loader: emulate immediate SSE4a instructions on macOS (#635)

* loader: emulate immediate SSE4a instructions on macOS

* tests: simplify macOS SSE4a test integration

Keep the Darwin target, CTest registration, and aggregate dependency in
one conditional block without a separate target list. Preserve the
standalone target when BUILD_TESTING is disabled.

Lowercase the test filename and diagnostic prefix, and drop unrelated
formatting changes from the instruction emulator.

Credit Brandon Strong's related SSE4a work in
https://github.com/KytyPS5/KytyPS5/pull/467.

Co-authored-by: Brandon Strong <brandostrong@gmail.com>

---------

Co-authored-by: nmzik <Nmzik@mail.ru>
Co-authored-by: Brandon Strong <brandostrong@gmail.com>

.github/workflows/build.yml

@@ -6,7 +6,12 @@ on:
     branches: [ main, master ]
   pull_request:
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+
 env:
+  SCCACHE_GHA_ENABLED: "true"
   KYTY_BUILD_ORIGIN: ${{ github.repository != 'KytyPS5/KytyPS5' && 'Fork' || github.event_name == 'push' && 'Official' || 'Source' }}
   KYTY_BUILD_REPOSITORY: ${{ github.repository }}
 
@@ -78,6 +83,9 @@ jobs:
           arch: win64_msvc2022_64
           cache: true
 
+      - name: Setup sccache
+        uses: mozilla-actions/sccache-action@v0.0.11
+
       - name: Verify toolchain
         shell: cmd
         run: |
@@ -100,6 +108,8 @@ jobs:
             -DKYTY_RELEASE_TAG=${{ needs.metadata.outputs.release_tag }} ^
             -DCMAKE_C_COMPILER=clang-cl ^
             -DCMAKE_CXX_COMPILER=clang-cl ^
+            -DCMAKE_C_COMPILER_LAUNCHER=sccache ^
+            -DCMAKE_CXX_COMPILER_LAUNCHER=sccache ^
             -DCMAKE_PREFIX_PATH="%Qt6_DIR%"
 
       - name: Build
@@ -156,6 +166,9 @@ jobs:
           arch: clang_64
           cache: true
 
+      - name: Setup sccache
+        uses: mozilla-actions/sccache-action@v0.0.11
+
       - name: Verify toolchain
         shell: bash
         run: |
@@ -180,6 +193,8 @@ jobs:
             -DCMAKE_OSX_ARCHITECTURES=x86_64 \
             -DCMAKE_C_COMPILER=clang \
             -DCMAKE_CXX_COMPILER=clang++ \
+            -DCMAKE_C_COMPILER_LAUNCHER=sccache \
+            -DCMAKE_CXX_COMPILER_LAUNCHER=sccache \
             -DCMAKE_PREFIX_PATH="$QT_ROOT_DIR"
 
       - name: Build
@@ -304,6 +319,7 @@ jobs:
             libxss-dev \
             lld \
             ninja-build \
+            pkg-config \
             wayland-protocols
 
       - name: Install Qt 6.10.3
@@ -315,6 +331,9 @@ jobs:
           arch: linux_gcc_64
           cache: true
 
+      - name: Setup sccache
+        uses: mozilla-actions/sccache-action@v0.0.11
+
       - name: Verify toolchain
         shell: bash
         run: |
@@ -339,6 +358,8 @@ jobs:
             -DKYTY_RELEASE_TAG="${{ needs.metadata.outputs.release_tag }}" \
             -DCMAKE_C_COMPILER=clang \
             -DCMAKE_CXX_COMPILER=clang++ \
+            -DCMAKE_C_COMPILER_LAUNCHER=sccache \
+            -DCMAKE_CXX_COMPILER_LAUNCHER=sccache \
             -DCMAKE_PREFIX_PATH="$Qt6_DIR" 2>&1 | tee _Build/configure.log
           exit "${PIPESTATUS[0]}"
 

.gitmodules

@@ -22,7 +22,7 @@
 	url = https://github.com/GPUOpen-LibrariesAndSDKs/VulkanMemoryAllocator.git
 [submodule "3rdparty/ffmpeg-core"]
 	path = 3rdparty/ffmpeg-core
-	url = https://github.com/shadps4-emu/ext-ffmpeg-core.git
+	url = https://github.com/KytyPS5/ext-ffmpeg-core.git
 	shallow = true
 [submodule "3rdparty/fmt"]
 	path = 3rdparty/fmt

CMakeLists.txt

@@ -317,24 +317,10 @@ endif()
 
 set(kyty_emulator_link_libraries common Vulkan::Headers spirv-tools-opt spirv-tools SDL2-static imgui xxhash FFmpeg::ffmpeg fmt::fmt nlohmann_json::nlohmann_json LibAtrac9 kyty_git_version Zydis::Zydis xbyak::xbyak)
 
-# Linux system libraries required by the static FFmpeg archive.
+# Linux system libraries.
 if(LINUX)
 	find_package(Threads REQUIRED)
 	list(APPEND kyty_emulator_link_libraries m ${CMAKE_DL_LIBS} Threads::Threads)
-
-	# Optional FFmpeg dependencies.
-	find_package(ZLIB)
-	if(ZLIB_FOUND)
-		list(APPEND kyty_emulator_link_libraries ZLIB::ZLIB)
-	endif()
-	find_library(KYTY_BZ2_LIBRARY bz2)
-	if(KYTY_BZ2_LIBRARY)
-		list(APPEND kyty_emulator_link_libraries ${KYTY_BZ2_LIBRARY})
-	endif()
-	find_library(KYTY_LZMA_LIBRARY lzma)
-	if(KYTY_LZMA_LIBRARY)
-		list(APPEND kyty_emulator_link_libraries ${KYTY_LZMA_LIBRARY})
-	endif()
 endif()
 
 set(inc_headers
@@ -615,6 +601,21 @@ if(BUILD_TESTING)
 	)
 endif()
 
+# Test Darwin signal-context restoration for SSE4a instructions.
+if(APPLE AND (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR
+		(NOT CMAKE_OSX_ARCHITECTURES AND CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|AMD64)$")))
+	add_executable(macos_sse4a_tests EXCLUDE_FROM_ALL
+		"${KYTY_TESTS_DIR}/macosSse4aTests.cpp"
+		"${KYTY_SOURCE_DIR}/loader/x64InstructionEmulator.cpp"
+	)
+	target_link_libraries(macos_sse4a_tests common fmt::fmt Zydis::Zydis)
+	target_include_directories(macos_sse4a_tests PRIVATE ${inc_headers})
+	if(BUILD_TESTING)
+		add_test(NAME macos_sse4a COMMAND $<TARGET_FILE:macos_sse4a_tests>)
+		set_tests_properties(macos_sse4a PROPERTIES TIMEOUT 15)
+		add_dependencies(kyty_tests macos_sse4a_tests)
+	endif()
+endif()
 
 add_executable(kyty_emulator "${KYTY_SOURCE_DIR}/main.cpp" ${kyty_emulator_src})
 configure_macos_guest_address_space(kyty_emulator)

README.md

@@ -171,7 +171,7 @@ build has no working sound and no gamepad hotplug:
 
 ```bash
 sudo apt-get install --no-install-recommends \
-  clang lld ninja-build cmake git glslang-tools \
+  clang lld ninja-build cmake git glslang-tools pkg-config \
   libgl1-mesa-dev libx11-dev libxcursor-dev libxext-dev libxfixes-dev \
   libxi-dev libxrandr-dev libxss-dev libxkbcommon-dev \
   libasound2-dev libpulse-dev libudev-dev libdbus-1-dev libwayland-dev wayland-protocols
@@ -192,13 +192,37 @@ cmake --install _Build/linux --prefix _Build/linux/install
 ```
 
 The install step copies the Qt libraries and plugins next to the binaries, so
-`_Build/linux/install` runs without a matching system Qt.
+`_Build/linux/install` runs without a matching system Qt. FFmpeg is linked statically
+from the pinned [KytyPS5 FFmpeg core](https://github.com/KytyPS5/ext-ffmpeg-core)
+release, including VP9 and WebM support. System FFmpeg packages are not required.
 
 As on Windows, the MSVC compiler is not used; Clang is required. `cl.exe` is rejected at configure
 time.
 
 The CMake source root is the repository root.
 
+### Building on NixOS
+
+A development shell provides Clang, CMake, Ninja, Qt 6, the Vulkan headers, and the SDL2 backend
+libraries. Enter it and configure exactly as on other Linux distributions; the shell exports
+`CMAKE_PREFIX_PATH` and `QT_PLUGIN_PATH`, so the `-DCMAKE_PREFIX_PATH="$Qt6_DIR"` argument is not
+needed:
+
+```bash
+nix-shell # or: nix develop
+git submodule update --init --recursive
+
+cmake -S . -B _Build/linux -G Ninja -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++
+
+cmake --build _Build/linux --target launcher --parallel
+cmake --install _Build/linux --prefix _Build/linux/install
+```
+
+The configure step downloads the FFmpeg prebuilts and the `xbyak`/`zydis` sources, so it needs
+network access; a fully sandboxed `nix build` would require vendoring those inputs. A Vulkan 1.3
+driver must be available at runtime (on NixOS, `hardware.graphics.enable = true`).
+
 ### Building on macOS
 
 macOS builds target x86-64 and run under Rosetta 2 on Apple Silicon, so the PS5's x86-64 game

flake.lock

@@ -0,0 +1,27 @@
+{
+  "nodes": {
+    "nixpkgs": {
+      "locked": {
+        "lastModified": 1789426278,
+        "narHash": "sha256-GnotcvKtbTMnoVx6G4E7ZdUX6487Dky74dqq3pU7iRk=",
+        "owner": "NixOS",
+        "repo": "nixpkgs",
+        "rev": "efe6f071ede9d21c37462d2d6682d5e670099684",
+        "type": "github"
+      },
+      "original": {
+        "owner": "NixOS",
+        "ref": "nixos-unstable",
+        "repo": "nixpkgs",
+        "type": "github"
+      }
+    },
+    "root": {
+      "inputs": {
+        "nixpkgs": "nixpkgs"
+      }
+    }
+  },
+  "root": "root",
+  "version": 7
+}

flake.nix

@@ -0,0 +1,20 @@
+{
+  description = "KytyPS5 development shell";
+
+  inputs.nixpkgs.url = "github:NixOS/nixpkgs/nixos-unstable";
+
+  outputs = { self, nixpkgs }:
+    let
+      systems = [ "x86_64-linux" ];
+      forAllSystems = nixpkgs.lib.genAttrs systems;
+    in
+    {
+      devShells = forAllSystems (system:
+        let
+          pkgs = import nixpkgs { inherit system; };
+        in
+        {
+          default = import ./nix/devshell.nix { inherit pkgs; };
+        });
+    };
+}

nix/devshell.nix

@@ -0,0 +1,84 @@
+{ pkgs }:
+
+let
+  # Build-time dependencies. Every -dev package the bundled SDL2 probes for must
+  # be listed here, otherwise SDL2 quietly drops the audio/Wayland/udev backends.
+  buildDeps = with pkgs; [
+    qt6.qtbase # Concurrent, Network, Widgets (launcher)
+    vulkan-headers
+    vulkan-loader
+    vulkan-tools
+    vulkan-validation-layers
+    libglvnd
+    # X11 / input
+    libX11
+    libXcursor
+    libXext
+    libXfixes
+    libXi
+    libXrandr
+    libXScrnSaver
+    libxkbcommon
+    # Wayland
+    wayland
+    wayland-protocols
+    libdecor
+    # audio + device/dbus
+    alsa-lib
+    libpulseaudio
+    systemd # libudev
+    dbus
+  ];
+
+  # Libraries the built binaries need at runtime.
+  runtimeLibs = with pkgs; [
+    stdenv.cc.cc.lib # libstdc++ / libgcc_s
+    libgcc.lib
+    qt6.qtbase # libQt6Concurrent/Network/Widgets
+    vulkan-loader
+    libglvnd
+    alsa-lib
+    libpulseaudio
+    libX11
+    libXcursor
+    libXext
+    libXfixes
+    libXi
+    libXrandr
+    libXScrnSaver
+    libxkbcommon
+    wayland
+    libdecor
+    systemd
+    dbus
+  ];
+in
+pkgs.mkShell {
+  name = "kytyps5";
+
+  nativeBuildInputs = with pkgs; [
+    clang
+    lld
+    cmake
+    ninja
+    pkg-config
+    git
+    glslang # provides glslangValidator for the bundled shaders
+  ];
+
+  buildInputs = buildDeps;
+
+  shellHook = ''
+    export CMAKE_PREFIX_PATH="${pkgs.qt6.qtbase}:''${CMAKE_PREFIX_PATH:-}"
+    export QT_PLUGIN_PATH="${pkgs.qt6.qtbase}/lib/qt-6/plugins:${pkgs.qt6.qtbase}/lib/qt6/plugins:''${QT_PLUGIN_PATH:-}"
+    export LD_LIBRARY_PATH="${pkgs.lib.makeLibraryPath runtimeLibs}:''${LD_LIBRARY_PATH:-}"
+
+    if [ -z "''${KYTY_SHELL_QUIET:-}" ]; then
+      echo "KytyPS5 dev shell"
+      echo "  cmake -S . -B _Build/linux -G Ninja -DCMAKE_BUILD_TYPE=Release \\"
+      echo "    -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++"
+      echo "  cmake --build _Build/linux --target launcher --parallel"
+      echo "  cmake --install _Build/linux --prefix _Build/linux/install"
+    fi
+  '';
+}

shell.nix

@@ -0,0 +1,4 @@
+let
+  pkgs = import <nixpkgs> { };
+in
+import ./nix/devshell.nix { inherit pkgs; }

src/common/emulatorConfig.cpp

@@ -43,6 +43,10 @@ int32_t GetUserId() {
 	return g_config->user_id;
 }
 
+const std::string& GetAudioInputDevice() {
+	return g_config->audio_input_device;
+}
+
 PresentMode GetPresentMode() {
 	return g_config->present_mode;
 }
@@ -55,6 +59,14 @@ bool FullscreenEnabled() {
 	return g_config->fullscreen_enabled;
 }
 
+bool VrEnabled() {
+	return g_config->vr_enabled;
+}
+
+bool AmdCpuEnabled() {
+	return g_config->amd_cpu_enabled;
+}
+
 uint32_t GetVblankFrequency() {
 	return std::clamp(g_config->vblank_frequency, 30u, 360u);
 }

src/common/emulatorConfig.h

@@ -43,9 +43,12 @@ struct ConfigOptions {
 	uint32_t               screen_height               = 720;
 	std::string            user_name                   = "Kyty";
 	int32_t                user_id                     = DEFAULT_USER_ID;
+	std::string            audio_input_device;
 	PresentMode            present_mode                = PresentMode::Mailbox;
 	int32_t                gpu_index                   = -1;
 	bool                   fullscreen_enabled          = false;
+	bool                   vr_enabled                  = false;
+	bool                   amd_cpu_enabled             = false;
 	uint32_t               vblank_frequency            = 60;
 	uint32_t               console_language            = DEFAULT_CONSOLE_LANGUAGE;
 	bool                   vulkan_validation_enabled   = false;
@@ -76,9 +79,12 @@ uint32_t GetScreenWidth();
 uint32_t GetScreenHeight();
 const std::string& GetUserName();
 int32_t  GetUserId();
+const std::string& GetAudioInputDevice();
 PresentMode GetPresentMode();
 int32_t GetGpuIndex();
 bool     FullscreenEnabled();
+bool     VrEnabled();
+bool     AmdCpuEnabled();
 uint32_t GetVblankFrequency();
 uint32_t GetConsoleLanguage();
 bool     VulkanValidationEnabled();

src/common/logging/log.cpp

@@ -129,6 +129,14 @@ static void WriteImpl(std::string_view text, fmt::text_style style = {}) {
 	}
 }
 
+void WriteToConsoleAndLog(std::string_view text) {
+	WriteImpl(text);
+	if (g_initialized && g_direction != Direction::Console) {
+		WriteStdout(text);
+	}
+	Flush();
+}
+
 void WriteFatal(std::string_view text) {
 	if (g_direction == Direction::Silent || !g_initialized) {
 		WriteStdout(text);

src/common/logging/log.h

@@ -25,6 +25,7 @@ Direction GetDirection();
 bool      IsSilent();
 void      Write(std::string_view text);
 void      Write(fmt::text_style style, std::string_view text);
+void      WriteToConsoleAndLog(std::string_view text);
 void      WriteFatal(std::string_view text);
 void      WriteFatal(fmt::text_style style, std::string_view text);
 void      Flush();

src/common/threads.cpp

@@ -8,7 +8,6 @@
 #include <chrono>             // IWYU pragma: keep
 #include <condition_variable> // IWYU pragma: keep
 #include <mutex>
-#include <vector>
 
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS && KYTY_COMPILER == KYTY_COMPILER_CLANG
 #define KYTY_WIN_CS
@@ -197,37 +196,7 @@ struct CondVarPrivate {
 #endif
 };
 
-static std::recursive_mutex                         g_cond_waiters_mutex;
-static std::vector<std::pair<int, CondVarPrivate*>> g_cond_waiters;
-static wait_poll_func_t                             g_cond_wait_poll_callback = nullptr;
-
-static void WakeCondVar(CondVarPrivate* cond_var) {
-#ifdef KYTY_WIN_CS
-	static auto func = ResolveWakeAllConditionVariable();
-	EXIT_NOT_IMPLEMENTED(func == nullptr);
-	func(&cond_var->m_cv);
-#else
-	cond_var->m_cv.notify_all();
-#endif
-}
-
-static void RegisterCondWaiter(CondVarPrivate* cond_var) {
-	std::lock_guard lock(g_cond_waiters_mutex);
-	g_cond_waiters.emplace_back(Thread::GetThreadIdUnique(), cond_var);
-}
-
-static void UnregisterCondWaiter(CondVarPrivate* cond_var) {
-	const auto      thread_id = Thread::GetThreadIdUnique();
-	std::lock_guard lock(g_cond_waiters_mutex);
-
-	const auto it = std::find_if(g_cond_waiters.begin(), g_cond_waiters.end(),
-	                             [thread_id, cond_var](const auto& waiter) {
-		                             return waiter.first == thread_id && waiter.second == cond_var;
-	                             });
-	if (it != g_cond_waiters.end()) {
-		g_cond_waiters.erase(it);
-	}
-}
+static wait_poll_func_t g_cond_wait_poll_callback = nullptr;
 
 struct ThreadPrivate {
 	ThreadPrivate(thread_func_t f, void* a): func(f), arg(a), m_thread(&Run, this) {}
@@ -365,7 +334,6 @@ CondVar::~CondVar() {
 }
 
 void CondVar::Wait(Mutex* mutex) {
-	RegisterCondWaiter(m_cond_var.get());
 #ifndef KYTY_WIN_CS
 	std::unique_lock<std::recursive_mutex> cpp_lock(mutex->m_mutex->m_mutex, std::adopt_lock_t());
 #endif
@@ -406,7 +374,6 @@ void CondVar::Wait(Mutex* mutex) {
 	}
 	cpp_lock.release();
 #endif
-	UnregisterCondWaiter(m_cond_var.get());
 }
 
 void CondVar::SetWaitPollCallback(wait_poll_func_t callback) {
@@ -415,7 +382,6 @@ void CondVar::SetWaitPollCallback(wait_poll_func_t callback) {
 
 bool CondVar::WaitFor(Mutex* mutex, uint32_t micros) {
 	bool ok = false;
-	RegisterCondWaiter(m_cond_var.get());
 #ifndef KYTY_WIN_CS
 	std::unique_lock<std::recursive_mutex> cpp_lock(mutex->m_mutex->m_mutex, std::adopt_lock_t());
 #endif
@@ -430,7 +396,6 @@ bool CondVar::WaitFor(Mutex* mutex, uint32_t micros) {
 	      std::cv_status::no_timeout);
 	cpp_lock.release();
 #endif
-	UnregisterCondWaiter(m_cond_var.get());
 	return ok;
 }
 
@@ -445,16 +410,13 @@ void CondVar::Signal() {
 }
 
 void CondVar::SignalAll() {
-	WakeCondVar(m_cond_var.get());
-}
-
-void CondVar::SignalThread(int thread_id) {
-	std::lock_guard lock(g_cond_waiters_mutex);
-	for (const auto& waiter: g_cond_waiters) {
-		if (waiter.first == thread_id) {
-			WakeCondVar(waiter.second);
-		}
-	}
+#ifdef KYTY_WIN_CS
+	static auto func = ResolveWakeAllConditionVariable();
+	EXIT_NOT_IMPLEMENTED(func == nullptr);
+	func(&m_cond_var->m_cv);
+#else
+	m_cond_var->m_cv.notify_all();
+#endif
 }
 
 int Thread::GetThreadIdUnique() {

src/common/threads.h

@@ -77,7 +77,6 @@ public:
 	void Signal();
 	void SignalAll();
 
-	static void SignalThread(int thread_id);
 	static void SetWaitPollCallback(wait_poll_func_t callback);
 
 	KYTY_CLASS_NO_COPY(CondVar);

src/emulator.cpp

@@ -196,6 +196,10 @@ void Run(const RunOptions& options) {
 	ClearDebugTextureFolder();
 
 	PrintSystemInfo();
+	std::string title_id;
+	if (Loader::SystemContentParamSfoGetString("TITLE_ID", &title_id) && !title_id.empty()) {
+		Log::WriteToConsoleAndLog(fmt::format("Title ID: {}\n", title_id));
+	}
 
 	int ok = atexit(KytyClose);
 	EXIT_NOT_IMPLEMENTED(ok != 0);

src/graphics/guest_gpu/command_processor/commandProcessor.h

@@ -32,11 +32,12 @@ private:
 
 	struct BufferCursor {
 		std::span<const uint32_t> commands;
-		uint32_t                  offset_dw           = 0;
-		uint32_t                  deferred_advance_dw = 0;
+		uint32_t                  offset_dw = 0;
 	};
 
 	std::vector<BufferCursor> m_buffer_stack;
+	std::span<const uint32_t> m_next_buffer;
+	bool                      m_chain         = false;
 	bool                      m_suspended     = false;
 	bool                      m_made_progress = false;
 };
@@ -129,7 +130,7 @@ public:
 	[[nodiscard]] bool ShouldSkipPredicatedPackets() const { return m_predicate_skip; }
 
 	Pm4ProcessResult Process(Pm4Execution& execution, std::span<const uint32_t> commands);
-	void             ProcessIndirectBuffer(std::span<const uint32_t> commands);
+	void             ProcessIndirectBuffer(std::span<const uint32_t> commands, bool chain);
 
 	void SetFlip(const FlipInfo& flip) { m_flip = flip; }
 
@@ -143,7 +144,7 @@ private:
 	                      uint32_t cache_action, uint32_t event_index, uint32_t event_write_source,
 	                      void* dst_gpu_addr, T value, uint32_t interrupt_selector,
 	                      uint32_t interrupt_context_id);
-	void ProcessPm4(Pm4Execution& execution, size_t stop_depth);
+	void ProcessPm4(Pm4Execution& execution);
 	void SuspendPm4();
 	CommandScheduler&   GetScheduler() const { return m_renderer.GetCommandScheduler(); }
 	CommandBuffer&      CurrentBuffer() { return GetScheduler().Current(); }

src/graphics/guest_gpu/command_processor/pm4Handlers.cpp

@@ -1495,10 +1495,10 @@ KYTY_CP_OP_PARSER(CpOpBranch) {
 	     reinterpret_cast<uint64_t>(else_buffer), else_num_dw);
 
 	if (take_then) {
-		cp.ProcessIndirectBuffer({then_buffer, then_num_dw});
+		cp.ProcessIndirectBuffer({then_buffer, then_num_dw}, true);
 	} else if (mode == 2 && else_num_dw != 0) {
 		EXIT_NOT_IMPLEMENTED(else_buffer == nullptr);
-		cp.ProcessIndirectBuffer({else_buffer, else_num_dw});
+		cp.ProcessIndirectBuffer({else_buffer, else_num_dw}, true);
 	}
 
 	return payload_dw;
@@ -1959,7 +1959,7 @@ KYTY_CP_OP_PARSER(CpOpIndirectBuffer) {
 
 	GraphicsDbgDumpDcb("ci", indirect_num_dw, indirect_buffer);
 
-	cp.ProcessIndirectBuffer({indirect_buffer, indirect_num_dw});
+	cp.ProcessIndirectBuffer({indirect_buffer, indirect_num_dw}, (control & (1u << 20u)) != 0);
 
 	return 3;
 }
@@ -3635,6 +3635,11 @@ void GraphicsInitJmpTablesShIndirect() {
 	g_hw_sh_indirect_func[Pm4::SPI_SHADER_PGM_RSRC4_GS] = [](KYTY_HW_SH_INDIRECT_ARGS) {
 		HwShIgnoreShaderRegister(cmd_offset, value);
 	};
+	// Toolkit state restoration emits the compiler's GS-front allocation
+	// metadata here. Native GS resource registers already carry that allocation.
+	g_hw_sh_indirect_func[0x0ca] = [](KYTY_HW_SH_INDIRECT_ARGS) {
+		HwShIgnoreShaderRegister(cmd_offset, value);
+	};
 	g_hw_sh_indirect_func[Pm4::SPI_GRAPHICS_SHADER_CONTROL_GS] = [](KYTY_HW_SH_INDIRECT_ARGS) {
 		HwShIgnoreShaderRegister(cmd_offset, value);
 	};
@@ -3654,12 +3659,13 @@ void GraphicsInitJmpTablesShIndirect() {
 	g_hw_sh_indirect_func[Pm4::SPI_GRAPHICS_SHADER_CONTROL_HS] = [](KYTY_HW_SH_INDIRECT_ARGS) {
 		HwShIgnoreShaderRegister(cmd_offset, value);
 	};
-	g_hw_sh_indirect_func[Pm4::SPI_SHADER_USER_DATA_ADDR_LO_HS] = [](KYTY_HW_SH_INDIRECT_ARGS) {
-		HwShIgnoreShaderRegister(cmd_offset, value);
-	};
-	g_hw_sh_indirect_func[Pm4::SPI_SHADER_USER_DATA_ADDR_HI_HS] = [](KYTY_HW_SH_INDIRECT_ARGS) {
-		HwShIgnoreShaderRegister(cmd_offset, value);
-	};
+	for (uint32_t offset = Pm4::SPI_SHADER_USER_DATA_ADDR_LO_HS;
+	     offset <= Pm4::SPI_SHADER_USER_DATA_ADDR_HI_HS; offset++) {
+		g_hw_sh_indirect_func[offset] = [](KYTY_HW_SH_INDIRECT_ARGS) {
+			cp.GetShCtx().SetHsUserDataAddress(cmd_offset - Pm4::SPI_SHADER_USER_DATA_ADDR_LO_HS,
+			                                   value);
+		};
+	}
 	g_hw_sh_indirect_func[Pm4::SPI_SHADER_PGM_LO_HS] = [](KYTY_HW_SH_INDIRECT_ARGS) {
 		auto base = cp.GetShCtx().GetVs().hs_regs.data_addr;
 		base &= 0xFFFFFF00000000FFull;
@@ -3818,6 +3824,25 @@ void GraphicsInitJmpTablesUcIndirect() {
 	for (auto& func: g_hw_uc_indirect_func) {
 		func = nullptr;
 	}
+	for (uint32_t i = 0; i < 4; ++i) {
+		g_hw_uc_indirect_func[Pm4::FSR_CONTROL_POINTS_LEFT_X + i] = [](KYTY_HW_UC_INDIRECT_ARGS) {
+			cp.GetUcfg().SetFsrControlPoint(0, cmd_offset - Pm4::FSR_CONTROL_POINTS_LEFT_X, value);
+		};
+		g_hw_uc_indirect_func[Pm4::FSR_CONTROL_POINTS_LEFT_Y + i] = [](KYTY_HW_UC_INDIRECT_ARGS) {
+			cp.GetUcfg().SetFsrControlPoint(1, cmd_offset - Pm4::FSR_CONTROL_POINTS_LEFT_Y, value);
+		};
+	}
+	for (uint32_t i = 0; i < 2; ++i) {
+		g_hw_uc_indirect_func[Pm4::FSR_ALPHA_LEFT_X + i] = [](KYTY_HW_UC_INDIRECT_ARGS) {
+			cp.GetUcfg().SetFsrAlpha(0, cmd_offset - Pm4::FSR_ALPHA_LEFT_X, value);
+		};
+		g_hw_uc_indirect_func[Pm4::FSR_ALPHA_LEFT_Y + i] = [](KYTY_HW_UC_INDIRECT_ARGS) {
+			cp.GetUcfg().SetFsrAlpha(1, cmd_offset - Pm4::FSR_ALPHA_LEFT_Y, value);
+		};
+		g_hw_uc_indirect_func[Pm4::FSR_WINDOW_LEFT + i] = [](KYTY_HW_UC_INDIRECT_ARGS) {
+			cp.GetUcfg().SetFsrWindow(cmd_offset - Pm4::FSR_WINDOW_LEFT, value);
+		};
+	}
 
 	g_hw_uc_indirect_func[Pm4::GE_CNTL] = [](KYTY_HW_UC_INDIRECT_ARGS) {
 		HW::GeControl r;

src/graphics/guest_gpu/graphicsRun.cpp

@@ -386,8 +386,12 @@ void CommandProcessor::WriteReferenceClock(uint64_t dst_address, uint32_t num_by
 	}
 	const auto value = Sync::ReadReferenceClock();
 	std::memcpy(reinterpret_cast<void*>(dst_address), &value, num_bytes);
-	LOGF("\t copy_data reference clock: dst=0x%016" PRIx64 " value=0x%016" PRIx64 " size=%u\n",
-	     dst_address, value, num_bytes);
+	static std::atomic<uint32_t> clock_log_count {0};
+	if (clock_log_count.fetch_add(1) < 64) {
+		LOGF("\t copy_data reference clock: dst=0x%016" PRIx64 " value=0x%016" PRIx64
+		     " size=%u\n",
+		     dst_address, value, num_bytes);
+	}
 }
 
 void CommandProcessor::DmaData(uint8_t engine, uint8_t dst_sel, uint8_t dst_cache_policy,
@@ -672,20 +676,16 @@ Pm4ProcessResult CommandProcessor::Process(Pm4Execution&             execution,
 		Pm4Execution*     previous_execution;
 	} execution_scope(*this, execution);
 
-	ProcessPm4(execution, 0);
+	ProcessPm4(execution);
 	return execution.m_buffer_stack.empty() ? Pm4ProcessResult::Complete
 	                                        : Pm4ProcessResult::Blocked;
 }
 
-void CommandProcessor::ProcessIndirectBuffer(std::span<const uint32_t> commands) {
+void CommandProcessor::ProcessIndirectBuffer(std::span<const uint32_t> commands, bool chain) {
 	EXIT_IF(g_current_execution == nullptr);
-	if (commands.empty()) {
-		return;
-	}
-	auto&      execution  = *g_current_execution;
-	const auto stop_depth = execution.m_buffer_stack.size();
-	execution.m_buffer_stack.push_back({commands});
-	ProcessPm4(execution, stop_depth);
+	EXIT_IF(!g_current_execution->m_next_buffer.empty());
+	g_current_execution->m_next_buffer = commands;
+	g_current_execution->m_chain       = chain;
 }
 
 void CommandProcessor::SuspendPm4() {
@@ -693,21 +693,13 @@ void CommandProcessor::SuspendPm4() {
 	g_current_execution->m_suspended = true;
 }
 
-void CommandProcessor::ProcessPm4(Pm4Execution& execution, size_t stop_depth) {
-	while (execution.m_buffer_stack.size() > stop_depth) {
+void CommandProcessor::ProcessPm4(Pm4Execution& execution) {
+	while (!execution.m_buffer_stack.empty()) {
 		if (g_gpu_state != nullptr) {
 			g_gpu_state->ProcessCommands();
 		}
-		const auto buffer_index = execution.m_buffer_stack.size() - 1;
-		auto&      cursor       = execution.m_buffer_stack[buffer_index];
+		auto& cursor = execution.m_buffer_stack.back();
 		EXIT_IF(cursor.offset_dw > cursor.commands.size());
-		if (cursor.deferred_advance_dw != 0) {
-			EXIT_IF(cursor.deferred_advance_dw > cursor.commands.size() - cursor.offset_dw);
-			cursor.offset_dw += cursor.deferred_advance_dw;
-			cursor.deferred_advance_dw = 0;
-			execution.m_made_progress  = true;
-			continue;
-		}
 		if (cursor.offset_dw == cursor.commands.size()) {
 			execution.m_buffer_stack.pop_back();
 			continue;
@@ -783,14 +775,19 @@ void CommandProcessor::ProcessPm4(Pm4Execution& execution, size_t stop_depth) {
 		    handler(*this, packet_header & ~1u, packet + 1, remaining_dw, total_dw) + 1;
 		EXIT_IF(packet_dw > remaining_dw);
 		if (execution.m_suspended) {
-			if (execution.m_buffer_stack.size() > buffer_index + 1) {
-				execution.m_buffer_stack[buffer_index].deferred_advance_dw = packet_dw;
-			}
 			return;
 		}
-		EXIT_IF(execution.m_buffer_stack.size() != buffer_index + 1);
-		execution.m_buffer_stack[buffer_index].offset_dw += packet_dw;
+		cursor.offset_dw += packet_dw;
 		execution.m_made_progress = true;
+		if (!execution.m_next_buffer.empty()) {
+			// Chains and taken branches reuse the fetcher; only calls retain a return cursor.
+			if (execution.m_chain) {
+				cursor = {execution.m_next_buffer};
+			} else {
+				execution.m_buffer_stack.push_back({execution.m_next_buffer});
+			}
+			execution.m_next_buffer = {};
+		}
 	}
 }
 
@@ -825,36 +822,55 @@ void CommandProcessor::SetNumInstances(uint32_t num_instances) {
 
 void CommandProcessor::SetPredication(uint32_t condition, uint32_t op, uint32_t wait_op,
                                       const volatile void* address, uint32_t count_in_dwords) {
-	if (wait_op != 0) {
-		BufferFlushAndWait();
-	}
-
 	(void)count_in_dwords;
+	uint64_t value = 0;
 
 	switch (op) {
-		case 0x00: {
+		case 0x00:
 			m_predicate_skip = false;
-		} break;
-		case 0x03: {
+			return;
+		case 0x01: {
 			EXIT_NOT_IMPLEMENTED(address == nullptr);
-
-			auto value = *reinterpret_cast<const volatile uint64_t*>(address);
-
-			switch (condition) {
-				case 0x00: m_predicate_skip = (value != 0); break;
-				case 0x01: m_predicate_skip = (value == 0); break;
-				default: EXIT("unknown predication condition: 0x%08" PRIx32 "\n", condition);
-			}
-			static std::atomic<uint32_t> log_count {0};
-			if (log_count.fetch_add(1) < 128) {
-				LOGF("\t bool predication: addr=0x%016" PRIx64 ", value=0x%016" PRIx64
-				     ", condition=%" PRIu32 ", skip=%u, wait_op=%" PRIu32 "\n",
-				     reinterpret_cast<uint64_t>(address), value, condition,
-				     m_predicate_skip ? 1u : 0u, wait_op);
+			// One begin/end pair per DB; bit 63 marks each counter ready.
+			constexpr uint64_t ready_bit = 1ull << 63u;
+			const auto* results = reinterpret_cast<const volatile uint64_t*>(address);
+			for (uint32_t db = 0; db < 16u; db++) {
+				const auto begin = results[db * 2u];
+				const auto end   = results[db * 2u + 1u];
+				if ((begin & end & ready_bit) == 0) {
+					if (wait_op == 0) {
+						SuspendPm4();
+					} else {
+						m_predicate_skip = false;
+					}
+					return;
+				}
+				value += end - begin;
 			}
 		} break;
+		case 0x03:
+			if (wait_op != 0) {
+				BufferFlushAndWait();
+			}
+			EXIT_NOT_IMPLEMENTED(address == nullptr);
+			value = *reinterpret_cast<const volatile uint64_t*>(address);
+			break;
 		default: EXIT("unknown predication op: 0x%08" PRIx32 "\n", op);
 	}
+	switch (condition) {
+		case 0x00: m_predicate_skip = (value != 0); break;
+		case 0x01: m_predicate_skip = (value == 0); break;
+		default: EXIT("unknown predication condition: 0x%08" PRIx32 "\n", condition);
+	}
+	if (op == 0x03) {
+		static std::atomic<uint32_t> log_count {0};
+		if (log_count.fetch_add(1) < 128) {
+			LOGF("\t bool predication: addr=0x%016" PRIx64 ", value=0x%016" PRIx64
+			     ", condition=%" PRIu32 ", skip=%u, wait_op=%" PRIu32 "\n",
+			     reinterpret_cast<uint64_t>(address), value, condition,
+			     m_predicate_skip ? 1u : 0u, wait_op);
+		}
+	}
 }
 
 void CommandProcessor::DrawIndex(DrawIndexArgs args) {
@@ -1090,17 +1106,6 @@ void CommandProcessor::DispatchDirect(uint32_t thread_group_x, uint32_t thread_g
 		// local_x        = std::max(cs.num_thread_x, 1u);
 		// local_y        = std::max(cs.num_thread_y, 1u);
 		// local_z        = std::max(cs.num_thread_z, 1u);
-		if (cs.wave_size == 64u) {
-			static std::atomic_bool logged_wave64_shader {false};
-			if (!logged_wave64_shader.exchange(true, std::memory_order_relaxed)) {
-				LOGF("warning: executing wave64 compute shader cs=0x%016" PRIx64 "\n",
-				     cs.data_addr);
-				std::printf("warning: executing wave64 compute shader cs=0x%016" PRIx64 "\n",
-				            cs.data_addr);
-				std::fflush(stdout);
-			}
-		}
-
 		m_renderer.GetRenderExecutor().DispatchDirect(m_submit_id, CurrentBuffer(), thread_group_x,
 		                                              thread_group_y, thread_group_z, mode);
 	}
@@ -1292,6 +1297,7 @@ void CommandProcessor::WriteAtEndOfPipe(uint32_t cache_policy, uint32_t event_wr
 								break;
 							case 0x14:
 							case 0x28:
+							case 0x2f:
 								if (event_index == 0x00) {
 									write64(false);
 									return;
@@ -1299,7 +1305,6 @@ void CommandProcessor::WriteAtEndOfPipe(uint32_t cache_policy, uint32_t event_wr
 								break;
 							case 0x2b:
 							case 0x2d:
-							case 0x2f:
 							case 0x30:
 								if (event_index == 0x00 && !with_interrupt) {
 									write64(false);

src/graphics/guest_gpu/hardwareContext.h

@@ -498,7 +498,8 @@ struct LsStageRegisters {
 };
 
 struct HsStageRegisters {
-	uint64_t          data_addr = 0;
+	uint64_t          data_addr      = 0;
+	uint64_t          user_data_addr = 0;
 	HsShaderResource1 rsrc1;
 	HsShaderResource2 rsrc2;
 };
@@ -935,6 +936,12 @@ private:
 	ShaderRegisters m_sh_regs;
 };
 
+struct FsrView {
+	uint32_t control_points[2][4] {};
+	uint32_t alphas[2][2] {};
+	uint32_t window[2] {};
+};
+
 class UserConfig {
 public:
 	UserConfig()  = default;
@@ -964,8 +971,17 @@ public:
 	void SetGdsOaCntl(uint32_t value) { m_gds_oa.cntl = value; }
 	void SetGdsOaCounter(uint32_t value) { m_gds_oa.counters[m_gds_oa.GetIndex()].counter = value; }
 	void SetGdsOaAddress(uint32_t value) { m_gds_oa.counters[m_gds_oa.GetIndex()].address = value; }
+	void SetFsrControlPoint(uint32_t axis, uint32_t index, uint32_t value) {
+		m_fsr_view.control_points[axis][index] = value;
+	}
+	void SetFsrAlpha(uint32_t axis, uint32_t index, uint32_t value) {
+		m_fsr_view.alphas[axis][index] = value;
+	}
+	void SetFsrWindow(uint32_t index, uint32_t value) { m_fsr_view.window[index] = value; }
+	[[nodiscard]] const FsrView& GetFsrView() const { return m_fsr_view; }
 
 private:
+	FsrView m_fsr_view;
 	Prospero::PrimitiveType m_prim_type               = Prospero::PrimitiveType::kNone;
 	uint32_t                m_index_offset            = 0;
 	uint32_t                m_object_id               = 0;
@@ -988,14 +1004,14 @@ public:
 	void SetEsShaderBase(uint64_t addr) { m_vs.es_regs.data_addr = addr; }
 	void SetLsShaderBase(uint64_t addr) { m_vs.ls_regs.data_addr = addr; }
 	void SetHsShaderBase(uint64_t addr) { m_vs.hs_regs.data_addr = addr; }
+	void SetHsUserDataAddress(uint32_t word, uint32_t value) {
+		SetUserDataAddressWord(m_vs.hs_regs.user_data_addr, word, value);
+	}
 	void SetHsShaderResource1(const HsShaderResource1& rsrc1) { m_vs.hs_regs.rsrc1 = rsrc1; }
 	void SetHsShaderResource2(const HsShaderResource2& rsrc2) { m_vs.hs_regs.rsrc2 = rsrc2; }
 	void SetGsShaderBase(uint64_t addr) { m_vs.gs_regs.data_addr = addr; }
 	void SetGsUserDataAddress(uint32_t word, uint32_t value) {
-		const auto shift   = word * 32u;
-		auto&      address = m_vs.gs_regs.user_data_addr;
-		address = (address & ~(uint64_t {0xffffffffu} << shift)) |
-		          (static_cast<uint64_t>(value) << shift);
+		SetUserDataAddressWord(m_vs.gs_regs.user_data_addr, word, value);
 	}
 	void SetGsShaderResource1(const GsShaderResource1& rsrc1) { m_vs.gs_regs.rsrc1 = rsrc1; }
 	void SetGsShaderResource2(const GsShaderResource2& rsrc2) { m_vs.gs_regs.rsrc2 = rsrc2; }
@@ -1037,6 +1053,12 @@ public:
 	[[nodiscard]] const ComputeShaderInfo& GetCs() const { return m_cs; }
 
 private:
+	static void SetUserDataAddressWord(uint64_t& address, uint32_t word, uint32_t value) {
+		const auto shift = word * 32u;
+		address = (address & ~(uint64_t {0xffffffffu} << shift)) |
+		          (static_cast<uint64_t>(value) << shift);
+	}
+
 	VertexShaderInfo  m_vs;
 	PixelShaderInfo   m_ps;
 	ComputeShaderInfo m_cs;

src/graphics/guest_gpu/pm4.h

@@ -987,6 +987,12 @@ constexpr uint32_t GE_USER_VGPR_EN_EN_USER_VGPR2_MASK  = 0x1;
 constexpr uint32_t GE_USER_VGPR_EN_EN_USER_VGPR3_SHIFT = 2;
 constexpr uint32_t GE_USER_VGPR_EN_EN_USER_VGPR3_MASK  = 0x1;
 
+constexpr uint32_t FSR_WINDOW_LEFT           = 0x2BC;
+constexpr uint32_t FSR_CONTROL_POINTS_LEFT_X = 0x2C4;
+constexpr uint32_t FSR_CONTROL_POINTS_LEFT_Y = 0x2CC;
+constexpr uint32_t FSR_ALPHA_LEFT_X          = 0x2E4;
+constexpr uint32_t FSR_ALPHA_LEFT_Y          = 0x2E8;
+
 constexpr uint32_t TA_CS_BC_BASE_ADDR       = 0x380;
 constexpr uint32_t TA_CS_BC_BASE_ADDR_HI    = 0x381;
 constexpr uint32_t TEXTURE_GRADIENT_FACTORS = 0x383;

src/graphics/guest_gpu/tile.h

@@ -86,6 +86,7 @@ struct TileMipLayout {
 	uint32_t padded_height = 0;
 	uint32_t tail_x        = 0;
 	uint32_t tail_y        = 0;
+	bool     operator==(const TileMipLayout&) const = default;
 };
 
 struct TileSurfaceLayout {

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

@@ -44,12 +44,7 @@ FaultManager::FaultManager(GraphicContext& graphics, CommandScheduler& scheduler
 	                                                &m_fault_process_desc_layout),
 	    "create fault-buffer descriptor layout");
 
-	vk::ShaderModuleCreateInfo module_info {};
-	module_info.codeSize = std::size(FAULT_BUFFER_PROCESS_SPV) * sizeof(uint32_t);
-	module_info.pCode    = FAULT_BUFFER_PROCESS_SPV;
-	vk::ShaderModule module = nullptr;
-	RequireVulkanSuccess(m_graphics.device.createShaderModule(&module_info, nullptr, &module),
-	                     "create fault-buffer shader module");
+	const auto module = CompileSPV(FAULT_BUFFER_PROCESS_SPV, m_graphics.device);
 
 	vk::PipelineLayoutCreateInfo pipeline_layout_info {};
 	pipeline_layout_info.setLayoutCount = 1;

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

@@ -33,22 +33,24 @@ namespace {
 
 constexpr uint64_t NumFramesBeforeRemoval = 32;
 
-[[nodiscard]] bool DecodeDccClear(const TextureCache::ImageDesc& desc, vk::Format format,
-                                  uint32_t fill, vk::ClearColorValue& clear) {
-	const auto code = static_cast<uint8_t>(fill);
-	if (fill != static_cast<uint32_t>(code) * 0x01010101u) {
-		return false;
+[[nodiscard]] bool DecodeDccClear(const TextureCache::ImageDesc& desc, uint8_t code,
+                                  vk::ClearColorValue& clear) {
+	switch (code) {
+		case 0x00:
+		case 0x20:
+		case 0x40:
+		case 0x80:
+		case 0xc0: break;
+		default: return false;
 	}
 	const auto& metadata = desc.info.metadata;
+	const auto  format   = desc.view_info.format;
 	if (code == 0x20) {
 		// Clear-to-register is a color-buffer operation; the texture pipe cannot decode it.
 		return desc.type == TextureCache::BindingType::RenderTarget &&
 		       metadata.dcc_clear_register_valid &&
 		       DecodePackedColorClear(format, metadata.dcc_clear_word, clear);
 	}
-	if (code != 0x00 && code != 0x40 && code != 0x80 && code != 0xc0) {
-		return false;
-	}
 	clear = {};
 	if (code == 0x00) {
 		return true;
@@ -206,6 +208,10 @@ bool TextureCache::SameBacking(const ImageInfo& cached, const ImageInfo& request
 	if (cached.extent != requested.extent) {
 		return false;
 	}
+	if (cached.resources.levels < requested.resources.levels ||
+	    cached.resources.layers < requested.resources.layers) {
+		return false;
+	}
 	if (cached.samples != requested.samples) {
 		return false;
 	}
@@ -323,10 +329,8 @@ void TextureCache::DeleteImage(ImageId id) {
 	if (image->info.HasMetadata()) {
 		const auto metadata = m_surface_metas.find(image->info.metadata.range.address);
 		if (metadata != m_surface_metas.end() &&
-		    ((image->info.metadata.kind == ImageMetadataKind::Dcc &&
-		      metadata->second.type == MetaDataInfo::Type::Dcc) ||
-		     (image->info.metadata.kind == ImageMetadataKind::Htile &&
-		      metadata->second.type == MetaDataInfo::Type::HTile))) {
+		    image->info.metadata.kind == ImageMetadataKind::Htile &&
+		    metadata->second.type == MetaDataInfo::Type::HTile) {
 			// A later binding may have reused this address for another metadata type.
 			m_surface_metas.erase(metadata);
 		}
@@ -810,17 +814,10 @@ TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& reques
 	const bool safe_to_delete =
 	    current_tick - std::min(current_tick, cached.tick_accessed_last) > NumFramesBeforeRemoval;
 
-	if (requested.data.address == cached.info.data.address) {
-		const uint32_t requested_block = requested.bytes_per_block * requested.samples;
-		const uint32_t cached_block    = cached.info.bytes_per_block * cached.info.samples;
-		if (requested.BlockExtent() != cached.info.BlockExtent() ||
-		    requested_block != cached_block) {
-			if (safe_to_delete) {
-				FreeImage(cached_id);
-			}
-			return {merged_id};
-		}
-
+	const uint32_t requested_block = requested.bytes_per_block * requested.samples;
+	const uint32_t cached_block    = cached.info.bytes_per_block * cached.info.samples;
+	if (requested.data.address == cached.info.data.address &&
+	    requested.BlockExtent() == cached.info.BlockExtent() && requested_block == cached_block) {
 		if (const auto depth_id = ResolveDepthOverlap(requested, binding, cached_id)) {
 			return {depth_id};
 		}
@@ -850,6 +847,16 @@ TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& reques
 		    ImageViewOps::FormatsCompatible(cached.info.pixel_format, requested.pixel_format)) {
 			return {ExpandImage(requested, cached_id)};
 		}
+		// PS5 mip tails can expose more levels without increasing the guest allocation.
+		if (requested.pixel_format == cached.info.pixel_format &&
+		    requested.type == cached.info.type && requested.resources > cached.info.resources &&
+		    (requested.data.size > cached.info.data.size ||
+		     (requested.data.size == cached.info.data.size &&
+		      requested.extent == cached.info.extent &&
+		      cached.info.resources.levels > 1 &&
+		      requested.resources.layers == cached.info.resources.layers))) {
+			return {ExpandImage(requested, cached_id)};
+		}
 		if (requested.pixel_format != cached.info.pixel_format ||
 		    requested.data.size <= cached.info.data.size) {
 			const auto result_id = merged_id ? merged_id : cached_id;
@@ -859,9 +866,6 @@ TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& reques
 			            ? result_id
 			            : ImageId {}};
 		}
-		if (requested.type == cached.info.type && requested.resources > cached.info.resources) {
-			return {ExpandImage(requested, cached_id)};
-		}
 		EXIT("TextureCache: unresolvable equal-address image overlap, address=0x%016" PRIx64
 		     " requested=%ux%u "
 		     "cached=%ux%u requested_size=0x%016" PRIx64 " cached_size=0x%016" PRIx64
@@ -873,38 +877,27 @@ TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& reques
 		     static_cast<uint32_t>(cached.info.tile_mode));
 	}
 
-	if (requested.data.address > cached.info.data.address) {
-		const int32_t mip = requested.MipOf(cached.info);
-		if (mip >= 0) {
-			const int32_t layer = requested.SliceOf(cached.info, mip);
-			if (layer >= 0) {
-				return {cached_id, mip, layer};
-			}
-		}
-		if (safe_to_delete) {
-			FreeImage(cached_id);
-		}
-		return {};
+	const int32_t requested_mip = requested.MipOf(cached.info);
+	if (requested_mip >= 0) {
+		const int32_t layer = requested.SliceOf(cached.info, requested_mip);
+		return {cached_id, requested_mip, layer};
 	}
 
 	const int32_t mip = cached.info.MipOf(requested);
 	if (mip >= 0) {
 		const int32_t layer = cached.info.SliceOf(requested, mip);
-		if (layer >= 0) {
-			if (cached.binding.is_target) {
-				cached.binding.needs_rebind = true;
-				if (merged_id) {
-					m_slot_images[merged_id].binding.is_target = true;
-				}
-				FreeImage(cached_id);
-				return {merged_id};
-			}
-			if (merged_id) {
-				CopyImageMip(merged_id, cached_id, static_cast<uint32_t>(mip),
-				             static_cast<uint32_t>(layer));
-				FreeImage(cached_id);
-			}
+		if (!merged_id) {
+			return {ExpandImage(requested, cached_id)};
 		}
+		cached.binding.needs_rebind |= cached.binding.is_bound || cached.binding.is_target;
+		m_slot_images[merged_id].binding.is_target |= cached.binding.is_target;
+		CopyImageMip(merged_id, cached_id, static_cast<uint32_t>(mip),
+		             static_cast<uint32_t>(layer));
+		FreeImage(cached_id);
+		return {merged_id};
+	}
+	if (requested.data.address >= cached.info.data.address && safe_to_delete) {
+		FreeImage(cached_id);
 	}
 	return {merged_id};
 }
@@ -919,7 +912,14 @@ ImageId TextureCache::ExpandImage(const ImageInfo& info, ImageId source_id) {
 		source.binding.needs_rebind = true;
 	}
 	InitializeImage(expanded_id);
-	CopyImage(expanded_id, source_id);
+	const int32_t mip = source.info.MipOf(info);
+	const int32_t layer = source.info.SliceOf(info, mip);
+	if (layer >= 0) {
+		CopyImageMip(expanded_id, source_id, static_cast<uint32_t>(mip),
+		             static_cast<uint32_t>(layer));
+	} else {
+		CopyImage(expanded_id, source_id);
+	}
 	FreeImage(source_id);
 	return expanded_id;
 }
@@ -1135,53 +1135,69 @@ void TextureCache::InitializeImage(ImageId id) {
 	}
 }
 
-void TextureCache::PrepareDccClear(ImageId id, const ImageDesc& desc) {
+void TextureCache::MaterializeDccClear(ImageId id, const ImageDesc& desc,
+                                       uint32_t metadata_base_layer) {
 	if (desc.info.metadata.kind != ImageMetadataKind::Dcc) {
 		return;
 	}
-	auto& image            = m_slot_images[id];
-	image.info.metadata    = desc.info.metadata;
-	auto [entry, inserted] = m_surface_metas.try_emplace(
-	    desc.info.metadata.range.address, MetaDataInfo {.type = MetaDataInfo::Type::Dcc});
-	auto& metadata = entry->second;
-	if (!inserted && metadata.type == MetaDataInfo::Type::PendingDcc) {
-		metadata.type = MetaDataInfo::Type::Dcc;
-	} else if (metadata.type != MetaDataInfo::Type::Dcc) {
-		EXIT("TextureCache: image reuses non-DCC metadata\n");
+	const auto range = desc.info.metadata.range;
+	{
+		std::scoped_lock lock {m_lock};
+		auto& image         = m_slot_images[id];
+		image.info.metadata = desc.info.metadata;
+		// A native DCC allocation must not retain a reused HTile/CMask/FMask interpretation.
+		m_surface_metas.erase(range.address);
+		if (range.size == 0 || desc.info.resources.levels != 1 || image.info.resources.levels != 1) {
+			return;
+		}
 	}
-	if (metadata.clear_mask == 0 || image.info.resources.levels != 1 ||
-	    desc.info.metadata.range.size == 0 || metadata.fill_size < desc.info.metadata.range.size) {
-		return;
-	}
-	vk::ClearValue clear {};
-	if (!DecodeDccClear(desc, image.backing.format, metadata.fill_value, clear.color)) {
-		return;
+	const auto layers = desc.info.TransferLayers();
+	// These one-mip surfaces use complete 4 KiB DCC metadata blocks.
+	constexpr uint64_t MetadataBlockSize = 0x1000;
+	if (!range.Valid() || range.address % MetadataBlockSize != 0 || layers == 0 ||
+	    range.size % layers != 0 || (range.size / layers) % MetadataBlockSize != 0) {
+		EXIT("TextureCache: DCC slices must contain aligned 4 KiB blocks\n");
 	}
 	const auto& view           = desc.view_info;
-	const bool  volume_texture = image.info.IsVolume() && view.type == vk::ImageViewType::e3D;
-	const auto  first          = volume_texture ? 0u : view.base_layer;
-	const auto  count = volume_texture ? std::max(image.info.extent.depth >> view.base_level, 1u)
-	                                   : view.layer_count;
-	if (first >= 32 || count > 32 - first) {
-		return;
+	const bool  volume_texture = desc.info.IsVolume() && view.type == vk::ImageViewType::e3D;
+	const auto  first          = volume_texture ? 0u : metadata_base_layer;
+	const auto  image_first    = volume_texture ? 0u : view.base_layer;
+	const auto  count          = volume_texture ? desc.info.extent.depth : view.layer_count;
+	if (first >= layers || count > layers - first) {
+		EXIT("TextureCache: DCC view exceeds its native metadata slices\n");
 	}
-	// The metadata fill covers the complete allocation. Consume each layer only after its
-	// native image contents exist; already materialized layers may have been rendered since.
-	for (uint32_t layer = first; layer < first + count;) {
-		if ((metadata.clear_mask & (1u << layer)) == 0) {
-			layer++;
+	// Finish native metadata writes before reading backing bytes. This can submit the scheduler,
+	// so discovery runs before final draw uploads and never holds the texture lock across it.
+	if (m_buffer_cache.IsRegionGpuModified(range.address, range.size)) {
+		m_buffer_cache.ReadMemory(range.address, range.size, false);
+	}
+	const auto slice_size = range.size / layers;
+	for (uint32_t slice = 0; slice < count; slice++) {
+		const auto address = range.address + slice_size * (first + slice);
+		uint8_t code = 0;
+		if (!LibKernel::Memory::TryReadBacking(address, &code, sizeof(code))) {
+			EXIT("TextureCache: failed to read DCC metadata backing\n");
+		}
+		vk::ClearValue clear {};
+		if (!DecodeDccClear(desc, code, clear.color)) {
 			continue;
 		}
-		const auto start = layer;
-		uint32_t   mask  = 0;
-		do {
-			mask |= 1u << layer++;
-		} while (layer < first + count && (metadata.clear_mask & (1u << layer)) != 0);
-		ClearImage(m_scheduler.Current(), id,
-		           {vk::ImageAspectFlagBits::eColor, view.base_level, view.level_count, start,
-		            layer - start},
-		           clear);
-		metadata.clear_mask &= ~mask;
+		std::vector<uint8_t> bytes(slice_size);
+		if (!LibKernel::Memory::TryReadBacking(address, bytes.data(), bytes.size())) {
+			EXIT("TextureCache: failed to read DCC metadata slice\n");
+		}
+		if (!std::all_of(bytes.begin(), bytes.end(), [code](uint8_t byte) { return byte == code; })) {
+			continue;
+		}
+		{
+			std::scoped_lock lock {m_lock};
+			ClearImage(m_scheduler.Current(), id, view.format,
+			           {vk::ImageAspectFlagBits::eColor, view.base_level, view.level_count,
+			            image_first + slice, 1}, clear);
+		}
+		// Native expanded keys own consumption. Existing buffer tracking publishes this CPU
+		// write to future GPU readers; FillBuffer can fault and must run outside the texture lock.
+		m_buffer_cache.FillBuffer(address, slice_size, UINT32_MAX, false);
 	}
 }
 
@@ -1221,7 +1237,8 @@ void TextureCache::AssociateStencil(ImageId depth_id, GuestRange stencil) {
 	ImageId association {};
 	for (const auto id: FindImagesInRegion(stencil.address, stencil.size, false)) {
 		const auto owner = m_slot_images.try_get(id);
-		if (owner != nullptr && owner->info.data.address == stencil.address) {
+		if (owner != nullptr && owner->info.data == stencil &&
+		    owner->info.extent == depth.info.extent) {
 			association = id;
 		}
 	}
@@ -1246,6 +1263,7 @@ ImageId TextureCache::FindImage(ImageDesc& desc, bool exact_format) {
 		std::scoped_lock lock {m_lock};
 		return GetNullImage(desc);
 	}
+	const auto metadata_base_layer = desc.view_info.base_layer;
 
 	ImageId result {};
 	{
@@ -1313,6 +1331,7 @@ ImageId TextureCache::FindImage(ImageDesc& desc, bool exact_format) {
 		image.tick_accessed_last = m_scheduler.CurrentTick();
 		TouchImage(image);
 	}
+	MaterializeDccClear(result, desc, metadata_base_layer);
 	return result;
 }
 
@@ -1376,7 +1395,6 @@ vk::ImageView TextureCache::FindTexture(ImageId id, const ImageDesc& desc) {
 		image.MarkGpuModified();
 	}
 	if (!image.info.data.Empty()) {
-		PrepareDccClear(id, desc);
 		RefreshImage(id);
 	}
 	switch (desc.type) {
@@ -1406,7 +1424,6 @@ vk::ImageView TextureCache::FindRenderTarget(ImageId id, const ImageDesc& desc)
 	TouchImage(image);
 	image.MarkGpuModified();
 	image.usage.render_target = true;
-	PrepareDccClear(id, desc);
 	RefreshImage(id);
 	CommitGpuWrite(image);
 	TrackImageDownload(id, image);
@@ -1430,16 +1447,9 @@ vk::ImageView TextureCache::FindDepthTarget(ImageId id, const ImageDesc& desc) {
 	RefreshImage(id);
 	if (desc.info.HasMetadata()) {
 		image.info.metadata = desc.info.metadata;
-		auto [metadata, inserted] =
-		    m_surface_metas.try_emplace(desc.info.metadata.range.address,
-		                                MetaDataInfo {.type       = MetaDataInfo::Type::HTile,
-		                                              .clear_mask = image.info.htile_clear_mask});
-		if (!inserted && metadata->second.type != MetaDataInfo::Type::HTile) {
-			// PS5 allocations can reuse DCC storage as HTile while the old color image is cached.
-			// The depth binding defines the new type; incompatible fill state cannot carry over.
-			metadata->second = {.type       = MetaDataInfo::Type::HTile,
-			                    .clear_mask = image.info.htile_clear_mask};
-		}
+		m_surface_metas.emplace(desc.info.metadata.range.address,
+		                        MetaDataInfo {.type       = MetaDataInfo::Type::HTile,
+		                                      .clear_mask = image.info.htile_clear_mask});
 	}
 	CommitGpuWrite(image);
 	if (desc.info.HasStencil()) {
@@ -1527,12 +1537,12 @@ bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address
 		}
 		clear.depthStencil.stencil = stencil_clear;
 	}
-	ClearImage(command, selected,
+	ClearImage(command, selected, image.backing.format,
 	           {aspect, 0, image.info.resources.levels, 0, image.info.TransferLayers()}, clear);
 	return true;
 }
 
-void TextureCache::ClearImage(CommandBuffer& command, ImageId id,
+void TextureCache::ClearImage(CommandBuffer& command, ImageId id, vk::Format format,
                               const vk::ImageSubresourceRange& range, const vk::ClearValue& clear) {
 	auto& image = m_slot_images[id];
 	const auto aspects = image.info.IsDepth() ? ImageViewOps::DepthAspectMask(image.backing.format)
@@ -1557,11 +1567,12 @@ void TextureCache::ClearImage(CommandBuffer& command, ImageId id,
 		}
 	}
 	command.EndRendering();
-	if (image.info.IsVolume() && !full_image) {
+	// Transfer clears use the backing format; aliased clears must encode through their view.
+	if (format != image.backing.format || (image.info.IsVolume() && !full_image)) {
 		EXIT_NOT_IMPLEMENTED(range.aspectMask != vk::ImageAspectFlagBits::eColor ||
 		                     range.levelCount != 1);
 		ImageViewInfo view {};
-		view.format = image.backing.format;
+		view.format = format;
 		view.type   = range.layerCount == 1 ? vk::ImageViewType::e2D : vk::ImageViewType::e2DArray;
 		view.base_level  = range.baseMipLevel;
 		view.base_layer  = range.baseArrayLayer;
@@ -1872,72 +1883,32 @@ void TextureCache::InvalidateCpuAliases(uint64_t address, uint64_t size) {
 bool TextureCache::IsMeta(uint64_t address) {
 	std::scoped_lock lock {m_lock};
 	const auto       found = m_surface_metas.find(address);
-	return found != m_surface_metas.end() && found->second.type != MetaDataInfo::Type::PendingDcc;
+	return found != m_surface_metas.end();
 }
 
-bool TextureCache::IsMetaCleared(uint64_t address, uint32_t slice, uint32_t* fill_value) {
+bool TextureCache::IsMetaCleared(uint64_t address, uint32_t slice) {
 	std::scoped_lock lock {m_lock};
 	const auto       found = m_surface_metas.find(address);
-	if (found == m_surface_metas.end() || found->second.type == MetaDataInfo::Type::PendingDcc ||
-	    slice >= 32) {
+	if (found == m_surface_metas.end() || slice >= 32) {
 		return false;
 	}
-	if (fill_value != nullptr) {
-		*fill_value = found->second.fill_value;
-	}
 	return (found->second.clear_mask & (1u << slice)) != 0;
 }
 
 bool TextureCache::ClearMeta(uint64_t address) {
 	std::scoped_lock lock {m_lock};
 	const auto       found = m_surface_metas.find(address);
-	if (found == m_surface_metas.end() || found->second.type == MetaDataInfo::Type::PendingDcc ||
-	    found->second.type == MetaDataInfo::Type::Dcc) {
-		// Preserve the broad metadata-clear operation for CMask/FMask/HTile. DCC requires a
-		// validated fill value, so an arbitrary compute write must not clear it.
+	if (found == m_surface_metas.end()) {
 		return false;
 	}
 	found->second.clear_mask = UINT32_MAX;
 	return true;
 }
 
-void TextureCache::TrackDccFill(uint64_t address, uint64_t size, uint32_t fill_value) {
-	if (!GuestRange {address, size}.Valid()) {
-		EXIT("TextureCache: invalid DCC fill range\n");
-	}
-	// DCC fills use a repeated byte code. Require all four bytes of the detected dword to agree,
-	// and mark only recognized deferred-clear encodings as logically clear.
-	const auto dcc_clear_mask = [fill_value] {
-		const auto code = static_cast<uint8_t>(fill_value);
-		if (fill_value != static_cast<uint32_t>(code) * 0x01010101u) {
-			return 0u;
-		}
-		switch (code) {
-			case 0x00:
-			case 0x20:
-			case 0x40:
-			case 0x80:
-			case 0xc0: return UINT32_MAX;
-			default: return 0u;
-		}
-	}();
-	std::scoped_lock lock {m_lock};
-	// The guest dispatch still writes metadata. An unknown address remains PendingDcc until an
-	// image descriptor confirms its role; never reinterpret CMask/FMask/HTile as DCC.
-	const auto found = m_surface_metas.try_emplace(address).first;
-	if (found->second.type == MetaDataInfo::Type::PendingDcc ||
-	    found->second.type == MetaDataInfo::Type::Dcc) {
-		found->second.clear_mask = dcc_clear_mask;
-		found->second.fill_value = fill_value;
-		found->second.fill_size  = size;
-	}
-}
-
 bool TextureCache::TouchMeta(uint64_t address, uint32_t slice, bool is_clear) {
 	std::scoped_lock lock {m_lock};
 	const auto       found = m_surface_metas.find(address);
-	if (found == m_surface_metas.end() || found->second.type == MetaDataInfo::Type::PendingDcc ||
-	    slice >= 32) {
+	if (found == m_surface_metas.end() || slice >= 32) {
 		return false;
 	}
 	if (is_clear) {
@@ -1953,10 +1924,13 @@ void TextureCache::UnmapMemory(uint64_t address, uint64_t size) {
 		EXIT("TextureCache: invalid unmap range\n");
 	}
 	std::scoped_lock lock {m_lock};
-	const auto       end = address + size;
-	for (auto metadata = m_surface_metas.lower_bound(address);
-	     metadata != m_surface_metas.end() && metadata->first < end;) {
-		metadata = m_surface_metas.erase(metadata);
+	for (auto metadata = m_surface_metas.begin(); metadata != m_surface_metas.end();) {
+		const auto base = metadata->first;
+		if (base >= address && base < address + size) {
+			metadata = m_surface_metas.erase(metadata);
+		} else {
+			++metadata;
+		}
 	}
 	auto images = FindImagesInRegion(address, size, false);
 	for (const auto id: images) {

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

@@ -64,11 +64,8 @@ public:
 	[[nodiscard]] bool IsRegionGpuModified(uint64_t address, uint64_t size);
 
 	[[nodiscard]] bool IsMeta(uint64_t address);
-	[[nodiscard]] bool IsMetaCleared(uint64_t address, uint32_t slice,
-	                                 uint32_t* fill_value = nullptr);
+	[[nodiscard]] bool IsMetaCleared(uint64_t address, uint32_t slice);
 	[[nodiscard]] bool ClearMeta(uint64_t address);
-	// Record deferred DCC state while the original guest dispatch writes the metadata.
-	void               TrackDccFill(uint64_t address, uint64_t size, uint32_t fill_value);
 	[[nodiscard]] bool TouchMeta(uint64_t address, uint32_t slice, bool is_clear);
 
 	void UnmapMemory(uint64_t address, uint64_t size);
@@ -81,17 +78,10 @@ private:
 	struct ImageDownload;
 
 	struct MetaDataInfo {
-		// A guest metadata-fill dispatch may initialize DCC before its render target is bound.
-		// PendingDcc retains that exact fill until an image binding classifies the address,
-		// without exposing an unconfirmed buffer address to the normal metadata heuristics.
-		// Keep all surface metadata in one entry so CMask/FMask can be
-		// registered beside HTile and DCC without introducing parallel tracking paths.
-		enum class Type : uint8_t { PendingDcc, CMask, FMask, HTile, Dcc };
+		enum class Type : uint8_t { CMask, FMask, HTile };
 
-		Type     type       = Type::PendingDcc;
-		uint32_t clear_mask = 0;
-		uint32_t fill_value = 0xffffffffu;
-		uint64_t fill_size  = 0;
+		Type     type;
+		uint32_t clear_mask = UINT32_MAX;
 	};
 
 	struct OverlapResult {
@@ -149,7 +139,8 @@ private:
 	                                                ImageId cached);
 	[[nodiscard]] ImageId       ExpandImage(const ImageInfo& info, ImageId source);
 	void                        RefreshImage(ImageId id);
-	void                        PrepareDccClear(ImageId id, const ImageDesc& desc);
+	void                        MaterializeDccClear(ImageId id, const ImageDesc& desc,
+	                                                uint32_t metadata_base_layer);
 	void                        InitializeImage(ImageId id);
 	[[nodiscard]] TextureTransfer
 	BuildTextureTransfer(const Image& image, BindingType binding, TransferDirection direction) const;
@@ -160,8 +151,8 @@ private:
 	void DownloadDepth(Image& image, Buffer& destination, uint64_t destination_offset);
 	void CommitGpuWrite(Image& image);
 	// Caller holds m_lock. Volume layer ranges select depth slices.
-	void ClearImage(CommandBuffer& command, ImageId id, const vk::ImageSubresourceRange& range,
-	                const vk::ClearValue& clear);
+	void ClearImage(CommandBuffer& command, ImageId id, vk::Format format,
+	                const vk::ImageSubresourceRange& range, const vk::ClearValue& clear);
 	void PrepareImageCopy(Image& image);
 	void RefreshCopySource(ImageId id);
 	[[nodiscard]] bool CopyD16(Image& destination, Image& source);

src/graphics/host_gpu/renderer/depthRenderTarget.cpp

@@ -19,6 +19,7 @@
 #include "graphics/host_gpu/vulkanCommon.h"
 
 #include <algorithm>
+#include <array>
 #include <atomic>
 #include <cmath>
 #include <cstdarg>
@@ -45,8 +46,12 @@ static vk::StencilOp ConvertStencilOp(uint8_t value, uint8_t write_mask, uint8_t
 	switch (static_cast<Prospero::StencilOp>(value)) {
 		case Prospero::StencilOp::kKeep: return vk::StencilOp::eKeep;
 		case Prospero::StencilOp::kZero: return vk::StencilOp::eZero;
-		case Prospero::StencilOp::kReplaceTest:
-		case Prospero::StencilOp::kReplaceOp: return vk::StencilOp::eReplace;
+		case Prospero::StencilOp::kReplaceTest: return vk::StencilOp::eReplace;
+		case Prospero::StencilOp::kReplaceOp:
+			if ((op_value & write_mask) == 0) {
+				return vk::StencilOp::eZero;
+			}
+			return vk::StencilOp::eReplace;
 		case Prospero::StencilOp::kAddClamp: return vk::StencilOp::eIncrementAndClamp;
 		case Prospero::StencilOp::kSubClamp: return vk::StencilOp::eDecrementAndClamp;
 		case Prospero::StencilOp::kInvert: return vk::StencilOp::eInvert;
@@ -66,9 +71,38 @@ static vk::StencilOp ConvertStencilOp(uint8_t value, uint8_t write_mask, uint8_t
 	}
 }
 
-static bool UsesStencilOpValue(uint8_t fail, uint8_t pass, uint8_t depth_fail) {
-	constexpr auto replace_op = static_cast<uint8_t>(Prospero::StencilOp::kReplaceOp);
-	return fail == replace_op || pass == replace_op || depth_fail == replace_op;
+static vk::StencilOpState ConvertStencilState(
+    uint8_t compare, const std::array<uint8_t, 3>& operations, uint8_t op_value,
+    const vk::StencilOpState& state) {
+	const auto test_value = state.reference;
+	auto reference       = test_value;
+	auto required_bits   = state.compareMask;
+	if (compare == static_cast<uint8_t>(vk::CompareOp::eAlways) ||
+	    compare == static_cast<uint8_t>(vk::CompareOp::eNever)) {
+		required_bits = 0;
+	}
+	std::array<vk::StencilOp, 3> converted {};
+	for (size_t i = 0; i < operations.size(); i++) {
+		converted[i] = ConvertStencilOp(operations[i], state.writeMask, op_value);
+		if (converted[i] != vk::StencilOp::eReplace) {
+			continue;
+		}
+		auto replacement = test_value;
+		if (static_cast<Prospero::StencilOp>(operations[i]) == Prospero::StencilOp::kReplaceOp) {
+			replacement = op_value;
+		}
+		if (((reference ^ replacement) & required_bits & state.writeMask) != 0) {
+			DepthFatal("unsupported stencil replacement: compare=%u, compare mask=0x%02" PRIx32
+			           ", write mask=0x%02" PRIx32 ", operation value=0x%02" PRIx8
+			           ", test value=0x%02" PRIx32,
+			           compare, state.compareMask, state.writeMask, op_value, test_value);
+		}
+		// Vulkan shares one reference between comparison and every replacement on this face.
+		reference = (reference & ~state.writeMask) | (replacement & state.writeMask);
+		required_bits |= state.writeMask;
+	}
+	return {converted[0], converted[1], converted[2], static_cast<vk::CompareOp>(compare),
+	        state.compareMask, state.writeMask, reference};
 }
 
 [[nodiscard]] static vk::Format ResolveHostDepthAttachmentFormat(const CommandBuffer&     buffer,
@@ -125,8 +159,8 @@ static TextureCache::ImageDesc MakeDepthTargetDesc(const CommandBuffer& buffer,
 			DepthFatal("invalid depth view: base=%u last=%u", z.depth_view.slice_start,
 			           z.depth_view.slice_max);
 	}
-	if (z.z_info.expclear_enabled || z.stencil_info.expclear_enabled ||
-	    z.z_info.partially_resident ||
+	// EXPCLEAR permits an HTile acceleration state; the host attachment is already expanded.
+	if (z.z_info.partially_resident ||
 	    z.stencil_info.partially_resident || z.z_info.max_mip_level != 0 ||
 	    z.depth_view.current_mip_level != 0 || unsupported_shading_rate_encoding ||
 	    depth_address == 0 || (depth_address & 0xffffu) != 0) {
@@ -296,31 +330,20 @@ void RenderExecutor::ResolveRenderDepthTarget(CommandBuffer& buffer, RenderDepth
 		const uint8_t back_write_mask  = stencil_ops_disabled ? 0 : sm.stencil_writemask_bf;
 		if (dc.stencilfunc > static_cast<uint8_t>(vk::CompareOp::eAlways) ||
 		    (dc.backface_enable &&
-		     dc.stencilfunc_bf > static_cast<uint8_t>(vk::CompareOp::eAlways)) ||
-		    (front_write_mask != 0 &&
-		     UsesStencilOpValue(sc.stencil_fail, sc.stencil_zpass, sc.stencil_zfail) &&
-		     sm.stencil_opval != sm.stencil_testval) ||
-		    (dc.backface_enable && back_write_mask != 0 &&
-		     UsesStencilOpValue(sc.stencil_fail_bf, sc.stencil_zpass_bf, sc.stencil_zfail_bf) &&
-		     sm.stencil_opval_bf != sm.stencil_testval_bf)) {
-			DepthFatal("unsupported stencil compare or replacement state");
+		     dc.stencilfunc_bf > static_cast<uint8_t>(vk::CompareOp::eAlways))) {
+			DepthFatal("unsupported stencil compare state");
 		}
-		r.stencil_static_front = {
-		    ConvertStencilOp(sc.stencil_fail, front_write_mask, sm.stencil_opval),
-		    ConvertStencilOp(sc.stencil_zpass, front_write_mask, sm.stencil_opval),
-		    ConvertStencilOp(sc.stencil_zfail, front_write_mask, sm.stencil_opval),
-		    static_cast<vk::CompareOp>(dc.stencilfunc)};
-		r.stencil_dynamic_front = {sm.stencil_mask, front_write_mask, sm.stencil_testval};
+		r.stencil_front = ConvertStencilState(
+		    dc.stencilfunc, {sc.stencil_fail, sc.stencil_zpass, sc.stencil_zfail},
+		    sm.stencil_opval, {.compareMask = sm.stencil_mask, .writeMask = front_write_mask,
+		                       .reference = sm.stencil_testval});
 		if (dc.backface_enable) {
-			r.stencil_static_back = {
-			    ConvertStencilOp(sc.stencil_fail_bf, back_write_mask, sm.stencil_opval_bf),
-			    ConvertStencilOp(sc.stencil_zpass_bf, back_write_mask, sm.stencil_opval_bf),
-			    ConvertStencilOp(sc.stencil_zfail_bf, back_write_mask, sm.stencil_opval_bf),
-			    static_cast<vk::CompareOp>(dc.stencilfunc_bf)};
-			r.stencil_dynamic_back = {sm.stencil_mask_bf, back_write_mask, sm.stencil_testval_bf};
+			r.stencil_back = ConvertStencilState(
+			    dc.stencilfunc_bf, {sc.stencil_fail_bf, sc.stencil_zpass_bf, sc.stencil_zfail_bf},
+			    sm.stencil_opval_bf, {.compareMask = sm.stencil_mask_bf, .writeMask = back_write_mask,
+			                          .reference = sm.stencil_testval_bf});
 		} else {
-			r.stencil_static_back  = r.stencil_static_front;
-			r.stencil_dynamic_back = r.stencil_dynamic_front;
+			r.stencil_back = r.stencil_front;
 		}
 	}
 	auto& cache = m_context.GetTextureCache();
@@ -406,14 +429,13 @@ vk::ImageAspectFlags RenderDepthInfo::AttachmentWriteAspects() const {
 		return writes;
 	}
 
-	const auto face_writes = [&](const PipelineStencilStaticState&  state,
-	                             const PipelineStencilDynamicState& dynamic) {
-		if (dynamic.writeMask == 0) {
+	const auto face_writes = [&](const vk::StencilOpState& state) {
+		if (state.writeMask == 0) {
 			return false;
 		}
 		bool can_pass = state.compareOp != vk::CompareOp::eNever;
 		bool can_fail = state.compareOp != vk::CompareOp::eAlways;
-		if (dynamic.compareMask == 0) {
+		if (state.compareMask == 0) {
 			switch (state.compareOp) {
 				case vk::CompareOp::eEqual:
 				case vk::CompareOp::eLessOrEqual:
@@ -439,8 +461,7 @@ vk::ImageAspectFlags RenderDepthInfo::AttachmentWriteAspects() const {
 		       (can_pass && depth_fail && state.depthFailOp != vk::StencilOp::eKeep);
 	};
 	if (stencil_clear_enable ||
-	    (stencil_test_enable && (face_writes(stencil_static_front, stencil_dynamic_front) ||
-	                             face_writes(stencil_static_back, stencil_dynamic_back)))) {
+	    (stencil_test_enable && (face_writes(stencil_front) || face_writes(stencil_back)))) {
 		writes |= vk::ImageAspectFlagBits::eStencil;
 	}
 	return writes;

src/graphics/host_gpu/renderer/depthRenderTarget.h

@@ -34,10 +34,8 @@ struct RenderDepthInfo {
 	bool                        stencil_clear_enable     = false;
 	uint8_t                     stencil_clear_value      = 0;
 	bool                        stencil_test_enable      = false;
-	PipelineStencilStaticState  stencil_static_front;
-	PipelineStencilStaticState  stencil_static_back;
-	PipelineStencilDynamicState stencil_dynamic_front;
-	PipelineStencilDynamicState stencil_dynamic_back;
+	vk::StencilOpState          stencil_front;
+	vk::StencilOpState          stencil_back;
 	ImageId                     image_id;
 
 	[[nodiscard]] vk::ImageAspectFlags AttachmentWriteAspects() const;

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

@@ -10,7 +10,6 @@
 
 #include <algorithm>
 #include <array>
-#include <iterator>
 
 namespace Libs::Graphics {
 
@@ -37,9 +36,8 @@ BlitHelper::BlitHelper(GraphicContext& graphics, CommandScheduler& scheduler)
 	    m_graphics.device.createPipelineLayout(&layout_info, nullptr, &m_pipeline_layout),
 	    "create BlitHelper pipeline layout");
 
-	m_vertex_shader = CreateShader(GPU_BLIT_FS_TRIANGLE_SPV, std::size(GPU_BLIT_FS_TRIANGLE_SPV));
-	m_fragment_shader =
-	    CreateShader(GPU_BLIT_COLOR_TO_MS_DEPTH_SPV, std::size(GPU_BLIT_COLOR_TO_MS_DEPTH_SPV));
+	m_vertex_shader   = CompileSPV(GPU_BLIT_FS_TRIANGLE_SPV, m_graphics.device);
+	m_fragment_shader = CompileSPV(GPU_BLIT_COLOR_TO_MS_DEPTH_SPV, m_graphics.device);
 }
 
 BlitHelper::~BlitHelper() {
@@ -60,17 +58,6 @@ BlitHelper::~BlitHelper() {
 	}
 }
 
-vk::ShaderModule BlitHelper::CreateShader(const uint32_t* code, size_t words) const {
-	EXIT_IF(code == nullptr || words == 0);
-	vk::ShaderModuleCreateInfo create {};
-	create.codeSize         = words * sizeof(uint32_t);
-	create.pCode            = code;
-	vk::ShaderModule module = nullptr;
-	RequireVulkanSuccess(m_graphics.device.createShaderModule(&create, nullptr, &module),
-	                     "create BlitHelper shader module");
-	return module;
-}
-
 vk::Pipeline BlitHelper::GetPipeline(PipelineKey key) {
 	const auto cached = std::ranges::find(m_pipelines, key, &Pipeline::key);
 	if (cached != m_pipelines.end()) {

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

@@ -37,8 +37,7 @@ private:
 		vk::Pipeline handle = nullptr;
 	};
 
-	[[nodiscard]] vk::ShaderModule CreateShader(const uint32_t* code, size_t words) const;
-	[[nodiscard]] vk::Pipeline     GetPipeline(PipelineKey key);
+	[[nodiscard]] vk::Pipeline GetPipeline(PipelineKey key);
 
 	GraphicContext&         m_graphics;
 	CommandScheduler&       m_scheduler;

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

@@ -94,9 +94,24 @@ struct ImageInfo {
 		       bytes_per_block == other.bytes_per_block;
 	}
 	[[nodiscard]] int32_t MipOf(const ImageInfo& container) const noexcept {
-		if (!IsCompatible(container) || tile_mode != container.tile_mode || resources.levels != 1 ||
-		    container.resources.layers == 0 ||
-		    container.resources.levels > container.mip_layout.size()) {
+		if (container.resources.levels > container.mip_layout.size()) {
+			return -1;
+		}
+		for (uint32_t level = 0; level < container.resources.levels; level++) {
+			if (SliceOf(container, static_cast<int32_t>(level)) >= 0) {
+				return static_cast<int32_t>(level);
+			}
+		}
+		return -1;
+	}
+	[[nodiscard]] int32_t SliceOf(const ImageInfo& container, int32_t mip) const noexcept {
+		if (!IsCompatible(container) || tile_mode != container.tile_mode || type != container.type ||
+		    IsVolume() || resources.levels != 1 || resources.layers == 0 ||
+		    container.resources.layers == 0 || mip < 0 ||
+		    static_cast<uint32_t>(mip) >= container.resources.levels ||
+		    container.resources.levels > container.mip_layout.size() ||
+		    !data.Valid() || !container.data.Valid() || data.address < container.data.address ||
+		    data.End() > container.data.End()) {
 			return -1;
 		}
 		if (HasStencil() != container.HasStencil() ||
@@ -105,75 +120,36 @@ struct ImageInfo {
 			return -1;
 		}
 
-		int32_t mip = -1;
-		for (uint32_t level = 0; level < container.resources.levels; level++) {
-			const auto& layout = container.mip_layout[level];
-			if (layout.size == 0 || layout.size % container.resources.layers != 0 ||
-			    container.data.address > UINT64_MAX - layout.offset) {
-				continue;
-			}
-			const auto mip_base   = container.data.address + layout.offset;
-			const auto slice_size = layout.size / container.resources.layers;
-			if (slice_size == 0 || mip_base > UINT64_MAX - layout.size) {
-				continue;
-			}
-			const auto mip_end = mip_base + layout.size;
-			if (data.address >= mip_base && data.address < mip_end &&
-			    (data.address - mip_base) % slice_size == 0) {
-				mip = static_cast<int32_t>(level);
-				break;
-			}
-		}
-		if (mip < 0) {
+		const auto  level  = static_cast<uint32_t>(mip);
+		const auto& layout = container.mip_layout[level];
+		if (extent.width != std::max(container.extent.width >> level, 1u) ||
+		    extent.height != std::max(container.extent.height >> level, 1u) ||
+		    extent.depth != container.extent.depth || mip_layout[0].pitch != layout.pitch ||
+		    mip_layout[0].height != layout.height || layout.size == 0 ||
+		    layout.size % container.resources.layers != 0 ||
+		    container.data.size % container.resources.layers != 0 ||
+		    data.size % resources.layers != 0) {
 			return -1;
 		}
 
-		const auto level = static_cast<uint32_t>(mip);
-		if (extent.width != std::max(container.extent.width >> level, 1u) ||
-		    extent.height != std::max(container.extent.height >> level, 1u)) {
+		// PS5 block slices contain the entire reversed mip chain.
+		const auto slice_stride = container.data.size / container.resources.layers;
+		const auto child_stride = data.size / resources.layers;
+		if (child_stride != layout.size / container.resources.layers ||
+		    layout.offset >= slice_stride || child_stride > slice_stride - layout.offset ||
+		    (resources.layers > 1 && child_stride != slice_stride)) {
 			return -1;
 		}
-		const auto mip_depth = std::max(container.extent.depth >> level, 1u);
-		if (container.type == Prospero::ImageType::kColor3D &&
-		    type == Prospero::ImageType::kColor2D) {
-			if (resources.layers != mip_depth) {
-				return -1;
-			}
-		} else if (type != container.type) {
+		const auto address_delta = data.address - container.data.address;
+		if (address_delta < layout.offset || (address_delta - layout.offset) % slice_stride != 0) {
 			return -1;
 		}
-		return mip;
-	}
-	[[nodiscard]] int32_t SliceOf(const ImageInfo& container, int32_t mip) const noexcept {
-		if (!IsCompatible(container) || type != container.type || mip < 0 ||
-		    static_cast<uint32_t>(mip) >= container.resources.levels ||
-		    container.resources.levels > container.mip_layout.size() ||
-		    container.resources.layers == 0 || data.size == 0) {
+		const auto layer = (address_delta - layout.offset) / slice_stride;
+		if (layer > INT32_MAX || layer >= container.resources.layers ||
+		    resources.layers > container.resources.layers - layer) {
 			return -1;
 		}
-		const auto level = static_cast<uint32_t>(mip);
-		if (extent.width != std::max(container.extent.width >> level, 1u) ||
-		    extent.height != std::max(container.extent.height >> level, 1u)) {
-			return -1;
-		}
-		const auto& layout = container.mip_layout[level];
-		if (layout.size == 0 || layout.size % container.resources.layers != 0 ||
-		    container.data.address > UINT64_MAX - layout.offset) {
-			return -1;
-		}
-		const auto slice_size = layout.size / container.resources.layers;
-		if (slice_size == 0 || data.size % slice_size != 0) {
-			return -1;
-		}
-		const auto mip_base = container.data.address + layout.offset;
-		if (data.address < mip_base) {
-			return -1;
-		}
-		const auto address_delta = data.address - mip_base;
-		if (address_delta % data.size != 0 || address_delta / data.size > INT32_MAX) {
-			return -1;
-		}
-		return static_cast<int32_t>(address_delta / data.size);
+		return static_cast<int32_t>(layer);
 	}
 };
 
@@ -183,6 +159,7 @@ struct ImageViewInfo {
 	vk::ImageAspectFlags aspect      = vk::ImageAspectFlagBits::eColor;
 	uint32_t             base_level  = 0;
 	uint32_t             level_count = 1;
+	uint32_t             min_lod     = 0; // U4.8 clamp relative to base_level.
 	uint32_t             base_layer  = 0;
 	uint32_t             layer_count = 1;
 	vk::ComponentMapping mapping     = {};
@@ -191,9 +168,10 @@ struct ImageViewInfo {
 	[[nodiscard]] bool operator==(const ImageViewInfo& rhs) const noexcept {
 		return format == rhs.format && type == rhs.type && aspect == rhs.aspect &&
 		       base_level == rhs.base_level && level_count == rhs.level_count &&
-		       base_layer == rhs.base_layer && layer_count == rhs.layer_count &&
-		       mapping.r == rhs.mapping.r && mapping.g == rhs.mapping.g &&
-		       mapping.b == rhs.mapping.b && mapping.a == rhs.mapping.a && usage == rhs.usage;
+		       min_lod == rhs.min_lod && base_layer == rhs.base_layer &&
+		       layer_count == rhs.layer_count && mapping.r == rhs.mapping.r &&
+		       mapping.g == rhs.mapping.g && mapping.b == rhs.mapping.b &&
+		       mapping.a == rhs.mapping.a && usage == rhs.usage;
 	}
 };
 

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

@@ -362,6 +362,12 @@ vk::ImageView Image::FindView(const ImageViewInfo& view_info) {
 	if (!is_storage) {
 		usage.usage &= ~vk::ImageUsageFlagBits::eStorage;
 	}
+	vk::ImageViewMinLodCreateInfoEXT min_lod {};
+	if (normalized.min_lod != 0) {
+		min_lod.minLod = static_cast<float>(normalized.base_level) +
+		                 static_cast<float>(normalized.min_lod) / 256.0f;
+		usage.pNext    = &min_lod;
+	}
 	vk::ImageViewCreateInfo create {};
 	create.pNext                           = &usage;
 	create.image                           = image.image;

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

@@ -260,12 +260,8 @@ vk::Pipeline TileManager::GetPipeline(uint32_t slot) {
 	const uint32_t                   values[] {1u << element_index, direction_index};
 	const vk::SpecializationMapEntry entries[] {{0, 0, 4}, {1, 4, 4}};
 	const vk::SpecializationInfo     specialization {2, entries, sizeof(values), values};
-	vk::ShaderModuleCreateInfo       module_info {};
-	module_info.codeSize    = shaders[family_index].words * sizeof(uint32_t);
-	module_info.pCode       = shaders[family_index].code;
-	vk::ShaderModule module = nullptr;
-	RequireVulkanSuccess(m_graphics.device.createShaderModule(&module_info, nullptr, &module),
-	                     "create TileManager shader module");
+	const auto module =
+	    CompileSPV({shaders[family_index].code, shaders[family_index].words}, m_graphics.device);
 	vk::PipelineShaderStageCreateInfo stage {};
 	stage.stage               = vk::ShaderStageFlagBits::eCompute;
 	stage.module              = module;
@@ -472,12 +468,7 @@ void TileManager::ConvertD16(Result source, Result target, D16Direction directio
 			code  = GPU_TILER_DEMOTE_D16_SPV;
 			words = std::size(GPU_TILER_DEMOTE_D16_SPV);
 		}
-		vk::ShaderModuleCreateInfo module_info {};
-		module_info.codeSize    = words * sizeof(uint32_t);
-		module_info.pCode       = code;
-		vk::ShaderModule module = nullptr;
-		RequireVulkanSuccess(m_graphics.device.createShaderModule(&module_info, nullptr, &module),
-		                     "create D16 conversion shader module");
+		const auto module = CompileSPV({code, words}, m_graphics.device);
 		vk::PipelineShaderStageCreateInfo stage {};
 		stage.stage               = vk::ShaderStageFlagBits::eCompute;
 		stage.module              = module;
@@ -611,12 +602,7 @@ void TileManager::ConvertD16(Result source, Result target, D16Direction directio
 
 void TileManager::SwapBgra16(Result input, Result output, uint32_t pixels) {
 	if (m_swap_bgra16 == nullptr) {
-		vk::ShaderModuleCreateInfo module_info {};
-		module_info.codeSize    = std::size(GPU_TILER_SWAP_BGRA16_SPV) * sizeof(uint32_t);
-		module_info.pCode       = GPU_TILER_SWAP_BGRA16_SPV;
-		vk::ShaderModule module = nullptr;
-		RequireVulkanSuccess(m_graphics.device.createShaderModule(&module_info, nullptr, &module),
-		                     "create BGRA16 swap shader module");
+		const auto module = CompileSPV(GPU_TILER_SWAP_BGRA16_SPV, m_graphics.device);
 		vk::PipelineShaderStageCreateInfo stage {};
 		stage.stage  = vk::ShaderStageFlagBits::eCompute;
 		stage.module = module;

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

@@ -15,6 +15,7 @@
 #include "graphics/guest_gpu/tile.h"
 #include "graphics/host_gpu/graphicContext.h"
 #include "graphics/host_gpu/hostMemory.h"
+#include "graphics/host_gpu/renderer/colorRenderTarget.h"
 #include "graphics/host_gpu/renderer/debug.h"
 #include "graphics/host_gpu/renderer/image/imageView.h"
 #include "graphics/host_gpu/renderer/image/textureCommon.h"
@@ -93,22 +94,15 @@ static const char* ShaderStageResourceName(ShaderType stage) {
 	switch (stage) {
 		case ShaderType::Vertex: return "Vertex";
 		case ShaderType::Mesh: return "Mesh";
+		case ShaderType::Local: return "Local";
+		case ShaderType::TessellationControl: return "Hull";
+		case ShaderType::TessellationEvaluation: return "Domain";
 		case ShaderType::Pixel: return "Pixel";
 		case ShaderType::Compute: return "Compute";
 		default: return "Unknown";
 	}
 }
 
-static vk::ShaderStageFlags NativeShaderStage(ShaderType stage) {
-	switch (stage) {
-		case ShaderType::Vertex: return vk::ShaderStageFlagBits::eVertex;
-		case ShaderType::Mesh: return vk::ShaderStageFlagBits::eMeshEXT;
-		case ShaderType::Pixel: return vk::ShaderStageFlagBits::eFragment;
-		case ShaderType::Compute: return vk::ShaderStageFlagBits::eCompute;
-		default: EXIT("unknown native shader stage\n");
-	}
-}
-
 static Prospero::ImageType TextureType(const ShaderTextureResource& descriptor) {
 	const auto type = descriptor.Type();
 	return type == Prospero::ImageType::kCube ? Prospero::ImageType::kColor2DArray : type;
@@ -154,7 +148,7 @@ NativeStorageBuffer(RenderContext& context, const PreparedBindings::BufferSource
 	}
 	buffer_offset = static_cast<uint32_t>(adjustment);
 	const vk::DescriptorBufferInfo result {buffer->Handle(), aligned_offset, size + adjustment};
-	if (resource.formatted && resource.written) {
+	if (resource.written) {
 		context.GetTextureCache().InvalidateMemoryFromGPU(address, size);
 	}
 	const char* access = "Read";
@@ -299,13 +293,8 @@ static bool IsSupportedStorageTextureDescriptor(const ShaderRecompiler::IR::Imag
 	const bool supported_swizzle =
 	    IsValidImageSwizzle(swizzle) &&
 	    (swizzle == DstSel(4, 5, 6, 7) || !resource.read || resource.atomic);
-	const auto max_mip = resource.r128 ? descriptor.LastLevel() : descriptor.MaxMip();
-	const auto view_last_level =
-	    resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::DynamicStorage
-	        ? descriptor.LastLevel()
-	        : std::min(descriptor.LastLevel(), max_mip);
 	return (is_1d || is_1d_array || is_2d || is_2d_array || is_3d) && supported_tile &&
-	       descriptor.BaseLevel() <= view_last_level && view_last_level <= max_mip &&
+	       descriptor.BaseLevel() <= descriptor.LastLevel() &&
 	       descriptor.MinLod() == 0 && supported_swizzle && descriptor.BCSwizzle() == 0 &&
 	       !descriptor.MsaaDepth();
 }
@@ -463,6 +452,14 @@ static ImageViewInfo TextureViewInfo(const ShaderRecompiler::IR::ImageResource&
 	view.aspect      = vk::ImageAspectFlagBits::eColor;
 	view.base_level  = descriptor.BaseLevel();
 	view.level_count = view_levels;
+	if (descriptor.MinLod() > descriptor.LastLevel() * 256u) {
+		EXIT("texture minimum LOD exceeds last mip level: min_lod=%u last_level=%u\n",
+		     descriptor.MinLod(), descriptor.LastLevel());
+	}
+	const auto base_lod = view.base_level * 256u;
+	if (descriptor.MinLod() > base_lod) {
+		view.min_lod = descriptor.MinLod() - base_lod;
+	}
 	view.usage = storage ? vk::ImageUsageFlagBits::eStorage : vk::ImageUsageFlagBits::eSampled;
 	view.mapping =
 	    storage || surface_format.conversion_format != Prospero::BufferFormat::kInvalid
@@ -513,6 +510,21 @@ static ImageViewInfo TextureViewInfo(const ShaderRecompiler::IR::ImageResource&
 	return view;
 }
 
+static bool TextureViewPreservesMipLayout(const TileSurfaceDescription& description,
+                                          uint32_t                      view_levels) {
+	TileSurfaceLayout physical {};
+	TileSurfaceLayout view {};
+	auto              view_description = description;
+	view_description.levels            = view_levels;
+	return TileGetTiledTextureLayout(description, physical) &&
+	       TileGetTiledTextureLayout(view_description, view) &&
+	       physical.first_tail_level == view.first_tail_level &&
+	       physical.block_slice_size == view.block_slice_size &&
+	       physical.total_size == view.total_size &&
+	       std::equal(std::begin(physical.mips), std::begin(physical.mips) + description.levels,
+	                  std::begin(view.mips));
+}
+
 TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageResource&   resource,
                                               const ShaderRecompiler::IR::DescriptorValue& value) {
 	auto descriptor = DecodeNativeDescriptor<ShaderTextureResource>(value);
@@ -529,24 +541,22 @@ TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageR
 		return {id, nullptr, std::move(desc)};
 	}
 
-	const auto address      = descriptor.Base40();
-	const auto width        = static_cast<uint32_t>(descriptor.Width5()) + 1u;
-	const auto height       = static_cast<uint32_t>(descriptor.Height5()) + 1u;
-	const auto base_level   = descriptor.BaseLevel();
-	const auto last_level   = descriptor.LastLevel();
-	const auto type         = TextureType(descriptor);
-	const bool multisampled = IsMultisampledTexture(type);
-	const auto max_mip      = resource.r128 ? last_level : descriptor.MaxMip();
-	const auto levels       = multisampled ? 1u : static_cast<uint32_t>(max_mip) + 1u;
-	const bool dynamic_storage =
-	    storage && resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::DynamicStorage;
-	const auto view_last_level =
-	    !multisampled && !dynamic_storage ? std::min(last_level, max_mip) : last_level;
+	const auto address         = descriptor.Base40();
+	const auto width           = static_cast<uint32_t>(descriptor.Width5()) + 1u;
+	const auto height          = static_cast<uint32_t>(descriptor.Height5()) + 1u;
+	const auto base_level      = descriptor.BaseLevel();
+	const auto last_level      = descriptor.LastLevel();
+	const auto type            = TextureType(descriptor);
+	const bool multisampled    = IsMultisampledTexture(type);
+	const auto max_mip         = resource.r128 ? last_level : descriptor.MaxMip();
+	const auto physical_levels = multisampled ? 1u : static_cast<uint32_t>(max_mip) + 1u;
+	const auto levels =
+	    multisampled ? 1u : std::max(physical_levels, static_cast<uint32_t>(last_level) + 1u);
 	const auto tile       = descriptor.TileMode();
 	const bool depth_tile = tile == Prospero::TileMode::kDepth;
 	const bool msaa_tile  = depth_tile || tile == Prospero::TileMode::kRenderTarget;
 	const bool msaa_array = type == Prospero::ImageType::kColor2DMsaaArray;
-	if ((!multisampled && (base_level > view_last_level || view_last_level >= levels)) ||
+	if ((!multisampled && base_level > last_level) ||
 	    (multisampled &&
 	     (base_level != 0 || last_level == 0 || last_level > 3 || max_mip != last_level ||
 	      !msaa_tile || (descriptor.MsaaDepth() && !depth_tile) ||
@@ -565,7 +575,7 @@ TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageR
 	}
 	const auto samples = multisampled ? 1u << last_level : 1u;
 	const auto view_levels =
-	    multisampled ? 1u : static_cast<uint32_t>(view_last_level - base_level) + 1u;
+	    multisampled ? 1u : static_cast<uint32_t>(last_level - base_level) + 1u;
 	const auto depth          = static_cast<uint32_t>(descriptor.Depth()) + 1u;
 	const auto format         = descriptor.Format();
 	const auto surface_format = TextureGetSurfaceFormatInfo(format);
@@ -584,7 +594,19 @@ TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageR
 	                             type == Prospero::ImageType::kColor2DArray ||
 	                             type == Prospero::ImageType::kColor2DMsaaArray;
 	const auto    image_layers = layered ? depth : 1u;
-	uint32_t      pitch        = 0;
+	if (levels > physical_levels) {
+		const TileSurfaceDescription physical {
+		    format, tile, volume ? TileSurfaceDimension::Dim3D : TileSurfaceDimension::Dim2D,
+		    width, height, volume ? depth : 1u, physical_levels, image_layers};
+		// Texture mip views take precedence over the resource count, but must keep its storage layout.
+		if (!TextureViewPreservesMipLayout(physical, levels)) {
+			EXIT("unsupported texture mip view changes physical layout: base=%u last=%u max=%u "
+			     "extent=%ux%ux%u tile=%u\n",
+			     base_level, last_level, max_mip, width, height, depth,
+			     static_cast<uint32_t>(tile));
+		}
+	}
+	uint32_t      pitch = 0;
 	TileSizeAlign size {};
 	if (multisampled) {
 		const auto bytes = Prospero::NumBytesPerElement(format);
@@ -597,8 +619,8 @@ TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageR
 		size.size *= image_layers;
 	} else {
 		pitch = TileGetTexturePitch(format, width, tile);
-		TileGetTextureTotalSize(format, width, height, volume ? depth : image_layers, levels, tile,
-		                        volume, size);
+		TileGetTextureTotalSize(format, width, height, volume ? depth : image_layers,
+		                        physical_levels, tile, volume, size);
 	}
 	EXIT_NOT_IMPLEMENTED(size.size == 0 || size.align == 0 ||
 	                     (address & (static_cast<uint64_t>(size.align) - 1u)) != 0);
@@ -635,7 +657,7 @@ TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageR
 	    !desc.info.IsDepth()) {
 		TileSizeAlign metadata_size {};
 		(void)TileGetDccSize(width, height, volume ? depth : image_layers,
-		                     desc.info.bytes_per_block, levels, tile, metadata_size,
+		                     desc.info.bytes_per_block, physical_levels, tile, metadata_size,
 		                     std::countr_zero(samples));
 		desc.info.metadata.kind          = ImageMetadataKind::Dcc;
 		desc.info.metadata.range         = {descriptor.MetaAddr() << 8u, metadata_size.size};
@@ -857,32 +879,39 @@ void RenderExecutor::RebindImages(PreparedBindings& prepared) {
 	}
 }
 
-RenderExecutor::GraphicsBindings
-RenderExecutor::PrepareGraphicsBindings(const ShaderStageRuntime& vertex,
-                                        const ShaderStageRuntime& pixel, bool pixel_active) {
-	GraphicsBindings bindings {
-	    .vertex = PrepareBindings(vertex),
-	};
-	if (pixel_active) {
-		bindings.pixel.emplace(PrepareBindings(pixel));
+void RenderExecutor::PrepareGraphicsBindings(std::span<PreparedBindings* const> stages,
+                                             std::span<RenderColorInfo> colors) {
+	bool uses_dma = false;
+	for (auto* stage: stages) {
+		FindBuffers(*stage);
+		uses_dma |= stage->runtime->program->info.uses_dma;
 	}
-	FindBuffers(bindings.vertex);
-	if (bindings.pixel) {
-		FindBuffers(*bindings.pixel);
-	}
-	if (bindings.vertex.runtime->program->info.uses_dma ||
-	    (bindings.pixel && bindings.pixel->runtime->program->info.uses_dma)) {
+	if (uses_dma) {
 		m_context.PrepareBda();
 	}
-	RebindBuffers(bindings.vertex);
-	if (bindings.pixel) {
-		RebindBuffers(*bindings.pixel);
+	for (auto* stage: stages) {
+		RebindImages(*stage);
 	}
-	RebindImages(bindings.vertex);
-	if (bindings.pixel) {
-		RebindImages(*bindings.pixel);
+	auto& cache = m_context.GetTextureCache();
+	for (auto& target: colors) {
+		EXIT_IF(!target.image_id);
+		const auto old_image = cache.m_slot_images.try_get(target.image_id);
+		if (old_image == nullptr || (!old_image->registered && !old_image->info.data.Empty()) ||
+		    old_image->binding.needs_rebind) {
+			if (old_image != nullptr) {
+				old_image->binding = {};
+			}
+			target.desc.view_info.base_level = target.guest_mip_level;
+			target.desc.view_info.base_layer = target.guest_array_layer;
+			target.image_id = cache.FindImage(target.desc);
+			BindRenderTarget(target.image_id);
+		}
+	}
+	// Discovery can read back PS5 metadata and submit the scheduler. Reserve draw buffers only
+	// after image identities are final; attachment layout transitions follow buffer alias copies.
+	for (auto* stage: stages) {
+		RebindBuffers(*stage);
 	}
-	return bindings;
 }
 
 void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
@@ -895,9 +924,10 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 	size_t write_count      = 0;
 	ShaderRecompiler::IR::PushData push_data;
 	bool                           has_push_data = false;
-	constexpr auto                 GraphicsStages = vk::ShaderStageFlagBits::eVertex |
-	                                                vk::ShaderStageFlagBits::eMeshEXT |
-	                                                vk::ShaderStageFlagBits::eFragment;
+	constexpr auto                 GraphicsStages =
+	    vk::ShaderStageFlagBits::eVertex | vk::ShaderStageFlagBits::eMeshEXT |
+	    vk::ShaderStageFlagBits::eTessellationControl |
+	    vk::ShaderStageFlagBits::eTessellationEvaluation | vk::ShaderStageFlagBits::eFragment;
 	vk::ShaderStageFlags push_stages = pipeline_bind_point == vk::PipelineBindPoint::eGraphics
 	                                       ? vk::ShaderStageFlagBits::eFragment
 	                                       : vk::ShaderStageFlags {};

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

@@ -89,12 +89,7 @@ template <typename... Args>
 void PipelineCacheLog(fmt::format_string<Args...> format, Args&&... args) {
 	auto message = fmt::format(format, std::forward<Args>(args)...);
 	message += '\n';
-	if (Log::GetDirection() != Log::Direction::Console) {
-		std::fwrite(message.data(), 1, message.size(), stdout);
-		std::fflush(stdout);
-	}
-	Log::Write(message);
-	Log::Flush();
+	Log::WriteToConsoleAndLog(message);
 }
 
 bool ReadShaderGuestMemory(void*, uint64_t address, uint32_t* value) {
@@ -236,6 +231,9 @@ struct PipelineCache::ProgramCache {
 		switch (options.stage) {
 			case ShaderType::Vertex: stage_name = "vs"; break;
 			case ShaderType::Mesh: stage_name = "ms"; break;
+			case ShaderType::Local: stage_name = "ls"; break;
+			case ShaderType::TessellationControl: stage_name = "hs"; break;
+			case ShaderType::TessellationEvaluation: stage_name = "ds"; break;
 			case ShaderType::Pixel: stage_name = "ps"; break;
 			case ShaderType::Compute: stage_name = "cs"; break;
 			default: EXIT("invalid pipeline shader stage\n");
@@ -250,12 +248,7 @@ struct PipelineCache::ProgramCache {
 		}
 		DumpShaderSpirv(stage_name, options.shader_hash, result.spirv);
 
-		vk::ShaderModuleCreateInfo create_info {};
-		create_info.codeSize    = result.spirv.size() * sizeof(uint32_t);
-		create_info.pCode       = result.spirv.data();
-		vk::ShaderModule module = nullptr;
-		RequireVulkanSuccess(device.createShaderModule(&create_info, nullptr, &module),
-		                     "create recompiled shader module");
+		const auto module = CompileSPV(result.spirv, device);
 		EXIT_IF(module == nullptr);
 		if (options.dump_ir) {
 			LOGF("%s SPIR-V words=%" PRIu64 " wave_size=%u\n", options.dump_label,
@@ -273,7 +266,7 @@ struct PipelineCache::ProgramCache {
 	                  uint32_t& push_data_cursor) {
 		ShaderType stage;
 		if constexpr (std::is_same_v<InputInfo, ShaderVertexInputInfo>) {
-			stage = input_info.mesh.threads_num[0] != 0 ? ShaderType::Mesh : ShaderType::Vertex;
+			stage = input_info.logical_stage;
 		} else if constexpr (std::is_same_v<InputInfo, ShaderPixelInputInfo>) {
 			stage = ShaderType::Pixel;
 		} else {
@@ -325,6 +318,9 @@ struct PipelineCache::ProgramCache {
 		switch (stage) {
 			case ShaderType::Vertex: label = "ShaderRecompiler VS"; break;
 			case ShaderType::Mesh: label = "ShaderRecompiler MS"; break;
+			case ShaderType::Local: label = "ShaderRecompiler LS"; break;
+			case ShaderType::TessellationControl: label = "ShaderRecompiler HS"; break;
+			case ShaderType::TessellationEvaluation: label = "ShaderRecompiler DS"; break;
 			case ShaderType::Pixel: label = "ShaderRecompiler PS"; break;
 			case ShaderType::Compute: label = "ShaderRecompiler CS"; break;
 			default: EXIT("invalid pipeline shader stage\n");
@@ -341,11 +337,11 @@ struct PipelineCache::ProgramCache {
 
 		if constexpr (std::is_same_v<InputInfo, ShaderVertexInputInfo>) {
 			options.user_data_base = 8;
-			if (stage == ShaderType::Mesh) {
+			if (stage == ShaderType::Mesh || stage == ShaderType::TessellationControl) {
 				options.user_data_base = 0;
-				options.wave_size      = input_info.mesh.wave_size;
+				options.wave_size = stage == ShaderType::Mesh ? input_info.mesh.wave_size : 64u;
 			}
-		} else if constexpr (std::is_same_v<InputInfo, ShaderComputeInputInfo>) {
+		} else {
 			options.wave_size = input_info.wave_size;
 		}
 		auto translated = ShaderRecompiler::TranslateProgram(params.code, options);
@@ -361,16 +357,19 @@ struct PipelineCache::ProgramCache {
 		input_info.stage = {.program = &permutation.program, .resources = std::move(resources)};
 		permutation.program.bindings.AdvancePushData(push_data_cursor);
 
-		std::array<size_t, static_cast<size_t>(ShaderType::Mesh) + 1> counts {};
+		std::array<size_t, static_cast<size_t>(ShaderType::TessellationEvaluation) + 1> counts {};
 		for (const auto& [key, source]: programs) {
 			counts[static_cast<size_t>(key.stage)] += source.permutations.size();
 		}
 		// Guest geometry shaders are compiled through the host mesh stage.
-		std::printf("Shaders: VS %zu | PS %zu | CS %zu | GS %zu\n",
+		std::printf("Shaders: VS %zu | PS %zu | CS %zu | GS %zu | LS %zu | HS %zu | TES %zu\n",
 		            counts[static_cast<size_t>(ShaderType::Vertex)],
 		            counts[static_cast<size_t>(ShaderType::Pixel)],
 		            counts[static_cast<size_t>(ShaderType::Compute)],
-		            counts[static_cast<size_t>(ShaderType::Mesh)]);
+		            counts[static_cast<size_t>(ShaderType::Mesh)],
+		            counts[static_cast<size_t>(ShaderType::Local)],
+		            counts[static_cast<size_t>(ShaderType::TessellationControl)],
+		            counts[static_cast<size_t>(ShaderType::TessellationEvaluation)]);
 		return permutation.handle;
 	}
 
@@ -564,12 +563,18 @@ PipelineCache::GraphicsPrograms PipelineCache::GetGraphicsPrograms(
     const HW::VertexShaderInfo& vertex_regs, const HW::PixelShaderInfo& pixel_regs,
     const HW::ShaderRegisters& sh, const HW::Context& context, const HW::UserConfig& user_config,
     std::span<const Prospero::ColorComponentMapping, 8> target_export_mapping, bool pixel_active,
-    ShaderVertexInputInfo& vertex_info, ShaderPixelInputInfo& pixel_info) {
-	const auto vertex_params = PrepareProgram(vertex_regs, context, user_config, vertex_info);
-	const bool mesh_active   = vertex_info.mesh.threads_num[0] != 0;
+    std::array<ShaderVertexInputInfo, 3>& vertex_info, ShaderPixelInputInfo& pixel_info) {
+	const bool tess_active = user_config.GetPrimType() == Prospero::PrimitiveType::kPatch;
+	std::array<ShaderParams, 3> vertex_params;
+	if (tess_active) {
+		vertex_params = PrepareTessellationPrograms(vertex_regs, context, vertex_info);
+	} else {
+		vertex_params[0] = PrepareProgram(vertex_regs, context, user_config, vertex_info[0]);
+	}
+	const bool mesh_active = vertex_info[0].logical_stage == ShaderType::Mesh;
 	if (mesh_active) {
 		EXIT_NOT_IMPLEMENTED(!m_graphics.mesh_shader_enabled);
-		auto& mesh              = vertex_info.mesh;
+		auto& mesh              = vertex_info[0].mesh;
 		mesh.host_subgroup_size = m_graphics.subgroup_size;
 		const auto& limits      = m_graphics.mesh_shader_properties;
 		const auto  logical_threads =
@@ -592,7 +597,7 @@ PipelineCache::GraphicsPrograms PipelineCache::GetGraphicsPrograms(
 	if (context.GetClipControl().clip_disable) {
 		const auto& viewport = context.GetScreenViewport().viewports[0];
 		const auto& limits   = m_graphics.GetPhysicalDeviceProperties().limits;
-		auto&       clip     = vertex_info.clip_space;
+		auto&       clip     = vertex_info[tess_active ? 2u : 0u].clip_space;
 		clip.scale[0]        = viewport.xscale;
 		clip.scale[1]        = viewport.yscale;
 		clip.offset[0]       = viewport.xoffset;
@@ -610,7 +615,9 @@ PipelineCache::GraphicsPrograms PipelineCache::GetGraphicsPrograms(
 	if (pixel_active) {
 		result.pixel = m_program_cache->Get(pixel_params, pixel_info, push_data_cursor);
 	}
-	result.vertex = m_program_cache->Get(vertex_params, vertex_info, push_data_cursor);
+	for (uint32_t i = 0; i < (tess_active ? 3u : 1u); i++) {
+		result.vertex[i] = m_program_cache->Get(vertex_params[i], vertex_info[i], push_data_cursor);
+	}
 	return result;
 }
 
@@ -630,10 +637,12 @@ bool PipelineStaticParameters::operator==(const PipelineStaticParameters& other)
 
 PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(
     std::span<const RenderColorInfo> colors, const RenderDepthInfo& depth,
-    const ShaderVertexInputInfo& vs_input_info, CommandBuffer& command,
+    std::span<const ShaderVertexInputInfo> vertex_info, CommandBuffer& command,
     const ShaderPixelInputInfo* ps_input_info, vk::PrimitiveTopology topology,
-    bool primitive_restart_enable, const ShaderProgram& vertex_program,
-    const ShaderProgram& pixel_program) {
+    bool primitive_restart_enable, const GraphicsPrograms& programs) {
+	const auto& vs_input_info  = vertex_info.front();
+	const auto& vertex_program = programs.vertex[0];
+	const auto& pixel_program  = programs.pixel;
 	KYTY_PROFILER_BLOCK("PipelineCache::CreatePipeline(Gfx)", profiler::colors::DeepOrangeA200);
 
 	EXIT_IF(colors.size() > RENDER_COLOR_ATTACHMENTS_MAX);
@@ -651,17 +660,22 @@ PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(
 	const auto ps_id = ps_active ? pixel_program.id : 0;
 
 	GraphicsPipelineKey key {};
-	key.vs_shader_id            = vs_id;
+	for (uint32_t i = 0; i < programs.vertex.size(); i++) {
+		key.vertex_shader_ids[i] = programs.vertex[i].id;
+	}
 	key.ps_shader_id            = ps_id;
 	auto& static_params         = key.static_params;
 	auto& rendering             = key.rendering;
-	rendering.color_count       = color_count;
+	rendering.color_count       = 0;
 	uint32_t attachment_samples = 0;
 	for (uint32_t i = 0; i < color_count; i++) {
+		const auto slot = colors[i].target_slot;
+		EXIT_IF(slot >= RENDER_COLOR_ATTACHMENTS_MAX);
+		rendering.color_count = std::max(rendering.color_count, slot + 1);
 		EXIT_IF(!colors[i].image_id || colors[i].desc.view_info.format == vk::Format::eUndefined);
-		static_params.color_mask[i] = colors[i].export_mapping.ApplyMask(
+		static_params.color_mask[slot] = colors[i].export_mapping.ApplyMask(
 		    render_target_mask_slot(ctx.GetRenderTargetMask(), colors[i].target_slot));
-		rendering.color_formats[i] = colors[i].desc.view_info.format;
+		rendering.color_formats[slot] = colors[i].desc.view_info.format;
 		if (attachment_samples == 0) {
 			attachment_samples = colors[i].desc.info.samples;
 		} else if (attachment_samples != colors[i].desc.info.samples) {
@@ -670,14 +684,14 @@ PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(
 		}
 		const auto& rt                        = ctx.GetRenderTarget(colors[i].target_slot);
 		const auto& bc                        = ctx.GetBlendControl(colors[i].target_slot);
-		static_params.color_srcblend[i]       = bc.color_srcblend;
-		static_params.color_comb_fcn[i]       = bc.color_comb_fcn;
-		static_params.color_destblend[i]      = bc.color_destblend;
-		static_params.alpha_srcblend[i]       = bc.alpha_srcblend;
-		static_params.alpha_comb_fcn[i]       = bc.alpha_comb_fcn;
-		static_params.alpha_destblend[i]      = bc.alpha_destblend;
-		static_params.separate_alpha_blend[i] = bc.separate_alpha_blend;
-		static_params.blend_enable[i]         = bc.enable && !rt.info.blend_bypass;
+		static_params.color_srcblend[slot]       = bc.color_srcblend;
+		static_params.color_comb_fcn[slot]       = bc.color_comb_fcn;
+		static_params.color_destblend[slot]      = bc.color_destblend;
+		static_params.alpha_srcblend[slot]       = bc.alpha_srcblend;
+		static_params.alpha_comb_fcn[slot]       = bc.alpha_comb_fcn;
+		static_params.alpha_destblend[slot]      = bc.alpha_destblend;
+		static_params.separate_alpha_blend[slot] = bc.separate_alpha_blend;
+		static_params.blend_enable[slot]         = bc.enable && !rt.info.blend_bypass;
 	}
 	const bool with_depth =
 	    depth.desc.view_info.format != vk::Format::eUndefined && static_cast<bool>(depth.image_id);
@@ -726,10 +740,8 @@ PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(
 	static_params.depth_bounds_test_enable = depth.depth_bounds_test_enable;
 	static_params.depth_min_bounds         = depth.depth_min_bounds;
 	static_params.depth_max_bounds         = depth.depth_max_bounds;
-	static_params.stencil_test_enable      = depth.stencil_test_enable;
-	static_params.stencil_front            = depth.stencil_static_front;
-	static_params.stencil_back             = depth.stencil_static_back;
-	const bool rect_list     = topology == vk::PrimitiveTopology::ePatchList;
+	const bool rect_list =
+	    command.GetUserConfig().GetPrimType() == Prospero::PrimitiveType::kRectList;
 	static_params.cull_back  = !rect_list && mc.cull_back;
 	static_params.cull_front = !rect_list && mc.cull_front;
 	static_params.face       = mc.face;
@@ -780,9 +792,8 @@ PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(
 
 	auto cached = std::make_unique<Pipeline>();
 	LogPipelineTrace("CreatePipelineInternal begin", vs_id, ps_id);
-	CreatePipelineInternal(m_graphics, *cached, rendering, key.vertex_input, vs_input_info,
-	                       vertex_program, ps_input_info, pixel_program, static_params,
-	                       m_driver_cache);
+	CreatePipelineInternal(m_graphics, *cached, rendering, key.vertex_input, vertex_info,
+	                       ps_input_info, programs, static_params, m_driver_cache);
 	LogPipelineTrace("CreatePipelineInternal done", vs_id, ps_id);
 
 	EXIT_NOT_IMPLEMENTED(cached->pipeline == nullptr);

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

@@ -42,9 +42,6 @@ struct PipelineStaticParameters {
 	bool                       depth_bounds_test_enable = false;
 	float                      depth_min_bounds         = 0.0f;
 	float                      depth_max_bounds         = 0.0f;
-	bool                       stencil_test_enable      = false;
-	PipelineStencilStaticState stencil_front;
-	PipelineStencilStaticState stencil_back;
 	uint32_t                   color_mask[RENDER_COLOR_ATTACHMENTS_MAX]           = {};
 	bool                       cull_front                                         = false;
 	bool                       cull_back                                          = false;
@@ -68,7 +65,7 @@ struct PipelineStaticParameters {
 static_assert(std::is_trivially_copyable_v<PipelineStaticParameters>);
 static_assert(std::is_standard_layout_v<PipelineStaticParameters>);
 static_assert(alignof(PipelineStaticParameters) == 1);
-static_assert(sizeof(PipelineStaticParameters) == 158);
+static_assert(sizeof(PipelineStaticParameters) == 125);
 
 struct PipelineRenderingState {
 	std::array<vk::Format, RENDER_COLOR_ATTACHMENTS_MAX> color_formats {};
@@ -121,8 +118,10 @@ public:
 	};
 
 	struct GraphicsPrograms {
-		ShaderProgram vertex;
+		std::array<ShaderProgram, 3> vertex;
 		ShaderProgram pixel;
+
+		[[nodiscard]] uint32_t VertexStageCount() const { return vertex[1] ? 3u : 1u; }
 	};
 
 	GraphicsPrograms
@@ -130,18 +129,18 @@ public:
 	                    const HW::PixelShaderInfo& pixel_regs, const HW::ShaderRegisters& sh,
 	                    const HW::Context& context, const HW::UserConfig& user_config,
 	                    std::span<const Prospero::ColorComponentMapping, 8> target_export_mapping,
-	                    bool pixel_active, ShaderVertexInputInfo& vertex_info,
+	                    bool pixel_active, std::array<ShaderVertexInputInfo, 3>& vertex_info,
 	                    ShaderPixelInputInfo& pixel_info);
 	ShaderProgram GetComputeProgram(const HW::ComputeShaderInfo& regs,
 	                                const HW::ShaderRegisters&   sh,
 	                                ShaderComputeInputInfo&      input_info);
 
-	Pipeline&
-	GetGraphicsPipeline(std::span<const RenderColorInfo> colors, const RenderDepthInfo& depth,
-	                    const ShaderVertexInputInfo& vs_input_info, CommandBuffer& command,
-	                    const ShaderPixelInputInfo* ps_input_info,
-	                    vk::PrimitiveTopology topology, bool primitive_restart_enable,
-	                    const ShaderProgram& vertex_program, const ShaderProgram& pixel_program);
+	Pipeline& GetGraphicsPipeline(std::span<const RenderColorInfo>       colors,
+	                              const RenderDepthInfo&                 depth,
+	                              std::span<const ShaderVertexInputInfo> vertex_info,
+	                              CommandBuffer& command, const ShaderPixelInputInfo* ps_input_info,
+	                              vk::PrimitiveTopology topology, bool primitive_restart_enable,
+	                              const GraphicsPrograms& programs);
 	Pipeline& GetComputePipeline(const ShaderComputeInputInfo& input_info,
 	                             const ShaderProgram&          compute_program);
 
@@ -150,13 +149,13 @@ private:
 
 	struct GraphicsPipelineKey {
 		PipelineRenderingState   rendering;
-		uint64_t                 vs_shader_id = 0;
+		std::array<uint64_t, 3>  vertex_shader_ids {};
 		uint64_t                 ps_shader_id = 0;
 		PipelineVertexInputState vertex_input;
 		PipelineStaticParameters static_params;
 
 		bool operator==(const GraphicsPipelineKey& other) const {
-			return rendering == other.rendering && vs_shader_id == other.vs_shader_id &&
+			return rendering == other.rendering && vertex_shader_ids == other.vertex_shader_ids &&
 			       ps_shader_id == other.ps_shader_id && vertex_input == other.vertex_input &&
 			       static_params == other.static_params;
 		}
@@ -189,7 +188,9 @@ private:
 		std::size_t operator()(const GraphicsPipelineKey& key) const {
 			std::size_t hash = 0;
 			PipelineKeyHash::MixRendering(hash, key.rendering);
-			PipelineKeyHash::Mix(hash, key.vs_shader_id);
+			for (const auto id: key.vertex_shader_ids) {
+				PipelineKeyHash::Mix(hash, id);
+			}
 			PipelineKeyHash::Mix(hash, key.ps_shader_id);
 			PipelineKeyHash::Mix(hash, key.vertex_input.binding_count);
 			for (uint32_t i = 0; i < key.vertex_input.binding_count; i++) {
@@ -219,12 +220,14 @@ private:
 };
 
 void LogPipelineTrace(const char* phase, uint64_t vertex_program_id, uint64_t pixel_program_id);
-void CreatePipelineInternal(
-    GraphicContext& graphics, PipelineCache::Pipeline& pipeline,
-    const PipelineRenderingState& rendering, const PipelineVertexInputState& vertex_input,
-    const ShaderVertexInputInfo& vs_input_info, const ShaderProgram& vertex_program,
-    const ShaderPixelInputInfo* ps_input_info, const ShaderProgram& pixel_program,
-    const PipelineStaticParameters& static_params, vk::PipelineCache driver_cache);
+void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& pipeline,
+                            const PipelineRenderingState&          rendering,
+                            const PipelineVertexInputState&        vertex_input,
+                            std::span<const ShaderVertexInputInfo> vertex_info,
+                            const ShaderPixelInputInfo*            ps_input_info,
+                            const PipelineCache::GraphicsPrograms& programs,
+                            const PipelineStaticParameters&        static_params,
+                            vk::PipelineCache                      driver_cache);
 void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& pipeline,
                             const ShaderComputeInputInfo& input_info,
                             vk::ShaderModule compute_module, vk::PipelineCache driver_cache);

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

@@ -8,6 +8,20 @@
 
 namespace Libs::Graphics {
 
+vk::ShaderStageFlagBits NativeShaderStage(ShaderType stage) {
+	switch (stage) {
+		case ShaderType::Local:
+		case ShaderType::Vertex: return vk::ShaderStageFlagBits::eVertex;
+		case ShaderType::Mesh: return vk::ShaderStageFlagBits::eMeshEXT;
+		case ShaderType::TessellationControl: return vk::ShaderStageFlagBits::eTessellationControl;
+		case ShaderType::TessellationEvaluation:
+			return vk::ShaderStageFlagBits::eTessellationEvaluation;
+		case ShaderType::Pixel: return vk::ShaderStageFlagBits::eFragment;
+		case ShaderType::Compute: return vk::ShaderStageFlagBits::eCompute;
+		default: EXIT("unknown native shader stage\n");
+	}
+}
+
 vk::PipelineStageFlags ShaderPipelineStages(vk::ShaderStageFlags stages) {
 	vk::PipelineStageFlags result = {};
 	if (stages & vk::ShaderStageFlagBits::eVertex) {
@@ -16,6 +30,12 @@ vk::PipelineStageFlags ShaderPipelineStages(vk::ShaderStageFlags stages) {
 	if (stages & vk::ShaderStageFlagBits::eMeshEXT) {
 		result |= vk::PipelineStageFlagBits::eMeshShaderEXT;
 	}
+	if (stages & vk::ShaderStageFlagBits::eTessellationControl) {
+		result |= vk::PipelineStageFlagBits::eTessellationControlShader;
+	}
+	if (stages & vk::ShaderStageFlagBits::eTessellationEvaluation) {
+		result |= vk::PipelineStageFlagBits::eTessellationEvaluationShader;
+	}
 	if (stages & vk::ShaderStageFlagBits::eFragment) {
 		result |= vk::PipelineStageFlagBits::eFragmentShader;
 	}

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

@@ -8,6 +8,7 @@ namespace Libs::Graphics {
 
 struct ShaderStageRuntime;
 
+vk::ShaderStageFlagBits NativeShaderStage(ShaderType stage);
 vk::PipelineStageFlags  ShaderPipelineStages(vk::ShaderStageFlags stages);
 VulkanMemoryBarrier     MakeShaderAccessDependency();
 VulkanMemoryBarrier     MakeShaderWriteHazardDependency();

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

@@ -8,6 +8,7 @@
 #include "graphics/host_gpu/renderer/debug.h"
 #include "graphics/host_gpu/renderer/pipeline/descriptors.h"
 #include "graphics/host_gpu/renderer/pipeline/pipelineCache.h"
+#include "graphics/host_gpu/renderer/pipeline/shaderResourceBarrier.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
 #include "graphics/host_gpu/renderer/renderTarget.h"
@@ -205,12 +206,18 @@ static void CreateDescriptorLayout(GraphicContext& graphics, PipelineCache::Pipe
 }
 
 // NOLINTNEXTLINE(readability-function-cognitive-complexity)
-void CreatePipelineInternal(
-    GraphicContext& graphics, PipelineCache::Pipeline& pipeline,
-    const PipelineRenderingState& rendering, const PipelineVertexInputState& vertex_input,
-    const ShaderVertexInputInfo& vs_input_info, const ShaderProgram& vertex_program,
-    const ShaderPixelInputInfo* ps_input_info, const ShaderProgram& pixel_program,
-    const PipelineStaticParameters& static_params, vk::PipelineCache driver_cache) {
+void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& pipeline,
+                            const PipelineRenderingState&          rendering,
+                            const PipelineVertexInputState&        vertex_input,
+                            std::span<const ShaderVertexInputInfo> vertex_info,
+                            const ShaderPixelInputInfo*            ps_input_info,
+                            const PipelineCache::GraphicsPrograms& programs,
+                            const PipelineStaticParameters&        static_params,
+                            vk::PipelineCache                      driver_cache) {
+	const auto& vs_input_info  = vertex_info.front();
+	const auto& vertex_program = programs.vertex[0];
+	const auto& pixel_program  = programs.pixel;
+	const bool  tessellation   = vertex_info.size() == 3;
 	const bool ps_active = ps_input_info != nullptr;
 	EXIT_IF(!vertex_program || (ps_active && !pixel_program));
 	const bool with_depth = rendering.depth_format != vk::Format::eUndefined ||
@@ -218,46 +225,39 @@ void CreatePipelineInternal(
 	EXIT_IF(!vs_input_info.stage);
 	const bool mesh = vs_input_info.stage.program->stage == ShaderType::Mesh;
 	EXIT_NOT_IMPLEMENTED(mesh && !graphics.mesh_shader_enabled);
-	const auto vertex_stage =
-	    mesh ? vk::ShaderStageFlagBits::eMeshEXT : vk::ShaderStageFlagBits::eVertex;
-
-	const bool rect_list = !mesh && static_params.topology == vk::PrimitiveTopology::ePatchList;
+	const bool rect_list =
+	    !mesh && !tessellation && static_params.topology == vk::PrimitiveTopology::ePatchList;
 
 	vk::ShaderModule tess_control_shader_module = nullptr;
 	vk::ShaderModule tess_eval_shader_module    = nullptr;
 
-	vk::ShaderModuleCreateInfo create_info {};
-	vk::Result result {};
 	if (rect_list) {
 		const auto shaders =
 		    BuildRectListShaders(vs_input_info, ps_active ? ps_input_info : nullptr);
-		create_info.codeSize = shaders.control.size() * 4;
-		create_info.pCode    = shaders.control.data();
-		result =
-		    graphics.device.createShaderModule(&create_info, nullptr, &tess_control_shader_module);
+		tess_control_shader_module = CompileSPV(shaders.control, graphics.device);
 		if (graphics_debug_dump_enabled()) {
-			LOGF("PipelineTrace: vkCreateShaderModule RectList TCS done result=%s module=%p\n",
-			     vk::to_string(result).c_str(), static_cast<void*>(tess_control_shader_module));
+			LOGF("PipelineTrace: vkCreateShaderModule RectList TCS done module=%p\n",
+			     static_cast<void*>(tess_control_shader_module));
 		}
-		EXIT_NOT_IMPLEMENTED(result != vk::Result::eSuccess);
 
-		create_info.codeSize = shaders.evaluation.size() * 4;
-		create_info.pCode    = shaders.evaluation.data();
-		result =
-		    graphics.device.createShaderModule(&create_info, nullptr, &tess_eval_shader_module);
+		tess_eval_shader_module = CompileSPV(shaders.evaluation, graphics.device);
 		if (graphics_debug_dump_enabled()) {
-			LOGF("PipelineTrace: vkCreateShaderModule RectList TES done result=%s module=%p\n",
-			     vk::to_string(result).c_str(), static_cast<void*>(tess_eval_shader_module));
+			LOGF("PipelineTrace: vkCreateShaderModule RectList TES done module=%p\n",
+			     static_cast<void*>(tess_eval_shader_module));
 		}
-		EXIT_NOT_IMPLEMENTED(result != vk::Result::eSuccess);
 	}
 
 	EXIT_NOT_IMPLEMENTED(
 	    rect_list && (tess_control_shader_module == nullptr || tess_eval_shader_module == nullptr));
 
-	vk::PipelineShaderStageCreateInfo shader_stages[4] = {
-	    {.stage = vertex_stage, .module = vertex_program.module, .pName = "main"}};
-	uint32_t shader_stage_count = 1;
+	vk::PipelineShaderStageCreateInfo shader_stages[4] {};
+	uint32_t                          shader_stage_count = 0;
+	for (uint32_t i = 0; i < vertex_info.size(); i++) {
+		shader_stages[shader_stage_count++] = {.stage =
+		                                           NativeShaderStage(vertex_info[i].logical_stage),
+		                                       .module = programs.vertex[i].module,
+		                                       .pName  = "main"};
+	}
 	if (rect_list) {
 		shader_stages[shader_stage_count++] = {.stage =
 		                                           vk::ShaderStageFlagBits::eTessellationControl,
@@ -432,15 +432,19 @@ void CreatePipelineInternal(
 	color_blending.pAttachments    = color_blend_attachment;
 
 	std::vector<vk::DescriptorSetLayoutBinding> descriptor_bindings;
-	AddLayoutBindings(descriptor_bindings, *vs_input_info.stage.program, vertex_stage);
+	vk::ShaderStageFlags graphics_stages = vk::ShaderStageFlagBits::eFragment;
+	for (const auto& stage: vertex_info) {
+		const auto native_stage = NativeShaderStage(stage.logical_stage);
+		AddLayoutBindings(descriptor_bindings, *stage.stage.program, native_stage);
+		graphics_stages |= native_stage;
+	}
 	if (ps_active) {
 		EXIT_IF(!ps_input_info->stage);
 		AddLayoutBindings(descriptor_bindings, *ps_input_info->stage.program,
 		                  vk::ShaderStageFlagBits::eFragment);
 	}
 	CreateDescriptorLayout(graphics, pipeline, descriptor_bindings);
-	const auto                  GraphicsStages = vertex_stage | vk::ShaderStageFlagBits::eFragment;
-	const vk::PushConstantRange push_constants {GraphicsStages, 0,
+	const vk::PushConstantRange push_constants {graphics_stages, 0,
 	                                            ShaderRecompiler::IR::NativePushConstantSize};
 
 	vk::PipelineLayoutCreateInfo pipeline_layout_info {};
@@ -456,8 +460,8 @@ void CreatePipelineInternal(
 		     " set_layouts=1 push_constants=%" PRIu32 "\n",
 		     vertex_program.id, ps_active ? pixel_program.id : 0, 1u);
 	}
-	result = graphics.device.createPipelineLayout(&pipeline_layout_info, nullptr,
-	                                              &pipeline.pipeline_layout);
+	auto result = graphics.device.createPipelineLayout(&pipeline_layout_info, nullptr,
+	                                                   &pipeline.pipeline_layout);
 	if (graphics_debug_dump_enabled()) {
 		LOGF("PipelineTrace: vkCreatePipelineLayout done result=%s layout=%p\n",
 		     vk::to_string(result).c_str(), static_cast<void*>(pipeline.pipeline_layout));
@@ -473,15 +477,6 @@ void CreatePipelineInternal(
 #else
 	    (static_params.depth_bounds_test_enable ? VK_TRUE : VK_FALSE);
 #endif
-	depth_stencil_info.stencilTestEnable = (static_params.stencil_test_enable ? VK_TRUE : VK_FALSE);
-	depth_stencil_info.front.failOp      = static_params.stencil_front.failOp;
-	depth_stencil_info.front.passOp      = static_params.stencil_front.passOp;
-	depth_stencil_info.front.depthFailOp = static_params.stencil_front.depthFailOp;
-	depth_stencil_info.front.compareOp   = static_params.stencil_front.compareOp;
-	depth_stencil_info.back.failOp       = static_params.stencil_back.failOp;
-	depth_stencil_info.back.passOp       = static_params.stencil_back.passOp;
-	depth_stencil_info.back.depthFailOp  = static_params.stencil_back.depthFailOp;
-	depth_stencil_info.back.compareOp    = static_params.stencil_back.compareOp;
 	depth_stencil_info.minDepthBounds    = static_params.depth_min_bounds;
 	depth_stencil_info.maxDepthBounds    = static_params.depth_max_bounds;
 
@@ -494,6 +489,8 @@ void CreatePipelineInternal(
 	    vk::DynamicState::eDepthCompareOp,
 	    vk::DynamicState::eDepthBiasEnable,
 	    vk::DynamicState::eDepthBias,
+	    vk::DynamicState::eStencilTestEnable,
+	    vk::DynamicState::eStencilOp,
 	    vk::DynamicState::eStencilCompareMask,
 	    vk::DynamicState::eStencilReference,
 	    vk::DynamicState::eStencilWriteMask,
@@ -524,8 +521,9 @@ void CreatePipelineInternal(
 	pipeline_info.pVertexInputState        = mesh ? nullptr : &vertex_input_info;
 	pipeline_info.pInputAssemblyState      = mesh ? nullptr : &input_assembly;
 	vk::PipelineTessellationStateCreateInfo tessellation_state {};
-	tessellation_state.patchControlPoints = 3;
-	pipeline_info.pTessellationState      = (rect_list ? &tessellation_state : nullptr);
+	tessellation_state.patchControlPoints =
+	    tessellation ? vs_input_info.tess.input_control_points : 3u;
+	pipeline_info.pTessellationState = (rect_list || tessellation) ? &tessellation_state : nullptr;
 	pipeline_info.pViewportState          = &viewport_state;
 	pipeline_info.pRasterizationState     = &rasterizer;
 	pipeline_info.pMultisampleState       = &multisampling;

src/graphics/host_gpu/renderer/render.h

@@ -170,15 +170,14 @@ private:
 	void DrawAuto(uint64_t submit_id, CommandBuffer& buffer, const DrawAutoArgs& args);
 
 	struct GraphicsBindings {
-		PreparedBindings                vertex;
+		std::array<PreparedBindings, 3> vertex;
 		std::optional<PreparedBindings> pixel;
 	};
 
 	[[nodiscard]] TextureBinding ResolveTexture(const ShaderRecompiler::IR::ImageResource& resource,
 	                                            const ShaderRecompiler::IR::DescriptorValue& value);
-	[[nodiscard]] GraphicsBindings PrepareGraphicsBindings(const ShaderStageRuntime& vertex,
-	                                                       const ShaderStageRuntime& pixel,
-	                                                       bool                      pixel_active);
+	void PrepareGraphicsBindings(std::span<PreparedBindings* const> stages,
+	                             std::span<RenderColorInfo> colors);
 	void ResolveRenderColorTarget(CommandBuffer& buffer, RenderColorInfo& target,
 	                              uint32_t render_target_slice_offset, uint32_t render_target_slot,
 	                              bool ignore_target_mask = false, bool exact_format = false);

src/graphics/host_gpu/renderer/renderCompute.cpp

@@ -172,7 +172,7 @@ bool RenderExecutor::TryConsumeComputeImageClear(const ShaderComputeInputInfo& i
 		                                       1, view.base_layer, view.layer_count};
 		vk::ClearValue clear {};
 		clear.depthStencil = vk::ClearDepthStencilValue {0.0f, fill.value};
-		cache.ClearImage(command, binding.image_id, range, clear);
+		cache.ClearImage(command, binding.image_id, image.backing.format, range, clear);
 		return true;
 	}
 	ShaderBufferResource descriptor;
@@ -183,14 +183,6 @@ bool RenderExecutor::TryConsumeComputeImageClear(const ShaderComputeInputInfo& i
 		return false;
 	}
 	if (!cache.ClearImageFromBuffer(command, descriptor.Base48(), size, packed_clear)) {
-		// Track deferred DCC state while the original dispatch writes the metadata allocation.
-		cache.TrackDccFill(descriptor.Base48(), size, packed_clear);
-		static std::atomic<uint32_t> logged_metadata_clears {0};
-		if (logged_metadata_clears.fetch_add(1, std::memory_order_relaxed) < 32) {
-			LOGF("GraphicsRenderDispatchDirect: metadata fill shader=0x%016" PRIx64
-			     " addr=0x%016" PRIx64 " size=0x%016" PRIx64 " value=0x%08" PRIx32 "\n",
-			     input.stage.program->shader_hash, descriptor.Base48(), size, packed_clear);
-		}
 		return false;
 	}
 	static std::atomic<uint32_t> logged_clears {0};
@@ -214,6 +206,17 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 	                    thread_group_x, thread_group_y, thread_group_z, mode,
 	                    sh_ctx.GetCs().cs_regs.data_addr);
 
+	if (thread_group_x == 0 || thread_group_y == 0 || thread_group_z == 0) {
+		static std::atomic<uint32_t> log_count {0};
+		if (log_count.fetch_add(1, std::memory_order_relaxed) < 32) {
+			LOGF("GraphicsRenderDispatchDirect: skipping zero-sized dispatch groups=%ux%ux%u "
+			     "mode=0x%08" PRIx32 " shader=0x%016" PRIx64 "\n",
+			     thread_group_x, thread_group_y, thread_group_z, mode,
+			     sh_ctx.GetCs().cs_regs.data_addr);
+		}
+		return;
+	}
+
 	Common::LockGuard lock(m_context.GetMutex());
 	if (sh_ctx.GetCs().cs_regs.data_addr == 0) {
 		LOGF("GraphicsRenderDispatchDirect: temporary: ignoring dispatch with null CS shader, "
@@ -355,17 +358,6 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 		}
 	}
 
-	if (thread_group_x == 0 || thread_group_y == 0 || thread_group_z == 0) {
-		static std::atomic<uint32_t> log_count {0};
-		if (log_count.fetch_add(1, std::memory_order_relaxed) < 32) {
-			LOGF("GraphicsRenderDispatchDirect: skipping zero-sized dispatch groups=%ux%ux%u "
-			     "mode=0x%08" PRIx32 " shader=0x%016" PRIx64 "\n",
-			     thread_group_x, thread_group_y, thread_group_z, mode,
-			     sh_ctx.GetCs().cs_regs.data_addr);
-		}
-		return;
-	}
-
 	buffer.EndRendering();
 	auto& pipeline =
 	    m_context.GetPipelineCache().GetComputePipeline(input_info, compute_program);
@@ -374,8 +366,8 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 	if (program.info.uses_dma) {
 		m_context.PrepareBda();
 	}
-	RebindBuffers(bindings);
 	RebindImages(bindings);
+	RebindBuffers(bindings);
 
 	auto              vk_buffer        = buffer.Handle();
 	PreparedBindings* descriptor_stage = &bindings;

src/graphics/host_gpu/renderer/renderDraw.cpp

@@ -15,6 +15,7 @@
 #include "graphics/host_gpu/renderer/colorRenderTarget.h"
 #include "graphics/host_gpu/renderer/debug.h"
 #include "graphics/host_gpu/renderer/depthRenderTarget.h"
+#include "graphics/host_gpu/renderer/image/textureCommon.h"
 #include "graphics/host_gpu/renderer/pipeline/pipelineCache.h"
 #include "graphics/host_gpu/renderer/pipeline/shaderResourceBarrier.h"
 #include "graphics/host_gpu/renderer/render.h"
@@ -308,24 +309,12 @@ static void LogDrawInputState(const CommandBuffer& buffer, const RenderColorInfo
 
 static void SetGraphicsDynamicParams(const CommandBuffer& buffer, vk::CommandBuffer vk_buffer,
                                      const ShaderVertexInputInfo& vs_input_info,
-                                     const RenderColorInfo* colors, uint32_t color_count,
-                                     const RenderDepthInfo& depth) {
+                                     const RenderDepthInfo& depth, const RenderState& rendering) {
 	KYTY_PROFILER_FUNCTION();
 
-	EXIT_IF(colors == nullptr);
 	const auto& ctx = buffer.GetRegisters();
-
-	const auto&  vp = ctx.GetScreenViewport();
-	vk::Extent2D framebuffer_extent {};
-	if (color_count > 0 && colors[0].image_id) {
-		framebuffer_extent = colors[0].Extent();
-	} else if (depth.image_id) {
-		framebuffer_extent = {depth.desc.info.extent.width, depth.desc.info.extent.height};
-	} else {
-		const auto& limits = buffer.GetGraphics().GetPhysicalDeviceProperties().limits;
-		framebuffer_extent = {limits.maxFramebufferWidth, limits.maxFramebufferHeight};
-	}
-
+	const auto&        vp  = ctx.GetScreenViewport();
+	const vk::Extent2D framebuffer_extent {rendering.width, rendering.height};
 	const auto& outputs = vs_input_info.stage.program->info.outputs;
 	const bool  indexed_viewports =
 	    std::any_of(outputs.begin(), outputs.end(), [](const auto& output) {
@@ -403,19 +392,16 @@ static void SetGraphicsDynamicParams(const CommandBuffer& buffer, vk::CommandBuf
 		vk_buffer.setDepthBias(constant_factor, poly_offset.clamp, slope_factor);
 	}
 
+	vk_buffer.setStencilTestEnable(depth.stencil_test_enable ? VK_TRUE : VK_FALSE);
 	if (depth.stencil_test_enable) {
-		vk_buffer.setStencilCompareMask(vk::StencilFaceFlagBits::eFront,
-		                                depth.stencil_dynamic_front.compareMask);
-		vk_buffer.setStencilCompareMask(vk::StencilFaceFlagBits::eBack,
-		                                depth.stencil_dynamic_back.compareMask);
-		vk_buffer.setStencilWriteMask(vk::StencilFaceFlagBits::eFront,
-		                              depth.stencil_dynamic_front.writeMask);
-		vk_buffer.setStencilWriteMask(vk::StencilFaceFlagBits::eBack,
-		                              depth.stencil_dynamic_back.writeMask);
-		vk_buffer.setStencilReference(vk::StencilFaceFlagBits::eFront,
-		                              depth.stencil_dynamic_front.reference);
-		vk_buffer.setStencilReference(vk::StencilFaceFlagBits::eBack,
-		                              depth.stencil_dynamic_back.reference);
+		const auto set_stencil = [&](vk::StencilFaceFlagBits face, const vk::StencilOpState& state) {
+			vk_buffer.setStencilOp(face, state.failOp, state.passOp, state.depthFailOp, state.compareOp);
+			vk_buffer.setStencilCompareMask(face, state.compareMask);
+			vk_buffer.setStencilWriteMask(face, state.writeMask);
+			vk_buffer.setStencilReference(face, state.reference);
+		};
+		set_stencil(vk::StencilFaceFlagBits::eFront, depth.stencil_front);
+		set_stencil(vk::StencilFaceFlagBits::eBack, depth.stencil_back);
 	}
 
 #if defined(__APPLE__)
@@ -423,15 +409,11 @@ static void SetGraphicsDynamicParams(const CommandBuffer& buffer, vk::CommandBuf
 	// eColorWriteEnableEXT dynamic state and relies on the static colorWriteMask instead.
 #else
 	vk::Bool32 enable[RENDER_COLOR_ATTACHMENTS_MAX] = {};
-	// Color-control operation selects special color-buffer paths, not the normal component write
-	// mask. Attachment availability therefore follows the target write mask.
-	for (uint32_t i = 0; i < color_count; i++) {
-		enable[i] = render_target_mask_slot(ctx.GetRenderTargetMask(), colors[i].target_slot) != 0
-		                ? VK_TRUE
-		                : VK_FALSE;
+	for (uint32_t slot = 0; slot < rendering.num_color_attachments; slot++) {
+		enable[slot] = rendering.color_attachments[slot].image_view != nullptr;
 	}
-	if (color_count != 0) {
-		vk_buffer.setColorWriteEnableEXT(color_count, enable);
+	if (rendering.num_color_attachments != 0) {
+		vk_buffer.setColorWriteEnableEXT(rendering.num_color_attachments, enable);
 	}
 #endif
 }
@@ -453,7 +435,7 @@ struct DrawRenderState {
 	RenderColorInfo       color_info[RENDER_COLOR_ATTACHMENTS_MAX] = {};
 	uint32_t              color_count                              = 0;
 	bool                  ps_active                                = true;
-	ShaderVertexInputInfo vs_input_info;
+	std::array<ShaderVertexInputInfo, 3> vertex_info;
 	ShaderPixelInputInfo  ps_input_info;
 	PipelineCache::GraphicsPrograms programs;
 };
@@ -477,19 +459,15 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
 	state.width                 = std::numeric_limits<uint32_t>::max();
 	state.height                = std::numeric_limits<uint32_t>::max();
 	state.num_layers            = std::numeric_limits<uint32_t>::max();
-	state.num_color_attachments = color_count;
+	state.num_color_attachments = 0;
 	uint32_t attachment_samples = 0;
 	for (uint32_t i = 0; i < color_count; i++) {
 		auto& target = colors[i];
 		EXIT_IF(!target.image_id);
-		const auto old_image = cache.m_slot_images.try_get(target.image_id);
-		if (old_image == nullptr || (!old_image->registered && !old_image->info.data.Empty()) ||
-		    old_image->binding.needs_rebind) {
-			if (old_image != nullptr) {
-				old_image->binding = {};
-			}
-			target.image_id = cache.FindImage(target.desc);
-			BindRenderTarget(target.image_id);
+		const auto owner = cache.m_slot_images.try_get(target.image_id);
+		if (owner == nullptr || (!owner->registered && !owner->info.data.Empty()) ||
+		    owner->binding.needs_rebind) {
+			EXIT("color target changed after render-state discovery\n");
 		}
 		const auto image_view = cache.FindRenderTarget(target.image_id, target.desc);
 		auto&      image      = cache.GetImage(target.image_id);
@@ -523,7 +501,8 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
 		state.width             = std::min(state.width, extent.width);
 		state.height            = std::min(state.height, extent.height);
 		state.num_layers        = std::min(state.num_layers, view.layer_count);
-		auto& attachment        = state.color_attachments[i];
+		state.num_color_attachments = std::max(state.num_color_attachments, target.target_slot + 1);
+		auto& attachment            = state.color_attachments[target.target_slot];
 		attachment.image_view   = image_view;
 		attachment.image_layout = layout;
 	}
@@ -709,6 +688,7 @@ struct PreparedVertexBuffers {
 
 	std::array<vk::Buffer, MaxBuffers>     buffers {};
 	std::array<vk::DeviceSize, MaxBuffers> offsets {};
+	std::array<vk::DeviceSize, MaxBuffers> sizes {};
 	uint32_t                               count = 0;
 };
 
@@ -793,6 +773,7 @@ static PreparedVertexBuffers AcquireVertexBuffers(CommandBuffer&               b
 
 		prepared.buffers[i] = range->binding.first->Handle();
 		prepared.offsets[i] = range->binding.second + vertex.addr - range->base_address;
+		prepared.sizes[i]   = std::min(size, range->acquired_end - vertex.addr);
 		SetVulkanObjectNameF(
 		    buffer.GetContext().GetGraphics().device, prepared.buffers[i],
 		    "Kyty.VertexBuffer[slot={} guest=0x{:016x} size=0x{:x} stride={} records={}]", i,
@@ -831,6 +812,7 @@ static bool GetDrawTopology(const HW::UserConfig& ucfg, bool auto_draw,
 		case Prospero::PrimitiveType::kTriStrip:
 			topology = vk::PrimitiveTopology::eTriangleStrip;
 			break;
+		case Prospero::PrimitiveType::kPatch:
 		case Prospero::PrimitiveType::kRectList:
 			topology = vk::PrimitiveTopology::ePatchList;
 			break;
@@ -895,42 +877,6 @@ static bool ResolvePrimitiveRestart(const CommandBuffer& buffer,
 	return false;
 }
 
-bool RenderExecutor::PrepareDrawRenderState(CommandBuffer& buffer, const DrawCallInfo& draw,
-                                            uint32_t            render_target_slice_offset,
-	                                        DrawRenderState& state) {
-	auto& ctx = buffer.GetRegisters();
-
-	if (ResolveColorTargets(buffer, render_target_slice_offset)) {
-		return false;
-	}
-	if (draw.IsIndexed()) {
-		LogDrawPhase(draw.Name(), "ResolveRenderColorTarget");
-	}
-	for (uint32_t slot = 0; slot < RENDER_COLOR_ATTACHMENTS_MAX; slot++) {
-		if (slot == 0 || (render_target_mask_slot(ctx.GetRenderTargetMask(), slot) != 0 &&
-		                  ctx.GetRenderTarget(slot).base.addr != 0)) {
-			ResolveRenderColorTarget(buffer, state.color_info[state.color_count],
-			                         render_target_slice_offset, slot);
-			if (state.color_info[state.color_count].image_id) {
-				state.color_count++;
-			}
-		}
-	}
-	if (draw.IsIndexed()) {
-		LogDrawPhase(draw.Name(), "ResolveRenderDepthTarget");
-	}
-	ResolveRenderDepthTarget(buffer, state.depth_info);
-
-	state.ps_active       = DrawHasActivePixelShader(buffer);
-	if (state.color_count == 0 && !state.depth_info.image_id && !state.ps_active) {
-		LogFramebufferSkip(draw.Name(), state.color_info[0], state.depth_info, buffer,
-		                   draw.index_count, 0);
-		return false;
-	}
-
-	return true;
-}
-
 static void RefreshShaders(CommandBuffer& buffer, const DrawCallInfo& draw,
                            DrawRenderState& state) {
 	auto& ctx    = buffer.GetRegisters();
@@ -944,8 +890,14 @@ static void RefreshShaders(CommandBuffer& buffer, const DrawCallInfo& draw,
 	state.ps_input_info = {};
 	std::array<Prospero::ColorComponentMapping, RENDER_COLOR_ATTACHMENTS_MAX>
 	    target_export_mapping {};
-	for (uint32_t i = 0; i < state.color_count; i++) {
-		target_export_mapping[state.color_info[i].target_slot] = state.color_info[i].export_mapping;
+	for (uint32_t slot = 0; slot < RENDER_COLOR_ATTACHMENTS_MAX; slot++) {
+		const auto& rt = ctx.GetRenderTarget(slot);
+		if (rt.base.addr != 0 && render_target_mask_slot(ctx.GetRenderTargetMask(), slot) != 0) {
+			target_export_mapping[slot] =
+			    TextureGetRenderTargetFormat(rt.info.format, rt.info.channel_type,
+			                                 rt.info.channel_order)
+			        .export_mapping;
+		}
 	}
 	auto& pipeline_cache = buffer.GetContext().GetPipelineCache();
 	if (draw.IsIndexed()) {
@@ -953,7 +905,46 @@ static void RefreshShaders(CommandBuffer& buffer, const DrawCallInfo& draw,
 	}
 	state.programs = pipeline_cache.GetGraphicsPrograms(
 	    vertex_shader_info, pixel_shader_info, shader_regs, ctx, buffer.GetUserConfig(),
-	    target_export_mapping, state.ps_active, state.vs_input_info, state.ps_input_info);
+	    target_export_mapping, state.ps_active, state.vertex_info, state.ps_input_info);
+}
+
+bool RenderExecutor::PrepareDrawRenderState(CommandBuffer& buffer, const DrawCallInfo& draw,
+                                            uint32_t            render_target_slice_offset,
+	                                        DrawRenderState& state) {
+	state.ps_active = DrawHasActivePixelShader(buffer);
+	RefreshShaders(buffer, draw, state);
+	uint32_t mrt_mask = 0;
+	if (state.ps_active) {
+		for (const auto& output: state.ps_input_info.stage.program->info.outputs) {
+			if (output.kind == ShaderRecompiler::IR::StageOutputKind::Mrt) {
+				mrt_mask |= 1u << output.index;
+			}
+		}
+	}
+	if (draw.IsIndexed()) {
+		LogDrawPhase(draw.Name(), "ResolveRenderColorTarget");
+	}
+	for (uint32_t slot = 0; slot < RENDER_COLOR_ATTACHMENTS_MAX; slot++) {
+		if ((mrt_mask & (1u << slot)) != 0) {
+			ResolveRenderColorTarget(buffer, state.color_info[state.color_count],
+			                         render_target_slice_offset, slot);
+			if (state.color_info[state.color_count].image_id) {
+				state.color_count++;
+			}
+		}
+	}
+	if (draw.IsIndexed()) {
+		LogDrawPhase(draw.Name(), "ResolveRenderDepthTarget");
+	}
+	ResolveRenderDepthTarget(buffer, state.depth_info);
+
+	if (state.color_count == 0 && !state.depth_info.image_id && !state.ps_active) {
+		LogFramebufferSkip(draw.Name(), state.color_info[0], state.depth_info, buffer,
+		                   draw.index_count, 0);
+		return false;
+	}
+
+	return true;
 }
 
 static PreparedIndexBuffer PrepareIndexBuffer(CommandBuffer&               buffer,
@@ -988,8 +979,9 @@ static void CommitVertexBuffers(vk::CommandBuffer            vk_buffer,
 		EXIT_IF(prepared.buffers[i] == nullptr);
 	}
 	if (prepared.count != 0) {
-		vk_buffer.bindVertexBuffers(0, prepared.count, prepared.buffers.data(),
-		                            prepared.offsets.data());
+		// Guest descriptor bounds must survive allocation merging in the cache.
+		vk_buffer.bindVertexBuffers2(0, prepared.count, prepared.buffers.data(),
+		                             prepared.offsets.data(), prepared.sizes.data(), nullptr);
 	}
 }
 
@@ -1016,7 +1008,7 @@ static void LogDrawStateIfNeeded(const CommandBuffer& buffer, const DrawCallInfo
 		LogDrawTargetState(draw.Name(), state.color_info[0], state.depth_info, buffer,
 		                   state.ps_input_info, draw.index_count, 0);
 	}
-	LogDrawInputState(buffer, state.color_info[0], state.vs_input_info, index_type_and_size,
+	LogDrawInputState(buffer, state.color_info[0], state.vertex_info[0], index_type_and_size,
 	                  draw.index_count, index_addr);
 }
 
@@ -1031,6 +1023,7 @@ static void EmitDrawPrimitives(const HW::UserConfig& ucfg, vk::CommandBuffer vk_
 		case Prospero::PrimitiveType::kTriFan:
 		case Prospero::PrimitiveType::kTriStrip:
 		case Prospero::PrimitiveType::kRectList:
+		case Prospero::PrimitiveType::kPatch:
 			if (draw.IsIndexed()) {
 				vk_buffer.drawIndexed(draw.index_count, draw.instance_count, 0, emit.vertex_offset,
 				                      emit.first_instance);
@@ -1069,11 +1062,19 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
                                          const DrawIndexBufferSource& index_source,
 	                                     bool primitive_restart_enable) {
 	auto& ucfg = buffer.GetUserConfig();
-	const bool mesh_active = state.vs_input_info.stage.program->stage == ShaderType::Mesh;
+	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;
 	uint32_t   mesh_groups = 0;
 	if (mesh_active) {
-		const auto& mesh = state.vs_input_info.mesh;
-		if (primitive_restart_enable || mesh.primitives_per_group == 0) {
+		const auto& mesh = state.vertex_info[0].mesh;
+		static std::atomic_bool restart_warned = false;
+		if (primitive_restart_enable && !restart_warned.exchange(true, std::memory_order_relaxed)) {
+			std::printf("Warning: primitive restart is not implemented for mesh shaders; "
+			            "continuing draw (primitive=%u indexed=%u)\n",
+			            static_cast<uint32_t>(ucfg.GetPrimType()), draw.IsIndexed());
+		}
+		if (mesh.primitives_per_group == 0) {
 			EXIT("unsupported mesh draw: primitive=%u indexed=%u restart=%u\n",
 			     static_cast<uint32_t>(ucfg.GetPrimType()), draw.IsIndexed(), primitive_restart_enable);
 		}
@@ -1100,13 +1101,24 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 		                              index_source.guest_element_size);
 	}
 	LogDrawPhase(draw.Name(), "PrepareBindings");
-	auto bindings = PrepareGraphicsBindings(state.vs_input_info.stage, state.ps_input_info.stage,
-	                                        state.ps_active);
+	GraphicsBindings                 bindings;
+	std::array<PreparedBindings*, 4> descriptor_stages {};
+	uint32_t                         stage_count = 0;
+	for (uint32_t i = 0; i < vertex_stages.size(); i++) {
+		bindings.vertex[i]               = PrepareBindings(state.vertex_info[i].stage);
+		descriptor_stages[stage_count++] = &bindings.vertex[i];
+	}
+	if (state.ps_active) {
+		bindings.pixel.emplace(PrepareBindings(state.ps_input_info.stage));
+		descriptor_stages[stage_count++] = &*bindings.pixel;
+	}
+	const auto stages = std::span {descriptor_stages.data(), stage_count};
+	PrepareGraphicsBindings(stages, std::span {state.color_info, state.color_count});
 	PreparedVertexBuffers vertex_bindings;
 	PreparedIndexBuffer   index_binding;
 	if (!mesh_active) {
 		LogDrawPhase(draw.Name(), "PrepareVertexBuffers");
-		vertex_bindings = AcquireVertexBuffers(buffer, state.vs_input_info);
+		vertex_bindings = AcquireVertexBuffers(buffer, state.vertex_info[0]);
 		index_binding   = PrepareIndexBuffer(buffer, index_source);
 	}
 	const auto rendering =
@@ -1117,9 +1129,9 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 		LogDrawPhase(draw.Name(), "CreatePipeline");
 	}
 	auto& pipeline = m_context.GetPipelineCache().GetGraphicsPipeline(
-	    std::span {state.color_info, state.color_count}, state.depth_info, state.vs_input_info, buffer,
+	    std::span {state.color_info, state.color_count}, state.depth_info, vertex_stages, buffer,
 	    state.ps_active ? &state.ps_input_info : nullptr, topology, primitive_restart_enable,
-	    state.programs.vertex, state.programs.pixel);
+	    state.programs);
 
 	// Resource preparation above may synchronously finish and restart the scheduler. From this
 	// point onward, every operation targets the current command buffer and cannot touch guest
@@ -1132,13 +1144,7 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 	if (bindings.pixel && !draw.IsIndexed()) {
 		SetDrawDebugPhase(buffer, submit_id, draw, 0x300u);
 	}
-	std::array<PreparedBindings*, 2> descriptor_stages {&bindings.vertex, nullptr};
-	const size_t                     descriptor_stage_count = bindings.pixel.has_value() ? 2u : 1u;
-	if (bindings.pixel) {
-		descriptor_stages[1] = &*bindings.pixel;
-	}
-	CommitBindings(buffer, vk::PipelineBindPoint::eGraphics, pipeline,
-	               std::span {descriptor_stages.data(), descriptor_stage_count});
+	CommitBindings(buffer, vk::PipelineBindPoint::eGraphics, pipeline, stages);
 	if (mesh_active) {
 		const uint32_t draw_data[] {
 		    draw.index_count,
@@ -1155,8 +1161,7 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 		CommitIndexBuffer(vk_buffer, index_binding);
 	}
 
-	SetGraphicsDynamicParams(buffer, vk_buffer, state.vs_input_info, state.color_info,
-	                         state.color_count, state.depth_info);
+	SetGraphicsDynamicParams(buffer, vk_buffer, vertex_stages.back(), state.depth_info, rendering);
 	if (m_context.GetGraphics().attachment_feedback_loop_enabled) {
 		vk_buffer.setAttachmentFeedbackLoopEnableEXT(
 		    rendering.depth_stencil_attachment.image_layout ==
@@ -1177,16 +1182,17 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 	if (mesh_active) {
 		vk_buffer.drawMeshTasksEXT(mesh_groups, draw.instance_count, 1);
 	} else {
-		EmitDrawPrimitives(ucfg, vk_buffer, state.vs_input_info, draw, emit);
+		EmitDrawPrimitives(ucfg, vk_buffer, state.vertex_info[0], draw, emit);
 	}
 
 	if (!draw.IsIndexed()) {
 		SetDrawDebugPhase(buffer, submit_id, draw, 0x600u);
 	}
 	vk::PipelineStageFlags shader_write_stages = {};
-	if (HasShaderBufferWrites(state.vs_input_info.stage)) {
-		shader_write_stages |= mesh_active ? vk::PipelineStageFlagBits::eMeshShaderEXT
-		                                   : vk::PipelineStageFlagBits::eVertexShader;
+	for (const auto& stage: vertex_stages) {
+		if (HasShaderBufferWrites(stage.stage)) {
+			shader_write_stages |= ShaderPipelineStages(NativeShaderStage(stage.logical_stage));
+		}
 	}
 	if (state.ps_active && HasShaderBufferWrites(state.ps_input_info.stage)) {
 		shader_write_stages |= vk::PipelineStageFlagBits::eFragmentShader;
@@ -1220,7 +1226,8 @@ void RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
 		return;
 	}
 
-	if (ConsumeMetadataColorOperation(buffer) || DepthStencilCopy(buffer)) {
+	if (ConsumeMetadataColorOperation(buffer) || DepthStencilCopy(buffer) ||
+	    ResolveColorTargets(buffer, args.render_target_slice_offset)) {
 		ResetBindings();
 		return;
 	}
@@ -1294,15 +1301,13 @@ void RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
 		return;
 	}
 
-	RefreshShaders(buffer, draw, state);
-
 	LogDrawStateIfNeeded(buffer, draw, state, args.index_type_and_size,
 	                     args.index_addr);
 
 	const bool indirect = args.offset_source == DrawOffsetSource::IndirectArgs;
 	const auto [vertex_offset, instance_offset] =
 	    indirect ? std::pair<int32_t, uint32_t> {0, args.first_instance}
-	             : ResolveDrawOffsets(ucfg.GetIndexOffset(), state.vs_input_info);
+	             : ResolveDrawOffsets(ucfg.GetIndexOffset(), state.vertex_info[0]);
 
 	DrawEmitInfo emit {};
 	emit.vertex_offset  = vertex_offset + args.base_vertex;
@@ -1332,7 +1337,8 @@ void RenderExecutor::DrawAuto(uint64_t submit_id, CommandBuffer& buffer, const D
 		return;
 	}
 
-	if (ConsumeMetadataColorOperation(buffer) || DepthStencilCopy(buffer)) {
+	if (ConsumeMetadataColorOperation(buffer) || DepthStencilCopy(buffer) ||
+	    ResolveColorTargets(buffer, args.render_target_slice_offset)) {
 		ResetBindings();
 		return;
 	}
@@ -1361,22 +1367,20 @@ void RenderExecutor::DrawAuto(uint64_t submit_id, CommandBuffer& buffer, const D
 	const DrawCallInfo draw {CommandBufferDebugOp::DrawIndexAuto,
 	                         args.vertex_count, args.instance_count, args.first_instance};
 
+	vk::PrimitiveTopology topology = vk::PrimitiveTopology::ePointList;
+	if (!GetDrawTopology(ucfg, true, topology)) {
+		ResetBindings();
+		return;
+	}
 	DrawRenderState state {};
 	if (!PrepareDrawRenderState(buffer, draw, args.render_target_slice_offset, state)) {
 		ResetBindings();
 		return;
 	}
 
-	vk::PrimitiveTopology topology = vk::PrimitiveTopology::ePointList;
-	if (!GetDrawTopology(ucfg, true, topology)) {
-		ResetBindings();
-		return;
-	}
-	RefreshShaders(buffer, draw, state);
-
-	const bool rect_list = topology == vk::PrimitiveTopology::ePatchList;
-	if (rect_list && state.vs_input_info.buffers_num == 0 &&
-	    state.vs_input_info.stage.program->param_export_mask == 0 &&
+	const bool rect_list = ucfg.GetPrimType() == Prospero::PrimitiveType::kRectList;
+	if (rect_list && state.vertex_info[0].buffers_num == 0 &&
+	    state.vertex_info[0].stage.program->param_export_mask == 0 &&
 	    state.ps_input_info.input_num != 0) {
 		if (graphics_debug_dump_enabled()) {
 			LOGF("DrawIndexAuto: skipping rect-list draw with no VS param exports and PS inputs: "
@@ -1393,7 +1397,7 @@ void RenderExecutor::DrawAuto(uint64_t submit_id, CommandBuffer& buffer, const D
 	const bool indirect = args.offset_source == DrawOffsetSource::IndirectArgs;
 	const auto [vertex_offset, instance_offset] =
 	    indirect ? std::pair<int32_t, uint32_t> {0, args.first_instance}
-	             : ResolveDrawOffsets(ucfg.GetIndexOffset(), state.vs_input_info);
+	             : ResolveDrawOffsets(ucfg.GetIndexOffset(), state.vertex_info[0]);
 	DrawEmitInfo emit {};
 	emit.first_vertex = static_cast<uint32_t>(vertex_offset + static_cast<int32_t>(args.first_vertex));
 	emit.first_instance = instance_offset;

src/graphics/host_gpu/renderer/renderTarget.h

@@ -5,7 +5,6 @@
 
 #include <array>
 #include <cstdint>
-#include <type_traits>
 
 namespace Libs::Graphics {
 
@@ -67,33 +66,6 @@ inline constexpr TargetViewInfo ResolveTargetViewInfo(uint32_t base_layer, uint3
 	        base_layer, last_layer - base_layer + 1u, last_layer + 1u};
 }
 
-#pragma pack(push, 1)
-
-struct PipelineStencilStaticState {
-	vk::StencilOp failOp      = vk::StencilOp::eKeep;
-	vk::StencilOp passOp      = vk::StencilOp::eKeep;
-	vk::StencilOp depthFailOp = vk::StencilOp::eKeep;
-	vk::CompareOp compareOp   = vk::CompareOp::eNever;
-};
-
-struct PipelineStencilDynamicState {
-	uint32_t compareMask = 0;
-	uint32_t writeMask   = 0;
-	uint32_t reference   = 0;
-};
-
-#pragma pack(pop)
-
-static_assert(std::is_trivially_copyable_v<PipelineStencilStaticState>);
-static_assert(std::is_standard_layout_v<PipelineStencilStaticState>);
-static_assert(alignof(PipelineStencilStaticState) == 1);
-static_assert(sizeof(PipelineStencilStaticState) ==
-              sizeof(vk::StencilOp) * 3 + sizeof(vk::CompareOp));
-static_assert(std::is_trivially_copyable_v<PipelineStencilDynamicState>);
-static_assert(std::is_standard_layout_v<PipelineStencilDynamicState>);
-static_assert(alignof(PipelineStencilDynamicState) == 1);
-static_assert(sizeof(PipelineStencilDynamicState) == sizeof(uint32_t) * 3);
-
 } // namespace Libs::Graphics
 
 #endif // EMULATOR_SRC_GRAPHICS_HOST_GPU_RENDERER_RENDERTARGET_H_

src/graphics/host_gpu/vulkanCommon.cpp

@@ -123,4 +123,14 @@ void RequireVulkanSuccess(vk::Result result, const char* operation) {
 	}
 }
 
+vk::ShaderModule CompileSPV(std::span<const uint32_t> code, vk::Device device) {
+	vk::ShaderModuleCreateInfo create_info {};
+	create_info.codeSize    = code.size_bytes();
+	create_info.pCode       = code.data();
+	vk::ShaderModule module = nullptr;
+	RequireVulkanSuccess(device.createShaderModule(&create_info, nullptr, &module),
+	                     "create SPIR-V shader module");
+	return module;
+}
+
 } // namespace Libs::Graphics

src/graphics/host_gpu/vulkanCommon.h

@@ -15,6 +15,7 @@
 
 #include <cstdint>
 #include <fmt/format.h>
+#include <span>
 #include <string>
 #include <utility>
 #include <vector>
@@ -27,6 +28,7 @@ using VulkanMemoryBarrier = vk::MemoryBarrier;
 
 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);
 
 template <typename Handle, typename... Args>
 void SetVulkanObjectNameF(vk::Device device, Handle handle, fmt::format_string<Args...> format,

src/graphics/presentation/videoOut.cpp

@@ -237,7 +237,9 @@ private:
 
 struct VideoOutDriver::Impl {
 public:
-	static constexpr int VIDEO_OUT_NUM_MAX = 4;
+	static constexpr std::array VIDEO_OUT_BUSES {
+	    VIDEO_OUT_BUS_TYPE_MAIN, VIDEO_OUT_BUS_TYPE_OVERLAY, VIDEO_OUT_BUS_TYPE_SUB};
+	static constexpr int VIDEO_OUT_NUM_MAX = VIDEO_OUT_BUSES.size() + 1;
 
 	Impl(uint32_t width, uint32_t height, Graphics::Presenter& presenter)
 	    : m_renderer(presenter.Renderer()), m_presenter(presenter), m_flip_queue(presenter) {
@@ -248,7 +250,7 @@ public:
 	~Impl();
 	KYTY_CLASS_NO_COPY(Impl);
 
-	int             Open(int bus_type);
+	int             Open(int bus_type, int index);
 	bool            Close(int handle);
 	VideoOutConfig* Get(int handle);
 	VideoOutConfig* Get(int handle, uint64_t& generation);
@@ -610,12 +612,17 @@ void VideoOutDriver::Impl::Init(uint32_t width, uint32_t height) {
 	}
 }
 
-int VideoOutDriver::Impl::Open(int bus_type) {
+int VideoOutDriver::Impl::Open(int bus_type, int index) {
+	const auto bus = std::find(VIDEO_OUT_BUSES.begin(), VIDEO_OUT_BUSES.end(), bus_type);
+	if (bus == VIDEO_OUT_BUSES.end()) {
+		return VIDEO_OUT_ERROR_INVALID_VALUE;
+	}
+	EXIT_NOT_IMPLEMENTED(index != 0);
 	Common::LockGuard lock(m_mutex);
 
-	const int handle = bus_type + 1;
+	const int handle = static_cast<int>(bus - VIDEO_OUT_BUSES.begin()) + 1;
 	if (m_video_out_ctx[handle].opened) {
-		return -1;
+		return VIDEO_OUT_ERROR_RESOURCE_BUSY;
 	}
 	auto&             config = m_video_out_ctx[handle];
 	Common::LockGuard config_lock(config.mutex);
@@ -1169,21 +1176,9 @@ KYTY_SYSV_ABI int VideoOutOpen(int user_id, int bus_type, int index, const void*
 	PRINT_NAME();
 
 	EXIT_NOT_IMPLEMENTED(user_id != 255 && user_id != 0);
-	if (bus_type != VIDEO_OUT_BUS_TYPE_MAIN && bus_type != VIDEO_OUT_BUS_TYPE_OVERLAY &&
-	    bus_type != VIDEO_OUT_BUS_TYPE_SUB) {
-		return VIDEO_OUT_ERROR_INVALID_VALUE;
-	}
-	EXIT_NOT_IMPLEMENTED(index != 0);
-
 	LOGF("\t param = 0x%016" PRIx64 "\n", reinterpret_cast<uint64_t>(param));
 
-	int handle = DriverState().Open(bus_type);
-
-	if (handle < 0) {
-		return VIDEO_OUT_ERROR_RESOURCE_BUSY;
-	}
-
-	return handle;
+	return DriverState().Open(bus_type, index);
 }
 
 KYTY_SYSV_ABI int VideoOutClose(int handle) {

src/graphics/presentation/window/hostInput.cpp

@@ -220,6 +220,8 @@ uint32_t DefaultKeyboardButton(int key_code) {
 		case SDLK_l: return Controller::PAD_BUTTON_CIRCLE;
 		case SDLK_q: return Controller::PAD_BUTTON_L1;
 		case SDLK_e: return Controller::PAD_BUTTON_R1;
+		case SDLK_z: return Controller::PAD_BUTTON_L2;
+		case SDLK_c: return Controller::PAD_BUTTON_R2;
 		case SDLK_LSHIFT: return Controller::PAD_BUTTON_L3;
 		case SDLK_LCTRL: return Controller::PAD_BUTTON_R3;
 		case SDLK_RETURN:

src/graphics/presentation/window/vulkanWindow.cpp

@@ -201,6 +201,8 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 		vk::PhysicalDeviceVulkan13Features features13 {};
 
 		vk::PhysicalDeviceColorWriteEnableFeaturesEXT color_write_ext {};
+		vk::PhysicalDeviceImageViewMinLodFeaturesEXT  image_view_min_lod {};
+		color_write_ext.pNext = &image_view_min_lod;
 
 		vk::PhysicalDeviceDepthClipEnableFeaturesEXT depth_clip_enable {};
 		depth_clip_enable.pNext = &color_write_ext;
@@ -235,6 +237,10 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 			skip_device = true;
 #endif
 		}
+		if (image_view_min_lod.minLod != VK_TRUE) {
+			LOGF("image view minLod is not supported\n");
+			skip_device = true;
+		}
 
 		if (depth_clip_control.depthClipControl != VK_TRUE) {
 			LOGF("depthClipControl is not supported\n");
@@ -514,11 +520,14 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 	depth_clip_enable.depthClipEnable = VK_TRUE;
 
 	vk::PhysicalDeviceDepthClipControlFeaturesEXT depth_clip_control {};
+	vk::PhysicalDeviceImageViewMinLodFeaturesEXT  image_view_min_lod {};
+	image_view_min_lod.minLod = VK_TRUE;
+	depth_clip_control.pNext  = &image_view_min_lod;
 	// MoltenVK lacks VK_EXT_depth_clip_enable and VK_EXT_color_write_enable, so drop those
 	// feature structs from the chain on macOS (the renderer falls back to default depth
 	// clipping and static color-write masks).
 #if !defined(__APPLE__)
-	depth_clip_control.pNext = &depth_clip_enable;
+	image_view_min_lod.pNext = &depth_clip_enable;
 #endif
 	depth_clip_control.depthClipControl = VK_TRUE;
 
@@ -984,7 +993,8 @@ void WindowContext::CreateVulkan() {
 
 	std::vector<const char*> device_extensions = {
 	    VK_KHR_SWAPCHAIN_EXTENSION_NAME, VK_EXT_DEPTH_CLIP_CONTROL_EXTENSION_NAME,
-	    VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME, "VK_KHR_maintenance1"};
+	    VK_EXT_IMAGE_VIEW_MIN_LOD_EXTENSION_NAME, VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME,
+	    "VK_KHR_maintenance1"};
 
 #if defined(__APPLE__)
 	// MoltenVK lacks VK_EXT_depth_clip_enable and VK_EXT_color_write_enable; the renderer

src/graphics/shader/recompiler/ShaderRecompiler.cpp

@@ -1,4 +1,5 @@
 #include "graphics/shader/recompiler/ShaderRecompiler.h"
+#include "graphics/shader/recompiler/Tessellation.h"
 
 #include "common/assert.h"
 #include "common/logging/log.h"
@@ -47,6 +48,9 @@ const char* StageName(ShaderType stage) {
 	switch (stage) {
 		case ShaderType::Compute: return "CS";
 		case ShaderType::Vertex: return "VS";
+		case ShaderType::Local: return "LS";
+		case ShaderType::TessellationControl: return "HS";
+		case ShaderType::TessellationEvaluation: return "TES";
 		case ShaderType::Mesh: return "MS";
 		case ShaderType::Pixel: return "PS";
 		default: return "unknown";
@@ -83,7 +87,6 @@ struct EmbeddedFetchSgprInfo {
 	EmbeddedFetchValueType type      = EmbeddedFetchValueType::Unknown;
 	int                    attrib_id = 0;
 	uint32_t               value     = 0;
-	std::vector<uint32_t>  prolog_loads;
 };
 
 using EmbeddedFetchVectorLanes = std::map<uint64_t, EmbeddedFetchSgprInfo>;
@@ -211,6 +214,8 @@ EmbeddedFetchData DetectEmbeddedVertexFetch(const Decoder::Program&      decoded
                                             const ShaderVertexInputInfo* input_info,
                                             uint32_t user_data_base, uint32_t user_data_count,
                                             uint32_t wave_size) {
+	const uint32_t    vertex_index_reg   = input_info->logical_stage == ShaderType::Local ? 2u : 5u;
+	const uint32_t    instance_index_reg = input_info->logical_stage == ShaderType::Local ? 5u : 8u;
 	EmbeddedFetchData data;
 	data.loads.reserve(input_info->resources_num);
 	int32_t vertex_offset_candidate   = -1;
@@ -251,16 +256,17 @@ EmbeddedFetchData DetectEmbeddedVertexFetch(const Decoder::Program&      decoded
 		// corresponding direct-draw offsets before fetching.
 		const bool vertex_index_accumulator =
 		    IsDecodedVgpr(inst.dst) &&
-		    (inst.dst.reg == 0 || (user_data_base == 8 && inst.dst.reg == 5));
+		    (inst.dst.reg == 0 || (user_data_base == 8 && inst.dst.reg == vertex_index_reg));
 		const bool instance_index_accumulator =
-		    IsDecodedVgpr(inst.dst) && (inst.dst.reg == (user_data_base == 8 ? 8u : 3u));
+		    IsDecodedVgpr(inst.dst) &&
+		    (inst.dst.reg == (user_data_base == 8 ? instance_index_reg : 3u));
 		uint32_t   sad_zero = 0;
 		const bool index_offset_add =
-		    (vertex_index_accumulator || instance_index_accumulator) &&
-		    IsDecodedSgpr(inst.src0) &&
+		    (vertex_index_accumulator || instance_index_accumulator) && IsDecodedSgpr(inst.src0) &&
 		    ((inst.opcode == Decoder::Opcode::V_ADD_I32 && IsDecodedVgpr(inst.src1) &&
 		      inst.src1.reg == inst.dst.reg) ||
-		     (user_data_base == 8 && (inst.dst.reg == 5 || inst.dst.reg == 8) &&
+		     (user_data_base == 8 &&
+		      (inst.dst.reg == vertex_index_reg || inst.dst.reg == instance_index_reg) &&
 		      inst.opcode == Decoder::Opcode::V_SAD_U32 && IsDecodedVgpr(inst.src2) &&
 		      inst.src2.reg == inst.dst.reg &&
 		      TryDecodedOperandConstant(sgprs, inst.src1, sad_zero) && sad_zero == 0));
@@ -319,7 +325,6 @@ EmbeddedFetchData DetectEmbeddedVertexFetch(const Decoder::Program&      decoded
 						auto& dst = sgprs[DecodedSgprReg(inst.dst)];
 						dst.type  = EmbeddedFetchValueType::Constant;
 						dst.value = value;
-						dst.prolog_loads.clear();
 					} else {
 						ClearEmbeddedFetchSgprs(sgprs, inst.dst, 1);
 					}
@@ -330,7 +335,6 @@ EmbeddedFetchData DetectEmbeddedVertexFetch(const Decoder::Program&      decoded
 					auto& dst = sgprs[DecodedSgprReg(inst.dst)];
 					dst.type  = EmbeddedFetchValueType::Constant;
 					dst.value = inst.src0.value;
-					dst.prolog_loads.clear();
 				}
 				break;
 			default:
@@ -347,7 +351,6 @@ EmbeddedFetchData DetectEmbeddedVertexFetch(const Decoder::Program&      decoded
 								auto& dst        = sgprs[register_id + i];
 								dst.type         = EmbeddedFetchValueType::Attrib;
 								dst.attrib_id    = index + static_cast<int>(i);
-								dst.prolog_loads = {inst.pc};
 							}
 						} else {
 							ClearEmbeddedFetchSgprs(sgprs, inst.dst, DecodedDstSize(inst));
@@ -365,7 +368,6 @@ EmbeddedFetchData DetectEmbeddedVertexFetch(const Decoder::Program&      decoded
 								dst.type  = EmbeddedFetchValueType::Buffer;
 								dst.attrib_id =
 								    BufferTableAttribFromOffset(raw_offset, static_cast<int>(i));
-								dst.prolog_loads = {inst.pc};
 							}
 						} else if (IsDecodedSgpr(inst.src1) &&
 						           DecodedSgprReg(inst.src1) < sgprs.size() &&
@@ -377,8 +379,6 @@ EmbeddedFetchData DetectEmbeddedVertexFetch(const Decoder::Program&      decoded
 								auto& dst        = sgprs[register_id + i];
 								dst.type         = EmbeddedFetchValueType::Buffer;
 								dst.attrib_id    = sgprs[DecodedSgprReg(inst.src1)].attrib_id;
-								dst.prolog_loads = sgprs[DecodedSgprReg(inst.src1)].prolog_loads;
-								dst.prolog_loads.push_back(inst.pc);
 							}
 						} else {
 							ClearEmbeddedFetchSgprs(sgprs, inst.dst, DecodedDstSize(inst));
@@ -391,8 +391,8 @@ EmbeddedFetchData DetectEmbeddedVertexFetch(const Decoder::Program&      decoded
 						ClearEmbeddedFetchVectorLanes(&vector_lanes, inst.dst.reg);
 					}
 					if (IsDecodedVgpr(inst.dst) && inst.dst.reg < vgpr_is_index.size() &&
-					    IsDecodedVgpr(inst.src0) && inst.src0.reg == 8 &&
-					    IsDecodedVgpr(inst.src1) && inst.src1.reg == 5) {
+					    IsDecodedVgpr(inst.src0) && inst.src0.reg == instance_index_reg &&
+					    IsDecodedVgpr(inst.src1) && inst.src1.reg == vertex_index_reg) {
 						vgpr_is_index[inst.dst.reg] = true;
 					}
 				} else if (IsEmbeddedFetchAttribPropagationAlu(inst)) {
@@ -417,7 +417,6 @@ EmbeddedFetchData DetectEmbeddedVertexFetch(const Decoder::Program&      decoded
 									break;
 								default: dst.value = src0 + src1; break;
 							}
-							dst.prolog_loads.clear();
 						} else {
 							ClearEmbeddedFetchSgprs(sgprs, inst.dst, 1);
 						}
@@ -440,7 +439,6 @@ EmbeddedFetchData DetectEmbeddedVertexFetch(const Decoder::Program&      decoded
 						load.pc           = inst.pc;
 						load.attrib_id    = buffer.attrib_id;
 						load.components   = DecodedDstSize(inst);
-						load.prolog_loads = buffer.prolog_loads;
 					}
 				}
 				break;
@@ -462,22 +460,9 @@ EmbeddedFetchData DetectEmbeddedVertexFetch(const Decoder::Program&      decoded
 Decoder::Program DecodeFusedProgram(std::span<const uint32_t> front, std::span<const uint32_t> back,
                                     std::vector<uint32_t>& joined_code) {
 	EXIT_IF(back.empty());
-	Decoder::Program result;
-	uint32_t         front_words = 0;
-	while (front_words < front.size()) {
-		auto& inst = result.instructions.emplace_back();
-		Decoder::DecodeInstruction(front, front_words, inst);
-		front_words += inst.word_count;
-		if (inst.opcode == Decoder::Opcode::S_SETPC_B64) {
-			EXIT_NOT_IMPLEMENTED(inst.src0.kind != Decoder::OperandKind::Sgpr ||
-			                     inst.src0.reg != 6u);
-			break;
-		}
-		EXIT_NOT_IMPLEMENTED(inst.opcode == Decoder::Opcode::S_ENDPGM);
-	}
-	EXIT_IF(result.instructions.empty() ||
-	        result.instructions.back().opcode != Decoder::Opcode::S_SETPC_B64);
-	joined_code.assign(front.begin(), front.begin() + front_words);
+	auto       result      = Decoder::DecodeFrontProgram(front);
+	const auto front_words = static_cast<uint32_t>(result.code.size());
+	joined_code.assign(result.code.begin(), result.code.end());
 	joined_code.insert(joined_code.end(), back.begin(), back.end());
 	// The merged-stage ABI passes the back shader in s[6:7]. Give that handoff an
 	// ordinary CFG edge, retaining both bodies in one register and LDS lifetime.
@@ -505,7 +490,9 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 		EXIT("shader recompiler input is empty\n");
 	}
 	if (options.stage != ShaderType::Compute && options.stage != ShaderType::Vertex &&
-	    options.stage != ShaderType::Pixel && options.stage != ShaderType::Mesh) {
+	    options.stage != ShaderType::Pixel && options.stage != ShaderType::Mesh &&
+	    options.stage != ShaderType::Local && options.stage != ShaderType::TessellationControl &&
+	    options.stage != ShaderType::TessellationEvaluation) {
 		EXIT("shader recompiler received unsupported stage %u\n",
 		     static_cast<unsigned>(options.stage));
 	}
@@ -525,6 +512,12 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 	std::vector<uint32_t> joined_code;
 	if (!options.back_code.empty()) {
 		decoded = DecodeFusedProgram(code, options.back_code, joined_code);
+	} else if (options.stage == ShaderType::Local) {
+		decoded = Decoder::DecodeFrontProgram(code);
+		// The separately compiled hull half runs in the next Vulkan stage.
+		auto& handoff     = decoded.instructions.back();
+		handoff.opcode    = Decoder::Opcode::S_ENDPGM;
+		handoff.src_count = 0;
 	} else {
 		Decoder::DecodeProgram(code, decoded);
 	}
@@ -568,8 +561,8 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 	}
 
 	EmbeddedFetchData embedded_fetch;
-	if (options.stage == ShaderType::Vertex && options.input_info.vertex != nullptr &&
-	    options.input_info.vertex->fetch_embedded) {
+	if ((options.stage == ShaderType::Vertex || options.stage == ShaderType::Local) &&
+	    options.input_info.vertex != nullptr && options.input_info.vertex->fetch_embedded) {
 		embedded_fetch = DetectEmbeddedVertexFetch(
 		    decoded, options.input_info.vertex, options.user_data_base,
 		    static_cast<uint32_t>(options.user_data.size()), options.wave_size);
@@ -608,6 +601,7 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 		IR::RemoveIdentities(ir.blocks);
 		IR::EliminateDeadCode(ir.blocks);
 	}
+	LowerTessellationMemory(ir, options);
 	IR::BuildSrtPlan(ir);
 	IR::EliminateDeadCode(ir.blocks);
 	IR::TrackResources(ir);

src/graphics/shader/recompiler/Tessellation.cpp

@@ -0,0 +1,330 @@
+#include "graphics/shader/recompiler/Tessellation.h"
+
+#include "common/assert.h"
+#include "common/logging/log.h"
+#include "graphics/shader/recompiler/ShaderRecompiler.h"
+#include "graphics/shader/recompiler/frontend/decode/ShaderDecoder.h"
+#include "graphics/shader/recompiler/ir/ShaderIR.h"
+#include "graphics/shader/recompiler/ir/passes/ConstantPropagation.h"
+#include "graphics/shader/recompiler/ir/passes/DeadCodeElimination.h"
+
+#include <algorithm>
+#include <array>
+#include <unordered_map>
+#include <utility>
+
+namespace Libs::Graphics::ShaderRecompiler {
+namespace {
+
+struct TessellationAddress {
+	enum class Kind { Unknown, Affine, PackedControlPoint };
+	Kind     kind        = Kind::Unknown;
+	uint32_t coefficient = 0;
+	uint32_t constant    = 0;
+};
+
+uint32_t ReflectTessellationStride(const Decoder::Program& program, bool local,
+                                   uint32_t control_points, uint32_t input_stride) {
+	using namespace Decoder;
+	using Address = TessellationAddress;
+	using Kind    = Address::Kind;
+	std::array<Address, IR::NumVectorRegs> registers {};
+	registers[local ? 3u : 1u] =
+	    local ? Address {Kind::Affine, 1u, 0u} : Address {Kind::PackedControlPoint};
+	const auto constant = [](uint32_t value) { return Address {Kind::Affine, 0u, value}; };
+	const auto read     = [&](const Operand& operand) -> Address {
+		if (operand.absolute || operand.negate || operand.sdwa_sext || operand.op_sel ||
+		    operand.op_sel_hi || operand.negate_hi || operand.dpp)
+			return {};
+		Address value;
+		if (operand.kind == OperandKind::IntegerInlineConstant ||
+		    operand.kind == OperandKind::LiteralConstant) {
+			value = constant(operand.value);
+		} else if (operand.kind == OperandKind::Vgpr) {
+			value = registers.at(operand.reg);
+		}
+		// The native packed HS ID's low byte is the patch ordinal. The logical
+		// interface addresses one patch, as in the frontend's v1 initialization.
+		if (value.kind == Kind::PackedControlPoint && operand.sdwa_sel == 0u) return constant(0u);
+		return operand.sdwa_sel == 6u ? value : Address {};
+	};
+	const auto affine = [](uint64_t coefficient, uint64_t offset) {
+		return coefficient <= UINT32_MAX && offset <= UINT32_MAX
+		           ? Address {Kind::Affine, static_cast<uint32_t>(coefficient),
+		                      static_cast<uint32_t>(offset)}
+		           : Address {};
+	};
+	const auto add = [&](Address lhs, Address rhs) {
+		return lhs.kind == Kind::Affine && rhs.kind == Kind::Affine
+		           ? affine(uint64_t {lhs.coefficient} + rhs.coefficient,
+		                    uint64_t {lhs.constant} + rhs.constant)
+		           : Address {};
+	};
+	const auto multiply = [&](Address lhs, Address rhs) {
+		if ((lhs.kind == Kind::Affine && lhs.coefficient == 0u && lhs.constant == 0u) ||
+		    (rhs.kind == Kind::Affine && rhs.coefficient == 0u && rhs.constant == 0u))
+			return constant(0u);
+		if (lhs.kind != Kind::Affine || rhs.kind != Kind::Affine) return Address {};
+		if (rhs.coefficient != 0u) std::swap(lhs, rhs);
+		return rhs.coefficient == 0u ? affine(uint64_t {lhs.coefficient} * rhs.constant,
+		                                      uint64_t {lhs.constant} * rhs.constant)
+		                             : Address {};
+	};
+	const auto low24 = [&](Address value) {
+		if (value.kind == Kind::Affine && value.coefficient == 0u)
+			return constant(value.constant & 0xffffffu);
+		return value.kind == Kind::Affine &&
+		               uint64_t {value.coefficient} * (control_points - 1u) + value.constant <=
+		                   0xffffffu
+		           ? value
+		           : Address {};
+	};
+	uint32_t stride = 0;
+	for (const auto& inst: program.instructions) {
+		// Stage exits preserve the active path's definitions. An internal join
+		// would require merging register definitions.
+		EXIT_NOT_IMPLEMENTED(IsDirectBranch(inst.opcode) &&
+		                     inst.branch_target != program.instructions.back().pc);
+		const bool local_store =
+		    inst.opcode == Opcode::DS_WRITE_B32 || inst.opcode == Opcode::DS_WRITE2_B32;
+		const bool buffer_store = inst.opcode == Opcode::BUFFER_STORE_DWORD ||
+		                          inst.opcode == Opcode::BUFFER_STORE_DWORDX2 ||
+		                          inst.opcode == Opcode::BUFFER_STORE_DWORDX3 ||
+		                          inst.opcode == Opcode::BUFFER_STORE_DWORDX4;
+		const bool control_store =
+		    !local && buffer_store && inst.src2.kind == OperandKind::Sgpr && inst.src2.reg == 2u;
+		const bool control_read =
+		    !local && (inst.opcode == Opcode::DS_READ_B32 || inst.opcode == Opcode::DS_READ2_B32);
+		if ((local && local_store) || control_store || control_read) {
+			const auto address = read(inst.src0);
+			if (address.kind != Kind::Affine) {
+				EXIT("%s tessellation address is not affine at pc 0x%08x\n", local ? "LS" : "HS",
+				     inst.pc);
+			}
+			if (address.coefficient != 0u) {
+				const auto expected = control_read ? input_stride : stride;
+				EXIT_NOT_IMPLEMENTED((address.coefficient & 3u) != 0u ||
+				                     (expected != 0u && expected != address.coefficient));
+				if (!control_read) stride = address.coefficient;
+			}
+		}
+		// Store vdata is a source, despite occupying the decoder's dst field.
+		if (local_store || buffer_store || inst.dst.kind != OperandKind::Vgpr) continue;
+		const auto lhs   = read(inst.src0);
+		const auto rhs   = read(inst.src1);
+		const auto third = read(inst.src2);
+		Address    value;
+		switch (inst.opcode) {
+			case Opcode::V_BFE_U32:
+				if (lhs.kind == Kind::PackedControlPoint && rhs.kind == Kind::Affine &&
+				    rhs.coefficient == 0u && third.kind == Kind::Affine &&
+				    third.coefficient == 0u) {
+					if (rhs.constant == 8u && third.constant == 5u) value = {Kind::Affine, 1u, 0u};
+					if (rhs.constant == 0u && third.constant == 8u) value = constant(0u);
+				}
+				break;
+			case Opcode::V_MUL_U32_U24: value = multiply(low24(lhs), low24(rhs)); break;
+			case Opcode::V_MAD_U32_U24: value = add(multiply(low24(lhs), low24(rhs)), third); break;
+			case Opcode::V_LSHL_ADD_U32:
+				if (rhs.kind == Kind::Affine && rhs.coefficient == 0u && rhs.constant < 32u)
+					value = add(multiply(lhs, constant(1u << rhs.constant)), third);
+				break;
+			case Opcode::V_LSHLREV_B32:
+				if (lhs.kind == Kind::Affine && lhs.coefficient == 0u && lhs.constant < 32u)
+					value = multiply(rhs, constant(1u << lhs.constant));
+				break;
+			case Opcode::V_SUB_NC_U32:
+				if (lhs.kind == Kind::Affine && rhs.kind == Kind::Affine &&
+				    lhs.coefficient >= rhs.coefficient && lhs.constant >= rhs.constant)
+					value = affine(lhs.coefficient - rhs.coefficient, lhs.constant - rhs.constant);
+				break;
+			default: break;
+		}
+		if (inst.dst.sdwa_sel != 6u || inst.dst.op_sel || inst.dst.omod || inst.dst.clamp ||
+		    inst.dst.dpp)
+			value = {};
+		for (uint32_t index = 0;
+		     index < std::max(inst.data_dwords, 1u) && inst.dst.reg + index < registers.size(); index++) {
+			registers[inst.dst.reg + index] = {};
+		}
+		registers.at(inst.dst.reg) = value;
+	}
+	EXIT_NOT_IMPLEMENTED(stride == 0u);
+	return stride;
+}
+
+const IR::Inst* TessellationBufferBase(const IR::Inst& inst) {
+	if (IR::BufferAccessOf(inst.GetOpcode()) == IR::BufferAccess::None || inst.NumArgs() <= 3u)
+		return nullptr;
+	const auto* base = inst.Arg(3).Resolve().TryInstruction();
+	return base != nullptr && base->GetOpcode() == IR::ValueOpcode::TessellationBase ? base
+	                                                                                 : nullptr;
+}
+
+IR::Value ActiveAddress(IR::Block& block, IR::Block::iterator before, IR::Value value,
+                        IR::Value predicate, std::unordered_map<IR::Inst*, IR::Value>& resolved) {
+	using namespace IR;
+	value        = value.Resolve();
+	auto* source = value.TryInstruction();
+	if (source == nullptr) return value;
+	if (const auto found = resolved.find(source); found != resolved.end()) return found->second;
+	if (source->GetOpcode() == ValueOpcode::SelectU32) {
+		return source->Arg(0).Resolve() == predicate.Resolve()
+		           ? ActiveAddress(block, before, source->Arg(1), predicate, resolved)
+		           : value;
+	}
+	if (source->GetOpcode() != ValueOpcode::IAdd32 && source->GetOpcode() != ValueOpcode::ISub32)
+		return value;
+	const auto lhs = ActiveAddress(block, before, source->Arg(0), predicate, resolved);
+	const auto rhs = ActiveAddress(block, before, source->Arg(1), predicate, resolved);
+	if (lhs != source->Arg(0).Resolve() || rhs != source->Arg(1).Resolve()) {
+		auto copy = block.PrependNewInst(before, source->GetOpcode(), {lhs, rhs},
+		                                 source->Flags<uint64_t>());
+		value     = Value(&*copy);
+	}
+	resolved.emplace(source, value);
+	return value;
+}
+
+} // namespace
+
+void AnalyzeTessellationPrograms(std::span<const uint32_t> local, std::span<const uint32_t> control,
+                                 ShaderTessellationInputInfo& info) {
+	const auto       local_program = Decoder::DecodeFrontProgram(local);
+	Decoder::Program control_program;
+	Decoder::DecodeProgram(control, control_program);
+	info.ls_stride = ReflectTessellationStride(local_program, true, info.input_control_points, 0u);
+	info.hs_stride = ReflectTessellationStride(control_program, false, info.output_control_points,
+	                                           info.ls_stride);
+	LOGF("Tessellation interface: input_cp=%u output_cp=%u ls_stride=%u hs_stride=%u\n",
+	     info.input_control_points, info.output_control_points, info.ls_stride, info.hs_stride);
+}
+
+void LowerTessellationMemory(IR::Program& program, const CompileOptions& options) {
+	using namespace IR;
+	if (options.stage != ShaderType::Local && options.stage != ShaderType::TessellationControl &&
+	    options.stage != ShaderType::TessellationEvaluation) {
+		return;
+	}
+	const auto& tess = options.input_info.vertex->tess;
+	// Ring addresses can reuse data VGPRs. Their inactive values are irrelevant to
+	// a store guarded by the same EXEC predicate, but must remain intact elsewhere.
+	for (auto* block: program.blocks) {
+		for (auto it = block->begin(); it != block->end(); ++it) {
+			if (TessellationBufferBase(*it) == nullptr) continue;
+			std::unordered_map<Inst*, Value> resolved;
+			it->SetArg(
+			    2, ActiveAddress(*block, it, it->Arg(2), it->Arg(it->NumArgs() - 1u), resolved));
+		}
+	}
+	ConstantPropagationPass(program.blocks);
+	uint32_t reads = 0, writes = 0, factors = 0;
+	for (auto* block: program.blocks) {
+		for (auto it = block->begin(); it != block->end(); ++it) {
+			auto&      inst   = *it;
+			const auto shared = SharedAccessOf(inst.GetOpcode());
+			const auto buffer = BufferAccessOf(inst.GetOpcode());
+			if (shared == SharedAccess::None && buffer == BufferAccess::None) {
+				continue;
+			}
+			const auto&           memory = program.memory_info.at(inst.Flags<MemoryFlags>().index);
+			bool                  write  = false;
+			uint32_t              components = 0;
+			TessellationAttribute kind;
+			Value                 address, predicate;
+			if (shared != SharedAccess::None && memory.kind == ResourceKind::Lds) {
+				write = shared == SharedAccess::Write;
+				EXIT_NOT_IMPLEMENTED(memory.data_bits != 32u ||
+				                     (options.stage == ShaderType::Local
+				                          ? !write
+				                          : options.stage != ShaderType::TessellationControl ||
+				                                shared != SharedAccess::Read));
+				kind       = write ? TessellationAttribute::LocalOutput
+				                   : TessellationAttribute::ControlInput;
+				components = SharedComponentCount(inst.GetOpcode());
+				address    = inst.Arg(0);
+				predicate  = inst.Arg(inst.NumArgs() - 1u);
+			} else if (buffer != BufferAccess::None && memory.kind == ResourceKind::Buffer) {
+				const auto* base = TessellationBufferBase(inst);
+				if (base == nullptr) {
+					continue;
+				}
+				EXIT_NOT_IMPLEMENTED(memory.data_bits != 32u || memory.formatted || memory.idxen ||
+				                     !memory.offen || buffer == BufferAccess::Atomic);
+				write      = buffer == BufferAccess::Write;
+				components = BufferComponentCount(inst.GetOpcode());
+				address    = inst.Arg(2).Resolve();
+				predicate  = inst.Arg(inst.NumArgs() - 1u);
+				if (base->Arg(0).U32() == 1u) {
+					EXIT_NOT_IMPLEMENTED(!write ||
+					                     options.stage != ShaderType::TessellationControl);
+					kind = TessellationAttribute::Factor;
+					factors += components;
+				} else {
+					kind = write ? TessellationAttribute::ControlOutput
+					             : TessellationAttribute::EvaluationInput;
+					EXIT_NOT_IMPLEMENTED(write
+					                         ? options.stage != ShaderType::TessellationControl
+					                         : options.stage != ShaderType::TessellationEvaluation);
+					if (address.IsImmediate() &&
+					    address.U32() >= tess.hs_stride * tess.output_control_points) {
+						EXIT_NOT_IMPLEMENTED(!write);
+						kind = TessellationAttribute::PatchOutput;
+					}
+				}
+			} else {
+				continue;
+			}
+			const auto emit = [&](ValueOpcode opcode, std::initializer_list<Value> args) {
+				return Value(&*block->PrependNewInst(it, opcode, args));
+			};
+			std::array<Value, 4> values;
+			for (uint32_t component = 0; component < components; component++) {
+				const auto offset = memory.offset + 4u * component;
+				const auto byte_address =
+				    offset == 0u ? address : emit(ValueOpcode::IAdd32, {address, Value(offset)});
+				if (write) {
+					Value data = shared != SharedAccess::None ? inst.Arg(component + 1u)
+					             : components == 1u
+					                 ? inst.Arg(4)
+					                 : emit(components == 2u   ? ValueOpcode::CompositeExtractU32x2
+					                        : components == 3u ? ValueOpcode::CompositeExtractU32x3
+					                                           : ValueOpcode::CompositeExtractU32x4,
+					                        {inst.Arg(4), Value(component)});
+					emit(ValueOpcode::SetTessellationAttribute,
+					     {Value(static_cast<uint32_t>(kind)), byte_address, data, predicate});
+					writes++;
+				} else {
+					values[component] =
+					    emit(ValueOpcode::GetTessellationAttribute,
+					         {Value(static_cast<uint32_t>(kind)), byte_address, predicate});
+					reads++;
+				}
+			}
+			if (!write) {
+				Value replacement = values[0];
+				if (components == 2u)
+					replacement =
+					    emit(ValueOpcode::CompositeConstructU32x2, {values[0], values[1]});
+				if (components == 3u)
+					replacement = emit(ValueOpcode::CompositeConstructU32x3,
+					                   {values[0], values[1], values[2]});
+				if (components == 4u)
+					replacement = emit(ValueOpcode::CompositeConstructU32x4,
+					                   {values[0], values[1], values[2], values[3]});
+				inst.ReplaceUsesWith(replacement);
+			}
+			inst.Invalidate();
+		}
+	}
+	ConstantPropagationPass(program.blocks);
+	RemoveIdentities(program.blocks);
+	EliminateDeadCode(program.blocks);
+	LOGF("%s tessellation lowering: reads=%u writes=%u factors=%u\n",
+	     options.stage == ShaderType::Local                 ? "LS"
+	     : options.stage == ShaderType::TessellationControl ? "HS"
+	                                                       : "TES",
+	     reads, writes, factors);
+}
+
+} // namespace Libs::Graphics::ShaderRecompiler

src/graphics/shader/recompiler/Tessellation.h

@@ -0,0 +1,21 @@
+#pragma once
+
+#include <cstdint>
+#include <span>
+
+namespace Libs::Graphics {
+struct ShaderTessellationInputInfo;
+}
+
+namespace Libs::Graphics::ShaderRecompiler {
+
+struct CompileOptions;
+namespace IR {
+struct Program;
+}
+
+void AnalyzeTessellationPrograms(std::span<const uint32_t> local, std::span<const uint32_t> control,
+                                 ShaderTessellationInputInfo& info);
+void LowerTessellationMemory(IR::Program& program, const CompileOptions& options);
+
+} // namespace Libs::Graphics::ShaderRecompiler

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

@@ -277,7 +277,8 @@ Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program
 					break;
 				}
 				case IR::ValueOpcode::LaneId:
-					requirements.subgroup_local_invocation_id = true;
+					requirements.subgroup_local_invocation_id |=
+					    program.stage != ShaderType::TessellationControl;
 					break;
 				case IR::ValueOpcode::ImageQueryLod: requirements.compute_derivatives = true; break;
 				case IR::ValueOpcode::ImageGatherRaw:
@@ -306,8 +307,10 @@ std::vector<uint32_t> EmitProgram(const IR::Program& program,
 	using namespace Emitter;
 
 	if (program.stage != ShaderType::Compute && program.stage != ShaderType::Vertex &&
-	    program.stage != ShaderType::Pixel && program.stage != ShaderType::Mesh) {
-		Fail(program, "binary SPIR-V emitter supports compute, vertex, and pixel shaders");
+	    program.stage != ShaderType::Pixel && program.stage != ShaderType::Mesh &&
+	    program.stage != ShaderType::Local && program.stage != ShaderType::TessellationControl &&
+	    program.stage != ShaderType::TessellationEvaluation) {
+		Fail(program, "binary SPIR-V emitter received an unsupported shader stage");
 	}
 	if (!program.srt_plan_complete || !program.resource_tracking_complete ||
 	    !program.shader_info_complete || !program.binding_layout_complete) {

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

@@ -250,7 +250,10 @@ void EmitStorageImageWrite(EmitterState& state, uint32_t resource, uint32_t mip_
 
 spv::ExecutionModel ExecutionModelForStage(ShaderType stage) {
 	switch (stage) {
+		case ShaderType::Local:
 		case ShaderType::Vertex: return spv::ExecutionModelVertex;
+		case ShaderType::TessellationControl: return spv::ExecutionModelTessellationControl;
+		case ShaderType::TessellationEvaluation: return spv::ExecutionModelTessellationEvaluation;
 		case ShaderType::Mesh: return spv::ExecutionModelMeshEXT; // MeshEXT
 		case ShaderType::Pixel: return spv::ExecutionModelFragment;
 		default: return spv::ExecutionModelGLCompute;

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

@@ -1,6 +1,9 @@
 #include "graphics/shader/recompiler/backend/spirv/spirvEmitterInstructions.h"
 
+#include "common/logging/log.h"
+
 #include <algorithm>
+#include <atomic>
 
 namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter {
 namespace {
@@ -38,7 +41,9 @@ uint32_t EmitBuiltinU32(EmitterState& state, IR::StageInputKind kind, uint32_t c
 		return EmitAddU32(state, local,
 		                  EmitBinaryU32(state, spv::OpIMul, group, ConstantU32(state, size)));
 	}
-	const auto variable = InputVariableForKind(state, kind);
+	const bool centroid = kind == IR::StageInputKind::BaryCoordSmoothCentroid;
+	const auto variable = InputVariableForKind(
+	    state, centroid ? IR::StageInputKind::BaryCoordSmooth : kind);
 	if (variable == 0) {
 		return ConstantU32(state, 0);
 	}
@@ -46,11 +51,13 @@ uint32_t EmitBuiltinU32(EmitterState& state, IR::StageInputKind kind, uint32_t c
 		const auto value = state.builder.AllocateId();
 		const auto bits  = state.builder.AllocateId();
 		state.builder.AddFunction(spv::OpLoad, TypeBool(state), value, variable);
-		state.builder.AddFunction(spv::OpSelect, TypeU32(state), bits, value, ConstantU32(state, 1),
-		                          ConstantU32(state, 0));
+		// PS5 initializes v_front_face with float +1.0/-1.0 bits.
+		state.builder.AddFunction(spv::OpSelect, TypeU32(state), bits, value,
+		                          ConstantU32(state, 0x3f800000u), ConstantU32(state, 0xbf800000u));
 		return bits;
 	}
 	if (kind == IR::StageInputKind::VertexIndex || kind == IR::StageInputKind::InstanceIndex ||
+	    kind == IR::StageInputKind::InvocationId || kind == IR::StageInputKind::PrimitiveId ||
 	    kind == IR::StageInputKind::Layer || kind == IR::StageInputKind::SampleId) {
 		const auto value = state.builder.AllocateId();
 		const auto bits  = state.builder.AllocateId();
@@ -58,7 +65,7 @@ uint32_t EmitBuiltinU32(EmitterState& state, IR::StageInputKind kind, uint32_t c
 		state.builder.AddFunction(spv::OpBitcast, TypeU32(state), bits, value);
 		return bits;
 	}
-	if (kind == IR::StageInputKind::FragCoord) {
+	if (kind == IR::StageInputKind::FragCoord || kind == IR::StageInputKind::TessCoord) {
 		const auto pointer = state.builder.AllocateId();
 		const auto value   = state.builder.AllocateId();
 		const auto bits    = state.builder.AllocateId();
@@ -69,15 +76,24 @@ uint32_t EmitBuiltinU32(EmitterState& state, IR::StageInputKind kind, uint32_t c
 		state.builder.AddFunction(spv::OpBitcast, TypeU32(state), bits, value);
 		return bits;
 	}
-	if (kind == IR::StageInputKind::BaryCoordSmooth ||
+	if (centroid || kind == IR::StageInputKind::BaryCoordSmooth ||
 	    kind == IR::StageInputKind::BaryCoordNoPerspective) {
-		const auto pointer = state.builder.AllocateId();
 		const auto value   = state.builder.AllocateId();
 		const auto bits    = state.builder.AllocateId();
-		state.builder.AddFunction(spv::OpAccessChain,
-		                          TypePointer(state, spv::StorageClassInput, TypeF32(state)),
-		                          pointer, variable, ConstantU32(state, component + 1u));
-		state.builder.AddFunction(spv::OpLoad, TypeF32(state), value, pointer);
+		if (centroid) {
+			const auto coordinates = state.builder.AllocateId();
+			state.builder.RequireCapability(spv::CapabilityInterpolationFunction);
+			state.builder.AddFunction(spv::OpExtInst, TypeF32Vector(state, 3), coordinates,
+			                          GlslStd450(state), GLSLstd450InterpolateAtCentroid, variable);
+			state.builder.AddFunction(spv::OpCompositeExtract, TypeF32(state), value,
+			                          coordinates, component + 1u);
+		} else {
+			const auto pointer = state.builder.AllocateId();
+			state.builder.AddFunction(spv::OpAccessChain,
+			                          TypePointer(state, spv::StorageClassInput, TypeF32(state)),
+			                          pointer, variable, ConstantU32(state, component + 1u));
+			state.builder.AddFunction(spv::OpLoad, TypeF32(state), value, pointer);
+		}
 		state.builder.AddFunction(spv::OpBitcast, TypeU32(state), bits, value);
 		return bits;
 	}
@@ -138,7 +154,7 @@ uint32_t EmitAttribute(EmitterState& state, uint32_t attr, uint32_t chan) {
 	if (input == nullptr || input->variable_id == 0) {
 		return ConstantU32(state, 0);
 	}
-	if (state.program.stage == ShaderType::Vertex) {
+	if (state.program.stage == ShaderType::Vertex || state.program.stage == ShaderType::Local) {
 		return EmitVertexParameterComponentU32(state, *input, chan & 3u);
 	}
 	const auto load_per_vertex = [&](uint32_t vertex) {
@@ -498,6 +514,31 @@ void EmitSetAttribute(ValueEmitContext& ctx, const IR::Inst& inst) {
 			state.builder.AddFunction(spv::OpStore, MeshOutputPointer(state, kind, exp.index),
 			                          value);
 		} else if (exp.kind == IR::ExportTargetKind::Position) {
+			if (state.invalid_position_clip_distance != UINT32_MAX) {
+				const auto zero = state.builder.Constant(spv::OpConstantNull, TypeF32Vector(state, 4));
+				const auto equal = state.builder.AllocateId();
+				const auto invalid = state.builder.AllocateId();
+				const auto distance = state.builder.AllocateId();
+				const auto distance_pointer = state.builder.AllocateId();
+				state.builder.AddFunction(spv::OpFOrdEqual, TypeBoolVector(state, 4), equal,
+				                          value, zero);
+				state.builder.AddFunction(spv::OpAll, TypeBool(state), invalid, equal);
+				// Zero at valid vertices makes a primitive containing an invalid position
+				// collapse to its remaining edge, before the undefined 0/0 perspective divide.
+				state.builder.AddFunction(spv::OpSelect, TypeF32(state), distance, invalid,
+				                          ConstantF32Value(state, -1.0f),
+				                          ConstantF32Value(state, 0.0f));
+				state.builder.AddFunction(
+				    spv::OpAccessChain, TypePointer(state, spv::StorageClassOutput, TypeF32(state)),
+				    distance_pointer, state.clip_distance_variable,
+				    ConstantU32(state, state.invalid_position_clip_distance));
+				state.builder.AddFunction(spv::OpStore, distance_pointer, distance);
+				static std::atomic_bool logged = false;
+				if (!logged.exchange(true, std::memory_order_relaxed)) {
+					Log::WriteToConsoleAndLog(
+					    "Shader: emitted zero-position clip guard\n");
+				}
+			}
 			const auto pointer = state.builder.AllocateId();
 			state.builder.AddFunction(
 			    spv::OpAccessChain,
@@ -517,11 +558,19 @@ uint32_t EmitIdentity(ValueEmitContext&, uint32_t value) {
 void EmitVoid(ValueEmitContext&) {}
 
 void EmitBarrier(EmitterState& state) {
-	const auto semantics =
-	    spv::MemorySemanticsAcquireReleaseMask | spv::MemorySemanticsWorkgroupMemoryMask;
+	const auto tessellation = state.program.stage == ShaderType::TessellationControl;
+	const auto memory_scope = tessellation ? spv::ScopeInvocation : spv::ScopeWorkgroup;
+	const auto semantics    = tessellation ? spv::MemorySemanticsMaskNone
+	                                       : spv::MemorySemanticsAcquireReleaseMask |
+	                                             spv::MemorySemanticsWorkgroupMemoryMask;
 	state.builder.AddFunction(spv::OpControlBarrier, ConstantU32(state, spv::ScopeWorkgroup),
-	                          ConstantU32(state, spv::ScopeWorkgroup),
-	                          ConstantU32(state, semantics));
+	                          ConstantU32(state, memory_scope), ConstantU32(state, semantics));
+}
+
+uint32_t EmitLaneId(EmitterState& state) {
+	return state.program.stage == ShaderType::TessellationControl
+	           ? EmitBuiltinU32(state, IR::StageInputKind::InvocationId, 0)
+	           : EmitSubgroupLocalInvocationId(state);
 }
 
 uint32_t EmitMeshDrawParameter(ValueEmitContext& ctx, const IR::Inst& inst) {

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

@@ -3,7 +3,9 @@
 #include "graphics/shader/recompiler/frontend/decode/ImageOps.h"
 
 #include <algorithm>
+#include <atomic>
 #include <bit>
+#include <cstdio>
 
 namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter {
 namespace {
@@ -646,6 +648,13 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 		const auto coord =
 		    CoordF32(ctx, mem, *address, layout.coord, dimension_info.coordinate_components);
 		if (op == IR::ValueOpcode::ImageGatherRaw) {
+			if (HasFlag(mem, Decoder::ImageSampleFlagLod)) {
+				static std::atomic_flag warned = ATOMIC_FLAG_INIT;
+				if (!warned.test_and_set(std::memory_order_relaxed)) {
+					std::fputs("Warning: approximating IMAGE_GATHER4_L at mip level 0; explicit LOD is ignored.\n",
+					           stderr);
+				}
+			}
 			if (dimension == ImageDimension::Dim1D) {
 				if (dref || !HasFlag(mem, Decoder::ImageSampleFlagLevelZero) ||
 				    HasFlag(mem, Decoder::ImageSampleFlagOffset) ||

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

@@ -155,6 +155,8 @@ inline constexpr auto EmitInstPrefetch = EmitVoid;
 void                  EmitBarrier(EmitterState& state);
 void                  EmitMeshAllocate(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitMeshDrawParameter(ValueEmitContext& ctx, const IR::Inst& inst);
+uint32_t              EmitGetTessellationAttribute(ValueEmitContext& ctx, const IR::Inst& inst);
+void                  EmitSetTessellationAttribute(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitGetUserData(EmitterState& state, IR::ScalarReg reg);
 uint32_t              EmitGetBuiltin(ValueEmitContext& ctx, IR::Value kind, IR::Value index);
 uint32_t              EmitUndefU1(EmitterState& state, const IR::Inst& inst);
@@ -165,7 +167,7 @@ inline constexpr auto EmitUndefU64 = EmitUndefU1;
 uint32_t              EmitDppMoveU32(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitDppUpdateU32(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitWqmU64(EmitterState& state, uint32_t value);
-inline constexpr auto EmitLaneId = EmitSubgroupLocalInvocationId;
+uint32_t              EmitLaneId(EmitterState& state);
 uint32_t              EmitBallot(ValueEmitContext& ctx, IR::Value predicate);
 uint32_t              EmitReadFirstLane(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitReadLane(ValueEmitContext& ctx, const IR::Inst& inst);
@@ -271,6 +273,7 @@ inline constexpr auto EmitImageAtomicOr32      = EmitImage;
 inline constexpr auto EmitImageAtomicXor32     = EmitImage;
 void                  EmitUnreachable(ValueEmitContext& ctx, const IR::Inst& inst);
 inline constexpr auto EmitPhi                        = EmitUnreachable;
+inline constexpr auto EmitTessellationBase           = EmitUnreachable;
 inline constexpr auto EmitGetThreadBitScalarRegister = EmitUnreachable;
 inline constexpr auto EmitSetThreadBitScalarRegister = EmitUnreachable;
 inline constexpr auto EmitGetScalarMaskTag           = EmitUnreachable;

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

@@ -80,6 +80,10 @@ struct EmitterState {
 	Builder                                          builder;
 	const IR::Program&                               program;
 	ShaderStageInputInfo                             input_info;
+	std::array<uint32_t, 6>                          tess_variables {};
+	uint32_t                                         tess_inner_variable = 0;
+	uint32_t                                         tess_patch_base     = 0;
+
 	const SpirvRequirements                          requirements;
 	uint32_t                                         lane_count              = 1;
 	uint32_t                                         lane_half               = 0;
@@ -112,6 +116,7 @@ struct EmitterState {
 	uint32_t                   per_vertex_variable                   = 0;
 	uint32_t                   point_size_variable                   = 0;
 	uint32_t                   clip_distance_variable                = 0;
+	uint32_t                   invalid_position_clip_distance        = UINT32_MAX;
 	uint32_t                   cull_distance_variable                = 0;
 	uint32_t                   layer_variable                        = 0;
 	uint32_t                   viewport_index_variable               = 0;
@@ -313,7 +318,11 @@ uint32_t ConstantU64(EmitterState& state, uint64_t value);
 
 uint32_t ConstantU32CompositeZero(EmitterState& state, uint32_t components);
 
+uint32_t DefineInterfaceVariable(EmitterState& state, uint32_t type, spv::StorageClass storage,
+                                 const char* name);
 void     DefineModule(EmitterState& state);
+void     DefineTessellationInterfaces(EmitterState& state);
+void     DefineTessellationExecutionModes(EmitterState& state);
 void     DefineMeshOutputs(EmitterState& state);
 void     EmitMeshEntryPoint(EmitterState& state);
 void     EmitMeshAllocate(ValueEmitContext& ctx, const IR::Inst& inst);

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

@@ -290,7 +290,8 @@ uint32_t GlslStd450(EmitterState& state) {
 }
 
 VertexInputScalarKind VertexParameterScalarKind(const EmitterState& state, uint32_t location) {
-	if (state.program.stage != ShaderType::Vertex || location >= ShaderVertexInputInfo::RES_MAX ||
+	if ((state.program.stage != ShaderType::Vertex && state.program.stage != ShaderType::Local) ||
+	    location >= ShaderVertexInputInfo::RES_MAX ||
 	    location >= static_cast<uint32_t>(state.input_info.vertex->resources_num)) {
 		return VertexInputScalarKind::Float;
 	}
@@ -333,8 +334,6 @@ uint32_t VertexParameterScalarType(EmitterState& state, VertexInputScalarKind ki
 	}
 }
 
-namespace {
-
 uint32_t DefineInterfaceVariable(EmitterState& state, uint32_t type, spv::StorageClass storage,
                                  const char* name) {
 	const auto variable =
@@ -344,9 +343,14 @@ uint32_t DefineInterfaceVariable(EmitterState& state, uint32_t type, spv::Storag
 	return variable;
 }
 
+namespace {
+
 uint32_t BuiltInForInput(IR::StageInputKind kind) {
 	switch (kind) {
 		case IR::StageInputKind::VertexIndex: return spv::BuiltInVertexIndex;
+		case IR::StageInputKind::InvocationId: return spv::BuiltInInvocationId;
+		case IR::StageInputKind::PrimitiveId: return spv::BuiltInPrimitiveId;
+		case IR::StageInputKind::TessCoord: return spv::BuiltInTessCoord;
 		case IR::StageInputKind::InstanceIndex: return spv::BuiltInInstanceIndex;
 		case IR::StageInputKind::FragCoord: return spv::BuiltInFragCoord;
 		case IR::StageInputKind::FrontFacing: return spv::BuiltInFrontFacing;
@@ -384,9 +388,24 @@ void DefineInputs(EmitterState& state) {
 		}
 	}
 	for (auto& input: state.inputs) {
+		if (state.program.stage == ShaderType::Pixel &&
+		    input.kind == IR::StageInputKind::Parameter) {
+			const auto location = PixelParameterLocation(state, input.location);
+			const auto alias = std::ranges::find_if(state.inputs, [&](const InputBinding& other) {
+				return other.kind == IR::StageInputKind::Parameter && other.variable_id != 0 &&
+				       PixelParameterLocation(state, other.location) == location;
+			});
+			if (alias != state.inputs.end()) {
+				EXIT_IF(alias->per_vertex != input.per_vertex);
+				input.variable_id = alias->variable_id;
+				continue;
+			}
+		}
 		uint32_t type = TypeU32(state);
 		switch (input.kind) {
 			case IR::StageInputKind::VertexIndex:
+			case IR::StageInputKind::InvocationId:
+			case IR::StageInputKind::PrimitiveId:
 			case IR::StageInputKind::InstanceIndex:
 			case IR::StageInputKind::Layer:
 			case IR::StageInputKind::SampleId: type = TypeI32(state); break;
@@ -394,11 +413,13 @@ void DefineInputs(EmitterState& state) {
 			case IR::StageInputKind::LocalInvocationId:
 			case IR::StageInputKind::GlobalInvocationId: type = TypeU32Vector(state, 3); break;
 			case IR::StageInputKind::FragCoord: type = TypeF32Vector(state, 4); break;
+			case IR::StageInputKind::TessCoord:
 			case IR::StageInputKind::BaryCoordSmooth:
 			case IR::StageInputKind::BaryCoordNoPerspective: type = TypeF32Vector(state, 3); break;
 			case IR::StageInputKind::FrontFacing: type = TypeBool(state); break;
 			case IR::StageInputKind::Parameter:
-				if (state.program.stage == ShaderType::Vertex) {
+				if (state.program.stage == ShaderType::Vertex ||
+				    state.program.stage == ShaderType::Local) {
 					type = VertexParameterScalarType(
 					    state, VertexParameterScalarKind(state, input.location));
 					const auto components = VertexParameterComponentCount(input);
@@ -468,6 +489,15 @@ void DefineOutputs(EmitterState& state) {
 		DefineMeshOutputs(state);
 		return;
 	}
+	if (state.program.stage == ShaderType::Vertex && clip_distance_count + cull_distance_count < 8u &&
+	    std::ranges::any_of(state.outputs, [](const OutputBinding& output) {
+		    return output.kind == IR::StageOutputKind::Position;
+	    })) {
+		// Reserve one plane for the enabled PA_CL_CLIP_CNTL clipping-error cull.
+		state.invalid_position_clip_distance = clip_distance_count++;
+		state.outputs.push_back({{IR::StageOutputKind::ClipDistance,
+		                          state.invalid_position_clip_distance, 0, "gl_ClipDistance"}});
+	}
 	const auto BuiltIn = [&](uint32_t& variable, uint32_t type, const char* name,
 	                         spv::BuiltIn builtin) {
 		if (variable == 0) {
@@ -543,6 +573,7 @@ void DefineModule(EmitterState& state) {
 	                                  state.program.info.outputs.size());
 	DefineInputs(state);
 	DefineOutputs(state);
+	DefineTessellationInterfaces(state);
 	DefineDescriptors(state);
 	if (state.requirements.function_lds) {
 		state.lds_variable = state.builder.AllocateId();
@@ -564,6 +595,10 @@ void DefineModule(EmitterState& state) {
 		state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeOutputPrimitivesEXT,
 		                               state.input_info.vertex->mesh.max_primitives);
 	}
+	if (state.program.stage == ShaderType::TessellationControl ||
+	    state.program.stage == ShaderType::TessellationEvaluation) {
+		DefineTessellationExecutionModes(state);
+	}
 	state.entry_label = state.builder.AllocateId();
 
 	state.builder.RequireCapability(spv::CapabilityShader);

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

@@ -0,0 +1,161 @@
+#include "graphics/shader/recompiler/backend/spirv/spirvEmitterInstructions.h"
+
+#include <algorithm>
+
+namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter {
+namespace {
+
+uint32_t TessellationPointer(ValueEmitContext& ctx, const IR::Inst& inst) {
+	auto& state          = ctx.state;
+	using Attribute      = IR::TessellationAttribute;
+	const auto  kind     = static_cast<Attribute>(inst.Arg(0).U32());
+	const auto& tess     = state.input_info.vertex->tess;
+	const auto  variable = state.tess_variables.at(static_cast<uint32_t>(kind));
+	EXIT_IF(variable == 0);
+	const auto input   = kind == Attribute::ControlInput || kind == Attribute::EvaluationInput;
+	const auto storage = input ? spv::StorageClassInput : spv::StorageClassOutput;
+	const auto pointer = state.builder.AllocateId();
+	if (kind == Attribute::Factor) {
+		EXIT_NOT_IMPLEMENTED(!inst.Arg(1).IsImmediate());
+		const auto index = inst.Arg(1).U32() / 4u;
+		const auto outer = index < 3u;
+		EXIT_NOT_IMPLEMENTED(index >= 4u);
+		state.builder.AddFunction(spv::OpAccessChain, TypePointer(state, storage, TypeF32(state)),
+		                          pointer, outer ? variable : state.tess_inner_variable,
+		                          ConstantU32(state, outer ? index : index - 3u));
+		return pointer;
+	}
+	auto address = ctx.Arg(inst, 1);
+	if (kind == Attribute::PatchOutput) {
+		address =
+		    EmitBinaryU32(state, spv::OpISub, address, ConstantU32(state, state.tess_patch_base));
+	}
+	const bool local  = kind == Attribute::LocalOutput || kind == Attribute::ControlInput;
+	const auto stride = local ? tess.ls_stride : tess.hs_stride;
+	const auto offset = kind == Attribute::PatchOutput ? address
+	                                                   : EmitBinaryU32(state, spv::OpUMod, address,
+	                                                                   ConstantU32(state, stride));
+	const auto attribute =
+	    EmitBinaryU32(state, spv::OpShiftRightLogical, offset, ConstantU32(state, 4));
+	const auto component =
+	    EmitBinaryU32(state, spv::OpBitwiseAnd,
+	                  EmitBinaryU32(state, spv::OpShiftRightLogical, offset, ConstantU32(state, 2)),
+	                  ConstantU32(state, 3));
+	const auto type = TypePointer(state, storage, TypeU32(state));
+	if (kind == Attribute::LocalOutput || kind == Attribute::PatchOutput) {
+		state.builder.AddFunction(spv::OpAccessChain, type, pointer, variable, attribute,
+		                          component);
+	} else {
+		const auto vertex =
+		    kind == Attribute::ControlOutput
+		        ? EmitLaneId(state)
+		        : EmitBinaryU32(state, spv::OpUDiv, address, ConstantU32(state, stride));
+		state.builder.AddFunction(spv::OpAccessChain, type, pointer, variable, vertex, attribute,
+		                          component);
+	}
+	return pointer;
+}
+
+} // namespace
+
+void DefineTessellationInterfaces(EmitterState& state) {
+	using Attribute = IR::TessellationAttribute;
+	std::array<bool, 6> used {};
+	uint32_t            patch_begin = UINT32_MAX, patch_end = 0;
+	for (const auto* block: state.program.blocks) {
+		for (const auto& inst: *block) {
+			if (inst.GetOpcode() != IR::ValueOpcode::GetTessellationAttribute &&
+			    inst.GetOpcode() != IR::ValueOpcode::SetTessellationAttribute)
+				continue;
+			const auto kind = inst.Arg(0).U32();
+			used.at(kind)   = true;
+			if (kind == static_cast<uint32_t>(Attribute::PatchOutput)) {
+				EXIT_NOT_IMPLEMENTED(!inst.Arg(1).IsImmediate());
+				patch_begin = std::min(patch_begin, inst.Arg(1).U32());
+				patch_end   = std::max(patch_end, inst.Arg(1).U32() + 4u);
+			}
+		}
+	}
+	if (std::ranges::none_of(used, [](bool value) { return value; })) return;
+	const auto& tess  = state.input_info.vertex->tess;
+	const auto  array = [&](uint32_t type, uint32_t count) {
+		return state.builder.Type(spv::OpTypeArray, type, ConstantU32(state, count));
+	};
+	for (uint32_t index = 0; index < used.size(); index++) {
+		if (!used[index]) continue;
+		const auto kind    = static_cast<Attribute>(index);
+		const bool input   = kind == Attribute::ControlInput || kind == Attribute::EvaluationInput;
+		const auto storage = input ? spv::StorageClassInput : spv::StorageClassOutput;
+		uint32_t   type;
+		if (kind == Attribute::Factor) {
+			type = array(TypeF32(state), 4u);
+		} else if (kind == Attribute::PatchOutput) {
+			state.tess_patch_base = patch_begin;
+			type = array(TypeU32Vector(state, 4), (patch_end - patch_begin + 15u) / 16u);
+		} else {
+			const bool local  = kind == Attribute::LocalOutput || kind == Attribute::ControlInput;
+			const auto stride = local ? tess.ls_stride : tess.hs_stride;
+			type              = array(TypeU32Vector(state, 4), (stride + 15u) / 16u);
+			if (kind != Attribute::LocalOutput) {
+				type = array(type, local ? tess.input_control_points : tess.output_control_points);
+			}
+		}
+		auto& variable = state.tess_variables[index];
+		variable       = DefineInterfaceVariable(state, type, storage, "tess_attributes");
+		if (kind == Attribute::Factor) {
+			state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationBuiltIn,
+			                            spv::BuiltInTessLevelOuter);
+			state.tess_inner_variable =
+			    DefineInterfaceVariable(state, array(TypeF32(state), 2u), storage, "tess_inner");
+			state.builder.AddAnnotation(spv::OpDecorate, state.tess_inner_variable,
+			                            spv::DecorationBuiltIn, spv::BuiltInTessLevelInner);
+			state.builder.AddAnnotation(spv::OpDecorate, state.tess_inner_variable,
+			                            spv::DecorationPatch);
+		} else {
+			state.builder.AddAnnotation(
+			    spv::OpDecorate, variable, spv::DecorationLocation,
+			    kind == Attribute::PatchOutput ? (tess.hs_stride + 15u) / 16u : 0u);
+		}
+		if (kind == Attribute::Factor || kind == Attribute::PatchOutput) {
+			state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationPatch);
+		}
+	}
+}
+
+void DefineTessellationExecutionModes(EmitterState& state) {
+	const auto& tess = state.input_info.vertex->tess;
+	EXIT_NOT_IMPLEMENTED(tess.domain != 1u || tess.partitioning != 2u ||
+	                     tess.output_topology != 2u);
+	state.builder.RequireCapability(spv::CapabilityTessellation);
+	if (state.program.stage == ShaderType::TessellationControl) {
+		state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeOutputVertices,
+		                               tess.output_control_points);
+	} else {
+		state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeTriangles);
+		state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeSpacingFractionalOdd);
+		state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeVertexOrderCw);
+	}
+}
+
+uint32_t EmitGetTessellationAttribute(ValueEmitContext& ctx, const IR::Inst& inst) {
+	return EmitValueOrZeroIfCondition(ctx.state, ctx.Arg(inst, 2), [&] {
+		const auto value = ctx.state.builder.AllocateId();
+		ctx.state.builder.AddFunction(spv::OpLoad, TypeU32(ctx.state), value,
+		                              TessellationPointer(ctx, inst));
+		return value;
+	});
+}
+
+void EmitSetTessellationAttribute(ValueEmitContext& ctx, const IR::Inst& inst) {
+	EmitIfCondition(ctx.state, ctx.Arg(inst, 3), [&] {
+		auto value = ctx.Arg(inst, 2);
+		if (inst.Arg(0).U32() == static_cast<uint32_t>(IR::TessellationAttribute::Factor)) {
+			const auto floating = ctx.state.builder.AllocateId();
+			ctx.state.builder.AddFunction(spv::OpBitcast, TypeF32(ctx.state), floating, value);
+			value = floating;
+		}
+		ctx.state.builder.AddFunction(spv::OpStore, TessellationPointer(ctx, inst), value);
+	});
+}
+
+} // namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter

src/graphics/shader/recompiler/frontend/cfg/ShaderCFG.cpp

@@ -1206,13 +1206,12 @@ uint32_t FindSelectionMerge(const Graph& graph, const BasicBlock& block) {
 					return false_target;
 				}
 				// A return can leave a selection without reaching its merge. Keep the
-				// continuing arm as the merge instead of joining live state with a return.
-				if (graph.Dominates(block.id, false_target) &&
-				    HasLinearPathToTerminal(graph, true_target)) {
+				// continuing arm as the merge. Shared continuations receive a dedicated
+				// gateway in SplitSharedMergeBlock before structured control is emitted.
+				if (HasLinearPathToTerminal(graph, true_target)) {
 					return false_target;
 				}
-				if (graph.Dominates(block.id, true_target) &&
-				    HasLinearPathToTerminal(graph, false_target)) {
+				if (HasLinearPathToTerminal(graph, false_target)) {
 					return true_target;
 				}
 			}

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

@@ -175,6 +175,7 @@ constexpr MimgSampleInfo MIMG_SAMPLE_OPCODE_LIST[] = {
 };
 
 constexpr MimgGatherInfo MIMG_GATHER_OPCODE_LIST[] = {
+    {0x44u, Opcode::IMAGE_GATHER4_L, ImageSampleFlagLod},
     {0x47u, Opcode::IMAGE_GATHER4_LZ, ImageSampleFlagLevelZero},
     {0x48u, Opcode::IMAGE_GATHER4_C, ImageSampleFlagCompare},
     {0x4fu, Opcode::IMAGE_GATHER4_C_LZ,

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

@@ -24,7 +24,8 @@ constexpr OpcodeMap SOP2_OPCODE_LIST[] = {
     {0x1cu, Opcode::S_XNOR_B32},        {0x1du, Opcode::S_XNOR_B64},
     {0x1eu, Opcode::S_LSHL_B32},        {0x1fu, Opcode::S_LSHL_B64},
     {0x20u, Opcode::S_LSHR_B32},        {0x21u, Opcode::S_LSHR_B64},
-    {0x22u, Opcode::S_ASHR_I32},        {0x24u, Opcode::S_BFM_B32},
+    {0x22u, Opcode::S_ASHR_I32},        {0x23u, Opcode::S_ASHR_I64},
+    {0x24u, Opcode::S_BFM_B32},
     {0x25u, Opcode::S_BFM_B64},         {0x26u, Opcode::S_MUL_I32},
     {0x27u, Opcode::S_BFE_U32},         {0x28u, Opcode::S_BFE_I32},
     {0x29u, Opcode::S_BFE_U64},         {0x2cu, Opcode::S_ABSDIFF_I32},
@@ -105,6 +106,7 @@ constexpr OpcodeMap SOPP_OPCODE_LIST[] = {
     {0x10u, Opcode::S_SENDMSG},
     {0x12u, Opcode::S_TRAP},
     {0x16u, Opcode::S_TTRACEDATA},
+    {0x17u, Opcode::S_CBRANCH_CDBGSYS},
     {0x20u, Opcode::S_INST_PREFETCH},
     {0x23u, Opcode::S_WAITCNT_DEPCTR},
 };
@@ -136,6 +138,10 @@ void DecodeSop1(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	inst.family    = Family::SOP1;
 	inst.opcode_id = opcode;
 	inst.opcode    = Detail::LookupOpcode(SOP1_OPS, opcode);
+	if (opcode == 0x21u && sdst == 125u) {
+		// S_SWAPPC_B64 with NULL discards the return PC, so it is a plain jump.
+		inst.opcode = Opcode::S_SETPC_B64;
+	}
 	SetRawWords(inst, code, word_index, 1);
 
 	if (inst.opcode == Opcode::UNSUPPORTED) {

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

@@ -144,6 +144,7 @@ std::string FormatMimg(const Instruction& inst) {
 	}
 	switch (inst.opcode) {
 		case Opcode::IMAGE_SAMPLE:
+		case Opcode::IMAGE_GATHER4_L:
 		case Opcode::IMAGE_GATHER4_LZ:
 		case Opcode::IMAGE_GATHER4_C:
 		case Opcode::IMAGE_GATHER4_C_LZ:
@@ -187,6 +188,8 @@ std::string FormatExp(const Instruction& inst) {
 
 bool IsConditionalBranch(Opcode opcode) {
 	switch (opcode) {
+		// DevKit NGG should be disabled.
+		case Opcode::S_CBRANCH_CDBGSYS: return false;
 		case Opcode::S_CBRANCH_SCC0:
 		case Opcode::S_CBRANCH_SCC1:
 		case Opcode::S_CBRANCH_VCCZ:
@@ -381,6 +384,25 @@ void DecodeInstruction(std::span<const uint32_t> code, uint32_t word_index, Inst
 	}
 }
 
+Program DecodeFrontProgram(std::span<const uint32_t> front) {
+	Program  result;
+	uint32_t front_words = 0;
+	while (front_words < front.size()) {
+		auto& inst = result.instructions.emplace_back();
+		DecodeInstruction(front, front_words, inst);
+		front_words += inst.word_count;
+		if (inst.opcode == Opcode::S_SETPC_B64) {
+			EXIT_NOT_IMPLEMENTED(inst.src0.kind != OperandKind::Sgpr || inst.src0.reg != 6u);
+			break;
+		}
+		EXIT_NOT_IMPLEMENTED(inst.opcode == Opcode::S_ENDPGM);
+	}
+	EXIT_IF(result.instructions.empty() ||
+	        result.instructions.back().opcode != Opcode::S_SETPC_B64);
+	result.code = front.first(front_words);
+	return result;
+}
+
 void DecodeProgram(std::span<const uint32_t> code, Program& program) {
 	program.instructions.clear();
 	program.instructions.reserve(code.size());
@@ -535,6 +557,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::S_CBRANCH_VCCNZ:
 		case Opcode::S_CBRANCH_EXECZ:
 		case Opcode::S_CBRANCH_EXECNZ:
+		case Opcode::S_CBRANCH_CDBGSYS:
 			return WithUnsupportedReason(inst, fmt::format("0x{:08x}: {} 0x{:08x}", inst.pc,
 			                                               magic_enum::enum_name(inst.opcode),
 			                                               inst.branch_target));
@@ -558,6 +581,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::IMAGE_LOAD_MIP:
 		case Opcode::IMAGE_GET_RESINFO:
 		case Opcode::IMAGE_GET_LOD:
+		case Opcode::IMAGE_GATHER4_L:
 		case Opcode::IMAGE_GATHER4_LZ:
 		case Opcode::IMAGE_GATHER4_C:
 		case Opcode::IMAGE_GATHER4_C_LZ:

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

@@ -106,6 +106,7 @@ enum class Opcode {
 	S_LSHR_B32,
 	S_LSHR_B64,
 	S_ASHR_I32,
+	S_ASHR_I64,
 	S_MUL_I32,
 	S_MUL_HI_U32,
 	S_MUL_HI_I32,
@@ -559,6 +560,7 @@ enum class Opcode {
 	IMAGE_ATOMIC_OR,
 	IMAGE_ATOMIC_XOR,
 	IMAGE_SAMPLE,
+	IMAGE_GATHER4_L,
 	IMAGE_GATHER4_LZ,
 	IMAGE_GATHER4_C,
 	IMAGE_GATHER4_C_LZ,
@@ -581,6 +583,7 @@ enum class Opcode {
 	S_CBRANCH_VCCNZ,
 	S_CBRANCH_EXECZ,
 	S_CBRANCH_EXECNZ,
+	S_CBRANCH_CDBGSYS,
 	S_SENDMSG,
 	S_SETREG_B32,
 	S_SLEEP,
@@ -722,6 +725,7 @@ struct Program {
 Family GetInstructionFamily(uint32_t word);
 // The output object must be freshly initialized.
 void DecodeInstruction(std::span<const uint32_t> code, uint32_t word_index, Instruction& inst);
+Program DecodeFrontProgram(std::span<const uint32_t> front);
 void DecodeProgram(std::span<const uint32_t> code, Program& program);
 bool IsConditionalBranch(Opcode opcode);
 bool IsDirectBranch(Opcode opcode);

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

@@ -516,7 +516,7 @@ struct Vop1SdwaRule {
 };
 
 constexpr Vop1SdwaRule VOP1_SDWA_RULES[] = {
-    {Opcode::V_MOV_B32, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), SdwaSelWords(),
+    {Opcode::V_MOV_B32, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), SdwaSelBytes() | SdwaSelWords(),
      SdwaSelWords() | SdwaSelFull(), false},
     {Opcode::V_CVT_F32_U32, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), 0, 0, false},
     {Opcode::V_CVT_F32_I32, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), 0, 0, false},

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

@@ -168,6 +168,21 @@ bool Translator::SimpleInteger(const Decoder::Instruction& inst, IR::ValueOpcode
 	return true;
 }
 
+bool Translator::S_ASHR_I64(const Decoder::Instruction& inst) {
+	// Signed 64-bit sources sign-extend literals; generic B64 operands zero-extend them.
+	const auto source =
+	    inst.src0.kind == Decoder::OperandKind::LiteralConstant
+	        ? ir.ConstructU64(
+	              ReadU32(inst.src0),
+	              IR::U32(IR::Value((inst.src0.value & 0x80000000u) ? 0xffffffffu : 0u)))
+	        : ReadU64(inst.src0);
+	const auto result =
+	    ir.Emit(IR::ValueOpcode::ShiftRightArithmetic64, {source, ReadU32(inst.src1)});
+	WriteOperand(inst.dst, result);
+	ir.SetScc(IR::U1(ir.Emit(IR::ValueOpcode::INotEqual64, {result, IR::Value(uint64_t {0})})));
+	return true;
+}
+
 bool Translator::ComposedIntegerBinary(const Decoder::Instruction& inst, IR::ValueOpcode opcode,
                                        bool negate_rhs, bool negate_result, bool update_scc) {
 	const auto lhs = ReadU32(inst.src0);

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

@@ -1034,6 +1034,7 @@ bool Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::IMAGE_STORE:
 		case Decoder::Opcode::IMAGE_STORE_MIP: return IMAGE_STORE(inst);
 		case Decoder::Opcode::IMAGE_SAMPLE: return IMAGE_SAMPLE(inst);
+		case Decoder::Opcode::IMAGE_GATHER4_L:
 		case Decoder::Opcode::IMAGE_GATHER4_LZ:
 		case Decoder::Opcode::IMAGE_GATHER4_C:
 		case Decoder::Opcode::IMAGE_GATHER4_C_LZ:

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

@@ -179,6 +179,7 @@ bool Translator::EmitScalar(const Decoder::Instruction& inst) {
 		case O::S_LSHR_B64:
 			return SimpleInteger(inst, IR::ValueOpcode::ShiftRightLogical64, IR::Type::U64, false,
 			                     false, true);
+		case O::S_ASHR_I64: return S_ASHR_I64(inst);
 
 		case O::S_ANDN2_B32:
 			return ComposedIntegerBinary(inst, IR::ValueOpcode::BitwiseAnd32, true, false, true);
@@ -227,6 +228,7 @@ bool Translator::EmitScalar(const Decoder::Instruction& inst) {
 		case O::S_CBRANCH_VCCNZ:
 		case O::S_CBRANCH_EXECZ:
 		case O::S_CBRANCH_EXECNZ:
+		case O::S_CBRANCH_CDBGSYS:
 		case O::S_ENDPGM: return true;
 		default: return false;
 	}

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

@@ -789,15 +789,6 @@ const EmbeddedFetchLoad* FindEmbeddedFetchLoad(const EmbeddedFetchPlan* plan, ui
 	return found != plan->loads.end() ? &*found : nullptr;
 }
 
-bool IsEmbeddedFetchPrologLoad(const EmbeddedFetchPlan* plan, uint32_t pc) {
-	if (plan == nullptr) {
-		return false;
-	}
-	return std::ranges::any_of(plan->loads, [pc](const auto& load) {
-		return std::ranges::find(load.prolog_loads, pc) != load.prolog_loads.end();
-	});
-}
-
 int ResolveEmbeddedFetchResource(const ShaderVertexInputInfo& input,
                                  const EmbeddedFetchLoad&     load) {
 	if (load.attrib_id >= 0 && load.attrib_id < input.resources_num &&
@@ -819,22 +810,6 @@ int ResolveEmbeddedFetchResource(const ShaderVertexInputInfo& input,
 	return -1;
 }
 
-bool IsScalarMemoryLoad(Decoder::Opcode opcode) {
-	switch (opcode) {
-		case Decoder::Opcode::S_LOAD_DWORD:
-		case Decoder::Opcode::S_LOAD_DWORDX2:
-		case Decoder::Opcode::S_LOAD_DWORDX4:
-		case Decoder::Opcode::S_LOAD_DWORDX8:
-		case Decoder::Opcode::S_LOAD_DWORDX16:
-		case Decoder::Opcode::S_BUFFER_LOAD_DWORD:
-		case Decoder::Opcode::S_BUFFER_LOAD_DWORDX2:
-		case Decoder::Opcode::S_BUFFER_LOAD_DWORDX4:
-		case Decoder::Opcode::S_BUFFER_LOAD_DWORDX8:
-		case Decoder::Opcode::S_BUFFER_LOAD_DWORDX16: return true;
-		default: return false;
-	}
-}
-
 bool IsBufferDwordLoad(Decoder::Opcode opcode) {
 	switch (opcode) {
 		case Decoder::Opcode::BUFFER_LOAD_DWORD:
@@ -898,11 +873,15 @@ void ValidateTranslateOptions(const TranslateOptions& options) {
 	if (options.wave_size != 32u && options.wave_size != 64u) {
 		EXIT("shader translation requires wave32 or wave64, got %u", options.wave_size);
 	}
-	if (options.embedded_fetch != nullptr && options.stage != ShaderType::Vertex) {
-		EXIT("embedded fetch requires the vertex shader stage");
+	if (options.embedded_fetch != nullptr && options.stage != ShaderType::Vertex &&
+	    options.stage != ShaderType::Local) {
+		EXIT("embedded fetch requires a vertex or local shader");
 	}
 	switch (options.stage) {
 		case ShaderType::Vertex:
+		case ShaderType::Local:
+		case ShaderType::TessellationControl:
+		case ShaderType::TessellationEvaluation:
 		case ShaderType::Mesh:
 			if (options.input_info.vertex == nullptr) {
 				EXIT("vertex shader translation has no vertex input metadata");
@@ -941,6 +920,9 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 	result.user_data_count     = options.user_data_count;
 	switch (options.stage) {
 		case ShaderType::Vertex:
+		case ShaderType::Local:
+		case ShaderType::TessellationControl:
+		case ShaderType::TessellationEvaluation:
 			result.scratch_dwords = options.input_info.vertex->scratch_size_dwords;
 			break;
 		case ShaderType::Mesh:
@@ -1120,26 +1102,32 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 			    entry_ir.BitwiseOr(wave_info, entry_ir.BitwiseOr(entry_ir.ShiftLeftLogical(
 			                                                         primitive_count, u32(8)),
 			                                                     vertex_count)));
-			// GS adjacency addresses local ES records in LDS. Strip winding alternates
-			// with the global primitive number, including across subgroup boundaries.
-			const auto parity = mesh.input_primitive ==
-			                            static_cast<uint32_t>(Prospero::PrimitiveType::kTriStrip)
-			                        ? entry_ir.BitwiseAnd(entry_ir.IAdd(primitive_chunk, local), u32(1))
-			                        : u32(0);
+			// GS adjacency addresses local ES records in LDS. Fans retain the draw's
+			// center in every subgroup; strip winding follows the global primitive.
 			const auto vertex = entry_ir.IMul(local, step);
-			const auto first  = entry_ir.IAdd(vertex, parity);
-			const auto second = mesh.InputPrimitiveSize() >= 2u
-			                        ? entry_ir.ISub(entry_ir.IAdd(vertex, u32(1)), parity)
-			                        : u32(0);
-			const auto third = mesh.InputPrimitiveSize() == 3u
-			                       ? entry_ir.IAdd(vertex, u32(2))
-			                       : u32(0);
+			auto       first  = vertex;
+			auto       second = u32(0);
+			auto       third  = u32(0);
+			if (mesh.InputPrimitiveSize() >= 2u) {
+				second = entry_ir.IAdd(vertex, u32(1));
+			}
+			if (mesh.InputPrimitiveSize() == 3u) {
+				third = entry_ir.IAdd(vertex, u32(2));
+			}
+			auto input_vertex = entry_ir.IAdd(chunk, local);
+			if (mesh.input_primitive == static_cast<uint32_t>(Prospero::PrimitiveType::kTriFan)) {
+				first = u32(0);
+				input_vertex = entry_ir.Select(entry_ir.IEqual(local, u32(0)), u32(0), input_vertex);
+			} else if (mesh.input_primitive == static_cast<uint32_t>(Prospero::PrimitiveType::kTriStrip)) {
+				const auto parity = entry_ir.BitwiseAnd(entry_ir.IAdd(primitive_chunk, local), u32(1));
+				first = entry_ir.IAdd(first, parity);
+				second = entry_ir.ISub(second, parity);
+			}
 			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(0),
 			                      entry_ir.BitwiseOr(entry_ir.ShiftLeftLogical(first, u32(2)),
 			                                         entry_ir.ShiftLeftLogical(second, u32(18))));
 			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(1),
 			                      entry_ir.ShiftLeftLogical(third, u32(2)));
-			const auto input_vertex = entry_ir.IAdd(chunk, local);
 			const auto index_bytes  = draw(3);
 			const auto indexed      = entry_ir.INotEqual(index_bytes, u32(0));
 			const auto index_low    = draw(4);
@@ -1164,15 +1152,53 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 			entry_ir.SetVectorReg(
 			    static_cast<IR::VectorReg>(8),
 			    entry_ir.IAdd(draw(2), builtin(IR::StageInputKind::WorkgroupId, 1)));
+		} else if (options.stage == ShaderType::Local) {
+			entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(3), IR::U32(IR::Value(64u)));
+			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(2),
+			                      builtin(IR::StageInputKind::VertexIndex));
+			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(3), IR::U32(IR::Value(0u)));
+			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(5),
+			                      builtin(IR::StageInputKind::InstanceIndex));
+		} else if (options.stage == ShaderType::TessellationControl) {
+			const auto& tess = options.input_info.vertex->tess;
+			entry_ir.SetScalarReg(
+			    static_cast<IR::ScalarReg>(2),
+			    IR::U32(entry_ir.Emit(IR::ValueOpcode::TessellationBase, {IR::Value(0u)})));
+			entry_ir.SetScalarReg(
+			    static_cast<IR::ScalarReg>(4),
+			    IR::U32(entry_ir.Emit(IR::ValueOpcode::TessellationBase, {IR::Value(1u)})));
+			entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(3),
+			                      IR::U32(IR::Value(0x81010000u | tess.input_control_points |
+			                                        (tess.output_control_points << 8u))));
+			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(0),
+			                      builtin(IR::StageInputKind::PrimitiveId));
+			entry_ir.SetVectorReg(
+			    static_cast<IR::VectorReg>(1),
+			    entry_ir.ShiftLeftLogical(builtin(IR::StageInputKind::InvocationId),
+			                              IR::U32(IR::Value(8u))));
+		} else if (options.stage == ShaderType::TessellationEvaluation) {
+			entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(3), IR::U32(IR::Value(64u)));
+			entry_ir.SetScalarReg(
+			    static_cast<IR::ScalarReg>(4),
+			    IR::U32(entry_ir.Emit(IR::ValueOpcode::TessellationBase, {IR::Value(0u)})));
+			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(5),
+			                      builtin(IR::StageInputKind::TessCoord, 0));
+			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(6),
+			                      builtin(IR::StageInputKind::TessCoord, 1));
+			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(7), IR::U32(IR::Value(0u)));
+			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(8),
+			                      builtin(IR::StageInputKind::PrimitiveId));
 		} else if (options.stage == ShaderType::Pixel) {
 			const auto* ps = options.input_info.pixel;
-			if (ps->ps_perspective_center_vgpr != UINT32_MAX) {
-				entry_ir.SetVectorReg(static_cast<IR::VectorReg>(ps->ps_perspective_center_vgpr),
-				                      builtin(IR::StageInputKind::BaryCoordSmooth, 0));
-				entry_ir.SetVectorReg(
-				    static_cast<IR::VectorReg>(ps->ps_perspective_center_vgpr + 1u),
-				    builtin(IR::StageInputKind::BaryCoordSmooth, 1));
-			}
+			const auto barycentric_pair = [&](uint32_t reg, IR::StageInputKind kind) {
+				if (reg != UINT32_MAX) {
+					entry_ir.SetVectorReg(static_cast<IR::VectorReg>(reg), builtin(kind, 0));
+					entry_ir.SetVectorReg(static_cast<IR::VectorReg>(reg + 1u), builtin(kind, 1));
+				}
+			};
+			barycentric_pair(ps->ps_perspective_center_vgpr, IR::StageInputKind::BaryCoordSmooth);
+			barycentric_pair(ps->ps_perspective_centroid_vgpr,
+			                 IR::StageInputKind::BaryCoordSmoothCentroid);
 			uint32_t reg = ps->ps_system_input_base;
 			if (ps->ps_pos_x) {
 				entry_ir.SetVectorReg(static_cast<IR::VectorReg>(reg++),
@@ -1215,10 +1241,6 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 			if (IsCodeTableLoad(cfg, instruction.pc)) {
 				continue;
 			}
-			if (IsScalarMemoryLoad(instruction.opcode) &&
-			    IsEmbeddedFetchPrologLoad(options.embedded_fetch, instruction.pc)) {
-				continue;
-			}
 			const auto* embedded = FindEmbeddedFetchLoad(options.embedded_fetch, instruction.pc);
 			if (embedded != nullptr && IsBufferDwordLoad(instruction.opcode) &&
 			    instruction.data_dwords == embedded->components &&

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

@@ -13,7 +13,6 @@ struct EmbeddedFetchLoad {
 	uint32_t              pc         = 0;
 	int                   attrib_id  = -1;
 	uint32_t              components = 0;
-	std::vector<uint32_t> prolog_loads;
 };
 
 struct EmbeddedFetchPlan {

src/graphics/shader/recompiler/frontend/translate/Translator.h

@@ -176,6 +176,7 @@ private:
 	                     bool negate_result);
 	bool SimpleInteger(const Decoder::Instruction& inst, IR::ValueOpcode opcode, IR::Type type,
 	                   bool reverse, bool mask_shift_count, bool update_scc);
+	bool S_ASHR_I64(const Decoder::Instruction& inst);
 	bool ComposedIntegerBinary(const Decoder::Instruction& inst, IR::ValueOpcode opcode,
 	                           bool negate_rhs, bool negate_result, bool update_scc);
 	bool V_AND_OR_B32(const Decoder::Instruction& inst);

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

@@ -112,6 +112,10 @@ ResourcePlan& ResourcePlan::operator=(ResourcePlan&& other) noexcept {
 }
 
 Program::~Program() {
+	// Planning expressions can refer to block values but outlive block storage in the base class.
+	for (auto& inst: value_storage) {
+		inst.Invalidate();
+	}
 	// Values may cross block boundaries. Detach all arguments before any block starts destroying
 	// its instruction storage so reverse-use links always point to live definitions.
 	for (auto* block: blocks) {
@@ -184,6 +188,22 @@ Value ResolveInvariantPhi(const ResourcePlan& program, Value value) {
 	return invariant;
 }
 
+bool HasShaderMemoryWrites(const Program& program) {
+	for (const auto* block: program.blocks) {
+		for (const auto& inst: *block) {
+			const auto op     = inst.GetOpcode();
+			const auto buffer = BufferAccessOf(op);
+			const auto image  = ImageOpcodeInfoOf(op).access;
+			if (buffer == BufferAccess::Write || buffer == BufferAccess::Atomic ||
+			    image == ImageAccess::Write || image == ImageAccess::Atomic ||
+			    AddressOpcodeInfoOf(op).access == AddressAccess::Write) {
+				return true;
+			}
+		}
+	}
+	return false;
+}
+
 void ValidateProgram(const Program& program, bool require_ssa) {
 	if (program.blocks.size() != program.block_info.size() ||
 	    program.blocks.size() != program.block_storage.size()) {

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

@@ -150,8 +150,20 @@ struct SampledResourcePair {
 	bool operator==(const SampledResourcePair& other) const = default;
 };
 
+enum class TessellationAttribute {
+	LocalOutput,
+	ControlInput,
+	ControlOutput,
+	EvaluationInput,
+	PatchOutput,
+	Factor
+};
+
 enum class StageInputKind {
 	VertexIndex,
+	InvocationId,
+	PrimitiveId,
+	TessCoord,
 	InstanceIndex,
 	FragCoord,
 	FrontFacing,
@@ -159,6 +171,7 @@ enum class StageInputKind {
 	Layer,
 	SampleId,
 	BaryCoordSmooth,
+	BaryCoordSmoothCentroid,
 	BaryCoordNoPerspective,
 	WorkgroupId,
 	LocalInvocationId,
@@ -287,8 +300,12 @@ static_assert(sizeof(PushData) == 128);
 constexpr uint32_t NativePushConstantSize = sizeof(PushData);
 
 [[nodiscard]] constexpr uint32_t NativeBinding(ShaderType stage, DescriptorBindingKind kind) {
+	const uint32_t group = stage == ShaderType::Pixel                    ? 1u
+	                       : stage == ShaderType::TessellationControl    ? 2u
+	                       : stage == ShaderType::TessellationEvaluation ? 3u
+	                                                                     : 0u;
 	return static_cast<uint32_t>(kind) +
-	       (stage == ShaderType::Pixel ? static_cast<uint32_t>(DescriptorBindingKind::Count) : 0u);
+	       group * static_cast<uint32_t>(DescriptorBindingKind::Count);
 }
 
 [[nodiscard]] constexpr ImageResourceClass ImageBindingResourceClass(DescriptorBindingKind kind) {
@@ -546,6 +563,7 @@ struct Program: ResourcePlan {
 };
 
 std::string ProgramToString(const Program& program);
+bool        HasShaderMemoryWrites(const Program& program);
 
 void  ValidateProgram(const Program& program, bool require_ssa);
 void  ResolveControlFlowIdentities(Program& program);

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

@@ -90,6 +90,7 @@ bool HasSideEffects(ValueOpcode opcode) {
 		case ValueOpcode::Reference:
 		case ValueOpcode::ReferenceU32:
 		case ValueOpcode::SetAttribute:
+		case ValueOpcode::SetTessellationAttribute:
 		case ValueOpcode::MeshAllocate:
 		case ValueOpcode::Barrier: return true;
 		default: return false;

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

@@ -10,6 +10,9 @@ VALUE_OPCODE(GetUserData, U32, ScalarReg)
 VALUE_OPCODE(GetShaderBase, U64)
 VALUE_OPCODE(MeshDrawParameter, U32, U32)
 VALUE_OPCODE(MeshAllocate, Void, U32)
+VALUE_OPCODE(TessellationBase, U32, U32)
+VALUE_OPCODE(GetTessellationAttribute, U32, U32, U32, U1)
+VALUE_OPCODE(SetTessellationAttribute, Void, U32, U32, U32, U1)
 VALUE_OPCODE(GetBuiltin, U32, U32, U32)
 VALUE_OPCODE(GetThreadBitScalarRegister, U1, ScalarReg)
 VALUE_OPCODE(SetThreadBitScalarRegister, Void, ScalarReg, U1)

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

@@ -228,6 +228,15 @@ void FoldInstruction(Block& block, Block::iterator instruction,
 			const auto offset = Arg(inst, 1);
 			const auto count  = Arg(inst, 2);
 			auto* source = value.TryInstruction();
+			if (source != nullptr && source->GetOpcode() == ValueOpcode::ShiftLeftLogical32 &&
+			    IsImmediate(offset, Type::U32) && IsImmediate(count, Type::U32)) {
+				const auto shift = Arg(*source, 1);
+				if (IsImmediate(shift, Type::U32) && shift.U32() < 32u &&
+				    offset.U32() <= shift.U32() && count.U32() <= shift.U32() - offset.U32()) {
+					Replace(inst, Value(0u));
+					return;
+				}
+			}
 			if (source != nullptr && source->GetOpcode() == ValueOpcode::GetBuiltin &&
 			    source->Arg(0) == Value(static_cast<uint32_t>(StageInputKind::PackedAncillary)) &&
 			    IsImmediate(offset, Type::U32) && IsImmediate(count, Type::U32) && count.U32() != 0u) {

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

@@ -647,7 +647,8 @@ bool BuildSamplerPlan(const ShaderInfo& base, const Images& images, SamplerPlan&
 }
 
 static std::vector<ResourceBlock> ResourceControlFlow(const Program& program) {
-	if (program.blocks.size() != program.block_info.size()) {
+	// Any shader write may alias a scalar predicate read, including on a later loop visit.
+	if (program.blocks.size() != program.block_info.size() || HasShaderMemoryWrites(program)) {
 		return {};
 	}
 	std::unordered_map<uint32_t, uint32_t> indices;
@@ -687,12 +688,6 @@ static std::vector<ResourceBlock> ResourceControlFlow(const Program& program) {
 			const auto op     = inst.GetOpcode();
 			const auto buffer = BufferAccessOf(op);
 			const auto image  = ImageOpcodeInfoOf(op);
-			// Any shader write may alias a scalar predicate read, including on a later loop visit.
-			if (buffer == BufferAccess::Write || buffer == BufferAccess::Atomic ||
-			    image.access == ImageAccess::Write || image.access == ImageAccess::Atomic ||
-			    AddressOpcodeInfoOf(op).access == AddressAccess::Write) {
-				return {};
-			}
 			if (buffer == BufferAccess::None && image.access == ImageAccess::None) {
 				continue;
 			}

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

@@ -88,6 +88,7 @@ public:
 		m_info.samplers.clear();
 		m_info.sampled_pairs.clear();
 		m_info.uses_dma = false;
+		m_shader_writes = HasShaderMemoryWrites(program);
 	}
 
 	void Run() {
@@ -168,15 +169,85 @@ private:
 		std::abort();
 	}
 
+	Value LowerDescriptorPhi(Value value) {
+		value           = value.Resolve();
+		const auto* phi = value.TryInstruction();
+		if (m_shader_writes || phi == nullptr || phi->GetOpcode() != ValueOpcode::Phi ||
+		    phi->NumArgs() != 2u || phi->NumPhiBlocks() != 2u || phi->GetType() != Type::U32 ||
+		    m_program.blocks.size() != m_program.block_info.size()) {
+			return value;
+		}
+		for (const auto& [original, selected]: m_descriptor_selections) {
+			if (original == phi) {
+				return selected;
+			}
+		}
+		const auto* merge  = phi->Parent();
+		const auto* branch = phi->PhiBlock(0);
+		if (merge == nullptr || branch == nullptr || phi->PhiBlock(1) == nullptr ||
+		    branch == phi->PhiBlock(1)) {
+			return value;
+		}
+		if (branch->ImmSuccessors().size() != 2u) {
+			if (branch->ImmPredecessors().size() != 1u) {
+				return value;
+			}
+			branch = branch->ImmPredecessors()[0];
+		}
+		if (branch == merge || branch->ImmSuccessors().size() != 2u) {
+			return value;
+		}
+		std::array<uint32_t, 2> target_ids;
+		for (uint32_t arm = 0; arm < 2; arm++) {
+			const auto* incoming = phi->PhiBlock(arm);
+			if (incoming == merge ||
+			    (incoming != branch &&
+			     (incoming->ImmPredecessors().size() != 1u ||
+			      incoming->ImmPredecessors()[0] != branch ||
+			      incoming->ImmSuccessors().size() != 1u ||
+			      incoming->ImmSuccessors()[0] != merge))) {
+				return value;
+			}
+			const auto* target = incoming == branch ? merge : incoming;
+			const auto  it     = std::ranges::find(m_program.blocks, target);
+			if (it == m_program.blocks.end()) {
+				return value;
+			}
+			target_ids[arm] = m_program.block_info[it - m_program.blocks.begin()].id;
+		}
+		const auto branch_it = std::ranges::find(m_program.blocks, branch);
+		if (branch_it == m_program.blocks.end()) {
+			return value;
+		}
+		const auto& info = m_program.block_info[branch_it - m_program.blocks.begin()];
+		const auto& term = info.terminator;
+		if (term.kind != CFG::TerminatorKind::ConditionalBranch ||
+		    !((term.true_block == target_ids[0] && term.false_block == target_ids[1]) ||
+		      (term.false_block == target_ids[0] && term.true_block == target_ids[1])) ||
+		    !ValidateRuntimeValue(m_program, info.condition, RuntimeValueType::Integer) ||
+		    !ValidateRuntimeValue(m_program, phi->Arg(0)) ||
+		    !ValidateRuntimeValue(m_program, phi->Arg(1))) {
+			return value;
+		}
+		// Retain a host expression; replacing the GPU Phi would break SSA dominance.
+		const auto true_arg = term.true_block == target_ids[0] ? 0u : 1u;
+		auto&      selected = m_program.value_storage.emplace_back(ValueOpcode::SelectU32);
+		selected.SetArg(0, info.condition);
+		selected.SetArg(1, phi->Arg(true_arg));
+		selected.SetArg(2, phi->Arg(true_arg ^ 1u));
+		m_descriptor_selections.emplace_back(phi, Value(&selected));
+		return Value(&selected);
+	}
+
 	void MakeSource(const Inst& handle, uint32_t width, bool sampler, bool sample_adjust,
-	                DescriptorSource& descriptor, uint32_t pc) const {
+	                DescriptorSource& descriptor, uint32_t pc) {
 		if (handle.NumArgs() != width) {
 			Fail(pc, fmt::format("{} has {} descriptor dwords, expected {}",
 			                     ValueOpcodeName(handle.GetOpcode()), handle.NumArgs(), width));
 		}
 		descriptor.dword_count = width;
 		for (uint32_t i = 0; i < width; i++) {
-			descriptor.dwords[i] = handle.Arg(i).Resolve();
+			descriptor.dwords[i] = LowerDescriptorPhi(handle.Arg(i));
 		}
 		if (sample_adjust) {
 			descriptor.dwords[3] = CanonicalizeSampleAdjustDword3(descriptor.dwords[3]);
@@ -744,12 +815,14 @@ private:
 		}
 	}
 
-	Program&                       m_program;
-	ShaderInfo                     m_info;
-	std::vector<DescriptorSource>  m_sources;
-	std::vector<HandlePatch>       m_handle_patches;
-	std::vector<MemoryPatch>       m_memory_patches;
-	std::vector<IndirectImagePlan> m_indirect_images;
+	Program&                                   m_program;
+	ShaderInfo                                 m_info;
+	std::vector<DescriptorSource>              m_sources;
+	std::vector<HandlePatch>                   m_handle_patches;
+	std::vector<MemoryPatch>                   m_memory_patches;
+	std::vector<IndirectImagePlan>             m_indirect_images;
+	std::vector<std::pair<const Inst*, Value>> m_descriptor_selections;
+	bool                                       m_shader_writes = false;
 };
 
 } // namespace

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

@@ -43,6 +43,9 @@ void AddOutput(ShaderInfo& info, StageOutputKind kind, uint32_t index, uint32_t
 void ValidateOptions(const Program& program, ShaderStageInputInfo input_info) {
 	switch (program.stage) {
 		case ShaderType::Vertex:
+		case ShaderType::Local:
+		case ShaderType::TessellationControl:
+		case ShaderType::TessellationEvaluation:
 		case ShaderType::Mesh:
 			if (input_info.vertex == nullptr) {
 				return Fail("vertex shader has no input metadata");
@@ -81,7 +84,8 @@ void ValidateValueReferences(const Program& program, ShaderStageInputInfo input_
 					    !inst.Arg(1).IsImmediate() || inst.Arg(1).GetType() != Type::U32) {
 						return Fail("typed attribute reference is not constant");
 					}
-					if (program.stage == ShaderType::Vertex &&
+					if ((program.stage == ShaderType::Vertex ||
+					     program.stage == ShaderType::Local) &&
 					    (inst.Arg(1).U32() >= 4u ||
 					     inst.Arg(0).U32() >=
 					         static_cast<uint32_t>(input_info.vertex->resources_num))) {
@@ -124,6 +128,8 @@ void ValidateValueReferences(const Program& program, ShaderStageInputInfo input_
 							}
 							break;
 						case StageInputKind::VertexIndex:
+						case StageInputKind::InvocationId:
+						case StageInputKind::PrimitiveId:
 						case StageInputKind::InstanceIndex:
 						case StageInputKind::FrontFacing:
 						case StageInputKind::LocalInvocationIndex:
@@ -137,11 +143,13 @@ void ValidateValueReferences(const Program& program, ShaderStageInputInfo input_
 							}
 							break;
 						case StageInputKind::BaryCoordSmooth:
+						case StageInputKind::BaryCoordSmoothCentroid:
 						case StageInputKind::BaryCoordNoPerspective:
 							if (component >= 2u) {
 								return Fail("typed barycentric component is out of range");
 							}
 							break;
+						case StageInputKind::TessCoord:
 						case StageInputKind::WorkgroupId:
 						case StageInputKind::LocalInvocationId:
 						case StageInputKind::GlobalInvocationId:
@@ -161,8 +169,10 @@ void ValidateValueReferences(const Program& program, ShaderStageInputInfo input_
 					}
 					const auto& exp = program.export_info[index];
 					if (exp.kind == ExportTargetKind::Position && exp.index != 0 && exp.en != 0 &&
-					    program.stage != ShaderType::Vertex && program.stage != ShaderType::Mesh) {
-						return Fail("auxiliary position export requires the vertex or mesh stage");
+					    program.stage != ShaderType::Vertex && program.stage != ShaderType::Mesh &&
+					    program.stage != ShaderType::TessellationEvaluation) {
+						return Fail("auxiliary position export requires a vertex, mesh, or "
+						            "tessellation evaluation shader");
 					}
 					break;
 				}
@@ -216,6 +226,18 @@ void CollectPixelInputs(const Program& program, const ShaderPixelInputInfo* pixe
 			}
 		}
 	}
+	// Aliases of a vertex output share one SPIR-V interface variable. If any
+	// alias reads raw vertices, interpolate the other aliases from those too.
+	for (uint32_t input = 0; input < pixel->input_num; input++) {
+		for (uint32_t alias = 0; alias < pixel->input_num; alias++) {
+			if (ShaderPixelParameterMappedLocation(*pixel, input) ==
+			        ShaderPixelParameterMappedLocation(*pixel, alias) &&
+			    ShaderPixelParameterIsFlat(*pixel, input) ==
+			        ShaderPixelParameterIsFlat(*pixel, alias)) {
+				per_vertex[input] = per_vertex[input] || per_vertex[alias];
+			}
+		}
+	}
 	for (uint32_t input = 0; input < pixel->input_num; input++) {
 		AddInput(info, StageInputKind::Parameter, input, 4, fmt::format("in_param_{}", input),
 		         per_vertex[input]);
@@ -249,6 +271,10 @@ void CollectComputeInputs(const ShaderComputeInputInfo* compute, ShaderInfo& inf
 void CollectBuiltinInputs(const Program& program, ShaderInfo& info) {
 	for (const auto* block: program.blocks) {
 		for (const auto& inst: *block) {
+			if (program.stage == ShaderType::TessellationControl &&
+			    inst.GetOpcode() == ValueOpcode::LaneId) {
+				AddInput(info, StageInputKind::InvocationId, 0, 1, "gl_InvocationID");
+			}
 			if (inst.GetOpcode() != ValueOpcode::GetBuiltin) {
 				continue;
 			}
@@ -260,6 +286,13 @@ void CollectBuiltinInputs(const Program& program, ShaderInfo& info) {
 				case StageInputKind::InstanceIndex:
 					AddInput(info, kind, 0, 1, "gl_InstanceIndex");
 					break;
+				case StageInputKind::InvocationId:
+					AddInput(info, kind, 0, 1, "gl_InvocationID");
+					break;
+				case StageInputKind::PrimitiveId:
+					AddInput(info, kind, 0, 1, "gl_PrimitiveID");
+					break;
+				case StageInputKind::TessCoord: AddInput(info, kind, 0, 3, "gl_TessCoord"); break;
 				case StageInputKind::FragCoord: AddInput(info, kind, 0, 4, "gl_FragCoord"); break;
 				case StageInputKind::FrontFacing:
 					AddInput(info, kind, 0, 1, "gl_FrontFacing");
@@ -267,7 +300,8 @@ void CollectBuiltinInputs(const Program& program, ShaderInfo& info) {
 				case StageInputKind::Layer: AddInput(info, kind, 0, 1, "gl_Layer"); break;
 				case StageInputKind::SampleId: AddInput(info, kind, 0, 1, "gl_SampleID"); break;
 				case StageInputKind::BaryCoordSmooth:
-					AddInput(info, kind, 0, 3, "gl_BaryCoordKHR");
+				case StageInputKind::BaryCoordSmoothCentroid:
+					AddInput(info, StageInputKind::BaryCoordSmooth, 0, 3, "gl_BaryCoordKHR");
 					break;
 				case StageInputKind::BaryCoordNoPerspective:
 					AddInput(info, kind, 0, 3, "gl_BaryCoordNoPerspKHR");
@@ -380,7 +414,10 @@ void CollectShaderInfo(Program& program, ShaderStageInputInfo input_info) {
 		    });
 	    });
 	switch (program.stage) {
-		case ShaderType::Vertex: CollectVertexInputs(program, input_info.vertex, next); break;
+		case ShaderType::Vertex:
+		case ShaderType::Local: CollectVertexInputs(program, input_info.vertex, next); break;
+		case ShaderType::TessellationControl:
+		case ShaderType::TessellationEvaluation:
 		case ShaderType::Mesh: break;
 		case ShaderType::Pixel: CollectPixelInputs(program, input_info.pixel, next); break;
 		case ShaderType::Compute: CollectComputeInputs(input_info.compute, next); break;

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

@@ -16,6 +16,13 @@ namespace {
 
 constexpr uint64_t AddressMask = 0x0000ffffffffffffull;
 
+SrtRuntime CleanRuntime(SrtRuntime runtime) {
+	runtime.read_memory = runtime.read_specialization_memory != nullptr
+	                          ? runtime.read_specialization_memory
+	                          : +[](void*, uint64_t, uint32_t*) { return false; };
+	return runtime;
+}
+
 const char* StageName(ShaderType stage) {
 	switch (stage) {
 		case ShaderType::Vertex: return "vertex";
@@ -646,8 +653,10 @@ private:
 			case ValueOpcode::GetShaderBase: result = m_runtime.shader_base; return true;
 			case ValueOpcode::Phi: return EvaluatePhi(inst, result);
 			case ValueOpcode::ReadFirstLane: {
-				Evaluator active(m_program, m_runtime, m_clean_flat_slots, m_clean_evaluator,
-				                 inst.Arg(1));
+				const auto clean_runtime = CleanRuntime(m_runtime);
+				Evaluator  clean_active(m_program, clean_runtime, {}, nullptr, inst.Arg(1));
+				Evaluator  active(m_program, m_runtime, m_clean_flat_slots, &clean_active,
+				                  inst.Arg(1));
 				return active.EvaluateWide(inst.Arg(0), result);
 			}
 			case ValueOpcode::BitCastU32F32:
@@ -892,12 +901,13 @@ private:
 			}
 			case ValueOpcode::SelectU32:
 			case ValueOpcode::SelectU1:
-			case ValueOpcode::SelectF32:
-				if (ternary()) {
-					result = a != 0u ? b : c;
-					return true;
+			case ValueOpcode::SelectF32: {
+				auto& predicate = m_clean_evaluator != nullptr ? *m_clean_evaluator : *this;
+				if (predicate.EvaluateWide(inst.Arg(0), a)) {
+					return Arg(inst, a != 0u ? 1u : 2u, result);
 				}
 				return false;
+			}
 			case ValueOpcode::IEqual32:
 				if (binary()) {
 					result = static_cast<uint32_t>(a) == static_cast<uint32_t>(b);
@@ -977,7 +987,7 @@ bool EvaluateRuntimeSourcesImpl(const ResourcePlan& program, std::span<const uin
                                 const SrtRuntime& runtime, std::vector<DescriptorValue>& results,
                                 std::vector<uint32_t>& flat, bool evaluate_flat,
                                 std::span<const uint8_t> clean_flat_slots,
-                                std::vector<uint8_t>& active_sources) {
+                                std::vector<uint8_t>&    active_sources) {
 	if (!program.srt_plan_complete) {
 		return false;
 	}
@@ -985,8 +995,7 @@ bool EvaluateRuntimeSourcesImpl(const ResourcePlan& program, std::span<const uin
 	    runtime.read_specialization_memory == nullptr) {
 		return false;
 	}
-	SrtRuntime clean_runtime  = runtime;
-	clean_runtime.read_memory = runtime.read_specialization_memory;
+	const auto           clean_runtime = CleanRuntime(runtime);
 	Evaluator            clean_evaluator(program, clean_runtime);
 	Evaluator            evaluator(program, runtime, clean_flat_slots, &clean_evaluator);
 	std::vector<uint8_t> active;
@@ -1081,11 +1090,8 @@ bool EvaluateUniformValues(const ResourcePlan& program, std::span<const Value> v
 	if (values.size() != results.size()) {
 		return false;
 	}
-	auto clean = runtime;
-	clean.read_memory = runtime.read_specialization_memory != nullptr
-	                        ? runtime.read_specialization_memory
-	                        : +[](void*, uint64_t, uint32_t*) { return false; };
-	Evaluator evaluator(program, clean);
+	const auto clean = CleanRuntime(runtime);
+	Evaluator  evaluator(program, clean);
 	for (size_t i = 0; i < values.size(); ++i) {
 		if (!evaluator.Evaluate(values[i], results[i])) {
 			return false;

src/graphics/shader/rectListShader.cpp

@@ -40,12 +40,15 @@ std::vector<Parameter> GetParameters(const ShaderVertexInputInfo& vertex_info,
 	}
 
 	std::vector<Parameter> parameters;
+	std::array<bool, 32> output_locations {};
 	for (const auto input: active_inputs) {
 		const auto input_location = ShaderPixelParameterMappedLocation(*pixel_info, input);
-		if ((vertex_info.stage.program->param_export_mask & (1u << input_location)) != 0) {
-			parameters.push_back({input_location,
-			                      ShaderPixelParameterLocation(*pixel_info, active_inputs, input),
+		const auto output_location = ShaderPixelParameterLocation(*pixel_info, active_inputs, input);
+		if ((vertex_info.stage.program->param_export_mask & (1u << input_location)) != 0 &&
+		    !output_locations[output_location]) {
+			parameters.push_back({input_location, output_location,
 			                      ShaderPixelParameterIsFlat(*pixel_info, input)});
+			output_locations[output_location] = true;
 		}
 	}
 	return parameters;

src/graphics/shader/shader.cpp

@@ -11,6 +11,8 @@
 #include "graphics/guest_gpu/graphicsRun.h"
 #include "graphics/guest_gpu/hardwareContext.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
+#include "graphics/shader/recompiler/ShaderRecompiler.h"
+#include "graphics/shader/recompiler/Tessellation.h"
 #include "graphics/shader/recompiler/frontend/decode/ShaderDecoder.h"
 #include "graphics/shader/shaderCompiler.h"
 #include "graphics/shader/shaderVertexMetadata.h"
@@ -511,33 +513,28 @@ static uint32_t ShaderCalcPsSystemInputBase(const HW::ShaderRegisters& regs) {
 	return reg;
 }
 
-static bool ShaderGetStaticInputInfoVS(const HW::VertexShaderInfo& regs,
-	                                   const HW::ShaderRegisters& sh,
-	                                   const ShaderMappedData& data,
-	                                   ShaderVertexInputInfo& info) {
+static bool ShaderGetStaticVertexInputInfo(uint64_t shader_addr, const HW::UserSgprInfo& user_sgpr,
+                                           uint32_t user_sgpr_num, const HW::ShaderRegisters& sh,
+                                           const ShaderMappedData& data,
+                                           ShaderVertexInputInfo&  info) {
 	KYTY_PROFILER_FUNCTION();
 
 	info = {};
 
 	info.pa_cl_vs_out_cntl = sh.m_paClVsOutCntl;
 
-	EXIT_NOT_IMPLEMENTED(regs.es_regs.data_addr == 0);
-
-	uint64_t                shader_addr   = regs.es_regs.data_addr;
-	const HW::UserSgprInfo& user_sgpr     = regs.gs_user_sgpr;
-	auto                    user_sgpr_num = regs.gs_regs.rsrc2.user_sgpr;
+	EXIT_NOT_IMPLEMENTED(shader_addr == 0);
 	info.scratch_size_dwords = data.scratch_size_dwords;
 
 	if (data.user_data == nullptr) {
-		LOGF("ShaderGetInputInfoVS(): no AGC user data for shader=0x%016" PRIx64 " es=0x%016" PRIx64
-		     " gs=0x%016" PRIx64 " user_sgpr_num=%u\n",
-		     shader_addr, regs.es_regs.data_addr, regs.gs_regs.data_addr,
-		     static_cast<uint32_t>(user_sgpr_num));
+		LOGF("ShaderGetVertexInputInfo(): no AGC user data for shader=0x%016" PRIx64
+		     " user_sgpr_num=%u\n",
+		     shader_addr, user_sgpr_num);
 	}
 	ShaderVertexMetadata metadata;
 	std::string          metadata_error;
 	if (!ShaderReadVertexMetadata(data, HW::UserSgprInfo::SGPRS_MAX, metadata, &metadata_error)) {
-		LOGF("ShaderGetInputInfoVS(): invalid AGC metadata shader=0x%016" PRIx64 ": %s\n",
+		LOGF("ShaderGetVertexInputInfo(): invalid AGC metadata shader=0x%016" PRIx64 ": %s\n",
 		     shader_addr, metadata_error.c_str());
 		return false;
 	}
@@ -556,7 +553,7 @@ static bool ShaderGetStaticInputInfoVS(const HW::VertexShaderInfo& regs,
 		    (static_cast<uint64_t>(user_sgpr.value[metadata.vertex_buffer_reg + 1]) << 32u));
 
 		if (attrib == nullptr || buffer == nullptr) {
-			LOGF("ShaderGetInputInfoVS(): null vertex table pointer shader=0x%016" PRIx64 "\n",
+			LOGF("ShaderGetVertexInputInfo(): null vertex table pointer shader=0x%016" PRIx64 "\n",
 			     shader_addr);
 			return false;
 		}
@@ -576,15 +573,20 @@ static void ShaderGetStaticInputInfoPS(
 	ps_info = {};
 	ps_info.scratch_size_dwords = data.scratch_size_dwords;
 
-	// SPI_PS_IN_CONTROL.NUM_INTERP occupies bits 5:0. Keep the remaining control
-	// flags in the hardware state and extract only the input count here.
+	// SPI_PS_IN_CONTROL: NUM_INTERP occupies bits 5:0 and PS_W32_EN is bit 15.
 	ps_info.input_num            = sh.ps_in_control & 0x3fu;
+	if ((sh.ps_in_control & 0x8000u) != 0) {
+		ps_info.wave_size = 32;
+	}
 	EXIT_NOT_IMPLEMENTED(ps_info.input_num > std::size(ps_info.interpolator_settings));
 	ps_info.ps_system_input_base = ShaderCalcPsSystemInputBase(sh);
 	const uint32_t active_inputs = sh.ps_input_ena & sh.ps_input_addr;
 	if ((active_inputs & 0x00000002u) != 0) {
 		ps_info.ps_perspective_center_vgpr = (active_inputs & 0x00000001u) != 0 ? 2u : 0u;
 	}
+	if ((active_inputs & 0x00000004u) != 0) {
+		ps_info.ps_perspective_centroid_vgpr = 2u * std::popcount(active_inputs & 0x3u);
+	}
 	for (uint32_t i = 0; i < data.num_input_semantics && i < ps_info.input_num && i < 32u; i++) {
 		const auto& semantic = data.input_semantics[i];
 		if (semantic.is_custom != 0 && semantic.is_f16 == 0) {
@@ -673,6 +675,12 @@ void BuildStageStaticKey(const ShaderVertexInputInfo& info, std::vector<uint32_t
 		                       mesh.primitives_per_group, mesh.vertices_per_group,
 		                       mesh.max_vertices, mesh.max_primitives, mesh.provoking_vertex});
 	}
+	key.push_back(info.tess.input_control_points);
+	if (info.tess.input_control_points != 0) {
+		const auto& tess = info.tess;
+		key.insert(key.end(), {tess.output_control_points, tess.ls_stride, tess.hs_stride,
+		                       tess.domain, tess.partitioning, tess.output_topology});
+	}
 
 	for (int i = 0; i < info.resources_num; i++) {
 		const auto& resource    = info.resources[i];
@@ -698,9 +706,11 @@ void BuildStageStaticKey(const ShaderPixelInputInfo& info, std::vector<uint32_t>
 	key.clear();
 	key.push_back(info.scratch_size_dwords);
 	key.push_back(info.input_num);
+	key.push_back(info.wave_size);
 	key.push_back(info.ps_system_input_base);
 	key.push_back(info.custom_interpolation_mask);
 	key.push_back(info.ps_perspective_center_vgpr);
+	key.push_back(info.ps_perspective_centroid_vgpr);
 	key.push_back(static_cast<uint32_t>(info.ps_pos_x));
 	key.push_back(static_cast<uint32_t>(info.ps_pos_y));
 	key.push_back(static_cast<uint32_t>(info.ps_pos_z));
@@ -750,7 +760,8 @@ ShaderParams PrepareProgram(const HW::VertexShaderInfo& regs, const HW::Context&
 	    GetDeclaredShaderHash(regs.es_regs.data_addr),
 	    std::span<const uint32_t>(regs.gs_user_sgpr.value, regs.gs_regs.rsrc2.user_sgpr), data);
 	if ((context.GetShaderStages() & 0x20u) == 0) {
-		if (!ShaderGetStaticInputInfoVS(regs, sh, data, info)) {
+		if (!ShaderGetStaticVertexInputInfo(regs.es_regs.data_addr, regs.gs_user_sgpr,
+		                                    regs.gs_regs.rsrc2.user_sgpr, sh, data, info)) {
 			EXIT("failed to prepare vertex shader program\n");
 		}
 		return params;
@@ -759,6 +770,7 @@ ShaderParams PrepareProgram(const HW::VertexShaderInfo& regs, const HW::Context&
 	// its user-data pointer in s0:s1.
 	params.user_data.insert(params.user_data.begin(), 8u, 0u);
 	info                     = {};
+	info.logical_stage       = ShaderType::Mesh;
 	info.pa_cl_vs_out_cntl   = sh.m_paClVsOutCntl;
 	auto& mesh               = info.mesh;
 	mesh.input_primitive     = static_cast<uint32_t>(user_config.GetPrimType());
@@ -785,6 +797,7 @@ ShaderParams PrepareProgram(const HW::VertexShaderInfo& regs, const HW::Context&
 	const auto& group = user_config.GetGeControl();
 	if ((user_config.GetPrimType() != Prospero::PrimitiveType::kPointList &&
 	     user_config.GetPrimType() != Prospero::PrimitiveType::kLineList &&
+	     user_config.GetPrimType() != Prospero::PrimitiveType::kTriFan &&
 	     user_config.GetPrimType() != Prospero::PrimitiveType::kTriStrip &&
 	     user_config.GetPrimType() != Prospero::PrimitiveType::kTriList) ||
 	    sh.m_vgtGsOutPrimType != 2u || sh.m_vgtGsMaxVertOut < 3u ||
@@ -806,6 +819,65 @@ ShaderParams PrepareProgram(const HW::VertexShaderInfo& regs, const HW::Context&
 	return params;
 }
 
+std::array<ShaderParams, 3>
+PrepareTessellationPrograms(const HW::VertexShaderInfo& regs, const HW::Context& context,
+                            std::array<ShaderVertexInputInfo, 3>& input_info) {
+	const auto& sh        = context.GetShaderRegisters();
+	const auto  local     = ShaderGetMappedData(regs.ls_regs.data_addr, "ShaderGetInputInfoLS():");
+	const auto  control   = ShaderGetMappedData(regs.hs_regs.data_addr, "ShaderGetInputInfoHS():");
+	const auto evaluation = ShaderGetMappedData(regs.es_regs.data_addr, "ShaderGetInputInfoTES():");
+	EXIT_NOT_IMPLEMENTED(local.type != Prospero::ShaderBinaryType::kHsFront ||
+	                     control.type != Prospero::ShaderBinaryType::kHsBack ||
+	                     evaluation.type != Prospero::ShaderBinaryType::kGs);
+	EXIT_NOT_IMPLEMENTED((context.GetShaderStages() & 0x4u) == 0 ||
+	                     (context.GetShaderStages() & 0x00600020u) != 0);
+	EXIT_IF(regs.hs_regs.user_data_addr == 0);
+
+	const auto local_users      = std::span(regs.hs_user_sgpr.value, regs.hs_regs.rsrc2.user_sgpr);
+	const auto evaluation_users = std::span(regs.gs_user_sgpr.value, regs.gs_regs.rsrc2.user_sgpr);
+	std::array<ShaderParams, 3> params {
+	    GetShaderParams(regs.ls_regs.data_addr, "ShaderRecompiler LS",
+	                    GetDeclaredShaderHash(regs.ls_regs.data_addr), local_users, local),
+	    GetShaderParams(regs.hs_regs.data_addr, "ShaderRecompiler HS",
+	                    GetDeclaredShaderHash(regs.hs_regs.data_addr), local_users, control),
+	    GetShaderParams(regs.es_regs.data_addr, "ShaderRecompiler TES",
+	                    GetDeclaredShaderHash(regs.es_regs.data_addr), evaluation_users,
+	                    evaluation),
+	};
+	// The fused HS back half receives its separate user-data address in s0:s1.
+	// RDNA2 reserves s0:s7 before the native HS user SGPRs.
+	params[1].user_data.insert(params[1].user_data.begin(), 8u, 0u);
+	params[1].user_data[0] = static_cast<uint32_t>(regs.hs_regs.user_data_addr);
+	params[1].user_data[1] = static_cast<uint32_t>(regs.hs_regs.user_data_addr >> 32u);
+
+	input_info = {};
+	if (!ShaderGetStaticVertexInputInfo(regs.ls_regs.data_addr, regs.hs_user_sgpr,
+	                                    regs.hs_regs.rsrc2.user_sgpr, sh, local, input_info[0])) {
+		EXIT("failed to prepare local shader program\n");
+	}
+	input_info[0].logical_stage       = ShaderType::Local;
+	input_info[1].logical_stage       = ShaderType::TessellationControl;
+	input_info[1].scratch_size_dwords = control.scratch_size_dwords;
+	input_info[2].logical_stage       = ShaderType::TessellationEvaluation;
+	input_info[2].scratch_size_dwords = evaluation.scratch_size_dwords;
+	input_info[2].pa_cl_vs_out_cntl   = sh.m_paClVsOutCntl;
+
+	ShaderTessellationInputInfo tess {
+	    .input_control_points  = (sh.m_vgtLsHsConfig >> 8u) & 0x3fu,
+	    .output_control_points = (sh.m_vgtLsHsConfig >> 14u) & 0x3fu,
+	    .domain                = sh.m_vgtTfParam & 0x3u,
+	    .partitioning          = (sh.m_vgtTfParam >> 2u) & 0x3u,
+	    .output_topology       = (sh.m_vgtTfParam >> 5u) & 0x3u,
+	};
+	EXIT_IF(tess.input_control_points == 0 || tess.input_control_points > 32 ||
+	        tess.output_control_points == 0 || tess.output_control_points > 32);
+	ShaderRecompiler::AnalyzeTessellationPrograms(params[0].code, params[1].code, tess);
+	for (auto& stage: input_info) {
+		stage.tess = tess;
+	}
+	return params;
+}
+
 ShaderParams PrepareProgram(
     const HW::PixelShaderInfo& regs, const HW::ShaderRegisters& sh,
     std::span<const Prospero::ColorComponentMapping, 8> target_export_mapping,
@@ -891,6 +963,7 @@ void ShaderDbgDumpInputInfo(const ShaderPixelInputInfo& info) {
 	     "\t ps_system_input_base = %u\n"
 	     "\t custom_interpolation_mask = 0x%08" PRIx32 "\n"
 	     "\t ps_perspective_center_vgpr = %" PRIu32 "\n"
+	     "\t ps_perspective_centroid_vgpr = %" PRIu32 "\n"
 	     "\t ps_pos_x             = %s\n"
 	     "\t ps_pos_y             = %s\n"
 	     "\t ps_pos_z             = %s\n"
@@ -903,7 +976,8 @@ void ShaderDbgDumpInputInfo(const ShaderPixelInputInfo& info) {
 	     "\t ps_early_z           = %s\n"
 	     "\t ps_execute_on_noop   = %s\n",
 	     info.input_num, info.ps_system_input_base, info.custom_interpolation_mask,
-	     info.ps_perspective_center_vgpr, info.ps_pos_x ? "true" : "false",
+	     info.ps_perspective_center_vgpr, info.ps_perspective_centroid_vgpr,
+	     info.ps_pos_x ? "true" : "false",
 	     info.ps_pos_y ? "true" : "false", info.ps_pos_z ? "true" : "false",
 	     info.ps_pos_w ? "true" : "false", info.ps_front_face ? "true" : "false",
 	     info.ps_ancillary ? "true" : "false",

src/graphics/shader/shader.h

@@ -33,7 +33,17 @@ struct ComputeShaderInfo;
 struct ShaderRegisters;
 } // namespace HW
 
-enum class ShaderType { Unknown, Vertex, Pixel, Fetch, Compute, Mesh };
+enum class ShaderType {
+	Unknown,
+	Vertex,
+	Pixel,
+	Fetch,
+	Compute,
+	Mesh,
+	Local,
+	TessellationControl,
+	TessellationEvaluation
+};
 
 namespace ShaderRecompiler::IR {
 struct CompiledShaderInfo;
@@ -87,8 +97,11 @@ struct ShaderMeshInputInfo: ShaderWorkgroupInputInfo {
 		}
 	}
 	[[nodiscard]] constexpr uint32_t InputPrimitiveStep() const {
-		return input_primitive == static_cast<uint32_t>(Prospero::PrimitiveType::kTriStrip)
-		           ? 1u : InputPrimitiveSize();
+		switch (static_cast<Prospero::PrimitiveType>(input_primitive)) {
+			case Prospero::PrimitiveType::kTriFan:
+			case Prospero::PrimitiveType::kTriStrip: return 1u;
+			default: return InputPrimitiveSize();
+		}
 	}
 	[[nodiscard]] constexpr uint32_t InputPrimitiveCount(uint32_t vertices) const {
 		const auto size = InputPrimitiveSize();
@@ -99,6 +112,16 @@ struct ShaderMeshInputInfo: ShaderWorkgroupInputInfo {
 	}
 };
 
+struct ShaderTessellationInputInfo {
+	uint32_t input_control_points  = 0;
+	uint32_t output_control_points = 0;
+	uint32_t ls_stride             = 0;
+	uint32_t hs_stride             = 0;
+	uint32_t domain                = 0;
+	uint32_t partitioning          = 0;
+	uint32_t output_topology       = 0;
+};
+
 struct ShaderVertexInputInfo {
 	static constexpr int RES_MAX = 32;
 
@@ -106,6 +129,7 @@ struct ShaderVertexInputInfo {
 	ShaderVertexDestination resources_dst[RES_MAX];
 	ShaderVertexInputBuffer buffers[RES_MAX];
 	ShaderStageRuntime      stage;
+	ShaderType                  logical_stage        = ShaderType::Vertex;
 	int                     resources_num       = 0;
 	int                     fetch_attrib_reg    = 0;
 	int                     fetch_buffer_reg    = 0;
@@ -114,6 +138,7 @@ struct ShaderVertexInputInfo {
 	uint32_t                pa_cl_vs_out_cntl    = 0;
 	ShaderClipSpaceTransform clip_space;
 	ShaderMeshInputInfo      mesh;
+	ShaderTessellationInputInfo tess;
 	bool                    fetch_external      = false;
 	bool                    fetch_embedded      = false;
 };
@@ -131,9 +156,11 @@ struct ShaderComputeInputInfo: ShaderWorkgroupInputInfo {
 struct ShaderPixelInputInfo {
 	uint32_t                                       interpolator_settings[32]    = {0};
 	uint32_t                                       input_num                    = 0;
+	uint32_t                                       wave_size                    = 64;
 	uint32_t                                       ps_system_input_base         = 0;
 	uint32_t                                       custom_interpolation_mask    = 0;
 	uint32_t                                       ps_perspective_center_vgpr   = UINT32_MAX;
+	uint32_t                                       ps_perspective_centroid_vgpr = UINT32_MAX;
 	uint8_t                                        target_output_mode[8]        = {};
 	std::array<Prospero::ColorComponentMapping, 8> target_export_mapping        = {};
 	uint32_t                                       scratch_size_dwords          = 0;

src/graphics/shader/shaderCompiler.h

@@ -30,6 +30,9 @@ void BuildStageStaticKey(const ShaderComputeInputInfo& input_info, std::vector<u
 
 ShaderParams PrepareProgram(const HW::VertexShaderInfo& regs, const HW::Context& context,
                             const HW::UserConfig& user_config, ShaderVertexInputInfo& input_info);
+std::array<ShaderParams, 3>
+PrepareTessellationPrograms(const HW::VertexShaderInfo& regs, const HW::Context& context,
+                            std::array<ShaderVertexInputInfo, 3>& input_info);
 ShaderParams PrepareProgram(
     const HW::PixelShaderInfo& regs, const HW::ShaderRegisters& sh,
     std::span<const Prospero::ColorComponentMapping, 8> target_export_mapping,

src/graphics/shader/shaderPixelParameter.cpp

@@ -19,19 +19,31 @@ uint32_t ShaderPixelParameterLocation(const ShaderPixelInputInfo& info,
                                       std::span<const uint32_t> active_inputs, uint32_t input) {
 	std::array<bool, 32> used_locations {};
 	for (const auto active_input: active_inputs) {
-		auto location = ShaderPixelParameterMappedLocation(info, active_input);
-		if (location < used_locations.size() && used_locations[location]) {
-			location = active_input;
-			while (location < used_locations.size() && used_locations[location]) {
-				location++;
+		used_locations[ShaderPixelParameterMappedLocation(info, active_input)] = true;
+	}
+	std::array<uint32_t, 64> group_locations;
+	group_locations.fill(UINT32_MAX);
+	for (const auto active_input: active_inputs) {
+		const auto mapped = ShaderPixelParameterMappedLocation(info, active_input);
+		const auto group  = mapped * 2u + ShaderPixelParameterIsFlat(info, active_input);
+		auto&      location = group_locations[group];
+		if (location == UINT32_MAX) {
+			location = mapped;
+			// Smooth and custom interpolation read the same vertex output. Only
+			// differing flat/smooth rectangle outputs need separate locations.
+			if (group_locations[group ^ 1u] != UINT32_MAX) {
+				location = 0;
+				while (location < used_locations.size() && used_locations[location]) {
+					location++;
+				}
+				EXIT_NOT_IMPLEMENTED(location >= used_locations.size());
 			}
-			EXIT_NOT_IMPLEMENTED(location >= used_locations.size());
+			used_locations[location] = true;
 		}
 
 		if (active_input == input) {
 			return location;
 		}
-		used_locations[location] = true;
 	}
 	return ShaderPixelParameterMappedLocation(info, input);
 }

src/kernel/fileSystem.cpp

@@ -458,6 +458,7 @@ int KYTY_SYSV_ABI KernelOpen(const char* path, int flags, uint16_t mode) {
 	file->real_name = g_mount_points->ResolvePath(file->name);
 
 	if (trunc && rw_mode == Common::File::Mode::Read) {
+		g_files->DeleteDescriptor(descriptor);
 		return KERNEL_ERROR_EACCES;
 	}
 

src/kernel/memoryAddressSpace.inc

@@ -346,7 +346,9 @@ public:
 				auto entry = std::prev(next);
 				if (vaddr + size <= entry->second.vaddr + entry->second.size) {
 					old = entry->second;
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 					m_maps.erase(entry);
+#endif
 				}
 			}
 		}
@@ -355,6 +357,12 @@ public:
 			return false;
 		}
 
+		const uint64_t left_size  = vaddr - old.vaddr;
+		const uint64_t right_addr = vaddr + size;
+		const uint64_t old_end    = old.vaddr + old.size;
+		const uint64_t right_size = old_end - right_addr;
+
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 		const bool exact_unmap = (vaddr == old.vaddr && size == old.size);
 		if (exact_unmap) {
 			const bool unmapped =
@@ -370,12 +378,6 @@ public:
 			return true;
 		}
 
-		const uint64_t left_size  = vaddr - old.vaddr;
-		const uint64_t right_addr = vaddr + size;
-		const uint64_t old_end    = old.vaddr + old.size;
-		const uint64_t right_size = old_end - right_addr;
-
-#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 		if (UnmapViewPreservePlaceholder(reinterpret_cast<void*>(old.vaddr), old.size)) {
 			bool ok = true;
 			if (left_size != 0 && left_size != old.size) {
@@ -418,39 +420,28 @@ public:
 		m_maps.emplace(old.vaddr, old);
 		return false;
 #else
-		if (!UnmapViewPreservePlaceholder(reinterpret_cast<void*>(old.vaddr), old.size)) {
-			std::lock_guard lock(m_mutex);
-			m_maps.emplace(old.vaddr, old);
+		// Replace only the requested range so concurrent accesses and GPU watch
+		// protections on the surviving mappings remain intact.
+		if (!UnmapViewPreservePlaceholder(reinterpret_cast<void*>(vaddr), size)) {
 			return false;
 		}
 
-		bool ok = true;
+		std::lock_guard lock(m_mutex);
+		m_maps.erase(old.vaddr);
 		if (left_size != 0) {
-			ok = MapExistingPlaceholderFixed(old.vaddr, left_size, old.backing_offset, old.mode) &&
-			     ok;
+			m_maps.emplace(old.vaddr,
+			               MapEntry {old.vaddr, left_size, old.backing_offset, old.mode});
 		}
 		if (right_size != 0) {
-			ok = MapExistingPlaceholderFixed(right_addr, right_size,
-			                                 old.backing_offset + (right_addr - old.vaddr),
-			                                 old.mode) &&
-			     ok;
-		}
-		if (!ok) {
-			if (left_size != 0 && Contains(old.vaddr, left_size)) {
-				EXIT_IF(!UnmapOne(old.vaddr, left_size));
-			}
-			if (right_size != 0 && Contains(right_addr, right_size)) {
-				EXIT_IF(!UnmapOne(right_addr, right_size));
-			}
-			EXIT_IF(
-			    !MapExistingPlaceholderFixed(old.vaddr, old.size, old.backing_offset, old.mode));
-			return false;
+			m_maps.emplace(right_addr,
+			               MapEntry {right_addr, right_size,
+			                         old.backing_offset + right_addr - old.vaddr, old.mode});
 		}
 
 		if (placeholder_preserved != nullptr) {
 			*placeholder_preserved = true;
 		}
-		return ok;
+		return true;
 #endif
 	}
 

src/kernel/pthread.cpp

@@ -3093,10 +3093,6 @@ Pthread PthreadSwapSelfForSignal(Pthread thread) {
 	return previous;
 }
 
-int PthreadGetUniqueId(Pthread thread) {
-	return thread != nullptr ? thread->unique_id : 0;
-}
-
 uint64_t PthreadGetHostThreadId(Pthread thread) {
 	return thread != nullptr ? thread->host_thread_id : 0;
 }

src/kernel/pthread.h

@@ -110,7 +110,6 @@ int KYTY_SYSV_ABI     PthreadRename(Pthread thread, const char* name);
 void KYTY_SYSV_ABI    PthreadYield();
 int KYTY_SYSV_ABI     PthreadGetthreadid();
 int KYTY_SYSV_ABI     PthreadGetaffinity(Pthread thread, KernelCpumask* mask);
-int                   PthreadGetUniqueId(Pthread thread);
 uint64_t              PthreadGetHostThreadId(Pthread thread);
 void                  PthreadWakeForSignal(Pthread thread);
 void                  PthreadQueuePendingSignal(Pthread thread, int signum);

src/kernel/syncOnAddress.cpp

@@ -47,11 +47,17 @@ struct WaitDeadline {
 	Clock::time_point end {};
 };
 
+[[nodiscard]] WaitDeadline MakeDeadline(std::chrono::nanoseconds timeout) {
+	const auto now = Clock::now();
+	const auto remaining = Clock::time_point::max() - now;
+	return {true, timeout >= remaining ? Clock::time_point::max() : now + timeout};
+}
+
 [[nodiscard]] WaitDeadline MakeDeadline(const uint32_t* timeout_micros) {
 	if (timeout_micros == nullptr) {
 		return {};
 	}
-	return {true, Clock::now() + std::chrono::microseconds(*timeout_micros)};
+	return MakeDeadline(std::chrono::microseconds(*timeout_micros));
 }
 
 [[nodiscard]] uint32_t GetWaitSliceMicros(const WaitDeadline& deadline, bool first_wait) {
@@ -65,7 +71,7 @@ struct WaitDeadline {
 	}
 
 	const auto remaining =
-	    std::chrono::duration_cast<std::chrono::microseconds>(deadline.end - now).count();
+	    std::chrono::ceil<std::chrono::microseconds>(deadline.end - now).count();
 	return static_cast<uint32_t>(std::min<int64_t>(remaining, SIGNAL_POLL_MICROS));
 }
 
@@ -78,9 +84,8 @@ void PollSignals(signal_poll_func_t signal_poll) {
 #if KYTY_PLATFORM == KYTY_PLATFORM_LINUX && !defined(__APPLE__)
 
 template <typename T>
-int WaitLinux(volatile T* address, T expected, const uint32_t* timeout_micros,
+int WaitLinux(volatile T* address, T expected, const WaitDeadline& deadline,
               signal_poll_func_t signal_poll) {
-	const auto deadline   = MakeDeadline(timeout_micros);
 	bool       first_wait = true;
 
 	for (;;) {
@@ -185,11 +190,10 @@ void UnregisterPortableWaiter(volatile void*                               addre
 }
 
 template <typename T>
-int WaitPortable(volatile T* address, T expected, const uint32_t* timeout_micros,
+int WaitPortable(volatile T* address, T expected, const WaitDeadline& deadline,
                  signal_poll_func_t signal_poll) {
 	PortableWaiter waiter;
 	auto           entry      = RegisterPortableWaiter(address, &waiter);
-	const auto     deadline   = MakeDeadline(timeout_micros);
 	bool           first_wait = true;
 	int            result     = OK;
 
@@ -250,7 +254,7 @@ int WakePortable(volatile void* address, int32_t count) {
 }
 
 template <typename T>
-int WaitImpl(volatile T* address, T expected, const uint32_t* timeout_micros,
+int WaitImpl(volatile T* address, T expected, const WaitDeadline& deadline,
              signal_poll_func_t signal_poll) {
 	if (!IsValidWaitAddress(address)) {
 		return KERNEL_ERROR_EINVAL;
@@ -258,9 +262,9 @@ int WaitImpl(volatile T* address, T expected, const uint32_t* timeout_micros,
 
 	int result = OK;
 #if KYTY_PLATFORM == KYTY_PLATFORM_LINUX && !defined(__APPLE__)
-	result = WaitLinux(address, expected, timeout_micros, signal_poll);
+	result = WaitLinux(address, expected, deadline, signal_poll);
 #else
-	result = WaitPortable(address, expected, timeout_micros, signal_poll);
+	result = WaitPortable(address, expected, deadline, signal_poll);
 #endif
 	PollSignals(signal_poll);
 	return result;
@@ -270,12 +274,17 @@ int WaitImpl(volatile T* address, T expected, const uint32_t* timeout_micros,
 
 int Wait32(volatile uint32_t* address, uint32_t expected, const uint32_t* timeout_micros,
            signal_poll_func_t signal_poll) {
-	return WaitImpl(address, expected, timeout_micros, signal_poll);
+	return WaitImpl(address, expected, MakeDeadline(timeout_micros), signal_poll);
 }
 
 int Wait64(volatile uint64_t* address, uint64_t expected, const uint32_t* timeout_micros,
            signal_poll_func_t signal_poll) {
-	return WaitImpl(address, expected, timeout_micros, signal_poll);
+	return WaitImpl(address, expected, MakeDeadline(timeout_micros), signal_poll);
+}
+
+int Wait64(volatile uint64_t* address, uint64_t expected, std::chrono::nanoseconds timeout,
+           signal_poll_func_t signal_poll) {
+	return WaitImpl(address, expected, MakeDeadline(timeout), signal_poll);
 }
 
 int Wake(volatile void* address, int32_t count) {

src/kernel/syncOnAddress.h

@@ -4,6 +4,8 @@
 #include "common/abi.h"
 #include "common/common.h"
 
+#include <chrono>
+
 namespace Libs::LibKernel::SyncOnAddress {
 
 using signal_poll_func_t = void (*)();
@@ -12,6 +14,8 @@ int Wait32(volatile uint32_t* address, uint32_t expected, const uint32_t* timeou
            signal_poll_func_t signal_poll = nullptr);
 int Wait64(volatile uint64_t* address, uint64_t expected, const uint32_t* timeout_micros,
            signal_poll_func_t signal_poll = nullptr);
+int Wait64(volatile uint64_t* address, uint64_t expected, std::chrono::nanoseconds timeout,
+           signal_poll_func_t signal_poll = nullptr);
 int Wake(volatile void* address, int32_t count);
 
 } // namespace Libs::LibKernel::SyncOnAddress

src/launcher/CMakeLists.txt

@@ -20,7 +20,7 @@ add_executable(launcher WIN32
 )
 
 
-target_link_libraries(launcher Qt6::Concurrent Qt6::Network Qt6::Widgets Vulkan::Headers kyty_git_version)
+target_link_libraries(launcher Qt6::Concurrent Qt6::Network Qt6::Widgets Vulkan::Headers SDL2-static kyty_git_version)
 if(WIN32)
 target_link_libraries(launcher setupapi)
 endif()

src/launcher/forms/configuration_edit_dialog.ui

@@ -75,7 +75,25 @@
        </layout>
       </widget>
      </item>
-     <item row="0" column="1" rowspan="2">
+     <item row="2" column="0">
+      <widget class="QGroupBox" name="audio_group">
+       <property name="title"><string>Audio</string></property>
+       <layout class="QFormLayout" name="audioLayout">
+        <property name="fieldGrowthPolicy">
+         <enum>QFormLayout::AllNonFixedFieldsGrow</enum>
+        </property>
+        <item row="0" column="0">
+         <widget class="QLabel" name="label_audio_input_device">
+          <property name="text"><string>Microphone:</string></property>
+         </widget>
+        </item>
+        <item row="0" column="1">
+         <widget class="QComboBox" name="comboBox_audio_input_device"/>
+        </item>
+       </layout>
+      </widget>
+     </item>
+     <item row="0" column="1" rowspan="3">
       <widget class="QGroupBox" name="graphics_group">
        <property name="title">
         <string>Graphics &amp;&amp; display</string>
@@ -152,7 +170,7 @@
        </layout>
       </widget>
      </item>
-     <item row="2" column="0" colspan="2">
+     <item row="3" column="0" colspan="2">
       <widget class="QGroupBox" name="debug_group">
        <property name="title">
         <string>Debugging &amp;&amp; logs</string>
@@ -253,21 +271,19 @@
        <property name="title">
         <string>Compatibility</string>
        </property>
-       <layout class="QHBoxLayout" name="compatibilityLayout">
+       <layout class="QVBoxLayout" name="compatibilityLayout">
+        <item>
+         <widget class="QCheckBox" name="checkBox_amd_cpu">
+          <property name="toolTip"><string>Patch CPU instructions that differ between AMD and the host CPU.</string></property>
+          <property name="text"><string>AMD CPU patch (EXPERIMENTAL)</string></property>
+         </widget>
+        </item>
         <item>
          <widget class="QCheckBox" name="checkBox_red_zone_protection">
           <property name="toolTip"><string>Helps prevent crashes in some games on Windows</string></property>
           <property name="text"><string>Windows SysV red zone crash protection (Experimental)</string></property>
          </widget>
         </item>
-        <item>
-         <spacer name="compatibilitySpacer">
-          <property name="orientation"><enum>Qt::Horizontal</enum></property>
-          <property name="sizeHint" stdset="0">
-           <size><width>40</width><height>20</height></size>
-          </property>
-         </spacer>
-        </item>
        </layout>
       </widget>
      </item>

src/launcher/forms/configuration_list_widget.ui

@@ -218,6 +218,11 @@
            <string>Firmware Version</string>
           </property>
          </column>
+         <column>
+          <property name="text">
+           <string>Size</string>
+          </property>
+         </column>
          <column>
           <property name="text">
            <string>Path</string>

src/launcher/include/configuration.h

@@ -87,6 +87,7 @@ public:
 	Resolution             screen_resolution           = Resolution::R1280X720;
 	QString                user_name                   = "Kyty";
 	int                    user_id                     = Config::DEFAULT_USER_ID;
+	QString                audio_input_device;
 	PresentMode            present_mode                = PresentMode::Mailbox;
 	int                    gpu_index                   = -1;
 	bool                   fullscreen_enabled          = false;
@@ -104,6 +105,7 @@ public:
 	QString                printf_output_file          = "_kyty.txt";
 	bool                   profiler_enabled            = false;
 	bool                   renderdoc_enabled           = false;
+	bool                   amd_cpu_enabled             = false;
 #if defined(_WIN32)
 	bool red_zone_protection_enabled = false;
 #endif
@@ -115,6 +117,7 @@ public:
 		screen_resolution           = other.screen_resolution;
 		user_name                   = other.user_name;
 		user_id                     = other.user_id;
+		audio_input_device          = other.audio_input_device;
 		present_mode                = other.present_mode;
 		gpu_index                   = other.gpu_index;
 		fullscreen_enabled          = other.fullscreen_enabled;
@@ -132,6 +135,7 @@ public:
 		printf_output_file          = other.printf_output_file;
 		profiler_enabled            = other.profiler_enabled;
 		renderdoc_enabled           = other.renderdoc_enabled;
+		amd_cpu_enabled             = other.amd_cpu_enabled;
 #if defined(_WIN32)
 		red_zone_protection_enabled = other.red_zone_protection_enabled;
 #endif
@@ -160,6 +164,7 @@ public:
 		KYTY_CFG_SET(screen_resolution);
 		KYTY_CFG_SET(user_name);
 		KYTY_CFG_SET(user_id);
+		KYTY_CFG_SET(audio_input_device);
 		KYTY_CFG_SET(present_mode);
 		KYTY_CFG_SET(gpu_index);
 		KYTY_CFG_SET(fullscreen_enabled);
@@ -177,6 +182,7 @@ public:
 		KYTY_CFG_SET(printf_output_file);
 		KYTY_CFG_SET(profiler_enabled);
 		KYTY_CFG_SET(renderdoc_enabled);
+		KYTY_CFG_SET(amd_cpu_enabled);
 #if defined(_WIN32)
 		KYTY_CFG_SET(red_zone_protection_enabled);
 #endif
@@ -196,6 +202,7 @@ public:
 		user_id            = user_id_ok && Config::IsConfiguredUserIdValid(saved_user_id)
 		                         ? saved_user_id
 		                         : Config::DEFAULT_USER_ID;
+		audio_input_device = s->value("audio_input_device", audio_input_device).toString();
 		KYTY_CFG_GET(present_mode);
 		gpu_index = s->value("gpu_index", -1).toInt();
 		if (EnumToText(present_mode).isEmpty()) {
@@ -219,6 +226,7 @@ public:
 		KYTY_CFG_GET(printf_output_file);
 		KYTY_CFG_GET(profiler_enabled);
 		KYTY_CFG_GET(renderdoc_enabled);
+		amd_cpu_enabled = s->value("amd_cpu_enabled", false).toBool();
 #if defined(_WIN32)
 		red_zone_protection_enabled =
 		    s->value("red_zone_protection_enabled", red_zone_protection_enabled).toBool();

src/launcher/src/configurationEditDialog.cpp

@@ -3,6 +3,7 @@
 #include "common/emulatorConfig.h"
 #include "configuration.h"
 #include "mandatoryLineEdit.h"
+#include "SDL.h"
 
 #include <QAbstractItemView>
 #include <QCheckBox>
@@ -167,6 +168,31 @@ void ConfigurationEditDialog::Init(const Configuration& info) {
 	m_ui->lineEdit_user_name->setMaxLength(static_cast<int>(Config::MAX_USER_NAME_LENGTH));
 	m_ui->lineEdit_user_name->setText(info.user_name);
 	m_ui->spinBox_user_id->setValue(info.user_id);
+	auto* microphone = m_ui->comboBox_audio_input_device;
+	microphone->clear();
+	microphone->addItem(tr("None"), QString {});
+	microphone->setToolTip(tr("Microphone used by games. None supplies silence."));
+	SDL_SetMainReady();
+	if (SDL_InitSubSystem(SDL_INIT_AUDIO) == 0) {
+		const int device_count = SDL_GetNumAudioDevices(SDL_TRUE);
+		for (int i = 0; i < device_count; i++) {
+			if (const auto* device = SDL_GetAudioDeviceName(i, SDL_TRUE); device != nullptr) {
+				const auto name = QString::fromUtf8(device);
+				if (microphone->findData(name) < 0) {
+					microphone->addItem(name, name);
+				}
+			}
+		}
+		SDL_QuitSubSystem(SDL_INIT_AUDIO);
+	} else {
+		microphone->setToolTip(tr("Microphones could not be listed: %1")
+		                           .arg(QString::fromUtf8(SDL_GetError())));
+	}
+	if (microphone->findData(info.audio_input_device) < 0) {
+		microphone->addItem(tr("%1 (unavailable)").arg(info.audio_input_device),
+		                    info.audio_input_device);
+	}
+	microphone->setCurrentIndex(microphone->findData(info.audio_input_device));
 	ListInit(m_ui->comboBox_screen_resolution, info.screen_resolution);
 	ListInit(m_ui->comboBox_present_mode, info.present_mode);
 	m_ui->comboBox_gpu->clear();
@@ -187,7 +213,7 @@ void ConfigurationEditDialog::Init(const Configuration& info) {
 #endif
 	QVulkanInstance instance;
 	instance.setApiVersion(QVersionNumber(1, 3, 0));
-#if !defined(__APPLE__)
+#if !defined(__APPLE__) && QT_VERSION >= QT_VERSION_CHECK(6, 6, 0)
 	instance.setFlags(QVulkanInstance::NoPortabilityDrivers);
 #endif
 	if (instance.create()) {
@@ -213,6 +239,10 @@ void ConfigurationEditDialog::Init(const Configuration& info) {
 	m_ui->checkBox_shader_validation->setChecked(info.shader_validation_enabled);
 	m_ui->checkBox_vulkan_validation->setChecked(info.vulkan_validation_enabled);
 	m_ui->checkBox_renderdoc_capture->setChecked(info.renderdoc_enabled);
+	m_ui->checkBox_amd_cpu->setChecked(info.amd_cpu_enabled);
+#if defined(__APPLE__)
+	m_ui->checkBox_amd_cpu->setVisible(false);
+#endif
 #if defined(_WIN32)
 	m_ui->checkBox_red_zone_protection->setChecked(info.red_zone_protection_enabled);
 #else
@@ -340,6 +370,7 @@ void ConfigurationEditDialog::resizeEvent(QResizeEvent* event) {
 static void UpdateInfo(Configuration& info, Ui::ConfigurationEditDialog& ui) {
 	info.user_name = ui.lineEdit_user_name->text().trimmed();
 	info.user_id   = ui.spinBox_user_id->value();
+	info.audio_input_device = ui.comboBox_audio_input_device->currentData().toString();
 	info.screen_resolution =
 	    TextToEnum<Configuration::Resolution>(ui.comboBox_screen_resolution->currentText());
 	info.present_mode =
@@ -352,6 +383,7 @@ static void UpdateInfo(Configuration& info, Ui::ConfigurationEditDialog& ui) {
 	info.vulkan_validation_enabled = ui.checkBox_vulkan_validation->isChecked();
 	info.shader_validation_enabled = ui.checkBox_shader_validation->isChecked();
 	info.renderdoc_enabled         = ui.checkBox_renderdoc_capture->isChecked();
+	info.amd_cpu_enabled           = ui.checkBox_amd_cpu->isChecked();
 #if defined(_WIN32)
 	info.red_zone_protection_enabled = ui.checkBox_red_zone_protection->isChecked();
 #endif

src/launcher/src/configurationItem.cpp

@@ -5,12 +5,15 @@
 #include <QApplication>
 #include <QComboBox>
 #include <QDir>
+#include <QDirIterator>
 #include <QFileInfo>
 #include <QFont>
+#include <QFutureWatcher>
 #include <QHBoxLayout>
 #include <QIcon>
 #include <QLabel>
 #include <QLineEdit>
+#include <QLocale>
 #include <QSize>
 #include <QStringList>
 #include <QStyle>
@@ -18,6 +21,7 @@
 #include <QVersionNumber>
 #include <QWheelEvent>
 #include <QWidget>
+#include <QtConcurrentRun>
 
 namespace {
 
@@ -26,6 +30,7 @@ enum Column {
 	SerialColumn,
 	GameVersionColumn,
 	FirmwareVersionColumn,
+	SizeColumn,
 	PathColumn,
 	StatusColumn,
 	CommentsColumn,
@@ -155,6 +160,34 @@ ConfigurationItem::ConfigurationItem(std::unique_ptr<Configuration> info, QTreeW
 
 	Update();
 	SetRunning(false);
+
+	setData(SizeColumn, Qt::UserRole, qint64(-1));
+	setText(SizeColumn, QStringLiteral("\u2014"));
+	auto* watcher = new QFutureWatcher<qint64>(this);
+	connect(watcher, &QFutureWatcher<qint64>::finished, this, [this, watcher]() {
+		const auto bytes = watcher->result();
+		setData(SizeColumn, Qt::UserRole, bytes);
+		if (bytes >= 0) {
+			setText(SizeColumn,
+			        QLocale().formattedDataSize(bytes, 2, QLocale::DataSizeTraditionalFormat));
+		}
+		watcher->deleteLater();
+	});
+	watcher->setFuture(QtConcurrent::run([path = m_info->basedir]() -> qint64 {
+		if (path.isEmpty() || !QDir(path).exists()) {
+			return -1;
+		}
+		qint64       bytes = 0;
+		QDirIterator files(path, QDir::Files | QDir::Hidden | QDir::System | QDir::NoSymLinks,
+		                   QDirIterator::Subdirectories);
+		while (files.hasNext()) {
+			files.next();
+			if (files.fileInfo().isFile()) {
+				bytes += files.fileInfo().size();
+			}
+		}
+		return bytes;
+	}));
 }
 
 ConfigurationItem::~ConfigurationItem() = default;
@@ -193,6 +226,9 @@ bool ConfigurationItem::operator<(const QTreeWidgetItem& other) const {
 	const int column = treeWidget() != nullptr ? treeWidget()->sortColumn() : NameColumn;
 	switch (column) {
 		case NameColumn: return GetSortText(*m_info) < GetSortText(*other_item->m_info);
+		case SizeColumn:
+			return data(SizeColumn, Qt::UserRole).toLongLong() <
+			       other.data(SizeColumn, Qt::UserRole).toLongLong();
 		case StatusColumn:
 			return GetStatusText(m_info->game_status) <
 			       GetStatusText(other_item->m_info->game_status);

src/launcher/src/configurationListWidget.cpp

@@ -61,9 +61,10 @@ constexpr int GAME_NAME_COLUMN             = 0;
 constexpr int GAME_SERIAL_COLUMN           = 1;
 constexpr int GAME_VERSION_COLUMN          = 2;
 constexpr int GAME_FIRMWARE_VERSION_COLUMN = 3;
-constexpr int GAME_PATH_COLUMN             = 4;
-constexpr int GAME_STATUS_COLUMN           = 5;
-constexpr int GAME_COMMENT_COLUMN          = 6;
+constexpr int GAME_SIZE_COLUMN             = 4;
+constexpr int GAME_PATH_COLUMN             = 5;
+constexpr int GAME_STATUS_COLUMN           = 6;
+constexpr int GAME_COMMENT_COLUMN          = 7;
 
 static QString NormalizeGameDirectory(const QString& dir) {
 	const auto trimmed = dir.trimmed();
@@ -217,6 +218,7 @@ ConfigurationListWidget::ConfigurationListWidget(QWidget* parent)
 	m_ui->cfgs_list->setColumnWidth(GAME_SERIAL_COLUMN, 110);
 	m_ui->cfgs_list->setColumnWidth(GAME_VERSION_COLUMN, 120);
 	m_ui->cfgs_list->setColumnWidth(GAME_FIRMWARE_VERSION_COLUMN, 150);
+	m_ui->cfgs_list->setColumnWidth(GAME_SIZE_COLUMN, 100);
 	m_ui->cfgs_list->setColumnWidth(GAME_PATH_COLUMN, 320);
 	m_ui->cfgs_list->setColumnWidth(GAME_STATUS_COLUMN, 150);
 	m_ui->cfgs_list->setColumnWidth(GAME_COMMENT_COLUMN, 240);

src/launcher/src/inputMappingDialog.cpp

@@ -46,8 +46,8 @@ constexpr PadControl PAD_CONTROLS[] = {
     {"Square", "Square", "K"},
     {"L1", "L1", "Q"},
     {"R1", "R1", "E"},
-    {"L2", "L2", ""},
-    {"R2", "R2", ""},
+    {"L2", "L2", "Z"},
+    {"R2", "R2", "C"},
     {"L3", "L3", "Left Shift"},
     {"R3", "R3", "Left Ctrl"},
     {"Options", "Options", "Return"},

src/launcher/src/mainDialog.cpp

@@ -228,6 +228,9 @@ static QStringList CreateEmulatorArgs(const Configuration& info) {
 	args << "--screen-height" << r.at(1);
 	args << "--user-name" << info.user_name;
 	args << "--user-id" << QString::number(info.user_id);
+	if (!info.audio_input_device.isEmpty()) {
+		args << "--mic" << info.audio_input_device;
+	}
 	args << "--present-mode" << EnumToText(info.present_mode);
 	if (info.gpu_index >= 0) {
 		args << "--gpu" << QString::number(info.gpu_index);
@@ -251,6 +254,9 @@ static QStringList CreateEmulatorArgs(const Configuration& info) {
 		args << "--profile";
 	}
 	args << "--spirv-debug-printf" << "false";
+	if (info.amd_cpu_enabled) {
+		args << "--amd-cpu";
+	}
 #if defined(_WIN32)
 	if (info.red_zone_protection_enabled) {
 		args << "--redzone";

src/libs/agc.cpp

@@ -262,35 +262,25 @@ struct CommandBuffer {
 		     reserved_dw);
 	}
 
-	[[nodiscard]] KYTY_SYSV_ABI uint32_t GetAvailableSizeDW() const {
-		if (cursor_up == nullptr || cursor_down == nullptr || cursor_down <= cursor_up) {
-			return 0;
-		}
-
-		auto available = static_cast<uint64_t>(cursor_down - cursor_up);
+	[[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) -
+		                                           reinterpret_cast<uintptr_t>(cursor_up));
+		const auto available = static_cast<uint64_t>(distance >> 2);
 		if (available <= reserved_dw) {
 			return 0;
 		}
-		if (available - reserved_dw > UINT32_MAX) {
-			LOGF_COLOR(
-			    Log::Color::Red,
-			    "\t command buffer has suspiciously large free space: cursor_up = 0x%016" PRIx64
-			    ", cursor_down = 0x%016" PRIx64 ", reserved_dw = %" PRIu32 "\n",
-			    reinterpret_cast<uint64_t>(cursor_up), reinterpret_cast<uint64_t>(cursor_down),
-			    reserved_dw);
-			return UINT32_MAX;
-		}
-		return static_cast<uint32_t>(available - reserved_dw);
+		return available - reserved_dw;
 	}
 
 	KYTY_SYSV_ABI bool ReserveDW(uint32_t num_dw) {
-		uint32_t remaining = GetAvailableSizeDW();
+		const uint64_t remaining = GetAvailableSizeDW();
 		if (num_dw > remaining) {
 			if (callback == nullptr) {
 				LOGF_COLOR(
 				    Log::Color::Red,
 				    "\t command buffer exhausted and has no grow callback: requested = %" PRIu32
-				    ", remaining = %" PRIu32 ", reserved_dw = %" PRIu32 "\n",
+				    ", remaining = %" PRIu64 ", reserved_dw = %" PRIu32 "\n",
 				    num_dw, remaining, reserved_dw);
 				DbgDump();
 				return false;
@@ -300,7 +290,7 @@ struct CommandBuffer {
 			if (!result) {
 				LOGF_COLOR(Log::Color::Red,
 				           "\t command buffer grow callback failed: requested = %" PRIu32
-				           ", remaining = %" PRIu32 ", reserved_dw = %" PRIu32 "\n",
+				           ", remaining = %" PRIu64 ", reserved_dw = %" PRIu32 "\n",
 				           num_dw, remaining, reserved_dw);
 				DbgDump();
 				return false;
@@ -308,7 +298,7 @@ struct CommandBuffer {
 			if (GetAvailableSizeDW() < num_dw) {
 				LOGF_COLOR(Log::Color::Red,
 				           "\t command buffer grow callback did not provide enough space: "
-				           "requested = %" PRIu32 ", remaining = %" PRIu32
+				           "requested = %" PRIu32 ", remaining = %" PRIu64
 				           ", reserved_dw = %" PRIu32 "\n",
 				           num_dw, GetAvailableSizeDW(), reserved_dw);
 				DbgDump();