex::Unlock() {
 }
 
 bool Mutex::TryLock() {
-#ifdef KYTY_WIN_CS
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 	return (TryEnterCriticalSection(&m_mutex->m_cs) != 0);
 #else
 	return m_mutex->m_mutex.try_lock();
@@ -292,7 +288,7 @@ CondVar::~CondVar() {
 }
 
 void CondVar::Wait(Mutex* mutex) {
-#ifndef KYTY_WIN_CS
+#if KYTY_PLATFORM != KYTY_PLATFORM_WINDOWS
 	std::unique_lock<std::recursive_mutex> cpp_lock(mutex->m_mutex->m_mutex, std::adopt_lock_t());
 #endif
 	auto poll_callback = [&] {
@@ -300,7 +296,7 @@ void CondVar::Wait(Mutex* mutex) {
 		if (callback == nullptr) {
 			return;
 		}
-#if defined(KYTY_WIN_CS)
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 		LeaveCriticalSection(&mutex->m_mutex->m_cs);
 		callback();
 		EnterCriticalSection(&mutex->m_mutex->m_cs);
@@ -310,7 +306,7 @@ void CondVar::Wait(Mutex* mutex) {
 		cpp_lock.lock();
 #endif
 	};
-#ifdef KYTY_WIN_CS
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 	static auto func = ResolveSleepConditionVariableCS();
 	EXIT_NOT_IMPLEMENTED(func == nullptr);
 	if (g_cond_wait_poll_callback == nullptr) {
@@ -340,10 +336,10 @@ void CondVar::SetWaitPollCallback(wait_poll_func_t callback) {
 
 bool CondVar::WaitFor(Mutex* mutex, uint32_t micros) {
 	bool ok = false;
-#ifndef KYTY_WIN_CS
+#if KYTY_PLATFORM != KYTY_PLATFORM_WINDOWS
 	std::unique_lock<std::recursive_mutex> cpp_lock(mutex->m_mutex->m_mutex, std::adopt_lock_t());
 #endif
-#ifdef KYTY_WIN_CS
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 	static auto func = ResolveSleepConditionVariableCS();
 	EXIT_NOT_IMPLEMENTED(func == nullptr);
 	ok = !(func(&m_cond_var->m_cv, &mutex->m_mutex->m_cs, (micros < 1000 ? 1 : micros / 1000)) ==
@@ -358,7 +354,7 @@ bool CondVar::WaitFor(Mutex* mutex, uint32_t micros) {
 }
 
 void CondVar::Signal() {
-#ifdef KYTY_WIN_CS
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 	static auto func = ResolveWakeConditionVariable();
 	EXIT_NOT_IMPLEMENTED(func == nullptr);
 	func(&m_cond_var->m_cv);
@@ -368,7 +364,7 @@ void CondVar::Signal() {
 }
 
 void CondVar::SignalAll() {
-#ifdef KYTY_WIN_CS
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 	static auto func = ResolveWakeAllConditionVariable();
 	EXIT_NOT_IMPLEMENTED(func == nullptr);
 	func(&m_cond_var->m_cv);

src/emulator.h

@@ -16,6 +16,7 @@ struct RunOptions {
 };
 
 void Run(const RunOptions& options);
+int LoadExec(const char* path, const char* const argv[]);
 
 } // namespace Emulator
 

src/graphics/guest_gpu/command_processor/pm4Handlers.cpp

@@ -1934,13 +1934,7 @@ KYTY_CP_OP_PARSER(CpOpIndirectBuffer) {
 
 	auto* indirect_buffer =
 	    reinterpret_cast<const uint32_t*>(buffer[0] | (static_cast<uint64_t>(buffer[1]) << 32u));
-	uint32_t                     indirect_num_dw = control & 0xfffffu;
-	static std::atomic<uint32_t> indirect_log_count {0};
-	if (indirect_log_count.fetch_add(1) < 128) {
-		LOGF("\t indirect buffer: addr=0x%016" PRIx64 ", num_dw=%" PRIu32 ", control=0x%08" PRIx32
-		     "\n",
-		     reinterpret_cast<uint64_t>(indirect_buffer), indirect_num_dw, control);
-	}
+	const uint32_t indirect_num_dw = control & 0xfffffu;
 
 	if (indirect_num_dw == 0) {
 		return 3;

src/graphics/guest_gpu/graphicsRun.cpp

@@ -576,12 +576,7 @@ bool GuestGpu::Process(Submission& submission) {
 				}
 			}
 			if (progressed) {
-				if (complete) {
-					m_renderer.RunGarbageCollector();
-				}
 				cp.BufferFlush();
-			} else if (complete) {
-				m_renderer.RunGarbageCollector();
 			}
 			break;
 		}
@@ -602,20 +597,16 @@ bool GuestGpu::Process(Submission& submission) {
 			complete = cp.Process(submission.command_execution, submission.commands) ==
 			           Pm4ProcessResult::Complete;
 			if (submission.command_execution.MadeProgress()) {
-				if (complete) {
-					m_renderer.RunGarbageCollector();
-				}
 				cp.BufferFlush();
-			} else if (complete) {
-				m_renderer.RunGarbageCollector();
 			}
 			break;
 		}
 		case SubmissionType::FlipPreparation:
-			m_renderer.RunGarbageCollector();
 			cp.PrepareCpuFlip(submission.flip_request_id);
 			break;
 		case SubmissionType::SuspendPoint:
+			// Age caches at the frame boundary, not once per command list.
+			m_renderer.RunGarbageCollector();
 			cp.EmitGlobalBarrier();
 			m_renderer.GetCommandScheduler().DeferPriorityOperation(
 			    [ready = m_suspend_point_ready] { ready->release(); });
@@ -1072,8 +1063,7 @@ void CommandProcessor::WriteAtEndOfPipe(uint32_t cache_policy, uint32_t event_wr
 		case 0x03: with_interrupt = false; break;
 		case 0x01:
 			if (!IsAsyncComputeQueue()) {
-				Sync::TriggerEopEventAtEndOfPipe(command, m_interrupt_event_id,
-				                                 interrupt_context_id);
+				TriggerEopEventAtEndOfPipe(interrupt_context_id);
 				return;
 			}
 			with_interrupt = true;
@@ -1082,24 +1072,15 @@ void CommandProcessor::WriteAtEndOfPipe(uint32_t cache_policy, uint32_t event_wr
 		default: EXIT("unknown interrupt_selector\n");
 	}
 
-	auto write32 = [&](bool with_writeback) {
+	auto write32 = [&] {
 		auto* dst  = static_cast<uint32_t*>(dst_gpu_addr);
 		auto  data = static_cast<uint32_t>(value);
+		EXIT_IF(dst == nullptr);
+		(void)command.Handle();
 		std::memcpy(dst, &data, sizeof(data));
 
 		if (with_interrupt) {
-			if (with_writeback) {
-				Sync::WriteAtEndOfPipeWithInterruptWriteBack32(m_submit_id, command, dst,
-				                                               data, m_interrupt_event_id,
-				                                               interrupt_context_id);
-			} else {
-				Sync::WriteAtEndOfPipeWithInterrupt32(m_submit_id, command, dst, data,
-				                                      m_interrupt_event_id, interrupt_context_id);
-			}
-		} else if (with_writeback) {
-			Sync::WriteAtEndOfPipeWithWriteBack32(m_submit_id, command, dst, data);
-		} else {
-			Sync::WriteAtEndOfPipe32(m_submit_id, command, dst, data);
+			TriggerEopEventAtEndOfPipe(interrupt_context_id);
 		}
 	};
 
@@ -1111,15 +1092,13 @@ void CommandProcessor::WriteAtEndOfPipe(uint32_t cache_policy, uint32_t event_wr
 					SynchronizeGpu();
 					Sync::ReadGds(*m_renderer.GetBufferCache().GetGdsBuffer(), dst, value & 0xffffu,
 					              value >> 16u);
-					Sync::WriteAtEndOfPipeGds32(m_submit_id, command, dst, value & 0xffffu,
-					                            value >> 16u);
 					if (with_interrupt) {
 						m_renderer.TriggerInterrupt(m_interrupt_event_id, interrupt_context_id);
 					}
 					return;
 				}
 			} else if (eop_event_type == 0x04 && cache_action == 0x00 && event_index == 0x05) {
-				write32(false);
+				write32();
 				return;
 			}
 			break;
@@ -1128,8 +1107,8 @@ void CommandProcessor::WriteAtEndOfPipe(uint32_t cache_policy, uint32_t event_wr
 			if constexpr (sizeof(T) == sizeof(uint32_t)) {
 				if (event_write_source == 0x02 && eop_event_type == 0x2f && event_index == 0x06) {
 					switch (cache_action) {
-						case 0x00: write32(false); return;
-						case 0x38: write32(true); return;
+						case 0x00:
+						case 0x38: write32(); return;
 						default: break;
 					}
 				}
@@ -1137,25 +1116,14 @@ void CommandProcessor::WriteAtEndOfPipe(uint32_t cache_policy, uint32_t event_wr
 				if (event_write_source == 0x04) {
 					value = Sync::ReadReferenceClock();
 				}
-				auto write64 = [&](bool with_writeback) {
+				auto write64 = [&] {
 					auto* dst = static_cast<uint64_t*>(dst_gpu_addr);
+					EXIT_IF(dst == nullptr);
+					(void)command.Handle();
 					std::memcpy(dst, &value, sizeof(value));
 
 					if (with_interrupt) {
-						if (with_writeback) {
-							Sync::WriteAtEndOfPipeWithInterruptWriteBack64(
-							    m_submit_id, command, dst, value, m_interrupt_event_id,
-							    interrupt_context_id);
-						} else {
-							Sync::WriteAtEndOfPipeWithInterrupt64(m_submit_id, command, dst,
-							                                      value, m_interrupt_event_id,
-							                                      interrupt_context_id);
-						}
-					} else if (with_writeback) {
-						Sync::WriteAtEndOfPipeWithWriteBack64(m_submit_id, command, dst,
-						                                      value);
-					} else {
-						Sync::WriteAtEndOfPipe64(m_submit_id, command, dst, value);
+						TriggerEopEventAtEndOfPipe(interrupt_context_id);
 					}
 				};
 
@@ -1164,7 +1132,7 @@ void CommandProcessor::WriteAtEndOfPipe(uint32_t cache_policy, uint32_t event_wr
 						switch (eop_event_type) {
 							case 0x04:
 								if (event_index == 0x05) {
-									write64(false);
+									write64();
 									return;
 								}
 								break;
@@ -1172,7 +1140,7 @@ void CommandProcessor::WriteAtEndOfPipe(uint32_t cache_policy, uint32_t event_wr
 							case 0x28:
 							case 0x2f:
 								if (event_index == 0x00) {
-									write64(false);
+									write64();
 									return;
 								}
 								break;
@@ -1180,7 +1148,7 @@ void CommandProcessor::WriteAtEndOfPipe(uint32_t cache_policy, uint32_t event_wr
 							case 0x2d:
 							case 0x30:
 								if (event_index == 0x00 && !with_interrupt) {
-									write64(false);
+									write64();
 									return;
 								}
 								break;
@@ -1195,20 +1163,20 @@ void CommandProcessor::WriteAtEndOfPipe(uint32_t cache_policy, uint32_t event_wr
 								if (((eop_event_type == 0x04 || eop_event_type == 0x28) &&
 								     event_index == 0x05) ||
 								    (event_index == 0x00)) {
-									write64(true);
+									write64();
 									return;
 								}
 								break;
 							case 0x2b:
 							case 0x2d:
 								if (event_index == 0x00 && !with_interrupt) {
-									write64(true);
+									write64();
 									return;
 								}
 								break;
 							case 0x2f:
 								if (event_index == 0x06 && !with_interrupt) {
-									write64(true);
+									write64();
 									return;
 								}
 								break;
@@ -1217,7 +1185,7 @@ void CommandProcessor::WriteAtEndOfPipe(uint32_t cache_policy, uint32_t event_wr
 						break;
 					case 0x3b:
 						if (eop_event_type == 0x04 && event_index == 0x05 && with_interrupt) {
-							write64(true);
+							write64();
 							return;
 						}
 						break;
@@ -1270,7 +1238,11 @@ void CommandProcessor::EmitGlobalBarrier() {
 }
 
 void CommandProcessor::TriggerEopEventAtEndOfPipe(uint32_t interrupt_context_id) {
-	Sync::TriggerEopEventAtEndOfPipe(CurrentBuffer(), m_interrupt_event_id, interrupt_context_id);
+	(void)CurrentBuffer().Handle();
+	GetScheduler().DeferPriorityOperation(
+	    [&renderer = m_renderer, event_id = m_interrupt_event_id, interrupt_context_id] {
+		    renderer.TriggerInterrupt(event_id, interrupt_context_id);
+	    });
 }
 
 void CommandProcessor::TriggerEvent(uint32_t event_type, uint32_t event_index,
@@ -1360,8 +1332,9 @@ void CommandProcessor::Flip() {
 	auto& command = CurrentBuffer();
 	auto request = Sync::PrepareVideoOutFlip(command, m_flip.handle, m_flip.index, m_flip.flip_mode,
 	                                         m_flip.flip_arg);
-	Sync::WriteAtEndOfPipeOnlyFlip(m_submit_id, command, m_flip.handle, m_flip.index,
-	                               m_flip.flip_mode, m_flip.flip_arg, request);
+	(void)command.Handle();
+	GetScheduler().DeferPriorityOperation(
+	    [&renderer = m_renderer, request] { renderer.GetVideoOut().CompleteFlip(request); });
 	GetScheduler().Flush();
 }
 
@@ -1375,12 +1348,13 @@ void CommandProcessor::Flip(void* dst_gpu_addr, uint32_t value) {
 		     reinterpret_cast<uint64_t>(dst_gpu_addr), value);
 	}
 
+	EXIT_IF(dst_gpu_addr == nullptr);
 	std::memcpy(dst_gpu_addr, &value, sizeof(value));
 	auto request = Sync::PrepareVideoOutFlip(command, m_flip.handle, m_flip.index, m_flip.flip_mode,
 	                                         m_flip.flip_arg);
-	Sync::WriteAtEndOfPipeWithFlip32(m_submit_id, command, static_cast<uint32_t*>(dst_gpu_addr),
-	                                 value, m_flip.handle, m_flip.index, m_flip.flip_mode,
-	                                 m_flip.flip_arg, request);
+	(void)command.Handle();
+	GetScheduler().DeferPriorityOperation(
+	    [&renderer = m_renderer, request] { renderer.GetVideoOut().CompleteFlip(request); });
 	GetScheduler().Flush();
 }
 
@@ -1400,12 +1374,16 @@ void CommandProcessor::FlipWithInterrupt(uint32_t eop_event_type, uint32_t cache
 	if (eop_event_type != 0x00000004 || cache_action != 0x00000038) {
 		EXIT("unknown event type\n");
 	}
+	EXIT_IF(dst_gpu_addr == nullptr);
 	std::memcpy(dst_gpu_addr, &value, sizeof(value));
 	auto request = Sync::PrepareVideoOutFlip(command, m_flip.handle, m_flip.index, m_flip.flip_mode,
 	                                         m_flip.flip_arg);
-	Sync::WriteAtEndOfPipeWithInterruptWriteBackFlip32(
-	    m_submit_id, command, static_cast<uint32_t*>(dst_gpu_addr), value, m_flip.handle,
-	    m_flip.index, m_flip.flip_mode, m_flip.flip_arg, request, m_interrupt_event_id);
+	(void)command.Handle();
+	GetScheduler().DeferPriorityOperation(
+	    [&renderer = m_renderer, event_id = m_interrupt_event_id, request] {
+		    renderer.GetVideoOut().CompleteFlip(request);
+		    renderer.TriggerInterrupt(event_id, 0);
+	    });
 	GetScheduler().Flush();
 }
 

src/graphics/host_gpu/graphicContext.h

@@ -33,6 +33,7 @@ struct GraphicContext {
 	// bool fp64_denorm_preserve = false; // Temporarily disabled.
 	bool                               attachment_feedback_loop_enabled      = false;
 	bool                               provoking_vertex_last_enabled         = false;
+	bool                               depth_color_copy_enabled               = false;
 	bool                               supports_block_texel_view              = false;
 	bool                                      mesh_shader_enabled                   = false;
 	vk::PhysicalDeviceMeshShaderPropertiesEXT mesh_shader_properties                = {};

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

@@ -36,7 +36,9 @@ constexpr uint64_t NumFramesBeforeRemoval = 32;
 [[nodiscard]] bool DecodeColorClear(const TextureCache::ImageDesc& desc, uint8_t code,
                                   vk::ClearColorValue& clear) {
 	const auto& metadata = desc.info.metadata;
-	const auto  format   = desc.view_info.format;
+	const auto  format   = desc.info.IsDepth()
+	                           ? TextureGetSurfaceFormatInfo(desc.info.guest_format).vk_format
+	                           : desc.view_info.format;
 	const bool  cmask    = metadata.kind == ImageMetadataKind::Cmask;
 	if (cmask ? code == 0 : code == 0x20) {
 		// Register clears belong to the color buffer; the texture pipe cannot decode them.
@@ -631,6 +633,9 @@ void TextureCache::CopyImage(ImageId destination_id, ImageId source_id) {
 void TextureCache::CopyImageMip(ImageId destination_id, ImageId source_id, uint32_t mip,
                                 uint32_t layer) {
 	RefreshCopySource(source_id);
+	if (m_slot_images[destination_id].IsBufferModified()) {
+		RefreshImage(destination_id);
+	}
 	auto& destination = m_slot_images[destination_id];
 	auto& source      = m_slot_images[source_id];
 	TrackImage(destination_id);
@@ -751,11 +756,63 @@ ImageId TextureCache::ResolveDepthOverlap(const ImageInfo& requested, BindingTyp
 	return replacement_id;
 }
 
+namespace BcAlias {
+
+bool Matches(const ImageInfo& requested, const ImageInfo& cached) {
+	return requested.guest_format != cached.guest_format &&
+	       requested.IsBlock() != cached.IsBlock() &&
+	       requested.tile_mode == cached.tile_mode &&
+	       requested.bytes_per_block == cached.bytes_per_block &&
+	       requested.samples == 1 && cached.samples == 1 &&
+	       (requested.data != cached.data || requested.resources != cached.resources ||
+	        requested.BlockExtent() != cached.BlockExtent()) &&
+	       !TextureCache::SameBacking(cached, requested, false);
+}
+
+bool Publish(TextureCache& cache, Image& image) {
+	if (!cache.CanDownload(image)) {
+		return false;
+	}
+	// ObtainBuffer preserves padding and records ownership before the GPU-only transfer.
+	auto [buffer, offset] =
+	    cache.m_buffer_cache.ObtainBuffer(image.info.data.address, image.info.data.size, true);
+	cache.DownloadImage(image, *buffer, offset, image.info.data.size, image.info.resources.levels);
+	image.ClearGpuModified();
+	image.MarkBufferModified();
+	return true;
+}
+
+void Resolve(TextureCache& cache, ImageId cached_id, ImageId merged_id) {
+	auto& cached = cache.m_slot_images[cached_id];
+	auto* merged = merged_id ? &cache.m_slot_images[merged_id] : nullptr;
+	// PS5 raw block images can span several packed BC mips. Transfer their complete
+	// guest bytes, including tail padding, before retiring the incompatible backing.
+	if (cached.SafeToDownload()) {
+		if (merged && merged->SafeToDownload() && !Publish(cache, *merged)) {
+			return;
+		}
+		if (!Publish(cache, cached)) {
+			return;
+		}
+		if (merged) {
+			merged->MarkBufferModified();
+		}
+	}
+	cached.binding.needs_rebind |= cached.binding.is_bound || cached.binding.is_target;
+	cache.FreeImage(cached_id);
+}
+
+} // namespace BcAlias
+
 TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& requested,
                                                          BindingType binding, ImageId cached_id,
                                                          ImageId merged_id) {
 	auto owner = m_slot_images.try_get(cached_id);
-	if (owner == nullptr) {
+	if (owner == nullptr || !owner->registered) {
+		return {merged_id};
+	}
+	if (BcAlias::Matches(requested, owner->info) && !owner->IsCpuDirty()) {
+		BcAlias::Resolve(*this, cached_id, merged_id);
 		return {merged_id};
 	}
 	auto&      cached       = *owner;
@@ -1059,8 +1116,15 @@ void TextureCache::MaterializeColorClear(ImageId id, const ImageDesc& desc,
 		}
 		{
 			std::scoped_lock lock {m_lock};
-			ClearImage(m_scheduler.Current(), id, view.format,
-			           {vk::ImageAspectFlagBits::eColor, view.base_level, view.level_count,
+			const auto& image = m_slot_images[id];
+			const bool depth = image.info.IsDepth();
+			if (depth) {
+				const auto value = clear.color.float32[0];
+				clear.depthStencil = {value, 0u};
+			}
+			ClearImage(m_scheduler.Current(), id, depth ? image.backing.format : view.format,
+			           {depth ? vk::ImageAspectFlagBits::eDepth : vk::ImageAspectFlagBits::eColor,
+			            view.base_level, view.level_count,
 			            image_first + slice, 1}, clear);
 		}
 		// Publish the conversion's expanded keys without treating them as guest writes
@@ -1166,6 +1230,7 @@ ImageId TextureCache::FindImage(ImageDesc& desc, bool exact_format) {
 		std::scoped_lock lock {m_lock};
 		const auto       candidates =
 		    FindImagesInRegion(desc.info.data.address, desc.info.data.size, false);
+		bool block_alias = false;
 
 		for (const auto id: candidates) {
 			const auto& image = m_slot_images[id];
@@ -1184,6 +1249,9 @@ ImageId TextureCache::FindImage(ImageDesc& desc, bool exact_format) {
 			}
 			if (SameBacking(image.info, desc.info, exact_format)) {
 				result = id;
+			} else if (image.SafeToDownload() &&
+			           BcAlias::Matches(desc.info, image.info)) {
+				block_alias = true;
 			}
 		}
 
@@ -1201,6 +1269,15 @@ ImageId TextureCache::FindImage(ImageDesc& desc, bool exact_format) {
 					view_layer = overlap.layer;
 				}
 			}
+		} else if (block_alias) {
+			for (const auto candidate: candidates) {
+				const auto* owner = m_slot_images.try_get(candidate);
+				if (candidate != result && owner != nullptr && owner->registered &&
+				    !owner->IsCpuDirty() &&
+				    BcAlias::Matches(m_slot_images[result].info, owner->info)) {
+					BcAlias::Resolve(*this, candidate, result);
+				}
+			}
 		}
 
 		if (result && !m_slot_images[result].depth_id) {
@@ -1923,6 +2000,9 @@ void TextureCache::RunGarbageCollector() {
 }
 
 void TextureCache::ProcessDownloadImages() {
+	if (!m_readback_linear_images) {
+		return;
+	}
 	std::scoped_lock lock {m_lock};
 	for (const auto id: m_download_images) {
 		const auto owner = m_slot_images.try_get(id);

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

@@ -26,8 +26,15 @@ class BufferCache;
 class CommandBuffer;
 class CommandScheduler;
 class RenderExecutor;
+class TextureCache;
 struct TextureCacheTestAccess;
 
+namespace BcAlias {
+[[nodiscard]] bool Matches(const ImageInfo& requested, const ImageInfo& cached);
+[[nodiscard]] bool Publish(TextureCache& cache, Image& image);
+void Resolve(TextureCache& cache, ImageId cached, ImageId merged);
+} // namespace BcAlias
+
 class TextureCache {
 public:
 	enum class BindingType : uint8_t { Texture, Storage, RenderTarget, DepthTarget, VideoOut };
@@ -181,6 +188,9 @@ private:
 	friend struct TextureCacheTestAccess;
 	friend class BufferCache;
 	friend class RenderExecutor;
+	friend bool BcAlias::Matches(const ImageInfo&, const ImageInfo&);
+	friend bool BcAlias::Publish(TextureCache&, Image&);
+	friend void BcAlias::Resolve(TextureCache&, ImageId, ImageId);
 };
 
 } // namespace Libs::Graphics

src/graphics/host_gpu/renderer/debug.cpp

@@ -394,26 +394,20 @@ static void RcPrint(const char* func, const HW::RenderControl& c) {
 
 	LOGF("\t depth_clear_enable       = %s\n"
 	     "\t stencil_clear_enable     = %s\n"
+	     "\t copy_depth_to_color      = %s\n"
+	     "\t copy_stencil_to_color    = %s\n"
 	     "\t resummarize_enable       = %s\n"
 	     "\t stencil_compress_disable = %s\n"
 	     "\t depth_compress_disable   = %s\n"
 	     "\t copy_centroid            = %s\n"
 	     "\t copy_sample              = %" PRIu8 "\n",
 	     c.depth_clear_enable ? "true" : "false", c.stencil_clear_enable ? "true" : "false",
+	     c.copy_depth_to_color ? "true" : "false", c.copy_stencil_to_color ? "true" : "false",
 	     c.resummarize_enable ? "true" : "false", c.stencil_compress_disable ? "true" : "false",
 	     c.depth_compress_disable ? "true" : "false", c.copy_centroid ? "true" : "false",
 	     c.copy_sample);
 }
 
-static void RcCheck(const HW::RenderControl& c) {
-	// EXIT_NOT_IMPLEMENTED(c.depth_clear_enable != false);
-	// EXIT_NOT_IMPLEMENTED(c.stencil_clear_enable != false);
-	// EXIT_NOT_IMPLEMENTED(c.stencil_compress_disable != false);
-	// EXIT_NOT_IMPLEMENTED(c.depth_compress_disable != false);
-	EXIT_NOT_IMPLEMENTED(c.copy_centroid != false);
-	EXIT_NOT_IMPLEMENTED(c.copy_sample != 0);
-}
-
 static void McPrint(const char* func, const HW::ModeControl& c) {
 	LOGF("%s\n", func);
 
@@ -869,9 +863,6 @@ void hw_check(const CommandBuffer& buffer) {
 	VpCheck(vp, smc);
 	log_phase("clip");
 	ClipCheck(c);
-	log_phase("rc");
-	RcCheck(rc);
-	log_phase("depth");
 	log_phase("mode");
 	McCheck(mc);
 	log_phase("blend");

src/graphics/host_gpu/renderer/depthRenderTarget.cpp

@@ -298,13 +298,20 @@ void RenderExecutor::ResolveRenderDepthTarget(CommandBuffer& buffer, RenderDepth
 		}
 		return;
 	}
-	if (rc.copy_depth_to_color || rc.copy_stencil_to_color || rc.copy_centroid ||
-	    rc.copy_sample != 0 || dc.zfunc > static_cast<uint8_t>(vk::CompareOp::eAlways) ||
+	if (rc.copy_stencil_to_color ||
+	    dc.zfunc > static_cast<uint8_t>(vk::CompareOp::eAlways) ||
 	    (!z.depth_view.depth_write_disable && z.z_write_base_addr != z.z_read_base_addr) ||
 	    (has_stencil && !z.depth_view.stencil_write_disable &&
 	     z.stencil_write_base_addr != z.stencil_read_base_addr)) {
 		DepthFatal("unsupported depth register state");
 	}
+	if (rc.copy_depth_to_color &&
+	    (z.z_info.num_samples != 0 || rc.copy_sample != 0 ||
+	     z.z_info.format != Prospero::DepthFormat::kZ32F ||
+	     dc.z_enable || dc.z_write_enable || dc.depth_bounds_enable || dc.stencil_enable ||
+	     rc.depth_clear_enable || rc.stencil_clear_enable)) {
+		DepthFatal("unsupported depth-to-color copy state");
+	}
 	r.desc = MakeDepthTargetDesc(buffer, z);
 	r.depth_clear_enable      = rc.depth_clear_enable;
 	r.depth_load_clear_enable = r.depth_clear_enable;

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

@@ -362,8 +362,17 @@ struct VideoOutFormatPolicy {
 };
 
 inline constexpr uint64_t VIDEO_OUT_PIXEL_FORMAT_R10_G10_B10_A2_BT2100_PQ = 0x8100070422000000ull;
+inline constexpr uint64_t VIDEO_OUT_PIXEL_FORMAT_B10_G10_R10_A2_BT2100_PQ = 0x8100070400000000ull;
 
-inline constexpr std::array<VideoOutFormatPolicy, 7> VIDEO_OUT_FORMAT_POLICIES {{
+[[nodiscard]] constexpr bool IsHdrVideoOutPixelFormat(uint64_t format) {
+	switch (format) {
+		case VIDEO_OUT_PIXEL_FORMAT_R10_G10_B10_A2_BT2100_PQ:
+		case VIDEO_OUT_PIXEL_FORMAT_B10_G10_R10_A2_BT2100_PQ: return true;
+		default: return false;
+	}
+}
+
+inline constexpr std::array<VideoOutFormatPolicy, 8> VIDEO_OUT_FORMAT_POLICIES {{
     {0x8000000022000000ull,
      {vk::Format::eR8G8B8A8Srgb, Prospero::BufferFormat::k8_8_8_8Srgb, 4, false}},
     {0x8000000000000000ull,
@@ -374,6 +383,8 @@ inline constexpr std::array<VideoOutFormatPolicy, 7> VIDEO_OUT_FORMAT_POLICIES {
      {vk::Format::eA2R10G10B10UnormPack32, Prospero::BufferFormat::k10_10_10_2UNorm, 4, false}},
     {VIDEO_OUT_PIXEL_FORMAT_R10_G10_B10_A2_BT2100_PQ,
      {vk::Format::eA2B10G10R10UnormPack32, Prospero::BufferFormat::k10_10_10_2UNorm, 4, false}},
+    {VIDEO_OUT_PIXEL_FORMAT_B10_G10_R10_A2_BT2100_PQ,
+     {vk::Format::eA2R10G10B10UnormPack32, Prospero::BufferFormat::k10_10_10_2UNorm, 4, false}},
     {0xc001000622000000ull,
      {vk::Format::eR16G16B16A16Sfloat, Prospero::BufferFormat::k16_16_16_16Float, 8, false}},
     {0xc001000600000000ull,

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

@@ -179,7 +179,9 @@ static void ValidateSampledDepthBinding(const ShaderRecompiler::IR::ImageResourc
                                         const ShaderTextureResource& descriptor, const Image& image,
                                         vk::Format view_format, uint64_t size) {
 	const bool resource_ok = IsSupportedSampledDepthResource(resource);
-	const bool encoding_ok = IsSupportedDepthTextureEncoding(descriptor, resource.r128);
+	// Comparison sampling changes the host format, not the guest metadata encoding.
+	const bool encoding_ok = descriptor.TileMode() != Prospero::TileMode::kDepth ||
+	                         IsSupportedDepthTextureEncoding(descriptor, resource.r128);
 	const bool view_ok =
 	    IsSupportedSampledDepthView(image.info.pixel_format, view_format, descriptor.DstSelXYZW());
 	if (resource_ok && encoding_ok && view_ok) {
@@ -622,8 +624,7 @@ TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageR
 	if (storage) {
 		ValidateStorageTexture(resource, descriptor, desc.info.data.size);
 	}
-	if (!resource.r128 && descriptor.MetaCompress() && tile != Prospero::TileMode::kDepth &&
-	    !desc.info.IsDepth()) {
+	if (!resource.r128 && descriptor.MetaCompress() && tile != Prospero::TileMode::kDepth) {
 		TileSizeAlign metadata_size {};
 		(void)TileGetDccSize(width, height, volume ? depth : image_layers,
 		                     desc.info.bytes_per_block, physical_levels, tile, metadata_size,

src/graphics/host_gpu/renderer/render.h

@@ -39,13 +39,7 @@ enum class CommandBufferDebugOp : uint32_t {
 	DispatchDirect,
 	DrawIndex,
 	DrawIndexAuto,
-	EopWrite,
-	EopInterrupt,
-	EopWriteBack,
-	EopFlip,
-	EopWriteBackFlip,
-	EopOnlyFlip,
-	DispatchIndirect,
+	DispatchIndirect = 9,
 	Unknown,
 };
 
@@ -186,6 +180,8 @@ private:
 	                              bool ignore_target_mask = false, bool exact_format = false);
 	void ResolveRenderDepthTarget(CommandBuffer& buffer, RenderDepthInfo& target);
 	[[nodiscard]] bool DepthStencilCopy(CommandBuffer& buffer);
+	[[nodiscard]] bool CopyDepthToColor(CommandBuffer& buffer, const DrawCallInfo& draw,
+	                                   const DrawEmitInfo& emit, DrawRenderState& state);
 	[[nodiscard]] bool PrepareDrawRenderState(CommandBuffer& buffer,
 	                                          const DrawCallInfo& draw,
 	                                          uint32_t            render_target_slice_offset,

src/graphics/host_gpu/renderer/renderDraw.cpp

@@ -1029,12 +1029,71 @@ static void EmitDrawPrimitives(const HW::UserConfig& ucfg, vk::CommandBuffer vk_
 	}
 }
 
+bool RenderExecutor::CopyDepthToColor(CommandBuffer& buffer, const DrawCallInfo& draw,
+                                       const DrawEmitInfo& emit, DrawRenderState& state) {
+	if (!buffer.GetRegisters().GetRenderControl().copy_depth_to_color) {
+		return false;
+	}
+	auto& ucfg = buffer.GetUserConfig();
+	const auto& hw = buffer.GetRegisters();
+	auto& depth = state.depth_info;
+	auto& color = state.color_info[0];
+	EXIT_NOT_IMPLEMENTED(!m_context.GetGraphics().depth_color_copy_enabled ||
+	                     !Prospero::IsRectList(ucfg.GetPrimType()) ||
+	                     draw.IsIndexed() || draw.index_count != 3 || draw.instance_count != 1 ||
+	                     emit.first_vertex != 0 || emit.first_instance != 0 ||
+	                     state.color_count != 1 || color.target_slot != 0 || !depth.image_id ||
+	                     color.desc.info.samples != 1 ||
+	                     depth.desc.info.extent.width != color.Extent().width ||
+	                     depth.desc.info.extent.height != color.Extent().height ||
+	                     depth.desc.view_info.layer_count != 1 ||
+	                     color.desc.view_info.layer_count != 1 ||
+	                     hw.GetBlendControl(0).enable || state.ps_input_info.ps_pixel_kill_enable ||
+	                     state.ps_input_info.ps_depth_export_enable ||
+	                     state.ps_input_info.ps_sample_mask_export_enable);
+	vk::ImageAspectFlags feedback;
+	const auto rendering = AcquireRenderTargets(buffer, &color, 1, depth, feedback);
+	if (depth.depth_load_clear_enable) {
+		m_context.GetCommandScheduler().BeginRendering(rendering);
+	}
+	auto& cache = m_context.GetTextureCache();
+	auto& source = cache.GetImage(depth.image_id);
+	auto& destination = cache.GetImage(color.image_id);
+	EXIT_NOT_IMPLEMENTED((source.backing.format != vk::Format::eD32Sfloat &&
+	                      source.backing.format != vk::Format::eD32SfloatS8Uint) ||
+	                     destination.backing.format != vk::Format::eR32Sfloat);
+	const auto rect = calc_final_scissor(hw.GetScreenViewport(), hw.GetScanModeControl(),
+	                                    {rendering.width, rendering.height}, 0);
+	if (rect.right > rect.left && rect.bottom > rect.top) {
+		const vk::ImageCopy region {
+		    {vk::ImageAspectFlagBits::eDepth, depth.desc.view_info.base_level,
+		     depth.desc.view_info.base_layer, 1}, {rect.left, rect.top, 0},
+		    {vk::ImageAspectFlagBits::eColor, color.desc.view_info.base_level,
+		     color.desc.view_info.base_layer, 1}, {rect.left, rect.top, 0},
+		    {static_cast<uint32_t>(rect.right - rect.left),
+		     static_cast<uint32_t>(rect.bottom - rect.top), 1}};
+		m_context.GetCommandScheduler().EndRendering();
+		const auto command = buffer.Handle();
+		source.Transit(vk::ImageLayout::eTransferSrcOptimal, vk::AccessFlagBits2::eTransferRead,
+		               {}, command);
+		destination.Transit(vk::ImageLayout::eTransferDstOptimal,
+		                    vk::AccessFlagBits2::eTransferWrite, {}, command);
+		command.copyImage(source.backing.image, vk::ImageLayout::eTransferSrcOptimal,
+		                  destination.backing.image, vk::ImageLayout::eTransferDstOptimal,
+		                  1, &region);
+	}
+	return true;
+}
+
 void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buffer,
                                          const DrawCallInfo& draw, DrawRenderState& state,
                                          vk::PrimitiveTopology topology, const DrawEmitInfo& emit,
                                          const DrawIndexBufferSource& index_source,
 	                                     bool primitive_restart_enable) {
 	auto& ucfg = buffer.GetUserConfig();
+	if (CopyDepthToColor(buffer, draw, emit, state)) {
+		return;
+	}
 	const auto vertex_stages =
 	    std::span {state.vertex_info.data(), state.programs.VertexStageCount()};
 	const bool mesh_active = state.vertex_info[0].stage.program->stage == ShaderType::Mesh;

src/graphics/host_gpu/renderer/sync.cpp

@@ -1,10 +1,6 @@
 #include "graphics/host_gpu/renderer/sync.h"
 
 #include "common/assert.h"
-#include "common/common.h"
-#include "common/logging/log.h"
-#include "common/threads.h"
-#include "graphics/host_gpu/graphicContext.h"
 #include "graphics/host_gpu/renderer/cache/bufferCache.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
@@ -53,109 +49,6 @@ uint64_t ReadReferenceClock() {
 	return value;
 }
 
-enum class EndOfPipeWriteSize : uint32_t { Dword = 4, Qword = 8 };
-enum class EndOfPipeWriteAction { Write, WriteBack, Interrupt, InterruptWriteBack };
-
-static CommandBufferDebugOp DebugOperation(EndOfPipeWriteAction action) {
-	switch (action) {
-		case EndOfPipeWriteAction::Write: return CommandBufferDebugOp::EopWrite;
-		case EndOfPipeWriteAction::WriteBack:
-		case EndOfPipeWriteAction::InterruptWriteBack: return CommandBufferDebugOp::EopWriteBack;
-		case EndOfPipeWriteAction::Interrupt: return CommandBufferDebugOp::EopInterrupt;
-	}
-	EXIT("unsupported end-of-pipe write action\n");
-	return CommandBufferDebugOp::Unknown;
-}
-
-static bool TriggersInterrupt(EndOfPipeWriteAction action) {
-	return action == EndOfPipeWriteAction::Interrupt ||
-	       action == EndOfPipeWriteAction::InterruptWriteBack;
-}
-
-static void RecordEndOfPipeWrite(uint64_t submit_id, CommandBuffer& buffer, uint64_t destination,
-                                 uint64_t value, EndOfPipeWriteSize size,
-                                 EndOfPipeWriteAction action, int interrupt_event_id = 0,
-                                 uint32_t context_id = 0) {
-	EXIT_IF(destination == 0);
-	(void)buffer.Handle();
-
-	const auto width      = static_cast<uint32_t>(size);
-	const auto value_low  = static_cast<uint32_t>(value);
-	const auto value_high = static_cast<uint32_t>(value >> 32u);
-	const auto operation  = static_cast<uint32_t>(DebugOperation(action));
-	if (TriggersInterrupt(action)) {
-		buffer.SetDebugInfo(operation, submit_id, width, context_id, value_low, value_high,
-		                    destination);
-		TriggerEopEventAtEndOfPipe(buffer, interrupt_event_id, context_id);
-	} else {
-		buffer.SetDebugInfo(operation, submit_id, width, value_low, value_high, 0, destination);
-	}
-}
-
-void WriteAtEndOfPipe32(uint64_t submit_id, CommandBuffer& buffer, uint32_t* dst_gpu_addr,
-                        uint32_t value) {
-	RecordEndOfPipeWrite(submit_id, buffer, reinterpret_cast<uint64_t>(dst_gpu_addr), value,
-	                     EndOfPipeWriteSize::Dword, EndOfPipeWriteAction::Write);
-}
-
-void WriteAtEndOfPipeGds32(uint64_t submit_id, CommandBuffer& buffer, uint32_t* dst_gpu_addr,
-                           uint32_t dw_offset, uint32_t dw_num) {
-	EXIT_IF(dst_gpu_addr == nullptr);
-	(void)buffer.Handle();
-	buffer.SetDebugInfo(static_cast<uint32_t>(CommandBufferDebugOp::EopWrite), submit_id,
-	                    dw_offset, dw_num, 0, 0, reinterpret_cast<uint64_t>(dst_gpu_addr));
-}
-
-void WriteAtEndOfPipe64(uint64_t submit_id, CommandBuffer& buffer, uint64_t* dst_gpu_addr,
-                        uint64_t value) {
-	RecordEndOfPipeWrite(submit_id, buffer, reinterpret_cast<uint64_t>(dst_gpu_addr), value,
-	                     EndOfPipeWriteSize::Qword, EndOfPipeWriteAction::Write);
-}
-
-void WriteAtEndOfPipeWithWriteBack64(uint64_t submit_id, CommandBuffer& buffer,
-                                     uint64_t* dst_gpu_addr, uint64_t value) {
-	RecordEndOfPipeWrite(submit_id, buffer, reinterpret_cast<uint64_t>(dst_gpu_addr), value,
-	                     EndOfPipeWriteSize::Qword, EndOfPipeWriteAction::WriteBack);
-}
-
-void WriteAtEndOfPipeWithWriteBack32(uint64_t submit_id, CommandBuffer& buffer,
-                                     uint32_t* dst_gpu_addr, uint32_t value) {
-	RecordEndOfPipeWrite(submit_id, buffer, reinterpret_cast<uint64_t>(dst_gpu_addr), value,
-	                     EndOfPipeWriteSize::Dword, EndOfPipeWriteAction::WriteBack);
-}
-
-void WriteAtEndOfPipeWithInterruptWriteBack64(uint64_t submit_id, CommandBuffer& buffer,
-                                              uint64_t* dst_gpu_addr, uint64_t value, int event_id,
-                                              uint32_t context_id) {
-	RecordEndOfPipeWrite(submit_id, buffer, reinterpret_cast<uint64_t>(dst_gpu_addr), value,
-	                     EndOfPipeWriteSize::Qword, EndOfPipeWriteAction::InterruptWriteBack,
-	                     event_id, context_id);
-}
-
-void WriteAtEndOfPipeWithInterruptWriteBack32(uint64_t submit_id, CommandBuffer& buffer,
-                                              uint32_t* dst_gpu_addr, uint32_t value, int event_id,
-                                              uint32_t context_id) {
-	RecordEndOfPipeWrite(submit_id, buffer, reinterpret_cast<uint64_t>(dst_gpu_addr), value,
-	                     EndOfPipeWriteSize::Dword, EndOfPipeWriteAction::InterruptWriteBack,
-	                     event_id, context_id);
-}
-
-void WriteAtEndOfPipeWithInterrupt64(uint64_t submit_id, CommandBuffer& buffer,
-                                     uint64_t* dst_gpu_addr, uint64_t value, int event_id,
-                                     uint32_t context_id) {
-	RecordEndOfPipeWrite(submit_id, buffer, reinterpret_cast<uint64_t>(dst_gpu_addr), value,
-	                     EndOfPipeWriteSize::Qword, EndOfPipeWriteAction::Interrupt, event_id,
-	                     context_id);
-}
-
-void WriteAtEndOfPipeWithInterrupt32(uint64_t submit_id, CommandBuffer& buffer,
-                                     uint32_t* dst_gpu_addr, uint32_t value, int event_id,
-                                     uint32_t context_id) {
-	RecordEndOfPipeWrite(submit_id, buffer, reinterpret_cast<uint64_t>(dst_gpu_addr), value,
-	                     EndOfPipeWriteSize::Dword, EndOfPipeWriteAction::Interrupt, event_id,
-	                     context_id);
-}
-
 uint64_t PrepareVideoOutFlip(CommandBuffer& buffer, int handle, int index, int flip_mode,
                              int64_t flip_arg) {
 	for (;;) {
@@ -176,65 +69,6 @@ uint64_t PrepareVideoOutFlip(CommandBuffer& buffer, int handle, int index, int f
 	}
 }
 
-void WriteAtEndOfPipeWithInterruptWriteBackFlip32(uint64_t submit_id, CommandBuffer& buffer,
-                                                  uint32_t* dst_gpu_addr, uint32_t value,
-                                                  int handle, int index, int flip_mode,
-                                                  int64_t flip_arg, uint64_t request_id,
-                                                  int event_id) {
-	EXIT_IF(dst_gpu_addr == nullptr);
-	(void)buffer.Handle();
-	buffer.SetDebugInfo(static_cast<uint32_t>(CommandBufferDebugOp::EopWriteBackFlip), submit_id,
-	                    static_cast<uint32_t>(handle), static_cast<uint32_t>(index),
-	                    static_cast<uint32_t>(flip_mode), value, static_cast<uint64_t>(flip_arg));
-
-	auto& renderer  = buffer.GetContext();
-	auto& scheduler = renderer.GetCommandScheduler();
-	EXIT_IF(!scheduler.Active() || &buffer != &scheduler.Current());
-	scheduler.DeferPriorityOperation([&renderer, event_id, request_id] {
-		renderer.GetVideoOut().CompleteFlip(request_id);
-		renderer.TriggerInterrupt(event_id, 0);
-	});
-}
-
-void WriteAtEndOfPipeWithFlip32(uint64_t submit_id, CommandBuffer& buffer, uint32_t* dst_gpu_addr,
-                                uint32_t value, int handle, int index, int flip_mode,
-                                int64_t flip_arg, uint64_t request_id) {
-	EXIT_IF(dst_gpu_addr == nullptr);
-	(void)buffer.Handle();
-	buffer.SetDebugInfo(static_cast<uint32_t>(CommandBufferDebugOp::EopFlip), submit_id,
-	                    static_cast<uint32_t>(handle), static_cast<uint32_t>(index),
-	                    static_cast<uint32_t>(flip_mode), value, static_cast<uint64_t>(flip_arg));
-
-	auto& renderer  = buffer.GetContext();
-	auto& scheduler = renderer.GetCommandScheduler();
-	EXIT_IF(!scheduler.Active() || &buffer != &scheduler.Current());
-	scheduler.DeferPriorityOperation(
-	    [&renderer, request_id] { renderer.GetVideoOut().CompleteFlip(request_id); });
-}
-
-void WriteAtEndOfPipeOnlyFlip(uint64_t submit_id, CommandBuffer& buffer, int handle, int index,
-                              int flip_mode, int64_t flip_arg, uint64_t request_id) {
-	(void)buffer.Handle();
-	buffer.SetDebugInfo(static_cast<uint32_t>(CommandBufferDebugOp::EopOnlyFlip), submit_id,
-	                    static_cast<uint32_t>(handle), static_cast<uint32_t>(index),
-	                    static_cast<uint32_t>(flip_mode), 0, static_cast<uint64_t>(flip_arg));
-
-	auto& renderer  = buffer.GetContext();
-	auto& scheduler = renderer.GetCommandScheduler();
-	EXIT_IF(!scheduler.Active() || &buffer != &scheduler.Current());
-	scheduler.DeferPriorityOperation(
-	    [&renderer, request_id] { renderer.GetVideoOut().CompleteFlip(request_id); });
-}
-
-void TriggerEopEventAtEndOfPipe(CommandBuffer& buffer, int event_id, uint32_t context_id) {
-	(void)buffer.Handle();
-	auto& renderer  = buffer.GetContext();
-	auto& scheduler = renderer.GetCommandScheduler();
-	EXIT_IF(!scheduler.Active() || &buffer != &scheduler.Current());
-	scheduler.DeferPriorityOperation(
-	    [&renderer, event_id, context_id] { renderer.TriggerInterrupt(event_id, context_id); });
-}
-
 static void InterruptEventResetFunc(LibKernel::EventQueue::KernelEqueueEvent* event) {
 	EXIT_IF(event == nullptr);
 	event->triggered    = false;

src/graphics/host_gpu/renderer/sync.h

@@ -17,44 +17,8 @@ namespace Sync {
                                            uint64_t& value);
 [[nodiscard]] uint64_t ReadReferenceClock();
 
-void TriggerEopEventAtEndOfPipe(CommandBuffer& buffer, int event_id, uint32_t context_id);
-
-void WriteAtEndOfPipe32(uint64_t submit_id, CommandBuffer& buffer, uint32_t* dst_gpu_addr,
-                        uint32_t value);
-void WriteAtEndOfPipe64(uint64_t submit_id, CommandBuffer& buffer, uint64_t* dst_gpu_addr,
-                        uint64_t value);
-void WriteAtEndOfPipeGds32(uint64_t submit_id, CommandBuffer& buffer, uint32_t* dst_gpu_addr,
-                           uint32_t dw_offset, uint32_t dw_num);
-void WriteAtEndOfPipeWithWriteBack32(uint64_t submit_id, CommandBuffer& buffer,
-                                     uint32_t* dst_gpu_addr, uint32_t value);
-void WriteAtEndOfPipeWithWriteBack64(uint64_t submit_id, CommandBuffer& buffer,
-                                     uint64_t* dst_gpu_addr, uint64_t value);
-void WriteAtEndOfPipeWithInterrupt32(uint64_t submit_id, CommandBuffer& buffer,
-                                     uint32_t* dst_gpu_addr, uint32_t value, int event_id,
-                                     uint32_t context_id = 0);
-void WriteAtEndOfPipeWithInterrupt64(uint64_t submit_id, CommandBuffer& buffer,
-                                     uint64_t* dst_gpu_addr, uint64_t value, int event_id,
-                                     uint32_t context_id = 0);
-void WriteAtEndOfPipeWithInterruptWriteBack32(uint64_t submit_id, CommandBuffer& buffer,
-                                              uint32_t* dst_gpu_addr, uint32_t value, int event_id,
-                                              uint32_t context_id = 0);
-void WriteAtEndOfPipeWithInterruptWriteBack64(uint64_t submit_id, CommandBuffer& buffer,
-                                              uint64_t* dst_gpu_addr, uint64_t value, int event_id,
-                                              uint32_t context_id = 0);
-
 [[nodiscard]] uint64_t PrepareVideoOutFlip(CommandBuffer& buffer, int handle, int index,
                                            int flip_mode, int64_t flip_arg);
-void WriteAtEndOfPipeOnlyFlip(uint64_t submit_id, CommandBuffer& buffer, int handle, int index,
-                              int flip_mode, int64_t flip_arg, uint64_t request_id);
-void WriteAtEndOfPipeWithFlip32(uint64_t submit_id, CommandBuffer& buffer, uint32_t* dst_gpu_addr,
-                                uint32_t value, int handle, int index, int flip_mode,
-                                int64_t flip_arg, uint64_t request_id);
-void WriteAtEndOfPipeWithInterruptWriteBackFlip32(uint64_t submit_id, CommandBuffer& buffer,
-                                                  uint32_t* dst_gpu_addr, uint32_t value,
-                                                  int handle, int index, int flip_mode,
-                                                  int64_t flip_arg, uint64_t request_id,
-                                                  int event_id);
-
 int  AddEqEvent(RenderContext& renderer, LibKernel::EventQueue::KernelEqueue eq, int id,
                 void* udata);
 int  DeleteEqEvent(RenderContext& renderer, LibKernel::EventQueue::KernelEqueue eq, int id);

src/graphics/host_gpu/shaders/gpu_video_out_overlay.frag

@@ -1,15 +1,53 @@
 #version 450
 
-layout(location = 0) in vec2 uv;
 layout(location = 0) out vec4 color;
+
+vec3 SdrToHdr10(vec3 rgb, bool source_linear) {
+	rgb = clamp(rgb, 0.0, 1.0);
+	vec3 linear = source_linear ? rgb :
+	    mix(rgb / 12.92, pow((rgb + 0.055) / 1.055, vec3(2.4)),
+	        greaterThan(rgb, vec3(0.04045)));
+	// Convert primaries and scale 250-nit reference white relative to 10,000 nits.
+	const mat3 to_rec2020 = (250.0 / 10000.0) *
+	    mat3(0.627404, 0.069097, 0.016391,
+	         0.329283, 0.919540, 0.088013,
+	         0.043313, 0.011362, 0.895595);
+	vec3 powered = pow(to_rec2020 * linear, vec3(2610.0 / 16384.0));
+	return pow((3424.0 / 4096.0 + (2413.0 / 128.0) * powered) /
+	               (1.0 + (2392.0 / 128.0) * powered), vec3(2523.0 / 32.0));
+}
+
+#ifdef IMGUI_HDR
+layout(set = 0, binding = 0) uniform texture2D image;
+layout(set = 1, binding = 0) uniform sampler image_sampler;
+layout(location = 0) in struct { vec4 Color; vec2 UV; } In;
+
+void main() {
+	vec4 sampled = In.Color * texture(sampler2D(image, image_sampler), In.UV);
+	color = vec4(SdrToHdr10(sampled.rgb, false), sampled.a);
+}
+#else
+layout(location = 0) in vec2 uv;
 layout(binding = 0) uniform sampler2D source;
 layout(push_constant) uniform AlphaControl {
-	uint premultiplied;
+	uint flags; // bit 0: premultiplied, bit 1: HDR target, bit 2: PQ source, bit 3: linear source
 };
 
 void main() {
-	color = texture(source, uv);
-	if (premultiplied == 0u) {
-		color.rgb *= color.a;
+	const uint HDR_TARGET = 2u;
+	const uint PQ_SOURCE = 4u;
+	vec4 sampled = texture(source, uv);
+	bool premultiplied = (flags & 1u) != 0u;
+	bool convert_to_hdr = (flags & (HDR_TARGET | PQ_SOURCE)) == HDR_TARGET;
+	if (convert_to_hdr && premultiplied) {
+		sampled.rgb = sampled.a > 0.0 ? sampled.rgb / sampled.a : vec3(0.0);
 	}
+	if (convert_to_hdr) {
+		sampled.rgb = SdrToHdr10(sampled.rgb, (flags & 8u) != 0u);
+	}
+	if (!premultiplied || convert_to_hdr) {
+		sampled.rgb *= sampled.a;
+	}
+	color = sampled;
 }
+#endif

src/graphics/host_gpu/vulkanCommon.h

@@ -28,6 +28,14 @@ vk::Format  VulkanFormat(Prospero::BufferFormat guest_format);
 void        RequireVulkanSuccess(vk::Result result, const char* operation);
 vk::ShaderModule CompileSPV(std::span<const uint32_t> code, vk::Device device);
 
+[[nodiscard]] constexpr bool IsHdrSwapchainFormat(vk::Format format) {
+	switch (format) {
+		case vk::Format::eA2B10G10R10UnormPack32:
+		case vk::Format::eA2R10G10B10UnormPack32: return true;
+		default: return false;
+	}
+}
+
 template <typename Handle, typename... Args>
 void SetVulkanObjectNameF(vk::Device device, Handle handle, fmt::format_string<Args...> format,
                           Args&&... args) {

src/graphics/presentation/presenter.h

@@ -26,11 +26,13 @@ public:
 	~Presenter();
 	KYTY_CLASS_NO_COPY(Presenter);
 
-	[[nodiscard]] Frame&         PrepareFrame(CommandBuffer& command, const ImageInfo& info);
+	[[nodiscard]] Frame&         PrepareFrame(CommandBuffer& command, const ImageInfo& info,
+	                                         bool hdr = false);
 	[[nodiscard]] Frame&         PrepareBlankFrame(uint32_t width, uint32_t height, bool opaque,
 	                                               CommandBuffer* producer = nullptr);
 	[[nodiscard]] bool           PresentLastFrame();
 	[[nodiscard]] bool           IsGuestPaused() const noexcept;
+	[[nodiscard]] bool           IsHdrSupported() const noexcept;
 	[[nodiscard]] bool           NeedsSystemOverlayRefresh() const noexcept;
 	[[nodiscard]] RenderContext& Renderer() const noexcept;
 	void                         Present(Frame& frame);

src/graphics/presentation/systemOverlay.cpp

@@ -7,6 +7,7 @@
 #include "common/file.h"
 #include "common/logging/log.h"
 #include "common/stringUtils.h"
+#include "gpu_blit_shaders/gpu_imgui_hdr_spv.h"
 #include "graphics/host_gpu/graphicContext.h"
 #include "graphics/presentation/ddsImage.h"
 #include "imgui.h"
@@ -890,6 +891,11 @@ struct SystemOverlay::Impl {
 		info.PipelineInfoMain.PipelineRenderingCreateInfo.colorAttachmentCount    = 1;
 		info.PipelineInfoMain.PipelineRenderingCreateInfo.pColorAttachmentFormats = &color_format;
 		info.CheckVkResultFn = CheckVulkanResult;
+		if (IsHdrSwapchainFormat(format)) {
+			info.CustomShaderFragCreateInfo.sType    = VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO;
+			info.CustomShaderFragCreateInfo.codeSize = sizeof(GPU_IMGUI_HDR_SPV);
+			info.CustomShaderFragCreateInfo.pCode    = GPU_IMGUI_HDR_SPV;
+		}
 		EXIT_IF(!ImGui_ImplVulkan_Init(&info));
 		vulkan_initialized = true;
 	}

src/graphics/presentation/videoOut.cpp

@@ -272,6 +272,7 @@ public:
 	void                     Init(uint32_t width, uint32_t height);
 	FlipQueue&               GetFlipQueue() { return m_flip_queue; }
 	Graphics::RenderContext& Renderer() const noexcept { return m_renderer; }
+	bool                     IsHdrSupported() const noexcept { return m_presenter.IsHdrSupported(); }
 
 	void VblankBegin();
 	void VblankEnd();
@@ -558,14 +559,6 @@ Graphics::ImageInfo BufferAttributeGroup::ImageInfo(const VideoOutBuffer& buffer
 	if (!Graphics::DecodeVideoOutPixelFormat(attribute.pixel_format, pixel_format)) {
 		EXIT("unsupported video-out pixel format: 0x%016" PRIx64 "\n", attribute.pixel_format);
 	}
-	if (attribute.pixel_format == Graphics::VIDEO_OUT_PIXEL_FORMAT_R10_G10_B10_A2_BT2100_PQ) {
-		static std::atomic_flag warned = ATOMIC_FLAG_INIT;
-		if (!warned.test_and_set(std::memory_order_relaxed)) {
-			Log::WriteToConsoleAndLog(fmt::format(
-			    "Warning: HDR-to-SDR conversion is not implemented; displaying PQ video-out "
-			    "pixels unchanged (format=0x{:016x}).\n", attribute.pixel_format));
-		}
-	}
 	const auto          tile_mode = Graphics::Prospero::TileMode::kRenderTarget;
 	Graphics::ImageInfo info {};
 	info.data            = {buffer.data_address, 0};
@@ -1071,31 +1064,31 @@ void FlipQueue::Prepare(uint64_t request_id, Graphics::CommandBuffer& buffer) {
 	uint32_t            height  = 0;
 	bool                current = false;
 	bool                premultiplied_alpha = false;
+	bool                hdr                 = false;
 	{
 		Common::LockGuard lock(cfg->mutex);
 		current = cfg->opened && !cfg->closing && cfg->generation == generation;
-		if (current) {
-			if (special) {
-				width  = cfg->width;
-				height = cfg->height;
-			} else {
-				const auto& surface = cfg->buffers[index];
-				if (!surface.Occupied()) {
-					EXIT("cannot prepare flip from an unregistered surface, id=%" PRIu64
-					     " index=%d\n",
-					     request_id, index);
-				}
-				if (surface.group_index >= VIDEO_OUT_BUFFER_ATTRIBUTE_NUM_MAX ||
-				    !cfg->groups[surface.group_index].occupied) {
-					EXIT("video-out surface references an unavailable attribute group, id=%" PRIu64
-					     " index=%d group=%d\n",
-					     request_id, index, surface.group_index);
-				}
-				source_info = cfg->groups[surface.group_index].ImageInfo(surface);
-				premultiplied_alpha =
-				    (cfg->groups[surface.group_index].attribute.option &
-				     VIDEO_OUT_BUFFER_ATTRIBUTE_OPTION_ALPHA_PREMULTIPLIED) != 0;
+		if (current && special) {
+			width  = cfg->width;
+			height = cfg->height;
+		} else if (current) {
+			const auto& surface = cfg->buffers[index];
+			if (!surface.Occupied()) {
+				EXIT("cannot prepare flip from an unregistered surface, id=%" PRIu64
+				     " index=%d\n",
+				     request_id, index);
 			}
+			if (surface.group_index >= VIDEO_OUT_BUFFER_ATTRIBUTE_NUM_MAX ||
+			    !cfg->groups[surface.group_index].occupied) {
+				EXIT("video-out surface references an unavailable attribute group, id=%" PRIu64
+				     " index=%d group=%d\n",
+				     request_id, index, surface.group_index);
+			}
+			const auto& group   = cfg->groups[surface.group_index];
+			source_info         = group.ImageInfo(surface);
+			hdr                 = Graphics::IsHdrVideoOutPixelFormat(group.attribute.pixel_format);
+			premultiplied_alpha = (group.attribute.option &
+			                       VIDEO_OUT_BUFFER_ATTRIBUTE_OPTION_ALPHA_PREMULTIPLIED) != 0;
 		}
 	}
 	if (!current) {
@@ -1108,12 +1101,18 @@ void FlipQueue::Prepare(uint64_t request_id, Graphics::CommandBuffer& buffer) {
 		}
 		return;
 	}
+	static std::atomic_flag warned = ATOMIC_FLAG_INIT;
+	if (hdr && !m_presenter.IsHdrSupported() && !warned.test_and_set(std::memory_order_relaxed)) {
+		Log::WriteToConsoleAndLog(
+		    "Warning: HDR output is unavailable; HDR-to-SDR conversion is not implemented, "
+		    "displaying PQ video-out pixels unchanged.\n");
+	}
 	Graphics::Presenter::Frame* frame = nullptr;
 	if (special) {
 		frame = &m_presenter.PrepareBlankFrame(width, height, index == VIDEO_OUT_BUFFER_INDEX_BLACK,
 		                                       &buffer);
 	} else {
-		frame = &m_presenter.PrepareFrame(buffer, source_info);
+		frame = &m_presenter.PrepareFrame(buffer, source_info, hdr);
 	}
 
 	Common::LockGuard lock(m_mutex);
@@ -1790,11 +1789,14 @@ KYTY_SYSV_ABI int VideoOutGetOutputStatus(int handle, VideoOutOutputStatus* stat
 
 	int32_t attribute3 = 0;
 	Loader::SystemContentParamSfoGetInt("ATTRIBUTE3", &attribute3);
+	int32_t attribute = 0;
+	const bool hdr_allowed = !Loader::SystemContentParamSfoGetInt("ATTRIBUTE", &attribute) ||
+	                         (attribute & 0x20000000) != 0;
 	ctx->mutex.Lock();
 	// Primary output reports 4K unless param.json Video-out Info enables resolution detection.
 	status->resolution =
 	    ((attribute3 & 4) != 0 && ctx->width < 3840 && ctx->height < 2160 ? 1u : 2u);
-	status->dynamicRange = 1;
+	status->dynamicRange = hdr_allowed && DriverState().IsHdrSupported() ? 2u : 1u;
 	status->refreshRate =
 	    (ctx->output_mode == VIDEO_OUT_OUTPUT_MODE_119_88HZ || Config::GetVblankFrequency() >= 119
 	         ? VIDEO_OUT_REFRESH_RATE_119_88HZ

src/graphics/presentation/window/swapchain.cpp

@@ -17,6 +17,7 @@
 
 #include <algorithm>
 #include <array>
+#include <atomic>
 #include <limits>
 #include <memory>
 #include <vector>
@@ -31,6 +32,8 @@ struct Presenter::Frame {
 	vk::ImageView view         = nullptr;
 	uint64_t      present_tick = 0;
 	bool          busy         = false;
+	bool          hdr          = false;
+	bool          blank        = false;
 
 	void Configure(GraphicContext& graphics, vk::Extent2D extent, vk::Format format);
 	void Transit(vk::CommandBuffer command, vk::ImageLayout layout, vk::AccessFlags2 access);
@@ -263,7 +266,7 @@ public:
 	void                 Create();
 	void                 Recreate(bool surface_lost = false);
 	[[nodiscard]] bool   NeedsResize() const;
-	[[nodiscard]] Status AcquireNextImage();
+	[[nodiscard]] Status AcquireNextImage(CommandScheduler& scheduler);
 	[[nodiscard]] bool   PrepareSystemOverlay();
 	void     RecordPresentCommands(CommandBuffer& command, Presenter::Frame* source,
 	                               const Presenter::Layer& overlay, bool draw_system_overlay);
@@ -274,20 +277,31 @@ public:
 		return static_cast<uint32_t>(m_images.size());
 	}
 	[[nodiscard]] vk::Format Format() const noexcept { return m_format; }
+	[[nodiscard]] bool IsHdrSupported() const noexcept { return m_hdr_supported.load(); }
+	bool SetHdr(bool hdr) {
+		const bool changed = m_hdr_active != (hdr && IsHdrSupported());
+		m_hdr_requested = hdr;
+		return changed;
+	}
 
 private:
+	vk::SurfaceFormatKHR SelectFormat();
 	void Destroy();
 	void DrawOverlay(vk::CommandBuffer command, const Presenter::Layer& layer);
 
 	WindowContext&   m_window;
 	vk::SwapchainKHR m_handle = nullptr;
 	vk::Format       m_format = vk::Format::eUndefined;
+	std::atomic_bool m_hdr_supported {false};
+	bool             m_hdr_requested = false;
+	bool             m_hdr_active    = false;
 	vk::Extent2D     m_extent {};
 	// Drawable pixel size observed when this swapchain was created.
 	vk::Extent2D                   m_window_extent {};
 	std::vector<vk::Image>         m_images;
 	std::vector<vk::ImageView>     m_image_views;
 	std::vector<vk::Semaphore>     m_image_acquired;
+	std::vector<uint64_t>          m_image_acquired_ticks;
 	std::vector<vk::Semaphore>     m_render_complete;
 	std::unique_ptr<SystemOverlay> m_system_overlay;
 	vk::DescriptorSetLayout        m_overlay_descriptors = nullptr;
@@ -346,6 +360,40 @@ struct Presenter::Impl {
 	std::atomic<uint64_t> presented_overlay_revision {0};
 };
 
+vk::SurfaceFormatKHR Swapchain::SelectFormat() {
+	const auto& formats    = m_window.surface_capabilities.formats;
+	auto&       graphics   = m_window.graphic_ctx;
+	const auto  hdr_format = std::ranges::find_if(formats, [&](const auto& candidate) {
+		return candidate.colorSpace == vk::ColorSpaceKHR::eHdr10St2084EXT &&
+		       IsHdrSwapchainFormat(candidate.format) &&
+		       static_cast<bool>(
+		           graphics.GetFormatProperties(candidate.format).optimalTilingFeatures &
+		           vk::FormatFeatureFlagBits::eBlitDst);
+	});
+	m_hdr_supported        = Config::HdrEnabled() && hdr_format != formats.end();
+	m_hdr_active           = m_hdr_requested && IsHdrSupported();
+	if (m_hdr_active) {
+		return *hdr_format;
+	}
+	if (Config::HdrEnabled() && !IsHdrSupported()) {
+		LOGF("HDR10 surface unavailable; using SDR output\n");
+	}
+	if (formats.size() == 1 && formats.front().format == vk::Format::eUndefined) {
+		return {vk::Format::eR8G8B8A8Unorm, vk::ColorSpaceKHR::eSrgbNonlinear};
+	}
+	for (const auto& candidate: formats) {
+		if (candidate.colorSpace != vk::ColorSpaceKHR::eSrgbNonlinear) {
+			continue;
+		}
+		switch (candidate.format) {
+			case vk::Format::eB8G8R8A8Unorm:
+			case vk::Format::eR8G8B8A8Unorm: return candidate;
+			default: break;
+		}
+	}
+	EXIT("no supported UNORM swapchain format\n");
+}
+
 void Swapchain::Create() {
 	auto& graphics = m_window.graphic_ctx;
 	EXIT_IF(graphics.device == nullptr);
@@ -385,18 +433,7 @@ void Swapchain::Create() {
 	        ? vk::CompositeAlphaFlagBitsKHR::eOpaque
 	        : vk::CompositeAlphaFlagBitsKHR::eInherit;
 
-	vk::SurfaceFormatKHR format {vk::Format::eR8G8B8A8Unorm, vk::ColorSpaceKHR::eSrgbNonlinear};
-	if (surface.formats.size() != 1 || surface.formats.front().format != vk::Format::eUndefined) {
-		const auto it = std::find_if(surface.formats.begin(), surface.formats.end(),
-		                             [](const vk::SurfaceFormatKHR& candidate) {
-			                             return candidate.format == vk::Format::eB8G8R8A8Unorm ||
-			                                    candidate.format == vk::Format::eR8G8B8A8Unorm;
-		                             });
-		if (it == surface.formats.end()) {
-			EXIT("no supported UNORM swapchain format\n");
-		}
-		format = *it;
-	}
+	const auto format = SelectFormat();
 	m_format                      = format.format;
 	const auto swapchain_features = graphics.GetFormatProperties(m_format).optimalTilingFeatures;
 	if (!static_cast<bool>(swapchain_features & vk::FormatFeatureFlagBits::eBlitDst)) {
@@ -436,6 +473,7 @@ void Swapchain::Create() {
 	RequireVulkanSuccess(graphics.device.createSwapchainKHR(&create_info, nullptr, &m_handle),
 	                     "vkCreateSwapchainKHR");
 	EXIT_IF(m_handle == nullptr);
+	LOGF("Vulkan presentation: %s\n", m_hdr_active ? "HDR10 (PQ)" : "SDR");
 
 	m_images = EnumerateVulkan<vk::Image>(
 	    "vkGetSwapchainImagesKHR", [&](uint32_t* count, vk::Image* images) {
@@ -464,6 +502,7 @@ void Swapchain::Create() {
 	vk::SemaphoreCreateInfo semaphore_info {};
 	semaphore_info.sType = vk::StructureType::eSemaphoreCreateInfo;
 	m_image_acquired.resize(m_images.size());
+	m_image_acquired_ticks.assign(m_images.size(), 0);
 	m_render_complete.resize(m_images.size());
 	for (size_t i = 0; i < m_images.size(); i++) {
 		RequireVulkanSuccess(
@@ -532,6 +571,7 @@ void Swapchain::Destroy() {
 	m_images.clear();
 	m_image_views.clear();
 	m_image_acquired.clear();
+	m_image_acquired_ticks.clear();
 	m_render_complete.clear();
 }
 
@@ -557,8 +597,11 @@ bool Swapchain::NeedsResize() const {
 	       m_window_extent.height != m_window.graphic_ctx.screen_height;
 }
 
-Swapchain::Status Swapchain::AcquireNextImage() {
+Swapchain::Status Swapchain::AcquireNextImage(CommandScheduler& scheduler) {
 	EXIT_IF(m_handle == nullptr || m_frame_index >= m_image_acquired.size());
+	// Repeated frames can wrap the semaphore ring without acquiring a new source frame.
+	// Wait for the submission that consumed this slot before signaling it again.
+	scheduler.Wait(m_image_acquired_ticks[m_frame_index]);
 	m_image_index     = static_cast<uint32_t>(-1);
 	const auto result = m_window.graphic_ctx.device.acquireNextImageKHR(
 	    m_handle, std::numeric_limits<uint64_t>::max(), m_image_acquired[m_frame_index], nullptr,
@@ -692,9 +735,11 @@ void Swapchain::DrawOverlay(vk::CommandBuffer command, const Presenter::Layer& l
 	write.descriptorType  = vk::DescriptorType::eCombinedImageSampler;
 	write.pImageInfo      = &image;
 	command.pushDescriptorSetKHR(vk::PipelineBindPoint::eGraphics, m_overlay_layout, 0, 1, &write);
-	const uint32_t premultiplied = layer.premultiplied_alpha;
+	const uint32_t flags = (layer.premultiplied_alpha ? 1u : 0u) |
+	                       (m_hdr_active ? 2u : 0u) | (frame.hdr ? 4u : 0u) |
+	                       (frame.image.format == vk::Format::eR16G16B16A16Sfloat ? 8u : 0u);
 	command.pushConstants(m_overlay_layout, vk::ShaderStageFlagBits::eFragment, 0,
-	                      sizeof(premultiplied), &premultiplied);
+	                      sizeof(flags), &flags);
 	command.bindPipeline(vk::PipelineBindPoint::eGraphics, m_overlay_pipeline);
 	vk::RenderingAttachmentInfo attachment {};
 	attachment.imageView   = m_image_views[m_image_index];
@@ -828,7 +873,9 @@ uint64_t Swapchain::Submit(CommandScheduler& scheduler) {
 	SubmitInfo submit;
 	submit.AddWait(m_image_acquired[m_frame_index], 1, vk::PipelineStageFlagBits::eTransfer);
 	submit.AddSignal(m_render_complete[m_image_index]);
-	return scheduler.Submit(submit);
+	const auto tick                       = scheduler.Submit(submit);
+	m_image_acquired_ticks[m_frame_index] = tick;
+	return tick;
 }
 
 Swapchain::Status Swapchain::Present() {
@@ -868,7 +915,7 @@ Presenter::Presenter(WindowContext& window): m_impl(std::make_unique<Impl>(windo
 
 Presenter::~Presenter() = default;
 
-Presenter::Frame& Presenter::PrepareFrame(CommandBuffer& buffer, const ImageInfo& info) {
+Presenter::Frame& Presenter::PrepareFrame(CommandBuffer& buffer, const ImageInfo& info, bool hdr) {
 	KYTY_PROFILER_FUNCTION();
 	EXIT_IF(buffer.IsInvalid());
 	auto frame_format = info.pixel_format;
@@ -878,6 +925,8 @@ Presenter::Frame& Presenter::PrepareFrame(CommandBuffer& buffer, const ImageInfo
 		default: break;
 	}
 	auto* frame = m_impl->frames.Acquire({info.extent.width, info.extent.height}, frame_format);
+	frame->hdr   = hdr;
+	frame->blank = false;
 	Common::LockGuard render_lock(m_impl->renderer.GetMutex());
 	auto&             image = m_impl->ResolveSurface(info);
 	if (image.backing.format == vk::Format::eUndefined) {
@@ -894,6 +943,8 @@ Presenter::Frame& Presenter::PrepareBlankFrame(uint32_t width, uint32_t height,
 	KYTY_PROFILER_FUNCTION();
 	auto              format = m_impl->frames.GetFormat();
 	auto*             frame  = m_impl->frames.Acquire({width, height}, format);
+	frame->hdr   = false;
+	frame->blank = true;
 	Common::LockGuard render_lock(m_impl->renderer.GetMutex());
 	frame->Configure(m_impl->window.graphic_ctx, {width, height}, format);
 	vk::ClearColorValue clear {};
@@ -922,6 +973,10 @@ bool Presenter::IsGuestPaused() const noexcept {
 	return m_impl->window.loop.paused.load(std::memory_order_acquire);
 }
 
+bool Presenter::IsHdrSupported() const noexcept {
+	return m_impl->swapchain.IsHdrSupported();
+}
+
 bool Presenter::NeedsSystemOverlayRefresh() const noexcept {
 	const auto visual = GetSystemOverlayVisualState();
 	return visual.active ||
@@ -967,12 +1022,16 @@ void Presenter::Impl::Present() {
 	KYTY_PROFILER_FUNCTION();
 
 	const auto overlay_visual = GetSystemOverlayVisualState();
+	// Keep black/blank flips in the current color space; only real main-bus frames switch it.
+	const auto* main_frame = layers[0].frame;
+	const bool hdr_changed = main_frame != nullptr && !main_frame->blank &&
+	                         swapchain.SetHdr(main_frame->hdr);
 	// Some window systems keep presenting an old swapchain after a resize.
-	if (swapchain.NeedsResize()) {
+	if (hdr_changed || swapchain.NeedsResize()) {
 		RecoverSwapchain(Swapchain::Status::Recreate);
 	}
 	for (uint32_t attempt = 0; attempt < 2; attempt++) {
-		auto status = swapchain.AcquireNextImage();
+		auto status = swapchain.AcquireNextImage(present_scheduler);
 		if (status != Swapchain::Status::Success) {
 			RecoverSwapchain(status);
 			continue;

src/graphics/presentation/window/vulkanWindow.cpp

@@ -478,6 +478,13 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 		provoking_vertex.pNext = supported_features2.pNext;
 		supported_features2.pNext = &provoking_vertex;
 	}
+	const bool maintenance8_extension =
+	    HasExtension(device_extensions, VK_KHR_MAINTENANCE_8_EXTENSION_NAME);
+	vk::PhysicalDeviceMaintenance8FeaturesKHR maintenance8 {};
+	if (maintenance8_extension) {
+		maintenance8.pNext = supported_features2.pNext;
+		supported_features2.pNext = &maintenance8;
+	}
 	const bool image_atomic_int64_extension =
 	    HasExtension(device_extensions, VK_EXT_SHADER_IMAGE_ATOMIC_INT64_EXTENSION_NAME);
 	vk::PhysicalDeviceShaderImageAtomicInt64FeaturesEXT image_atomic_int64 {};
@@ -542,6 +549,9 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 	graphics.attachment_feedback_loop_enabled =
 	    feedback_extensions && feedback_layout.attachmentFeedbackLoopLayout &&
 	    feedback_dynamic.attachmentFeedbackLoopDynamicState;
+	graphics.depth_color_copy_enabled = maintenance8_extension && maintenance8.maintenance8;
+	LOGF("Vulkan depth/color copy support: %s\n",
+	     graphics.depth_color_copy_enabled ? "true" : "false");
 	LOGF("Vulkan depth feedback support: %s\n",
 	     graphics.attachment_feedback_loop_enabled ? "true" : "false");
 	if (graphics.mesh_shader_enabled) {
@@ -638,6 +648,10 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 		image_atomic_int64.sparseImageInt64Atomics = VK_FALSE;
 		create_info.pNext = &image_atomic_int64;
 	}
+	if (graphics.depth_color_copy_enabled) {
+		maintenance8.pNext = const_cast<void*>(create_info.pNext);
+		create_info.pNext = &maintenance8;
+	}
 	create_info.pQueueCreateInfos       = &queue_create_info;
 	create_info.queueCreateInfoCount    = 1;
 	create_info.enabledExtensionCount   = static_cast<uint32_t>(device_extensions.size());
@@ -671,6 +685,11 @@ static void VulkanGetExtensions(VulkanExtensions& r) {
 	        });
 
 	r.enable_validation_layers = Config::VulkanValidationEnabled();
+	if (Config::HdrEnabled() &&
+	    HasExtension(r.available_extensions, VK_EXT_SWAPCHAIN_COLOR_SPACE_EXTENSION_NAME) &&
+	    !HasExtension(r.required_extensions, VK_EXT_SWAPCHAIN_COLOR_SPACE_EXTENSION_NAME)) {
+		r.required_extensions.push_back(VK_EXT_SWAPCHAIN_COLOR_SPACE_EXTENSION_NAME);
+	}
 
 	if (HasExtension(r.available_extensions, VK_EXT_DEBUG_UTILS_EXTENSION_NAME)) {
 		r.required_extensions.push_back(VK_EXT_DEBUG_UTILS_EXTENSION_NAME);
@@ -1002,6 +1021,7 @@ void WindowContext::CreateVulkan() {
 			graphic_ctx.memory_budget_ext_enabled = true;
 		}
 		for (const auto* extension: {VK_EXT_ROBUSTNESS_2_EXTENSION_NAME,
+		                             VK_KHR_MAINTENANCE_8_EXTENSION_NAME,
 		                             VK_EXT_SHADER_IMAGE_ATOMIC_INT64_EXTENSION_NAME,
 		                             VK_EXT_PROVOKING_VERTEX_EXTENSION_NAME,
 		                             VK_EXT_MESH_SHADER_EXTENSION_NAME,

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

@@ -323,25 +323,25 @@ uint32_t QueryDimensions(ValueEmitContext& ctx, const IR::MemoryInfo& mem,
 	return vector;
 }
 
-uint32_t PackedOffset(ValueEmitContext& ctx, const IR::MemoryInfo& mem, const IR::Inst& address,
-                      const ImageSampleLayout& layout, ImageDimension dimension) {
+uint32_t PackedOffset(ValueEmitContext& ctx, IR::Value value, ImageDimension dimension,
+                      uint32_t& mask) {
 	const auto components = ImageDimensionInfoFor(dimension).spatial_components;
-	const auto zero       = ConstantI32(ctx.state, 0);
-	if (layout.offset == NoImageComponent || mem.image_address_components <= layout.offset) {
-		if (components == 1u) return zero;
-		const auto result = ctx.state.builder.AllocateId();
-		if (components == 3u) {
-			ctx.state.builder.AddFunction(spv::OpCompositeConstruct, TypeI32Vector(ctx.state, 3),
-			                              result, zero, zero, zero);
-		} else {
-			ctx.state.builder.AddFunction(spv::OpCompositeConstruct, TypeI32Vector(ctx.state, 2),
-			                              result, zero, zero);
+	uint32_t values[3] {};
+	if (value.IsImmediate()) {
+		// RDNA2 packs signed six-bit texel offsets into the low bits of each byte.
+		for (uint32_t index = 0; index < components; index++) {
+			const auto bits = (value.U32() >> (index * 8u)) & 0x3fu;
+			values[index] = ConstantI32(ctx.state, static_cast<int32_t>(bits ^ 0x20u) - 0x20);
 		}
-		return result;
+		mask |= spv::ImageOperandsConstOffsetMask;
+		return components == 1u
+		           ? values[0]
+		           : ctx.state.builder.Constant(spv::OpConstantComposite,
+		                                        TypeI32Vector(ctx.state, components),
+		                                        std::span(values).first(components));
 	}
-	const auto packed    = Unary(ctx.state, spv::OpBitcast, TypeI32(ctx.state),
-	                             AddressU32(ctx, mem, address, layout.offset));
-	uint32_t   values[3] = {zero, zero, zero};
+	mask |= spv::ImageOperandsOffsetMask;
+	const auto packed = Unary(ctx.state, spv::OpBitcast, TypeI32(ctx.state), ctx.Def(value));
 	for (uint32_t index = 0; index < components; index++) {
 		values[index] = ctx.state.builder.AllocateId();
 		ctx.state.builder.AddFunction(spv::OpBitFieldSExtract, TypeI32(ctx.state), values[index],
@@ -350,13 +350,8 @@ uint32_t PackedOffset(ValueEmitContext& ctx, const IR::MemoryInfo& mem, const IR
 	}
 	if (components == 1u) return values[0];
 	const auto result = ctx.state.builder.AllocateId();
-	if (components == 3u) {
-		ctx.state.builder.AddFunction(spv::OpCompositeConstruct, TypeI32Vector(ctx.state, 3),
-		                              result, values[0], values[1], values[2]);
-	} else {
-		ctx.state.builder.AddFunction(spv::OpCompositeConstruct, TypeI32Vector(ctx.state, 2),
-		                              result, values[0], values[1]);
-	}
+	ctx.state.builder.AddFunction(spv::OpCompositeConstruct, TypeI32Vector(ctx.state, components),
+	                              result, std::span(values).first(components));
 	return result;
 }
 
@@ -794,21 +789,28 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 		const auto  layout         = Layout(mem);
 		const auto  numeric_class  = image.numeric_class;
 		const bool  dref           = HasFlag(mem, Decoder::ImageSampleFlagCompare);
+		IR::Value offset_value;
+		if (layout.offset != NoImageComponent) {
+			const auto value = address->Arg(layout.offset).Resolve();
+			offset_value = IR::ResolveActiveU32(value, inst.Arg(inst.NumArgs() - 1u).Resolve());
+			if (offset_value.IsEmpty()) offset_value = value;
+		}
 		if (dref && state.program.info.images[mem.resource].conversion_format !=
 		                Prospero::BufferFormat::kInvalid) {
 			ctx.Fail(inst, "uses depth comparison with a converted image");
 			return;
 		}
 		if (op == IR::ValueOpcode::ImageGatherRaw) {
-			const auto coord = CoordF32(ctx, mem, *address, layout.coord,
-			                            dimension_info.coordinate_components, image.cube);
 			if (dimension == ImageDimension::Dim1D) {
-				if (dref || !HasFlag(mem, Decoder::ImageSampleFlagLevelZero) ||
+				if (image.indirect_root == mem.resource || dref ||
+				    !HasFlag(mem, Decoder::ImageSampleFlagLevelZero) ||
 				    HasFlag(mem, Decoder::ImageSampleFlagOffset) ||
 				    HasFlag(mem, Decoder::ImageSampleFlagGatherHorizontal)) {
 					ctx.Fail(inst, "has an unsupported 1D gather variant");
 					return;
 				}
+				const auto coord = CoordF32(ctx, mem, *address, layout.coord,
+				                            dimension_info.coordinate_components, image.cube);
 				const auto sample = EmitOneDimensionalGatherLz(ctx, mem, coord, numeric_class);
 				ctx.Define(inst, ResultVector(ctx, UnpackImageGather(ctx, mem, sample),
 				                              numeric_class, false, mem, true));
@@ -829,32 +831,38 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 			} else {
 				component_or_dref = ConstantU32(state, ImageGatherSource(state, mem));
 			}
-			uint32_t operand_mask = 0;
-			uint32_t offset = 0;
-			if (HasFlag(mem, Decoder::ImageSampleFlagGatherHorizontal)) {
-				operand_mask = spv::ImageOperandsConstOffsetsMask;
-				offset = HorizontalOffsets(state, dimension);
-			} else if (layout.offset != NoImageComponent) {
-				operand_mask = spv::ImageOperandsOffsetMask;
-				offset = PackedOffset(ctx, mem, *address, layout, dimension);
-			}
 			const auto sampler_id = LoadSamplerDescriptor(state, mem.sampler);
-			const auto EmitGather = [&](uint32_t mip) {
-				const auto sampled = MakeSampledImage(state, mem.resource, sampler_id, mip);
-				const auto sample = state.builder.AllocateId();
-				const std::array operands {operand_mask, offset};
-				state.builder.AddFunction(
-				    dref ? spv::OpImageDrefGather : spv::OpImageGather, result_type,
-				    sample, sampled, coord, component_or_dref,
-				    std::span(operands).first(operand_mask != 0u ? 2u : 0u));
-				return sample;
+			const auto EmitGather = [&](uint32_t resource, uint32_t array_index) {
+				const auto& candidate = state.program.info.images[resource];
+				const auto dimension = candidate.dimension;
+				const auto coord = CoordF32(
+				    ctx, mem, *address, layout.coord,
+				    ImageDimensionInfoFor(dimension).coordinate_components, candidate.cube);
+				uint32_t operand_mask = 0;
+				uint32_t offset = 0;
+				if (HasFlag(mem, Decoder::ImageSampleFlagGatherHorizontal)) {
+					operand_mask = spv::ImageOperandsConstOffsetsMask;
+					offset = HorizontalOffsets(state, dimension);
+				} else if (layout.offset != NoImageComponent) {
+					offset = PackedOffset(ctx, offset_value, dimension, operand_mask);
+				}
+				const auto EmitMip = [&](uint32_t mip) {
+					const auto sampled = MakeSampledImage(state, resource, sampler_id, mip, array_index);
+					const auto sample = state.builder.AllocateId();
+					const std::array operands {operand_mask, offset};
+					state.builder.AddFunction(
+					    dref ? spv::OpImageDrefGather : spv::OpImageGather, result_type,
+					    sample, sampled, coord, component_or_dref,
+					    std::span(operands).first(operand_mask != 0u ? 2u : 0u));
+					return sample;
+				};
+				return candidate.mip_mode == IR::ImageMipMode::Dynamic && candidate.mip_count > 1u
+				           ? EmitIndexSwitch(
+				                 state, GatherMip(ctx, inst, mem, *address, layout, candidate.mip_count),
+				                 candidate.mip_count, result_type, EmitMip)
+				           : EmitMip(0u);
 			};
-			const auto sample = image.mip_mode == IR::ImageMipMode::Dynamic && image.mip_count > 1u
-			                        ? EmitIndexSwitch(
-			                              state, GatherMip(ctx, inst, mem, *address, layout,
-			                                               image.mip_count),
-			                              image.mip_count, result_type, EmitGather)
-			                        : EmitGather(0);
+			const auto sample = EmitImageAccess(ctx, inst, result_type, EmitGather);
 			ctx.Define(inst, ResultVector(ctx, UnpackImageGather(ctx, mem, sample),
 			                              result_numeric_class, false, mem, true));
 			return;
@@ -878,7 +886,7 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 				dref_value = AddressF32(ctx, mem, *address, layout.dref);
 			}
 		}
-		std::array<uint32_t, 3> operands {};
+		std::array<uint32_t, 4> operands {};
 		size_t operand_count = 0;
 		if (HasFlag(mem, Decoder::ImageSampleFlagDerivative)) {
 			operands[0] = spv::ImageOperandsGradMask;
@@ -908,9 +916,20 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 			    state, ctx.Def(handle->Arg(0)), static_cast<uint32_t>(samplers.size()),
 			    TypeU32(state), [&](uint32_t ordinal) { return ConstantU32(state, samplers[ordinal]); });
 		}
+		if (layout.offset != NoImageComponent &&
+		    !offset_value.IsImmediate()) {
+			ctx.Fail(inst, "has a nonconstant sample texel offset");
+		}
 		const auto sampler_id = LoadSamplerDescriptor(state, mem.sampler, sampler_index);
 		const auto EmitSample = [&](uint32_t resource, uint32_t array_index) {
 			const auto& candidate = state.program.info.images[resource];
+			auto sample_operands = operands;
+			auto sample_operand_count = operand_count;
+			if (layout.offset != NoImageComponent) {
+				sample_operand_count = std::max(size_t{1}, sample_operand_count);
+				sample_operands[sample_operand_count++] = PackedOffset(
+				    ctx, offset_value, candidate.dimension, sample_operands[0]);
+			}
 			const auto coord =
 			    CoordF32(ctx, mem, *address, layout.coord,
 			             ImageDimensionInfoFor(candidate.dimension).coordinate_components,
@@ -920,7 +939,7 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 			const auto sample = state.builder.AllocateId();
 			state.builder.AddFunction(opcode, result_type, sample, sampled, coord,
 			                          std::span(&dref_value, dref ? 1u : 0u),
-			                          std::span<const uint32_t>(operands).first(operand_count));
+			                          std::span<const uint32_t>(sample_operands).first(sample_operand_count));
 			return sample;
 		};
 		auto result = EmitImageAccess(ctx, inst, result_type, EmitSample);

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

@@ -687,7 +687,8 @@ void Translator::IMAGE_SAMPLE(const Decoder::Instruction& inst) {
 	const auto resource = GetImageResource(memory);
 	const auto sampler  = GetSamplerResource(memory);
 	const auto address  = MakeImageAddress(inst, MemorySourceAt(inst, 0));
-	const auto result   = ir.Emit(IR::ValueOpcode::ImageSampleRaw, {resource, sampler, address},
+	const auto result   = ir.Emit(IR::ValueOpcode::ImageSampleRaw,
+	                              {resource, sampler, address, ir.GetExec()},
 	                              AddMemoryInfo(memory, inst.pc));
 	const bool dref     = (memory.image_sample_flags & Decoder::ImageSampleFlagCompare) != 0u;
 	if (dref && memory.data_bits != 16u) {
@@ -714,7 +715,7 @@ void Translator::IMAGE_GATHER(const Decoder::Instruction& inst) {
 	     has_lod ? resource.Instruction()->Arg(1) : IR::Value(0u),
 	     has_lod ? resource.Instruction()->Arg(3) : IR::Value(0u),
 	     has_lod ? sampler.Instruction()->Arg(1) : IR::Value(0u),
-	     has_lod ? sampler.Instruction()->Arg(2) : IR::Value(0u)},
+	     has_lod ? sampler.Instruction()->Arg(2) : IR::Value(0u), ir.GetExec()},
 	    AddMemoryInfo(memory, inst.pc));
 	for (uint32_t index = 0; index < memory.data_dwords; index++) {
 		WriteOperand(OffsetOperand(inst.dst, index),

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

@@ -3,6 +3,7 @@
 #include "graphics/shader/recompiler/BufferFormat.h"
 
 #include <fmt/format.h>
+#include <xbyak/xbyak.h>
 #include <map>
 #include <new>
 #include <unordered_map>
@@ -99,6 +100,9 @@ bool EquivalentValue(const ResourcePlan& program, Value left, Value right,
 
 } // namespace
 
+ResourcePlan::ResourcePlan() = default;
+ResourcePlan::ResourcePlan(ResourcePlan&&) noexcept = default;
+
 ResourcePlan::~ResourcePlan() {
 	for (auto& inst: value_storage) {
 		inst.Invalidate();

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

@@ -21,6 +21,10 @@
 #include <string_view>
 #include <vector>
 
+namespace Xbyak {
+class CodeGenerator;
+}
+
 namespace Libs::Graphics::ShaderRecompiler::IR {
 
 enum class ResourceKind {
@@ -573,12 +577,12 @@ struct ResourcePlan {
 		uint64_t           generation = 0;
 	};
 
-	ResourcePlan() = default;
+	ResourcePlan();
 	~ResourcePlan();
 
 	ResourcePlan(const ResourcePlan&)            = delete;
 	ResourcePlan& operator=(const ResourcePlan&) = delete;
-	ResourcePlan(ResourcePlan&&) noexcept         = default;
+	ResourcePlan(ResourcePlan&&) noexcept;
 	ResourcePlan& operator=(ResourcePlan&& other) noexcept;
 
 	ShaderType                    stage           = ShaderType::Unknown;
@@ -597,6 +601,7 @@ struct ResourcePlan {
 	bool                                resource_tracking_complete = false;
 	ShaderInfo                          info;
 	UniformFillPlan                     uniform_fill;
+	std::unique_ptr<Xbyak::CodeGenerator> srt_code;
 	// GPU-thread scratch for nested clean/EXEC memos, activity and material keys.
 	mutable std::deque<EvaluationContext> evaluation_contexts;
 	mutable uint32_t                       evaluation_value_count = 0;

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

@@ -317,8 +317,8 @@ VALUE_OPCODE(BvhIntersect, U32x4, U32x4, ImageAddress, U1)
 VALUE_OPCODE(ImageQueryLod, U32x4, ImageResource, SamplerResource, ImageAddress)
 VALUE_OPCODE(ImageRead, U32x4, ImageResource, ImageAddress, U1)
 VALUE_OPCODE(ImageWrite, Void, ImageResource, ImageAddress, U32x4, U1)
-VALUE_OPCODE(ImageSampleRaw, U32x4, ImageResource, SamplerResource, ImageAddress)
-VALUE_OPCODE(ImageGatherRaw, U32x4, ImageResource, SamplerResource, ImageAddress, U32, U32, U32, U32)
+VALUE_OPCODE(ImageSampleRaw, U32x4, ImageResource, SamplerResource, ImageAddress, U1)
+VALUE_OPCODE(ImageGatherRaw, U32x4, ImageResource, SamplerResource, ImageAddress, U32, U32, U32, U32, U1)
 VALUE_OPCODE(ImageAtomicCompareSwap32, U32, ImageResource, ImageAddress, U32, U32, U1)
 VALUE_OPCODE(ImageAtomicSwap32, U32, ImageResource, ImageAddress, U32, U1)
 VALUE_OPCODE(ImageAtomicSwap64, U64, ImageResource, ImageAddress, U64, U1)

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

@@ -989,23 +989,29 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 	// Keep only that Boolean structure, never the varying shader dependency graph.
 	Value unknown;
 	bool capture_control_reads = false;
+	std::unordered_map<const Inst*, Value> predicates;
 	std::function<Value(Value)> ClonePredicate = [&](Value value) -> Value {
 		value = value.Resolve();
 		if (value.IsEmpty()) return {};
+		const auto* inst = value.TryInstruction();
+		if (inst == nullptr)
+			return ValidateRuntimeValue(program, value, RuntimeValueType::Integer) ? value : Value {};
+		// Preserve shared EXEC predicates, including subgraphs with no uniform factor.
+		const auto [found, inserted] = predicates.try_emplace(inst);
+		if (!inserted) return found->second;
+		auto& result = found->second;
 		if (ValidateRuntimeValue(program, value, RuntimeValueType::Integer)) {
 			capture_control_reads |=
 			    !ValidateRuntimeValue(program, value, RuntimeValueType::ImmutableInteger);
-			return Clone(value);
+			return result = Clone(value);
 		}
-		const auto* inst = value.TryInstruction();
-		if (inst == nullptr) return {};
 		const auto op = inst->GetOpcode();
 		if (op == ValueOpcode::ConditionRef || op == ValueOpcode::LogicalNot) {
 			const auto operand = ClonePredicate(inst->Arg(0));
-			if (operand.IsEmpty() || op == ValueOpcode::ConditionRef) return operand;
+			if (operand.IsEmpty() || op == ValueOpcode::ConditionRef) return result = operand;
 			auto& node = plan.value_storage.emplace_back(op);
 			node.SetArg(0, operand);
-			return Value(&node);
+			return result = Value(&node);
 		}
 		if (op != ValueOpcode::LogicalAnd && op != ValueOpcode::LogicalOr) return {};
 		auto left = ClonePredicate(inst->Arg(0));
@@ -1019,7 +1025,7 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 		auto& node = plan.value_storage.emplace_back(op);
 		node.SetArg(0, left);
 		node.SetArg(1, right);
-		return Value(&node);
+		return result = Value(&node);
 	};
 	plan.control_flow = ResourceControlFlow(program, ClonePredicate);
 	plan.uniform_fill = AnalyzeUniformFill(program);
@@ -1064,6 +1070,7 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 	}
 	if (capture_indirect_reads || !plan.source_reads.empty())
 		plan.resource_tracking_complete &= !program.has_address_writes;
+	CompileSrtWalker(plan);
 	return plan;
 }
 
@@ -1405,6 +1412,7 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
 			}
 			EXIT_IF(image.indirect_root == memory.resource &&
 			        inst.GetOpcode() != ValueOpcode::ImageSampleRaw &&
+			        inst.GetOpcode() != ValueOpcode::ImageGatherRaw &&
 			        inst.GetOpcode() != ValueOpcode::ImageRead);
 		}
 	}

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

@@ -1,12 +1,17 @@
 #include "graphics/shader/recompiler/ir/passes/SrtWalker.h"
 
+#include "common/abi.h"
 #include "common/assert.h"
 #include "graphics/shader/recompiler/ir/ShaderIR.h"
 
+#include <xbyak/xbyak.h>
+
 #include <algorithm>
 #include <bit>
 #include <cmath>
 #include <cstring>
+#include <map>
+#include <unordered_map>
 #include <unordered_set>
 
 namespace Libs::Graphics::ShaderRecompiler::IR {
@@ -359,6 +364,131 @@ private:
 
 } // namespace
 
+void CompileSrtWalker(ResourcePlan& program) {
+	if (!program.resource_tracking_complete || !program.srt_plan_complete ||
+	    program.capture_specialization_reads || !program.control_flow.empty() ||
+	    program.srt_reads.empty()) return;
+
+	using Reads = std::map<int32_t, const SrtRead*>;
+	std::map<uint32_t, Reads> roots;
+	std::unordered_map<const Inst*, Reads> pointer_uses;
+	std::unordered_map<const Inst*, int32_t> offsets;
+	for (const auto& read: program.srt_reads) {
+		const auto* inst = read.value.Resolve().TryInstruction();
+		if (inst == nullptr || inst->GetOpcode() != ValueOpcode::LoadAddressU32 ||
+		    inst->NumArgs() < 2 || !IsRawRead(program, *inst) ||
+		    read.flat_offset >= program.srt_reads.size()) return;
+		const auto offset = inst->Arg(1).Resolve();
+		if (!offset.IsImmediate() || offset.GetType() != Type::U32) return;
+		const auto& memory = program.memory_info[inst->Flags<SrtReadFlags>().index];
+		const auto relative =
+		    (static_cast<int64_t>(static_cast<int32_t>(memory.offset)) & ~int64_t {3}) +
+		    static_cast<int64_t>(offset.U32() & ~3u);
+		if (relative < INT32_MIN || relative > INT32_MAX ||
+		    !offsets.emplace(inst, static_cast<int32_t>(relative)).second) return;
+	}
+	uint32_t user_data_count = 0;
+	for (const auto& read: program.srt_reads) {
+		const auto* inst = read.value.ResolveInstruction();
+		const auto* handle = inst->Arg(0).Resolve().TryInstruction();
+		if (handle == nullptr || handle->GetOpcode() != ValueOpcode::GetAddressResource ||
+		    handle->NumArgs() != 2) return;
+		const auto* low = handle->Arg(0).Resolve().TryInstruction();
+		const auto* high = handle->Arg(1).Resolve().TryInstruction();
+		if (low == nullptr || high == nullptr) return;
+		Reads* uses = nullptr;
+		if (low->GetOpcode() == ValueOpcode::GetUserData &&
+		    high->GetOpcode() == ValueOpcode::GetUserData && low->NumArgs() == 1 &&
+		    high->NumArgs() == 1 && low->Arg(0).GetType() == Type::ScalarReg &&
+		    high->Arg(0).GetType() == Type::ScalarReg) {
+			const auto reg = RegIndex(low->Arg(0).ScalarRegister());
+			if (reg < program.user_data_base ||
+			    RegIndex(high->Arg(0).ScalarRegister()) != reg + 1u) return;
+			const auto base = reg - program.user_data_base;
+			user_data_count = std::max(user_data_count, base + 2u);
+			uses = &roots[base];
+		} else {
+			const auto lo = offsets.find(low);
+			const auto hi = offsets.find(high);
+			if (lo == offsets.end() || hi == offsets.end() ||
+			    static_cast<int64_t>(lo->second) + 4 != hi->second ||
+			    !EquivalentValue(program, low->Arg(0), high->Arg(0))) return;
+			uses = &pointer_uses[low];
+		}
+		if (!uses->emplace(offsets.at(inst), &read).second) return;
+	}
+	std::unordered_set<const Inst*> visited;
+	const auto reachable = [&](auto&& self, const Reads& reads) -> bool {
+		for (const auto& [offset, read]: reads) {
+			const auto* inst = read->value.ResolveInstruction();
+			const auto child = pointer_uses.find(inst);
+			if (child != pointer_uses.end() &&
+			    (!visited.insert(inst).second || !self(self, child->second))) return false;
+		}
+		return true;
+	};
+	for (const auto& [reg, reads]: roots) {
+		if (!reachable(reachable, reads)) return;
+	}
+	if (visited.size() != pointer_uses.size()) return;
+
+	program.srt_code = std::make_unique<Xbyak::CodeGenerator>(4096, Xbyak::AutoGrow);
+	auto& code = *program.srt_code;
+	Xbyak::Label failed, finished;
+	code.push(code.rbp);
+	code.mov(code.rbp, code.rsp);
+	code.cmp(code.r8, user_data_count);
+	code.jb(failed, Xbyak::CodeGenerator::T_NEAR);
+	code.mov(code.r9, AddressMask & ~uint64_t {3});
+	// Follow shadPS4's VisitPointer: copy one table level in offset order, then
+	// descend through its child pointers. Seed the descriptor evaluator's memo.
+	const auto visit = [&](auto&& self, int32_t pointer_offset, const Reads& reads) -> void {
+		code.push(code.rdi);
+		code.mov(code.rdi, code.qword[code.rdi + pointer_offset]);
+		code.and_(code.rdi, code.r9);
+		const auto first = reads.begin()->first;
+		const auto last = reads.rbegin()->first;
+		if (first < 0) {
+			code.mov(code.r10, -static_cast<int64_t>(first));
+			code.cmp(code.rdi, code.r10);
+			code.jb(failed, Xbyak::CodeGenerator::T_NEAR);
+		}
+		if (last > 0) {
+			code.mov(code.r10, AddressMask - last);
+			code.cmp(code.rdi, code.r10);
+			code.ja(failed, Xbyak::CodeGenerator::T_NEAR);
+		}
+		for (const auto& [offset, read]: reads) {
+			const auto index = read->value.ResolveInstruction()->EvaluationIndex(program.evaluation_value_count);
+			using Entry = ResourcePlan::EvaluationContext::Entry;
+			const auto slot = code.rdx + index * sizeof(Entry);
+			code.mov(code.r10d, code.dword[code.rdi + offset]);
+			code.mov(code.dword[code.rsi + read->flat_offset * sizeof(uint32_t)], code.r10d);
+			code.mov(code.qword[slot + offsetof(Entry, value)], code.r10);
+			code.mov(code.qword[slot + offsetof(Entry, generation)], code.rcx);
+		}
+		for (const auto& [offset, read]: reads) {
+			const auto child = pointer_uses.find(read->value.ResolveInstruction());
+			if (child == pointer_uses.end()) continue;
+			self(self, offset, child->second);
+		}
+		code.pop(code.rdi);
+	};
+	for (const auto& [reg, reads]: roots) {
+		visit(visit, reg * sizeof(uint32_t), reads);
+	}
+	code.mov(code.eax, 1);
+	code.jmp(finished, Xbyak::CodeGenerator::T_NEAR);
+	code.L(failed);
+	code.xor_(code.eax, code.eax);
+	code.L(finished);
+	code.mov(code.rsp, code.rbp);
+	code.pop(code.rbp);
+	code.ret();
+	code.readyRE();
+	EXIT_IF(Xbyak::GetError() != 0);
+}
+
 SrtWalker::SrtWalker(const ResourcePlan& program, const SrtRuntime& runtime,
                      SrtWalker* clean_evaluator, Value active_mask)
     : m_program(program), m_runtime(runtime), m_clean_evaluator(clean_evaluator),
@@ -954,6 +1084,15 @@ bool SrtWalker::EvaluateDescriptor(uint32_t source, DescriptorValue& result) {
 
 bool SrtWalker::RefreshFlatBuffer(std::vector<uint32_t>& flat) {
 	if (!m_program.srt_plan_complete) return false;
+	if (m_program.srt_code && m_runtime.read_memory == nullptr) {
+		flat.resize(m_program.srt_reads.size());
+		m_program.active_sources.clear();
+		m_context.values.resize(m_program.evaluation_value_count);
+		using Walker = bool KYTY_SYSV_ABI (*)(const uint32_t*, uint32_t*,
+		    ResourcePlan::EvaluationContext::Entry*, uint64_t, size_t);
+		return m_program.srt_code->getCode<Walker>()(m_runtime.user_data.data(), flat.data(),
+		    m_context.values.data(), m_context.generation, m_runtime.user_data.size());
+	}
 	const auto refresh = [&](uint32_t slot) {
 		if (slot >= m_program.srt_reads.size()) return false;
 		const auto& read = m_program.srt_reads[slot];

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

@@ -45,6 +45,7 @@ enum class RuntimeValueType { Any, Integer, ImmutableInteger };
 
 bool ValidateRuntimeValue(const ResourcePlan& program, Value value,
                           RuntimeValueType type = RuntimeValueType::Any);
+void CompileSrtWalker(ResourcePlan& program);
 // Uses the strict reader for values that affect shader specialization.
 SrtRuntime CleanRuntime(SrtRuntime runtime);
 

src/launcher/forms/configuration_edit_dialog.ui

@@ -180,6 +180,12 @@
               <property name="text"><string>Accurate texture synchronization</string></property>
              </widget>
             </item>
+            <item row="8" column="1">
+             <widget class="QCheckBox" name="checkBox_hdr">
+              <property name="toolTip"><string>Enable HDR10 output for supported games. Requires HDR enabled on the host display.</string></property>
+              <property name="text"><string>Enable HDR display support</string></property>
+             </widget>
+            </item>
            </layout>
           </widget>
          </item>
@@ -323,6 +329,23 @@
        </item>
       </layout>
      </widget>
+     <widget class="QWidget" name="network_tab">
+      <attribute name="title"><string>Network</string></attribute>
+      <layout class="QVBoxLayout" name="networkTabLayout">
+       <item>
+        <widget class="QCheckBox" name="checkBox_network_connected">
+         <property name="toolTip"><string>Let games use your computer's local network connection.</string></property>
+         <property name="text"><string>Enable network connection</string></property>
+        </widget>
+       </item>
+       <item>
+        <spacer name="networkSpacer">
+         <property name="orientation"><enum>Qt::Vertical</enum></property>
+         <property name="sizeHint" stdset="0"><size><width>20</width><height>40</height></size></property>
+        </spacer>
+       </item>
+      </layout>
+     </widget>
      <widget class="QWidget" name="debug_tab">
       <attribute name="title"><string>Debug</string></attribute>
       <layout class="QVBoxLayout" name="debugTabLayout">

src/launcher/include/configuration.h

@@ -117,12 +117,14 @@ public:
 	PresentMode            present_mode                = PresentMode::Mailbox;
 	int                    gpu_index                   = -1;
 	bool                   fullscreen_enabled          = false;
+	bool                   hdr_enabled                 = false;
 	bool                   hide_cursor_enabled         = false;
 	bool                   readback_linear_images      = false;
 	bool                   sync_raw_image_buffers      = false;
 	bool                   tessellation_enabled        = false;
 	bool                   trophy_enabled              = true;
 	bool                   skip_notice_screen          = false;
+	bool                   network_connected           = false;
 	int                    vblank_frequency            = 60;
 	int                    console_language            = DEFAULT_CONSOLE_LANGUAGE;
 	bool                   vulkan_validation_enabled   = false;
@@ -152,12 +154,14 @@ public:
 		present_mode                = other.present_mode;
 		gpu_index                   = other.gpu_index;
 		fullscreen_enabled          = other.fullscreen_enabled;
+		hdr_enabled                 = other.hdr_enabled;
 		hide_cursor_enabled         = other.hide_cursor_enabled;
 		readback_linear_images      = other.readback_linear_images;
 		sync_raw_image_buffers      = other.sync_raw_image_buffers;
 		tessellation_enabled        = other.tessellation_enabled;
 		trophy_enabled              = other.trophy_enabled;
 		skip_notice_screen          = other.skip_notice_screen;
+		network_connected           = other.network_connected;
 		vblank_frequency            = other.vblank_frequency;
 		console_language            = other.console_language;
 		vulkan_validation_enabled   = other.vulkan_validation_enabled;

src/launcher/src/configuration.cpp

@@ -10,12 +10,14 @@ QVariantMap Configuration::GameSettings() const {
 	KYTY_CFG_SET(present_mode);
 	KYTY_CFG_SET(gpu_index);
 	KYTY_CFG_SET(fullscreen_enabled);
+	KYTY_CFG_SET(hdr_enabled);
 	KYTY_CFG_SET(hide_cursor_enabled);
 	KYTY_CFG_SET(readback_linear_images);
 	KYTY_CFG_SET(sync_raw_image_buffers);
 	KYTY_CFG_SET(tessellation_enabled);
 	KYTY_CFG_SET(trophy_enabled);
 	KYTY_CFG_SET(skip_notice_screen);
+	KYTY_CFG_SET(network_connected);
 	KYTY_CFG_SET(vblank_frequency);
 	KYTY_CFG_SET(console_language);
 	KYTY_CFG_SET(vulkan_validation_enabled);
@@ -55,12 +57,14 @@ void Configuration::ReadGameSettingsValues(const Settings& s) {
 		present_mode = PresentMode::Mailbox;
 	}
 	KYTY_CFG_GET(fullscreen_enabled);
+	KYTY_CFG_GET(hdr_enabled);
 	KYTY_CFG_GET(hide_cursor_enabled);
 	KYTY_CFG_GET(readback_linear_images);
 	KYTY_CFG_GET(sync_raw_image_buffers);
 	KYTY_CFG_GET(tessellation_enabled);
 	trophy_enabled   = s.value("trophy_enabled", trophy_enabled).toBool();
 	KYTY_CFG_GET(skip_notice_screen);
+	KYTY_CFG_GET(network_connected);
 	vblank_frequency = s.value("vblank_frequency", vblank_frequency).toInt();
 	console_language = s.value("console_language", console_language).toInt();
 	if (console_language < 0 || console_language > MAX_CONSOLE_LANGUAGE) {

src/launcher/src/configurationEditDialog.cpp

@@ -273,11 +273,13 @@ void ConfigurationEditDialog::Init(const Configuration& info) {
 	    info.gpu_index >= 0 && info.gpu_index < m_ui->comboBox_gpu->count() - 1 ? info.gpu_index + 1
 	                                                                            : 0);
 	m_ui->checkBox_fullscreen->setChecked(info.fullscreen_enabled);
+	m_ui->checkBox_hdr->setChecked(info.hdr_enabled);
 	m_ui->checkBox_hide_cursor->setChecked(info.hide_cursor_enabled);
 	m_ui->checkBox_readback->setChecked(info.readback_linear_images);
 	m_ui->checkBox_sync_raw_image_buffers->setChecked(info.sync_raw_image_buffers);
 	m_ui->checkBox_tessellation->setChecked(info.tessellation_enabled);
 	m_ui->checkBox_skip_notice_screen->setChecked(info.skip_notice_screen);
+	m_ui->checkBox_network_connected->setChecked(info.network_connected);
 	m_ui->checkBox_trophy_notifications->setChecked(info.trophy_enabled);
 	m_ui->spinBox_vblank_frequency->setValue(info.vblank_frequency);
 	m_ui->comboBox_console_language->clear();
@@ -434,11 +436,13 @@ static void UpdateInfo(Configuration& info, Ui::ConfigurationEditDialog& ui, boo
 	    TextToEnum<Configuration::PresentMode>(ui.comboBox_present_mode->currentText());
 	info.gpu_index                 = ui.comboBox_gpu->currentIndex() - 1;
 	info.fullscreen_enabled        = ui.checkBox_fullscreen->isChecked();
+	info.hdr_enabled               = ui.checkBox_hdr->isChecked();
 	info.hide_cursor_enabled       = ui.checkBox_hide_cursor->isChecked();
 	info.readback_linear_images    = ui.checkBox_readback->isChecked();
 	info.sync_raw_image_buffers    = ui.checkBox_sync_raw_image_buffers->isChecked();
 	info.tessellation_enabled      = ui.checkBox_tessellation->isChecked();
 	info.skip_notice_screen        = ui.checkBox_skip_notice_screen->isChecked();
+	info.network_connected         = ui.checkBox_network_connected->isChecked();
 	info.trophy_enabled            = ui.checkBox_trophy_notifications->isChecked();
 	info.vblank_frequency          = ui.spinBox_vblank_frequency->value();
 	info.console_language          = ui.comboBox_console_language->currentIndex();

src/launcher/src/mainDialog.cpp

@@ -245,6 +245,9 @@ static QStringList CreateEmulatorArgs(const Configuration& info) {
 	if (info.fullscreen_enabled) {
 		args << "--fullscreen";
 	}
+	if (info.hdr_enabled) {
+		args << "--hdr";
+	}
 	if (info.hide_cursor_enabled) {
 		args << "--hide-cursor";
 	}
@@ -252,6 +255,7 @@ static QStringList CreateEmulatorArgs(const Configuration& info) {
 	args << "--sync-raw-image-buffers" << BoolArg(info.sync_raw_image_buffers);
 	args << "--trophy-notifications" << BoolArg(info.trophy_enabled);
 	args << "--skip-notice-screen" << BoolArg(info.skip_notice_screen);
+	args << "--network-connected" << BoolArg(info.network_connected);
 	if (info.tessellation_enabled) {
 		args << "--tessellation";
 	}

src/libs/agc.cpp

@@ -259,29 +259,6 @@ struct CommandBuffer {
 	void*     user_data;
 	uint32_t  reserved_dw;
 
-	void DbgDump() const {
-		if (!Config::GraphicsDebugDumpEnabled() ||
-		    Config::GetPrintfDirection() == Config::LogDirection::Silent) {
-			return;
-		}
-		static std::atomic<uint32_t> log_count {0};
-		if (log_count.fetch_add(1) >= 64) {
-			return;
-		}
-
-		LOGF("\t bottom      = 0x%016" PRIx64 "\n"
-		     "\t top         = 0x%016" PRIx64 "\n"
-		     "\t cursor_up   = 0x%016" PRIx64 "\n"
-		     "\t cursor_down = 0x%016" PRIx64 "\n"
-		     "\t callback    = 0x%016" PRIx64 "\n"
-		     "\t user_data   = 0x%016" PRIx64 "\n"
-		     "\t reserved_dw = %" PRIu32 "\n",
-		     reinterpret_cast<uint64_t>(bottom), reinterpret_cast<uint64_t>(top),
-		     reinterpret_cast<uint64_t>(cursor_up), reinterpret_cast<uint64_t>(cursor_down),
-		     reinterpret_cast<uint64_t>(callback), reinterpret_cast<uint64_t>(user_data),
-		     reserved_dw);
-	}
-
 	[[nodiscard]] KYTY_SYSV_ABI uint64_t GetAvailableSizeDW() const {
 		// Interpret the signed cursor distance as an unsigned 64-bit DWORD count.
 		const auto distance = static_cast<int64_t>(reinterpret_cast<uintptr_t>(cursor_down) -
@@ -302,7 +279,6 @@ struct CommandBuffer {
 				    "\t command buffer exhausted and has no grow callback: requested = %" PRIu32
 				    ", remaining = %" PRIu64 ", reserved_dw = %" PRIu32 "\n",
 				    num_dw, remaining, reserved_dw);
-				DbgDump();
 				return false;
 			}
 
@@ -312,7 +288,6 @@ struct CommandBuffer {
 				           "\t command buffer grow callback failed: requested = %" PRIu32
 				           ", remaining = %" PRIu64 ", reserved_dw = %" PRIu32 "\n",
 				           num_dw, remaining, reserved_dw);
-				DbgDump();
 				return false;
 			}
 			if (GetAvailableSizeDW() < num_dw) {
@@ -321,7 +296,6 @@ struct CommandBuffer {
 				           "requested = %" PRIu32 ", remaining = %" PRIu64
 				           ", reserved_dw = %" PRIu32 "\n",
 				           num_dw, GetAvailableSizeDW(), reserved_dw);
-				DbgDump();
 				return false;
 			}
 		}
@@ -1598,8 +1572,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbContextStateOp(CommandBuffer* buf, uint32_t operat
 		return nullptr;
 	}
 
-	buf->DbgDump();
-
 	uint32_t*  first  = nullptr;
 	const auto append = [&](uint32_t segment_size_dw) {
 		auto* segment = buf->AllocateDW(segment_size_dw);
@@ -1902,8 +1874,6 @@ uint32_t* KYTY_SYSV_ABI AgcCbSetShRegisterRangeDirect(CommandBuffer* buf, uint32
 
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(num_values + 2);
 
 	if (cmd == nullptr) {
@@ -1945,8 +1915,6 @@ uint32_t* KYTY_SYSV_ABI AgcCbSetShRegistersDirect(CommandBuffer*
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 	EXIT_NOT_IMPLEMENTED(regs == nullptr);
 
-	buf->DbgDump();
-
 	std::vector<ShaderRegister> local_regs(num_regs);
 	for (uint32_t i = 0; i < num_regs; i++) {
 		local_regs[i].offset = regs[i].offset;
@@ -2012,8 +1980,6 @@ uint32_t* KYTY_SYSV_ABI AgcCbSetUcRegistersDirect(CommandBuffer*
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 	EXIT_NOT_IMPLEMENTED(regs == nullptr);
 
-	buf->DbgDump();
-
 	// Original implementation stages register values in a temporary DWORD array. It reads each new run only
 	// after the preceding run has been allocated, so preserve that ordering around grow callbacks.
 	std::vector<uint32_t> values(num_regs);
@@ -2104,8 +2070,6 @@ uint32_t* KYTY_SYSV_ABI AgcCbReleaseMem(CommandBuffer* buf, uint8_t action, uint
 	                     data_sel != 5);
 	EXIT_NOT_IMPLEMENTED(interrupt > 4);
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(8);
 
 	if (cmd == nullptr) {
@@ -2158,8 +2122,6 @@ uint32_t* KYTY_SYSV_ABI AgcAcbResetQueue(CommandBuffer* buf, uint32_t op) {
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 	EXIT_NOT_IMPLEMENTED((op & ~0x1c2u) != 0);
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(2);
 
 	EXIT_NOT_IMPLEMENTED(cmd == nullptr);
@@ -2180,8 +2142,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbResetQueue(CommandBuffer* buf, uint32_t op, uint32
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 	EXIT_NOT_IMPLEMENTED((op & ~0xfffu) != 0);
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(2);
 
 	EXIT_NOT_IMPLEMENTED(cmd == nullptr);
@@ -2203,8 +2163,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbWaitUntilSafeForRendering(CommandBuffer* buf,
 
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(7);
 
 	EXIT_NOT_IMPLEMENTED(cmd == nullptr);
@@ -2257,8 +2215,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbSetWorkloadsActive(CommandBuffer* buf, uint32_t st
 		}
 	}
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(WORKLOAD_ACTIVE_PACKET_SIZE_DW);
 	if (cmd == nullptr) {
 		return nullptr;
@@ -2294,8 +2250,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbSetWorkloadComplete(CommandBuffer* buf, uint32_t s
 		}
 	}
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(WORKLOAD_COMPLETE_PACKET_SIZE_DW);
 	if (cmd == nullptr) {
 		return nullptr;
@@ -2319,8 +2273,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbSetShRegisterDirect(CommandBuffer* buf, ShaderRegi
 
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(3);
 
 	if (cmd == nullptr) {
@@ -2340,8 +2292,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbSetCxRegisterDirect(CommandBuffer* buf, ShaderRegi
 
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(3);
 
 	if (cmd == nullptr) {
@@ -2367,8 +2317,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbSetUcRegisterDirect(CommandBuffer* buf, ShaderRegi
 
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(3);
 
 	if (cmd == nullptr) {
@@ -2394,8 +2342,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbSetCxRegistersIndirect(CommandBuffer*
 
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(RegIndirectPacketSizeDw);
 
 	EXIT_NOT_IMPLEMENTED(cmd == nullptr);
@@ -2423,8 +2369,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbSetShRegistersIndirect(CommandBuffer*
 
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(RegIndirectPacketSizeDw);
 
 	EXIT_NOT_IMPLEMENTED(cmd == nullptr);
@@ -2453,8 +2397,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbSetUcRegistersIndirect(CommandBuffer*
 
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(RegIndirectPacketSizeDw);
 
 	EXIT_NOT_IMPLEMENTED(cmd == nullptr);
@@ -2475,8 +2417,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbSetIndexSize(CommandBuffer* buf, uint8_t index_siz
 	     index_size, cache_policy);
 
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(3);
 
 	EXIT_NOT_IMPLEMENTED(cmd == nullptr);
@@ -2496,8 +2436,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbSetIndexBuffer(CommandBuffer* buf, uint64_t index_
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 	EXIT_NOT_IMPLEMENTED((index_addr & 1u) != 0);
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(3);
 
 	EXIT_NOT_IMPLEMENTED(cmd == nullptr);
@@ -2516,8 +2454,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbSetIndexCount(CommandBuffer* buf, uint32_t index_c
 
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(2);
 
 	EXIT_NOT_IMPLEMENTED(cmd == nullptr);
@@ -2537,8 +2473,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbSetNumInstances(CommandBuffer* buf, uint32_t num_i
 		return nullptr;
 	}
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(2);
 
 	if (cmd == nullptr) {
@@ -2635,8 +2569,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbDrawIndex(CommandBuffer* buf, uint32_t index_count
 	EXIT_NOT_IMPLEMENTED(index_addr == nullptr);
 	EXIT_NOT_IMPLEMENTED((reinterpret_cast<uint64_t>(index_addr) & 1u) != 0);
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(6);
 
 	EXIT_NOT_IMPLEMENTED(cmd == nullptr);
@@ -2678,8 +2610,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbDrawIndexMultiInstanced(CommandBuffer* buf, uint32
 	EXIT_NOT_IMPLEMENTED(object_ids == nullptr);
 	EXIT_NOT_IMPLEMENTED((reinterpret_cast<uint64_t>(index_addr) & 1u) != 0);
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(9);
 
 	EXIT_NOT_IMPLEMENTED(cmd == nullptr);
@@ -2747,8 +2677,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbDrawIndexAuto(CommandBuffer* buf, uint32_t index_c
 
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(3);
 
 	EXIT_NOT_IMPLEMENTED(cmd == nullptr);
@@ -2777,8 +2705,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbDrawIndexOffset(CommandBuffer* buf, uint32_t index
 
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(5);
 
 	EXIT_NOT_IMPLEMENTED(cmd == nullptr);
@@ -2808,8 +2734,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbSetBaseIndirectArgs(CommandBuffer* buf, uint32_t s
 
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(4);
 
 	EXIT_NOT_IMPLEMENTED(cmd == nullptr);
@@ -2834,8 +2758,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbDrawIndexIndirect(CommandBuffer* buf, uint32_t dat
 
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(5);
 
 	EXIT_NOT_IMPLEMENTED(cmd == nullptr);
@@ -2863,8 +2785,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbDrawIndirect(CommandBuffer* buf, uint32_t data_off
 
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(5);
 
 	EXIT_NOT_IMPLEMENTED(cmd == nullptr);
@@ -2901,8 +2821,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbDrawIndirectMulti(CommandBuffer*       buf,
 		return nullptr;
 	}
 
-	buf->DbgDump();
-
 	constexpr uint32_t draw_packet_size_dw = 10u;
 	constexpr uint32_t total_size_dw =
 	    AGC_INTERNAL_DATA_PACKET_SIZE_DW + draw_packet_size_dw +
@@ -2963,8 +2881,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbDrawIndexIndirectMulti(CommandBuffer*       buf,
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 	EXIT_NOT_IMPLEMENTED((count_indirect & ~1u) != 0);
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(10);
 
 	EXIT_NOT_IMPLEMENTED(cmd == nullptr);
@@ -3013,8 +2929,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbDispatchIndirect(CommandBuffer* buf, uint32_t data
 
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(3);
 
 	EXIT_NOT_IMPLEMENTED(cmd == nullptr);
@@ -3046,8 +2960,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbEventWrite(CommandBuffer* buf, uint8_t event_type,
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 	EXIT_NOT_IMPLEMENTED(event_type > 0x3fu);
 
-	buf->DbgDump();
-
 	const bool addressed_event = ((event_type & 0xfeu) == 0x38u);
 	auto*      cmd             = buf->AllocateDW(addressed_event ? 4u : 2u);
 
@@ -3143,8 +3055,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbAcquireMem(CommandBuffer* buf, uint8_t engine, uin
 		}
 	}
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(8);
 
 	if (cmd == nullptr) {
@@ -3203,8 +3113,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbCondExec(CommandBuffer* buf, const volatile uint32
 		return nullptr;
 	}
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(5);
 
 	if (cmd == nullptr) {
@@ -3376,8 +3284,6 @@ uint32_t* KYTY_SYSV_ABI AgcAcbDispatchIndirect(CommandBuffer*       buf,
 		return nullptr;
 	}
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(4);
 
 	if (cmd == nullptr) {
@@ -3426,8 +3332,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbStallCommandBufferParser(CommandBuffer* buf) {
 
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(2);
 
 	EXIT_NOT_IMPLEMENTED(cmd == nullptr);
@@ -3714,6 +3618,20 @@ int KYTY_SYSV_ABI AgcSetPacketPredication(uint32_t* packet, uint32_t predication
 	return OK;
 }
 
+int KYTY_SYSV_ABI AgcSetNop(uint32_t* packet) {
+	PRINT_NAME();
+
+	const auto size = AgcGetPacketSize(packet);
+
+	packet[0] = (packet[0] & 0xffff0003u) | (Pm4::IT_NOP << 8u);
+	// Prevent the disabled payload from being treated as an executable marker.
+	if (size > 1 && (packet[1] & 0xffff0000u) == 0x68750000u) {
+		packet[1] &= 0xffffu;
+	}
+
+	return OK;
+}
+
 int KYTY_SYSV_ABI AgcSetRangePredication(uint32_t* start, const volatile uint32_t* end,
                                          uint32_t predication) {
 	PRINT_NAME();
@@ -3833,8 +3751,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbWriteData(CommandBuffer* buf, uint8_t dst, uint8_t
 		return nullptr;
 	}
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(4 + num_dwords);
 
 	if (cmd == nullptr) {
@@ -3887,8 +3803,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbGetLodStats(CommandBuffer* buf, uint8_t cache_poli
 		return nullptr;
 	}
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(5);
 
 	if (cmd == nullptr) {
@@ -4091,8 +4005,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbWaitRegMem(CommandBuffer* buf, uint8_t size, uint8
 		           cache_policy);
 	}
 
-	buf->DbgDump();
-
 	auto address_value = reinterpret_cast<uint64_t>(address);
 	bool wait32        = (size == 0);
 	auto poll          = wait_reg_mem_poll_cycles_to_packet(poll_cycles);
@@ -4222,8 +4134,6 @@ uint32_t* KYTY_SYSV_ABI AgcDcbSetFlip(CommandBuffer* buf, uint32_t video_out_han
 
 	EXIT_NOT_IMPLEMENTED(buf == nullptr);
 
-	buf->DbgDump();
-
 	auto* cmd = buf->AllocateDW(6);
 
 	EXIT_NOT_IMPLEMENTED(cmd == nullptr);

src/libs/agc.h

@@ -270,6 +270,7 @@ uint32_t KYTY_SYSV_ABI  AgcGetPacketSize(uint32_t* packet);
 int KYTY_SYSV_ABI       AgcSetPacketPredication(uint32_t* packet, uint32_t predication);
 int KYTY_SYSV_ABI       AgcSetRangePredication(uint32_t* start, const volatile uint32_t* end,
                                                uint32_t predication);
+int KYTY_SYSV_ABI       AgcSetNop(uint32_t* packet);
 int KYTY_SYSV_ABI       AgcRewindPatchSetRewindState(uint32_t* cmd, uint8_t state);
 int KYTY_SYSV_ABI       AgcCondExecPatchSetEnd(uint32_t* cmd, const volatile uint32_t* buffer);
 int KYTY_SYSV_ABI       AgcCondExecPatchSetCommandAddress(uint32_t*                cmd,

src/libs/libAgcDriver.cpp

@@ -143,6 +143,7 @@ LIB_DEFINE(InitAgcDriver_1) {
 	LIB_FUNC("Lkf86B98qPc", Gen5::AgcGetPacketSize);
 	LIB_FUNC("w6Dj1VJt5qY", Gen5::AgcSetPacketPredication);
 	LIB_FUNC("n8vgpaQg6dA", Gen5::AgcSetRangePredication);
+	LIB_FUNC("K2mciNVxUCE", Gen5::AgcSetNop);
 	LIB_FUNC("i1jyy49AjXU", Gen5::AgcDcbWriteData);
 	LIB_FUNC("p9tI+yTvx68", Gen5::AgcDcbWriteDataGetSize);
 	LIB_FUNC("vuSXe69VILM", Gen5::AgcDcbGetLodStats);

src/libs/libAmpr.cpp

@@ -868,6 +868,10 @@ static uint64_t AprReadFileRecordSize(uint64_t file_offset) {
 	return (file_offset >> 32u) != 0 ? READ_FILE_RECORD_SIZE_EXT : READ_FILE_RECORD_SIZE;
 }
 
+static uint64_t WriteAddressRecordSize(uint64_t value) {
+	return value <= 3 ? 8 : value < (uint64_t {1} << 34) ? 12 : 16;
+}
+
 static uint64_t AprReadGatherRecordSize(uint64_t file_offset) {
 	return file_offset > 0x3ffffu ? READ_GATHER_RECORD_SIZE_EXT : READ_GATHER_RECORD_SIZE;
 }
@@ -1617,11 +1621,10 @@ static uint64_t KYTY_SYSV_ABI MeasureCommandSizeFixed32(uint64_t, uint64_t, uint
 	return 0x20;
 }
 
-static uint64_t KYTY_SYSV_ABI MeasureCommandSizeWriteAddressOnCompletion(volatile uint64_t*,
-                                                                         uint64_t) {
+static uint64_t KYTY_SYSV_ABI MeasureCommandSizeWriteAddress(volatile uint64_t*, uint64_t value) {
 	PRINT_NAME();
 
-	return 0x20;
+	return WriteAddressRecordSize(value);
 }
 
 static uint64_t KYTY_SYSV_ABI
@@ -1950,7 +1953,7 @@ static int KYTY_SYSV_ABI CommandBufferWaitOnCounter(void* command_buffer, uint8_
 }
 
 static int AppendWriteAddressCommand(void* command_buffer, volatile uint64_t* address,
-                                     uint64_t value) {
+                                     uint64_t value, uint64_t record_size) {
 	if (command_buffer == nullptr || address == nullptr ||
 	    (reinterpret_cast<uint64_t>(address) & 7) != 0) {
 		return LibKernel::KERNEL_ERROR_EINVAL;
@@ -1958,7 +1961,7 @@ static int AppendWriteAddressCommand(void* command_buffer, volatile uint64_t* ad
 	return AppendCommandRecord(
 	           reinterpret_cast<uint64_t>(command_buffer),
 	           CommandBufferState::WriteAddressCommand {reinterpret_cast<uint64_t>(address), value},
-	           0x20)
+	           record_size)
 	           ? OK
 	           : LibKernel::KERNEL_ERROR_EBUSY;
 }
@@ -1967,7 +1970,7 @@ static int KYTY_SYSV_ABI CommandBufferWriteAddress(void* command_buffer, volatil
                                                    uint64_t value, uint32_t) {
 	PRINT_NAME();
 
-	return AppendWriteAddressCommand(command_buffer, address, value);
+	return AppendWriteAddressCommand(command_buffer, address, value, WriteAddressRecordSize(value));
 }
 
 static int KYTY_SYSV_ABI CommandBufferWriteAddressOnCompletion(void*              command_buffer,
@@ -1975,7 +1978,7 @@ static int KYTY_SYSV_ABI CommandBufferWriteAddressOnCompletion(void*
                                                                uint64_t           value) {
 	PRINT_NAME();
 
-	return AppendWriteAddressCommand(command_buffer, address, value);
+	return AppendWriteAddressCommand(command_buffer, address, value, WriteAddressRecordSize(value));
 }
 
 static int KYTY_SYSV_ABI CommandBufferWriteCounter(void* command_buffer, uint8_t index,
@@ -2007,14 +2010,14 @@ static int KYTY_SYSV_ABI CommandBufferWriteAddressFromTimeCounter(void*
                                                                   uint32_t) {
 	PRINT_NAME();
 
-	return AppendWriteAddressCommand(command_buffer, address, uint64_t {0});
+	return AppendWriteAddressCommand(command_buffer, address, uint64_t {0}, 0x20);
 }
 
 static int KYTY_SYSV_ABI CommandBufferWriteAddressFromTimeCounterOnCompletion(
     void* command_buffer, volatile uint64_t* address) {
 	PRINT_NAME();
 
-	return AppendWriteAddressCommand(command_buffer, address, uint64_t {0});
+	return AppendWriteAddressCommand(command_buffer, address, uint64_t {0}, 0x20);
 }
 
 static int KYTY_SYSV_ABI CommandBufferWriteAddressFromCounter(void*              command_buffer,
@@ -2022,14 +2025,14 @@ static int KYTY_SYSV_ABI CommandBufferWriteAddressFromCounter(void*
                                                               uint32_t) {
 	PRINT_NAME();
 
-	return AppendWriteAddressCommand(command_buffer, address, uint64_t {0});
+	return AppendWriteAddressCommand(command_buffer, address, uint64_t {0}, 0x20);
 }
 
 static int KYTY_SYSV_ABI CommandBufferWriteAddressFromCounterOnCompletion(
     void* command_buffer, volatile uint64_t* address, uint8_t) {
 	PRINT_NAME();
 
-	return AppendWriteAddressCommand(command_buffer, address, uint64_t {0});
+	return AppendWriteAddressCommand(command_buffer, address, uint64_t {0}, 0x20);
 }
 
 static int KYTY_SYSV_ABI CommandBufferWriteAddressFromCounterPair(void*              command_buffer,
@@ -2037,14 +2040,14 @@ static int KYTY_SYSV_ABI CommandBufferWriteAddressFromCounterPair(void*
                                                                   uint8_t, uint32_t) {
 	PRINT_NAME();
 
-	return AppendWriteAddressCommand(command_buffer, address, uint64_t {0});
+	return AppendWriteAddressCommand(command_buffer, address, uint64_t {0}, 0x20);
 }
 
 static int KYTY_SYSV_ABI CommandBufferWriteAddressFromCounterPairOnCompletion(
     void* command_buffer, volatile uint64_t* address, uint8_t) {
 	PRINT_NAME();
 
-	return AppendWriteAddressCommand(command_buffer, address, uint64_t {0});
+	return AppendWriteAddressCommand(command_buffer, address, uint64_t {0}, 0x20);
 }
 
 static int KYTY_SYSV_ABI AprCommandBufferReadFileGather(void*    command_buffer, uint64_t, uint64_t,
@@ -2452,14 +2455,14 @@ LIB_DEFINE(InitAmpr_1) {
 	LIB_FUNC("vWU-odnS+fU", Ampr::MeasureCommandSizeReadFile);
 	LIB_FUNC("sSAUCCU1dv4", Ampr::MeasureCommandSizeWriteKernelEventQueue);
 	LIB_FUNC("Zi3dBUjgyXI", Ampr::MeasureCommandSizeWriteKernelEventQueue);
-	LIB_FUNC("C+IEj+BsAFM", Ampr::MeasureCommandSizeWriteAddressOnCompletion);
+	LIB_FUNC("C+IEj+BsAFM", Ampr::MeasureCommandSizeWriteAddress);
 	LIB_FUNC("x7SQEXfeovg", Ampr::MeasureCommandSizeWriteAddressFromTimeCounterOnCompletion);
 	LIB_FUNC("32AcaTaBPSY", Ampr::MeasureCommandSizeWriteAddressFromCounterOnCompletion);
 	LIB_FUNC("vxC58+DRk-U", Ampr::MeasureCommandSizeWriteAddressFromCounterPairOnCompletion);
 	LIB_FUNC("4muPEJ-x5N8", Ampr::MeasureCommandSizeWriteCounterOnCompletion);
 	LIB_FUNC("0BMj1hgG+kE", Ampr::MeasureCommandSizeFixed32);
 	LIB_FUNC("ClnsFLLLcss", Ampr::MeasureCommandSizeFixed32);
-	LIB_FUNC("4fgtGfXDrFc", Ampr::MeasureCommandSizeFixed32);
+	LIB_FUNC("4fgtGfXDrFc", Ampr::MeasureCommandSizeWriteAddress);
 	LIB_FUNC("gAtc79UTt5E", Ampr::MeasureCommandSizeFixed32);
 	LIB_FUNC("JYd9g9L+TmE", Ampr::MeasureCommandSizeFixed32);
 	LIB_FUNC("2Hw8gjMdwSY", Ampr::MeasureCommandSizeFixed32);

src/libs/libC.cpp

@@ -54,11 +54,7 @@ struct CContext {
 	std::list<atexit_func_t> atexit;
 };
 
-struct InitEnvParams {
-	int         argc;
-	uint32_t    pad;
-	const char* argv[3];
-};
+using InitEnvParams = Loader::EntryParams;
 
 static int                g_argc = 0;
 static const char* const* g_argv = nullptr;
@@ -271,9 +267,7 @@ static KYTY_SYSV_ABI void init_env(const InitEnvParams* params) {
 		return;
 	}
 
-	constexpr int argv_capacity = static_cast<int>(sizeof(params->argv) / sizeof(params->argv[0]));
-
-	EXIT_NOT_IMPLEMENTED(params->argc < 0 || params->argc >= argv_capacity);
+	EXIT_NOT_IMPLEMENTED(params->argc < 0);
 
 	g_argc = params->argc;
 	g_argv = params->argv;

src/libs/libKernel.cpp

@@ -3117,6 +3117,7 @@ LIB_DEFINE(InitLibKernel_1_FS) {
 	LIB_FUNC("mBd4AfLP+u8", FileSystem::KernelPwritev);
 	LIB_FUNC("eV9wAD2riIA", FileSystem::KernelStat);
 	LIB_FUNC("kBwCPsYX-m4", FileSystem::KernelFstat);
+	LIB_FUNC("VW3TVZiM4-E", FileSystem::KernelFtruncate);
 	LIB_FUNC("AUXVxWeJU-A", FileSystem::KernelUnlink);
 	LIB_FUNC("52NcYU9+lEo", FileSystem::KernelRename);
 	LIB_FUNC("taRWhTJFTgE", FileSystem::KernelGetdirentries);

src/libs/libSystemService.cpp

@@ -4,6 +4,7 @@
 #include "common/emulatorConfig.h"
 #include "common/logging/log.h"
 #include "common/stringUtils.h"
+#include "emulator.h"
 #include "graphics/presentation/systemOverlay.h"
 #include "libs/dialog.h"
 #include "libs/errno.h"
@@ -99,6 +100,11 @@ static int KYTY_SYSV_ABI SystemServiceHideSplashScreen() {
 	return OK;
 }
 
+int KYTY_SYSV_ABI SystemServiceLoadExec(const char* path, const char* const argv[]) {
+	PRINT_NAME();
+	return Emulator::LoadExec(path, argv);
+}
+
 static int KYTY_SYSV_ABI SystemServiceParamGetInt(int param_id, int* value) {
 	PRINT_NAME();
 
@@ -559,6 +565,7 @@ LIB_DEFINE(InitSystemGesture_1) {
 } // namespace SystemGesture
 
 LIB_DEFINE(InitSystemService_1) {
+	LIB_FUNC("JoBqSQt1yyA", SystemService::SystemServiceLoadExec);
 	LIB_FUNC("Vo5V8KAwCmk", SystemService::SystemServiceHideSplashScreen);
 	LIB_FUNC("fZo48un7LK4", SystemService::SystemServiceParamGetInt);
 	LIB_FUNC("SsC-m-S9JTA", SystemService::SystemServiceParamGetString);

src/libs/network.cpp

@@ -15,6 +15,9 @@
 #else
 #include <arpa/inet.h>
 #include <fcntl.h>
+#include <ifaddrs.h>
+#include <net/if.h>
+#include <net/route.h>
 #include <netdb.h>
 #include <netinet/tcp.h>
 #include <sys/select.h>
@@ -41,10 +44,12 @@
 #include <cstring>
 #include <fmt/format.h>
 #include <fmt/ranges.h>
+#include <fstream>
 #include <limits>
 #include <memory>
 #include <mutex>
 #include <optional>
+#include <sstream>
 #include <string>
 #include <utility>
 #include <vector>
@@ -3386,7 +3391,6 @@ constexpr int NET_CTL_STATE_DISCONNECTED  = 0;
 constexpr int NET_CTL_STATE_IPOBTAINED    = 3;
 constexpr int NET_CTL_DEVICE_WIRED        = 0;
 constexpr int NET_CTL_DEVICE_WIRELESS     = 1;
-constexpr int NET_CTL_LINK_DISCONNECTED   = 0;
 constexpr int NET_CTL_LINK_CONNECTED      = 1;
 constexpr int NET_CTL_EVENT_DISCONNECTED  = 1;
 constexpr int NET_CTL_EVENT_IPOBTAINED    = 3;
@@ -3448,8 +3452,6 @@ struct NetCtlCallbackSlot {
 
 static Common::Mutex                     g_net_ctl_callbacks_mutex;
 static std::array<NetCtlCallbackSlot, 8> g_net_ctl_callbacks;
-static std::atomic_bool                  g_net_ctl_connected          = false;
-static std::atomic_bool                  g_net_ctl_status_initialized = false;
 
 struct HostNetworkInfo {
 	bool         connected = false;
@@ -3463,22 +3465,16 @@ struct HostNetworkInfo {
 };
 
 static void CopyIpv4String(char* dst, size_t dst_size, const sockaddr* addr) {
-#if defined(_WIN32)
 	if (dst == nullptr || dst_size == 0 || addr == nullptr || addr->sa_family != AF_INET) {
 		return;
 	}
 
 	const auto* in = reinterpret_cast<const sockaddr_in*>(addr);
 	inet_ntop(AF_INET, &in->sin_addr, dst, static_cast<socklen_t>(dst_size));
-#else
-	(void)dst;
-	(void)dst_size;
-	(void)addr;
-#endif
 }
 
-static void CopyIpv4Netmask(char* dst, size_t dst_size, uint8_t prefix_len) {
 #if defined(_WIN32)
+static void CopyIpv4Netmask(char* dst, size_t dst_size, uint8_t prefix_len) {
 	if (dst == nullptr || dst_size == 0 || prefix_len > 32) {
 		return;
 	}
@@ -3487,26 +3483,22 @@ static void CopyIpv4Netmask(char* dst, size_t dst_size, uint8_t prefix_len) {
 	in_addr        in {};
 	in.S_un.S_addr = htonl(mask);
 	inet_ntop(AF_INET, &in, dst, static_cast<socklen_t>(dst_size));
-#else
-	(void)dst;
-	(void)dst_size;
-	(void)prefix_len;
-#endif
 }
+#endif
 
 static HostNetworkInfo QueryHostNetworkInfo() {
 	HostNetworkInfo info;
 #if defined(_WIN32)
-	ULONG flags = GAA_FLAG_SKIP_ANYCAST | GAA_FLAG_SKIP_MULTICAST;
+	ULONG flags = GAA_FLAG_SKIP_ANYCAST | GAA_FLAG_SKIP_MULTICAST | GAA_FLAG_INCLUDE_GATEWAYS;
 	ULONG size  = 0;
-	auto  ret   = GetAdaptersAddresses(AF_UNSPEC, flags, nullptr, nullptr, &size);
+	auto  ret   = GetAdaptersAddresses(AF_INET, flags, nullptr, nullptr, &size);
 	if (ret != ERROR_BUFFER_OVERFLOW || size == 0) {
 		return info;
 	}
 
 	std::vector<uint8_t> buffer(size);
 	auto*                adapters = reinterpret_cast<IP_ADAPTER_ADDRESSES*>(buffer.data());
-	ret = GetAdaptersAddresses(AF_UNSPEC, flags, nullptr, adapters, &size);
+	ret = GetAdaptersAddresses(AF_INET, flags, nullptr, adapters, &size);
 	if (ret != NO_ERROR) {
 		return info;
 	}
@@ -3557,40 +3549,76 @@ static HostNetworkInfo QueryHostNetworkInfo() {
 					}
 					return info;
 				}
-			} else if (addr->sa_family == AF_INET6) {
-				const auto* in6 = reinterpret_cast<const sockaddr_in6*>(addr);
-				if (!IN6_IS_ADDR_LOOPBACK(&in6->sin6_addr) &&
-				    !IN6_IS_ADDR_UNSPECIFIED(&in6->sin6_addr)) {
-					info.connected = true;
-				}
 			}
 		}
 	}
 
 	return info;
 #else
-	info.connected = true;
+	ifaddrs* addresses = nullptr;
+	if (getifaddrs(&addresses) != 0) {
+		return info;
+	}
+	const auto usable = [](const ifaddrs* address) {
+		return address->ifa_addr != nullptr && address->ifa_addr->sa_family == AF_INET &&
+		       (address->ifa_flags & (IFF_UP | IFF_RUNNING | IFF_LOOPBACK)) ==
+		           (IFF_UP | IFF_RUNNING) &&
+		       reinterpret_cast<const sockaddr_in*>(address->ifa_addr)->sin_addr.s_addr != 0;
+	};
+	const ifaddrs* selected = nullptr;
+	for (auto* address = addresses; address != nullptr; address = address->ifa_next) {
+		if (usable(address)) {
+			selected = address;
+			break;
+		}
+	}
+	std::ifstream routes("/proc/net/route");
+	std::string line;
+	uint32_t best_metric = std::numeric_limits<uint32_t>::max();
+	uint32_t gateway = 0;
+	bool has_default_route = false;
+	while (std::getline(routes, line)) {
+		std::istringstream fields(line);
+		std::string name;
+		uint32_t destination = 0, candidate_gateway = 0, flags = 0, refs = 0, uses = 0;
+		uint32_t metric = 0, mask = 0;
+		if (!(fields >> name >> std::hex >> destination >> candidate_gateway >> flags >>
+		      std::dec >> refs >> uses >> metric >> std::hex >> mask) ||
+		    destination != 0 || mask != 0 || (flags & (RTF_UP | RTF_REJECT)) != RTF_UP ||
+		    (has_default_route && metric >= best_metric)) {
+			continue;
+		}
+		for (auto* address = addresses; address != nullptr; address = address->ifa_next) {
+			if (usable(address) && name == address->ifa_name) {
+				selected = address;
+				gateway = candidate_gateway;
+				best_metric = metric;
+				has_default_route = true;
+				break;
+			}
+		}
+	}
+	if (selected != nullptr) {
+		CopyIpv4String(info.ip_address, sizeof(info.ip_address), selected->ifa_addr);
+		CopyIpv4String(info.netmask, sizeof(info.netmask), selected->ifa_netmask);
+		info.connected = info.ip_address[0] != '\0';
+		if (has_default_route) {
+			in_addr address {gateway};
+			inet_ntop(AF_INET, &address, info.default_route, sizeof(info.default_route));
+		}
+	}
+	freeifaddrs(addresses);
 	return info;
 #endif
 }
 
 static bool NetCtlConnected() {
-	if (!g_net_ctl_status_initialized.load()) {
-		g_net_ctl_connected          = false;
-		g_net_ctl_status_initialized = true;
-		LOGF("\t host network connected = false (forced offline)\n");
-	}
-
-	return g_net_ctl_connected.load();
+	return Config::IsNetworkConnected() && QueryHostNetworkInfo().connected;
 }
 
 int KYTY_SYSV_ABI NetCtlInit() {
 	PRINT_NAME();
 
-	g_net_ctl_connected          = false;
-	g_net_ctl_status_initialized = true;
-	LOGF("\t host network connected = false (forced offline)\n");
-
 	return OK;
 }
 
@@ -3651,7 +3679,7 @@ int KYTY_SYSV_ABI NetCtlGetStateV6(int* state) {
 		return NET_CTL_ERROR_INVALID_ADDR;
 	}
 
-	*state = (NetCtlConnected() ? NET_CTL_STATE_IPOBTAINED : NET_CTL_STATE_DISCONNECTED);
+	*state = NET_CTL_STATE_DISCONNECTED;
 
 	return OK;
 }
@@ -3720,24 +3748,28 @@ int KYTY_SYSV_ABI NetCtlGetResult(int event_type, int* error_code) {
 int KYTY_SYSV_ABI NetCtlGetInfo(int code, NetCtlInfo* info) {
 	PRINT_NAME();
 
-	EXIT_NOT_IMPLEMENTED(info == nullptr);
-
+	if (info == nullptr) {
+		return NET_CTL_ERROR_INVALID_ADDR;
+	}
+	if (code < 1 || code > 21) {
+		return NET_CTL_ERROR_INVALID_CODE;
+	}
+	if (!Config::IsNetworkConnected()) {
+		return NET_CTL_ERROR_NOT_CONNECTED;
+	}
+	const auto host = QueryHostNetworkInfo();
+	if (!host.connected || (code == 16 && host.default_route[0] == '\0')) {
+		return NET_CTL_ERROR_NOT_CONNECTED;
+	}
 	memset(info, 0, sizeof(NetCtlInfo));
 
-	// Online/host-backed info responses are left below for testing, but disabled
-	// to preserve the disconnected-console behavior.
-	return NET_CTL_ERROR_NOT_CONNECTED;
-
 	switch (code) {
 		case 1:
-			info->device =
-			    (QueryHostNetworkInfo().wireless ? NET_CTL_DEVICE_WIRELESS : NET_CTL_DEVICE_WIRED);
+			info->device = (host.wireless ? NET_CTL_DEVICE_WIRELESS : NET_CTL_DEVICE_WIRED);
 			break;
-		case 2: info->ether_addr = QueryHostNetworkInfo().ether_addr; break;
+		case 2: info->ether_addr = host.ether_addr; break;
 		case 3: info->mtu = 1500; break;
-		case 4:
-			info->link = (NetCtlConnected() ? NET_CTL_LINK_CONNECTED : NET_CTL_LINK_DISCONNECTED);
-			break;
+		case 4: info->link = NET_CTL_LINK_CONNECTED; break;
 		case 5: break;
 		case 6: break;
 		case 7: info->wifi_security = 0; break;
@@ -3748,28 +3780,25 @@ int KYTY_SYSV_ABI NetCtlGetInfo(int code, NetCtlInfo* info) {
 		case 12: break;
 		case 13: break;
 		case 14:
-			std::strncpy(info->ip_address, QueryHostNetworkInfo().ip_address,
+			std::strncpy(info->ip_address, host.ip_address,
 			             sizeof(info->ip_address) - 1);
 			break;
 		case 15:
-			std::strncpy(info->netmask, QueryHostNetworkInfo().netmask, sizeof(info->netmask) - 1);
+			std::strncpy(info->netmask, host.netmask, sizeof(info->netmask) - 1);
 			break;
 		case 16:
-			std::strncpy(info->default_route, QueryHostNetworkInfo().default_route,
+			std::strncpy(info->default_route, host.default_route,
 			             sizeof(info->default_route) - 1);
 			break;
 		case 17:
-			std::strncpy(info->primary_dns, QueryHostNetworkInfo().primary_dns,
-			             sizeof(info->primary_dns) - 1);
+			std::strncpy(info->primary_dns, host.primary_dns, sizeof(info->primary_dns) - 1);
 			break;
 		case 18:
-			std::strncpy(info->secondary_dns, QueryHostNetworkInfo().secondary_dns,
-			             sizeof(info->secondary_dns) - 1);
+			std::strncpy(info->secondary_dns, host.secondary_dns, sizeof(info->secondary_dns) - 1);
 			break;
 		case 19: info->http_proxy_config = NET_CTL_HTTP_PROXY_OFF; break;
 		case 20: break;
 		case 21: info->http_proxy_port = 0; break;
-		default: LOGF("\t unknown NetCtl info code: %d\n", code); return NET_CTL_ERROR_INVALID_CODE;
 	}
 
 	return OK;

src/libs/ngs2.cpp

@@ -30,6 +30,10 @@ LIB_NAME("Ngs2", "Ngs2");
 
 constexpr int32_t NGS2_ERROR_INVALID_OUT_ADDRESS =
     static_cast<int32_t>(0x804a8010u);
+constexpr int32_t NGS2_ERROR_INVALID_NUM_WAVEFORM_BLOCKS =
+    static_cast<int32_t>(0x804a8053u);
+constexpr int32_t NGS2_ERROR_FULL_WAVEFORM_BLOCKS =
+    static_cast<int32_t>(0x804a8054u);
 constexpr int32_t NGS2_ERROR_INVALID_WAVEFORM_DATA =
     static_cast<int32_t>(0x804a8430u);
 constexpr int32_t NGS2_ERROR_INVALID_WAVEFORM_FORMAT =
@@ -549,6 +553,7 @@ struct Ngs2VoiceInternal {
 	uint64_t                            sample_step        = 0;
 	uint64_t                            decoded_samples    = 0;
 	uint64_t                            decoded_bytes      = 0;
+	uint64_t                            source_generation  = 0;
 	uint32_t                            duration_remaining = 0;
 	uint32_t                            skip_remaining     = 0;
 	const uint8_t*                      waveform_end       = nullptr;
@@ -574,6 +579,19 @@ struct Ngs2VoiceInternal {
 		filters.clear();
 	}
 
+	void ClearWaveformBlocks() {
+		++source_generation;
+		blocks.clear();
+		sample_phase       = 0;
+		duration_remaining = 0;
+		skip_remaining     = 0;
+		compressed_input.clear();
+		frame_cursor       = 0;
+		frame_samples      = 0;
+		next_frame_samples = 0;
+		waveform_end       = nullptr;
+	}
+
 	void SetupSampler(const Ngs2WaveformFormat& format) {
 		const bool deferred = format.waveform_type == 0 && format.num_channels == 0 &&
 		                      format.sample_rate == 0 && format.config_data == 0 &&
@@ -582,23 +600,15 @@ struct Ngs2VoiceInternal {
 		                     (format.frame_margin != 0 || format.frame_offset != 0 ||
 		                      format.sample_rate == 0 || format.num_channels == 0));
 		ResetSetupState();
-		channels           = format.num_channels;
-		sample_rate        = format.sample_rate;
-		sample_phase       = 0;
-		sample_step        = uint64_t(sample_rate) << 32u;
-		decoded_samples    = 0;
-		decoded_bytes      = 0;
-		duration_remaining = 0;
-		skip_remaining     = 0;
-		waveform_end       = nullptr;
-		blocks.clear();
+		ClearWaveformBlocks();
+		channels        = format.num_channels;
+		sample_rate     = format.sample_rate;
+		sample_step     = uint64_t(sample_rate) << 32u;
+		decoded_samples = 0;
+		decoded_bytes   = 0;
 		decoder.reset();
 		decoded_frame.clear();
 		next_decoded_frame.clear();
-		compressed_input.clear();
-		frame_cursor       = 0;
-		frame_samples      = 0;
-		next_frame_samples = 0;
 		accepts_blocks = true;
 		if (deferred) {
 			return;
@@ -625,6 +635,63 @@ struct Ngs2VoiceInternal {
 		}
 	}
 
+	int AddWaveformBlocks(const uint8_t* data, uint32_t flags, uint32_t count,
+	                      const Ngs2WaveformBlock* waveform_blocks) {
+		const auto capacity = rack->type == Ngs2RackType::Sampler
+		                          ? rack->option.sampler.max_waveform_blocks
+		                          : rack->option.custom_sampler.max_waveform_blocks;
+		if (count > capacity) {
+			return NGS2_ERROR_INVALID_NUM_WAVEFORM_BLOCKS;
+		}
+		EXIT_NOT_IMPLEMENTED(!accepts_blocks || (flags & ~0x17u) != 0);
+		const bool clear     = (flags & 4u) != 0;
+		const bool only_data = (flags & 2u) != 0;
+		const bool join      = only_data && (flags & 0x10u) == 0;
+		EXIT_NOT_IMPLEMENTED(only_data && decoder == nullptr);
+		const auto retained = clear ? 0 : blocks.size();
+		Block      tail     = retained == 0 ? Block {} : blocks.back();
+		std::vector<Block> pending;
+		pending.reserve(count);
+		for (uint32_t i = 0; i < count; ++i) {
+			auto info = waveform_blocks[i];
+			if (!only_data && info.num_samples == 0 && info.data_size == 0) {
+				continue;
+			}
+			EXIT_NOT_IMPLEMENTED(!only_data && info.num_samples == 0);
+			if (only_data) {
+				info.num_repeats      = 0;
+				info.num_skip_samples = 0;
+				info.num_samples      = 0;
+			}
+			EXIT_NOT_IMPLEMENTED(info.num_repeats != 0 && rack->type != Ngs2RackType::Sampler);
+			EXIT_NOT_IMPLEMENTED(decoder == nullptr &&
+			                     (uint64_t(info.num_skip_samples) + info.num_samples) *
+			                             channels * sizeof(int16_t) >
+			                         info.data_size);
+			Block block {data + info.data_offset, info};
+			auto* previous = pending.empty() ? (retained == 0 ? nullptr : &tail)
+			                                 : &pending.back();
+			if (join && previous != nullptr &&
+			    previous->data + previous->info.data_size == block.data) {
+				previous->info.data_size += block.info.data_size;
+			} else {
+				pending.push_back(block);
+			}
+		}
+		// Validate the complete addition before clearing, extending, or closing the stream.
+		if (retained + pending.size() > capacity) {
+			return NGS2_ERROR_FULL_WAVEFORM_BLOCKS;
+		}
+		if (clear) {
+			ClearWaveformBlocks();
+		} else if (retained != 0) {
+			blocks.back().info.data_size = tail.info.data_size;
+		}
+		blocks.insert(blocks.end(), pending.begin(), pending.end());
+		accepts_blocks = (flags & 1u) != 0;
+		return OK;
+	}
+
 	void SetMatrix(uint32_t index, const float* levels, uint32_t count) {
 		EXIT_NOT_IMPLEMENTED(index >= matrices.size());
 		matrices[index].assign(levels, levels + count);
@@ -1425,9 +1492,10 @@ static_assert(offsetof(Ngs2VoiceCallbackInfo, size) == 40 &&
               offsetof(Ngs2VoiceCallbackInfo, repeats) == 44 &&
               offsetof(Ngs2VoiceCallbackInfo, attributes) == 48);
 
-static void Ngs2FinishBlock(Ngs2VoiceInternal& voice) {
-	auto&      block   = voice.blocks.front();
-	const bool repeat  = block.info.num_repeats != 0;
+static bool Ngs2FinishBlock(Ngs2VoiceInternal& voice) {
+	const auto generation = voice.source_generation;
+	auto&      block      = voice.blocks.front();
+	const bool repeat     = block.info.num_repeats != 0;
 	voice.waveform_end = voice.WaveformData();
 	if (repeat) {
 		if (block.info.num_repeats != UINT32_MAX) {
@@ -1458,32 +1526,41 @@ static void Ngs2FinishBlock(Ngs2VoiceInternal& voice) {
 		reinterpret_cast<void KYTY_SYSV_ABI (*)(const Ngs2VoiceCallbackInfo*)>(voice.callback)(
 		    &info);
 	}
+	return voice.source_generation == generation;
 }
 
-static void Ngs2AdvancePcm(Ngs2VoiceInternal& voice) {
-	const auto output_rate = uint64_t(voice.rack->ngs->option.sample_rate) << 32u;
-	// Completion callbacks may interrupt advancement across multiple blocks.
+static bool Ngs2DecodePcmFrame(Ngs2VoiceInternal& voice, std::vector<float>& frame,
+                               uint32_t& frame_samples) {
+	// PCM has no encoded frame size. Use one system grain as a bounded staging
+	// quantum, keeping source decoding separate from resampled output consumption.
+	const auto quantum = std::max(voice.rack->ngs->option.num_grain_samples, 1u);
+	frame.resize(size_t(quantum) * voice.channels);
+	frame_samples = 0;
 	while (voice.state == Ngs2VoicePlayState::Playing && !voice.blocks.empty() &&
-	       voice.sample_phase >= output_rate) {
-		auto&      block     = voice.blocks.front();
-		const auto available = block.info.num_samples - block.cursor;
-		const auto advance   = std::min(voice.sample_phase / output_rate,
-		                                uint64_t(available));
-		block.cursor += static_cast<uint32_t>(advance);
-		voice.decoded_samples += advance;
-		voice.decoded_bytes += advance * voice.channels * sizeof(int16_t);
-		voice.sample_phase -= advance * output_rate;
-		if (block.cursor == block.info.num_samples) {
-			Ngs2FinishBlock(voice);
+	       frame_samples < quantum) {
+		auto& block = voice.blocks.front();
+		const auto count = std::min(quantum - frame_samples, block.info.num_samples - block.cursor);
+		const auto* pcm = reinterpret_cast<const int16_t*>(voice.WaveformData());
+		for (size_t i = 0; i < size_t(count) * voice.channels; ++i) {
+			frame[size_t(frame_samples) * voice.channels + i] = pcm[i] / 32768.0f;
+		}
+		block.cursor += count;
+		frame_samples += count;
+		voice.decoded_samples += count;
+		voice.decoded_bytes += size_t(count) * voice.channels * sizeof(int16_t);
+		if (block.cursor == block.info.num_samples && !Ngs2FinishBlock(voice)) {
+			return false;
 		}
 	}
-	if (voice.blocks.empty() && !voice.accepts_blocks) {
-		voice.sample_phase = 0;
-	}
+	return frame_samples != 0;
 }
 
 static bool Ngs2DecodeFrame(Ngs2VoiceInternal& voice, std::vector<float>& frame,
                             uint32_t& frame_samples) {
+	if (voice.decoder == nullptr) {
+		return Ngs2DecodePcmFrame(voice, frame, frame_samples);
+	}
+	frame_samples = 0;
 	while (voice.state == Ngs2VoicePlayState::Playing && !voice.blocks.empty()) {
 		auto& block = voice.blocks.front();
 		if (block.info.num_samples != 0 && block.data_cursor == 0) {
@@ -1491,7 +1568,9 @@ static bool Ngs2DecodeFrame(Ngs2VoiceInternal& voice, std::vector<float>& frame,
 			voice.skip_remaining     = block.info.num_skip_samples;
 		}
 		if (voice.duration_remaining == 0) {
-			Ngs2FinishBlock(voice);
+			if (!Ngs2FinishBlock(voice)) {
+				return false;
+			}
 			continue;
 		}
 		const auto skip =
@@ -1512,7 +1591,9 @@ static bool Ngs2DecodeFrame(Ngs2VoiceInternal& voice, std::vector<float>& frame,
 			block.data_cursor += bytes;
 			voice.decoded_bytes += bytes;
 			if (voice.compressed_input.size() < info.next_frame_size) {
-				Ngs2FinishBlock(voice);
+				if (!Ngs2FinishBlock(voice)) {
+					return false;
+				}
 				continue;
 			}
 		}
@@ -1529,7 +1610,9 @@ static bool Ngs2DecodeFrame(Ngs2VoiceInternal& voice, std::vector<float>& frame,
 			                              block.data + block.info.data_size);
 			voice.decoded_bytes += block.info.data_size - block.data_cursor;
 			block.data_cursor = block.info.data_size;
-			Ngs2FinishBlock(voice);
+			if (!Ngs2FinishBlock(voice)) {
+				return false;
+			}
 			continue;
 		}
 		EXIT_NOT_IMPLEMENTED(result.result != OK || result.input_consumed == 0);
@@ -1547,13 +1630,16 @@ static bool Ngs2DecodeFrame(Ngs2VoiceInternal& voice, std::vector<float>& frame,
 			EXIT_NOT_IMPLEMENTED(decoded_samples > voice.duration_remaining);
 			frame_samples = decoded_samples;
 			voice.duration_remaining -= decoded_samples;
+			voice.decoded_samples += decoded_samples;
 		}
 		if (voice.duration_remaining == 0 ||
 		    (block.data_cursor == block.info.data_size && voice.compressed_input.empty())) {
 			if (voice.duration_remaining == 0) {
 				voice.compressed_input.clear();
 			}
-			Ngs2FinishBlock(voice);
+			if (!Ngs2FinishBlock(voice)) {
+				return false;
+			}
 		}
 		if (decoded_samples != 0) {
 			return true;
@@ -1562,8 +1648,10 @@ static bool Ngs2DecodeFrame(Ngs2VoiceInternal& voice, std::vector<float>& frame,
 	return false;
 }
 
-static bool Ngs2CurrentAtracFrame(Ngs2VoiceInternal& voice) {
+static bool Ngs2CurrentFrame(Ngs2VoiceInternal& voice) {
 	if (voice.frame_cursor == voice.frame_samples) {
+		voice.frame_cursor  = 0;
+		voice.frame_samples = 0;
 		if (voice.next_frame_samples != 0) {
 			std::swap(voice.decoded_frame, voice.next_decoded_frame);
 			voice.frame_samples      = voice.next_frame_samples;
@@ -1571,21 +1659,21 @@ static bool Ngs2CurrentAtracFrame(Ngs2VoiceInternal& voice) {
 		} else if (!Ngs2DecodeFrame(voice, voice.decoded_frame, voice.frame_samples)) {
 			return false;
 		}
-		voice.frame_cursor = 0;
 	}
 	return voice.state == Ngs2VoicePlayState::Playing &&
 	       voice.frame_cursor < voice.frame_samples;
 }
 
-static void Ngs2AdvanceAtrac9(Ngs2VoiceInternal& voice) {
+static void Ngs2AdvanceSamples(Ngs2VoiceInternal& voice) {
 	const auto output_rate = uint64_t(voice.rack->ngs->option.sample_rate) << 32u;
 	while (voice.state == Ngs2VoicePlayState::Playing && voice.sample_phase >= output_rate) {
-		if (!Ngs2CurrentAtracFrame(voice)) {
+		if (!Ngs2CurrentFrame(voice)) {
 			break;
 		}
-		++voice.frame_cursor;
-		++voice.decoded_samples;
-		voice.sample_phase -= output_rate;
+		const auto advance = std::min(voice.sample_phase / output_rate,
+		                              uint64_t(voice.frame_samples - voice.frame_cursor));
+		voice.frame_cursor += static_cast<uint32_t>(advance);
+		voice.sample_phase -= advance * output_rate;
 	}
 	if (voice.blocks.empty() && voice.frame_cursor == voice.frame_samples &&
 	    voice.next_frame_samples == 0 && !voice.accepts_blocks) {
@@ -1594,42 +1682,14 @@ static void Ngs2AdvanceAtrac9(Ngs2VoiceInternal& voice) {
 }
 
 static void Ngs2ConsumeSamples(Ngs2VoiceInternal& voice, uint32_t grain) {
-	uint32_t output = 0;
+	const auto generation = voice.source_generation;
+	uint32_t   output     = 0;
 	const auto output_rate = uint64_t(voice.rack->ngs->option.sample_rate) << 32u;
 	while (voice.state == Ngs2VoicePlayState::Playing && output < grain &&
 	       (!voice.blocks.empty() || voice.frame_cursor < voice.frame_samples ||
 	        voice.next_frame_samples != 0)) {
-		if (voice.decoder == nullptr) {
-			Ngs2AdvancePcm(voice);
-			if (voice.state != Ngs2VoicePlayState::Playing || voice.blocks.empty()) {
-				break;
-			}
-			const auto& block = voice.blocks.front();
-			const auto* pcm   = reinterpret_cast<const int16_t*>(voice.WaveformData());
-			const auto* next = pcm;
-			if (block.cursor + 1 < block.info.num_samples) {
-				next += voice.channels;
-			} else if (block.info.num_repeats != 0) {
-				next = reinterpret_cast<const int16_t*>(block.data) +
-				       block.info.num_skip_samples * voice.channels;
-			} else if (voice.blocks.size() > 1) {
-				const auto& next_block = voice.blocks[1];
-				next = reinterpret_cast<const int16_t*>(next_block.data) +
-				       next_block.info.num_skip_samples * voice.channels;
-			}
-			const float fraction = static_cast<float>(voice.sample_phase) /
-			                       static_cast<float>(output_rate);
-			for (uint32_t c = 0; c < voice.channels; ++c) {
-				voice.samples[c * grain + output] =
-				    std::lerp(static_cast<float>(pcm[c]), static_cast<float>(next[c]), fraction) /
-				    32768.0f;
-			}
-			voice.sample_phase += voice.sample_step;
-			++output;
-			continue;
-		}
-		Ngs2AdvanceAtrac9(voice);
-		if (!Ngs2CurrentAtracFrame(voice)) {
+		Ngs2AdvanceSamples(voice);
+		if (voice.source_generation != generation || !Ngs2CurrentFrame(voice)) {
 			break;
 		}
 		if (voice.frame_cursor + 1 == voice.frame_samples && voice.sample_phase != 0) {
@@ -1639,7 +1699,8 @@ static void Ngs2ConsumeSamples(Ngs2VoiceInternal& voice, uint32_t grain) {
 				}
 				Ngs2DecodeFrame(voice, voice.next_decoded_frame, voice.next_frame_samples);
 			}
-			if (voice.state != Ngs2VoicePlayState::Playing ||
+			if (voice.source_generation != generation ||
+			    voice.state != Ngs2VoicePlayState::Playing ||
 			    voice.frame_cursor >= voice.frame_samples) {
 				break;
 			}
@@ -1657,10 +1718,13 @@ static void Ngs2ConsumeSamples(Ngs2VoiceInternal& voice, uint32_t grain) {
 		voice.sample_phase += voice.sample_step;
 		++output;
 	}
-	if (voice.decoder == nullptr) {
-		Ngs2AdvancePcm(voice);
-	} else {
-		Ngs2AdvanceAtrac9(voice);
+	if (voice.source_generation == generation) {
+		Ngs2AdvanceSamples(voice);
+	}
+	if (voice.source_generation != generation) {
+		// A callback replaced the source, possibly changing its channel layout.
+		voice.samples.assign(size_t(grain) * voice.channels, 0.0f);
+		output = 0;
 	}
 	voice.has_samples = output != 0;
 }
@@ -2370,41 +2434,10 @@ int KYTY_SYSV_ABI Ngs2VoiceControl(uintptr_t voice_handle, const Ngs2VoiceParamH
 							const Ngs2WaveformBlock* blocks;
 						};
 						const auto& blocks = *reinterpret_cast<const BlocksParam*>(param);
-						EXIT_NOT_IMPLEMENTED(!voice->accepts_blocks ||
-						                     (blocks.flags & ~0x17u) != 0);
-						if ((blocks.flags & 0x4u) != 0) {
-							voice->blocks.clear();
-							voice->sample_phase       = 0;
-							voice->duration_remaining = 0;
-							voice->skip_remaining     = 0;
-							voice->compressed_input.clear();
-							voice->frame_cursor       = 0;
-							voice->frame_samples      = 0;
-							voice->next_frame_samples = 0;
-							voice->waveform_end  = nullptr;
-						}
-						voice->accepts_blocks = (blocks.flags & 1u) != 0;
-						const bool only_data  = (blocks.flags & 2u) != 0;
-						EXIT_NOT_IMPLEMENTED(only_data && voice->decoder == nullptr);
-						for (uint32_t i = 0; i < blocks.count; ++i) {
-							auto block = blocks.blocks[i];
-							if (!only_data && block.num_samples == 0 && block.data_size == 0) {
-								continue;
-							}
-							EXIT_NOT_IMPLEMENTED(!only_data && block.num_samples == 0);
-							if (only_data) {
-								block.num_repeats      = 0;
-								block.num_skip_samples = 0;
-								block.num_samples      = 0;
-							}
-							EXIT_NOT_IMPLEMENTED(block.num_repeats != 0 &&
-							                     voice->rack->type != Ngs2RackType::Sampler);
-							EXIT_NOT_IMPLEMENTED(
-							    voice->decoder == nullptr &&
-							    (uint64_t(block.num_skip_samples) + block.num_samples) *
-							            voice->channels * 2 >
-							        block.data_size);
-							voice->blocks.push_back({blocks.data + block.data_offset, block});
+						const auto result = voice->AddWaveformBlocks(
+						    blocks.data, blocks.flags, blocks.count, blocks.blocks);
+						if (result != OK) {
+							return result;
 						}
 						break;
 					}
@@ -2511,6 +2544,15 @@ int KYTY_SYSV_ABI Ngs2VoiceRunCommands(uintptr_t voice_handle, const void* comma
 	return OK;
 }
 
+template <typename State>
+static void Ngs2FillSamplerState(const Ngs2VoiceInternal& voice, State& state) {
+	state.voice_state.state_flags = Ngs2GetStateFlags(&voice);
+	state.num_decoded_samples     = voice.decoded_samples;
+	state.decoded_data_size       = voice.decoded_bytes;
+	state.waveform_data           = voice.WaveformData();
+	state.user_data               = voice.blocks.empty() ? 0 : voice.blocks.front().info.user_data;
+}
+
 int KYTY_SYSV_ABI Ngs2VoiceGetState(uintptr_t voice_handle, Ngs2VoiceState* state,
                                     size_t state_size) {
 	PRINT_NAME();
@@ -2547,7 +2589,7 @@ int KYTY_SYSV_ABI Ngs2VoiceGetState(uintptr_t voice_handle, Ngs2VoiceState* stat
 			EXIT_NOT_IMPLEMENTED(state_size != configured_size);
 			std::memset(state, 0, state_size);
 			auto* sampler = reinterpret_cast<Ngs2CustomSamplerVoiceState*>(state);
-			sampler->voice_state.state_flags = Ngs2GetStateFlags(voice);
+			Ngs2FillSamplerState(*voice, *sampler);
 			break;
 		}
 		case Ngs2RackType::Sampler: {
@@ -2568,12 +2610,7 @@ int KYTY_SYSV_ABI Ngs2VoiceGetState(uintptr_t voice_handle, Ngs2VoiceState* stat
 			sampler->envelope_height     = 1.0f;
 			sampler->peak_height         = 0.0f;
 			sampler->reserved            = 0;
-			sampler->num_decoded_samples = voice->decoded_samples;
-			sampler->decoded_data_size   = voice->decoded_bytes;
-			sampler->waveform_data       = voice->WaveformData();
-			if (!voice->blocks.empty()) {
-				sampler->user_data = voice->blocks.front().info.user_data;
-			}
+			Ngs2FillSamplerState(*voice, *sampler);
 			break;
 		}
 		default: EXIT("unknown type: %s\n", magic_enum::enum_name(voice->rack->type));

src/loader/loadExec.cpp

@@ -0,0 +1,130 @@
+#include "common/file.h"
+#include "common/logging/log.h"
+#include "emulator.h"
+#include "kernel/fileSystem.h"
+#include "libs/errno.h"
+
+#include <cerrno>
+#include <cstdlib>
+#include <cstring>
+#include <fstream>
+#include <string_view>
+#include <vector>
+
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+#elif defined(__linux__)
+#include <linux/close_range.h>
+#include <unistd.h>
+#endif
+
+namespace Emulator {
+
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
+static std::wstring QuoteArgument(std::wstring_view argument) {
+	std::wstring quoted  = L"\"";
+	size_t       slashes = 0;
+	for (wchar_t c: argument) {
+		if (c == L'\\') {
+			++slashes;
+			continue;
+		}
+		quoted.append(c == L'"' ? slashes * 2 + 1 : slashes, L'\\');
+		quoted += c;
+		slashes = 0;
+	}
+	quoted.append(slashes * 2, L'\\');
+	return quoted + L'"';
+}
+#endif
+
+int LoadExec(const char* path, const char* const argv[]) {
+	using namespace Libs::SystemService;
+	if (path == nullptr || !std::string_view(path).starts_with("/app0/") ||
+	    !Common::File::IsFileExisting(Libs::LibKernel::FileSystem::GetRealFilename(path))) {
+		return SYSTEM_SERVICE_ERROR_PARAMETER;
+	}
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS || defined(__linux__)
+	std::vector<std::string> arguments {"--guest-exec", path};
+	size_t                   bytes = 0;
+	for (size_t i = 0; argv != nullptr && argv[i] != nullptr; ++i) {
+		const auto size = strnlen(argv[i], 4096 - bytes) + 1;
+		if (size > 4096 - bytes) {
+			return SYSTEM_SERVICE_ERROR_PARAMETER;
+		}
+		bytes += size;
+		arguments.emplace_back(argv[i], size - 1);
+	}
+	LOGF("LoadExec: %s\n", path);
+	Log::Flush();
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
+	wchar_t    executable[32768] {};
+	const auto length = GetModuleFileNameW(nullptr, executable, 32768);
+	if (length == 0 || length >= 32768) {
+		return SYSTEM_SERVICE_ERROR_INTERNAL;
+	}
+	if (__wargv == nullptr) {
+		return SYSTEM_SERVICE_ERROR_INTERNAL;
+	}
+	std::wstring command = QuoteArgument(executable);
+	for (int i = 1; i < __argc && __wargv[i] != nullptr; ++i) {
+		if (std::wstring_view(__wargv[i]) == L"--guest-exec") {
+			break;
+		}
+		command += L" " + QuoteArgument(__wargv[i]);
+	}
+	for (const auto& argument: arguments) {
+		const int size =
+		    MultiByteToWideChar(CP_UTF8, MB_ERR_INVALID_CHARS, argument.c_str(), -1, nullptr, 0);
+		if (size == 0) {
+			return SYSTEM_SERVICE_ERROR_PARAMETER;
+		}
+		std::wstring wide(size, L'\0');
+		if (MultiByteToWideChar(CP_UTF8, MB_ERR_INVALID_CHARS, argument.c_str(), -1, wide.data(),
+		                        size) == 0) {
+			return SYSTEM_SERVICE_ERROR_PARAMETER;
+		}
+		wide.resize(size - 1);
+		command += L" " + QuoteArgument(wide);
+	}
+	STARTUPINFOW startup {};
+	startup.cb = sizeof(startup);
+	PROCESS_INFORMATION process {};
+	if (!CreateProcessW(executable, command.data(), nullptr, nullptr, FALSE, CREATE_NO_WINDOW,
+	                    nullptr, nullptr, &startup, &process)) {
+		LOGF("LoadExec: process creation failed: %lu\n", GetLastError());
+		return SYSTEM_SERVICE_ERROR_INTERNAL;
+	}
+	CloseHandle(process.hThread);
+	CloseHandle(process.hProcess);
+	std::fflush(nullptr);
+	std::quick_exit(0);
+#else
+	std::ifstream            command("/proc/self/cmdline", std::ios::binary);
+	std::vector<std::string> host_arguments;
+	for (std::string argument;
+	     std::getline(command, argument, '\0') && argument != "--guest-exec";) {
+		host_arguments.push_back(std::move(argument));
+	}
+	if (host_arguments.empty()) {
+		return SYSTEM_SERVICE_ERROR_INTERNAL;
+	}
+	host_arguments.insert(host_arguments.end(), arguments.begin(), arguments.end());
+	std::vector<char*> host_argv;
+	for (auto& argument: host_arguments) {
+		host_argv.push_back(argument.data());
+	}
+	host_argv.push_back(nullptr);
+	if (close_range(3, ~0u, CLOSE_RANGE_UNSHARE | CLOSE_RANGE_CLOEXEC) == 0) {
+		execvp(host_argv[0], host_argv.data());
+	}
+	LOGF("LoadExec failed: %s\n", std::strerror(errno));
+#endif
+	return SYSTEM_SERVICE_ERROR_INTERNAL;
+#else
+	return SYSTEM_SERVICE_ERROR_UNAVAILABLE;
+#endif
+}
+
+} // namespace Emulator

src/loader/runtimeLinker.cpp

@@ -25,6 +25,7 @@
 
 #include <algorithm>
 #include <atomic>
+#include <cstddef>
 #include <cstdio>
 #include <cstdlib>
 #include <cstring>
@@ -88,16 +89,6 @@ ThreadLocalStorage::~ThreadLocalStorage() {
 	}
 }
 
-#pragma pack(1)
-
-struct EntryParams {
-	int         argc;
-	uint32_t    pad;
-	const char* argv[3];
-};
-
-#pragma pack()
-
 using atexit_func_t = KYTY_SYSV_ABI void (*)();
 using entry_func_t  = KYTY_SYSV_ABI void (*)(EntryParams* params, atexit_func_t atexit_func);
 using module_ini_fini_func_t = KYTY_SYSV_ABI int (*)(size_t args, const void* argp,
@@ -1283,11 +1274,16 @@ void RuntimeLinker::Execute(const std::filesystem::path& game_patch) {
 	LOGF_COLOR(Log::Color::BrightYellow, "---\n--- Execute: %s\n---\n", "Main");
 
 	if (auto entry = GetEntry(); entry != 0) {
+		const auto& arguments = Config::GetGuestArguments();
+		const auto params_size = offsetof(EntryParams, argv) +
+		                         (arguments.size() + 2) * sizeof(const char*);
 		auto* params = reinterpret_cast<EntryParams*>(
-		    (reinterpret_cast<uintptr_t>(main_stack_top) - 0x100u) & ~static_cast<uintptr_t>(0x0f));
-		std::memset(params, 0, sizeof(EntryParams));
-		params->argc    = 1;
-		params->argv[0] = "KytyEmu";
+		    (reinterpret_cast<uintptr_t>(main_stack_top) - params_size - 0x100u) &
+		    ~static_cast<uintptr_t>(0x0f));
+		std::memset(params, 0, params_size);
+		for (const auto& argument: arguments) {
+			params->argv[params->argc++] = argument.c_str();
+		}
 
 		LOGF("stack_addr = %" PRIx64 "\n", reinterpret_cast<uint64_t>(params));
 

src/loader/runtimeLinker.h

@@ -14,6 +14,12 @@
 
 namespace Loader {
 
+struct EntryParams {
+	int         argc;
+	uint32_t    pad;
+	const char* argv[1];
+};
+
 class Elf64;
 struct Elf64_Sym;
 struct Elf64_Rela;

src/loader/systemContent.cpp

@@ -164,6 +164,8 @@ void Psf::OpenJson(const std::filesystem::path& file_name) {
 	add_string_param("contentId", "CONTENT_ID");
 	add_string_param("contentVersion", "APP_VER");
 	add_string_param("appVersion", "APP_VER");
+	AddParamInt("ATTRIBUTE", 0);
+	add_int_param("attribute", "ATTRIBUTE");
 	add_int_param("attribute3", "ATTRIBUTE3");
 	add_int_param("userDefinedParam1", "USER_DEFINED_PARAM_1");
 	add_int_param("userDefinedParam2", "USER_DEFINED_PARAM_2");
@@ -360,10 +362,12 @@ void PlayGo::Open(const std::filesystem::path& file_name) {
 	}
 
 	uint32_t magic1 = 0;
+	constexpr uint32_t PLGO_MAGIC = 0x6f676c70;
+	constexpr uint32_t PLGX_MAGIC = 0x78676c70;
 
 	m_f.Read(&magic1, 4);
 
-	if (magic1 != 0x6f676c70) {
+	if (magic1 != PLGO_MAGIC && magic1 != PLGX_MAGIC) {
 		LOGF("invalid file: magic1 = %08" PRIx32 "\n", magic1);
 		return;
 	}

src/main.cpp

@@ -61,6 +61,7 @@ static void PrintUsage() {
 	::printf(
 	    "  --gpu <index>                        Vulkan physical device index. Default: auto.\n");
 	::printf("  --fullscreen                         Run in borderless desktop fullscreen.\n");
+	::printf("  --hdr                                Enable HDR10 output when supported.\n");
 	::printf(
 	    "  --hide-cursor                        Hide the cursor after 2 s idle. Default: off.\n");
 	::printf("  --no-splash                          Disable the startup splash screen.\n");
@@ -69,6 +70,7 @@ static void PrintUsage() {
 	::printf("  --vblank-frequency <num>             Virtual vblank frequency. Default: 60.\n");
 	::printf("  --console-language <0-29>            Console language. Default: 1 (English US).\n");
 	::printf("  --vulkan-validation <true|false>     Enable Vulkan validation.\n");
+	::printf("  --network-connected <true|false>     Use the host local network (default: false).\n");
 	::printf("  --gpu-assisted-validation <t|f>      Bounds-check shader accesses on the GPU.\n"
 	         "                                       Implies --vulkan-validation; very slow.\n");
 	::printf("  --shader-validation <true|false>     Enable shader validation.\n");
@@ -215,6 +217,11 @@ static bool ParseArgs(int argc, char* argv[], RunOptions& options, bool& show_he
 			continue;
 		}
 
+		if (arg == "--hdr") {
+			options.config.hdr_enabled = true;
+			continue;
+		}
+
 		if (arg == "--hide-cursor") {
 			options.config.hide_cursor_enabled = true;
 			continue;
@@ -300,6 +307,10 @@ static bool ParseArgs(int argc, char* argv[], RunOptions& options, bool& show_he
 				         value.c_str());
 				return false;
 			}
+		} else if (arg == "--guest-exec") {
+			options.elf = Common::PathFromUtf8(Common::FixFilenameSlash(value));
+			options.config.guest_arguments.assign(argv + i + 1, argv + argc);
+			break;
 		} else if (arg == "--game-patch") {
 			if (!options.game_patch.empty()) {
 				::printf("--game-patch can only be specified once\n");
@@ -378,6 +389,11 @@ static bool ParseArgs(int argc, char* argv[], RunOptions& options, bool& show_he
 				::printf("invalid console language: %s\n", value.c_str());
 				return false;
 			}
+		} else if (arg == "--network-connected") {
+			if (!ParseBool(value, options.config.network_connected)) {
+				::printf("invalid boolean for %s: %s\n", arg.c_str(), value.c_str());
+				return false;
+			}
 		} else if (arg == "--vulkan-validation") {
 			if (!ParseBool(value, options.config.vulkan_validation_enabled)) {
 				::printf("invalid boolean for %s: %s\n", arg.c_str(), value.c_str());

tests/KernelFileSystemTests.cpp

@@ -35,6 +35,9 @@
 #include <vector>
 
 #if defined(__linux__)
+#include <arpa/inet.h>
+#include <ifaddrs.h>
+#include <net/if.h>
 #include <sys/mman.h>
 #include <unistd.h>
 #endif
@@ -898,6 +901,11 @@ void CheckAmprOrdering(Loader::SymbolDatabase &symbols, uint32_t file_id) {
   using WriteCounter = int (KYTY_SYSV_ABI *)(void *, uint8_t, uint8_t, uint64_t,
                                            uint8_t, uint32_t);
   using WriteAddress = int (KYTY_SYSV_ABI *)(void *, volatile uint64_t *, uint64_t);
+  using WriteImmediately = int (KYTY_SYSV_ABI *)(void *, volatile uint64_t *, uint64_t,
+                                               uint32_t);
+  using MeasureWriteAddress = uint64_t (KYTY_SYSV_ABI *)(volatile uint64_t *, uint64_t);
+  using MeasureWriteImmediately = uint64_t (KYTY_SYSV_ABI *)(volatile uint64_t *, uint64_t,
+                                                           uint32_t);
   using ReadFile = int (KYTY_SYSV_ABI *)(void *, uint64_t, uint64_t, uint32_t,
                                        void *, uint64_t, uint64_t);
   using WriteKernelEvent = int (KYTY_SYSV_ABI *)(void *, uint64_t, uint64_t, uint64_t,
@@ -916,6 +924,9 @@ void CheckAmprOrdering(Loader::SymbolDatabase &symbols, uint32_t file_id) {
   const auto wait_counter = reinterpret_cast<WaitCounter>(find("cQb8Zr8Q0Y0"));
   const auto write_counter = reinterpret_cast<WriteCounter>(find("jK+yuYCI7MA"));
   const auto write_address = reinterpret_cast<WriteAddress>(find("sJXyWHjP-F8"));
+  const auto write_immediately = reinterpret_cast<WriteImmediately>(find("j0+3uJMxYJY"));
+  const auto measure_write_address = reinterpret_cast<MeasureWriteAddress>(find("C+IEj+BsAFM"));
+  const auto measure_write_immediately = reinterpret_cast<MeasureWriteImmediately>(find("4fgtGfXDrFc"));
   const auto read_file = reinterpret_cast<ReadFile>(find("mQ16-QdKv7k"));
   const auto write_event = reinterpret_cast<WriteKernelEvent>(find("H896Pt-yB4I"));
   const auto submit_apr = reinterpret_cast<SubmitApr>(find("ASoW5WE-UPo"));
@@ -927,7 +938,7 @@ void CheckAmprOrdering(Loader::SymbolDatabase &symbols, uint32_t file_id) {
     std::array<uint32_t, 256> data {};
   };
   std::array<Buffer, 4> buffers;
-  const auto reset = [&](size_t apr_count) {
+  const auto reset = [&](size_t apr_count, uint32_t data_size = sizeof(Buffer::data)) {
     for (size_t i = 0; i < buffers.size(); ++i) {
       auto &buffer = buffers[i];
       construct(buffer.header.data());
@@ -936,10 +947,60 @@ void CheckAmprOrdering(Loader::SymbolDatabase &symbols, uint32_t file_id) {
       } else {
         construct_amm(buffer.header.data());
       }
-      Check(set_buffer(buffer.header.data(), buffer.data.data(), sizeof(buffer.data)) == OK,
+      Check(set_buffer(buffer.header.data(), buffer.data.data(), data_size) == OK,
             "initialize AMPR command buffer");
     }
   };
+  struct LiteralWrite { uint64_t value; uint32_t size; };
+  constexpr std::array literal_writes {
+      LiteralWrite{0, 8}, LiteralWrite{3, 8}, LiteralWrite{4, 12},
+      LiteralWrite{(uint64_t{1} << 34) - 1, 12},
+      LiteralWrite{uint64_t{1} << 34, 16}, LiteralWrite{UINT64_MAX, 16}};
+  for (const auto &literal : literal_writes) {
+    for (const bool immediately : std::array{false, true}) {
+      reset(1, literal.size);
+      auto *command_buffer = buffers[0].header.data();
+      uint64_t fence = ~literal.value;
+      const auto append = [&](uint64_t value) {
+        return immediately ? write_immediately(command_buffer, &fence, value, 1)
+                           : write_address(command_buffer, &fence, value);
+      };
+      Check(measure_write_address(&fence, literal.value) == literal.size &&
+                measure_write_immediately(&fence, literal.value, 1) == literal.size &&
+                append(literal.value) == OK && offset(command_buffer) == literal.size,
+            "literal write fits exactly its measured command size");
+      Check(append(~literal.value) == Libs::LibKernel::KERNEL_ERROR_EBUSY &&
+                offset(command_buffer) == literal.size,
+            "rejected write preserves the full command buffer offset");
+      Result result {1234, 5678};
+      uint32_t submission = 0;
+      Check(submit_apr(command_buffer, 3, &result, &submission) == OK &&
+                wait_apr(submission) == OK && result.result == OK &&
+                fence == literal.value,
+            "bounded write preserves all sixty-four literal bits");
+    }
+  }
+  reset(1, 768);
+  std::array<std::array<char, 3>, 16> batch_output {};
+  std::array<uint64_t, 16> batch_fences {};
+  auto *batch = buffers[0].header.data();
+  for (size_t i = 0; i < batch_fences.size(); ++i) {
+    Check(read_file(batch, reinterpret_cast<uint64_t>(&buffers[0].header[3]),
+                    reinterpret_cast<uint64_t>(&buffers[0].header[4]), file_id,
+                    batch_output[i].data(), batch_output[i].size(), 0) == OK &&
+              write_address(batch, &batch_fences[i], i + 1) == OK,
+          "sixteen APR read/fence pairs fit the guest's 768-byte command buffer");
+  }
+  Result batch_result {1234, 5678};
+  uint32_t batch_submission = 0;
+  Check(submit_apr(batch, 3, &batch_result, &batch_submission) == OK &&
+            wait_apr(batch_submission) == OK && batch_result.result == OK,
+        "complete the bounded APR read batch");
+  for (size_t i = 0; i < batch_fences.size(); ++i) {
+    Check(batch_fences[i] == i + 1 &&
+              std::memcmp(batch_output[i].data(), "APR", 3) == 0,
+          "every read publishes its data and completion, including requests fourteen and fifteen");
+  }
   // WaitCompare order: ==, unsigned >/<, !=, wrapped >=, signed >/<.
   struct Comparison { uint8_t compare; uint64_t blocked, reference, released; };
   constexpr std::array comparisons {
@@ -1401,6 +1462,84 @@ void CheckSocketReceiveBuffer(int reader, int writer) {
 }
 #endif
 
+#if defined(__linux__)
+void KYTY_SYSV_ABI RecordNetCtlEvent(int event, void* arg) {
+  auto& events = *static_cast<std::array<int, 2>*>(arg);
+  ++events[0];
+  events[1] = event;
+}
+
+void TestNetCtlHostIpv4() {
+  using namespace Libs::Network::NetCtl;
+  ifaddrs* interfaces = nullptr;
+  Check(getifaddrs(&interfaces) == 0, "enumerate host IPv4 addresses");
+  std::vector<uint32_t> host_addresses;
+  for (auto* interface = interfaces; interface != nullptr; interface = interface->ifa_next) {
+    if (interface->ifa_addr == nullptr || interface->ifa_addr->sa_family != AF_INET ||
+        (interface->ifa_flags & IFF_UP) == 0 || (interface->ifa_flags & IFF_RUNNING) == 0 ||
+        (interface->ifa_flags & IFF_LOOPBACK) != 0) {
+      continue;
+    }
+    const auto address = reinterpret_cast<const sockaddr_in*>(interface->ifa_addr)->sin_addr;
+    if (address.s_addr != 0 && (ntohl(address.s_addr) >> 24) != 127) {
+      host_addresses.push_back(address.s_addr);
+    }
+  }
+  freeifaddrs(interfaces);
+
+  Config::ConfigOptions options;
+  options.printf_direction = Config::LogDirection::Silent;
+  Check(!options.network_connected, "host networking requires explicit opt-in");
+  alignas(uint32_t) std::array<char, 256> storage {};
+  auto* info = reinterpret_cast<NetCtlInfo*>(storage.data());
+  constexpr int NotConnected = static_cast<int>(0x80412108u);
+  Check(NetCtlGetInfo(14, nullptr) == static_cast<int>(0x80412107u) &&
+            NetCtlGetInfo(0, info) == static_cast<int>(0x80412106u) &&
+            NetCtlGetInfo(22, info) == static_cast<int>(0x80412106u),
+        "NetCtl reports invalid output addresses and information codes");
+  for (bool enabled : {false, true}) {
+    options.network_connected = enabled;
+    Config::Load(options);
+    Check(NetCtlInit() == OK, "initialize NetCtl for selected network mode");
+    const bool connected = enabled && !host_addresses.empty();
+    int state = -1;
+    Check(NetCtlGetState(&state) == OK && state == (connected ? 3 : 0),
+          "NetCtl IPv4 state matches opt-in and host connectivity");
+    Check(NetCtlGetStateV6(&state) == OK && state == 0,
+          "IPv4 connectivity does not advertise IPv6 support");
+
+    Check(NetCtlGetInfo(14, info) == (connected ? OK : NotConnected),
+          "NetCtl IPv4 query respects network mode and host state");
+    in_addr address {};
+    if (connected) {
+      Check(inet_pton(AF_INET, storage.data(), &address) == 1 &&
+                std::find(host_addresses.begin(), host_addresses.end(), address.s_addr) !=
+                    host_addresses.end(),
+            "NetCtl reports an IPv4 address assigned to an active host interface");
+    }
+    Check(NetCtlGetInfo(15, info) == (connected ? OK : NotConnected) &&
+              (!connected || inet_pton(AF_INET, storage.data(), &address) == 1),
+          "NetCtl returns the active IPv4 netmask only in connected mode");
+    const int gateway_result = NetCtlGetInfo(16, info);
+    Check(gateway_result == NotConnected ||
+              (connected && gateway_result == OK &&
+               inet_pton(AF_INET, storage.data(), &address) == 1),
+          "NetCtl reports a valid IPv4 gateway or its absence");
+
+    std::array<int, 2> events {};
+    int callback_id = -1;
+    Check(NetCtlRegisterCallback(RecordNetCtlEvent, &events, &callback_id) == OK &&
+              NetCtlCheckCallback() == OK && events[0] == 1 &&
+              events[1] == (connected ? 3 : 1) && NetCtlCheckCallback() == OK && events[0] == 1,
+          "NetCtl delivers the matching connection event once while state is unchanged");
+    Check(NetCtlUnregisterCallback(callback_id) == OK, "release NetCtl callback");
+    NetCtlTerm();
+  }
+  options.network_connected = false;
+  Config::Load(options);
+}
+#endif
+
 void CheckEtherAddressFormatting() {
   Loader::SymbolDatabase symbols;
   Libs::LibNet::InitNet_1_Net(&symbols);
@@ -1732,6 +1871,9 @@ int main(int, char**) {
   FileSystem::Shutdown();
   CheckSocketWakeup();
   CheckEtherAddressFormatting();
+#if defined(__linux__)
+  TestNetCtlHostIpv4();
+#endif
   TestNpWebApi2Memory();
   TestNpCommerceDialog();
   graphics.reset();

tests/Ngs2SamplerTests.cpp

@@ -24,9 +24,11 @@ struct Fixture {
 
 	Fixture(uint32_t rate, uint32_t channels = 1, Ngs2RackType type = Ngs2RackType::Sampler) {
 		system.option.sample_rate = 48000;
+		system.option.num_grain_samples = 256;
 		rack.ngs                  = &system;
 		rack.type                 = type;
 		voice.rack                = &rack;
+		SetCapacity(4);
 		struct Setup {
 			Ngs2VoiceParamHeader header {40, 0, 0x10000000};
 			Ngs2WaveformFormat   format;
@@ -39,20 +41,51 @@ struct Fixture {
 		voice.state = Ngs2VoicePlayState::Playing;
 	}
 
-	void Queue(const int16_t* data, uint32_t frames, uint32_t flags, uint32_t repeats = 0,
-	           uint32_t skip = 0) {
-		Ngs2WaveformBlock block {
-		    0,  size_t(frames + skip) * voice.channels * sizeof(int16_t), repeats, skip, frames, 0,
-		    123};
+	void SetCapacity(uint32_t capacity) {
+		if (rack.type == Ngs2RackType::Sampler) {
+			rack.option.sampler.max_waveform_blocks = capacity;
+		} else {
+			rack.option.custom_sampler.max_waveform_blocks = capacity;
+			rack.option.custom_sampler.custom_rack_option.state_size = sizeof(Ngs2CustomSamplerVoiceState);
+		}
+	}
+
+	Ngs2SamplerVoiceState State() {
+		Ngs2SamplerVoiceState state {};
+		if (rack.type == Ngs2RackType::CustomSampler) {
+			Ngs2CustomSamplerVoiceState custom {};
+			Check(Ngs2VoiceGetState(reinterpret_cast<uintptr_t>(&voice), &custom.voice_state,
+			                       sizeof(custom)) == OK, "custom sampler state failed");
+			state.voice_state = custom.voice_state;
+			state.num_decoded_samples = custom.num_decoded_samples;
+			state.decoded_data_size = custom.decoded_data_size;
+			state.waveform_data = custom.waveform_data;
+			state.user_data = custom.user_data;
+		} else {
+			Check(Ngs2VoiceGetState(reinterpret_cast<uintptr_t>(&voice), &state.voice_state,
+			                       sizeof(state)) == OK, "sampler state failed");
+		}
+		return state;
+	}
+
+	int QueueBlocks(const void* data, const Ngs2WaveformBlock* blocks, uint32_t count,
+	                uint32_t flags) {
 		struct Blocks {
 			Ngs2VoiceParamHeader     header {32, 0, 0x10000001};
 			const void*              data;
 			uint32_t                 flags, count;
 			const Ngs2WaveformBlock* blocks;
-		} param {{32, 0, 0x10000001}, data, flags, 1, &block};
+		} param {{32, 0, 0x10000001}, data, flags, count, blocks};
 		param.header.id = rack.type == Ngs2RackType::Sampler ? 0x10000001 : 0x40010001;
-		Check(Ngs2VoiceControl(reinterpret_cast<uintptr_t>(&voice), &param.header) == OK,
-		      "sampler queue failed");
+		return Ngs2VoiceControl(reinterpret_cast<uintptr_t>(&voice), &param.header);
+	}
+
+	void Queue(const int16_t* data, uint32_t frames, uint32_t flags, uint32_t repeats = 0,
+	           uint32_t skip = 0) {
+		const Ngs2WaveformBlock block {
+		    0, size_t(frames + skip) * voice.channels * sizeof(int16_t), repeats, skip, frames, 0,
+		    123};
+		Check(QueueBlocks(data, &block, 1, flags) == OK, "sampler queue failed");
 	}
 
 	void Render(uint32_t frames) {
@@ -61,6 +94,190 @@ struct Fixture {
 	}
 };
 
+void TestWaveformQueueCapacity() {
+	constexpr int32_t full = static_cast<int32_t>(0x804a8054u);
+	constexpr int32_t invalid_count = static_cast<int32_t>(0x804a8053u);
+	for (auto rack: {Ngs2RackType::Sampler, Ngs2RackType::CustomSampler}) {
+		for (uint32_t capacity: {1u, 4u, 16u}) {
+			Fixture f(24000, 1, rack);
+			f.SetCapacity(capacity);
+			f.system.option.num_grain_samples = 1;
+			const int16_t pcm[] = {8192, 16384};
+			const int16_t replacement[] = {24576, 4096};
+			std::vector<Ngs2WaveformBlock> blocks(capacity + 1);
+			for (uint32_t i = 0; i < blocks.size(); ++i) {
+				blocks[i] = {0, sizeof(pcm), 0, 0, 2, 0, i + 1};
+			}
+			Check(f.QueueBlocks(pcm, blocks.data(), capacity, 1) == OK,
+			      "configured waveform capacity was not available");
+			f.Render(1);
+			const auto phase = f.voice.sample_phase;
+			Check(f.QueueBlocks(replacement, blocks.data(), 1, 0) == full,
+			      "full waveform queue accepted another guest buffer");
+			Check(f.QueueBlocks(replacement, blocks.data(), capacity + 1, 4) == invalid_count,
+			      "oversized waveform request did not reject before CLEAR");
+			Check(f.voice.blocks.size() == capacity && f.voice.accepts_blocks &&
+			          f.voice.sample_phase == phase && f.voice.decoded_samples == 1 &&
+			          f.voice.decoded_bytes == sizeof(int16_t),
+			      "rejected waveform request changed playback or stream continuity");
+			for (uint32_t i = 0; i < capacity; ++i) {
+				Check(f.voice.blocks[i].data == reinterpret_cast<const uint8_t*>(pcm) &&
+				          f.voice.blocks[i].info.user_data == i + 1 &&
+				          f.voice.blocks[i].cursor == (i == 0 ? 1u : 0u),
+				      "rejected waveform request replaced an outstanding block");
+			}
+			f.Render(3);
+			Check(f.voice.samples == std::vector<float> {0.375f, 0.5f,
+			                                            capacity == 1 ? 0.5f : 0.375f} &&
+			          f.voice.blocks.size() == capacity - 1,
+			      "queue rejection changed partially played source audio");
+			Check(f.QueueBlocks(replacement, blocks.data(), 1, 1) == OK &&
+			          f.voice.blocks.size() == capacity,
+			      "completed waveform block did not release its queue slot");
+			const auto progress = f.State();
+			Check(f.QueueBlocks(replacement, blocks.data(), 1, 5) == OK &&
+			          f.voice.blocks.size() == 1 && f.voice.sample_phase == 0 &&
+			          f.voice.decoded_samples == progress.num_decoded_samples &&
+			          f.voice.decoded_bytes == progress.decoded_data_size,
+			      "CLEAR failed to replace a full queue while preserving progress");
+			f.Render(2);
+			Check(f.voice.samples == std::vector<float> {0.75f, 0.4375f},
+			      "CLEAR retained an old source or fractional playback position");
+		}
+	}
+}
+
+struct PcmRing {
+	static constexpr uint32_t block_frames = 7, block_count = 64;
+	Fixture& fixture;
+	std::array<std::array<int16_t, block_frames * 2>, 4> buffers {};
+	uint32_t next_block = 0, completed = 0;
+
+	void QueueNext(uint32_t slot) {
+		auto& buffer = buffers[slot];
+		for (uint32_t i = 0; i < block_frames; ++i) {
+			buffer[2 * i] = static_cast<int16_t>(100 + (next_block * block_frames + i) * 16);
+			buffer[2 * i + 1] = -buffer[2 * i];
+		}
+		const Ngs2WaveformBlock block {0, sizeof(buffer), 0, 0, block_frames, 0, slot};
+		++next_block;
+		Check(fixture.QueueBlocks(buffer.data(), &block, 1, next_block == block_count ? 0 : 1) == OK,
+		      "released ring buffer could not be queued again");
+	}
+};
+
+void KYTY_SYSV_ABI RefillPcmRing(const Ngs2VoiceCallbackInfo* info) {
+	auto& ring = *reinterpret_cast<PcmRing*>(info->data);
+	const auto slot = ring.completed++ % ring.buffers.size();
+	Check(info->flag == 1 && info->attributes == 0 && info->user == slot &&
+	          info->block_data == ring.buffers[slot].data() &&
+	          info->size == sizeof(ring.buffers[slot]),
+	      "ring completion callback released the wrong source buffer");
+	for (const auto& block: ring.fixture.voice.blocks) {
+		Check(block.data != reinterpret_cast<const uint8_t*>(ring.buffers[slot].data()),
+		      "completion callback retained the buffer about to be overwritten");
+	}
+	if (ring.next_block < PcmRing::block_count) {
+		ring.QueueNext(static_cast<uint32_t>(slot));
+	}
+}
+
+void TestWaveformRingReuse() {
+	for (auto rack: {Ngs2RackType::Sampler, Ngs2RackType::CustomSampler}) {
+		Fixture f(48000, 2, rack);
+		PcmRing ring {f};
+		f.voice.callback = reinterpret_cast<uintptr_t>(&RefillPcmRing);
+		f.voice.callback_data = reinterpret_cast<uintptr_t>(&ring);
+		f.voice.callback_flags = 1;
+		for (uint32_t slot = 0; slot < ring.buffers.size(); ++slot) {
+			ring.QueueNext(slot);
+		}
+		constexpr std::array<uint32_t, 6> grains {1, 16, 3, 29, 2, 11};
+		constexpr auto total_frames = PcmRing::block_frames * PcmRing::block_count;
+		uint32_t rendered = 0, grain_index = 0;
+		while (rendered < total_frames) {
+			const auto grain = std::min(grains[grain_index++ % grains.size()], total_frames - rendered);
+			f.Render(grain);
+			for (uint32_t i = 0; i < grain; ++i) {
+				const float expected = (100 + (rendered + i) * 16) / 32768.0f;
+				Check(f.voice.samples[i] == expected && f.voice.samples[grain + i] == -expected,
+				      "ring buffer reuse replayed overwritten data or mixed stereo channels");
+			}
+			rendered += grain;
+			const auto decoded = f.State().num_decoded_samples;
+			Check(f.voice.blocks.size() <= ring.buffers.size() &&
+			          decoded >= rendered && decoded - rendered < 256 &&
+			          f.voice.decoded_bytes == decoded * 4 &&
+			          ring.completed == decoded / PcmRing::block_frames,
+			      "ring playback lost queue capacity or decoded progress");
+			if (decoded < total_frames) {
+				const auto slot = (decoded / PcmRing::block_frames) % ring.buffers.size();
+				Check(f.voice.WaveformData() == reinterpret_cast<const uint8_t*>(
+				          ring.buffers[slot].data() + (decoded % PcmRing::block_frames) * 2),
+				      "ring read address did not track the active source buffer");
+			}
+		}
+		Check(f.voice.blocks.empty() && f.voice.state == Ngs2VoicePlayState::Empty &&
+		          ring.completed == PcmRing::block_count,
+		      "finite ring stream did not release all buffers and finish");
+	}
+}
+
+void TestProgressDrivenWaveformRing() {
+	const auto source_sample = [](uint64_t frame) {
+		return static_cast<int16_t>(static_cast<int32_t>((frame * 43) % 32768) - 16384);
+	};
+	for (auto rack: {Ngs2RackType::Sampler, Ngs2RackType::CustomSampler}) {
+		Fixture f(44100, 2, rack);
+		std::array<std::array<int16_t, 1024 * 2>, 4> buffers {};
+		uint64_t submitted = 0, decoded_bytes = 0, rendered = 0;
+		uint32_t slot = 0;
+		const auto refill = [&](uint32_t frames) {
+			auto& buffer = buffers[slot];
+			for (const auto& block: f.voice.blocks) {
+				Check(block.data != reinterpret_cast<const uint8_t*>(buffer.data()),
+				      "decoded progress allowed the guest to overwrite an outstanding ring buffer");
+			}
+			for (uint32_t i = 0; i < frames; ++i) {
+				buffer[i * 2] = source_sample(submitted + i);
+				buffer[i * 2 + 1] = -buffer[i * 2];
+			}
+			const Ngs2WaveformBlock block {0, frames * 4u, 0, 0, frames, 0, slot};
+			Check(f.QueueBlocks(buffer.data(), &block, 1, 1) == OK,
+			      "progress-driven ring exhausted its waveform slots");
+			submitted += frames;
+			slot = (slot + 1) % buffers.size();
+		};
+		refill(1024);
+		for (uint32_t grain = 0; grain < 128; ++grain) {
+			f.Render(256);
+			for (uint32_t i = 0; i < 256; ++i) {
+				const uint64_t phase = (rendered + i) * 44100;
+				const auto frame = phase / 48000;
+				const float fraction = static_cast<float>(phase % 48000) / 48000.0f;
+				const float expected = std::lerp(static_cast<float>(source_sample(frame)),
+				                                 static_cast<float>(source_sample(frame + 1)), fraction) /
+				                       32768.0f;
+				Check(std::abs(f.voice.samples[i] - expected) < 0.00001f &&
+				          std::abs(f.voice.samples[256 + i] + expected) < 0.00001f,
+				      "progress-driven ring replayed overwritten music samples");
+			}
+			rendered += 256;
+			const auto state = f.State();
+			const auto delta = state.decoded_data_size - decoded_bytes;
+			Check(state.decoded_data_size >= decoded_bytes && delta <= sizeof(buffers[0]) &&
+			          delta % 4 == 0 && state.num_decoded_samples * 4 == state.decoded_data_size &&
+			          state.num_decoded_samples <= submitted,
+			      "published sampler progress does not describe decoded source data");
+			decoded_bytes = state.decoded_data_size;
+			if (delta != 0) {
+				refill(static_cast<uint32_t>(delta / 4));
+			}
+		}
+		Check(decoded_bytes > 28 * sizeof(buffers[0]), "ring test did not exercise repeated buffer reuse");
+	}
+}
+
 void TestStreamingResample() {
 	Fixture              f(44100, 2);
 	std::vector<int16_t> pcm(441 * 2);
@@ -178,6 +395,7 @@ void TestAtrac9ResampleAcrossFrames() {
 		f->voice.next_decoded_frame[0] = 1.0f;
 		f->voice.next_decoded_frame[1] = 1.5f;
 		f->voice.next_frame_samples = 2;
+		f->voice.decoded_samples = 4;
 		f->voice.accepts_blocks = false;
 		return f;
 	};
@@ -218,6 +436,7 @@ void TestAtrac9ResampleAfterStarvation() {
 	f.voice.decoded_frame[0] = 0.0f;
 	f.voice.decoded_frame[1] = 0.5f;
 	f.voice.frame_samples    = 2;
+	f.voice.decoded_samples  = 2;
 	f.Render(4);
 	Check(f.voice.samples == std::vector<float> {0.0f, 0.25f, 0.5f, 0.5f} &&
 	          f.voice.decoded_samples == 2,
@@ -230,6 +449,7 @@ void TestAtrac9ResampleAfterStarvation() {
 	f.voice.next_decoded_frame[0] = 1.0f;
 	f.voice.next_decoded_frame[1] = 1.5f;
 	f.voice.next_frame_samples    = 2;
+	f.voice.decoded_samples       += 2;
 	f.Render(4);
 	Check(f.voice.samples == std::vector<float> {1.0f, 1.25f, 1.5f, 1.5f} &&
 	          f.voice.decoded_samples == 4,
@@ -773,8 +993,8 @@ void TestLoopCallbacksAndExit() {
 	Ngs2VoiceParamHeader exit_loop {sizeof(Ngs2VoiceParamHeader), 0, 0x10000004};
 	Ngs2VoiceControl(reinterpret_cast<uintptr_t>(&active.voice), &exit_loop);
 	active.voice.SetEvent(32);
-	active.Render(3);
-	Check(active.voice.samples == std::vector<float> {0.25f, 0.5f, 0.75f} &&
+	active.Render(5);
+	Check(active.voice.samples == std::vector<float> {0.25f, 0.5f, 0.25f, 0.5f, 0.75f} &&
 	          active.voice.blocks.empty() && log.events.size() == 1,
 	      "exiting infinite loop lost its final iteration or ignored callback flags");
 
@@ -858,8 +1078,69 @@ void TestAtrac9Looping() {
 	      "compressed loop exit failed to finish the active iteration");
 }
 
+void TestAtrac9WaveformJoining() {
+	std::array<uint8_t, 192> encoded {};
+	for (size_t i = 0; i < 4; ++i) {
+		std::copy(ATRAC9_MONO_BLOCK.begin(), ATRAC9_MONO_BLOCK.end(),
+		          encoded.begin() + i * ATRAC9_MONO_BLOCK.size());
+		if (i != 0) encoded[i * ATRAC9_MONO_BLOCK.size()] |= 0x80;
+	}
+	for (auto rack: {Ngs2RackType::Sampler, Ngs2RackType::CustomSampler}) {
+		Fixture reference(24000, 1, rack);
+		reference.voice.SetupSampler({NGS2_WAVEFORM_TYPE_ATRAC9, 1, 24000, 0xfe4005f0});
+		reference.voice.state = Ngs2VoicePlayState::Playing;
+		const Ngs2WaveformBlock whole {0, encoded.size(), 0, 0, 512, 0, 123};
+		Check(reference.QueueBlocks(encoded.data(), &whole, 1, 0) == OK,
+		      "compressed reference block could not be queued");
+		reference.Render(1024);
+		for (bool nojoin: {false, true}) {
+			Fixture f(24000, 1, rack);
+			f.SetCapacity(1);
+			f.voice.SetupSampler({NGS2_WAVEFORM_TYPE_ATRAC9, 1, 24000, 0xfe4005f0});
+			f.voice.state = Ngs2VoicePlayState::Playing;
+			LoopCallbacks callbacks;
+			f.voice.callback = reinterpret_cast<uintptr_t>(&RecordLoopCallback);
+			f.voice.callback_data = reinterpret_cast<uintptr_t>(&callbacks);
+			f.voice.callback_flags = 1;
+			const Ngs2WaveformBlock first {0, 36, 0, 0, 512, 0, 123};
+			const Ngs2WaveformBlock rest {36, encoded.size() - 36, 99, 99, 99, 0, 123};
+			Check(f.QueueBlocks(encoded.data(), &first, 1, 1) == OK,
+			      "compressed stream start could not be queued");
+			const uint32_t flags = nojoin ? 0x12 : 2;
+			if (nojoin) {
+				Check(f.QueueBlocks(encoded.data(), &rest, 1, flags) ==
+				          static_cast<int32_t>(0x804a8054u) &&
+				          f.voice.blocks.size() == 1 && f.voice.blocks.front().info.data_size == 36 &&
+				          f.voice.accepts_blocks,
+				      "NOJOINT bypassed capacity or changed the full stream");
+				f.SetCapacity(2);
+			}
+			Check(f.QueueBlocks(encoded.data(), &rest, 1, flags) == OK &&
+			          f.voice.blocks.size() == (nojoin ? 2u : 1u),
+			      "consecutive ONLY_DATA blocks did not follow the joining flag");
+			f.Render(1024);
+			Check(f.voice.samples == reference.voice.samples && f.voice.decoded_samples == 512 &&
+			          f.voice.decoded_bytes == encoded.size() && f.voice.blocks.empty() &&
+			          callbacks.events.size() == (nojoin ? 2u : 1u),
+			      "compressed joining changed audio, progress, or block completion callbacks");
+		}
+		Fixture full(24000, 1, rack);
+		full.SetCapacity(2);
+		full.voice.SetupSampler({NGS2_WAVEFORM_TYPE_ATRAC9, 1, 24000, 0xfe4005f0});
+		const Ngs2WaveformBlock initial[] = {{0, 36, 0, 0, 512}, {36, 36, 0, 0, 512}};
+		const Ngs2WaveformBlock addition[] = {{72, 36}, {144, 48}};
+		Check(full.QueueBlocks(encoded.data(), initial, 2, 1) == OK &&
+		          full.QueueBlocks(encoded.data(), addition, 2, 2) ==
+		              static_cast<int32_t>(0x804a8054u) &&
+		          full.voice.blocks.size() == 2 && full.voice.blocks.back().info.data_size == 36 &&
+		          full.voice.accepts_blocks,
+		      "overflow after a tentative join partially extended or closed the stream");
+	}
+}
+
 void TestWaveformReadAddress() {
 	Fixture       f(48000, 2);
+	f.system.option.num_grain_samples = 1;
 	const int16_t pcm[] = {-100, -100, 8192, 4096, 16384, 8192};
 	f.Queue(pcm, 2, 1, 1, 1);
 	const auto next_data = [&]() {
@@ -887,11 +1168,12 @@ struct CallbackEvent {
 	Ngs2VoiceInternal* voice;
 	uint32_t           event;
 	uint32_t           calls = 0;
+	uint32_t           trigger = 1;
 };
 
 void KYTY_SYSV_ABI ApplyCallbackEvent(const Ngs2VoiceCallbackInfo* info) {
 	auto& context = *reinterpret_cast<CallbackEvent*>(info->data);
-	if (++context.calls == 1) {
+	if (++context.calls == context.trigger) {
 		context.voice->SetEvent(context.event);
 	}
 }
@@ -910,14 +1192,15 @@ void TestCallbackStopsConsumption() {
 			f.Render(2);
 			Check(f.voice.state != Ngs2VoicePlayState::Playing && context.calls == 1,
 			      "completion callback failed to interrupt playback");
-			Check(f.voice.samples[0] == 0.25f && f.voice.samples[1] == 0.0f &&
+			Check(f.voice.samples[0] == 0.0f && f.voice.samples[1] == 0.0f &&
 			          f.voice.decoded_samples == 1 && f.voice.blocks.size() == 1 &&
 			          f.voice.blocks.front().cursor == 0,
 			      "callback pause/stop consumed the next queued block");
 			if (event == 16) {
 				f.voice.SetEvent(32);
-				f.Render(1);
-				Check(f.voice.samples[0] == 0.5f && f.voice.decoded_samples == 2,
+				f.Render(2);
+				Check(f.voice.samples == std::vector<float> {0.25f, 0.5f} &&
+				          f.voice.decoded_samples == 2,
 				      "callback pause failed to preserve queued audio for resume");
 			}
 		}
@@ -936,15 +1219,95 @@ void TestCallbackPauseDuringLargeStep() {
 	f.voice.callback_data  = reinterpret_cast<uintptr_t>(&context);
 	f.voice.callback_flags = 1;
 	f.Render(2);
-	Check(context.calls == 1 && f.voice.decoded_samples == 1 &&
-	          f.voice.sample_phase == (uint64_t(72000) << 32u),
+	Check(context.calls == 1 && f.voice.decoded_samples == 1 && f.voice.sample_phase == 0 &&
+	          !f.voice.has_samples,
 	      "pause advanced across additional blocks in a large resampling step");
 	f.voice.SetEvent(32);
-	f.Render(1);
-	Check(std::abs(f.voice.samples[0] - 0.4375f) < 1e-7f && context.calls == 3,
+	f.Render(2);
+	Check(f.voice.samples[0] == 0.25f && std::abs(f.voice.samples[1] - 0.4375f) < 1e-7f &&
+	          context.calls == 3,
 	      "resume lost fractional position or read past the next short block");
 }
 
+void TestCallbackPauseDuringLookahead() {
+	for (auto rack: {Ngs2RackType::Sampler, Ngs2RackType::CustomSampler}) {
+		Fixture f(24000, 1, rack);
+		f.system.option.num_grain_samples = 1;
+		int16_t pcm[] = {8192, 16384};
+		f.Queue(pcm, 1, 1);
+		f.Queue(pcm + 1, 1, 0);
+		CallbackEvent context {&f.voice, 16, 0, 2};
+		f.voice.callback = reinterpret_cast<uintptr_t>(&ApplyCallbackEvent);
+		f.voice.callback_data = reinterpret_cast<uintptr_t>(&context);
+		f.voice.callback_flags = 1;
+		f.Render(1);
+		Check(f.voice.samples[0] == 0.25f, "initial source sample was not rendered");
+		f.Render(2);
+		Check(!f.voice.has_samples && f.voice.state == Ngs2VoicePlayState::Paused &&
+		          context.calls == 2 && f.State().num_decoded_samples == 2 &&
+		          f.voice.sample_phase == (uint64_t(24000) << 32u),
+		      "lookahead callback pause consumed output or lost decoded progress");
+		pcm[0] = pcm[1] = -32768;
+		f.voice.SetEvent(32);
+		f.Render(3);
+		Check(f.voice.samples == std::vector<float> {0.375f, 0.5f, 0.5f} &&
+		          f.voice.state == Ngs2VoicePlayState::Empty,
+		      "lookahead pause failed to preserve owned samples or fractional playback position");
+	}
+}
+
+struct CallbackReplacement {
+	Fixture& fixture;
+	bool setup;
+	uint32_t trigger, calls = 0;
+	const std::array<int16_t, 4> pcm {16384, 8192, 4096, 2048};
+};
+
+void KYTY_SYSV_ABI ReplaceCallbackSource(const Ngs2VoiceCallbackInfo* info) {
+	auto& replacement = *reinterpret_cast<CallbackReplacement*>(info->data);
+	if (++replacement.calls != replacement.trigger) {
+		return;
+	}
+	auto& f = replacement.fixture;
+	f.voice.callback_flags = 0;
+	if (replacement.setup) {
+		f.voice.SetupSampler({0x12, 1, 48000});
+	}
+	f.Queue(replacement.pcm.data(), 2, replacement.setup ? 0 : 4);
+	f.voice.SetEvent(1);
+}
+
+void TestCallbackReplacesSource() {
+	for (auto rack: {Ngs2RackType::Sampler, Ngs2RackType::CustomSampler}) {
+		for (bool setup: {false, true}) {
+			for (uint32_t trigger: {1u, 2u}) {
+				Fixture f(24000, 2, rack);
+				f.system.option.num_grain_samples = 1;
+				CallbackReplacement replacement {f, setup, trigger};
+				const int16_t pcm[] = {8192, 4096};
+				for (uint32_t i = 0; i < trigger; ++i) {
+					f.Queue(pcm, 1, 1);
+				}
+				f.voice.callback = reinterpret_cast<uintptr_t>(&ReplaceCallbackSource);
+				f.voice.callback_data = reinterpret_cast<uintptr_t>(&replacement);
+				f.voice.callback_flags = 1;
+				f.Render(4);
+				Check(!f.voice.has_samples && f.voice.samples.size() == (setup ? 4u : 8u) &&
+				          std::ranges::all_of(f.voice.samples, [](float value) { return value == 0; }) &&
+				          f.voice.blocks.size() == 1 && f.voice.blocks.front().cursor == 0,
+				      "callback source replacement used an old frame or channel layout");
+				f.Render(4);
+				const std::vector<float> expected = setup
+				    ? std::vector<float> {0.5f, 0.25f, 0, 0}
+				    : std::vector<float> {0.5f, 0.3125f, 0.125f, 0.125f,
+				                          0.25f, 0.15625f, 0.0625f, 0.0625f};
+				Check(f.voice.samples == expected && f.voice.state == Ngs2VoicePlayState::Empty,
+				      "callback replacement failed to play the new source on the next render");
+			}
+		}
+	}
+}
+
 void TestUpstreamCustomSamplerControls() {
 	Fixture f(24000, 1, Ngs2RackType::CustomSampler);
 	const int16_t pcm[] = {8192, 16384, 24576, 4096};
@@ -953,7 +1316,7 @@ void TestUpstreamCustomSamplerControls() {
 	Check(f.voice.samples[0] == 0.25f, "custom sampler source rate was rejected");
 	f.Queue(pcm + 2, 2, 4);
 	f.Render(1);
-	Check(f.voice.samples[0] == 0.75f && f.voice.blocks.size() == 1,
+	Check(f.voice.samples[0] == 0.75f && f.voice.blocks.empty() && f.voice.frame_samples == 2,
 	      "replacement blocks retained old data or fractional position");
 	struct Pitch {
 		Ngs2VoiceParamHeader header {16, 0, 0x40010005};
@@ -1058,6 +1421,9 @@ void TestFiniteFilterTail() {
 } // namespace
 
 int main() {
+	TestProgressDrivenWaveformRing();
+	TestWaveformQueueCapacity();
+	TestWaveformRingReuse();
 	TestAtrac9Vibration();
 	TestPcm16RenderOutput();
 	TestSubmixerSetupAndRouting();
@@ -1069,6 +1435,7 @@ int main() {
 	TestDeferredSamplerSetup();
 	TestLoopCallbacksAndExit();
 	TestAtrac9Looping();
+	TestAtrac9WaveformJoining();
 	TestWaveformReadAddress();
 	TestFilterTailDuringStarvation();
 	TestFiniteFilterTail();
@@ -1079,6 +1446,8 @@ int main() {
 	TestUpstreamCustomSamplerControls();
 	TestCallbackStopsConsumption();
 	TestCallbackPauseDuringLargeStep();
+	TestCallbackPauseDuringLookahead();
+	TestCallbackReplacesSource();
 	TestStreamingResample();
 	TestPitchLoopAndSkip();
 	TestPitchClamps();

tests/ResourceMaterializationTests.cpp

@@ -168,6 +168,117 @@ void TestMappedSrtUsesDirectReaderByDefault() {
         "cache rematerialization did not use the direct reader by default");
 }
 
+void TestNativeNestedSrtRefresh() {
+  using namespace Libs::Graphics::ShaderRecompiler::IR;
+  std::array<uint32_t, 3> first{0u, 0x12345678u, 0xabcdef01u};
+  std::array<uint32_t, 3> second{0u, 0x87654321u, 0x10fedcbau};
+  const auto pointer_words = [](const void *pointer) {
+    const auto address = reinterpret_cast<uint64_t>(pointer);
+    return std::array<uint32_t, 2>{static_cast<uint32_t>(address) | 3u,
+                                  static_cast<uint32_t>(address >> 32u) | 0xabcd0000u};
+  };
+  auto table = pointer_words(first.data());
+  const auto user_data = pointer_words(table.data());
+  auto program = SrtProgram(0u);
+  program.user_data_base = 4;
+  program.user_data_count = 2;
+  auto &block = *program.blocks[0];
+  const auto emit = [&](ValueOpcode opcode, std::initializer_list<Value> args) {
+    return Value(&block.AppendNewInst(opcode, args));
+  };
+  const auto low = emit(ValueOpcode::GetUserData, {Value(ScalarReg(4))});
+  const auto high = emit(ValueOpcode::GetUserData, {Value(ScalarReg(5))});
+  const auto root = emit(ValueOpcode::GetAddressResource, {low, high});
+  const auto read = [&](Value address, uint32_t offset) {
+    const auto value = emit(ValueOpcode::LoadAddressU32,
+                            {address, Value(offset), Value(0u), Value(true)});
+    value.Instruction()->SetFlags(MemoryFlags{.index = 0});
+    program.srt_reads.push_back(
+        {value, static_cast<uint32_t>(program.srt_reads.size())});
+    return value;
+  };
+  const auto pointer_low = read(root, 0u);
+  const auto pointer_high = read(root, 4u);
+  auto *leaf = program.srt_reads[0].value.Instruction();
+  leaf->Arg(0).Instruction()->SetArg(0, pointer_low);
+  leaf->Arg(0).Instruction()->SetArg(1, pointer_high);
+  leaf->SetArg(1, Value(7u));
+  read(leaf->Arg(0), 11u);
+  auto plan = ExtractResourcePlan(program);
+  Check(plan.srt_code != nullptr, "nested user-data tree did not compile an x86 walker");
+  ResourceSnapshot snapshot;
+  ResourceSpecialization specialization;
+  const SrtRuntime runtime{.user_data = user_data};
+  {
+    SrtWalker walker(plan, runtime);
+    Check(walker.RefreshFlatBuffer(snapshot.flattened_srt),
+          "native SRT refresh failed");
+    const auto loaded = first[1];
+    first[1] ^= 0xffffffffu;
+    uint32_t value = 0;
+    Check(walker.Evaluate(plan.srt_reads[0].value, value) && value == loaded,
+          "descriptor evaluation reloaded a value already copied by the native walker");
+    first[1] = loaded;
+  }
+  for (const auto *data : {&first, &second}) {
+    table = pointer_words(data->data());
+    Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
+              snapshot.flattened_srt ==
+                  std::vector<uint32_t>{(*data)[1], table[0], table[1], (*data)[2]},
+          "native SRT refresh lost pointer masking, alignment, or current values");
+  }
+  Check(!MaterializeResources(plan, {.user_data = std::span(user_data).first(1)},
+                              snapshot, specialization),
+        "native SRT walk accepted truncated user data");
+  const std::array<uint32_t, 2> fake_user_data{0x1003u, 0xabcd0000u};
+  uint32_t reads = 0;
+  const SrtRuntime callback_runtime{
+      .user_data = fake_user_data,
+      .read_memory = +[](void *data, uint64_t address, std::span<uint32_t> words) {
+        ++*static_cast<uint32_t *>(data);
+        switch (address) {
+          case 0x1000u: words[0] = 0x2003u; break;
+          case 0x1004u: words[0] = 0xabcd0000u; break;
+          case 0x2004u: words[0] = 0x1234u; break;
+          case 0x2008u: words[0] = 0x5678u; break;
+          default: return false;
+        }
+        return true;
+      },
+      .userdata = &reads};
+  Check(MaterializeResources(plan, callback_runtime, snapshot, specialization) && reads == 4 &&
+            snapshot.flattened_srt ==
+                std::vector<uint32_t>{0x1234u, 0x2003u, 0xabcd0000u, 0x5678u},
+        "native-eligible SRT bypassed its callback or repeated nested pointer reads");
+  leaf->SetArg(0, root);
+  leaf->SetArg(1, Value(0u));
+  program.srt_reads.resize(1);
+  for (const int32_t offset : {-1, 7}) {
+    program.memory_info[0].offset = static_cast<uint32_t>(offset);
+    auto bounded = ExtractResourcePlan(program);
+    auto base = pointer_words(second.data() + 1);
+    Check(bounded.srt_code != nullptr &&
+              MaterializeResources(bounded, {.user_data = base}, snapshot, specialization) &&
+              snapshot.flattened_srt[0] == second[offset < 0 ? 0 : 2],
+          "native SRT applied a signed immediate offset incorrectly");
+    base = offset < 0 ? std::array<uint32_t, 2>{0u, 0u}
+                      : std::array<uint32_t, 2>{0xfffffffcu, 0xffffu};
+    Check(!MaterializeResources(bounded, {.user_data = base}, snapshot, specialization),
+          "native SRT accepted address underflow or 48-bit overflow");
+  }
+  high.Instruction()->SetArg(0, Value(ScalarReg(6)));
+  Check(ExtractResourcePlan(program).srt_code == nullptr,
+        "native SRT accepted nonadjacent user-data pointer words");
+  high.Instruction()->SetArg(0, Value(ScalarReg(5)));
+  program.memory_info[0].offset = 0;
+  program.srt_reads.push_back({pointer_low, 1});
+  program.srt_reads.push_back({pointer_high, 2});
+  leaf->SetArg(0, emit(ValueOpcode::GetAddressResource, {pointer_low, pointer_high}));
+  pointer_high.Instruction()->SetArg(1, Value(8u));
+  Check(ExtractResourcePlan(program).srt_code == nullptr,
+        "native SRT accepted nonadjacent nested pointer words");
+}
+
 void TestIntegerRuntimeValueFollowsSrtReads() {
   using namespace Libs::Graphics::ShaderRecompiler::IR;
   auto plan = SrtProgram(0x10000);
@@ -256,6 +367,8 @@ void TestSrtAliasesRetainReadPolicy() {
             plan.srt_reads[1].value.Instruction()->Flags<SrtReadFlags>().clean == 0u &&
             plan.srt_reads[2].value.Instruction()->Flags<SrtReadFlags>().clean == 1u,
         "alias normalization lost read identity or retained discarded EXEC provenance");
+  Check(plan.srt_code == nullptr,
+        "captured strict reads compiled an unused x86 walker");
   struct Reads { uint32_t strict = 0; uint32_t ordinary = 0; bool dirty = false; } reads;
   const SrtRuntime runtime{
       .read_memory = +[](void *data, uint64_t address, std::span<uint32_t> words) {
@@ -527,6 +640,64 @@ void TestFailedMaterializationRejectsStage() {
         "missing runtime user data did not reject the cached stage");
 }
 
+void TestSharedControlPredicates() {
+  using namespace Libs::Graphics::ShaderRecompiler::IR;
+  using Libs::Graphics::ShaderRecompiler::CFG::TerminatorKind;
+  constexpr uint32_t depth = 8;
+  for (const auto operation : {ValueOpcode::LogicalAnd, ValueOpcode::LogicalOr}) {
+    Program program;
+    program.user_data_count = 1;
+    program.srt_plan_complete = program.resource_tracking_complete = true;
+    auto &block = AddValueBlock(program);
+    const auto emit = [&](ValueOpcode opcode, std::initializer_list<Value> args) {
+      return Value(&block.AppendNewInst(opcode, args));
+    };
+    const auto user = emit(ValueOpcode::GetUserData, {Value(ScalarReg(0))});
+    const auto uniform = emit(ValueOpcode::INotEqual32, {user, Value(0u)});
+    const auto lane = emit(ValueOpcode::GetBuiltin,
+        {Value(static_cast<uint32_t>(StageInputKind::LocalInvocationId)), Value(0u)});
+    auto varying = emit(ValueOpcode::INotEqual32, {lane, Value(0u)});
+    // Both failed and partially uniform subgraphs share their operands exponentially.
+    for (uint32_t i = 0; i < depth; ++i)
+      varying = emit(ValueOpcode::LogicalOr, {varying, varying});
+    auto condition = emit(operation, {varying, uniform});
+    for (uint32_t i = 0; i < depth; ++i)
+      condition = emit(ValueOpcode::LogicalOr, {condition, condition});
+    condition = emit(ValueOpcode::LogicalNot, {condition});
+    condition = emit(ValueOpcode::ConditionRef, {condition});
+    for (uint32_t i = 0; i < 3; ++i) AddValueBlock(program);
+    for (uint32_t i = 0; i < 2; ++i) {
+      program.blocks[i]->condition = condition;
+      program.blocks[i]->terminator.kind = TerminatorKind::ConditionalBranch;
+      program.blocks[i]->terminator.true_block = program.blocks[i == 0 ? 1 : 3];
+      program.blocks[i]->terminator.false_block = program.blocks[2];
+    }
+    auto plan = ExtractResourcePlan(program);
+    Check(plan.value_storage.size() == depth + 5 &&
+              plan.control_flow.size() == 4 &&
+              plan.control_flow[0].condition == plan.control_flow[1].condition &&
+              !plan.capture_specialization_reads,
+          "shared predicates expanded into trees or retained varying dependencies");
+    for (const uint32_t flag : {0u, 1u}) {
+      const std::array<uint32_t, 1> user_data{flag};
+      SrtWalker walker(plan, {.user_data = user_data});
+      uint32_t result = 99;
+      const bool conjunction = operation == ValueOpcode::LogicalAnd;
+      const bool known = conjunction ? flag == 0 : flag != 0;
+      Check(walker.Evaluate(plan.control_flow[0].condition, result) == known &&
+                (!known || result == uint32_t{conjunction}),
+            "shared predicate changed its uniform factor or cached runtime values");
+      std::vector<uint32_t> flat;
+      Check(walker.RefreshFlatBuffer(flat) &&
+                plan.visited_blocks == (known
+                    ? (conjunction ? std::vector<uint8_t>{1, 1, 0, 1}
+                                   : std::vector<uint8_t>{1, 0, 1, 0})
+                    : std::vector<uint8_t>{1, 1, 1, 1}),
+            "shared predicate skipped an unknown branch or followed a rejected branch");
+    }
+  }
+}
+
 void TestFiniteImageRefreshReusesScalarReads() {
   using namespace Libs::Graphics::ShaderRecompiler::IR;
   Program program;
@@ -671,6 +842,7 @@ void DbgExit(int) { std::abort(); }
 
 int main() {
   TestMappedSrtUsesDirectReaderByDefault();
+  TestNativeNestedSrtRefresh();
   TestIntegerRuntimeValueFollowsSrtReads();
   TestSrtAliasesRetainReadPolicy();
   TestUniformVectorDescriptorRead();
@@ -679,6 +851,7 @@ int main() {
   for (const bool memory_condition : {false, true})
     TestWrittenDescriptorPredicateReads(memory_condition);
   TestFailedMaterializationRejectsStage();
+  TestSharedControlPredicates();
   TestFiniteImageRefreshReusesScalarReads();
   TestMixedSamplerVariantsShareRuntimeDescriptor();
   std::puts("ResourceMaterializationTests: all cases passed");

tests/ResourceTrackingTests.cpp

@@ -270,7 +270,7 @@ MakeIndirectImageFixture(bool malformed, uint32_t table_stride = 32, uint32_t ta
   sample.kind = ResourceKind::Image;
   sample.image_dimension = Decoder::ImageDimension::Dim2D;
   const auto sampled = fixture->Emit(ValueOpcode::ImageSampleRaw,
-                                     {image, sampler, fixture->ImageAddress()},
+                                     {image, sampler, fixture->ImageAddress(), Value(true)},
                                      fixture->AddMemory(sample, 0x10f0));
   const auto sampled_x =
       fixture->Emit(ValueOpcode::CompositeExtractU32x4, {sampled, Value(0u)});
@@ -486,7 +486,8 @@ void TestWaterfallImageTable() {
       memory.kind = ResourceKind::Image;
       memory.image_dimension = Decoder::ImageDimension::Dim2D;
       const auto value = fixture->Emit(ValueOpcode::ImageSampleRaw,
-          {image, sampler, fixture->ImageAddress()}, fixture->AddMemory(memory, 0x92cu), block);
+          {image, sampler, fixture->ImageAddress(), block == entry ? active : guard},
+          fixture->AddMemory(memory, 0x92cu), block);
       fixture->Emit(ValueOpcode::ReferenceU32,
           {fixture->Emit(ValueOpcode::CompositeExtractU32x4, {value, Value(0u)}, 0, block)}, 0, block);
     };
@@ -641,7 +642,7 @@ void TestGuardedDirectImageTable() {
     MemoryInfo memory;
     memory.kind = ResourceKind::Image;
     memory.image_dimension = Decoder::ImageDimension::Dim2D;
-    fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, fixture.ImageAddress()},
+    fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, fixture.ImageAddress(), Value(true)},
                  fixture.AddMemory(memory, 0x128));
     fixture.PlanAndTrack();
     const auto source = fixture.program.info.images[0].source;
@@ -989,7 +990,7 @@ void TestBoundedComputeImageLoop() {
     sample.kind = ResourceKind::Image;
     sample.image_dimension = Decoder::ImageDimension::Dim2D;
     fixture.Emit(ValueOpcode::ImageSampleRaw,
-                 {image, sampler, fixture.ImageAddress()},
+                 {image, sampler, fixture.ImageAddress(), Value(true)},
                  fixture.AddMemory(sample, 0x29c));
     fixture.PlanAndTrack();
     const auto source = fixture.program.info.images[0].source;
@@ -1325,7 +1326,9 @@ void TestUniformizedMaterialImageKeys() {
     MemoryInfo image_memory;
     image_memory.kind = ResourceKind::Image;
     image_memory.image_dimension = Decoder::ImageDimension::Dim2D;
-    const auto sampled = fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, address},
+    const auto sampled = fixture.Emit(ValueOpcode::ImageSampleRaw,
+        {image, sampler, address,
+         variant == Variant::EmptyFirstEntry ? material_guard : sample_guard},
         fixture.AddMemory(image_memory, 0x1aecu), sample);
     if (first_lane) {
       const auto component = fixture.Emit(
@@ -1477,7 +1480,7 @@ void TestImageDescriptorFields() {
     memory.kind = ResourceKind::Image;
     memory.image_dimension = Decoder::ImageDimension::Dim2D;
     memory.image_r128 = r128;
-    fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, fixture.ImageAddress()},
+    fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, fixture.ImageAddress(), Value(true)},
                  fixture.AddMemory(memory, 0x80));
     fixture.PlanAndTrack();
     auto plan = ExtractResourcePlan(fixture.program);
@@ -1571,7 +1574,7 @@ void TestUniformScalarBufferImage() {
   sample.kind = ResourceKind::Image;
   sample.image_dimension = Decoder::ImageDimension::Dim2D;
   fixture.Emit(ValueOpcode::ImageSampleRaw,
-                {image, sampler, fixture.ImageAddress()}, fixture.AddMemory(sample, 0x228));
+                {image, sampler, fixture.ImageAddress(), Value(true)}, fixture.AddMemory(sample, 0x228));
   fixture.PlanAndTrack();
   const auto plan = ExtractResourcePlan(fixture.program);
 
@@ -1866,7 +1869,7 @@ void TestImagesSamplersAndAliases() {
     memory.kind = ResourceKind::Image;
     memory.image_dimension = Decoder::ImageDimension::Dim2D;
     memory.image_sample_flags = sample_flags;
-    fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, image_address},
+    fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, image_address, Value(true)},
                  fixture.AddMemory(memory, pc));
     return std::pair{image, sampler};
   };
@@ -1938,14 +1941,14 @@ void TestSampleAdjustSamplerScratch() {
   memory.image_dimension = Decoder::ImageDimension::Dim2D;
   memory.image_sample_flags = Decoder::ImageSampleFlagAdjust;
   fixture.Emit(ValueOpcode::ImageSampleRaw,
-               {image, sampler, fixture.ImageAddress()},
+               {image, sampler, fixture.ImageAddress(), Value(true)},
                fixture.AddMemory(memory, 0x1ec));
   // PPSA24156's zero border payload leaves the reserved scratch shift as
   // the entire DWORD after the frontend folds its OR with zero.
   const auto shift_sampler = fixture.Sampler(
       {Value(0x36u), Value(0xfff000u), Value(0x02500000u), scratch}, 0x200);
   fixture.Emit(ValueOpcode::ImageSampleRaw,
-               {image, shift_sampler, fixture.ImageAddress()},
+               {image, shift_sampler, fixture.ImageAddress(), Value(true)},
                fixture.AddMemory(memory, 0x200));
   fixture.PlanAndTrack();
   auto plan = ExtractResourcePlan(fixture.program);
@@ -1994,7 +1997,7 @@ void TestSampleAdjustSamplerScratch() {
     rejected_memory.image_dimension = Decoder::ImageDimension::Dim2D;
     rejected_memory.image_sample_flags = flags;
     rejected.Emit(ValueOpcode::ImageSampleRaw,
-                  {rejected_image, rejected_sampler, rejected.ImageAddress()},
+                  {rejected_image, rejected_sampler, rejected.ImageAddress(), Value(true)},
                   rejected.AddMemory(rejected_memory, 0x200));
 
     CheckFatal([&] { rejected.PlanAndTrack(); },
@@ -2112,7 +2115,7 @@ void TestGatherLodSamplerValidation() {
         fixture.Emit(ValueOpcode::ImageGatherRaw,
                      {fixture.Image(image_words), fixture.Sampler(sampler_words),
                       fixture.ImageAddress(), image_words[1], image_words[3],
-                      sampler_words[1], sampler_words[2]},
+                      sampler_words[1], sampler_words[2], Value(true)},
                      fixture.AddMemory(memory, i * 4));
       }
       fixture.PlanAndTrack();
@@ -2514,7 +2517,7 @@ ResourcePlan ConditionalSamplerPlan(bool diamond, bool reverse, bool reverse_phi
     sample.kind = ResourceKind::Image;
     sample.image_dimension = Decoder::ImageDimension::Dim2D;
     const auto result =
-        fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, address},
+        fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, address, Value(true)},
                      fixture.AddMemory(sample, 0x4d4));
     fixture.Emit(ValueOpcode::ReferenceU32,
                  {fixture.Emit(ValueOpcode::CompositeExtractU32x4,
@@ -2619,7 +2622,7 @@ void TestFiniteImagePhiCycle() {
   MemoryInfo memory;
   memory.kind = ResourceKind::Image;
   memory.image_dimension = Decoder::ImageDimension::Dim2D;
-  fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, fixture.ImageAddress()},
+  fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, fixture.ImageAddress(), Value(true)},
                fixture.AddMemory(memory, 4));
   fixture.PlanAndTrack();
   auto plan = ExtractResourcePlan(fixture.program);
@@ -2685,12 +2688,12 @@ void TestFiniteInlineSampler() {
     memory.kind = ResourceKind::Image;
     memory.image_dimension = Decoder::ImageDimension::Dim2D;
     if (format == F::k32Float)
-      fixture.Emit(ValueOpcode::ImageSampleRaw, {image, previous, fixture.ImageAddress()},
+      fixture.Emit(ValueOpcode::ImageSampleRaw, {image, previous, fixture.ImageAddress(), Value(true)},
                    fixture.AddMemory(memory, 0x440u));
-    fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, fixture.ImageAddress()},
+    fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, fixture.ImageAddress(), Value(true)},
                  fixture.AddMemory(memory, 0x1dd0u));
     if (format == F::k32Float)
-      fixture.Emit(ValueOpcode::ImageSampleRaw, {image, direct, fixture.ImageAddress()},
+      fixture.Emit(ValueOpcode::ImageSampleRaw, {image, direct, fixture.ImageAddress(), Value(true)},
                    fixture.AddMemory(memory, 0x1dd8u));
   }
   fixture.PlanAndTrack();
@@ -2764,7 +2767,7 @@ void TestFiniteInlineSampler() {
     memory.kind = ResourceKind::Image;
     memory.image_dimension = Decoder::ImageDimension::Dim2D;
     memory.image_sample_flags = flags;
-    adjusted.Emit(ValueOpcode::ImageSampleRaw, {image, shared, adjusted.ImageAddress()},
+    adjusted.Emit(ValueOpcode::ImageSampleRaw, {image, shared, adjusted.ImageAddress(), Value(true)},
                   adjusted.AddMemory(memory, 0x1dd0u));
   }
   CheckFatal([&] { adjusted.PlanAndTrack(); }, "not a valid runtime value",
@@ -2817,7 +2820,7 @@ void TestFiniteImageBitScanSentinel() {
     MemoryInfo memory;
     memory.kind = ResourceKind::Image;
     memory.image_dimension = Decoder::ImageDimension::Dim2D;
-    fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, fixture.ImageAddress()},
+    fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, fixture.ImageAddress(), Value(true)},
                  fixture.AddMemory(memory, 4));
     if (!nonzero) {
       CheckFatal([&] { fixture.PlanAndTrack(); }, "not a valid runtime value",

tests/ShaderRecompilerComputeTests.cpp

@@ -159,6 +159,8 @@ struct BufferCacheTestAccess {
   static_assert(std::same_as<decltype(BufferCache::m_slot_buffers),
                              Common::SlotVector<Buffer>>);
 
+  static uint64_t GcTick(const BufferCache &cache) { return cache.m_gc_tick; }
+
   static void SetGarbageCollectionThresholds(BufferCache &cache,
                                              uint64_t trigger,
                                              uint64_t critical) {
@@ -224,6 +226,8 @@ struct TextureCacheTestAccess {
   static_assert(TextureCache::ImagePageTable::kAddressSpaceBits == 44);
   static_assert(TextureCache::ImagePageTable::kFirstLevelBits == 14);
 
+  static uint64_t GcTick(const TextureCache &cache) { return cache.m_gc_tick; }
+
   static std::unique_lock<TrackingSpinLock> Lock(TextureCache &cache) {
     return std::unique_lock(cache.m_lock);
   }
@@ -2614,6 +2618,31 @@ public:
       packet[4] = value;
     };
 
+    const auto cache_ticks = [&] {
+      std::pair<uint64_t, uint64_t> ticks;
+      gpu.SendCommandSync([&] {
+        ticks = {TextureCacheTestAccess::GcTick(context.GetTextureCache()),
+                 BufferCacheTestAccess::GcTick(context.GetBufferCache())};
+      });
+      return ticks;
+    };
+    const auto initial_cache_ticks = cache_ticks();
+    uint32_t graphics_value = 0;
+    uint32_t async_value = 0;
+    std::array<uint32_t, 5> graphics_list{}, compute_list{};
+    write(graphics_list.data(), &graphics_value, 17);
+    write(compute_list.data(), &async_value, 29);
+    // A frame can contain many lists; completing them must not age either cache.
+    for (uint32_t i = 0; i < 32; ++i) {
+      gpu.Submit(graphics_list, {});
+      gpu.SubmitCompute(0x20, compute_list);
+    }
+    finish();
+    Require(name, "command-list cache lifetime",
+            graphics_value == 17 && async_value == 29 &&
+                cache_ticks() == initial_cache_ticks,
+            "individual graphics or compute lists aged the caches within a frame");
+
     gpu.SendCommandSync([&] {
       setup.BufferInit();
       SubmitInfo blocked;
@@ -2648,6 +2677,10 @@ public:
     parser.join();
     Require(name, "marker insertion", parser_nonblocking,
             "inserting a suspend point blocked the GPU command thread");
+    Require(name, "frame cache maintenance",
+            cache_ticks() == std::pair{initial_cache_ticks.first + 1,
+                                      initial_cache_ticks.second + 1},
+            "a suspend point did not age both caches exactly once");
 
     std::binary_semaphore second_entered{0};
     std::binary_semaphore second_returned{0};
@@ -2685,6 +2718,10 @@ public:
             "or the host command lane");
     Require(name, "frame count", gpu.GetFrameNum() == 2,
             "suspend-point insertion did not advance the frame counter once");
+    Require(name, "async cache lifetime",
+            cache_ticks() == std::pair{initial_cache_ticks.first + 2,
+                                      initial_cache_ticks.second + 2},
+            "asynchronous work aged the caches between frame boundaries");
 
     // A marker must stay behind a stalled graphics stream, then clear its
     // predicate before the next stream. Async compute releases that stream.
@@ -6741,6 +6778,99 @@ public:
               "to the image");
       texture_cache.MarkGpuWritten(mirror_refresh);
 
+      // Unity copies BC mip regions and packed mip tails through integer images.
+      // The CPU allocation remains white; later array sampling must see GPU writes.
+      for (uint32_t format_index = 0; format_index < 3; ++format_index) {
+        constexpr std::array formats{vk::Format::eBc4UnormBlock,
+                                     vk::Format::eBc5UnormBlock,
+                                     vk::Format::eBc7UnormBlock};
+        constexpr std::array guest_formats{Prospero::BufferFormat::kBc4UNorm,
+                                           Prospero::BufferFormat::kBc5UNorm,
+                                           Prospero::BufferFormat::kBc7UNorm};
+        constexpr uint32_t width = 128, levels = 8, layers = 3;
+        const uint32_t block_bytes = format_index == 0 ? 8 : 16;
+        const uint64_t address = base + 0x1900000 + format_index * 0x40000;
+        auto array = MakeLinearDesc(address, 0, formats[format_index],
+            guest_formats[format_index], Prospero::ImageType::kColor2D,
+            {width, width, 1}, layers, block_bytes, 1);
+        array.info.resources.levels = levels;
+        array.info.tile_mode = Prospero::TileMode::kStandard4KB;
+        array.info.UpdateSize();
+        array.view_info.level_count = levels;
+        std::memset(memory + address - base, 0xff, array.info.data.size);
+        auto value_for = [&](uint32_t layer, uint64_t offset) {
+          return 0x7c000000u | (format_index << 16) | (layer << 8) |
+                 static_cast<uint32_t>(offset / 4096);
+        };
+        auto write_raw = [&](uint32_t layer, uint32_t level, uint32_t value) {
+          const auto &mip = array.info.mip_layout[level];
+          auto raw = MakeLinearDesc(
+              address + layer * array.info.tiled_slice_stride + mip.offset, 0,
+              block_bytes == 8 ? vk::Format::eR32G32Uint
+                               : vk::Format::eR32G32B32A32Uint,
+              block_bytes == 8 ? Prospero::BufferFormat::k32_32UInt
+                               : Prospero::BufferFormat::k32_32_32_32UInt,
+              Prospero::ImageType::kColor2D, {mip.pitch, mip.height, 1},
+              1, block_bytes, 1);
+          raw.type = BindingType::Storage;
+          raw.info.tile_mode = array.info.tile_mode;
+          raw.info.UpdateSize();
+          raw.view_info.usage = vk::ImageUsageFlagBits::eStorage;
+          const auto producer = texture_cache.FindImage(raw);
+          (void)texture_cache.FindTexture(producer, raw);
+          vk::ClearValue clear{};
+          std::fill_n(clear.color.uint32.data(), 4, value);
+          TextureCacheTestAccess::ClearImage(texture_cache, scheduler.Current(),
+              producer, {vk::ImageAspectFlagBits::eColor, 0, 1, 0, 1}, clear);
+        };
+        for (const uint32_t layer : {0u, 2u}) {
+          uint64_t previous_offset = UINT64_MAX;
+          for (uint32_t level = 0; level < levels; ++level) {
+            const auto offset = array.info.mip_layout[level].offset;
+            if (offset == previous_offset) continue;
+            previous_offset = offset;
+            write_raw(layer, level, value_for(layer, offset));
+          }
+        }
+        std::vector<vk::BufferImageCopy> probes;
+        std::vector<uint32_t> expected;
+        for (uint32_t layer = 0; layer < layers; ++layer) {
+          for (uint32_t level = 0; level < levels; ++level) {
+            vk::BufferImageCopy copy{};
+            copy.bufferOffset = expected.size() * sizeof(uint32_t);
+            copy.imageSubresource = {vk::ImageAspectFlagBits::eColor, level, layer, 1};
+            copy.imageExtent = {std::min(width >> level, 4u),
+                                std::min(width >> level, 4u), 1};
+            probes.push_back(copy);
+            expected.insert(expected.end(), block_bytes / sizeof(uint32_t),
+                layer == 1 ? UINT32_MAX
+                           : value_for(layer, array.info.mip_layout[level].offset));
+          }
+        }
+        auto verify_array = [&] {
+          const auto consumer = texture_cache.FindImage(array);
+          (void)texture_cache.FindTexture(consumer, array);
+          auto readback = CreateHostBuffer(name, expected.size() * sizeof(uint32_t),
+              vk::BufferUsageFlagBits::eTransferDst, std::vector<uint32_t>(expected.size()));
+          texture_cache.GetImage(consumer).Download(probes, readback.buffer, 0, readback.size);
+          HostReadBarrier(readback.buffer, readback.size, vk::PipelineStageFlagBits::eTransfer,
+                          vk::AccessFlagBits::eTransferWrite);
+          scheduler.Finish();
+          Require(name, "raw image producers populate BC array",
+                  ReadBuffer(name, readback, expected.size()) == expected,
+                  "BC array lost GPU mip/tail writes or overwrote an untouched layer");
+          DestroyBuffer(&readback);
+        };
+        verify_array();
+        constexpr uint32_t replacement = 0x13579bdfu;
+        write_raw(2, array.info.first_tail_level, replacement);
+        for (uint32_t level = array.info.first_tail_level; level < levels; ++level) {
+          const auto word = (2 * levels + level) * block_bytes / sizeof(uint32_t);
+          std::fill_n(expected.begin() + word, block_bytes / sizeof(uint32_t), replacement);
+        }
+        verify_array();
+      }
+
       // A write through a separate buffer must not replace untouched image
       // texels with stale backing bytes. Exact-base writes still refresh it.
       constexpr uint64_t buffer_alias_offset = 0x1a00000;
@@ -11083,6 +11213,78 @@ public:
       context.ShutdownGpu();
       LibKernel::Memory::InstallGpuResources(nullptr);
     }
+    {
+      // PPSA01593 samples a DCC R32 color surface through a comparison sampler.
+      constexpr uint64_t comparison_dcc = base + 0x10000;
+      std::memset(mapped, 0, 0x10000);
+      std::memset(reinterpret_cast<void *>(comparison_dcc), 0xff, 0x1000);
+      RenderContext context(m_runtime_context);
+      context.InitializeGpu(nullptr);
+      LibKernel::Memory::InstallGpuResources(&context);
+      auto &scheduler = context.GetCommandScheduler();
+      HW::Context registers{};
+      HW::UserConfig user_config{};
+      HW::Shader shaders{};
+      scheduler.Begin(registers, user_config, shaders);
+      context.MapMemory(base, allocation_size);
+      auto &cache = context.GetTextureCache();
+      auto &executor = context.GetRenderExecutor();
+      const ShaderTextureResource descriptor{{static_cast<uint32_t>(base >> 8u),
+          0xc1600000u, 0x000fc00fu, 0x91b00204u, 0, 0x00700000u,
+          0x002b0000u | (static_cast<uint32_t>(comparison_dcc >> 8u) << 24u),
+          static_cast<uint32_t>(comparison_dcc >> 16u)}};
+      ShaderRecompiler::IR::DescriptorValue value{};
+      value.dword_count = 8;
+      std::copy_n(descriptor.fields, 8, value.dwords.begin());
+      ShaderRecompiler::IR::ImageResource resource{};
+      resource.resource_class = ShaderRecompiler::IR::ImageResourceClass::Storage;
+      resource.numeric_class = Prospero::TextureNumericClass::Float;
+      resource.dimension = ShaderRecompiler::Decoder::ImageDimension::Dim2D;
+      resource.written = true;
+      const auto color = RenderExecutorTestAccess::ResolveTexture(executor, resource, value);
+      (void)cache.FindTexture(color.image_id, color.desc);
+      vk::ClearValue paint{};
+      paint.color.float32 = std::array{0.375f, 0.f, 0.f, 0.f};
+      TextureCacheTestAccess::ClearImage(cache, scheduler.Current(), color.image_id,
+          {vk::ImageAspectFlagBits::eColor, 0, 1, 0, 1}, paint);
+      resource.resource_class = ShaderRecompiler::IR::ImageResourceClass::Sampled;
+      resource.written = false;
+      resource.read = true;
+      resource.depth_compare = true;
+      auto comparison = RenderExecutorTestAccess::ResolveTexture(executor, resource, value);
+      (void)cache.FindTexture(comparison.image_id, comparison.desc);
+      const auto &depth = cache.GetImage(comparison.image_id);
+      Require(name, "expanded DCC comparison alias",
+              comparison.image_id != color.image_id &&
+                  depth.info.pixel_format == vk::Format::eD32Sfloat &&
+                  depth.info.guest_format == Prospero::BufferFormat::k32Float &&
+                  depth.info.tile_mode == Prospero::TileMode::kRenderTarget &&
+                  depth.info.extent == vk::Extent3D{64, 64, 1} &&
+                  depth.info.metadata.kind == ImageMetadataKind::Dcc &&
+                  depth.info.metadata.range == GuestRange{comparison_dcc, 0x1000} &&
+                  ReadCachedTexel(name, context, comparison.image_id) ==
+                      std::vector<u32>{std::bit_cast<u32>(0.375f)},
+              "comparison promotion lost native color data or guest DCC layout");
+      for (const u32 code : {0x40404040u, 0x80808080u}) {
+        WriteMetadata(context, comparison_dcc, 0x1000, code);
+        comparison = RenderExecutorTestAccess::ResolveTexture(executor, resource, value);
+        (void)cache.FindTexture(comparison.image_id, comparison.desc);
+        Require(name, "DCC clear on a comparison backing",
+                ReadCachedTexel(name, context, comparison.image_id, {63, 63, 0}) ==
+                    std::vector<u32>{code == 0x40404040u ? 0u : 0x3f800000u},
+                "DCC clear used the native depth format or color aspect");
+        std::array<uint8_t, 0x1000> metadata{};
+        Require(name, "comparison clear expands metadata",
+                LibKernel::Memory::TryReadBacking(comparison_dcc, metadata.data(),
+                                                 metadata.size()) &&
+                    std::ranges::all_of(metadata, [](uint8_t byte) { return byte == 0xff; }),
+                "comparison clear retained the native DCC clear key");
+      }
+      context.UnmapMemory(base, allocation_size);
+      scheduler.Finish();
+      context.ShutdownGpu();
+      LibKernel::Memory::InstallGpuResources(nullptr);
+    }
     Require(name, "unmap",
             Libs::LibKernel::Memory::KernelMunmap(base, allocation_size) == 0,
             "captured scene mapping release failed");
@@ -17925,9 +18127,9 @@ public:
                   captured.mip_layout[0].pitch == 512 &&
                   captured.mip_layout[0].height == 512,
               "R64 BC1 texture lost its compressed blocks or array stride");
-      const auto small =
+      const auto small_array =
           MakeTilingImage(format, 67, 51, 1, layers, tile, 0, false);
-      check_round_trip("R64 BC1 array", small.data.size, small);
+      check_round_trip("R64 BC1 array", small_array.data.size, small_array);
     }
 
     {
@@ -34888,12 +35090,11 @@ TestCase ImageSampleOpcodeAliasUsesNormalCoords() {
   return test;
 }
 
-TestCase ImageSampleA16OffsetKeepsTexelOffset32BitOnGpu() {
+template <u32 wave_size> TestCase ImageSampleA16OffsetKeepsTexelOffset32BitOnGpu() {
   using O = ShaderOpcode;
 
   std::vector<u32> code;
-  AppendVMovU32(&code, 20,
-                1); // Non-constant +1 X offset is not a SPIR-V ConstOffset.
+  AppendVMovU32(&code, 20, 1); // +1 X remains a full dword with A16 coordinates.
   AppendVMovLiteral(&code, 21, 0x36003900u); // x=0.625, y=0.375 packed as f16.
   AppendVMovU32(&code, 22, 0);
   code.push_back(EncodeMimg0(0x30, 0xf));
@@ -34904,18 +35105,63 @@ TestCase ImageSampleA16OffsetKeepsTexelOffset32BitOnGpu() {
   AppendEnd(&code);
 
   auto image = MakeRgbaImage(4, 4);
-  SetRgbaPixel(&image, 4, 2, 1, 0x3f800000u, 0x40000000u, 0x40400000u,
+  SetRgbaPixel(&image, 4, 2, 1, 0x3f000000u, 0, 0, 0);
+  SetRgbaPixel(&image, 4, 3, 1, 0x3f800000u, 0x40000000u, 0x40400000u,
                0x40800000u);
 
   TestCase test;
-  test.name = "ImageSampleA16OffsetKeepsTexelOffset32BitOnGpu";
+  test.name = wave_size == 32 ? "ImageSampleA16OffsetKeepsTexelOffset32BitOnGpu32"
+                              : "ImageSampleA16OffsetKeepsTexelOffset32BitOnGpu64";
   test.code = code;
+  test.compute_info.wave_size = wave_size;
+  test.has_compute_info = true;
   test.expected = {0x3f800000u, 0x40000000u, 0x40400000u, 0x40800000u};
   test.opcodes = {O::V_MOV_B32, O::IMAGE_SAMPLE, O::BUFFER_STORE_DWORD,
                   O::S_ENDPGM};
   test.sampled_image_rgba = image;
-  test.required_spirv = {"UnpackHalf2x16"};
-  test.forbidden_spirv = {"OpBitFieldSExtract"};
+  test.required_spirv = {"UnpackHalf2x16", "Lod|ConstOffset"};
+  test.forbidden_spirv = {"OpBitFieldSExtract", "OpImageQuerySizeLod", "OpFDiv"};
+  return test;
+}
+
+template <u32 wave_size> TestCase ImageSampleOffsetsReadFourNeighbors() {
+  using O = ShaderOpcode;
+  TestCase test;
+  test.name = wave_size == 32 ? "ImageSampleOffsetsReadFourNeighbors32"
+                              : "ImageSampleOffsetsReadFourNeighbors64";
+  test.compute_info.wave_size = wave_size;
+  test.has_compute_info = true;
+  constexpr std::array offsets{0x01u, 0x3fu, 0x3f00u, 0x0100u};
+  constexpr std::array opcodes{0x37u, 0x37u, 0x34u, 0x32u};
+  for (u32 sample = 0; sample < offsets.size(); ++sample) {
+    AppendVMovLiteral(&test.code, 20, offsets[sample]);
+    u32 coordinate = 21;
+    if (opcodes[sample] == 0x32u) {
+      for (float gradient : {0.25f, 0.f, 0.f, 0.25f}) {
+        AppendVMovLiteral(&test.code, coordinate++, std::bit_cast<u32>(gradient));
+      }
+    }
+    // All instructions address the center of texel (2, 1) before applying the offset.
+    AppendVMovLiteral(&test.code, coordinate, std::bit_cast<u32>(0.625f));
+    AppendVMovLiteral(&test.code, coordinate + 1, std::bit_cast<u32>(0.375f));
+    if (opcodes[sample] == 0x34u) AppendVMovU32(&test.code, coordinate + 2, 0);
+    test.code.push_back(EncodeMimg0(opcodes[sample], 0x1));
+    test.code.push_back(EncodeMimg1(sample, 20));
+    AppendStoreVgpr(&test.code, sample, sample);
+  }
+  AppendEnd(&test.code);
+  test.sampled_image_rgba = MakeRgbaImage(4, 4);
+  SetRgbaPixel(&test.sampled_image_rgba, 4, 2, 1, std::bit_cast<u32>(16.f), 0, 0, 0);
+  constexpr std::array<std::array<u32, 2>, 4> neighbors{{{3, 1}, {1, 1}, {2, 0}, {2, 2}}};
+  for (u32 sample = 0; sample < neighbors.size(); ++sample) {
+    const auto value = std::bit_cast<u32>(float(sample + 1));
+    SetRgbaPixel(&test.sampled_image_rgba, 4, neighbors[sample][0], neighbors[sample][1],
+                 value, 0, 0, 0);
+    test.expected.push_back(value);
+  }
+  test.opcodes = {O::V_MOV_B32, O::IMAGE_SAMPLE, O::BUFFER_STORE_DWORD, O::S_ENDPGM};
+  test.required_spirv = {"Lod|ConstOffset", "Grad|ConstOffset"};
+  test.forbidden_spirv = {"OpImageQuerySizeLod", "OpFDiv", "OpBitFieldSExtract", "OpSwitch"};
   return test;
 }
 
@@ -34986,7 +35232,8 @@ TestCase ImageGatherCompareOpcodes() {
       O::V_MOV_B32,         O::IMAGE_GATHER4_C,      O::IMAGE_GATHER4_C_LZ,
       O::IMAGE_GATHER4_C_O, O::IMAGE_GATHER4_C_LZ_O, O::BUFFER_STORE_DWORD,
       O::S_ENDPGM};
-  test.required_spirv = {"OpImageDrefGather", "OpBitFieldSExtract"};
+  test.required_spirv = {"OpImageDrefGather", "ConstOffset"};
+  test.forbidden_spirv = {"OpBitFieldSExtract"};
   test.compile_only = true;
   return test;
 }
@@ -35538,7 +35785,7 @@ void CheckIndirectImageKeySwitch(VulkanHarness &vulkan) {
   std::memcpy(&memory_flag_bits, &memory_flags, sizeof(memory_flags));
   auto &sample = block->AppendNewInst(
       ValueOpcode::ImageSampleRaw,
-      {Value(&image), Value(&sampler), Value(&address)}, memory_flag_bits);
+      {Value(&image), Value(&sampler), Value(&address), Value(true)}, memory_flag_bits);
   auto &sample_x = block->AppendNewInst(ValueOpcode::CompositeExtractU32x4,
                                         {Value(&sample), Value(0u)});
   block->AppendNewInst(ValueOpcode::ReferenceU32, {Value(&sample_x)});
@@ -35749,29 +35996,63 @@ void CheckIndirectImageKeySwitch(VulkanHarness &vulkan) {
         resource == 4u ? vk::ImageViewType::e2DArray : vk::ImageViewType::e2D, layers));
   }
   const auto native_sampler = vulkan.CreateSampler(test.name);
-  for (const bool fetch : {false, true}) {
+  enum class Access { Sample, Gather, GatherHomogeneous, Fetch };
+  auto *access = &sample;
+  for (const auto mode : {Access::Sample, Access::Gather,
+                          Access::GatherHomogeneous, Access::Fetch}) {
+    const bool fetch = mode == Access::Fetch;
+    const bool gather = mode == Access::Gather || mode == Access::GatherHomogeneous;
+    const bool homogeneous = mode == Access::GatherHomogeneous;
+    if (mode == Access::Gather) {
+      test.name = "IndirectImageGatherContiguousRuns";
+      const auto position = std::ranges::find_if(*block, [&](const Inst &inst) {
+        return &inst == access;
+      });
+      auto &read = *block->PrependNewInst(
+          position, ValueOpcode::ImageGatherRaw,
+          {Value(&image), Value(&sampler), Value(&address), Value(0u), Value(0u),
+           Value(0u), Value(0u), Value(true)}, memory_flag_bits);
+      access->ReplaceUsesWith(Value(&read), false);
+      access = &read;
+      compiled.program.info.images[4].cube = false;
+      auto &gather_memory = compiled.program.memory_info[0];
+      gather_memory.dmask = 1;
+      gather_memory.image_sample_flags = ShaderRecompiler::Decoder::ImageSampleFlagLevelZero |
+          ShaderRecompiler::Decoder::ImageSampleFlagOffset;
+      // GATHER4_LZ_O places its packed signed offset before the coordinates.
+      address.SetArg(0, Value(0x3fu));
+      address.SetArg(1, Value(std::bit_cast<u32>(0.75f)));
+      address.SetArg(2, Value(std::bit_cast<u32>(0.25f)));
+      address.SetArg(3, Value(0u));
+    }
+    auto &selected_resources = compiled.program.info.images[0].indirect_resources;
+    selected_resources = homogeneous ? std::vector<u32>{0u, 2u, 3u, 5u, 6u}
+                                     : root.indirect_resources;
+    if (homogeneous) test.name = "IndirectImageGatherHomogeneous";
+    compiled.resources.flattened_srt[6] = homogeneous ? 4u : ordinals[5];
     if (fetch) {
       test.name = "IndirectImageFetchContiguousRuns";
       const auto position = std::ranges::find_if(*block, [&](const Inst &inst) {
-        return &inst == &sample;
+        return &inst == access;
       });
       auto &read = *block->PrependNewInst(
           position, ValueOpcode::ImageRead,
           {Value(&image), Value(&address), Value(true)}, memory_flag_bits);
-      sample.ReplaceUsesWith(Value(&read), false);
+      access->ReplaceUsesWith(Value(&read), false);
       sampler.Invalidate();
       compiled.program.info.samplers.clear();
       compiled.program.info.sampled_pairs.clear();
       compiled.program.info.images[4].cube = false;
+      compiled.program.memory_info[0].dmask = 0xf;
       compiled.program.memory_info[0].image_sample_flags = 0;
       compiled.program.memory_info[0].image_has_mip = true;
       for (u32 component = 0; component < 4u; ++component)
         address.SetArg(component, Value(0u));
       // Native DIM=2DArray places LOD after the layer, even for a 2D candidate.
       address.SetArg(3, Value(1u));
-      compiled.program.shader_info_complete = false;
-      CollectShaderInfo(compiled.program, {.compute = &compute});
     }
+    compiled.program.shader_info_complete = false;
+    CollectShaderInfo(compiled.program, {.compute = &compute});
     for (const u32 wave_size : {32u, 64u}) {
       compute.wave_size = wave_size;
       compute.host_subgroup_size = vulkan.SubgroupSize();
@@ -35782,31 +36063,38 @@ void CheckIndirectImageKeySwitch(VulkanHarness &vulkan) {
       compiled.packed_user_data.resize(compiled.program.bindings.ShaderDataDwords());
       ValidateSpirv(test.name, compiled.spirv);
       const auto halves = wave_size > compute.host_subgroup_size ? 2u : 1u;
+      const auto image_opcode = fetch ? "OpImageFetch" : gather ? "OpImageGather"
+                                                          : "OpImageSampleExplicitLod";
       Require(test.name, "native array image runs", tools.Disassemble(compiled.spirv, &text) &&
-                  CountText(text, fetch ? "OpImageFetch" : "OpImageSampleExplicitLod") == 3u * halves,
+                  CountText(text, image_opcode) == (homogeneous ? 1u : 3u) * halves,
               "mixed runs expanded into per-image accesses");
-      if (fetch) {
+      if (homogeneous) {
+        Require(test.name, "homogeneous gather array", text.find("OpSwitch") == std::string::npos,
+                "homogeneous gather emitted a descriptor switch");
+      }
+      if (fetch || gather) {
         std::vector<u32> decorated;
-        std::vector<u32> fetched_images;
+        std::vector<u32> image_operands;
         for (size_t offset = 5; offset < compiled.spirv.size();) {
           const auto words = std::span<const u32>(compiled.spirv).subspan(
               offset, compiled.spirv[offset] >> 16u);
           const auto opcode = static_cast<spv::Op>(words[0] & 0xffffu);
           if (opcode == spv::OpDecorate && words[2] == spv::DecorationNonUniform)
             decorated.push_back(words[1]);
-          if (opcode == spv::OpImageFetch) fetched_images.push_back(words[3]);
+          if (opcode == (fetch ? spv::OpImageFetch : spv::OpImageGather))
+            image_operands.push_back(words[3]);
           offset += words.size();
         }
-        Require(test.name, "nonuniform fetch operands",
-                std::ranges::count_if(fetched_images, [&](u32 operand) {
+        Require(test.name, "nonuniform image operands",
+                std::ranges::count_if(image_operands, [&](u32 operand) {
                   return std::ranges::find(decorated, operand) != decorated.end();
-                }) == 2u * halves,
-                "a dynamically selected fetch image lacks its nonuniform decoration");
+                }) == (homogeneous ? 1u : 2u) * halves,
+                "a dynamically selected image operand lacks its nonuniform decoration");
       }
       test.initial.assign(wave_size, 0xdeadbeefu);
       test.expected.assign(wave_size, 0u);
       for (u32 lane = 0; lane < ordinals.size(); ++lane) {
-        const auto resource = root.indirect_resources[ordinals[lane]];
+        const auto resource = selected_resources[compiled.resources.flattened_srt[1u + lane]];
         test.expected[lane] = std::bit_cast<u32>(float(resource + (fetch && resource != 0u ? 10u : 0u)));
       }
       auto buffer = vulkan.CreateStorageBuffer(test.name, test.initial, test.initial.size());
@@ -35814,7 +36102,7 @@ void CheckIndirectImageKeySwitch(VulkanHarness &vulkan) {
                       fetch ? vk::Sampler{} : native_sampler, textures);
       const auto actual = vulkan.ReadBuffer(test.name, buffer, test.expected.size());
       vulkan.DestroyBuffer(&buffer);
-      CompareWords(test, "root, native slots, cube and mapping bounds", test.expected, actual);
+      CompareWords(test, "root, native slots, array runs and mapping bounds", test.expected, actual);
     }
     std::printf("[compute] %-32s ok\n", test.name);
   }
@@ -37659,7 +37947,10 @@ std::vector<TestCase> MakeCases() {
   AddCase(ImageD16GatherPacksHalfPairs);
   AddCase(ImageSampleA16SamplerCoordsOnGpu);
   AddCase(ImageSampleOpcodeAliasUsesNormalCoords);
-  AddCase(ImageSampleA16OffsetKeepsTexelOffset32BitOnGpu);
+  AddCase(ImageSampleA16OffsetKeepsTexelOffset32BitOnGpu<32>);
+  AddCase(ImageSampleA16OffsetKeepsTexelOffset32BitOnGpu<64>);
+  AddCase(ImageSampleOffsetsReadFourNeighbors<32>);
+  AddCase(ImageSampleOffsetsReadFourNeighbors<64>);
   AddCase(ImageSampleA16CompareBiasRdna2AddressOrder);
   AddCase(ImageGatherCompareOpcodes);
   AddCase(ImageStoreVariants);
@@ -40070,7 +40361,7 @@ void CheckImageSamplerSpecialization() {
               sizeof(signed_memory_flags));
   mixed_sampler_block->AppendNewInst(
       ValueOpcode::ImageSampleRaw,
-      {Value(&image_handle), Value(&sampler_handle), Value(&image_address)},
+      {Value(&image_handle), Value(&sampler_handle), Value(&image_address), Value(true)},
       signed_memory_flag_bits);
 
   ShaderRecompiler::IR::DescriptorValue native_image_descriptor{};
@@ -42630,6 +42921,43 @@ void CheckPm4IndirectControlFlow(RenderContext &renderer) {
           Gen5::AgcCbNop(reinterpret_cast<Gen5::CommandBuffer*>(&payload_dcb), 3) == payload.data() &&
               payload == std::array<uint32_t, 3>{0xc0011000u, 0x12345678u, 0x9abcdef0u},
           "NOP emitter overwrote the caller's reserved payload");
+  alignas(8) uint64_t released = 0;
+  std::array<uint32_t, 20> disabled_commands{};
+  const auto disabled_write = write(&returned, 99);
+  std::copy(disabled_write.begin(), disabled_write.end(), disabled_commands.begin());
+  disabled_commands[5] = KYTY_PM4(2, Pm4::IT_NUM_INSTANCES, 0);
+  disabled_commands[6] = 0x68750123u;
+  CommandBufferLayout release_dcb{disabled_commands.data() + 7, disabled_commands.data() + 15,
+      disabled_commands.data() + 7, disabled_commands.data() + 15, nullptr, nullptr, 0};
+  auto* release = Gen5::AgcCbReleaseMem(
+      reinterpret_cast<Gen5::CommandBuffer*>(&release_dcb), 0x28, 0, 0, 0,
+      reinterpret_cast<const volatile Gen5::Label*>(&released), 2, 77, 0, 0, 0, 0);
+  const auto enabled_write = write(&selected, 88);
+  std::copy(enabled_write.begin(), enabled_write.end(), disabled_commands.begin() + 15);
+  const auto unpatched_commands = disabled_commands;
+  Require(name, "SetNop packets",
+          release == disabled_commands.data() + 7 &&
+              Gen5::AgcSetNop(disabled_commands.data()) == 0 &&
+              Gen5::AgcSetNop(disabled_commands.data() + 5) == 0 && Gen5::AgcSetNop(release) == 0 &&
+              disabled_commands[0] == ((disabled_write[0] & 0xffff00ffu) | (Pm4::IT_NOP << 8u)) &&
+              disabled_commands[5] == KYTY_PM4(2, Pm4::IT_NOP, Pm4::R_ZERO) &&
+              disabled_commands[6] == 0x123u &&
+              release[0] == KYTY_PM4(8, Pm4::IT_NOP, Pm4::R_ZERO) &&
+              unpatched_commands[7] == KYTY_PM4(8, Pm4::IT_NOP, Pm4::R_RELEASE_MEM) &&
+              Gen5::AgcGetPacketSize(disabled_commands.data()) == disabled_write.size() &&
+              Gen5::AgcGetPacketSize(disabled_commands.data() + 5) == 2 &&
+              Gen5::AgcGetPacketSize(release) == 8 &&
+              std::equal(disabled_commands.begin() + 1, disabled_commands.begin() + 5,
+                         unpatched_commands.begin() + 1) &&
+              std::equal(disabled_commands.begin() + 8, disabled_commands.end(),
+                         unpatched_commands.begin() + 8),
+          "SetNop changed a packet length, kept its operation or marker, or touched its payload");
+  selected = returned = 0;
+  Pm4Execution disabled_execution;
+  Require(name, "SetNop execution",
+          processor.Process(disabled_execution, disabled_commands) == Pm4ProcessResult::Complete &&
+              returned == 0 && released == 0 && selected == 88,
+          "a disabled packet still executed or hid the packet after it");
   const auto then_commands = write(&selected, 11);
   const auto else_commands = write(&selected, 33);
   const auto branch_suffix = write(&selected, 44);
@@ -43290,6 +43618,16 @@ int main(int argc, char **argv) {
     RunCase(&vulkan, FlatVirtualAddressRebasesGuestAllocation());
     return 0;
   }
+  if (argc == 2 && std::strcmp(argv[1], "--image-sample-offset-only") == 0) {
+    VulkanHarness vulkan;
+    RunCase(&vulkan, ImageSampleOffsetsReadFourNeighbors<32>());
+    RunCase(&vulkan, ImageSampleOffsetsReadFourNeighbors<64>());
+    RunCase(&vulkan, ImageSampleA16OffsetKeepsTexelOffset32BitOnGpu<32>());
+    RunCase(&vulkan, ImageSampleA16OffsetKeepsTexelOffset32BitOnGpu<64>());
+    RunCase(&vulkan, ImageSampleAndGather());
+    RunCase(nullptr, ImageGatherCompareOpcodes());
+    return 0;
+  }
   if (argc == 2 && std::strcmp(argv[1], "--image-gather-lod-only") == 0) {
     VulkanHarness vulkan;
     RunCase(&vulkan, ImageGatherExplicitLod());

tests/shaderCfgTests.cpp

@@ -1316,7 +1316,7 @@ void TestShaderFeatureCollection() {
                        {Value(0u), Value(0u), Value(0u)});
   block->AppendNewInst(ValueOpcode::ImageGatherRaw,
                        {Value(0u), Value(0u), Value(0u), Value(0u),
-                        Value(0u), Value(0u), Value(0u)});
+                        Value(0u), Value(0u), Value(0u), Value(true)});
   auto &shared = block->AppendNewInst(ValueOpcode::LoadSharedU32,
                                       {Value(0u), Value(true)});
   shared.SetFlags(MemoryFlags{.index = 0});