texture cache: convert D16<->D24/D32 depth copies instead of crashing (#1873)

gpu: convert guest D16 depth between native D16 and stenciled four-byte host depth

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

.github/workflows/build.yml

@@ -32,7 +32,7 @@ jobs:
   windows:
     name: Build KytyPS5 (Windows)
     needs: metadata
-    runs-on: windows-2022
+    runs-on: windows-2025
 
     steps:
       - name: Checkout repository

3rdparty/CMakeLists.txt

@@ -42,6 +42,28 @@ if(NOT KYTY_CLANG_CL)
 endif()
 install(FILES "${openxr_SOURCE_DIR}/LICENSE" DESTINATION "licenses/openxr")
 
+# Noto Sans v2.015 (SIL Open Font License 1.1) is libFont's fallback font.
+set(KYTY_NOTO_SANS_REPOSITORY "https://raw.githubusercontent.com/notofonts/notofonts.github.io")
+set(KYTY_NOTO_SANS_COMMIT 42af67ccd12bc0e31ed85384bf73e39248265374)
+set(KYTY_NOTO_SANS_BASE_URL "${KYTY_NOTO_SANS_REPOSITORY}/${KYTY_NOTO_SANS_COMMIT}/fonts")
+set(KYTY_NOTO_SANS_DIR "${FETCHCONTENT_BASE_DIR}/noto-sans")
+function(kyty_fetch_noto_sans path sha256)
+	get_filename_component(name "${path}" NAME)
+	file(DOWNLOAD "${KYTY_NOTO_SANS_BASE_URL}/${path}" "${KYTY_NOTO_SANS_DIR}/${name}"
+		EXPECTED_HASH SHA256=${sha256})
+endfunction()
+kyty_fetch_noto_sans(LICENSE f2095b08bed08b23a6fe26112fcd679a2bee3f002eef077eb05d215ed1051bd8)
+kyty_fetch_noto_sans(NotoSans/hinted/ttf/NotoSans-Regular.ttf
+	478c558ea716033cd60c03438f628dfa75694dcf6b5f6d505a2f05fd2b4f3823)
+kyty_fetch_noto_sans(NotoSans/hinted/ttf/NotoSans-Bold.ttf
+	1df075a380fc7cb898acf64c1f7b3b4dd780de3caa860178bf929de35817a913)
+kyty_fetch_noto_sans(NotoSans/hinted/ttf/NotoSans-Italic.ttf
+	467e3f89eeca4108bb8710a2b9e0cf2281ac56d5b0609211a83776d0505eecb5)
+kyty_fetch_noto_sans(NotoSans/hinted/ttf/NotoSans-BoldItalic.ttf
+	1b602a9d6353be42c91df097a4857b69fa2696f26703d7a33b54a15d87c2622c)
+install(FILES "${KYTY_NOTO_SANS_DIR}/LICENSE" DESTINATION "licenses/noto-sans")
+set(KYTY_NOTO_SANS_DIR "${KYTY_NOTO_SANS_DIR}" PARENT_SCOPE)
+
 # The guest instruction patcher needs a decoder/encoder and a small
 # trampoline assembler. Pin the same revisions used by shadPS4's implementation.
 set(XBYAK_BUILD_TESTS OFF CACHE BOOL "" FORCE)

CMakeLists.txt

@@ -337,7 +337,7 @@ if (KYTY_CLANG_CL)
 	set_source_files_properties(${kyty_emulator_src} PROPERTIES COMPILE_FLAGS "-Wno-pragma-pack -Wno-deprecated-declarations -D_TIMESPEC_DEFINED")
 endif()
 
-set(kyty_emulator_link_libraries common Vulkan::Headers spirv-tools-opt spirv-tools SDL3::SDL3-static imgui xxhash FFmpeg::ffmpeg fmt::fmt nlohmann_json::nlohmann_json LibAtrac9 kyty_git_version Zydis::Zydis xbyak::xbyak opus OpenXR::openxr_loader)
+set(kyty_emulator_link_libraries common Vulkan::Headers spirv-tools-opt spirv-tools SDL3::SDL3-static imgui xxhash FFmpeg::ffmpeg fmt::fmt nlohmann_json::nlohmann_json LibAtrac9 kyty_git_version Zydis::Zydis xbyak::xbyak opus OpenXR::openxr_loader libzstd_static)
 
 # Linux system libraries.
 if(LINUX)
@@ -407,6 +407,8 @@ add_executable(scalar_provenance_tests EXCLUDE_FROM_ALL
 target_link_libraries(scalar_provenance_tests fmt::fmt xbyak::xbyak)
 target_include_directories(scalar_provenance_tests PRIVATE ${inc_headers})
 
+add_executable(build_flags_tests EXCLUDE_FROM_ALL "${KYTY_TESTS_DIR}/BuildFlagsTests.cpp")
+
 add_executable(page_manager_tests EXCLUDE_FROM_ALL
 	"${KYTY_TESTS_DIR}/PageManagerTests.cpp"
 	"${KYTY_SOURCE_DIR}/graphics/host_gpu/pageManager.cpp"
@@ -451,6 +453,11 @@ add_executable(lru_cache_tests EXCLUDE_FROM_ALL
 )
 target_include_directories(lru_cache_tests PRIVATE ${inc_headers})
 
+add_executable(slot_vector_tests EXCLUDE_FROM_ALL
+	"${KYTY_TESTS_DIR}/SlotVectorTests.cpp"
+)
+target_include_directories(slot_vector_tests PRIVATE ${inc_headers})
+
 add_executable(ime_dialog_tests EXCLUDE_FROM_ALL
 	"${KYTY_TESTS_DIR}/ImeDialogTests.cpp"
 	"${KYTY_SOURCE_DIR}/libs/imeCommon.cpp"
@@ -645,6 +652,7 @@ if(BUILD_TESTING)
 	add_test(NAME image_page_table COMMAND $<TARGET_FILE:image_page_table_tests>)
 	add_test(NAME memory_tracker COMMAND $<TARGET_FILE:memory_tracker_tests>)
 	add_test(NAME page_manager COMMAND $<TARGET_FILE:page_manager_tests>)
+	add_test(NAME build_flags COMMAND $<TARGET_FILE:build_flags_tests>)
 	add_test(NAME archive_file COMMAND $<TARGET_FILE:archive_file_tests>)
 	add_test(NAME trophy_system COMMAND $<TARGET_FILE:trophy_system_tests>)
 	add_test(NAME avplayer_file COMMAND $<TARGET_FILE:avplayer_file_tests>)
@@ -738,6 +746,7 @@ if(BUILD_TESTING)
 		image_page_table_tests
 		memory_tracker_tests
 		page_manager_tests
+		build_flags_tests
 		archive_file_tests
 		trophy_system_tests
 		avplayer_file_tests
@@ -806,6 +815,19 @@ add_custom_target(kyty_touch_cursors
 add_dependencies(kyty_emulator kyty_touch_cursors)
 install(DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/assets/cursors" DESTINATION assets)
 
+set(KYTY_NOTO_SANS_FONTS
+	"${KYTY_NOTO_SANS_DIR}/NotoSans-Regular.ttf"
+	"${KYTY_NOTO_SANS_DIR}/NotoSans-Bold.ttf"
+	"${KYTY_NOTO_SANS_DIR}/NotoSans-Italic.ttf"
+	"${KYTY_NOTO_SANS_DIR}/NotoSans-BoldItalic.ttf")
+add_custom_target(kyty_noto_sans_fonts
+	COMMAND ${CMAKE_COMMAND} -E make_directory "$<TARGET_FILE_DIR:kyty_emulator>/assets/fonts"
+	COMMAND ${CMAKE_COMMAND} -E copy_if_different ${KYTY_NOTO_SANS_FONTS}
+		"$<TARGET_FILE_DIR:kyty_emulator>/assets/fonts"
+	VERBATIM)
+add_dependencies(kyty_emulator kyty_noto_sans_fonts)
+install(FILES ${KYTY_NOTO_SANS_FONTS} DESTINATION assets/fonts)
+
 target_link_libraries(kyty_emulator ${kyty_emulator_link_libraries})
 if (WIN32)
 	target_link_libraries(kyty_emulator opengl32)

src/common/additionalContent.cpp

@@ -109,4 +109,23 @@ std::vector<AddContentEntry> FindAddContent(const std::filesystem::path& dir,
 	return entries;
 }
 
+std::optional<AddContentEntry> ParseAddContent(std::string_view value) {
+	static const std::regex label_pattern("^[A-Z0-9]{16}$");
+	const auto              eq    = value.find('=');
+	const auto              label = value.substr(0, eq);
+	if (!std::regex_match(label.begin(), label.end(), label_pattern)) {
+		return std::nullopt;
+	}
+
+	AddContentEntry entry {{}, std::string(label), {}, {}};
+	if (eq != std::string_view::npos) {
+		entry.mount_dir = PathFromUtf8(FixFilenameSlash(value.substr(eq + 1)));
+		std::error_code ec;
+		if (!std::filesystem::is_directory(entry.mount_dir, ec)) {
+			return std::nullopt;
+		}
+	}
+	return entry;
+}
+
 } // namespace Common

src/common/additionalContent.h

@@ -2,7 +2,9 @@
 #define KYTY_COMMON_ADDITIONAL_CONTENT_H_
 
 #include <filesystem>
+#include <optional>
 #include <string>
+#include <string_view>
 #include <vector>
 
 namespace Common {
@@ -10,7 +12,7 @@ namespace Common {
 struct AddContentEntry {
 	std::string           content_id;
 	std::string           entitlement_label;
-	std::filesystem::path mount_dir;
+	std::filesystem::path mount_dir; // Empty for an add-on without extra data.
 	std::string           title;
 };
 
@@ -18,6 +20,9 @@ struct AddContentEntry {
 std::vector<AddContentEntry> FindAddContent(const std::filesystem::path& dir,
                                             const std::string&           title_id);
 
+// Parses a user add-on: "LABEL" without extra data, or "LABEL=DIR" with its data in DIR.
+std::optional<AddContentEntry> ParseAddContent(std::string_view value);
+
 } // namespace Common
 
 #endif /* KYTY_COMMON_ADDITIONAL_CONTENT_H_ */

src/common/emulatorConfig.cpp

@@ -133,6 +133,10 @@ bool ShaderValidationEnabled() {
 	return g_config->shader_validation_enabled;
 }
 
+bool ShaderCacheEnabled() {
+	return g_config->shader_cache_enabled;
+}
+
 ShaderOptimizationType GetShaderOptimizationType() {
 	return g_config->shader_optimization_type;
 }

src/common/emulatorConfig.h

@@ -72,6 +72,7 @@ struct ConfigOptions {
 	uint32_t               console_language            = DEFAULT_CONSOLE_LANGUAGE;
 	bool                   vulkan_validation_enabled   = false;
 	bool                   shader_validation_enabled   = false;
+	bool                   shader_cache_enabled        = false;
 	ShaderOptimizationType shader_optimization_type    = ShaderOptimizationType::None;
 	LogDirection           shader_log_direction        = LogDirection::Silent;
 	std::filesystem::path  shader_log_folder           = "_Shaders";
@@ -132,6 +133,7 @@ uint32_t GetConsoleLanguage();
 bool     VulkanValidationEnabled();
 
 bool                   ShaderValidationEnabled();
+bool                   ShaderCacheEnabled();
 ShaderOptimizationType GetShaderOptimizationType();
 LogDirection           GetShaderLogDirection();
 std::filesystem::path  GetShaderLogFolder();

src/common/slotVector.h

@@ -1,15 +1,15 @@
 #ifndef EMULATOR_SRC_COMMON_SLOTVECTOR_H_
 #define EMULATOR_SRC_COMMON_SLOTVECTOR_H_
 
-#include "common/assert.h"
-#include "common/abi.h"
+#include "common/common.h"
 
 #include <compare>
 #include <cstdint>
-#include <deque>
 #include <functional>
 #include <limits>
-#include <optional>
+#include <memory>
+#include <new>
+#include <numeric>
 #include <utility>
 #include <vector>
 
@@ -19,105 +19,105 @@ struct SlotId {
 	static constexpr uint32_t INVALID_INDEX = std::numeric_limits<uint32_t>::max();
 
 	constexpr SlotId() noexcept = default;
-	constexpr SlotId(uint32_t value) noexcept: index(value), generation(1) {}
-	constexpr SlotId(uint32_t value, uint32_t slot_generation) noexcept
-	    : index(value), generation(slot_generation) {}
+	constexpr SlotId(uint32_t index) noexcept: index(index) {}
 
-	[[nodiscard]] constexpr explicit operator bool() const noexcept { return index != INVALID_INDEX; }
+	[[nodiscard]] constexpr explicit operator bool() const noexcept {
+		return index != INVALID_INDEX;
+	}
 	constexpr auto operator<=>(const SlotId&) const noexcept = default;
 
 	uint32_t index = INVALID_INDEX;
-	uint32_t generation = 0;
 };
 
-// Stable-address slot storage for cache resources that are intentionally non-movable.
+// Fixed-size chunks keep the stored, non-movable resources in place. Occupancy is a byte per
+// slot, not a bitset: a deferred erase runs while fault threads read other slots.
 template <typename T>
 class SlotVector {
+	static constexpr uint32_t ChunkSize = 256;
+
 public:
 	SlotVector() = default;
 	KYTY_CLASS_NO_COPY(SlotVector);
 
+	~SlotVector() noexcept {
+		ForEach([](SlotId, T& value) { value.~T(); });
+	}
+
 	[[nodiscard]] T& operator[](SlotId id) noexcept {
-		EXIT_IF(!is_allocated(id));
-		return *m_values[id.index].value;
+		return m_chunks[id.index / ChunkSize][id.index % ChunkSize].object;
 	}
 
 	[[nodiscard]] const T& operator[](SlotId id) const noexcept {
-		EXIT_IF(!is_allocated(id));
-		return *m_values[id.index].value;
+		return m_chunks[id.index / ChunkSize][id.index % ChunkSize].object;
 	}
 
 	[[nodiscard]] T* try_get(SlotId id) noexcept {
-		return is_allocated(id) ? &*m_values[id.index].value : nullptr;
+		return is_allocated(id) ? &(*this)[id] : nullptr;
 	}
 
 	[[nodiscard]] const T* try_get(SlotId id) const noexcept {
-		return is_allocated(id) ? &*m_values[id.index].value : nullptr;
+		return is_allocated(id) ? &(*this)[id] : nullptr;
 	}
 
 	[[nodiscard]] bool is_allocated(SlotId id) const noexcept {
-		return id && id.index < m_values.size() &&
-		       m_values[id.index].generation == id.generation && m_values[id.index].value.has_value();
+		return id.index < m_stored.size() && m_stored[id.index] != 0;
 	}
 
 	template <typename... Args>
 	[[nodiscard]] SlotId insert(Args&&... args) {
-		uint32_t index = 0;
 		if (m_free_list.empty()) {
-			index = static_cast<uint32_t>(m_values.size());
-			m_values.emplace_back();
-			m_values.back().value.emplace(std::forward<Args>(args)...);
-		} else {
-			index = m_free_list.back();
-			m_free_list.pop_back();
-			EXIT_IF(m_values[index].value.has_value());
-			m_values[index].value.emplace(std::forward<Args>(args)...);
+			Reserve();
 		}
-		++m_size;
-		return SlotId {index, m_values[index].generation};
+		const uint32_t index = m_free_list.back();
+		m_free_list.pop_back();
+		new (&(*this)[index]) T(std::forward<Args>(args)...);
+		m_stored[index] = 1;
+		return SlotId {index};
 	}
 
 	void erase(SlotId id) noexcept {
-		EXIT_IF(!is_allocated(id));
-		auto& slot = m_values[id.index];
-		slot.value.reset();
-		if (++slot.generation == 0) {
-			slot.generation = 1;
-		}
+		(*this)[id].~T();
+		m_stored[id.index] = 0;
 		m_free_list.push_back(id.index);
-		--m_size;
 	}
 
-	[[nodiscard]] size_t size() const noexcept { return m_size; }
-	[[nodiscard]] size_t capacity() const noexcept { return m_values.size(); }
-
 	template <typename F>
 	void ForEach(F&& fn) {
-		for (uint32_t index = 0; index < m_values.size(); ++index) {
-			if (m_values[index].value) {
-				fn(SlotId {index, m_values[index].generation}, *m_values[index].value);
+		for (uint32_t index = 0; index < m_stored.size(); ++index) {
+			if (m_stored[index] != 0) {
+				fn(SlotId {index}, (*this)[index]);
 			}
 		}
 	}
 
 	template <typename F>
 	void ForEach(F&& fn) const {
-		for (uint32_t index = 0; index < m_values.size(); ++index) {
-			if (m_values[index].value) {
-				fn(SlotId {index, m_values[index].generation}, *m_values[index].value);
+		for (uint32_t index = 0; index < m_stored.size(); ++index) {
+			if (m_stored[index] != 0) {
+				fn(SlotId {index}, (*this)[index]);
 			}
 		}
 	}
 
 private:
-	struct Slot {
-		std::optional<T> value;
-		uint32_t         generation = 1;
+	union Slot {
+		Slot() noexcept {}
+		~Slot() noexcept {}
+
+		T object;
 	};
 
-	std::deque<Slot>             m_values;
-	std::vector<uint32_t>        m_free_list;
-	size_t                       m_size = 0;
+	void Reserve() {
+		const auto first = static_cast<uint32_t>(m_stored.size());
+		m_chunks.push_back(std::make_unique<Slot[]>(ChunkSize));
+		m_stored.resize(first + ChunkSize);
+		m_free_list.resize(ChunkSize);
+		std::iota(m_free_list.rbegin(), m_free_list.rend(), first);
+	}
+
+	std::vector<std::unique_ptr<Slot[]>> m_chunks;
+	std::vector<uint8_t>                 m_stored;
+	std::vector<uint32_t>                m_free_list;
 };
 
 } // namespace Common
@@ -125,7 +125,7 @@ private:
 template <>
 struct std::hash<Common::SlotId> {
 	[[nodiscard]] size_t operator()(Common::SlotId id) const noexcept {
-		return std::hash<uint64_t> {}((static_cast<uint64_t>(id.generation) << 32u) | id.index);
+		return std::hash<uint32_t> {}(id.index);
 	}
 };
 

src/emulator.cpp

@@ -99,10 +99,6 @@ static void Init(const Config::ConfigOptions& cfg, const std::filesystem::path&
 
 	if (Common::File::IsFileExisting(param_json)) {
 		Loader::SystemContentLoadParamSfo(param_json);
-		if (const auto flexible_memory_size = Loader::SystemContentGetFlexibleMemorySize();
-		    flexible_memory_size != 0) {
-			Libs::LibKernel::Memory::SetFlexibleMemorySize(flexible_memory_size);
-		}
 	}
 	if (cfg.splash_enabled) {
 		Libs::Graphics::InitializeSplashScreen(param_json.parent_path());
@@ -188,7 +184,7 @@ void Run(const RunOptions& options) {
 
 	MountSandboxDirs();
 
-	Loader::AddContentLoad(options.addcont_dir, title_id);
+	Loader::AddContentLoad(options.addcont_dir, options.addcont, title_id);
 	auto* rt = Common::Singleton<Loader::RuntimeLinker>::Instance();
 	Libs::InitAll(rt->Symbols());
 

src/emulator.h

@@ -1,6 +1,7 @@
 #ifndef EMULATOR_INCLUDE_EMULATOR_EMULATOR_H_
 #define EMULATOR_INCLUDE_EMULATOR_EMULATOR_H_
 
+#include "common/additionalContent.h"
 #include "common/emulatorConfig.h"
 #include "common/stringUtils.h"
 
@@ -9,11 +10,12 @@
 namespace Emulator {
 
 struct RunOptions {
-	Config::ConfigOptions config;
-	std::filesystem::path app0_dir;
-	std::filesystem::path elf;
-	std::filesystem::path game_patch;
-	std::filesystem::path addcont_dir;
+	Config::ConfigOptions                config;
+	std::filesystem::path                app0_dir;
+	std::filesystem::path                elf;
+	std::filesystem::path                game_patch;
+	std::filesystem::path                addcont_dir;
+	std::vector<Common::AddContentEntry> addcont;
 };
 
 void Run(const RunOptions& options);

src/graphics/guest_gpu/command_processor/commandProcessor.h

@@ -27,6 +27,7 @@ struct IndirectDrawRegisters {
 	uint32_t vertex_offset;
 	uint32_t instance_offset;
 	uint32_t index_offset;
+	uint32_t draw_index;
 };
 
 class Pm4Execution {
@@ -97,6 +98,7 @@ public:
 
 	template <typename T>
 	void WaitRegMem(uint32_t func, const T* addr, T ref, T mask, uint32_t poll, uint32_t wait_op);
+	void MemSemaphore(uint64_t* counter, bool signal, bool write);
 	void WriteData(uint32_t* dst, const uint32_t* src, uint32_t dw_num, uint32_t write_control);
 	void WriteReferenceClock(uint64_t dst_address, uint32_t num_bytes);
 	void DmaData(uint8_t engine, uint8_t dst_sel, uint8_t dst_cache_policy,

src/graphics/guest_gpu/command_processor/pm4Dispatch.cpp

@@ -202,6 +202,7 @@ constexpr auto MakeOpcodeDispatchTable() {
 	g_cp_op_func[Pm4::IT_DMA_DATA]                  = CpOpDmaData;
 	g_cp_op_func[Pm4::IT_ACQUIRE_MEM]               = CpOpAcquireMem;
 	g_cp_op_func[Pm4::IT_RELEASE_MEM]               = CpOpReleaseMem;
+	g_cp_op_func[Pm4::IT_MEM_SEMAPHORE]             = CpOpMemSemaphore;
 	g_cp_op_func[Pm4::IT_REWIND]                    = CpOpRewind;
 	g_cp_op_func[Pm4::IT_SET_CONTEXT_REG]           = CpOpSetContextReg;
 	g_cp_op_func[Pm4::IT_SET_SH_REG]                = CpOpSetShaderReg;

src/graphics/guest_gpu/command_processor/pm4Dispatch.h

@@ -136,6 +136,7 @@ uint32_t CpOpEventWrite(CommandProcessor&, uint32_t, const uint32_t*, uint32_t,
 uint32_t CpOpDmaData(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpAcquireMem(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpReleaseMem(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
+uint32_t CpOpMemSemaphore(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpSetContextReg(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpSetShaderReg(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpSetUconfigReg(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);

src/graphics/guest_gpu/command_processor/pm4Handlers.cpp

@@ -1582,7 +1582,7 @@ KYTY_CP_OP_PARSER(CpOpDrawIndirect) {
 	// The native indexed packet enables the high-half first-index destination in word 3.
 	const IndirectDrawRegisters registers {
 	    buffer[1] & 0xffffu, buffer[2] & 0xffffu,
-	    indexed && (buffer[2] & (1u << 28u)) != 0 ? buffer[1] >> 16u : Pm4::SH_NOP};
+	    indexed && (buffer[2] & (1u << 28u)) != 0 ? buffer[1] >> 16u : Pm4::SH_NOP, Pm4::SH_NOP};
 
 	cp.DrawIndirect(data_offset, registers, draw_initiator, indexed);
 
@@ -1602,10 +1602,12 @@ KYTY_CP_OP_PARSER(CpOpDrawIndirectMulti) {
 	const auto stride_in_bytes = buffer[7];
 	const auto draw_initiator  = buffer[8];
 	const bool indexed         = (cmd_id == 0xc0083800);
-	// Multi-draw moves the first-index enable to word 4; the destinations stay in 2/3.
+	// Multi-draw moves the first-index enable to word 4; the destinations stay in 2/3. Word 4
+	// bit 31 enables the subdraw index, written to the register in its low half.
 	const IndirectDrawRegisters registers {
 	    buffer[1] & 0xffffu, buffer[2] & 0xffffu,
-	    indexed && (buffer[3] & (1u << 28u)) != 0 ? buffer[1] >> 16u : Pm4::SH_NOP};
+	    indexed && (buffer[3] & (1u << 28u)) != 0 ? buffer[1] >> 16u : Pm4::SH_NOP,
+	    (buffer[3] >> 31u) != 0 ? buffer[3] & 0xffffu : Pm4::SH_NOP};
 
 	if (count_indirect == 0) {
 		count_addr = nullptr;
@@ -2058,6 +2060,21 @@ KYTY_CP_OP_PARSER(CpOpReleaseMem) {
 	return 7;
 }
 
+KYTY_CP_OP_PARSER(CpOpMemSemaphore) {
+	KYTY_PROFILER_FUNCTION();
+
+	const auto control = buffer[2];
+	EXIT_NOT_IMPLEMENTED(cmd_id != 0xc0023900 || (control & ~0xe0110000u) != 0 ||
+	                     (control >> 29u) < 6u);
+
+	auto* counter = reinterpret_cast<uint64_t*>(
+	    (buffer[0] & ~0x7u) | (static_cast<uint64_t>(buffer[1] & 0x3ffffu) << 32u));
+	// A signal updates the counter before the next packet, so its mailbox stall never waits.
+	cp.MemSemaphore(counter, (control >> 29u) == 6u, ((control >> 20u) & 0x1u) != 0);
+
+	return 3;
+}
+
 KYTY_CP_OP_PARSER(CpOpSetContextReg) {
 	KYTY_PROFILER_FUNCTION();
 

src/graphics/guest_gpu/graphicsRun.cpp

@@ -281,6 +281,16 @@ template void CommandProcessor::WaitRegMem<uint32_t>(uint32_t, const uint32_t*,
 template void CommandProcessor::WaitRegMem<uint64_t>(uint32_t, const uint64_t*, uint64_t, uint64_t,
                                                      uint32_t, uint32_t);
 
+void CommandProcessor::MemSemaphore(uint64_t* counter, bool signal, bool write) {
+	if (signal) {
+		*counter = write ? 1u : *counter + 1u;
+	} else if (*counter == 0) {
+		SuspendPm4();
+	} else {
+		(*counter)--;
+	}
+}
+
 void CommandProcessor::WriteData(uint32_t* dst, const uint32_t* src, uint32_t dw_num,
                                  uint32_t write_control) {
 	const uint32_t dst_sel = ((write_control >> 30u) & 0x1u) | ((write_control >> 7u) & 0x1eu);
@@ -789,7 +799,7 @@ void CommandProcessor::DrawIndexOffset(uint32_t index_offset, uint32_t index_cou
 
 static void PatchIndirectDrawOffsets(CommandProcessor& cp, IndirectDrawRegisters registers,
                                      uint32_t vertex_offset, uint32_t instance_offset,
-                                     uint32_t first_index) {
+                                     uint32_t first_index, uint32_t draw_index) {
 	const auto write = [&](uint32_t location, uint32_t value) {
 		if (location == Pm4::SH_NOP) {
 			return;
@@ -801,6 +811,28 @@ static void PatchIndirectDrawOffsets(CommandProcessor& cp, IndirectDrawRegisters
 	write(registers.vertex_offset, vertex_offset);
 	write(registers.instance_offset, instance_offset);
 	write(registers.index_offset, first_index);
+	write(registers.draw_index, draw_index);
+}
+
+// Writes the GS user-data slots the packet patches; false when the shaders cannot read the
+// offsets from GPU records: a first-index or subdraw-index destination, or another stage's
+// user data.
+static bool IndirectDrawOffsets(IndirectDrawRegisters registers, ShaderIndirectOffsets& offsets) {
+	if (registers.index_offset != Pm4::SH_NOP || registers.draw_index != Pm4::SH_NOP) {
+		return false;
+	}
+	for (const auto [location, slot]: {std::pair {registers.vertex_offset, &offsets.vertex},
+	                                   std::pair {registers.instance_offset, &offsets.instance}}) {
+		if (location == Pm4::SH_NOP) {
+			continue;
+		}
+		const uint32_t index = location - Pm4::SPI_SHADER_USER_DATA_GS_0;
+		if (index >= HW::UserSgprInfo::SGPRS_MAX) {
+			return false;
+		}
+		*slot = static_cast<uint8_t>(index);
+	}
+	return true;
 }
 
 void CommandProcessor::DrawIndirect(uint32_t data_offset, IndirectDrawRegisters registers,
@@ -834,12 +866,11 @@ void CommandProcessor::DrawIndirectMulti(uint32_t data_offset, uint32_t max_coun
 	// NUM_INSTANCES keeps the last draw's instance count.
 	m_num_instances_addr = args_addr + static_cast<uint64_t>(draw_count - 1) * stride_in_bytes +
 	                       offsetof(vk::DrawIndirectCommand, instanceCount);
-	// With no SGPRs to patch, the GPU reads the records if the shaders allow. Shaders apply
-	// the records' offsets only through those SGPRs; Vulkan applies them, which matches
-	// while they are zero.
-	const DrawIndirectArgs records {args_addr, draw_count, stride_in_bytes};
-	if (registers.vertex_offset == Pm4::SH_NOP && registers.instance_offset == Pm4::SH_NOP &&
-	    registers.index_offset == Pm4::SH_NOP && (!indexed || m_index_buffer_size != 0) &&
+	// The GPU reads the records if the shaders allow; the VS prolog loads the patched SGPRs from
+	// Vulkan's bases. With no SGPR patched, Vulkan applies the offsets, which matches while they
+	// are zero.
+	DrawIndirectArgs records {args_addr, draw_count, stride_in_bytes};
+	if (IndirectDrawOffsets(registers, records.offsets) && (!indexed || m_index_buffer_size != 0) &&
 	    (indexed ? executor.DrawIndex(m_submit_id, CurrentBuffer(),
 	                                  {.index_count         = m_index_buffer_size,
 	                                   .index_addr          = IndexAddress(0),
@@ -854,7 +885,7 @@ void CommandProcessor::DrawIndirectMulti(uint32_t data_offset, uint32_t max_coun
 		const auto addr = args_addr + static_cast<uint64_t>(i) * stride_in_bytes;
 		if (!indexed) {
 			const auto& args = *reinterpret_cast<const vk::DrawIndirectCommand*>(addr);
-			PatchIndirectDrawOffsets(*this, registers, args.firstVertex, args.firstInstance, 0);
+			PatchIndirectDrawOffsets(*this, registers, args.firstVertex, args.firstInstance, 0, i);
 			executor.DrawAuto(m_submit_id, CurrentBuffer(),
 			                  {.vertex_count   = args.vertexCount,
 			                   .instance_count = args.instanceCount,
@@ -866,7 +897,7 @@ void CommandProcessor::DrawIndirectMulti(uint32_t data_offset, uint32_t max_coun
 
 		const auto& args = *reinterpret_cast<const vk::DrawIndexedIndirectCommand*>(addr);
 		PatchIndirectDrawOffsets(*this, registers, static_cast<uint32_t>(args.vertexOffset),
-		                         args.firstInstance, args.firstIndex);
+		                         args.firstInstance, args.firstIndex, i);
 		executor.DrawIndex(m_submit_id, CurrentBuffer(),
 		                   {.index_count         = std::min(args.indexCount, max_indices),
 		                    .index_addr          = IndexAddress(args.firstIndex),
@@ -964,16 +995,28 @@ void CommandProcessor::ReleaseMem(uint32_t data_sel, uint32_t interrupt_selector
 	// Dumped occlusion counters are written before any end-of-pipe signal reaches the guest.
 	m_renderer.GetOcclusionQueries().Publish();
 
+	// A label beside GPU-written bytes goes to the backing store, so they are not downloaded.
+	// Other labels go through the guest mapping: its fault keeps a clean cached copy current and
+	// downloads GPU-written bytes under the label before the store.
+	const auto write_label = [this, address](const void* value, uint64_t size) {
+		const auto vaddr   = reinterpret_cast<uint64_t>(address);
+		auto&      buffers = m_renderer.GetBufferCache();
+		if (!GuestRange {vaddr, size}.Valid() || !buffers.IsRegionGpuModified(vaddr, size) ||
+		    buffers.HasGpuDirtyBytes(vaddr, size) ||
+		    !LibKernel::Memory::TryWriteBacking(vaddr, value, size)) {
+			std::memcpy(address, value, size);
+		}
+	};
 	// EndOfPipeData: kNone, kImmediate32b, kImmediate64b, kRefClk; 5 reads GDS.
 	switch (data_sel) {
 		case 0: break;
 		case 1: {
 			const auto value = static_cast<uint32_t>(data);
-			std::memcpy(address, &value, sizeof(value));
+			write_label(&value, sizeof(value));
 			break;
 		}
 		case 3: data = Sync::ReadReferenceClock(); [[fallthrough]];
-		case 2: std::memcpy(address, &data, sizeof(data)); break;
+		case 2: write_label(&data, sizeof(data)); break;
 		case 5: {
 			GetScheduler().Finish();
 			const auto gds = static_cast<uint32_t>(data);

src/graphics/host_gpu/memoryTracker.cpp

@@ -66,18 +66,6 @@ RegionManager* MemoryTracker::GetOrCreateRegion(uint64_t index) {
 	return ptr;
 }
 
-void MemoryTracker::QueueBdaSync(uint64_t vaddr, uint64_t size) {
-	const auto begin = Common::AlignDown(vaddr, TRACKER_PAGE_SIZE);
-	const auto end   = Common::AlignUp(vaddr + size, TRACKER_PAGE_SIZE);
-	std::lock_guard lock(m_bda_sync_mutex);
-	m_bda_sync.Add(begin, end - begin);
-}
-
-RangeSet MemoryTracker::TakeBdaSync() {
-	std::lock_guard lock(m_bda_sync_mutex);
-	return std::exchange(m_bda_sync, RangeSet {});
-}
-
 bool MemoryTracker::IsRegionCpuModified(uint64_t vaddr, uint64_t size) {
 	CheckNotInUploadCallback();
 	return Iterate<true>(vaddr, size, [](RegionManager* manager, uint64_t offset, uint64_t bytes) {
@@ -100,7 +88,6 @@ void MemoryTracker::MarkRegionAsCpuModified(uint64_t vaddr, uint64_t size) {
 		std::scoped_lock lock(manager->lock);
 		manager->ChangeState<DirtySource::Cpu, true>(manager->GetCpuAddr() + offset, bytes);
 	});
-	QueueBdaSync(vaddr, size);
 }
 
 void MemoryTracker::MarkRegionAsGpuModified(uint64_t vaddr, uint64_t size) {
@@ -141,8 +128,6 @@ void MemoryTracker::UntrackMemory(uint64_t vaddr, uint64_t size) {
 	Iterate<false>(vaddr, size, [](RegionManager* manager, uint64_t offset, uint64_t bytes) {
 		manager->ChangeState<DirtySource::Cpu, true>(manager->GetCpuAddr() + offset, bytes);
 	});
-	locks.clear();
-	QueueBdaSync(vaddr, size);
 }
 
 } // namespace Libs::Graphics

src/graphics/host_gpu/memoryTracker.h

@@ -29,10 +29,6 @@ public:
 	void               MarkRegionAsGpuModified(uint64_t vaddr, uint64_t size);
 	void               UnmarkRegionAsGpuModified(uint64_t vaddr, uint64_t size);
 	void               UntrackMemory(uint64_t vaddr, uint64_t size);
-	// Queue after changing ownership; a concurrent drain must not consume an event
-	// before the dirty pages it describes become visible.
-	void                   QueueBdaSync(uint64_t vaddr, uint64_t size);
-	[[nodiscard]] RangeSet TakeBdaSync();
 	// Removes protection from a range and flushes GPU-owned data when required.
 	template <typename Flush>
 	void InvalidateRegion(uint64_t vaddr, uint64_t size, Flush&& on_flush) noexcept {
@@ -55,7 +51,6 @@ public:
 				on_flush();
 			}
 		});
-		QueueBdaSync(vaddr, size);
 	}
 #if KYTY_BUILD == KYTY_BUILD_DEBUG
 	void ValidateGpuDirtyPages(const RangeSet& dirty, uint64_t vaddr, uint64_t size,
@@ -154,8 +149,6 @@ private:
 	std::unique_ptr<std::atomic<RegionManager*>[]> m_regions;
 	std::vector<std::unique_ptr<RegionManager>>    m_region_storage;
 	std::mutex                                     m_region_mutex;
-	std::mutex                                     m_bda_sync_mutex;
-	RangeSet                                       m_bda_sync;
 	PageManager&                                   m_page_manager;
 };
 

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

@@ -127,7 +127,7 @@ void BufferCache::ChangeRegister(BufferId id) {
 		}
 		WriteDataBuffer(m_bda_pagetable_buffer, table_offset,
 		                addresses.data(), addresses.size() * sizeof(vk::DeviceAddress));
-		m_memory_tracker.QueueBdaSync(buffer.CpuAddress(), buffer.Size());
+		QueueBdaSync(buffer.CpuAddress(), buffer.Size());
 	} else {
 		const auto found = m_buffers.find(buffer.CpuAddress());
 		EXIT_IF(found == m_buffers.end() || found->second != id);
@@ -299,6 +299,7 @@ void BufferCache::InvalidateMemory(uint64_t vaddr, uint64_t size) {
 	}
 	m_memory_tracker.InvalidateRegion(vaddr, size,
 	                                  [this, vaddr, size] { ReadMemory(vaddr, size, true); });
+	QueueBdaSync(vaddr, size);
 }
 
 void BufferCache::ReadMemory(uint64_t vaddr, uint64_t size, bool is_write) {
@@ -318,6 +319,7 @@ void BufferCache::ReadMemory(uint64_t vaddr, uint64_t size, bool is_write) {
 		}
 		if (is_write) {
 			m_memory_tracker.MarkRegionAsCpuModified(vaddr, size);
+			QueueBdaSync(vaddr, size);
 		}
 	});
 }
@@ -719,14 +721,16 @@ void BufferCache::ProcessFaultBuffer() {
 	m_fault_manager.ProcessFaultBuffer();
 }
 
-void BufferCache::SynchronizeBda(const RangeSet& mapped) {
-	// Take before walking: writes queued during synchronization remain pending.
-	const auto pending = m_memory_tracker.TakeBdaSync();
-	pending.ForEach([&](uint64_t begin, uint64_t end) {
-		mapped.ForEachInRange(begin, end - begin, [this](uint64_t start, uint64_t finish) {
-			SynchronizeBuffersInRange(start, finish - start);
-		});
-	});
+void BufferCache::QueueBdaSync(uint64_t vaddr, uint64_t size) {
+	const auto begin = Common::AlignDown(vaddr, TRACKER_PAGE_SIZE);
+	const auto end   = Common::AlignUp(vaddr + size, TRACKER_PAGE_SIZE);
+	std::lock_guard lock(m_bda_sync_mutex);
+	m_bda_sync.Add(begin, end - begin);
+}
+
+RangeSet BufferCache::TakeBdaSync() {
+	std::lock_guard lock(m_bda_sync_mutex);
+	return std::exchange(m_bda_sync, {});
 }
 
 void BufferCache::SynchronizeBuffersInRange(uint64_t vaddr, uint64_t size) {

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

@@ -13,6 +13,7 @@
 
 #include <array>
 #include <map>
+#include <mutex>
 #include <span>
 #include <utility>
 #include <vector>
@@ -85,19 +86,16 @@ public:
 	void               MarkColorMetadataChecked(uint64_t vaddr, uint64_t size, uint32_t keys);
 	[[nodiscard]] bool IsColorMetadataChecked(uint64_t vaddr, uint64_t size, uint32_t keys) const;
 	void               ProcessFaultBuffer();
-	void               SynchronizeBda(const RangeSet& mapped);
-	void               QueueBdaSync(uint64_t vaddr, uint64_t size) {
-		m_memory_tracker.QueueBdaSync(vaddr, size);
-	}
+	void               SynchronizeBuffersInRange(uint64_t vaddr, uint64_t size);
+	// Queue after the range's state change becomes visible, or a drain could consume it first.
+	void                   QueueBdaSync(uint64_t vaddr, uint64_t size);
+	[[nodiscard]] RangeSet TakeBdaSync();
 	void               RunGarbageCollector();
 
 private:
 	friend struct BufferCacheTestAccess;
 
-	bool IsBufferInvalid(BufferId id) const {
-		const auto* buffer = m_slot_buffers.try_get(id);
-		return buffer == nullptr || buffer->is_deleted;
-	}
+	bool IsBufferInvalid(BufferId id) const { return !id || m_slot_buffers[id].is_deleted; }
 
 	using BufferMap = std::map<uint64_t, BufferId>;
 	struct OverlapResult {
@@ -122,7 +120,6 @@ private:
 	void DeleteBuffer(BufferId id);
 	[[nodiscard]] bool SynchronizeBuffer(Buffer& buffer, uint64_t vaddr, uint64_t size,
 	                                     bool is_written, bool is_texel_buffer);
-	void SynchronizeBuffersInRange(uint64_t vaddr, uint64_t size);
 	[[nodiscard]] vk::Buffer UploadCopies(Buffer& buffer, std::span<vk::BufferCopy> copies,
 	                                      uint64_t total_size);
 	[[nodiscard]] bool SynchronizeBufferFromImage(Buffer& buffer, uint64_t vaddr, uint64_t size);
@@ -144,6 +141,8 @@ private:
 	// One per TileManager::ColorClearSlots key slot.
 	std::array<RangeSet, 8>                           m_checked_color_metadata;
 	MemoryTracker                                     m_memory_tracker;
+	std::mutex                                        m_bda_sync_mutex;
+	RangeSet                                          m_bda_sync;
 	StreamBuffer                                      m_staging_buffer;
 	StreamBuffer                                      m_stream_buffer;
 	StreamBuffer                                      m_download_buffer;

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

@@ -8,6 +8,7 @@
 #include <cstddef>
 #include <cstdint>
 #include <cstring>
+#include <iterator>
 #include <memory>
 #include <type_traits>
 #include <utility>
@@ -118,6 +119,13 @@ public:
 
 	class const_iterator final {
 	public:
+		// libstdc++'s algorithms need these. Input, not forward, because operator* returns a copy.
+		using iterator_category = std::input_iterator_tag;
+		using value_type        = OwnerT;
+		using difference_type   = std::ptrdiff_t;
+		using pointer           = void;
+		using reference         = OwnerT;
+
 		[[nodiscard]] OwnerT operator*() const noexcept { return m_owners->At(m_index); }
 		const_iterator&      operator++() noexcept {
 			++m_index;

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

@@ -251,24 +251,24 @@ void TextureCache::UnregisterImage(ImageId id) {
 }
 
 void TextureCache::DeleteImage(ImageId id) {
-	auto* image = m_slot_images.try_get(id);
-	if (image == nullptr || !image->registered) {
+	auto& image = m_slot_images[id];
+	if (!image.registered) {
 		return;
 	}
-	if (image->depth_id) {
-		std::erase(m_slot_images[image->depth_id].stencil_ids, id);
+	if (image.depth_id) {
+		std::erase(m_slot_images[image.depth_id].stencil_ids, id);
 	} else {
 		// Detach first: deleting each association also unlinks it from its depth owner.
-		for (const auto association: std::exchange(image->stencil_ids, {})) {
+		for (const auto association: std::exchange(image.stencil_ids, {})) {
 			FreeImage(association);
 		}
 	}
-	if (image->IsGpuModified()) {
+	if (image.IsGpuModified()) {
 		EXIT("TextureCache: deleting a GPU-modified image without resolving its contents\n");
 	}
 	m_download_images.erase(id);
-	if (image->info.metadata.kind == ImageMetadataKind::Htile) {
-		m_surface_metas.erase(image->info.metadata.range.address);
+	if (image.info.metadata.kind == ImageMetadataKind::Htile) {
+		m_surface_metas.erase(image.info.metadata.range.address);
 	}
 	UnregisterImage(id);
 	if (m_scheduler.Active()) {
@@ -429,12 +429,6 @@ TextureCache::ImageIds TextureCache::FindImagesInRegion(uint64_t address, uint64
 		return {};
 	}
 
-	uint32_t query_epoch = ++m_image_query_epoch;
-	if (query_epoch == 0) {
-		m_slot_images.ForEach([](ImageId, const Image& image) { image.query_epoch = 0; });
-		query_epoch = ++m_image_query_epoch;
-	}
-
 	ImageIds result;
 	ForEachPage(address, size, [&](uint64_t page) {
 		const auto* owners = m_image_page_table.Find(page);
@@ -442,19 +436,17 @@ TextureCache::ImageIds TextureCache::FindImagesInRegion(uint64_t address, uint64
 			return;
 		}
 		owners->ForEach([&](ImageId id) {
-			auto* image = m_slot_images.try_get(id);
-			if (image == nullptr) {
+			const auto& image = m_slot_images[id];
+			if (image.picked || !image.Overlaps(address, size, page_overlap)) {
 				return;
 			}
-			if (image->query_epoch == query_epoch) {
-				return;
-			}
-			image->query_epoch = query_epoch;
-			if (image->Overlaps(address, size, page_overlap)) {
-				result.push_back(id);
-			}
+			image.picked = true;
+			result.push_back(id);
 		});
 	});
+	for (const auto id: result) {
+		m_slot_images[id].picked = false;
+	}
 	return result;
 }
 
@@ -496,16 +488,15 @@ void TextureCache::RefreshCopySource(ImageId id) {
 }
 
 bool TextureCache::CopyD16(Image& destination, Image& source) {
-	const bool source_depth      = source.info.IsDepth();
-	const bool destination_depth = destination.info.IsDepth();
-	if (source_depth == destination_depth) {
+	// Converts guest D16 between a four-byte host depth aspect and two-byte color or D16 depth.
+	const bool source_wide = DepthAspectTransferBytes(source.backing.format) == sizeof(uint32_t);
+	if (source_wide == (DepthAspectTransferBytes(destination.backing.format) == sizeof(uint32_t))) {
 		return false;
 	}
-	auto&      depth          = source_depth ? source : destination;
-	auto&      color          = source_depth ? destination : source;
-	const auto transfer_bytes = DepthAspectTransferBytes(depth.backing.format);
+	auto& depth  = source_wide ? source : destination;
+	auto& narrow = source_wide ? destination : source;
 	if (depth.info.bytes_per_block != sizeof(uint16_t) ||
-	    color.info.bytes_per_block != sizeof(uint16_t) || transfer_bytes != sizeof(uint32_t)) {
+	    narrow.info.bytes_per_block != sizeof(uint16_t)) {
 		return false;
 	}
 	EXIT_IF(source.backing.samples != 1 || destination.backing.samples != 1 ||
@@ -515,42 +506,42 @@ bool TextureCache::CopyD16(Image& destination, Image& source) {
 
 	const auto     layers = depth.info.resources.layers;
 	const uint64_t depth_slice =
-	    static_cast<uint64_t>(depth.info.pitch) * depth.info.extent.height * transfer_bytes;
-	const uint64_t color_slice =
-	    static_cast<uint64_t>(color.info.pitch) * color.info.extent.height * sizeof(uint16_t);
-	EXIT_IF(layers == 0 || depth_slice > UINT64_MAX / layers || color_slice > UINT64_MAX / layers);
-	const auto                       depth_size = depth_slice * layers;
-	const auto                       color_size = color_slice * layers;
-	const std::array                 depth_copies {
+	    static_cast<uint64_t>(depth.info.pitch) * depth.info.extent.height * sizeof(uint32_t);
+	const uint64_t narrow_slice =
+	    static_cast<uint64_t>(narrow.info.pitch) * narrow.info.extent.height * sizeof(uint16_t);
+	EXIT_IF(layers == 0 || depth_slice > UINT64_MAX / layers || narrow_slice > UINT64_MAX / layers);
+	const std::array depth_copies {
 	    vk::BufferImageCopy {0,
 	                         depth.info.pitch,
 	                         depth.info.extent.height,
 	                         {vk::ImageAspectFlagBits::eDepth, 0, 0, layers},
 	                         {},
 	                         depth.info.extent}};
-	const std::array color_copies {
+	const std::array narrow_copies {
 	    vk::BufferImageCopy {0,
-	                         color.info.pitch,
-	                         color.info.extent.height,
-	                         {vk::ImageAspectFlagBits::eColor, 0, 0, layers},
+	                         narrow.info.pitch,
+	                         narrow.info.extent.height,
+	                         {narrow.info.IsDepth() ? vk::ImageAspectFlagBits::eDepth
+	                                                : vk::ImageAspectFlagBits::eColor,
+	                          0, 0, layers},
 	                         {},
-	                         color.info.extent}};
+	                         narrow.info.extent}};
 
-	auto                         depth_buffer = m_tiler.GetScratchBuffer(depth_size);
-	auto                         color_buffer = m_tiler.GetScratchBuffer(color_size, depth_buffer.buffer);
+	auto depth_buffer  = m_tiler.GetScratchBuffer(depth_slice * layers);
+	auto narrow_buffer = m_tiler.GetScratchBuffer(narrow_slice * layers, depth_buffer.buffer);
 	const TileManager::D16Layout promote_layout {
 	    .width               = depth.info.extent.width,
 	    .height              = depth.info.extent.height,
 	    .layers              = layers,
-	    .source_row_stride   = static_cast<uint64_t>(color.info.pitch) * sizeof(uint16_t),
-	    .target_row_stride   = static_cast<uint64_t>(depth.info.pitch) * transfer_bytes,
-	    .source_slice_stride = color_slice,
+	    .source_row_stride   = static_cast<uint64_t>(narrow.info.pitch) * sizeof(uint16_t),
+	    .target_row_stride   = static_cast<uint64_t>(depth.info.pitch) * sizeof(uint32_t),
+	    .source_slice_stride = narrow_slice,
 	    .target_slice_stride = depth_slice,
 	};
 	const bool d32 = DepthAspectTransferFormat(depth.backing.format) == vk::Format::eD32Sfloat;
-	if (source_depth) {
+	if (source_wide) {
 		source.Download(depth_copies, depth_buffer.buffer, depth_buffer.offset, depth_buffer.size);
-		m_tiler.ConvertD16(depth_buffer, color_buffer, TileManager::D16Direction::Demote, d32,
+		m_tiler.ConvertD16(depth_buffer, narrow_buffer, TileManager::D16Direction::Demote, d32,
 		                   {.width               = promote_layout.width,
 		                    .height              = promote_layout.height,
 		                    .layers              = promote_layout.layers,
@@ -558,11 +549,12 @@ bool TextureCache::CopyD16(Image& destination, Image& source) {
 		                    .target_row_stride   = promote_layout.source_row_stride,
 		                    .source_slice_stride = promote_layout.target_slice_stride,
 		                    .target_slice_stride = promote_layout.source_slice_stride});
-		destination.Upload(color_copies, color_buffer.buffer, color_buffer.offset,
-		                   color_buffer.size);
+		destination.Upload(narrow_copies, narrow_buffer.buffer, narrow_buffer.offset,
+		                   narrow_buffer.size);
 	} else {
-		source.Download(color_copies, color_buffer.buffer, color_buffer.offset, color_buffer.size);
-		m_tiler.ConvertD16(color_buffer, depth_buffer, TileManager::D16Direction::Promote, d32,
+		source.Download(narrow_copies, narrow_buffer.buffer, narrow_buffer.offset,
+		                narrow_buffer.size);
+		m_tiler.ConvertD16(narrow_buffer, depth_buffer, TileManager::D16Direction::Promote, d32,
 		                   promote_layout);
 		destination.Upload(depth_copies, depth_buffer.buffer, depth_buffer.offset,
 		                   depth_buffer.size);
@@ -787,15 +779,22 @@ void Resolve(TextureCache& cache, ImageId cached_id, ImageId merged_id) {
 TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& requested,
                                                          BindingType binding, ImageId cached_id,
                                                          ImageId merged_id) {
-	auto owner = m_slot_images.try_get(cached_id);
-	if (owner == nullptr || !owner->registered) {
+	auto& cached = m_slot_images[cached_id];
+	if (!cached.registered) {
 		return {merged_id};
 	}
-	if (BcAlias::Matches(requested, owner->info) && !owner->IsCpuDirty()) {
+	// PS5: image heaps keep BC images and raw block views of their layers resident together, so a
+	// raw read of a layer subset with the same levels views the BC image instead of evicting it.
+	const bool raw_view = binding == BindingType::Texture && !requested.IsBlock() &&
+	                      requested.type == cached.info.type &&
+	                      requested.data.address == cached.info.data.address &&
+	                      requested.BlockExtent() == cached.info.BlockExtent() &&
+	                      requested.resources.levels == cached.info.resources.levels &&
+	                      requested.resources.layers < cached.info.resources.layers;
+	if (!raw_view && BcAlias::Matches(requested, cached.info) && !cached.IsCpuDirty()) {
 		BcAlias::Resolve(*this, cached_id, merged_id);
 		return {merged_id};
 	}
-	auto&      cached       = *owner;
 	const auto current_tick = m_scheduler.CurrentTick();
 	const bool safe_to_delete =
 	    current_tick - std::min(current_tick, cached.tick_accessed_last) > NumFramesBeforeRemoval &&
@@ -869,9 +868,8 @@ TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& reques
 		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;
-			const auto result    = m_slot_images.try_get(result_id);
-			return {result != nullptr && ImageViewOps::FormatsCompatible(result->info.pixel_format,
-			                                                             requested.pixel_format)
+			return {ImageViewOps::FormatsCompatible(m_slot_images[result_id].info.pixel_format,
+			                                        requested.pixel_format)
 			            ? result_id
 			            : ImageId {}};
 		}
@@ -1218,9 +1216,8 @@ ImageId 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 == stencil &&
-		    owner->info.extent == depth.info.extent) {
+		const auto& owner = m_slot_images[id];
+		if (owner.info.data == stencil && owner.info.extent == depth.info.extent) {
 			association = id;
 		}
 	}
@@ -1313,12 +1310,24 @@ ImageId TextureCache::FindImage(ImageDesc& desc, bool exact_format) {
 					view_layer = overlap.layer;
 				}
 			}
+			// PS5: image heaps hold a descriptor per dynamic-resolution extent of a render target,
+			// and only one extent at an address holds data, so heap lookups create no second one.
+			if (!result && desc.heap &&
+			    std::any_of(candidates.begin(), candidates.end(), [&](ImageId id) {
+				    const auto* image = m_slot_images.try_get(id);
+				    return image != nullptr && image->registered &&
+					       image->heap_refs.load(std::memory_order_relaxed) != 0u &&
+					       image->info.data.address == desc.info.data.address &&
+					       image->info.type == desc.info.type &&
+					       image->info.extent != desc.info.extent;
+			    })) {
+				return {};
+			}
 		} else if (block_alias) {
 			for (const auto candidate: candidates) {
-				const auto* owner = m_slot_images.try_get(candidate);
-				if (candidate != result && owner != nullptr && owner->registered &&
-				    !owner->IsCpuDirty() &&
-				    BcAlias::Matches(m_slot_images[result].info, owner->info)) {
+				const auto& owner = m_slot_images[candidate];
+				if (candidate != result && owner.registered && !owner.IsCpuDirty() &&
+				    BcAlias::Matches(m_slot_images[result].info, owner.info)) {
 					BcAlias::Resolve(*this, candidate, result);
 				}
 			}
@@ -1381,14 +1390,14 @@ ImageId TextureCache::FindImageFromRange(uint64_t address, uint64_t size, bool e
 	std::scoped_lock lock {m_lock};
 	ImageIds         matches;
 	for (const auto id: FindImagesInRegion(address, size, false)) {
-		auto owner = m_slot_images.try_get(id);
-		if (owner == nullptr || owner->info.data.address != address) {
+		const auto* owner = &m_slot_images[id];
+		if (owner->info.data.address != address) {
 			continue;
 		}
 		if (ensure_valid && owner->depth_id) {
-			owner = m_slot_images.try_get(owner->depth_id);
+			owner = &m_slot_images[owner->depth_id];
 		}
-		if (owner == nullptr || (ensure_valid && !owner->SafeToDownload())) {
+		if (ensure_valid && !owner->SafeToDownload()) {
 			continue;
 		}
 		matches.push_back(id);
@@ -1405,11 +1414,8 @@ ImageId TextureCache::FindImageFromRange(uint64_t address, uint64_t size, bool e
 			}
 		}
 	}
-	if (selected && ensure_valid) {
-		const auto owner = m_slot_images.try_get(selected);
-		if (owner != nullptr && owner->depth_id) {
-			selected = owner->depth_id;
-		}
+	if (selected && ensure_valid && m_slot_images[selected].depth_id) {
+		selected = m_slot_images[selected].depth_id;
 	}
 	return selected;
 }
@@ -1430,9 +1436,8 @@ vk::ImageView TextureCache::FindTexture(ImageId id, const ImageDesc& desc) {
 	    desc.info.data.End() <= image.info.stencil.End()) {
 		for (const auto stencil_id:
 		     FindImagesInRegion(image.info.stencil.address, image.info.stencil.size, false)) {
-			const auto* stencil = m_slot_images.try_get(stencil_id);
-			if (stencil != nullptr && stencil->depth_id == id &&
-			    stencil->info.data == image.info.stencil) {
+			const auto& stencil = m_slot_images[stencil_id];
+			if (stencil.depth_id == id && stencil.info.data == image.info.stencil) {
 				RefreshImage(stencil_id);
 				break;
 			}
@@ -1543,27 +1548,24 @@ bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address
 	vk::ImageAspectFlags  aspect {};
 	ImageSubresourceRange subresources;
 	for (const auto id: FindImagesInRegion(address, size, false)) {
-		auto owner = m_slot_images.try_get(id);
-		if (owner == nullptr) {
-			continue;
-		}
+		const auto&          owner = m_slot_images[id];
 		vk::ImageAspectFlags candidate {};
 		ImageId              candidate_id = id;
 		ImageSubresourceRange candidate_subresources {
-		    0, owner->info.resources.levels, 0, owner->info.TransferLayers()};
-		if (owner->depth_id && owner->info.data.address == address &&
-		    owner->info.data.size == size) {
+		    0, owner.info.resources.levels, 0, owner.info.TransferLayers()};
+		if (owner.depth_id && owner.info.data.address == address &&
+		    owner.info.data.size == size) {
 			candidate              = vk::ImageAspectFlagBits::eStencil;
-			candidate_id           = owner->depth_id;
-			candidate_subresources = owner->stencil_subresources;
-			owner        = m_slot_images.try_get(candidate_id);
-			if (owner == nullptr || owner->backing.image == nullptr || !owner->info.HasStencil()) {
+			candidate_id           = owner.depth_id;
+			candidate_subresources = owner.stencil_subresources;
+			const auto& depth      = m_slot_images[candidate_id];
+			if (depth.backing.image == nullptr || !depth.info.HasStencil()) {
 				continue;
 			}
-		} else if (!owner->depth_id && owner->info.data.address == address &&
-		           owner->info.data.size == size) {
-			candidate = owner->info.IsDepth() ? vk::ImageAspectFlagBits::eDepth
-			                                  : vk::ImageAspectFlagBits::eColor;
+		} else if (!owner.depth_id && owner.info.data.address == address &&
+		           owner.info.data.size == size) {
+			candidate = owner.info.IsDepth() ? vk::ImageAspectFlagBits::eDepth
+			                                 : vk::ImageAspectFlagBits::eColor;
 		}
 		if (!candidate) {
 			continue;
@@ -1904,23 +1906,20 @@ void TextureCache::InvalidateCpuAliases(uint64_t address, uint64_t size) {
 	const auto page_begin = Common::AlignDown(address, TRACKER_PAGE_SIZE);
 	const auto page_end   = Common::AlignUp(address + size, TRACKER_PAGE_SIZE);
 	for (const auto id: FindImagesInRegion(address, size, true)) {
-		auto owner = m_slot_images.try_get(id);
-		if (owner == nullptr) {
-			continue;
-		}
-		if (owner->Overlaps(address, size)) {
-			owner->MarkCpuDirty();
+		auto& owner = m_slot_images[id];
+		if (owner.Overlaps(address, size)) {
+			owner.MarkCpuDirty();
 			UntrackImage(id);
 			continue;
 		}
-		const auto image_begin = owner->info.data.address;
-		const auto image_end   = owner->info.data.End();
+		const auto image_begin = owner.info.data.address;
+		const auto image_end   = owner.info.data.End();
 		if (page_end < image_end) {
 			UntrackImageHead(id);
 		} else if (image_begin < page_begin) {
 			UntrackImageTail(id);
 		} else {
-			MarkAsMaybeDirty(id, *owner);
+			MarkAsMaybeDirty(id, owner);
 		}
 	}
 }
@@ -2060,8 +2059,7 @@ void TextureCache::ProcessDownloadImages() {
 	}
 	std::scoped_lock lock {m_lock};
 	for (const auto id: m_download_images) {
-		const auto owner = m_slot_images.try_get(id);
-		if (owner != nullptr && owner->registered && owner->IsGpuModified()) {
+		if (m_slot_images[id].IsGpuModified()) {
 			(void)DownloadImageMemory(id);
 		}
 	}

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

@@ -43,6 +43,7 @@ public:
 		ImageInfo     info;
 		ImageViewInfo view_info;
 		BindingType   type = BindingType::Texture;
+		bool          heap = false; // PS5: resolved for a resident image heap.
 	};
 
 	TextureCache(GraphicContext& graphics, CommandScheduler& scheduler, PageManager& page_manager,
@@ -184,7 +185,6 @@ private:
 	uint64_t                                          m_pressure_gc_memory = 1536ull * 1024 * 1024;
 	uint64_t         m_critical_gc_memory     = 3ull * 1024 * 1024 * 1024;
 	uint64_t         m_gc_tick                = 0;
-	mutable uint32_t m_image_query_epoch      = 0;
 	bool             m_readback_linear_images = false;
 	bool             m_readback_tiled_images  = false;
 

src/graphics/host_gpu/renderer/debug.cpp

@@ -595,13 +595,6 @@ void LogDrawPhase(const char* draw_name, const char* phase) {
 	}
 }
 
-void LogPipelineTrace(const char* phase, uint64_t vertex_program_id, uint64_t pixel_program_id) {
-	if (graphics_debug_dump_enabled()) {
-		LOGF("PipelineTrace: %s VS=%" PRIu64 " PS=%" PRIu64 "\n", phase, vertex_program_id,
-		     pixel_program_id);
-	}
-}
-
 static void VpPrint(const char* func, const HW::ScreenViewport& vp,
                     const HW::ScanModeControl& smc) {
 	LOGF("%s\n", func);

src/graphics/host_gpu/renderer/dynamicState.cpp

@@ -125,6 +125,14 @@ void DynamicState::Commit(const GraphicContext& graphics, vk::CommandBuffer cmdb
 			}
 		}
 	}
+	if (dirty_state.cull_mode) {
+		dirty_state.cull_mode = false;
+		cmdbuf.setCullMode(cull_mode);
+	}
+	if (dirty_state.front_face) {
+		dirty_state.front_face = false;
+		cmdbuf.setFrontFace(front_face);
+	}
 	if (dirty_state.blend_constants) {
 		dirty_state.blend_constants = false;
 		cmdbuf.setBlendConstants(blend_constants.data());

src/graphics/host_gpu/renderer/dynamicState.h

@@ -52,6 +52,9 @@ struct DynamicState {
 		bool stencil_back_write_mask    : 1;
 		bool stencil_back_compare_mask  : 1;
 
+		bool cull_mode  : 1;
+		bool front_face : 1;
+
 		bool blend_constants       : 1;
 		bool color_write_enables   : 1;
 		bool line_width            : 1;
@@ -86,6 +89,9 @@ struct DynamicState {
 	uint32_t   stencil_back_write_mask {};
 	uint32_t   stencil_back_compare_mask {};
 
+	vk::CullModeFlags cull_mode {};
+	vk::FrontFace     front_face {};
+
 	std::array<float, 4>                               blend_constants {};
 	std::array<vk::Bool32, RENDER_COLOR_ATTACHMENTS_MAX> color_write_enables {};
 	uint32_t                                           num_color_write_enables {};
@@ -221,6 +227,20 @@ struct DynamicState {
 		}
 	}
 
+	void SetCullMode(const vk::CullModeFlags cull_mode_) {
+		if (cull_mode != cull_mode_) {
+			cull_mode             = cull_mode_;
+			dirty_state.cull_mode = true;
+		}
+	}
+
+	void SetFrontFace(const vk::FrontFace front_face_) {
+		if (front_face != front_face_) {
+			front_face             = front_face_;
+			dirty_state.front_face = true;
+		}
+	}
+
 	void SetBlendConstants(const std::array<float, 4> blend_constants_) {
 		if (blend_constants != blend_constants_) {
 			blend_constants             = blend_constants_;

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

@@ -425,15 +425,15 @@ void Image::Resolve(Image& source, const ImageSubresourceRange& source_range,
 	const auto destination_height =
 	    std::max(backing.extent.height >> destination_range.base_level, 1u);
 	const bool copy = source.backing.samples == 1;
-	EXIT_IF(layers == 0 || info.extent.width > source_width || info.extent.height > source_height ||
-	        info.extent.width > destination_width || info.extent.height > destination_height ||
+	EXIT_IF(layers == 0 ||
 	        (copy ? !ImageViewOps::FormatsCompatible(source.backing.format, backing.format)
 	              : source.backing.format != backing.format));
 	auto resolved_source_range             = source_range;
 	auto resolved_destination_range        = destination_range;
 	resolved_source_range.layer_count      = layers;
 	resolved_destination_range.layer_count = layers;
-	const vk::Extent3D resolve_extent {info.extent.width, info.extent.height, 1};
+	const vk::Extent3D resolve_extent {std::min(source_width, destination_width),
+	                                   std::min(source_height, destination_height), 1};
 
 	m_scheduler.EndRendering();
 	auto command = m_scheduler.Current().Handle();

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

@@ -139,14 +139,17 @@ public:
 	}
 
 	static inline std::atomic<uint64_t> heap_epoch {0};
+	static inline std::atomic<uint64_t> serials {0};
 
+	// Unique per image, unlike ImageId: freed slots are reused.
+	const uint64_t   serial = serials.fetch_add(1u, std::memory_order_relaxed) + 1u;
 	ImageInfo        info;
 	VulkanImage      backing;
 	std::vector<CachedImageView> views;
 	ImageUsage       usage;
 	ImageBinding     binding;
 	bool             registered     = false;
-	mutable uint32_t query_epoch    = 0;
+	mutable bool     picked         = false; // Set only while FindImagesInRegion collects it.
 	uint64_t         track_addr     = 0;
 	uint64_t         track_addr_end = 0;
 	ImageId          depth_id {};

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

@@ -378,6 +378,9 @@ inline constexpr std::array<VideoOutFormatPolicy, 8> VIDEO_OUT_FORMAT_POLICIES {
 		const auto encoded = static_cast<float>(value & 0xffu) / 255.0f;
 		return encoded <= 0.04045f ? encoded / 12.92f : std::pow((encoded + 0.055f) / 1.055f, 2.4f);
 	};
+	const auto unorm16 = [](uint64_t value) {
+		return static_cast<float>(value & 0xffffu) / 65535.0f;
+	};
 	const auto float16 = [](uint32_t value) {
 		const auto sign = (value & 0x8000u) << 16u;
 		const auto exponent = (value >> 10u) & 0x1fu;
@@ -403,19 +406,49 @@ inline constexpr std::array<VideoOutFormatPolicy, 8> VIDEO_OUT_FORMAT_POLICIES {
 			next.float32[1] = float16(packed >> 16u);
 			[[fallthrough]];
 		case vk::Format::eR16Sfloat: next.float32[0] = float16(packed); break;
-		case vk::Format::eR16G16Unorm:
-			next.float32[1] = static_cast<float>(packed >> 16u) / 65535.0f;
+		case vk::Format::eR16G16B16A16Unorm:
+			next.float32[2] = unorm16(clear_words >> 32u);
+			next.float32[3] = unorm16(clear_words >> 48u);
 			[[fallthrough]];
-		case vk::Format::eR16Unorm:
-			next.float32[0] = static_cast<float>(packed & 0xffffu) / 65535.0f;
+		case vk::Format::eR16G16Unorm:
+			next.float32[1] = unorm16(packed >> 16u);
+			[[fallthrough]];
+		case vk::Format::eR16Unorm: next.float32[0] = unorm16(packed); break;
+		case vk::Format::eR16G16B16A16Uint:
+			for (uint32_t i = 0; i < 4; i++) {
+				next.uint32[i] = static_cast<uint16_t>(clear_words >> (16u * i));
+			}
 			break;
-		// Each clear register stores one raw 32-bit float channel.
+		case vk::Format::eR16G16B16A16Sint:
+			for (uint32_t i = 0; i < 4; i++) {
+				next.int32[i] = static_cast<int16_t>(clear_words >> (16u * i));
+			}
+			break;
+		// Each clear register stores one raw 32-bit channel.
 		case vk::Format::eR32G32Sfloat:
 			next.float32[1] = std::bit_cast<float>(static_cast<uint32_t>(clear_words >> 32u));
 			[[fallthrough]];
 		case vk::Format::eR32Sfloat: next.float32[0] = std::bit_cast<float>(packed); break;
-		case vk::Format::eR32Uint: next.uint32[0] = packed; break;
-		case vk::Format::eR32Sint: next.int32[0] = static_cast<int32_t>(packed); break;
+		case vk::Format::eR32G32Uint:
+		case vk::Format::eR32G32Sint:
+			next.uint32[1] = static_cast<uint32_t>(clear_words >> 32u);
+			[[fallthrough]];
+		case vk::Format::eR32Uint:
+		case vk::Format::eR32Sint: next.uint32[0] = packed; break;
+		// Narrow targets hold their channels in the low bits of the first word. UFC 6 (PPSA23566)
+		// fast-clears an R8_UINT map through CMASK every frame and then writes only part of it;
+		// without these the clear was dropped and old contents piled up.
+		case vk::Format::eR8Uint: next.uint32[0] = packed & 0xffu; break;
+		case vk::Format::eR8Unorm: next.float32[0] = unorm8(packed); break;
+		case vk::Format::eR8G8Uint:
+			next.uint32[0] = packed & 0xffu;
+			next.uint32[1] = (packed >> 8u) & 0xffu;
+			break;
+		case vk::Format::eR8G8Unorm:
+			next.float32[0] = unorm8(packed);
+			next.float32[1] = unorm8(packed >> 8u);
+			break;
+		case vk::Format::eR16Uint: next.uint32[0] = packed & 0xffffu; break;
 		case vk::Format::eR8G8B8A8Srgb:
 			next.float32[0] = srgb8(packed);
 			next.float32[1] = srgb8(packed >> 8u);
@@ -462,10 +495,16 @@ inline constexpr std::array<VideoOutFormatPolicy, 8> VIDEO_OUT_FORMAT_POLICIES {
 // Keep this separate: the first register word alone cannot describe a 64-bit color.
 [[nodiscard]] inline bool DecodeColorDwordFill(vk::Format format, uint32_t packed,
                                                vk::ClearColorValue& clear) {
-	if ((format == vk::Format::eR16Sfloat || format == vk::Format::eR16Unorm) &&
+	if ((format == vk::Format::eR16Sfloat || format == vk::Format::eR16Unorm ||
+	     format == vk::Format::eR16Uint || format == vk::Format::eR8G8Unorm ||
+	     format == vk::Format::eR8G8Uint) &&
 	    (packed & 0xffffu) != (packed >> 16u)) {
 		return false;
 	}
+	if ((format == vk::Format::eR8Unorm || format == vk::Format::eR8Uint) &&
+	    packed != (packed & 0xffu) * 0x01010101u) {
+		return false;
+	}
 	if (format == vk::Format::eR16G16B16A16Sfloat) {
 		if (packed != 0) {
 			return false;

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

@@ -484,7 +484,7 @@ void TileManager::ConvertD16(Result source, Result target, D16Direction directio
 	                            vk::AccessFlagBits::eHostWrite |
 	                            vk::AccessFlagBits::eTransferWrite |
 	                            vk::AccessFlagBits::eShaderWrite;
-	barriers[1].dstAccessMask = vk::AccessFlagBits::eShaderRead | vk::AccessFlagBits::eShaderWrite;
+	barriers[1].dstAccessMask = vk::AccessFlagBits::eShaderWrite;
 	command.pipelineBarrier(
 	    vk::PipelineStageFlagBits::eAllCommands | vk::PipelineStageFlagBits::eHost,
 	    vk::PipelineStageFlagBits::eComputeShader, {}, 0, nullptr, 2, barriers, 0, nullptr);

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

@@ -604,8 +604,13 @@ TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageR
 	                                 view_levels, desc.info.resources.layers);
 	desc.view_info.base_level = view_base;
 	desc.type = storage ? TextureCache::BindingType::Storage : TextureCache::BindingType::Texture;
+	desc.heap = resource.heap;
 
-	auto       id                  = texture_cache.FindImage(desc, shader_conversion);
+	auto id = texture_cache.FindImage(desc, shader_conversion);
+	if (!id) {
+		// Another extent at the address holds it; the heap retries once an image is created.
+		return {.desc = std::move(desc), .layout = vk::ImageLayout::eGeneral};
+	}
 	auto*      image               = &texture_cache.GetImage(id);
 	const bool stencil_association = static_cast<bool>(image->depth_id);
 	if (stencil_association) {
@@ -672,9 +677,7 @@ void RenderExecutor::BindRenderTarget(ImageId id) {
 
 void RenderExecutor::ResetBindings() {
 	for (const auto id: m_bound_images) {
-		if (auto* image = m_context.GetTextureCache().m_slot_images.try_get(id); image != nullptr) {
-			image->binding = {};
-		}
+		m_context.GetTextureCache().m_slot_images[id].binding = {};
 	}
 	m_bound_images.clear();
 }
@@ -683,7 +686,7 @@ void RenderExecutor::PrepareBindings(const ShaderStageRuntime& runtime,
                                      PreparedBindings& prepared) {
 	KYTY_PROFILER_FUNCTION();
 	EXIT_IF(!runtime);
-	const auto& program  = *runtime.program;
+	const auto& program  = runtime.module->info;
 	const auto& snapshot = *runtime.resources;
 	prepared.runtime = &runtime;
 	prepared.gds = {nullptr, 0, VK_WHOLE_SIZE};
@@ -725,7 +728,7 @@ void RenderExecutor::FindBuffers(std::span<PreparedBindings* const> stages) {
 	for (auto* stage: stages) {
 		auto& prepared = *stage;
 		EXIT_IF(prepared.runtime == nullptr || !*prepared.runtime);
-		const auto& program  = *prepared.runtime->program;
+		const auto& program  = prepared.runtime->module->info;
 		const auto& snapshot = *prepared.runtime->resources;
 		const auto& layout = program.bindings;
 		prepared.buffer_sources.resize(layout.descriptor_counts[static_cast<size_t>(BindingKind::Buffers)]);
@@ -782,7 +785,7 @@ void RenderExecutor::FindBuffers(std::span<PreparedBindings* const> stages) {
 void RenderExecutor::RebindBuffers(PreparedBindings& prepared) {
 	KYTY_PROFILER_FUNCTION();
 	EXIT_IF(prepared.runtime == nullptr || !*prepared.runtime);
-	const auto& program   = *prepared.runtime->program;
+	const auto& program   = prepared.runtime->module->info;
 	const auto& snapshot  = *prepared.runtime->resources;
 	const auto& layout    = program.bindings;
 	EXIT_IF(prepared.buffer_sources.size() != layout.descriptor_counts[static_cast<size_t>(BindingKind::Buffers)]);
@@ -803,6 +806,9 @@ void RenderExecutor::RebindBuffers(PreparedBindings& prepared) {
 		prepared.buffers[index] = NativeStorageBuffer(m_context, prepared.buffer_sources[index],
 		                                              resource, buffer_offset);
 		pack_memory_offset(index, buffer_offset);
+		// Null bindings allow 0 bytes, not the shared null buffer's 16, so no store reaches it.
+		prepared.shader_data[layout.MemorySizeDword() + index] =
+		    static_cast<uint32_t>(prepared.buffer_sources[index].size + buffer_offset);
 	}
 	if (layout.descriptor_counts[static_cast<size_t>(BindingKind::FlattenedSrt)] != 0) {
 		prepared.flattened_srt = NativeUpload(m_context, snapshot.flattened_srt);
@@ -815,7 +821,7 @@ void RenderExecutor::RebindBuffers(PreparedBindings& prepared) {
 void RenderExecutor::RebindImages(PreparedBindings& prepared) {
 	KYTY_PROFILER_FUNCTION();
 	EXIT_IF(prepared.runtime == nullptr || !*prepared.runtime);
-	const auto& program  = *prepared.runtime->program;
+	const auto& program  = prepared.runtime->module->info;
 	const auto& snapshot = *prepared.runtime->resources;
 	auto&       images   = prepared.images;
 	EXIT_IF(images.size() != program.info.images.size());
@@ -823,12 +829,9 @@ void RenderExecutor::RebindImages(PreparedBindings& prepared) {
 	// Acquire each view before a later overlapping descriptor can replace its image.
 	for (uint32_t i = 0; i < program.info.images.size(); i++) {
 		if (!images[i].image_id) continue;
-		const auto old_image = texture_cache.m_slot_images.try_get(images[i].image_id);
-		if (old_image == nullptr || !old_image->registered ||
-		    old_image->binding.needs_rebind) {
-			if (old_image != nullptr) {
-				old_image->binding = {};
-			}
+		auto& old_image = texture_cache.m_slot_images[images[i].image_id];
+		if (!old_image.registered || old_image.binding.needs_rebind) {
+			old_image.binding = {};
 			images[i] = ResolveTexture(program.info.images[i], snapshot.images[i]);
 			BindImage(images[i].image_id,
 			          images[i].desc.type == TextureCache::BindingType::Storage);
@@ -864,8 +867,8 @@ void RenderExecutor::PrepareGraphicsBindings(std::span<PreparedBindings* const>
 	bool uses_image_heap = false;
 	FindBuffers(stages);
 	for (auto* stage: stages) {
-		uses_dma |= stage->runtime->program->info.uses_dma;
-		uses_image_heap |= stage->runtime->program->info.uses_image_heap;
+		uses_dma |= stage->runtime->module->info.info.uses_dma;
+		uses_image_heap |= stage->runtime->module->info.info.uses_image_heap;
 	}
 	if (uses_dma) {
 		m_context.PrepareBda();
@@ -879,12 +882,9 @@ void RenderExecutor::PrepareGraphicsBindings(std::span<PreparedBindings* const>
 	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->binding.needs_rebind) {
-			if (old_image != nullptr) {
-				old_image->binding = {};
-			}
+		auto& old_image = cache.m_slot_images[target.image_id];
+		if (!old_image.registered || old_image.binding.needs_rebind) {
+			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);
@@ -917,7 +917,7 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 	                                       : vk::ShaderStageFlags {};
 	for (const auto* prepared: prepared_bindings) {
 		EXIT_IF(prepared == nullptr || prepared->runtime == nullptr || !*prepared->runtime);
-		const auto& program = *prepared->runtime->program;
+		const auto& program = prepared->runtime->module->info;
 		write_count += std::popcount(program.bindings.descriptor_mask);
 		for (auto mask = program.bindings.descriptor_mask; mask != 0; mask &= mask - 1) {
 			const auto index = std::countr_zero(mask);
@@ -938,17 +938,16 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 		const auto& reads = reader->runtime->resources->specialization_reads;
 		if (reads.empty()) continue;
 		for (const auto* writer: prepared_bindings) {
-			if (writer->runtime->program->has_address_writes) {
+			if (writer->runtime->module->info.has_address_writes) {
 				EXIT("scalar resource reads cannot be proven disjoint from shader address writes\n");
 			}
 		}
 		for (const auto [address, size]: reads) {
 			for (const auto id: m_bound_images) {
-				const auto* image = m_context.GetTextureCache().m_slot_images.try_get(id);
-				if (image == nullptr ||
-				    (!image->binding.shader_write && !image->binding.is_target)) continue;
-				for (const auto written: {image->info.data, image->info.stencil,
-				                          image->info.metadata.range}) {
+				const auto& image = m_context.GetTextureCache().m_slot_images[id];
+				if (!image.binding.shader_write && !image.binding.is_target) continue;
+				for (const auto written: {image.info.data, image.info.stencil,
+				                          image.info.metadata.range}) {
 					if (written.size != 0 && ImageRangeOverlaps(address, size,
 					                                          written.address, written.size)) {
 						EXIT("scalar resource reads overlap an image or attachment write\n");
@@ -956,7 +955,7 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 				}
 			}
 			for (const auto* writer: prepared_bindings) {
-				const auto& program = *writer->runtime->program;
+				const auto& program = writer->runtime->module->info;
 				for (const auto& resource: program.info.buffers) {
 					if (!resource.written || resource.descriptor_index == UINT32_MAX) continue;
 					const auto& written = writer->buffer_sources[resource.descriptor_index];
@@ -982,7 +981,7 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 	std::array<vk::DescriptorBufferInfo, 2> address_buffers;
 
 	for (auto* prepared: prepared_bindings) {
-		const auto& program       = *prepared->runtime->program;
+		const auto& program       = prepared->runtime->module->info;
 		auto&       descriptors   = *prepared;
 		const auto  shader_stage  = NativeShaderStage(program.stage);
 		const auto  shader_stages = ShaderPipelineStages(shader_stage);

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

@@ -12,6 +12,7 @@
 #include "kernel/memory.h"
 
 #include <algorithm>
+#include <bit>
 #include <xxhash.h>
 
 namespace Libs::Graphics {
@@ -19,7 +20,7 @@ namespace {
 
 namespace IR = ShaderRecompiler::IR;
 
-// Smaller tables are cheap to materialize on each draw and are often transient.
+// Smaller tables are often written for one draw, so they enter on a second draw of a frame.
 constexpr uint64_t MinTableBytes = 64ull * 1024ull;
 constexpr uint32_t ElementMask   = (1u << IR::ImageHeapKindShift) - 1u;
 
@@ -44,6 +45,20 @@ void Widen(size_t& begin, size_t& end, size_t first, size_t last) {
 	end              = empty ? last : std::max(end, last);
 }
 
+// Probes at first + k * step whose first dword lies within a table of size bytes.
+uint64_t Probes(uint64_t size, uint64_t first, uint64_t step) {
+	return size >= first + 4u ? (size - first - 4u) / step + 1u : 0u;
+}
+
+// Tables grow a little on most frames, so their slot words get room to double.
+std::unique_ptr<Buffer> SlotBuffer(RenderContext& context, uint64_t size) {
+	return std::make_unique<Buffer>(
+	    context.GetGraphics(), context.GetCommandScheduler(), MemoryUsage::DeviceLocal, 0,
+	    vk::BufferUsageFlagBits::eTransferDst | vk::BufferUsageFlagBits::eStorageBuffer |
+	        vk::BufferUsageFlagBits::eShaderDeviceAddress,
+	    std::bit_ceil(size + 4u));
+}
+
 // Guest bytes of a table that its tracker pages [begin, end) cover.
 std::pair<uint64_t, uint64_t> PageRange(uint64_t base, uint64_t size, size_t begin, size_t end) {
 	const auto origin = Common::AlignDown(base, TRACKER_PAGE_SIZE);
@@ -118,26 +133,25 @@ ImageHeap::~ImageHeap() {
 bool ImageHeap::Acquire(const IR::ImageHeapRequest& request, IR::ImageHeapView& result) {
 	KYTY_PROFILER_FUNCTION();
 	const auto size = request.size & ~uint64_t {3};
-	if (size < MinTableBytes || request.count == 0u || request.step == 0u ||
+	if (request.count == 0u || request.step == 0u ||
 	    ((request.base | request.first | request.step) & 3u) != 0u) {
 		return false;
 	}
-	auto memory = std::ranges::find_if(m_memories, [&](const auto& candidate) {
-		return candidate->base == request.base && candidate->size == size;
-	});
-	if (memory == m_memories.end()) {
+	auto memory = std::ranges::find_if(
+	    m_memories, [&](const auto& candidate) { return candidate->base == request.base; });
+	if (memory == m_memories.end() && size < MinTableBytes &&
+	    m_sightings.try_emplace(request.base, m_draw).first->second == m_draw) {
+		return false;
+	}
+	// A table follows its descriptor's size, as the first program of a draw acquires it.
+	if (memory == m_memories.end() || ((*memory)->size != size && (*memory)->draw != m_draw)) {
 		if (!m_context.IsMapped(request.base, size)) return false;
-		auto created   = std::make_unique<Memory>();
-		created->base  = request.base;
-		created->size  = size;
-		created->words.resize(size / 4u);
-		created->dead.resize(size / 4u);
-		created->dirty_pages.assign((Common::AlignUp(request.base + size, TRACKER_PAGE_SIZE) -
-		                             Common::AlignDown(request.base, TRACKER_PAGE_SIZE)) /
-		                                TRACKER_PAGE_SIZE,
-		                            uint8_t {1});
-		std::scoped_lock lock {m_mutex};
-		memory = m_memories.insert(m_memories.end(), std::move(created));
+		if (memory == m_memories.end()) {
+			std::scoped_lock lock {m_mutex};
+			memory = m_memories.insert(m_memories.end(),
+			                           std::make_unique<Memory>(Memory {.base = request.base}));
+		}
+		Resize(**memory, size);
 	}
 	// A draw refreshes a table and a view takes its changes once, so all programs of the draw see
 	// one snapshot and one set of kinds.
@@ -146,27 +160,28 @@ bool ImageHeap::Acquire(const IR::ImageHeapRequest& request, IR::ImageHeapView&
 		(*memory)->draw = m_draw;
 	}
 	auto&      views   = (*memory)->views;
+	const auto bytes   = (*memory)->size;
 	const auto profile = IR::ImageHeapResource(*request.root, 0u);
+	const auto probes  = Probes(size, request.first, request.step);
+	const auto limit   = request.count < probes ? request.count : UINT64_MAX;
 	auto       view    = std::ranges::find_if(views, [&](const auto& candidate) {
 		return candidate->step == request.step && candidate->first == request.first &&
-		       candidate->count == request.count && candidate->profile == profile;
+		       candidate->limit == limit && candidate->profile == profile;
 	});
 	if (view == views.end()) {
 		const auto epoch   = Image::heap_epoch.load(std::memory_order_relaxed);
+		const auto count   = std::min(limit, Probes(bytes, request.first, request.step));
 		auto       created = std::make_unique<View>(View {.memory        = **memory,
 		                                                  .step          = request.step,
 		                                                  .first         = request.first,
-		                                                  .count         = request.count,
+		                                                  .count         = count,
+		                                                  .limit         = limit,
 		                                                  .profile       = profile,
 		                                                  .resolve_epoch = epoch,
 		                                                  .layout_epoch  = epoch});
-		created->slots.resize(size / 4u + 1u);
-		created->buffer = std::make_unique<Buffer>(
-		    m_context.GetGraphics(), m_context.GetCommandScheduler(), MemoryUsage::DeviceLocal, 0,
-		    vk::BufferUsageFlagBits::eTransferDst | vk::BufferUsageFlagBits::eStorageBuffer |
-		        vk::BufferUsageFlagBits::eShaderDeviceAddress,
-		    size + 4u);
-		UpdateProbes(*created, 0u, size);
+		created->slots.resize(bytes / 4u + 1u);
+		created->buffer = SlotBuffer(m_context, bytes);
+		UpdateProbes(*created, 0u, bytes);
 		// The first upload also clears the slots no probe selects.
 		created->dirty_begin = 0;
 		created->dirty_end   = created->slots.size();
@@ -254,10 +269,12 @@ void ImageHeap::Unwatch(const Memory& memory) {
 void ImageHeap::InvalidateMemory(uint64_t address, uint64_t size) {
 	std::scoped_lock lock {m_mutex};
 	for (auto& memory: m_memories) {
-		const auto begin = std::max(address, memory->base);
-		const auto end   = std::min(address + size, memory->base + memory->size);
-		if (begin >= end) continue;
+		// Watches cover whole pages, so a write next to the table must unwatch its page too.
 		const auto origin = Common::AlignDown(memory->base, TRACKER_PAGE_SIZE);
+		const auto begin  = std::max(address, origin);
+		const auto end    = std::min(address + size,
+		                             origin + memory->dirty_pages.size() * TRACKER_PAGE_SIZE);
+		if (begin >= end) continue;
 		MarkDirty(*memory, (begin - origin) / TRACKER_PAGE_SIZE,
 		          (end - 1u - origin) / TRACKER_PAGE_SIZE + 1u);
 	}
@@ -350,12 +367,17 @@ void ImageHeap::Release(View& view, uint32_t entry) {
 }
 
 void ImageHeap::Unreference(Element& element) {
-	if (auto* image = m_context.GetTextureCache().m_slot_images.try_get(element.binding.image_id)) {
+	if (auto* image = ImageOf(element)) {
 		image->heap_refs.fetch_sub(1u, std::memory_order_relaxed);
 		image->tick_accessed_last = m_context.GetCommandScheduler().CurrentTick();
 	}
 }
 
+Image* ImageHeap::ImageOf(const Element& element) const {
+	auto* image = m_context.GetTextureCache().m_slot_images.try_get(element.binding.image_id);
+	return image != nullptr && image->serial == element.serial ? image : nullptr;
+}
+
 uint32_t ImageHeap::Allocate() {
 	auto& scheduler = m_context.GetCommandScheduler();
 	while (!m_retired.empty() && scheduler.IsFree(m_retired.front().first)) {
@@ -376,15 +398,56 @@ uint32_t ImageHeap::Allocate() {
 }
 
 void ImageHeap::Retire(View& view) {
+	if (view.released.empty()) return;
+	// A retired index is reused once its tick ends, so no held entry may keep it.
+	std::erase_if(view.held, [&](uint32_t index) { return m_elements[index].refs == 0u; });
 	const auto tick = m_context.GetCommandScheduler().CurrentTick();
 	for (const auto index: view.released) m_retired.emplace_back(tick, index);
 	view.released.clear();
 }
 
+void ImageHeap::Resize(Memory& memory, uint64_t size) {
+	{
+		std::scoped_lock lock {m_mutex};
+		Unwatch(memory);
+		memory.size = size;
+		// The next refresh rereads the whole table.
+		memory.dirty_pages.assign((Common::AlignUp(memory.base + size, TRACKER_PAGE_SIZE) -
+		                           Common::AlignDown(memory.base, TRACKER_PAGE_SIZE)) /
+		                              TRACKER_PAGE_SIZE,
+		                          uint8_t {1});
+	}
+	memory.words.resize(size / 4u);
+	memory.dead.resize(size / 4u);
+	auto& scheduler = m_context.GetCommandScheduler();
+	for (auto& view: memory.views) {
+		const auto count = std::min(view->limit, Probes(size, view->first, view->step));
+		for (auto probe = count; probe < view->count; ++probe) {
+			Assign(*view, (view->first + probe * view->step) / 4u, 0u);
+		}
+		view->count       = count;
+		const auto slots  = view->slots.size();
+		view->slots.resize(size / 4u + 1u);
+		view->dirty_end   = std::min(view->dirty_end, view->slots.size());
+		view->dirty_begin = std::min(view->dirty_begin, view->dirty_end);
+		if (view->slots.size() * sizeof(uint32_t) > view->buffer->Size()) {
+			if (scheduler.Active()) {
+				scheduler.DeferOperation([buffer = std::move(view->buffer)] {});
+			}
+			view->buffer = SlotBuffer(m_context, size);
+			Widen(view->dirty_begin, view->dirty_end, 0u, view->slots.size());
+		} else if (slots < view->slots.size()) {
+			// Device words past the old end may hold stale slots.
+			Widen(view->dirty_begin, view->dirty_end, slots, view->slots.size());
+		}
+	}
+}
+
 void ImageHeap::Rebind(RenderExecutor& executor) {
 	KYTY_PROFILER_FUNCTION();
 	auto&      textures = m_context.GetTextureCache();
 	const auto epoch    = Image::heap_epoch.load(std::memory_order_relaxed);
+	const auto serial   = Image::serials.load(std::memory_order_relaxed);
 	for (auto* view: m_active) {
 		bool       stale = false;
 		const auto mark  = [&](uint32_t index) { m_elements[index].stale = stale = true; };
@@ -393,7 +456,7 @@ void ImageHeap::Rebind(RenderExecutor& executor) {
 			for (const auto& [value, index]: view->elements) {
 				const auto& binding = m_elements[index].binding;
 				if (!binding.image_id) continue;
-				const auto* image = textures.m_slot_images.try_get(binding.image_id);
+				const auto* image = ImageOf(m_elements[index]);
 				if (image == nullptr || !image->registered || image->depth_id ||
 				    image->binding.needs_rebind) {
 					mark(index);
@@ -413,6 +476,19 @@ void ImageHeap::Rebind(RenderExecutor& executor) {
 			mark(index);
 			return true;
 		});
+		// Held elements retry once an image is created, such as a render target of their extent.
+		if (view->held_serial != serial) {
+			view->held_serial = serial;
+			std::erase_if(view->held, [&](uint32_t index) {
+				const auto& element = m_elements[index];
+				if (element.refs == 0u || element.stale) return true;
+				const auto binding = executor.ResolveTexture(
+				    IR::ImageHeapResource(view->profile, element.kind), element.value);
+				if (binding.image_id) mark(index);
+				// An unmapped texture stays null.
+				return binding.image_id || binding.desc.info.data.Empty();
+			});
+		}
 		if (stale) {
 			// A changed image needs a new element; pending draws still read the old one. The new
 			// element keeps the old value, since a view takes table changes only when acquired.
@@ -434,12 +510,15 @@ void ImageHeap::Rebind(RenderExecutor& executor) {
 			if (binding.image_id) {
 				auto& image = textures.GetImage(binding.image_id);
 				image.heap_refs.fetch_add(1u, std::memory_order_relaxed);
+				m_elements[index].serial = image.serial;
 				binding.layout     = image.ReadLayout();
 				binding.image_view = textures.FindTexture(binding.image_id, binding.desc);
 				if (binding.desc.info.metadata.kind == ImageMetadataKind::Dcc) {
 					view->dcc.push_back(index);
 				}
 				view->away.push_back(index);
+			} else if (!binding.desc.info.data.Empty()) {
+				view->held.push_back(index);
 			}
 			m_elements[index].binding = std::move(binding);
 			m_writes.push_back(index);
@@ -471,8 +550,7 @@ void ImageHeap::WriteDescriptors() {
 void ImageHeap::Commit(CommandBuffer& buffer, vk::PipelineBindPoint bind_point,
                        vk::PipelineLayout layout) {
 	KYTY_PROFILER_FUNCTION();
-	auto&      textures = m_context.GetTextureCache();
-	const auto epoch    = Image::heap_epoch.load(std::memory_order_relaxed);
+	const auto epoch = Image::heap_epoch.load(std::memory_order_relaxed);
 	for (auto* view: m_active) {
 		if (view->dirty_begin != view->dirty_end) {
 			const auto slots = std::span(view->slots).subspan(view->dirty_begin,
@@ -493,7 +571,7 @@ void ImageHeap::Commit(CommandBuffer& buffer, vk::PipelineBindPoint bind_point,
 		// this draw's discovery replaced lives until the tick ends, and its slots still select it.
 		std::erase_if(view->away, [&](uint32_t index) {
 			const auto& element = m_elements[index];
-			auto*       image   = textures.m_slot_images.try_get(element.binding.image_id);
+			auto*       image   = ImageOf(element);
 			if (element.refs == 0u || image == nullptr) return true;
 			if (image->binding.is_target || image->binding.shader_write ||
 			    image->binding.force_general) {
@@ -519,26 +597,42 @@ void ImageHeap::UnmapMemory(uint64_t address, uint64_t size) {
 	} else {
 		m_unmapped.push_back({address, size});
 	}
-	auto&           scheduler = m_context.GetCommandScheduler();
 	std::scoped_lock lock {m_mutex};
 	std::erase_if(m_memories, [&](const std::unique_ptr<Memory>& memory) {
 		if (address >= memory->base + memory->size || memory->base >= address + size) {
 			return false;
 		}
-		Unwatch(*memory);
-		for (auto& view: memory->views) {
-			std::erase(m_active, view.get());
-			for (const auto& [value, index]: view->elements) {
-				Unreference(m_elements[index]);
-				view->released.push_back(index);
-			}
-			Retire(*view);
-			if (scheduler.Active()) {
-				scheduler.DeferOperation([buffer = std::move(view->buffer)] {});
-			}
-		}
+		Drop(*memory);
 		return true;
 	});
 }
 
+void ImageHeap::RunGarbageCollector() {
+	// A smaller table rebuilt each frame moves to a new base. Larger ones keep their dead words.
+	std::scoped_lock lock {m_mutex};
+	std::erase_if(m_memories, [&](const std::unique_ptr<Memory>& memory) {
+		if (memory->size >= MinTableBytes || memory->draw > m_suspend_draw) return false;
+		Drop(*memory);
+		return true;
+	});
+	m_suspend_draw = m_draw;
+	m_sightings.clear();
+}
+
+void ImageHeap::Drop(Memory& memory) {
+	auto& scheduler = m_context.GetCommandScheduler();
+	Unwatch(memory);
+	for (auto& view: memory.views) {
+		std::erase(m_active, view.get());
+		for (const auto& [value, index]: view->elements) {
+			Unreference(m_elements[index]);
+			view->released.push_back(index);
+		}
+		Retire(*view);
+		if (scheduler.Active()) {
+			scheduler.DeferOperation([buffer = std::move(view->buffer)] {});
+		}
+	}
+}
+
 } // namespace Libs::Graphics

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

@@ -22,10 +22,11 @@ class RenderExecutor;
 struct ImageHeapTestAccess;
 
 // Guest image tables that shaders index by GPU-computed offsets, kept resident in one bindless
-// sampled-image array. Each view of a table owns device slot words, one per table dword, that
-// select an array element. CPU writes to a table fault in its snapshot, which each draw refreshes
-// once; a view takes the changes when a draw first acquires it. Changes to the images behind
-// elements bump Image::heap_epoch.
+// sampled-image array. A table is keyed by its base and follows its descriptor's size. Each view
+// of a table owns device slot words, one per table dword, that select an array element. CPU
+// writes to a table fault in its snapshot, which each draw refreshes once; a view takes the
+// changes when a draw first acquires it. Changes to the images behind elements bump
+// Image::heap_epoch. A smaller table leaves at a suspend point if the frame did not acquire it.
 class ImageHeap final: public ShaderRecompiler::IR::ImageHeapSource {
 public:
 	ImageHeap(RenderContext& context, PageManager& page_manager);
@@ -46,6 +47,8 @@ public:
 	// CPU write fault, on any thread.
 	void InvalidateMemory(uint64_t address, uint64_t size);
 	void UnmapMemory(uint64_t address, uint64_t size);
+	// At a suspend point.
+	void RunGarbageCollector();
 
 private:
 	using DescriptorValue = ShaderRecompiler::IR::DescriptorValue;
@@ -57,9 +60,10 @@ private:
 	struct Element {
 		DescriptorValue value;
 		TextureBinding  binding;
-		uint32_t        kind  = 0;
-		uint32_t        refs  = 0; // Slots of the owning view that select it.
-		bool            stale = false;
+		uint32_t        kind   = 0;
+		uint32_t        refs   = 0; // Slots of the owning view that select it.
+		bool            stale  = false;
+		uint64_t        serial = 0; // Image::serial of binding.image_id.
 	};
 
 	struct Memory;
@@ -69,6 +73,7 @@ private:
 		uint64_t                            step  = 0;
 		uint64_t                            first = 0;
 		uint64_t                            count = 0;
+		uint64_t                            limit = 0; // Probes a key domain allows, if fewer.
 		ShaderRecompiler::IR::ImageResource profile;
 		std::vector<uint32_t>               slots; // Plus a trailing null slot.
 		std::vector<uint32_t>               kind_keys;
@@ -78,6 +83,7 @@ private:
 		std::vector<uint32_t>   released; // No slot selects them after the next upload.
 		std::vector<uint32_t>   away;     // Not yet moved to the read layout.
 		std::vector<uint32_t>   dcc;      // Color metadata clears apply on each discovery.
+		std::vector<uint32_t>   held;     // Another extent holds their address.
 		std::unique_ptr<Buffer> buffer;
 		size_t                  dirty_begin   = 0; // Slots to upload.
 		size_t                  dirty_end     = 0;
@@ -85,6 +91,7 @@ private:
 		size_t                  changed_end   = 0;
 		uint64_t                resolve_epoch = 0;
 		uint64_t                layout_epoch  = 0;
+		uint64_t                held_serial   = 0; // Image::serials when held was last retried.
 	};
 
 	struct Memory {
@@ -102,6 +109,7 @@ private:
 	template <bool track>
 	void     Watch(const Memory& memory, size_t begin, size_t end);
 	void     Unwatch(const Memory& memory);
+	void     Drop(Memory& memory);
 	void     UpdateProbes(View& view, uint64_t begin, uint64_t end);
 	void     UpdateProbe(View& view, uint64_t probe);
 	void     Assign(View& view, size_t slot, uint32_t entry);
@@ -109,8 +117,10 @@ private:
 	uint32_t Intern(View& view, uint32_t kind, const DescriptorValue& value);
 	void     Release(View& view, uint32_t entry);
 	void     Unreference(Element& element);
+	Image*   ImageOf(const Element& element) const;
 	uint32_t Allocate();
 	void     Retire(View& view);
+	void     Resize(Memory& memory, uint64_t size);
 	void     WriteDescriptors();
 
 	RenderContext&                       m_context;
@@ -118,9 +128,11 @@ private:
 	std::mutex                           m_mutex; // Fault-visible Memory state and m_memories.
 	std::vector<std::unique_ptr<Memory>> m_memories;
 	std::vector<View*>                   m_active; // Acquired since the last commit.
-	uint64_t                             m_draw = 1;
+	uint64_t                             m_draw         = 1;
+	uint64_t                             m_suspend_draw = 0; // m_draw at the last suspend point.
 	std::vector<Element>                 m_elements;
 	std::vector<uint32_t>                m_free;
+	std::unordered_map<uint64_t, uint64_t>    m_sightings; // Smaller tables: base, first draw.
 	std::deque<std::pair<uint64_t, uint32_t>> m_retired; // Tick, element.
 	std::vector<std::pair<size_t, size_t>>    m_runs;
 	std::vector<uint32_t>                     m_scratch;

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

@@ -4,50 +4,62 @@
 #include "common/abi.h"
 #include "common/assert.h"
 #include "common/common.h"
+#include "common/file.h"
+#include "common/uniqueFunction.h"
+#include "graphics/host_gpu/renderer/colorRenderTarget.h"
+#include "graphics/host_gpu/renderer/depthRenderTarget.h"
 #include "graphics/host_gpu/renderer/renderTarget.h"
 #include "graphics/host_gpu/vulkanCommon.h"
+#include "graphics/shader/recompiler/ir/passes/ResourceMaterialization.h"
+#include "graphics/shader/rectListShader.h"
 #include "graphics/shader/shader.h"
 
 #include <chrono>
+#include <condition_variable>
 #include <cstddef>
 #include <filesystem>
 #include <memory>
+#include <mutex>
+#include <queue>
 #include <span>
+#include <thread>
 #include <type_traits>
 #include <unordered_map>
+#include <vector>
 
 namespace Libs::Graphics {
 
 struct GraphicContext;
-struct RenderColorInfo;
-struct RenderDepthInfo;
+struct ShaderParams;
 class CommandBuffer;
 
-namespace HW {
-class Context;
-class Shader;
-class UserConfig;
-struct ComputeShaderInfo;
-} // namespace HW
+namespace ShaderRecompiler {
+struct CompileOptions;
+struct ShaderSource;
+struct TranslateResult;
+} // namespace ShaderRecompiler
 
 namespace ShaderRecompiler::IR {
 class ImageHeapSource;
 } // namespace ShaderRecompiler::IR
 
+namespace ShaderRecompiler::Spirv {
+struct Profile;
+} // namespace ShaderRecompiler::Spirv
+
 #pragma pack(push, 1)
 
 struct PipelineStaticParameters {
 	bool                       negative_one_to_one      = false;
 	bool                       depth_clip_enable        = true;
 	vk::PrimitiveTopology      topology                 = vk::PrimitiveTopology::ePointList;
+	// 0 without tessellation, 3 for a rectangle list, else the input control points.
+	uint32_t                   patch_control_points     = 0;
 	bool                       primitive_restart_enable = false;
 	uint32_t                   samples                  = 1;
 	bool                       sample_shading_enable    = false;
 	bool                       alpha_to_coverage_enable = false;
 	uint32_t                   color_mask[RENDER_COLOR_ATTACHMENTS_MAX]           = {};
-	bool                       cull_front                                         = false;
-	bool                       cull_back                                          = false;
-	bool                       face                                               = false;
 	bool                       provoking_vtx_last                                 = false;
 	vk::PolygonMode            polygon_mode                                       = vk::PolygonMode::eFill;
 	uint8_t                    color_srcblend[RENDER_COLOR_ATTACHMENTS_MAX]       = {};
@@ -58,57 +70,110 @@ struct PipelineStaticParameters {
 	uint8_t                    alpha_destblend[RENDER_COLOR_ATTACHMENTS_MAX]      = {};
 	bool                       separate_alpha_blend[RENDER_COLOR_ATTACHMENTS_MAX] = {};
 	bool                       blend_enable[RENDER_COLOR_ATTACHMENTS_MAX]         = {};
-
-	bool operator==(const PipelineStaticParameters& other) const noexcept;
 };
 
-#pragma pack(pop)
-
-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) == 117);
-
 struct PipelineRenderingState {
 	std::array<vk::Format, RENDER_COLOR_ATTACHMENTS_MAX> color_formats {};
 	vk::Format                                           depth_format   = vk::Format::eUndefined;
 	vk::Format                                           stencil_format = vk::Format::eUndefined;
 	uint32_t                                             color_count    = 0;
-
-	bool operator==(const PipelineRenderingState&) const = default;
 };
 
 struct PipelineVertexInputState {
 	struct Binding {
-		uint32_t stride                           = 0;
-		bool     instance                         = false;
-		bool     operator==(const Binding&) const = default;
+		uint32_t stride   = 0;
+		uint8_t  instance = 0;
 	};
 	struct Attribute {
-		uint32_t offset                             = 0;
-		uint8_t  binding                            = 0;
-		bool     operator==(const Attribute&) const = default;
+		uint32_t offset  = 0;
+		uint8_t  binding = 0;
 	};
 
 	std::array<Binding, ShaderVertexInputInfo::RES_MAX>   bindings {};
 	std::array<Attribute, ShaderVertexInputInfo::RES_MAX> attributes {};
 	uint8_t                                               binding_count   = 0;
 	uint8_t                                               attribute_count = 0;
-
-	bool operator==(const PipelineVertexInputState&) const = default;
 };
 
-struct ShaderProgram {
-	uint64_t         id     = 0;
-	vk::ShaderModule module = nullptr;
+// Everything a graphics pipeline is made of besides the stage modules its ids name. Aligned so
+// that rendering.color_formats can be passed to Vulkan.
+struct alignas(4) GraphicsPipelineKey {
+	// VS, LS or mesh; HS; TES; PS. 0 marks an absent stage.
+	std::array<uint64_t, 4>  stage_ids {};
+	PipelineRenderingState   rendering;
+	PipelineVertexInputState vertex_input;
+	PipelineStaticParameters static_params;
 
-	explicit operator bool() const { return id != 0 && module != nullptr; }
+	bool operator==(const GraphicsPipelineKey& other) const;
+};
+
+#pragma pack(pop)
+
+// Keys are compared and hashed as bytes.
+static_assert(std::has_unique_object_representations_v<GraphicsPipelineKey>);
+
+// A compiled stage with what pipeline creation needs from it besides the key.
+struct ShaderModule {
+	uint64_t                                     id     = 0;
+	vk::ShaderModule                             module = nullptr;
+	ShaderRecompiler::IR::CompiledShaderInfo     info;
+	ShaderRecompiler::IR::ResourceSpecialization specialization;
+	std::vector<uint32_t>                        function_code;
+	// VS or LS: the host format of each vertex input attribute.
+	std::vector<vk::Format>                      vertex_formats;
+	// Rectangle-list PS: the parameters it reads.
+	std::vector<RectListParameter>               rect_parameters;
+};
+
+// A draw's topology, shader stages (from PipelineCache::GetGraphicsPrograms) and targets.
+struct DrawRenderState {
+	RenderDepthInfo                      depth_info;
+	RenderColorInfo                      color_info[RENDER_COLOR_ATTACHMENTS_MAX] = {};
+	uint32_t                             color_count                              = 0;
+	// Color targets the pixel program writes.
+	uint32_t                             mrt_mask                                 = 0;
+	vk::PrimitiveTopology                topology = vk::PrimitiveTopology::ePointList;
+	bool                                 primitive_restart_enable                 = false;
+	bool                                 ps_active                                = false;
+	std::array<ShaderVertexInputInfo, 3> vertex_info;
+	ShaderPixelInputInfo                 ps_input_info;
+	// Input: the user-data slots a GPU-read indirect draw patches.
+	ShaderIndirectOffsets                indirect_offsets;
+
+	[[nodiscard]] std::span<const ShaderVertexInputInfo> VertexStages() const {
+		return {vertex_info.data(), vertex_info[1].stage ? 3u : 1u};
+	}
+};
+
+// Writes the shader store on its own thread, running jobs in the order they were posted.
+class ShaderStoreWriter {
+public:
+	// Keeps the first `kept` bytes of the store at `path`, or starts it over with `header`. The
+	// thread runs only if that succeeded.
+	void Open(const std::filesystem::path& path, std::span<const uint8_t> header, uint64_t kept);
+	explicit operator bool() const { return m_thread.joinable(); }
+	// Compresses `record` and appends it on the store thread, in call order.
+	void Append(std::vector<uint8_t>&& record);
+
+private:
+	friend struct PipelineCacheTestAccess;
+
+	void Post(Common::UniqueFunction<void>&& job);
+
+	Common::File                             m_file;
+	std::mutex                               m_mutex;
+	std::condition_variable_any              m_ready;
+	std::queue<Common::UniqueFunction<void>> m_jobs;
+	// Declared last, so it drains the queue and joins before the rest is destroyed.
+	std::jthread m_thread;
 };
 
 // The owning renderer serializes access, including saves while the GPU is running.
 class PipelineCache {
 public:
-	PipelineCache(GraphicContext& graphics, ShaderRecompiler::IR::ImageHeapSource* image_heap);
+	// Without a shader store path nothing is kept across runs.
+	PipelineCache(GraphicContext& graphics, ShaderRecompiler::IR::ImageHeapSource* image_heap,
+	              const std::filesystem::path& shader_store = {});
 	~PipelineCache();
 	KYTY_CLASS_NO_COPY(PipelineCache);
 	void Save();
@@ -121,87 +186,106 @@ public:
 		bool                    uses_image_heap       = false;
 	};
 
-	struct GraphicsPrograms {
-		std::array<ShaderProgram, 3> vertex;
-		ShaderProgram pixel;
+	// Reads the shaders and registers bound in `command` into the state's shader stages.
+	void GetGraphicsPrograms(const CommandBuffer& command, DrawRenderState& state);
+	// Null skips the dispatch.
+	const ShaderModule* GetComputeProgram(const HW::ComputeShaderInfo& regs,
+	                                      ShaderComputeInputInfo&      input_info);
 
-		[[nodiscard]] uint32_t VertexStageCount() const { return vertex[1] ? 3u : 1u; }
-	};
-
-	GraphicsPrograms
-	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, 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,
-	                              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);
+	Pipeline& GetGraphicsPipeline(const CommandBuffer& command, const DrawRenderState& state);
+	Pipeline& GetComputePipeline(const ShaderComputeInputInfo& input_info);
 
 private:
-	struct ProgramCache;
+	friend struct PipelineCacheTestAccess;
 
-	struct GraphicsPipelineKey {
-		PipelineRenderingState   rendering;
-		std::array<uint64_t, 3>  vertex_shader_ids {};
-		uint64_t                 ps_shader_id = 0;
-		PipelineVertexInputState vertex_input;
-		PipelineStaticParameters static_params;
+	struct Program {
+		std::unique_ptr<ShaderRecompiler::ShaderSource> call_source;
+		ShaderRecompiler::IR::ResourcePlan           resource_plan;
+		ShaderRecompiler::IR::ResourceSnapshot       resources;
+		ShaderRecompiler::IR::ResourceSpecialization specialization;
+		std::vector<ShaderModule>                    modules;
+		bool                                         skip_dispatch = false;
 
-		bool operator==(const GraphicsPipelineKey& other) const {
-			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;
-		}
+		[[nodiscard]] std::span<const uint32_t> FunctionCode() const;
+		bool RefreshCallSource(const ShaderRecompiler::IR::SrtRuntime& runtime);
+		template <bool HasCalls>
+		ShaderModule* FindModule(const ShaderRecompiler::IR::SrtRuntime& runtime,
+		                         uint32_t                                push_data_cursor);
 	};
 
-	struct PipelineKeyHash {
-		static void Mix(std::size_t& hash, std::size_t value) {
-			hash ^= value + static_cast<std::size_t>(0x9e3779b97f4a7c15ull) + (hash << 6u) +
-			        (hash >> 2u);
-		}
+	struct ProgramKey {
+		ShaderType            stage           = ShaderType::Unknown;
+		uint64_t              hash            = 0;
+		uint32_t              user_data_count = 0;
+		std::vector<uint32_t> static_state;
+
+		bool operator==(const ProgramKey&) const = default;
+	};
+
+	// Variants of one source share a bucket; ProgramKey equality compares the state.
+	struct ProgramKeyHash {
+		std::size_t operator()(const ProgramKey& key) const { return key.hash; }
 	};
 
 	struct GraphicsPipelineKeyHash {
 		std::size_t operator()(const GraphicsPipelineKey& key) const;
 	};
 
-	GraphicContext&               m_graphics;
-	std::unique_ptr<ProgramCache> m_program_cache;
+	using ModuleMap = std::unordered_map<uint64_t, const ShaderModule*>;
+
+	GraphicContext&                        m_graphics;
+	ShaderRecompiler::IR::ImageHeapSource* m_image_heap;
+	std::unordered_map<ProgramKey, Program, ProgramKeyHash> m_program_cache;
+	// Modules compiled per stage.
+	std::array<uint32_t, static_cast<size_t>(ShaderType::TessellationEvaluation) + 1> m_compiled {};
+	ProgramKey                             m_lookup_key;
 	vk::PipelineCache             m_driver_cache = nullptr;
 	std::filesystem::path         m_driver_cache_path;
 	std::chrono::steady_clock::time_point m_next_driver_cache_save;
 	bool                          m_driver_cache_dirty = false;
-	std::unordered_map<GraphicsPipelineKey, std::unique_ptr<Pipeline>, GraphicsPipelineKeyHash>
-	                                                        m_graphics_pipelines;
-	std::unordered_map<uint64_t, std::unique_ptr<Pipeline>> m_compute_pipelines;
+	std::unordered_map<GraphicsPipelineKey, Pipeline, GraphicsPipelineKeyHash> m_graphics_pipelines;
+	std::unordered_map<uint64_t, Pipeline>                                     m_compute_pipelines;
+	std::unordered_map<uint64_t, ShaderTessellationInputInfo>                  m_tess_layouts;
+	// Records everything first created; closed without a shader store.
+	ShaderStoreWriter m_store;
 
+	template <typename InputInfo>
+	const ShaderModule* GetProgram(const ShaderParams& params, InputInfo& input_info,
+	                               uint32_t& push_data_cursor);
+	ShaderModule& CompileModule(Program& program, const ProgramKey& key,
+	                            ShaderRecompiler::CompileOptions  options,
+	                            ShaderRecompiler::TranslateResult translated,
+	                            uint32_t                          push_data_cursor);
 	void InitializeDriverCache();
 	void CheckpointDriverCache();
+
+	// pipelineSerialization.cpp
+	void WarmUp(const std::filesystem::path& path);
+	bool LoadRecord(std::span<const uint8_t>& records, std::vector<GraphicsPipelineKey>& pipelines);
+	void StoreModule(const ProgramKey& key, const Program& program, const ShaderModule& module,
+	                 const std::vector<uint32_t>& spirv);
+	void StorePipeline(const GraphicsPipelineKey& key);
+	void StoreLayout(uint64_t key, const ShaderTessellationInputInfo& layout);
+	// Every module by id, with id 0 naming an absent stage.
+	[[nodiscard]] ModuleMap MapModules() const;
+	// False when a stage id of the key is not in `modules`.
+	bool CreateGraphicsPipeline(Pipeline& pipeline, const GraphicsPipelineKey& key,
+	                            const ModuleMap& modules);
 };
 
-void LogPipelineTrace(const char* phase, uint64_t vertex_program_id, uint64_t pixel_program_id);
+// The shader store of the running title; empty unless --shader-cache is on in a Release build
+// without debug dumps, which change the SPIR-V.
+std::filesystem::path ShaderStorePath();
+// Host capabilities that change the emitted SPIR-V.
+ShaderRecompiler::Spirv::Profile SpirvProfile(const GraphicContext& graphics);
+
+// Stages are indexed like GraphicsPipelineKey::stage_ids.
 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 GraphicsPipelineKey&                 key,
+                            const std::array<const ShaderModule*, 4>& stages,
+                            vk::PipelineCache                          driver_cache);
 void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& pipeline,
-                            const ShaderComputeInputInfo& input_info,
-                            vk::ShaderModule compute_module, vk::PipelineCache driver_cache);
+                            const ShaderModule& module, vk::PipelineCache driver_cache);
 
 } // namespace Libs::Graphics
 

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

@@ -0,0 +1,329 @@
+#include "graphics/host_gpu/renderer/pipeline/pipelineCache.h"
+
+#include "common/logging/log.h"
+#include "common/stringUtils.h"
+#include "graphics/host_gpu/vulkanCommon.h"
+#include "graphics/shader/recompiler/ShaderRecompiler.h"
+#include "graphics/shader/recompiler/ir/Serialize.h"
+
+#include <algorithm>
+#include <chrono>
+#include <cstring>
+#include <fmt/format.h>
+#include <ranges>
+#include <xxhash.h>
+#include <zstd.h>
+
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+#elif defined(__APPLE__)
+#include <mach-o/dyld.h>
+#endif
+
+// The shader store: a header, then zstd-compressed records of everything a run created first.
+// Records are only appended, so a store cut short by a crash keeps every complete record.
+namespace Libs::Graphics {
+
+namespace {
+
+using ShaderRecompiler::IR::IoAll;
+using ShaderRecompiler::IR::Reader;
+using ShaderRecompiler::IR::Writer;
+
+enum class Record : uint8_t { Module, GraphicsPipeline, TessellationLayout };
+
+// Precedes each record's zstd frame.
+struct Frame {
+	uint32_t compressed_size = 0;
+	uint32_t raw_size        = 0;
+	uint64_t checksum        = 0; // XXH3 of the compressed bytes.
+};
+
+// Changes with every build, so records never meet code that encodes them differently. Zero when
+// the executable cannot be found.
+uint64_t BuildStamp() {
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
+	std::array<wchar_t, 32768> name {};
+	GetModuleFileNameW(nullptr, name.data(), static_cast<DWORD>(name.size()));
+	const std::filesystem::path path = name.data();
+#elif defined(__APPLE__)
+	std::array<char, 4096> name {};
+	auto size = static_cast<uint32_t>(name.size());
+	_NSGetExecutablePath(name.data(), &size);
+	const std::filesystem::path path = name.data();
+#else
+	const std::filesystem::path path = "/proc/self/exe";
+#endif
+	std::error_code error;
+	const auto      time = std::filesystem::last_write_time(path, error).time_since_epoch().count();
+	return error ? 0 : static_cast<uint64_t>(time);
+}
+
+// What every record depends on besides its own fields; empty when the build is unknown.
+std::vector<uint8_t> StoreHeader(const GraphicContext& graphics) {
+	const auto stamp = BuildStamp();
+	if (stamp == 0) {
+		return {};
+	}
+	Writer header;
+	IoAll(header, stamp, SpirvProfile(graphics),
+	      graphics.GetPhysicalDeviceProperties().limits.maxComputeSharedMemorySize);
+	return std::move(header.bytes);
+}
+
+// Record fields, shared by writing and reading.
+template <class S, class Key> void IoKey(S& io, Key& key) {
+	auto& [stage, hash, user_data_count, static_state] = key;
+	IoAll(io, stage, hash, user_data_count, static_state);
+}
+
+template <class S, class Module> void IoModule(S& io, Module& stage) {
+	// The handle comes from the SPIR-V; stored programs have no function code.
+	auto& [id, handle, info, specialization, function_code, vertex_formats, rect_parameters] = stage;
+	IoAll(io, id, info, specialization, vertex_formats, rect_parameters);
+}
+
+} // namespace
+
+void ShaderStoreWriter::Open(const std::filesystem::path& path, std::span<const uint8_t> header,
+                             uint64_t kept) {
+	Common::File::CreateDirectories(path.parent_path());
+	if (kept == 0) {
+		if (m_file.Create(path)) {
+			m_file.Write(header.data(), static_cast<uint32_t>(header.size()));
+		}
+	} else if (m_file.Open(path, Common::File::Mode::ReadWrite)) {
+		m_file.Truncate(kept);
+		m_file.Seek(kept);
+	}
+	if (m_file.IsInvalid()) {
+		return;
+	}
+	m_thread = std::jthread([this](std::stop_token stop) {
+		std::unique_lock lock(m_mutex);
+		// Once stop is requested the wait returns false only when the queue is empty.
+		while (m_ready.wait(lock, stop, [this] { return !m_jobs.empty(); })) {
+			const auto job = std::move(m_jobs.front());
+			m_jobs.pop();
+			lock.unlock();
+			job();
+			lock.lock();
+		}
+	});
+}
+
+void ShaderStoreWriter::Append(std::vector<uint8_t>&& record) {
+	Post([this, record = std::move(record)] {
+		std::vector<uint8_t> bytes(sizeof(Frame) + ZSTD_compressBound(record.size()));
+		const auto size = ZSTD_compress(bytes.data() + sizeof(Frame), bytes.size() - sizeof(Frame),
+		                                record.data(), record.size(), 3);
+		EXIT_IF(ZSTD_isError(size));
+		const Frame frame {.compressed_size = static_cast<uint32_t>(size),
+		                   .raw_size        = static_cast<uint32_t>(record.size()),
+		                   .checksum        = XXH3_64bits(bytes.data() + sizeof(Frame), size)};
+		std::memcpy(bytes.data(), &frame, sizeof(frame));
+		// A failed write leaves a partial record; the next run drops it and everything after it.
+		m_file.Write(bytes.data(), static_cast<uint32_t>(sizeof(Frame) + size));
+	});
+}
+
+void ShaderStoreWriter::Post(Common::UniqueFunction<void>&& job) {
+	{
+		std::scoped_lock lock(m_mutex);
+		m_jobs.push(std::move(job));
+	}
+	m_ready.notify_one();
+}
+
+void PipelineCache::WarmUp(const std::filesystem::path& path) {
+	if (path.empty()) {
+		return;
+	}
+	const auto header = StoreHeader(m_graphics);
+	if (header.empty()) {
+		return;
+	}
+	const auto           start = std::chrono::steady_clock::now();
+	std::vector<uint8_t> bytes;
+	if (Common::File file;
+	    Common::File::IsFileExisting(path) && file.Open(path, Common::File::Mode::Read)) {
+		bytes.resize(file.Size());
+		uint32_t read = 0;
+		file.Read(bytes.data(), static_cast<uint32_t>(bytes.size()), &read);
+		bytes.resize(read);
+	}
+	Log::WriteToConsoleAndLog(
+	    fmt::format("Shader cache: compiling shaders and pipelines from {} ({:.1f} MB)\n",
+	                Common::PathToString(path), static_cast<double>(bytes.size()) / 1'000'000.0));
+	std::vector<GraphicsPipelineKey> pipelines;
+	auto                             records = std::span<const uint8_t>(bytes);
+	const bool reusable =
+	    records.size() >= header.size() && std::ranges::equal(records.first(header.size()), header);
+	if (reusable) {
+		records = records.subspan(header.size());
+		while (LoadRecord(records, pipelines)) {
+		}
+	}
+	// A damaged record ends the store; the records before it stay.
+	m_store.Open(path, header, reusable ? bytes.size() - records.size() : 0);
+
+	const auto modules = MapModules();
+	for (const auto& key: pipelines) {
+		Pipeline pipeline;
+		if (CreateGraphicsPipeline(pipeline, key, modules)) {
+			m_graphics_pipelines.emplace(key, pipeline);
+		}
+	}
+	// A compute pipeline depends only on its module.
+	for (const auto* module: modules | std::views::values) {
+		if (module != nullptr && module->info.stage == ShaderType::Compute) {
+			CreatePipelineInternal(m_graphics, m_compute_pipelines[module->id], *module,
+			                       m_driver_cache);
+		}
+	}
+	if (!m_graphics_pipelines.empty() || !m_compute_pipelines.empty()) {
+		CheckpointDriverCache();
+	}
+	Log::WriteToConsoleAndLog(fmt::format(
+	    "Shader cache: loaded {} programs, {} modules, {} pipelines in {} ms\n",
+	    m_program_cache.size(), modules.size() - 1,
+	    m_graphics_pipelines.size() + m_compute_pipelines.size(),
+	    std::chrono::duration_cast<std::chrono::milliseconds>(std::chrono::steady_clock::now() -
+	                                                          start)
+	        .count()));
+}
+
+bool PipelineCache::LoadRecord(std::span<const uint8_t>&        records,
+                               std::vector<GraphicsPipelineKey>& pipelines) {
+	Frame frame;
+	if (records.size() < sizeof(frame)) {
+		return false;
+	}
+	std::memcpy(&frame, records.data(), sizeof(frame));
+	if (records.size() - sizeof(frame) < frame.compressed_size) {
+		return false;
+	}
+	const auto compressed = records.subspan(sizeof(frame), frame.compressed_size);
+	if (XXH3_64bits(compressed.data(), compressed.size()) != frame.checksum ||
+	    ZSTD_getFrameContentSize(compressed.data(), compressed.size()) != frame.raw_size) {
+		return false;
+	}
+	std::vector<uint8_t> raw(frame.raw_size);
+	if (ZSTD_decompress(raw.data(), raw.size(), compressed.data(), compressed.size()) != raw.size()) {
+		return false;
+	}
+	Reader     reader {raw};
+	auto       type     = Record::Module;
+	const auto complete = [&] { return reader.ok && reader.bytes.empty(); };
+	IoAll(reader, type);
+	switch (type) {
+		case Record::Module: {
+			ProgramKey            key;
+			std::vector<uint8_t>  plan;
+			ShaderModule          module;
+			std::vector<uint32_t> spirv;
+			IoKey(reader, key);
+			IoAll(reader, plan);
+			IoModule(reader, module);
+			IoAll(reader, spirv);
+			if (!complete()) {
+				return false;
+			}
+			const auto [program, inserted] = m_program_cache.try_emplace(std::move(key));
+			if (inserted &&
+			    !ShaderRecompiler::IR::LoadResourcePlan(plan, program->second.resource_plan)) {
+				m_program_cache.erase(program);
+				return false;
+			}
+			module.module = CompileSPV(spirv, m_graphics.device);
+			if (module.module == nullptr) {
+				return false;
+			}
+			program->second.modules.push_back(std::move(module));
+			break;
+		}
+		case Record::GraphicsPipeline: {
+			GraphicsPipelineKey key;
+			reader.Raw(&key, sizeof(key));
+			if (!complete()) {
+				return false;
+			}
+			pipelines.push_back(key);
+			break;
+		}
+		case Record::TessellationLayout: {
+			uint64_t                    key = 0;
+			ShaderTessellationInputInfo layout;
+			IoAll(reader, key, layout);
+			if (!complete()) {
+				return false;
+			}
+			m_tess_layouts.try_emplace(key, layout);
+			break;
+		}
+		default: return false;
+	}
+	records = records.subspan(sizeof(frame) + compressed.size());
+	return true;
+}
+
+// Records are encoded here, since programs and modules change later.
+void PipelineCache::StoreModule(const ProgramKey& key, const Program& program,
+                                const ShaderModule& module, const std::vector<uint32_t>& spirv) {
+	if (!m_store) {
+		return;
+	}
+	Writer record;
+	IoAll(record, Record::Module);
+	IoKey(record, key);
+	IoAll(record, ShaderRecompiler::IR::SaveResourcePlan(program.resource_plan));
+	IoModule(record, module);
+	IoAll(record, spirv);
+	m_store.Append(std::move(record.bytes));
+}
+
+void PipelineCache::StorePipeline(const GraphicsPipelineKey& key) {
+	if (!m_store) {
+		return;
+	}
+	Writer record;
+	IoAll(record, Record::GraphicsPipeline);
+	record.Raw(&key, sizeof(key));
+	m_store.Append(std::move(record.bytes));
+}
+
+void PipelineCache::StoreLayout(uint64_t key, const ShaderTessellationInputInfo& layout) {
+	if (!m_store) {
+		return;
+	}
+	Writer record;
+	IoAll(record, Record::TessellationLayout, key, layout);
+	m_store.Append(std::move(record.bytes));
+}
+
+PipelineCache::ModuleMap PipelineCache::MapModules() const {
+	ModuleMap modules {{0, nullptr}};
+	for (const auto& program: m_program_cache | std::views::values) {
+		for (const auto& module: program.modules) {
+			modules.emplace(module.id, &module);
+		}
+	}
+	return modules;
+}
+
+bool PipelineCache::CreateGraphicsPipeline(Pipeline& pipeline, const GraphicsPipelineKey& key,
+                                           const ModuleMap& modules) {
+	std::array<const ShaderModule*, 4> stages {};
+	for (size_t i = 0; i < stages.size(); i++) {
+		const auto found = modules.find(key.stage_ids[i]);
+		if (found == modules.end()) {
+			return false;
+		}
+		stages[i] = found->second;
+	}
+	CreatePipelineInternal(m_graphics, pipeline, key, stages, m_driver_cache);
+	return true;
+}
+
+} // namespace Libs::Graphics

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

@@ -1,6 +1,7 @@
 #include "graphics/host_gpu/renderer/pipeline/shaderResourceBarrier.h"
 
 #include "common/assert.h"
+#include "graphics/host_gpu/renderer/pipeline/pipelineCache.h"
 #include "graphics/shader/shader.h"
 #include "graphics/shader/shaderBindings.h"
 
@@ -90,7 +91,7 @@ vk::BufferMemoryBarrier MakeGdsDependency(vk::Buffer buffer) {
 
 bool HasShaderBufferWrites(const ShaderStageRuntime& runtime) {
 	EXIT_IF(!runtime);
-	const auto& program   = *runtime.program;
+	const auto& program   = runtime.module->info;
 	const auto& resources = *runtime.resources;
 	EXIT_IF(resources.buffers.size() != program.info.buffers.size());
 	// Address writes reach memory outside the bound buffer descriptors.

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

@@ -13,7 +13,6 @@
 #include "graphics/host_gpu/renderer/renderContext.h"
 #include "graphics/host_gpu/renderer/renderTarget.h"
 #include "graphics/host_gpu/vulkanCommon.h"
-#include "graphics/shader/recompiler/BufferFormat.h"
 #include "graphics/shader/recompiler/ir/ShaderIR.h"
 #include "graphics/shader/rectListShader.h"
 #include "graphics/shader/shader.h"
@@ -21,124 +20,11 @@
 #include <algorithm>
 #include <array>
 #include <bit>
-#include <limits>
 #include <span>
 #include <vector>
 
 namespace Libs::Graphics {
 
-// IDK: maybe we can remove it?
-constexpr uint8_t kTemporaryVertexAttribFormat113 =
-    static_cast<uint8_t>(Prospero::VertexAttribFormat::k16_16SInt);
-constexpr uint32_t kTemporaryPs5BufferFormat121 = 121u;
-
-static bool NarrowInputFormat(vk::Format& format, uint32_t& size, uint32_t used_components) {
-	if (used_components == 0 || used_components >= size) {
-		return false;
-	}
-
-	switch (format) {
-		case vk::Format::eR32G32B32A32Sfloat:
-			switch (used_components) {
-				case 1: format = vk::Format::eR32Sfloat; break;
-				case 2: format = vk::Format::eR32G32Sfloat; break;
-				case 3: format = vk::Format::eR32G32B32Sfloat; break;
-				default: return false;
-			}
-			size = used_components;
-			return true;
-		case vk::Format::eR32G32B32Sfloat:
-			switch (used_components) {
-				case 1: format = vk::Format::eR32Sfloat; break;
-				case 2: format = vk::Format::eR32G32Sfloat; break;
-				default: return false;
-			}
-			size = used_components;
-			return true;
-		case vk::Format::eR16G16B16A16Sfloat:
-			switch (used_components) {
-				case 1: format = vk::Format::eR16Sfloat; break;
-				case 2: format = vk::Format::eR16G16Sfloat; break;
-				default: return false;
-			}
-			size = used_components;
-			return true;
-		case vk::Format::eR8G8B8A8Unorm:
-			switch (used_components) {
-				case 1: format = vk::Format::eR8Unorm; break;
-				case 2: format = vk::Format::eR8G8Unorm; break;
-				default: return false;
-			}
-			size = used_components;
-			return true;
-		case vk::Format::eR8G8B8A8Snorm:
-			if (used_components != 2) {
-				return false;
-			}
-			format = vk::Format::eR8G8Snorm;
-			size   = 2;
-			return true;
-		case vk::Format::eR8G8B8A8Uint:
-			switch (used_components) {
-				case 1: format = vk::Format::eR8Uint; break;
-				case 2: format = vk::Format::eR8G8Uint; break;
-				default: return false;
-			}
-			size = used_components;
-			return true;
-		default: break;
-	}
-
-	return false;
-}
-
-static void GetInputFormat(const ShaderBufferResource& res, vk::Format& format, uint32_t& size,
-                           uint32_t used_components) {
-	const auto fmt        = res.Format();
-	const auto raw_format = res.RawFormat();
-	if (raw_format == kTemporaryVertexAttribFormat113) {
-		static bool logged_113 = false;
-		if (!logged_113) {
-			LOGF("InputFormat: temporary: accepting invalid PS5 buffer format 113 as "
-			     "vk::Format::eR32G32B32A32Sfloat\n");
-			logged_113 = true;
-		}
-		format = vk::Format::eR32G32B32A32Sfloat;
-		size   = 4;
-		if (NarrowInputFormat(format, size, used_components)) {
-			LOGF("InputFormat: narrowing fmt=%u to %s for used_components=%u\n", raw_format,
-			     vk::to_string(format).c_str(), used_components);
-		}
-		return;
-	}
-	if (raw_format == kTemporaryPs5BufferFormat121) {
-		static bool logged_121 = false;
-		if (!logged_121) {
-			LOGF("InputFormat: accepting PS5 buffer format 121 as vk::Format::eR16G16Sfloat\n");
-			logged_121 = true;
-		}
-		format = vk::Format::eR16G16Sfloat;
-		size   = 2;
-		return;
-	}
-
-	format = VulkanFormat(fmt);
-	size   = ShaderRecompiler::Format::GetFormatInfo(fmt).component_count;
-	if (format == vk::Format::eUndefined || size == 0) {
-		EXIT("unknown vertex format: fmt = %u\n", raw_format);
-	}
-
-	if (NarrowInputFormat(format, size, used_components)) {
-		static std::atomic<uint64_t> log_count = 0;
-		auto                         log_id    = log_count.fetch_add(1);
-		if (log_id < 32) {
-			LOGF("VertexInput: narrowed vertex format to %" PRIu32
-			     " component(s) for shader fetch\n",
-			     used_components);
-		}
-	}
-}
-
 static vk::BlendFactor GetBlendFactor(uint32_t factor) {
 	switch (static_cast<Prospero::BlendFactor>(factor)) {
 		case Prospero::BlendFactor::kZero: return vk::BlendFactor::eZero;
@@ -242,33 +128,27 @@ 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,
-                            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 ||
-	                        rendering.stencil_format != vk::Format::eUndefined;
-	EXIT_IF(!vs_input_info.stage);
-	const bool mesh = vs_input_info.stage.program->stage == ShaderType::Mesh;
+                            const GraphicsPipelineKey&                 key,
+                            const std::array<const ShaderModule*, 4>& stages,
+                            vk::PipelineCache                          driver_cache) {
+	const auto& rendering     = key.rendering;
+	const auto& vertex_input  = key.vertex_input;
+	const auto& static_params = key.static_params;
+	const auto& vertex        = *stages[0];
+	const auto* pixel         = stages[3];
+	const bool  with_depth    = rendering.depth_format != vk::Format::eUndefined ||
+	                            rendering.stencil_format != vk::Format::eUndefined;
+	const bool  mesh          = vertex.info.stage == ShaderType::Mesh;
 	EXIT_NOT_IMPLEMENTED(mesh && !graphics.mesh_shader_enabled);
-	const bool rect_list =
-	    !mesh && !tessellation && static_params.topology == vk::PrimitiveTopology::ePatchList;
+	const bool rect_list = static_params.patch_control_points != 0 && stages[1] == nullptr;
 
 	vk::ShaderModule tess_control_shader_module = nullptr;
 	vk::ShaderModule tess_eval_shader_module    = nullptr;
 
 	if (rect_list) {
-		const auto shaders =
-		    BuildRectListShaders(vs_input_info, ps_active ? ps_input_info : nullptr);
+		std::span<const RectListParameter> parameters;
+		if (pixel != nullptr) parameters = pixel->rect_parameters;
+		const auto shaders = BuildRectListShaders(parameters, vertex.info.param_export_mask);
 		tess_control_shader_module = CompileSPV(shaders.control, graphics.device);
 		if (graphics_debug_dump_enabled()) {
 			LOGF("PipelineTrace: vkCreateShaderModule RectList TCS done module=%p\n",
@@ -287,11 +167,12 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 
 	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"};
+	for (const auto* stage: stages) {
+		if (stage != nullptr) {
+			shader_stages[shader_stage_count++] = {.stage  = NativeShaderStage(stage->info.stage),
+			                                       .module = stage->module,
+			                                       .pName  = "main"};
+		}
 	}
 	if (rect_list) {
 		shader_stages[shader_stage_count++] = {.stage =
@@ -303,11 +184,6 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 		                                       .module = tess_eval_shader_module,
 		                                       .pName  = "main"};
 	}
-	if (ps_active) {
-		shader_stages[shader_stage_count++] = {.stage  = vk::ShaderStageFlagBits::eFragment,
-		                                       .module = pixel_program.module,
-		                                       .pName  = "main"};
-	}
 
 	vk::VertexInputAttributeDescription input_attr[ShaderVertexInputInfo::RES_MAX] {};
 	vk::VertexInputBindingDescription   input_desc[ShaderVertexInputInfo::RES_MAX] {};
@@ -323,46 +199,7 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 		input_attr[index].binding  = vertex_input.attributes[index].binding;
 		input_attr[index].location = index;
 		input_attr[index].offset   = vertex_input.attributes[index].offset;
-
-		uint32_t   attr_size     = 4;
-		const auto registers_num = vs_input_info.resources_dst[index].registers_num;
-		const auto compiled_components =
-		    vs_input_info.stage.program->info.vertex_fetch_components[index];
-		const auto used_components =
-		    compiled_components > 0 ? static_cast<int>(compiled_components) : registers_num;
-		GetInputFormat(vs_input_info.resources[index], input_attr[index].format, attr_size,
-		               static_cast<uint32_t>(used_components));
-
-		if (graphics_debug_dump_enabled()) {
-			static std::atomic_uint log_count = 0;
-			const auto              log_id    = log_count.fetch_add(1, std::memory_order_relaxed);
-			if (log_id < 128) {
-				LOGF("VertexInputState[%u]: attr=%u binding=%u offset=%u stride=%u fmt=%d "
-				     "src_fmt=%u dst=v%u regs=%u"
-				     " fetched_components=%u attr_size=%u swizzle=%u,%u,%u,%u\n",
-				     log_id, index, input_attr[index].binding, input_attr[index].offset,
-				     input_desc[input_attr[index].binding].stride,
-				     static_cast<int>(input_attr[index].format),
-				     static_cast<uint32_t>(vs_input_info.resources[index].Format()),
-				     static_cast<uint32_t>(vs_input_info.resources_dst[index].register_start),
-				     static_cast<uint32_t>(registers_num), static_cast<uint32_t>(used_components),
-				     attr_size, static_cast<uint32_t>(vs_input_info.resources[index].DstSelX()),
-				     static_cast<uint32_t>(vs_input_info.resources[index].DstSelY()),
-				     static_cast<uint32_t>(vs_input_info.resources[index].DstSelZ()),
-				     static_cast<uint32_t>(vs_input_info.resources[index].DstSelW()));
-			}
-		}
-
-		if (vs_input_info.resources[index].OutOfBounds() != 0) {
-			static bool logged = false;
-			if (!logged) {
-				LOGF("VertexInput: temporary: accepting PS5 out-of-bounds behavior %" PRIu8 "\n",
-				     vs_input_info.resources[index].OutOfBounds());
-				logged = true;
-			}
-		}
-
-		EXIT_IF(registers_num < 1 || registers_num > 4);
+		input_attr[index].format   = vertex.vertex_formats[index];
 	}
 
 	vk::PipelineVertexInputStateCreateInfo vertex_input_info {};
@@ -382,17 +219,6 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	vk::PipelineViewportStateCreateInfo viewport_state {};
 	viewport_state.pNext = &depth_clip_control;
 
-	vk::CullModeFlags cull_mode = vk::CullModeFlagBits::eNone;
-	if (static_params.cull_back) {
-		cull_mode |= vk::CullModeFlagBits::eBack;
-	}
-	if (static_params.cull_front) {
-		cull_mode |= vk::CullModeFlagBits::eFront;
-	}
-
-	vk::FrontFace front_face =
-	    (static_params.face ? vk::FrontFace::eClockwise : vk::FrontFace::eCounterClockwise);
-
 	vk::PipelineRasterizationDepthClipStateCreateInfoEXT clip_ext {};
 	clip_ext.depthClipEnable = static_params.depth_clip_enable ? VK_TRUE : VK_FALSE;
 
@@ -413,8 +239,6 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 		provoking_vertex.pNext = rasterizer.pNext;
 		rasterizer.pNext = &provoking_vertex;
 	}
-	rasterizer.cullMode  = cull_mode;
-	rasterizer.frontFace = front_face;
 	rasterizer.polygonMode = static_params.polygon_mode;
 	rasterizer.lineWidth = 1.0f;
 
@@ -470,17 +294,13 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 
 	std::vector<vk::DescriptorSetLayoutBinding> descriptor_bindings;
 	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);
-		pipeline.uses_image_heap |= stage.stage.program->info.uses_image_heap;
-		graphics_stages |= native_stage;
-	}
-	if (ps_active) {
-		EXIT_IF(!ps_input_info->stage);
-		AddLayoutBindings(descriptor_bindings, *ps_input_info->stage.program,
-		                  vk::ShaderStageFlagBits::eFragment);
-		pipeline.uses_image_heap |= ps_input_info->stage.program->info.uses_image_heap;
+	for (const auto* stage: stages) {
+		if (stage != nullptr) {
+			const auto native_stage = NativeShaderStage(stage->info.stage);
+			AddLayoutBindings(descriptor_bindings, stage->info, native_stage);
+			pipeline.uses_image_heap |= stage->info.info.uses_image_heap;
+			graphics_stages |= native_stage;
+		}
 	}
 	CreateDescriptorLayout(graphics, pipeline, descriptor_bindings);
 	const vk::PushConstantRange push_constants {graphics_stages, 0,
@@ -498,7 +318,7 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	if (graphics_debug_dump_enabled()) {
 		LOGF("PipelineTrace: vkCreatePipelineLayout begin VS=%" PRIu64 " PS=%" PRIu64
 		     " set_layouts=1 push_constants=%" PRIu32 "\n",
-		     vertex_program.id, ps_active ? pixel_program.id : 0, 1u);
+		     key.stage_ids[0], key.stage_ids[3], 1u);
 	}
 	auto result = graphics.device.createPipelineLayout(&pipeline_layout_info, nullptr,
 	                                                   &pipeline.pipeline_layout);
@@ -508,14 +328,14 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	}
 	EXIT_NOT_IMPLEMENTED(result != vk::Result::eSuccess);
 
-	EXIT_NOT_IMPLEMENTED(pipeline.pipeline_layout == nullptr);
-
 	vk::PipelineDepthStencilStateCreateInfo depth_stencil_info {};
 
 	std::vector<vk::DynamicState> dynamic_states {
 	    vk::DynamicState::eViewportWithCount,
 	    vk::DynamicState::eScissorWithCount,
 	    vk::DynamicState::eLineWidth,
+	    vk::DynamicState::eCullMode,
+	    vk::DynamicState::eFrontFace,
 	    vk::DynamicState::eDepthTestEnable,
 	    vk::DynamicState::eDepthWriteEnable,
 	    vk::DynamicState::eDepthCompareOp,
@@ -555,9 +375,9 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	pipeline_info.pVertexInputState        = mesh ? nullptr : &vertex_input_info;
 	pipeline_info.pInputAssemblyState      = mesh ? nullptr : &input_assembly;
 	vk::PipelineTessellationStateCreateInfo tessellation_state {};
-	tessellation_state.patchControlPoints =
-	    tessellation ? vs_input_info.tess.input_control_points : 3u;
-	pipeline_info.pTessellationState = (rect_list || tessellation) ? &tessellation_state : nullptr;
+	tessellation_state.patchControlPoints = static_params.patch_control_points;
+	pipeline_info.pTessellationState =
+	    static_params.patch_control_points != 0 ? &tessellation_state : nullptr;
 	pipeline_info.pViewportState          = &viewport_state;
 	pipeline_info.pRasterizationState     = &rasterizer;
 	pipeline_info.pMultisampleState       = &multisampling;
@@ -573,7 +393,7 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 		LOGF("PipelineTrace: vkCreateGraphicsPipelines begin VS=%" PRIu64 " PS=%" PRIu64
 		     " topology=%" PRIu32 " color_mask=0x%08" PRIx32
 		     " depth=%s blend=%s dyn_states=%" PRIu32 "\n",
-		     vertex_program.id, ps_active ? pixel_program.id : 0,
+		     key.stage_ids[0], key.stage_ids[3],
 		     static_cast<uint32_t>(static_params.topology), static_params.color_mask[0],
 		     (with_depth ? "true" : "false"), (static_params.blend_enable[0] ? "true" : "false"),
 		     dynamic_state.dynamicStateCount);
@@ -586,29 +406,18 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	}
 	EXIT_NOT_IMPLEMENTED(result != vk::Result::eSuccess);
 
-	EXIT_NOT_IMPLEMENTED(pipeline.pipeline == nullptr);
-
-	if (tess_control_shader_module != nullptr) {
-		graphics.device.destroyShaderModule(tess_control_shader_module, nullptr);
-	}
-	if (tess_eval_shader_module != nullptr) {
-		graphics.device.destroyShaderModule(tess_eval_shader_module, nullptr);
-	}
+	graphics.device.destroyShaderModule(tess_control_shader_module, nullptr);
+	graphics.device.destroyShaderModule(tess_eval_shader_module, nullptr);
 }
 
-// NOLINTNEXTLINE(readability-function-cognitive-complexity)
 void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& pipeline,
-                            const ShaderComputeInputInfo& input_info,
-                            vk::ShaderModule compute_module, vk::PipelineCache driver_cache) {
-	EXIT_IF(compute_module == nullptr);
-
+                            const ShaderModule& module, vk::PipelineCache driver_cache) {
 	vk::PipelineShaderStageCreateInfo                     comp_shader_stage_info {};
 	vk::PipelineShaderStageRequiredSubgroupSizeCreateInfo comp_subgroup_size {};
 	comp_shader_stage_info.stage  = vk::ShaderStageFlagBits::eCompute;
-	comp_shader_stage_info.module = compute_module;
+	comp_shader_stage_info.module = module.module;
 	comp_shader_stage_info.pName  = "main";
-	EXIT_IF(!input_info.stage);
-	const auto wave_size = input_info.stage.program->wave_size;
+	const auto wave_size = module.info.wave_size;
 	if (graphics.compute_subgroup_size_control_enabled &&
 	    wave_size >= graphics.min_subgroup_size && wave_size <= graphics.max_subgroup_size) {
 		comp_subgroup_size.requiredSubgroupSize = wave_size;
@@ -616,9 +425,8 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	}
 
 	std::vector<vk::DescriptorSetLayoutBinding> descriptor_bindings;
-	AddLayoutBindings(descriptor_bindings, *input_info.stage.program,
-	                  vk::ShaderStageFlagBits::eCompute);
-	pipeline.uses_image_heap = input_info.stage.program->info.uses_image_heap;
+	AddLayoutBindings(descriptor_bindings, module.info, vk::ShaderStageFlagBits::eCompute);
+	pipeline.uses_image_heap = module.info.info.uses_image_heap;
 	CreateDescriptorLayout(graphics, pipeline, descriptor_bindings);
 	const vk::PushConstantRange push_constants {vk::ShaderStageFlagBits::eCompute, 0,
 	                                            ShaderRecompiler::IR::NativePushConstantSize};
@@ -632,18 +440,12 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 
 	EXIT_IF(pipeline.pipeline_layout != nullptr);
 
-	LOGF("PipelineTrace: vkCreatePipelineLayout CS begin set_layouts=1 push_constants=%u\n",
-	     1u);
 	auto result = graphics.device.createPipelineLayout(&pipeline_layout_info, nullptr,
 	                                                  &pipeline.pipeline_layout);
-	LOGF("PipelineTrace: vkCreatePipelineLayout CS done result=%s layout=%p\n",
-	     vk::to_string(result).c_str(), static_cast<void*>(pipeline.pipeline_layout));
 	EXIT_NOT_IMPLEMENTED(result != vk::Result::eSuccess);
 
-	EXIT_NOT_IMPLEMENTED(pipeline.pipeline_layout == nullptr);
-
 	vk::ComputePipelineCreateInfo info {};
-	if (input_info.stage.program->info.uses_gds) {
+	if (module.info.info.uses_gds) {
 		info.flags = vk::PipelineCreateFlagBits::eDispatchBase;
 	}
 	info.stage             = comp_shader_stage_info;
@@ -652,15 +454,9 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 
 	EXIT_IF(pipeline.pipeline != nullptr);
 
-	LOGF("PipelineTrace: vkCreateComputePipelines begin layout=%p\n",
-	     static_cast<void*>(pipeline.pipeline_layout));
 	result = graphics.device.createComputePipelines(driver_cache, 1, &info, nullptr,
 	                                                &pipeline.pipeline);
-	LOGF("PipelineTrace: vkCreateComputePipelines done result=%s pipeline=%p\n",
-	     vk::to_string(result).c_str(), static_cast<void*>(pipeline.pipeline));
 	EXIT_NOT_IMPLEMENTED(result != vk::Result::eSuccess);
-
-	EXIT_NOT_IMPLEMENTED(pipeline.pipeline == nullptr);
 }
 
 } // namespace Libs::Graphics

src/graphics/host_gpu/renderer/render.h

@@ -23,14 +23,9 @@ class Shader;
 } // namespace HW
 
 struct GraphicContext;
-struct ShaderBufferResource;
-struct ShaderComputeInputInfo;
-struct RenderDepthInfo;
-struct RenderColorInfo;
 struct DrawCallInfo;
 struct DrawEmitInfo;
 struct DrawIndexBufferSource;
-struct DrawRenderState;
 class RenderContext;
 class CommandScheduler;
 struct RenderExecutorTestAccess;
@@ -50,9 +45,10 @@ enum class DrawOffsetSource : uint8_t {
 
 // Nonzero address: the GPU reads count Vulkan indirect draw records, stride bytes apart.
 struct DrawIndirectArgs {
-	uint64_t address = 0;
-	uint32_t count   = 0;
-	uint32_t stride  = 0;
+	uint64_t              address = 0;
+	uint32_t              count   = 0;
+	uint32_t              stride  = 0;
+	ShaderIndirectOffsets offsets;
 };
 
 struct DrawIndexArgs {
@@ -198,9 +194,8 @@ private:
 	                                          uint32_t            render_target_slice_offset,
 	                                          DrawRenderState& state);
 	void ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buffer, const DrawCallInfo& draw,
-	                         DrawRenderState& state, vk::PrimitiveTopology topology,
-	                         const DrawEmitInfo& emit, const DrawIndexBufferSource& index_source,
-	                         bool primitive_restart_enable);
+	                         DrawRenderState& state, const DrawEmitInfo& emit,
+	                         const DrawIndexBufferSource& index_source);
 	[[nodiscard]] RenderState AcquireRenderTargets(CommandBuffer& buffer, RenderColorInfo* colors,
 	                                               uint32_t color_count, RenderDepthInfo& depth,
 	                                               vk::ImageAspectFlags& feedback_aspects,

src/graphics/host_gpu/renderer/renderCompute.cpp

@@ -53,7 +53,7 @@ static bool FillSourcesDisjoint(std::span<const ShaderRecompiler::IR::Descriptor
 
 bool RenderExecutor::TryConsumeComputeMetaClear(const ShaderComputeInputInfo& input,
                                                 const CommandBuffer&          buffer) {
-	const auto& program   = *input.stage.program;
+	const auto& program   = input.stage.module->info;
 	const auto& resources = *input.stage.resources;
 	if (resources.buffers.size() != program.info.buffers.size()) {
 		EXIT("compute runtime buffer count does not match shader metadata\n");
@@ -135,7 +135,7 @@ bool ResolveComputeBufferFill(const ShaderComputeInputInfo& input, uint32_t grou
 bool RenderExecutor::TryConsumeComputeFill(const ShaderComputeInputInfo& input,
                                           CommandBuffer& command, uint32_t group_x,
                                           uint32_t group_y, uint32_t group_z, uint32_t mode) {
-	const auto& program   = *input.stage.program;
+	const auto& program   = input.stage.module->info;
 	const auto& resources = *input.stage.resources;
 	const auto& fill      = resources.uniform_fill;
 	auto&       cache     = command.GetContext().GetTextureCache();
@@ -208,7 +208,7 @@ bool RenderExecutor::TryConsumeComputeFill(const ShaderComputeInputInfo& input,
 	if (logged_clears.fetch_add(1, std::memory_order_relaxed) < 32) {
 		LOGF("GraphicsRenderDispatchDirect: compute %s shader=0x%016" PRIx64
 		     " addr=0x%016" PRIx64 " size=0x%016" PRIx64 " value=0x%08" PRIx32 "\n",
-		     image_clear ? "image clear" : "buffer fill", input.stage.program->shader_hash,
+		     image_clear ? "image clear" : "buffer fill", input.stage.module->info.shader_hash,
 		     descriptor.Base48(), size, packed_clear);
 	}
 	return true;
@@ -216,7 +216,7 @@ bool RenderExecutor::TryConsumeComputeFill(const ShaderComputeInputInfo& input,
 
 static void BindSharedMemory(RenderContext& context, ShaderComputeInputInfo& input,
                              PreparedBindings& bindings, uint64_t indirect_args = 0) {
-	if (input.stage.program->bindings.descriptor_counts[
+	if (input.stage.module->info.bindings.descriptor_counts[
 	        static_cast<size_t>(ShaderRecompiler::IR::DescriptorBindingKind::SharedMemory)] == 0) {
 		return;
 	}
@@ -248,7 +248,6 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
                                     uint32_t thread_group_z, uint32_t mode, bool async_compute) {
 	EXIT_IF(buffer.IsInvalid());
 	m_context.GetCommandScheduler().PopPendingOperations();
-	auto& ctx    = buffer.GetRegisters();
 	auto& sh_ctx = buffer.GetShaders();
 
 	buffer.SetDebugInfo(static_cast<uint32_t>(CommandBufferDebugOp::DispatchDirect), submit_id,
@@ -293,7 +292,6 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 	}
 
 	const auto& cs_regs = sh_ctx.GetCs();
-	const auto& sh_regs = ctx.GetShaderRegisters();
 
 	ShaderComputeInputInfo input_info {};
 	input_info.async_compute = async_compute;
@@ -312,16 +310,14 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 		}
 	}
 	m_context.GetImageHeap().BeginDraw();
-	const auto compute_program =
-	    m_context.GetPipelineCache().GetComputeProgram(cs_regs, sh_regs, input_info);
-	if (!compute_program) return;
+	if (!m_context.GetPipelineCache().GetComputeProgram(cs_regs, input_info)) return;
 	if (use_thread_dimensions) {
 		input_info.dispatch_threads_num[0]    = thread_group_x;
 		input_info.dispatch_threads_num[1]    = thread_group_y;
 		input_info.dispatch_threads_num[2]    = thread_group_z;
 	}
 
-	const auto& program   = *input_info.stage.program;
+	const auto& program   = input_info.stage.module->info;
 	const auto& resources = *input_info.stage.resources;
 	if (resources.specialization_reads.empty() &&
 	    (TryConsumeComputeMetaClear(input_info, buffer) ||
@@ -351,8 +347,7 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 	}
 
 	buffer.EndRendering();
-	auto& pipeline =
-	    m_context.GetPipelineCache().GetComputePipeline(input_info, compute_program);
+	auto& pipeline = m_context.GetPipelineCache().GetComputePipeline(input_info);
 	auto& bindings = m_compute_bindings;
 	PrepareBindings(input_info.stage, bindings);
 	if (program.bindings.dispatch_thread_dword != ShaderRecompiler::IR::PushData::NoStart) {
@@ -502,16 +497,14 @@ void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
 	input_info.async_compute = async_compute;
 	input_info.indirect_args_address = args_addr;
 	m_context.GetImageHeap().BeginDraw();
-	const auto compute_program = m_context.GetPipelineCache().GetComputeProgram(
-	    cs_regs, buffer.GetRegisters().GetShaderRegisters(), input_info);
-	if (!compute_program) return;
+	if (!m_context.GetPipelineCache().GetComputeProgram(cs_regs, input_info)) return;
 	buffer.EndRendering();
-	auto& pipeline = m_context.GetPipelineCache().GetComputePipeline(input_info, compute_program);
+	auto& pipeline = m_context.GetPipelineCache().GetComputePipeline(input_info);
 	auto& bindings = m_compute_bindings;
 	PrepareBindings(input_info.stage, bindings);
 	PreparedBindings* descriptor_stage = &bindings;
 	FindBuffers(std::span {&descriptor_stage, 1u});
-	const auto& program = *input_info.stage.program;
+	const auto& program = input_info.stage.module->info;
 	if (program.info.uses_dma) {
 		m_context.PrepareBda();
 	}

src/graphics/host_gpu/renderer/renderContext.cpp

@@ -10,10 +10,10 @@
 
 namespace Libs::Graphics {
 
-RenderContext::RenderContext(GraphicContext& graphics)
+RenderContext::RenderContext(GraphicContext& graphics, const std::filesystem::path& shader_store)
     : m_graphics(graphics), m_render_executor(*this), m_command_scheduler(*this, graphics),
       m_descriptor_heap(graphics, m_command_scheduler.GetMasterSemaphore()),
-      m_pipeline_cache(graphics, &m_image_heap), m_sampler_cache(graphics),
+      m_pipeline_cache(graphics, &m_image_heap, shader_store), m_sampler_cache(graphics),
       m_buffer_cache(graphics, m_command_scheduler, m_page_manager, m_texture_cache),
       m_texture_cache(graphics, m_command_scheduler, m_page_manager, m_buffer_cache),
       m_image_heap(*this, m_page_manager), m_occlusion_queries(graphics, m_command_scheduler) {
@@ -134,7 +134,11 @@ void RenderContext::PrepareBda() {
 		m_bda_logged = true;
 	}
 	std::shared_lock lock(m_mapped_ranges_mutex);
-	m_buffer_cache.SynchronizeBda(m_mapped_ranges);
+	m_buffer_cache.TakeBdaSync().ForEach([this](uint64_t begin, uint64_t end) {
+		m_mapped_ranges.ForEachInRange(begin, end - begin, [this](uint64_t start, uint64_t finish) {
+			m_buffer_cache.SynchronizeBuffersInRange(start, finish - start);
+		});
+	});
 	m_fault_process_pending = true;
 }
 
@@ -143,6 +147,7 @@ void RenderContext::RunGarbageCollector() {
 		m_fault_process_pending = false;
 		m_buffer_cache.ProcessFaultBuffer();
 	}
+	m_image_heap.RunGarbageCollector();
 	m_texture_cache.ProcessDownloadImages();
 	m_texture_cache.RunGarbageCollector();
 	m_buffer_cache.RunGarbageCollector();

src/graphics/host_gpu/renderer/renderContext.h

@@ -31,7 +31,8 @@ class GuestGpu;
 
 class RenderContext {
 public:
-	explicit RenderContext(GraphicContext& graphics);
+	explicit RenderContext(GraphicContext&              graphics,
+	                       const std::filesystem::path& shader_store = ShaderStorePath());
 	~RenderContext();
 	KYTY_CLASS_NO_COPY(RenderContext);
 

src/graphics/host_gpu/renderer/renderDraw.cpp

@@ -17,7 +17,6 @@
 #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"
@@ -58,7 +57,7 @@ std::pair<int32_t, uint32_t> ResolveDrawOffsets(uint32_t index_offset,
 	}
 
 	EXIT_IF(!vs_input_info.stage);
-	const auto& program   = *vs_input_info.stage.program;
+	const auto& program   = vs_input_info.stage.module->info;
 	const auto& resources = *vs_input_info.stage.resources;
 	if (index_offset == 0 &&
 	    program.info.vertex_offset_sgpr >= static_cast<int32_t>(program.user_data_base)) {
@@ -198,7 +197,7 @@ static void LogDrawTargetState(const char* draw_name, const RenderColorInfo& col
 	const auto& dc             = ctx.GetDepthControl();
 	const auto& vp             = ctx.GetScreenViewport();
 	const auto& vp0            = vp.viewports[0];
-	const auto& ps_resources   = ps_input_info.stage.program->info;
+	const auto& ps_resources   = ps_input_info.stage.module->info.info;
 	const auto  sampled_images = std::count_if(
 	    ps_resources.images.begin(), ps_resources.images.end(), [](const auto& image) {
 		    return image.resource_class == ShaderRecompiler::IR::ImageResourceClass::Sampled;
@@ -321,7 +320,7 @@ static void UpdateGraphicsDynamicState(const CommandBuffer& buffer, DynamicState
 	const auto& ctx = buffer.GetRegisters();
 	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 auto& outputs = vs_input_info.stage.module->info.info.outputs;
 	const bool  indexed_viewports =
 	    std::any_of(outputs.begin(), outputs.end(), [](const auto& output) {
 		    return output.kind == ShaderRecompiler::IR::StageOutputKind::ViewportIndex;
@@ -404,6 +403,20 @@ static void UpdateGraphicsDynamicState(const CommandBuffer& buffer, DynamicState
 		dynamic_state.SetDepthBias(constant_factor, poly_offset.clamp, slope_factor);
 	}
 
+	// Rect lists are drawn without culling.
+	vk::CullModeFlags cull_mode {};
+	if (!Prospero::IsRectList(buffer.GetUserConfig().GetPrimType())) {
+		if (mode.cull_front) {
+			cull_mode |= vk::CullModeFlagBits::eFront;
+		}
+		if (mode.cull_back) {
+			cull_mode |= vk::CullModeFlagBits::eBack;
+		}
+	}
+	dynamic_state.SetCullMode(cull_mode);
+	dynamic_state.SetFrontFace(mode.face ? vk::FrontFace::eClockwise
+	                                     : vk::FrontFace::eCounterClockwise);
+
 	dynamic_state.SetStencilTestEnabled(depth.stencil_test_enable);
 	if (depth.stencil_test_enable) {
 		const auto& front = depth.stencil_front;
@@ -433,22 +446,6 @@ static bool DrawHasValidVertexShader(const HW::Shader& sh_ctx) {
 	return vs.es_regs.data_addr != 0;
 }
 
-static bool PixelShaderHasDepthOrCoverageSideEffects(const HW::ShaderRegisters& sh_regs) {
-	const auto& db = sh_regs.db_shader_control;
-	return db.shader_kill_enable || db.shader_z_export_enable || db.shader_mask_export_enable ||
-	       db.shader_dual_export_enable || db.shader_execute_on_noop || sh_regs.AlphaToMask();
-}
-
-struct DrawRenderState {
-	RenderDepthInfo       depth_info;
-	RenderColorInfo       color_info[RENDER_COLOR_ATTACHMENTS_MAX] = {};
-	uint32_t              color_count                              = 0;
-	bool                  ps_active                                = true;
-	std::array<ShaderVertexInputInfo, 3> vertex_info;
-	ShaderPixelInputInfo  ps_input_info;
-	PipelineCache::GraphicsPrograms programs;
-};
-
 struct DrawCallInfo {
 	CommandBufferDebugOp debug_op       = CommandBufferDebugOp::DrawIndex;
 	uint32_t             index_count    = 0;
@@ -474,9 +471,8 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
 	for (uint32_t i = 0; i < color_count; i++) {
 		auto& target = colors[i];
 		EXIT_IF(!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) {
+		const auto& owner = cache.m_slot_images[target.image_id];
+		if ((!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);
@@ -502,8 +498,8 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
 		attachment.image_layout = layout;
 	}
 	if (depth.image_id) {
-		const auto owner = cache.m_slot_images.try_get(depth.image_id);
-		if (owner == nullptr || !owner->registered || owner->binding.needs_rebind) {
+		const auto& owner = cache.m_slot_images[depth.image_id];
+		if (!owner.registered || owner.binding.needs_rebind) {
 			EXIT("depth target changed after render-state discovery\n");
 		}
 		const auto  image_view = cache.FindDepthTarget(depth.image_id, depth.desc);
@@ -589,23 +585,6 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
 	return state;
 }
 
-static uint32_t DrawColorOutputMask(const HW::Context& ctx) {
-	const auto& sh_regs     = ctx.GetShaderRegisters();
-	const auto  write_mask  = ctx.GetRenderTargetMask() & sh_regs.m_cbShaderMask;
-	uint32_t    output_mask = ctx.HasObjectPrimitiveIdOutput() ? 1u << 7u : 0;
-	for (uint32_t index = 0; index < RENDER_COLOR_ATTACHMENTS_MAX; index++) {
-		const auto slot = ShaderPixelExportTarget(sh_regs.m_cbShaderMask, index);
-		if (slot >= RENDER_COLOR_ATTACHMENTS_MAX) {
-			break;
-		}
-		if (sh_regs.target_output_mode[index] != 0 &&
-		    render_target_mask_slot(write_mask, slot) != 0) {
-			output_mask |= 1u << slot;
-		}
-	}
-	return output_mask;
-}
-
 enum class CbColorMode : uint8_t {
 	Disable            = 0,
 	Normal             = 1,
@@ -870,62 +849,24 @@ static bool ResolvePrimitiveRestart(const CommandBuffer& buffer,
 	return false;
 }
 
-static void RefreshShaders(CommandBuffer& buffer, const DrawCallInfo& draw,
-                           uint32_t color_output_mask, DrawRenderState& state) {
-	auto& ctx    = buffer.GetRegisters();
-	auto& sh_ctx = buffer.GetShaders();
-
-	const auto& vertex_shader_info = sh_ctx.GetVs();
-	const auto& pixel_shader_info  = sh_ctx.GetPs();
-	const auto& shader_regs        = ctx.GetShaderRegisters();
-
-	state.programs      = {};
-	state.ps_input_info = {};
-	std::array<Prospero::ColorComponentMapping, RENDER_COLOR_ATTACHMENTS_MAX>
-	    target_export_mapping {};
-	for (uint32_t slot = 0; slot < RENDER_COLOR_ATTACHMENTS_MAX; slot++) {
-		const auto& rt = ctx.GetRenderTarget(slot);
-		if ((color_output_mask & (1u << slot)) != 0 && rt.base.addr != 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();
-	buffer.GetContext().GetImageHeap().BeginDraw();
-	if (draw.IsIndexed()) {
-		LogDrawPhase(draw.Name(), "GetGraphicsPrograms");
-	}
-	state.programs = pipeline_cache.GetGraphicsPrograms(
-	    vertex_shader_info, pixel_shader_info, shader_regs, ctx, buffer.GetUserConfig(),
-	    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) {
-	const auto& shader_regs       = buffer.GetRegisters().GetShaderRegisters();
-	const auto  color_output_mask = DrawColorOutputMask(buffer.GetRegisters());
-	state.ps_active = buffer.GetShaders().GetPs().ps_regs.data_addr != 0 &&
-	                  (color_output_mask != 0 ||
-	                   PixelShaderHasDepthOrCoverageSideEffects(shader_regs));
-	RefreshShaders(buffer, draw, color_output_mask, 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 ||
-			    output.kind == ShaderRecompiler::IR::StageOutputKind::ObjectPrimitiveId) {
-				mrt_mask |= 1u << output.location;
-			}
-		}
+	m_context.GetImageHeap().BeginDraw();
+	if (draw.IsIndexed()) {
+		LogDrawPhase(draw.Name(), "GetGraphicsPrograms");
+	}
+	m_context.GetPipelineCache().GetGraphicsPrograms(buffer, state);
+	// Vertex programs beyond the host mesh limits or with unsupported call targets are not
+	// compiled.
+	if (state.vertex_info[0].stage.module == nullptr) {
+		return false;
 	}
-	mrt_mask &= color_output_mask;
 	if (draw.IsIndexed()) {
 		LogDrawPhase(draw.Name(), "ResolveRenderColorTarget");
 	}
 	for (uint32_t slot = 0; slot < RENDER_COLOR_ATTACHMENTS_MAX; slot++) {
-		if ((mrt_mask & (1u << slot)) != 0) {
+		if ((state.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) {
@@ -1124,16 +1065,15 @@ bool RenderExecutor::CopyDepthToColor(CommandBuffer& buffer, const DrawCallInfo&
 
 void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buffer,
                                          const DrawCallInfo& draw, DrawRenderState& state,
-                                         vk::PrimitiveTopology topology, const DrawEmitInfo& emit,
-                                         const DrawIndexBufferSource& index_source,
-	                                     bool primitive_restart_enable) {
-	auto& ucfg = buffer.GetUserConfig();
+                                         const DrawEmitInfo&          emit,
+                                         const DrawIndexBufferSource& index_source) {
+	auto&      ucfg                     = buffer.GetUserConfig();
+	const bool primitive_restart_enable = state.primitive_restart_enable;
 	if (CopyDepthToColor(buffer, draw, emit, state)) {
 		return;
 	}
-	const auto vertex_stages =
-	    std::span {state.vertex_info.data(), state.programs.VertexStageCount()};
-	const bool mesh_active = state.vertex_info[0].stage.program->stage == ShaderType::Mesh;
+	const auto vertex_stages = state.VertexStages();
+	const bool mesh_active = state.vertex_info[0].stage.module->info.stage == ShaderType::Mesh;
 	uint32_t   mesh_groups = 0;
 	if (mesh_active) {
 		const auto& mesh = state.vertex_info[0].mesh;
@@ -1203,10 +1143,7 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 	if (draw.IsIndexed()) {
 		LogDrawPhase(draw.Name(), "CreatePipeline");
 	}
-	auto& pipeline = m_context.GetPipelineCache().GetGraphicsPipeline(
-	    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);
+	auto& pipeline = m_context.GetPipelineCache().GetGraphicsPipeline(buffer, state);
 	vk::ImageAspectFlags feedback_aspects;
 	const auto rendering =
 	    AcquireRenderTargets(buffer, state.color_info, state.color_count, state.depth_info,
@@ -1238,7 +1175,6 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 		if (mesh_active) {
 			draw_data = {draw.index_count,
 			             static_cast<uint32_t>(emit.vertex_offset),
-			             emit.first_instance,
 			             index_source.guest_element_size,
 			             static_cast<uint32_t>(index_source.address),
 			             static_cast<uint32_t>(index_source.address >> 32u)};
@@ -1314,8 +1250,9 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 	}
 }
 
-// Vulkan offsets for a prepared draw, or nullopt when its indirect records need CPU counts, as
-// mesh and quad-list draws do. Indirect records carry their own offsets.
+// Vulkan offsets for a prepared draw, or nullopt when its indirect records need the CPU: mesh
+// and quad-list draws for counts, tessellation draws for patched SGPRs. Indirect records carry
+// their own offsets.
 static std::optional<DrawEmitInfo> ResolveDrawEmit(const HW::UserConfig&        ucfg,
                                                    const ShaderVertexInputInfo& vs,
                                                    DrawOffsetSource             source,
@@ -1323,8 +1260,9 @@ static std::optional<DrawEmitInfo> ResolveDrawEmit(const HW::UserConfig&
                                                    uint32_t                     first_instance,
                                                    const DrawIndirectArgs&      indirect) {
 	if (indirect.address != 0) {
-		if (vs.stage.program->stage == ShaderType::Mesh ||
-		    ucfg.GetPrimType() == Prospero::PrimitiveType::kQuadListLegacy) {
+		if (vs.stage.module->info.stage == ShaderType::Mesh ||
+		    ucfg.GetPrimType() == Prospero::PrimitiveType::kQuadListLegacy ||
+		    vs.indirect_offsets != indirect.offsets) {
 			return std::nullopt;
 		}
 		return DrawEmitInfo {.indirect = indirect};
@@ -1388,8 +1326,8 @@ bool RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
 
 	hw_check(buffer);
 
-	vk::PrimitiveTopology topology = vk::PrimitiveTopology::ePointList;
-	if (!GetDrawTopology(ucfg, topology)) {
+	DrawRenderState state {.indirect_offsets = args.indirect.offsets};
+	if (!GetDrawTopology(ucfg, state.topology)) {
 		return true;
 	}
 
@@ -1410,7 +1348,7 @@ bool RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
 		default: EXIT("unknown index_type_and_size: %u\n", args.index_type_and_size);
 	}
 	index_source.size = static_cast<uint64_t>(args.index_count) * index_source.guest_element_size;
-	const bool primitive_restart = ResolvePrimitiveRestart(buffer, index_source);
+	state.primitive_restart_enable = ResolvePrimitiveRestart(buffer, index_source);
 
 	std::vector<uint16_t> expanded_indices;
 	if (index_source.guest_element_size == 1) {
@@ -1418,7 +1356,8 @@ bool RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
 		const auto* src = static_cast<const uint8_t*>(args.index_addr);
 		expanded_indices.resize(args.index_count);
 		for (uint32_t i = 0; i < args.index_count; i++) {
-			expanded_indices[i] = primitive_restart && src[i] == 0xffu ? 0xffffu : src[i];
+			expanded_indices[i] =
+			    state.primitive_restart_enable && src[i] == 0xffu ? 0xffffu : src[i];
 		}
 		index_source.host_data = expanded_indices.data();
 		index_source.size      = expanded_indices.size() * sizeof(uint16_t);
@@ -1426,7 +1365,6 @@ bool RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
 
 	const DrawCallInfo draw {CommandBufferDebugOp::DrawIndex, args.index_count,
 	                        args.instance_count, args.first_instance};
-	DrawRenderState state {};
 	if (!PrepareDrawRenderState(buffer, draw, args.render_target_slice_offset, state)) {
 		ResetBindings();
 		return true;
@@ -1437,8 +1375,7 @@ bool RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
 	const auto emit = ResolveDrawEmit(ucfg, state.vertex_info[0], args.offset_source,
 	                                  args.base_vertex, args.first_instance, args.indirect);
 	if (emit) {
-		ExecutePreparedDraw(submit_id, buffer, draw, state, topology, *emit, index_source,
-		                    primitive_restart);
+		ExecutePreparedDraw(submit_id, buffer, draw, state, *emit, index_source);
 	}
 	ResetBindings();
 	return emit.has_value();
@@ -1494,12 +1431,11 @@ bool 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, topology)) {
+	DrawRenderState state {.indirect_offsets = args.indirect.offsets};
+	if (!GetDrawTopology(ucfg, state.topology)) {
 		ResetBindings();
 		return true;
 	}
-	DrawRenderState state {};
 	if (!PrepareDrawRenderState(buffer, draw, args.render_target_slice_offset, state)) {
 		ResetBindings();
 		return true;
@@ -1507,7 +1443,7 @@ bool RenderExecutor::DrawAuto(uint64_t submit_id, CommandBuffer& buffer, const D
 
 	const bool rect_list = Prospero::IsRectList(ucfg.GetPrimType());
 	if (rect_list && state.vertex_info[0].buffers_num == 0 &&
-	    state.vertex_info[0].stage.program->param_export_mask == 0 &&
+	    state.vertex_info[0].stage.module->info.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: "
@@ -1524,7 +1460,7 @@ bool RenderExecutor::DrawAuto(uint64_t submit_id, CommandBuffer& buffer, const D
 	                                  static_cast<int32_t>(args.first_vertex), args.first_instance,
 	                                  args.indirect);
 	if (emit) {
-		ExecutePreparedDraw(submit_id, buffer, draw, state, topology, *emit, {}, false);
+		ExecutePreparedDraw(submit_id, buffer, draw, state, *emit, {});
 	}
 	ResetBindings();
 	return emit.has_value();
@@ -1559,8 +1495,9 @@ bool RenderExecutor::ResolveColorTargets(CommandBuffer& buffer, uint32_t render_
 	cache.MarkGpuWritten(dst.image_id);
 	auto& source      = cache.GetImage(src.image_id);
 	auto& destination = cache.GetImage(dst.image_id);
-	destination.Resolve(source, {src.guest_mip_level, 1, src.guest_array_layer, 1},
-	                    {dst.guest_mip_level, 1, dst.guest_array_layer, 1});
+	destination.Resolve(source,
+	                    {src.desc.view_info.base_level, 1, src.desc.view_info.base_layer, 1},
+	                    {dst.desc.view_info.base_level, 1, dst.desc.view_info.base_layer, 1});
 	return true;
 }
 

src/graphics/host_gpu/shaders/gpu_tiler_demote_d16.comp

@@ -1,10 +1,11 @@
 #version 450
+#extension GL_EXT_shader_16bit_storage : require
 
 layout(constant_id = 0) const uint D32 = 0u;
 layout(local_size_x = 64) in;
 
 layout(set = 0, binding = 0, std430) readonly buffer Input { uint data[]; } input_buffer;
-layout(set = 0, binding = 1, std430) buffer Output { uint data[]; } output_buffer;
+layout(set = 0, binding = 1, std430) buffer Output { uint16_t data[]; } output_buffer;
 
 layout(push_constant, std430) uniform Conversion {
 	uint src_base;
@@ -32,9 +33,5 @@ void main() {
 	}
 	uint src = params.src_base + y * params.source_pitch + x * 4u;
 	uint dst = params.dst_base + y * params.target_pitch + x * 2u;
-	uint shift = (dst & 2u) * 8u;
-	uint mask = 0xffffu << shift;
-	uint word = dst >> 2u;
-	atomicAnd(output_buffer.data[word], ~mask);
-	atomicOr(output_buffer.data[word], decode(input_buffer.data[src >> 2u]) << shift);
+	output_buffer.data[dst >> 1u] = uint16_t(decode(input_buffer.data[src >> 2u]));
 }

src/graphics/presentation/performanceOverlay.cpp

@@ -533,16 +533,16 @@ void DrawMemoryAndIo(const Snapshot& shown) {
 		    Bytes(guest.pooled_allocated) + " (" + Bytes(guest.pool_committed) + " committed)");
 		Row("Flexible", Bytes(guest.flexible_used) + " / " + Bytes(guest.flexible_total));
 		Row("Extended", Bytes(guest.automatic_allocated));
-		// Past the pool size the guest sees no page-table entries available.
-		const auto entries = std::max(guest.cpu_page_entries, guest.gpu_page_entries);
+		// Past its total a side sees no page-table entries available.
+		const auto percent = std::max(guest.cpu_page_entries * 100 / guest.cpu_page_entries_total,
+		                              guest.gpu_page_entries * 100 / guest.gpu_page_entries_total);
 		Row("Page table (2 MB)",
-		    Fmt("CPU %llu  GPU %llu  / %llu",
+		    Fmt("CPU %llu / %llu  GPU %llu / %llu",
 		        static_cast<unsigned long long>(guest.cpu_page_entries),
+		        static_cast<unsigned long long>(guest.cpu_page_entries_total),
 		        static_cast<unsigned long long>(guest.gpu_page_entries),
-		        static_cast<unsigned long long>(guest.page_entries_total)),
-		    entries > guest.page_entries_total            ? BAD
-		    : entries * 10 > guest.page_entries_total * 9 ? MEDIUM
-		                                                  : GOOD);
+		        static_cast<unsigned long long>(guest.gpu_page_entries_total)),
+		    LevelColor(static_cast<double>(percent), 90, 100));
 		ImGui::EndTable();
 	}
 	if (BeginSection("Emulator process")) {

src/graphics/presentation/window/vulkanWindow.cpp

@@ -51,6 +51,7 @@ struct VulkanExtensions {
 vk::PhysicalDeviceVulkan11Features WindowContext::RequiredVulkan11Features() noexcept {
 	vk::PhysicalDeviceVulkan11Features features {};
 	features.storageBuffer16BitAccess = VK_TRUE;
+	features.shaderDrawParameters     = VK_TRUE;
 	return features;
 }
 
@@ -278,6 +279,8 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 		              required_features12.samplerMirrorClampToEdge);
 		check_feature(features11.storageBuffer16BitAccess, "storageBuffer16BitAccess",
 		              WindowContext::RequiredVulkan11Features().storageBuffer16BitAccess);
+		check_feature(features11.shaderDrawParameters, "shaderDrawParameters",
+		              WindowContext::RequiredVulkan11Features().shaderDrawParameters);
 		check_feature(features12.storageBuffer8BitAccess, "storageBuffer8BitAccess",
 		              required_features12.storageBuffer8BitAccess);
 		check_feature(features12.timelineSemaphore, "timelineSemaphore",
@@ -337,6 +340,7 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 
 		check_feature(device_features2.features.shaderImageGatherExtended,
 		              "shaderImageGatherExtended");
+		check_feature(device_features2.features.shaderResourceMinLod, "shaderResourceMinLod");
 
 		check_feature(device_features2.features.independentBlend, "independentBlend");
 		check_feature(device_features2.features.tessellationShader, "tessellationShader");
@@ -636,6 +640,7 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 	device_features.shaderStorageImageMultisample        =
 	    supported_features2.features.shaderStorageImageMultisample;
 	device_features.shaderImageGatherExtended            = VK_TRUE;
+	device_features.shaderResourceMinLod                 = VK_TRUE;
 	device_features.independentBlend                     = VK_TRUE;
 	device_features.dualSrcBlend                         = VK_TRUE;
 	device_features.tessellationShader                   = VK_TRUE;
@@ -733,6 +738,15 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 		conditional_rendering.pNext = const_cast<void*>(create_info.pNext);
 		create_info.pNext = &conditional_rendering;
 	}
+	// The renderer only uses the Vulkan 1.3 core dynamic states, but RADV (Mesa 26.3) creates its
+	// PS epilog cache only when this EXT feature is enabled and crashes in vkCmdDraw when a
+	// depth- or stencil-writing fragment shader needs an epilog.
+	vk::PhysicalDeviceExtendedDynamicStateFeaturesEXT extended_dynamic_state {};
+	extended_dynamic_state.extendedDynamicState = VK_TRUE;
+	if (HasExtension(device_extensions, VK_EXT_EXTENDED_DYNAMIC_STATE_EXTENSION_NAME)) {
+		extended_dynamic_state.pNext = const_cast<void*>(create_info.pNext);
+		create_info.pNext = &extended_dynamic_state;
+	}
 	create_info.pQueueCreateInfos       = &queue_create_info;
 	create_info.queueCreateInfoCount    = 1;
 	create_info.enabledExtensionCount   = static_cast<uint32_t>(device_extensions.size());
@@ -1111,7 +1125,8 @@ void WindowContext::CreateVulkan() {
 		                             VK_KHR_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_EXTENSION_NAME,
 		                             VK_EXT_DEPTH_RANGE_UNRESTRICTED_EXTENSION_NAME,
 		                             VK_KHR_FRAGMENT_SHADER_BARYCENTRIC_EXTENSION_NAME,
-		                             VK_EXT_CONDITIONAL_RENDERING_EXTENSION_NAME}) {
+		                             VK_EXT_CONDITIONAL_RENDERING_EXTENSION_NAME,
+		                             VK_EXT_EXTENDED_DYNAMIC_STATE_EXTENSION_NAME}) {
 			if (HasExtension(available_extensions, extension)) {
 				device_extensions.push_back(extension);
 			}

src/graphics/shader/recompiler/ShaderRecompiler.cpp

@@ -30,10 +30,6 @@ namespace Libs::Graphics::ShaderRecompiler {
 
 namespace {
 
-const char* GetDumpLabel(const CompileOptions& options) {
-	return options.dump_label != nullptr ? options.dump_label : "ShaderRecompiler";
-}
-
 std::string MakeIrDump(std::string_view cfg, const IR::Program& ir) {
 	std::string dump = "CFG:\n";
 	dump += cfg;
@@ -44,30 +40,17 @@ std::string MakeIrDump(std::string_view cfg, const IR::Program& ir) {
 	return dump;
 }
 
-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";
-	}
-}
-
 void LogDispatcherFallback(const CompileOptions& options, const CFG::Graph& cfg, const char* phase) {
 	const auto* block        = cfg.FindBlock(cfg.failure_block);
 	const auto  start        = block != nullptr ? block->start_pc : UINT32_MAX;
 	const auto  end          = block != nullptr ? block->end_pc : UINT32_MAX;
 	const auto  predecessors = block != nullptr ? block->predecessors.size() : 0u;
 	const auto  successors   = block != nullptr ? block->successors.size() : 0u;
-	LOGF("%s CFG dispatcher fallback: stage=%s hash=0x%016" PRIx64
+	LOGF("ShaderRecompiler %s CFG dispatcher fallback: hash=0x%016" PRIx64
 	     " phase=%s failure=%s block=%" PRIu32 " pc=0x%08" PRIx32 "..0x%08" PRIx32 " preds=%" PRIu64
 	     " succs=%" PRIu64 " blocks=%" PRIu64 " loops=%" PRIu64 " back_edges=%" PRIu64
 	     " reason=%s\n",
-	     GetDumpLabel(options), StageName(options.stage), options.shader_hash, phase,
+	     StageName(options.stage), options.shader_hash, phase,
 	     CFG::FailureKindToString(cfg.failure_kind).c_str(), cfg.failure_block, start, end,
 	     static_cast<uint64_t>(predecessors), static_cast<uint64_t>(successors),
 	     static_cast<uint64_t>(cfg.blocks.size()), static_cast<uint64_t>(cfg.natural_loops.size()),
@@ -484,6 +467,7 @@ Decoder::Program DecodeFusedProgram(std::span<const uint32_t> front, std::span<c
 void PrepareCallTargets(ShaderSource& source, const CompileOptions& options) {
 	using namespace IR;
 	EXIT_NOT_IMPLEMENTED(options.stage != ShaderType::Compute);
+	source.wave_size = options.wave_size;
 	for (uint32_t i = 0; i < source.decoded.instructions.size(); ++i) {
 		if (source.decoded.instructions[i].opcode != Decoder::Opcode::S_SWAPPC_B64) continue;
 		source.calls.push_back({.instruction = i});
@@ -520,7 +504,8 @@ void PrepareCallTargets(ShaderSource& source, const CompileOptions& options) {
 	};
 	Register = [&](uint32_t before, uint32_t code) -> Value {
 		std::vector<uint32_t> crossed_calls;
-		const auto index = CFG::FindScalarDefinition(source.decoded, graph, before, code, crossed_calls);
+		const auto index = CFG::FindScalarDefinition(source.decoded, graph, before, code,
+		                                             source.wave_size, crossed_calls);
 		for (const auto crossed: crossed_calls) {
 			std::ranges::find(source.calls, crossed, &ShaderSource::Call::instruction)->preserved_registers.set(code);
 		}
@@ -584,14 +569,30 @@ void PrepareCallTargets(ShaderSource& source, const CompileOptions& options) {
 	for (auto& inst: block) {
 		inst.SetParent(nullptr);
 		if (inst.GetOpcode() == ValueOpcode::LoadAddressU32 || inst.GetOpcode() == ValueOpcode::ReadConstBuffer)
-			inst.SetFlags(SrtReadFlags {.index = inst.Flags<MemoryFlags>().index});
+			inst.SetFlags(MemoryFlags {.index = inst.Flags<MemoryFlags>().index});
 	}
 	query.value_storage.splice(query.value_storage.end(), block.Instructions());
 	source.call_targets = std::move(static_cast<ResourcePlan&>(query));
+	std::vector<Value> targets;
+	for (const auto& site: source.calls) targets.insert(targets.end(), site.target.begin(), site.target.end());
+	CompileSrtWalker(source.call_targets, targets);
 }
 
 } // namespace
 
+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";
+	}
+}
+
 ShaderSource PrepareShaderSource(std::span<const uint32_t> code, const CompileOptions& options) {
 	EXIT_IF(code.empty());
 	ShaderSource source;
@@ -617,12 +618,18 @@ const Decoder::Program& RefreshShaderSource(ShaderSource& source, const IR::SrtR
 	EXIT_IF(source.unsupported_call_pc != UINT32_MAX);
 	auto& reads = source.reads;
 	reads.clear();
-	IR::SrtReadCapture capture(runtime, reads);
-	const auto clean = IR::CleanRuntime(capture.ObservedRuntime());
-	IR::SrtWalker walker(source.call_targets, clean);
+	std::vector<uint32_t> flat;
+	const bool walked = IR::WalkSrt(source.call_targets, runtime, flat, &reads);
+	const auto read_code = [&](uint64_t address, std::span<uint32_t> words) {
+		if (runtime.read_specialization_memory == nullptr ||
+		    !runtime.read_specialization_memory(runtime.userdata, address, words)) return false;
+		reads.emplace_back(address, words.size_bytes());
+		return true;
+	};
 	for (auto& site: source.calls) {
 		uint32_t low = 0, high = 0;
-		if (!walker.Evaluate(site.target[0], low) || !walker.Evaluate(site.target[1], high))
+		if (!walked || !IR::WalkedValue(source.call_targets, site.target[0], low) ||
+		    !IR::WalkedValue(source.call_targets, site.target[1], high))
 			EXIT("shader call at pc 0x%08x has an unavailable scalar target",
 			     source.decoded.instructions[site.instruction].pc);
 		// SWAPPC reads the old aliased pair and ignores the target's low two bits.
@@ -646,7 +653,7 @@ const Decoder::Program& RefreshShaderSource(ShaderSource& source, const IR::SrtR
 				break;
 			}
 			linked.observed_function.resize(function.size);
-			if (!clean.read_memory(clean.userdata, source.calls[i].address, linked.observed_function) ||
+			if (!read_code(source.calls[i].address, linked.observed_function) ||
 			    !std::ranges::equal(expected, linked.observed_function)) {
 				unchanged = false;
 				break;
@@ -676,8 +683,7 @@ const Decoder::Program& RefreshShaderSource(ShaderSource& source, const IR::SrtR
 				uint32_t count = 1;
 				Decoder::Instruction inst;
 				for (;;) {
-					if (!clean.read_memory(clean.userdata, address + uint64_t {index} * sizeof(uint32_t),
-					                       std::span(words).first(count)))
+					if (!read_code(address + uint64_t {index} * sizeof(uint32_t), std::span(words).first(count)))
 						EXIT("shader call at pc 0x%08x cannot read leaf code", call.pc);
 					inst = {};
 					Decoder::DecodeInstruction(words, 0u, inst);
@@ -698,10 +704,11 @@ const Decoder::Program& RefreshShaderSource(ShaderSource& source, const IR::SrtR
 		EXIT_IF(function_pc > UINT32_MAX - function_size);
 		const auto function_offset = static_cast<uint32_t>(linked.code.size());
 		for (auto inst: function) {
-			EXIT_IF(CFG::MayWriteScalarRegister(inst, call.dst.reg) ||
-			        CFG::MayWriteScalarRegister(inst, call.dst.reg + 1u));
+			EXIT_IF(CFG::MayWriteScalarRegister(inst, call.dst.reg, source.wave_size) ||
+			        CFG::MayWriteScalarRegister(inst, call.dst.reg + 1u, source.wave_size));
 			for (uint32_t code = 0; code < site.preserved_registers.size(); ++code) {
-				if (site.preserved_registers[code] && CFG::MayWriteScalarRegister(inst, code))
+				if (site.preserved_registers[code] &&
+				    CFG::MayWriteScalarRegister(inst, code, source.wave_size))
 					EXIT("shader call at pc 0x%08x clobbers scalar target dependency s%u", call.pc, code);
 			}
 			linked.code.insert(linked.code.end(), inst.raw, inst.raw + inst.word_count);
@@ -756,6 +763,7 @@ TranslateResult TranslateProgram(const Decoder::Program& decoded, const CompileO
 		     static_cast<unsigned>(options.stage));
 	}
 
+	const auto* stage        = StageName(options.stage);
 	const auto compile_begin = std::chrono::steady_clock::now();
 	const auto phase_ms = [&compile_begin]() {
 		return static_cast<uint64_t>(std::chrono::duration_cast<std::chrono::milliseconds>(
@@ -765,27 +773,25 @@ TranslateResult TranslateProgram(const Decoder::Program& decoded, const CompileO
 	std::string decoded_dump;
 	if (options.dump_ir) {
 		decoded_dump = Decoder::ProgramToString(decoded);
-		if (options.early_dump) {
-			LOGF("%s decoded RDNA2 (early):\n%s", GetDumpLabel(options), decoded_dump.c_str());
-		}
+		LOGF("ShaderRecompiler %s decoded RDNA2:\n%s", stage, decoded_dump.c_str());
 	}
 
-	LOGF("%s phase begin: stage=%s hash=0x%016" PRIx64 " CFG BuildGraph\n", GetDumpLabel(options),
-	     StageName(options.stage), options.shader_hash);
+	LOGF("ShaderRecompiler %s phase begin: hash=0x%016" PRIx64 " CFG BuildGraph\n", stage,
+	     options.shader_hash);
 	auto native_cfg = CFG::BuildGraph(decoded);
 	CFG::Graph structured_cfg;
 	auto* selected_cfg = &native_cfg;
-	LOGF("%s phase end: stage=%s hash=0x%016" PRIx64 " CFG BuildGraph blocks=%" PRIu64
+	LOGF("ShaderRecompiler %s phase end: hash=0x%016" PRIx64 " CFG BuildGraph blocks=%" PRIu64
 	     " loops=%" PRIu64 " back_edges=%" PRIu64 " elapsed_ms=%" PRIu64 "\n",
-	     GetDumpLabel(options), StageName(options.stage), options.shader_hash,
+	     stage, options.shader_hash,
 	     static_cast<uint64_t>(native_cfg.blocks.size()),
 	     static_cast<uint64_t>(native_cfg.natural_loops.size()),
 	     static_cast<uint64_t>(native_cfg.back_edges.size()), phase_ms());
 	if (native_cfg.irreducible) {
 		LogDispatcherFallback(options, native_cfg, "build");
 	} else {
-		LOGF("%s phase begin: stage=%s hash=0x%016" PRIx64 " CFG Structurize\n",
-		     GetDumpLabel(options), StageName(options.stage), options.shader_hash);
+		LOGF("ShaderRecompiler %s phase begin: hash=0x%016" PRIx64 " CFG Structurize\n", stage,
+		     options.shader_hash);
 		structured_cfg = CFG::Structurize(native_cfg);
 		if (structured_cfg.unsupported) {
 			native_cfg.unsupported = true;
@@ -795,12 +801,12 @@ TranslateResult TranslateProgram(const Decoder::Program& decoded, const CompileO
 			LogDispatcherFallback(options, native_cfg, "structurize");
 		} else {
 			selected_cfg = &structured_cfg;
-			LOGF("%s structured CFG success: blocks=%" PRIu64 "\n", GetDumpLabel(options),
+			LOGF("ShaderRecompiler %s structured CFG success: blocks=%" PRIu64 "\n", stage,
 			     static_cast<uint64_t>(selected_cfg->blocks.size()));
 		}
-		LOGF("%s phase end: stage=%s hash=0x%016" PRIx64 " CFG Structurize blocks=%" PRIu64
+		LOGF("ShaderRecompiler %s phase end: hash=0x%016" PRIx64 " CFG Structurize blocks=%" PRIu64
 		     " loops=%" PRIu64 " elapsed_ms=%" PRIu64 "\n",
-		     GetDumpLabel(options), StageName(options.stage), options.shader_hash,
+		     stage, options.shader_hash,
 		     static_cast<uint64_t>(selected_cfg->blocks.size()),
 		     static_cast<uint64_t>(selected_cfg->natural_loops.size()), phase_ms());
 	}
@@ -813,7 +819,7 @@ TranslateResult TranslateProgram(const Decoder::Program& decoded, const CompileO
 		    decoded, *options.input_info.vertex, options.user_data_base,
 		    static_cast<uint32_t>(options.user_data.size()), options.wave_size);
 		if (!embedded_fetch.loads.empty()) {
-			LOGF("%s embedded vertex fetch plan: detected=%" PRIu64 "\n", GetDumpLabel(options),
+			LOGF("ShaderRecompiler %s embedded vertex fetch plan: detected=%" PRIu64 "\n", stage,
 			     static_cast<uint64_t>(embedded_fetch.loads.size()));
 		}
 	}
@@ -826,12 +832,12 @@ TranslateResult TranslateProgram(const Decoder::Program& decoded, const CompileO
 	    .input_info       = options.input_info,
 	    .embedded_fetch   = embedded_fetch.loads.empty() ? nullptr : &embedded_fetch,
 	};
-	LOGF("%s phase begin: stage=%s hash=0x%016" PRIx64 " IR TranslateProgram\n",
-	     GetDumpLabel(options), StageName(options.stage), options.shader_hash);
+	LOGF("ShaderRecompiler %s phase begin: hash=0x%016" PRIx64 " IR TranslateProgram\n", stage,
+	     options.shader_hash);
 	auto ir = Frontend::TranslateProgram(decoded, cfg, translate_options);
-	LOGF("%s phase end: stage=%s hash=0x%016" PRIx64 " IR TranslateProgram blocks=%" PRIu64
+	LOGF("ShaderRecompiler %s phase end: hash=0x%016" PRIx64 " IR TranslateProgram blocks=%" PRIu64
 	     " elapsed_ms=%" PRIu64 "\n",
-	     GetDumpLabel(options), StageName(options.stage), options.shader_hash,
+	     stage, options.shader_hash,
 	     static_cast<uint64_t>(ir.blocks.size()), phase_ms());
 	IR::RewriteToSsa(ir.blocks);
 	const auto simplify = [&ir] {
@@ -842,8 +848,9 @@ TranslateResult TranslateProgram(const Decoder::Program& decoded, const CompileO
 	simplify();
 	const auto read_lane_stats = IR::EliminateReadLane(ir, ir.wave_size);
 	if (read_lane_stats.rewritten_reads != 0 || read_lane_stats.rewritten_reductions != 0) {
-		LOGF("%s read-lane elimination: reads=%" PRIu32 " reductions=%" PRIu32 "\n",
-		     GetDumpLabel(options), read_lane_stats.rewritten_reads, read_lane_stats.rewritten_reductions);
+		LOGF("ShaderRecompiler %s read-lane elimination: reads=%" PRIu32
+		     " reductions=%" PRIu32 "\n",
+		     stage, read_lane_stats.rewritten_reads, read_lane_stats.rewritten_reductions);
 	}
 	// Cleanup also exposes profitable folds by removing dead predicate consumers.
 	simplify();
@@ -857,10 +864,8 @@ TranslateResult TranslateProgram(const Decoder::Program& decoded, const CompileO
 	std::string cfg_dump;
 	if (options.dump_ir) {
 		cfg_dump = CFG::GraphToString(cfg);
-		if (options.early_dump) {
-			LOGF("%s native IR before resource tracking:\n%s", GetDumpLabel(options),
-			     MakeIrDump(cfg_dump, ir).c_str());
-		}
+		LOGF("ShaderRecompiler %s native IR before resource tracking:\n%s", stage,
+		     MakeIrDump(cfg_dump, ir).c_str());
 	}
 	IR::TrackResources(ir, decoded, native_cfg, options.input_info);
 	TranslateResult result;
@@ -877,6 +882,7 @@ CompileResult CompileProgram(TranslateResult translated, const CompileOptions& o
                              uint32_t push_data_start_dword) {
 	const auto emit_begin = std::chrono::steady_clock::now();
 	auto& ir = translated.program;
+	const auto* stage = StageName(ir.stage);
 	IR::ApplyResourceSpecialization(ir, specialization);
 	// The resource plan owns host descriptor evaluation now. Keep only dependencies consumed
 	// by GPU memory operations; bound descriptor dwords must not retain shader instructions.
@@ -923,19 +929,16 @@ CompileResult CompileProgram(TranslateResult translated, const CompileOptions& o
 	std::string ir_dump;
 	if (options.dump_ir) {
 		ir_dump = MakeIrDump(translated.cfg_dump, ir);
-		if (options.early_dump) {
-			LOGF("%s native IR (early):\n%s", GetDumpLabel(options), ir_dump.c_str());
-		}
+		LOGF("ShaderRecompiler %s native IR:\n%s", stage, ir_dump.c_str());
 	}
 
-	LOGF("%s phase begin: stage=%s hash=0x%016" PRIx64 " SPIR-V EmitProgram\n",
-	     GetDumpLabel(options), StageName(ir.stage), ir.shader_hash);
+	LOGF("ShaderRecompiler %s phase begin: hash=0x%016" PRIx64 " SPIR-V EmitProgram\n", stage,
+	     ir.shader_hash);
 	auto spirv = Spirv::EmitProgram(ir, options.input_info, push_data_start_dword, *options.profile,
 	                                options.emit_debug_info);
-	LOGF("%s phase end: stage=%s hash=0x%016" PRIx64 " SPIR-V EmitProgram words=%" PRIu64
-	     " elapsed_ms=%" PRIu64 "\n",
-	     GetDumpLabel(options), StageName(ir.stage), ir.shader_hash,
-	     static_cast<uint64_t>(spirv.size()),
+	LOGF("ShaderRecompiler %s phase end: hash=0x%016" PRIx64 " SPIR-V EmitProgram words=%" PRIu64
+	     " wave_size=%u elapsed_ms=%" PRIu64 "\n",
+	     stage, ir.shader_hash, static_cast<uint64_t>(spirv.size()), ir.wave_size,
 	     static_cast<uint64_t>(std::chrono::duration_cast<std::chrono::milliseconds>(
 	                               std::chrono::steady_clock::now() - emit_begin)
 	                               .count()));

src/graphics/shader/recompiler/ShaderRecompiler.h

@@ -22,10 +22,8 @@ struct CompileOptions {
 	uint32_t                    user_data_base  = 0;
 	uint64_t                    shader_hash     = 0;
 	bool                        dump_ir                    = true;
-	bool                        early_dump                 = false;
 	bool                        emit_debug_info            = false;
 	const Spirv::Profile*       profile                    = &Spirv::DefaultProfile;
-	const char*                 dump_label                 = nullptr;
 	std::span<const uint32_t>   user_data;
 	std::span<const uint32_t>   back_code;
 	ShaderStageInputInfo        input_info;
@@ -64,11 +62,13 @@ struct ShaderSource {
 	IR::ResourcePlan call_targets;
 	std::vector<Call> calls;
 	uint32_t unsupported_call_pc = UINT32_MAX;
+	uint32_t wave_size = 64;
 	std::optional<Linked> linked;
 	uint64_t revision = 0;
 	std::vector<std::pair<uint64_t, uint64_t>> reads;
 };
 
+[[nodiscard]] const char* StageName(ShaderType stage);
 [[nodiscard]] ShaderSource PrepareShaderSource(std::span<const uint32_t> code,
 	                                           const CompileOptions& options);
 const Decoder::Program& RefreshShaderSource(ShaderSource& source, const IR::SrtRuntime& runtime);

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

@@ -54,9 +54,30 @@ uint32_t EmitExt(EmitterState& state, uint32_t type, uint32_t opcode,
 	return result;
 }
 
-uint32_t EmitF32ToF16RtzBits(EmitterState& state, uint32_t value) {
+uint32_t EmitF32ToU32(EmitterState& state, uint32_t src, bool signed_value) {
+	// Clamping first keeps the truncating conversion defined; NaN and values past the largest
+	// float below the integer limit are selected afterwards.
+	const auto clamped =
+	    EmitExt(state, TypeF32(state), GLSLstd450NClamp,
+	            {src, ConstantF32(state, signed_value ? 0xcf000000u : 0u),
+	             ConstantF32(state, signed_value ? 0x4effffffu : 0x4f7fffffu)});
+	const auto converted = signed_value
+	                           ? Unary(state, spv::OpBitcast, TypeU32(state),
+	                                   Unary(state, spv::OpConvertFToS, TypeI32(state), clamped))
+	                           : Unary(state, spv::OpConvertFToU, TypeU32(state), clamped);
+	const auto above     = Binary(state, spv::OpFOrdGreaterThanEqual, TypeBool(state), src,
+	                              ConstantF32(state, signed_value ? 0x4f000000u : 0x4f800000u));
+	const auto high =
+	    Select(state, TypeU32(state), above,
+	           ConstantU32(state, signed_value ? 0x7fffffffu : 0xffffffffu), converted);
+	return Select(state, TypeU32(state), Unary(state, spv::OpIsNan, TypeBool(state), src),
+	              ConstantU32(state, 0), high);
+}
+
+} // namespace
+
+uint32_t EmitF32BitsToF16RtzBits(EmitterState& state, uint32_t bits) {
 	// Native half conversion causes regressions; use compact integer RTZ conversion.
-	const auto bits = EmitBitcastF32ToU32(state, value);
 	const auto abs  = EmitAndConstant(state, bits, 0x7fffffffu);
 	const auto sign = EmitAndConstant(
 	    state, Binary(state, spv::OpShiftRightLogical, TypeU32(state), bits, ConstantU32(state, 16)),
@@ -92,28 +113,6 @@ uint32_t EmitF32ToF16RtzBits(EmitterState& state, uint32_t value) {
 	return EmitOrU32(state, sign, magnitude);
 }
 
-uint32_t EmitF32ToU32(EmitterState& state, uint32_t src, bool signed_value) {
-	// Clamping first keeps the truncating conversion defined; NaN and values past the largest
-	// float below the integer limit are selected afterwards.
-	const auto clamped =
-	    EmitExt(state, TypeF32(state), GLSLstd450NClamp,
-	            {src, ConstantF32(state, signed_value ? 0xcf000000u : 0u),
-	             ConstantF32(state, signed_value ? 0x4effffffu : 0x4f7fffffu)});
-	const auto converted = signed_value
-	                           ? Unary(state, spv::OpBitcast, TypeU32(state),
-	                                   Unary(state, spv::OpConvertFToS, TypeI32(state), clamped))
-	                           : Unary(state, spv::OpConvertFToU, TypeU32(state), clamped);
-	const auto above     = Binary(state, spv::OpFOrdGreaterThanEqual, TypeBool(state), src,
-	                              ConstantF32(state, signed_value ? 0x4f000000u : 0x4f800000u));
-	const auto high =
-	    Select(state, TypeU32(state), above,
-	           ConstantU32(state, signed_value ? 0x7fffffffu : 0xffffffffu), converted);
-	return Select(state, TypeU32(state), Unary(state, spv::OpIsNan, TypeBool(state), src),
-	              ConstantU32(state, 0), high);
-}
-
-} // namespace
-
 uint32_t EmitFPFma32(EmitterState& state, uint32_t a, uint32_t b, uint32_t c) {
 	const auto result = EmitGlsl<GLSLstd450Fma, IR::Type::F32>(state, a, b, c);
 	state.builder.AddAnnotation(spv::OpDecorate, result, spv::DecorationNoContraction);
@@ -199,7 +198,8 @@ uint32_t EmitCompositeExtractU32x2(EmitterState& state, uint32_t arg0, IR::Value
 
 uint32_t EmitPackFloat2x16Rtz(EmitterState& state, uint32_t arg0, uint32_t arg1) {
 	return EmitNative<spv::OpBitFieldInsert, IR::Type::U32>(
-	    state, EmitF32ToF16RtzBits(state, arg0), EmitF32ToF16RtzBits(state, arg1),
+	    state, EmitF32BitsToF16RtzBits(state, EmitBitcastF32ToU32(state, arg0)),
+	    EmitF32BitsToF16RtzBits(state, EmitBitcastF32ToU32(state, arg1)),
 	    ConstantU32(state, 16), ConstantU32(state, 16));
 }
 

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

@@ -30,6 +30,13 @@ bool PixelParameterIsCustom(const EmitterState& state, uint32_t attr) {
 	       ShaderPixelParameterIsCustom(*state.input_info.pixel, attr);
 }
 
+uint32_t MrtOutputType(EmitterState& state, uint32_t index) {
+	const auto& pixel = *state.input_info.pixel;
+	if ((pixel.target_sint_mask & (1u << index)) != 0u) return TypeI32Vector(state, 4);
+	if ((pixel.target_uint_mask & (1u << index)) != 0u) return TypeU32Vector(state, 4);
+	return TypeF32Vector(state, 4);
+}
+
 uint32_t OutputVariableForExport(const EmitterState& state, const IR::ExportInfo& exp) {
 	if (exp.kind == IR::ExportTargetKind::Position && exp.index == 0) {
 		return state.per_vertex_variable;

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

@@ -63,7 +63,8 @@ uint32_t EmitBuiltinU32(EmitterState& state, IR::StageInputKind kind, uint32_t c
 	}
 	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) {
+	    kind == IR::StageInputKind::Layer || kind == IR::StageInputKind::SampleId ||
+	    kind == IR::StageInputKind::BaseVertex || kind == IR::StageInputKind::BaseInstance) {
 		const auto value = state.builder.AllocateId();
 		const auto bits  = state.builder.AllocateId();
 		state.builder.AddFunction(spv::OpLoad, TypeI32(state), value, variable);
@@ -114,6 +115,18 @@ uint32_t EmitBuiltinU32(EmitterState& state, IR::StageInputKind kind, uint32_t c
 	return EmitInputComponentU32(state, kind, component);
 }
 
+// Whether EXEC enables a source lane within the caller's 32-lane half.
+uint32_t EmitSourceLaneActive(ValueEmitContext& ctx, uint32_t exec, uint32_t lane) {
+	auto&      state  = ctx.state;
+	const auto scope  = ConstantU32(state, spv::ScopeSubgroup);
+	const auto ballot = EmitNative<spv::OpGroupNonUniformBallot, IR::Type::U32x4>(state, scope, exec);
+	if (ctx.half != 0) {
+		lane = EmitAndConstant(state, lane, 31);
+	}
+	return EmitNative<spv::OpGroupNonUniformBallotBitExtract, IR::Type::U1>(state, scope, ballot,
+	                                                                        lane);
+}
+
 uint32_t EmitDppWriteCondition(ValueEmitContext& ctx, const IR::DppMoveFlags& flags,
                                uint32_t exec) {
 	auto&    state    = ctx.state;
@@ -135,9 +148,24 @@ uint32_t EmitDppWriteCondition(ValueEmitContext& ctx, const IR::DppMoveFlags& fl
 	// Only row shifts (0x101-0x10f, 0x111-0x11f) can vacate a lane; other controls always read one.
 	const bool row_shift = flags.control > 0x100u && flags.control < 0x120u &&
 	                       flags.control != 0x110u;
-	if (!flags.bound_control && row_shift) {
+	if (flags.bound_control || (flags.fetch_inactive && !row_shift)) {
+		return writable;
+	}
+	// Without BOUND_CTRL, a vacated source or, without FI, an EXEC-disabled one disables the write.
+	const auto target = EmitDppTargetLane(state, flags);
+	if (row_shift) {
+		writable = Binary(state, spv::OpLogicalAnd, TypeBool(state), writable, target.valid);
+	}
+	if (!flags.fetch_inactive) {
+		// Vacated lanes target outside their row; wrapping keeps the ballot index in range.
+		auto source = target.lane;
+		if (row_shift) {
+			lane   = lane != 0 ? lane : EmitSubgroupLocalInvocationId(state);
+			source = EmitNative<spv::OpBitFieldInsert, IR::Type::U32>(
+			    state, lane, source, ConstantU32(state, 0), ConstantU32(state, 4));
+		}
 		writable = Binary(state, spv::OpLogicalAnd, TypeBool(state), writable,
-		                  EmitDppTargetLane(state, flags).valid);
+		                  EmitSourceLaneActive(ctx, exec, source));
 	}
 	return writable;
 }
@@ -264,9 +292,10 @@ uint32_t ExportRawComponent(ValueEmitContext& ctx, uint32_t vector, uint32_t com
 }
 
 uint32_t ExportVector(ValueEmitContext& ctx, uint32_t data, const IR::ExportInfo& exp,
-                      bool uint_output) {
-	auto& state = ctx.state;
-	if (exp.compr && !uint_output) {
+                      uint32_t vector_type) {
+	auto&      state          = ctx.state;
+	const bool integer_output = vector_type != TypeF32Vector(state, 4);
+	if (exp.compr && !integer_output) {
 		auto unpack = GLSLstd450UnpackHalf2x16;
 		switch (MrtOutputMode(state, exp)) {
 			case 5: unpack = GLSLstd450UnpackUnorm2x16; break;
@@ -303,9 +332,11 @@ uint32_t ExportVector(ValueEmitContext& ctx, uint32_t data, const IR::ExportInfo
 		    ConstantU32(state, 0),
 		    ConstantU32(state, 0),
 		    ConstantU32(state, 0),
-		    ConstantU32(state, uint_output ? 1u : 0x3f800000u),
+		    ConstantU32(state, integer_output ? 1u : 0x3f800000u),
 		};
 		if (exp.compr) {
+			const auto extract = vector_type == TypeU32Vector(state, 4) ? spv::OpBitFieldUExtract
+			                                                            : spv::OpBitFieldSExtract;
 			for (uint32_t pair = 0; pair < 2u; pair++) {
 				if ((exp.en & (3u << (pair * 2u))) == 0u) {
 					continue;
@@ -317,8 +348,8 @@ uint32_t ExportVector(ValueEmitContext& ctx, uint32_t data, const IR::ExportInfo
 						continue;
 					}
 					raw[component] = state.builder.AllocateId();
-					state.builder.AddFunction(spv::OpBitFieldUExtract, TypeU32(state), raw[component],
-					                          packed, ConstantU32(state, lane * 16u),
+					state.builder.AddFunction(extract, TypeU32(state), raw[component], packed,
+					                          ConstantU32(state, lane * 16u),
 					                          ConstantU32(state, 16));
 				}
 			}
@@ -333,12 +364,10 @@ uint32_t ExportVector(ValueEmitContext& ctx, uint32_t data, const IR::ExportInfo
 		state.builder.AddFunction(spv::OpCompositeConstruct, TypeU32Vector(state, 4), data,
 		                          raw[0], raw[1], raw[2], raw[3]);
 	}
-	if (uint_output) {
+	if (vector_type == TypeU32Vector(state, 4)) {
 		return data;
 	}
-	const auto vector = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpBitcast, TypeF32Vector(state, 4), vector, data);
-	return vector;
+	return Unary(state, spv::OpBitcast, vector_type, data);
 }
 
 uint32_t DistancePointer(EmitterState& state, IR::StageOutputKind kind, uint32_t index) {
@@ -536,11 +565,11 @@ void EmitSetAttribute(ValueEmitContext& ctx, const IR::Inst& inst) {
 		if (state.program.stage != ShaderType::Mesh && variable == 0) {
 			return;
 		}
-		const bool uint_output = MrtOutputMode(state, exp) == 7u;
-		const auto vector_type = uint_output ? TypeU32Vector(state, 4) : TypeF32Vector(state, 4);
-		auto       value       = ExportVector(ctx, data, exp, uint_output);
+		const auto vector_type = MrtOutputMode(state, exp) != 0u ? MrtOutputType(state, exp.index)
+		                                                         : TypeF32Vector(state, 4);
+		auto       value       = ExportVector(ctx, data, exp, vector_type);
 		if (state.program.stage == ShaderType::Pixel && exp.kind == IR::ExportTargetKind::Mrt &&
-		    exp.index == 0 && !uint_output &&
+		    exp.index == 0 && vector_type == TypeF32Vector(state, 4) &&
 		    state.input_info.pixel->alpha_blend_source != ShaderAlphaBlendSource::None) {
 			// Broadcast logical alpha before swizzling the primary output.
 			const auto blend_output =
@@ -684,9 +713,16 @@ void EmitBarrier(EmitterState& state) {
 		return;
 	}
 	const auto memory_scope = tessellation ? spv::ScopeInvocation : spv::ScopeWorkgroup;
-	const auto memory = state.lds_storage_class == spv::StorageClassStorageBuffer
-	                        ? spv::MemorySemanticsUniformMemoryMask
-	                        : spv::MemorySemanticsWorkgroupMemoryMask;
+	auto memory = state.lds_storage_class == spv::StorageClassStorageBuffer
+	                  ? spv::MemorySemanticsUniformMemoryMask
+	                  : spv::MemorySemanticsWorkgroupMemoryMask;
+	// Shaders read other waves' buffer and image stores with plain loads right after S_BARRIER.
+	const auto written = [](const auto& resource) { return resource.written; };
+	if (std::ranges::any_of(state.program.info.buffers, written) ||
+	    std::ranges::any_of(state.program.info.images, written)) {
+		memory = memory | spv::MemorySemanticsUniformMemoryMask |
+		         spv::MemorySemanticsImageMemoryMask;
+	}
 	const auto semantics = tessellation ? spv::MemorySemanticsMaskNone
 	                                    : spv::MemorySemanticsAcquireReleaseMask | memory;
 	state.builder.AddFunction(spv::OpControlBarrier, ConstantU32(state, spv::ScopeWorkgroup),
@@ -784,6 +820,13 @@ uint32_t EmitBallot(ValueEmitContext& ctx, IR::Value predicate) {
 	return ctx.Ballot(predicate);
 }
 
+uint32_t EmitLiveMask(EmitterState& state) {
+	const auto helper =
+	    EmitNative<spv::OpLoad, IR::Type::U1>(state, state.helper_invocation_variable);
+	return EmitNative<spv::OpGroupNonUniformBallot, IR::Type::U32x4>(
+	    state, ConstantU32(state, spv::ScopeSubgroup), EmitLogicalNotBool(state, helper));
+}
+
 uint32_t EmitReadFirstLane(ValueEmitContext& ctx, const IR::Inst& inst) {
 	const auto ballot = ctx.Ballot(inst.Arg(1));
 	const auto lane   = ctx.FirstLane(ballot);
@@ -794,18 +837,28 @@ uint32_t EmitReadLane(ValueEmitContext& ctx, const IR::Inst& inst) {
 	return ctx.Shuffle(inst, 0, ctx.Arg(inst, 1));
 }
 
-uint32_t EmitSubgroupUMinU32(ValueEmitContext& ctx, const IR::Inst& inst) {
+template <spv::Op Reduce, auto Combine>
+static uint32_t EmitSubgroupReduceU32(ValueEmitContext& ctx, const IR::Inst& inst) {
 	auto& state = ctx.state;
 	auto value = ctx.Arg(inst, 0);
 	if (ctx.other_half != nullptr) {
-		value = EmitUMin32(state, ctx.HalfArg(inst, 0, 0), ctx.HalfArg(inst, 0, 1));
+		value = Combine(state, ctx.HalfArg(inst, 0, 0), ctx.HalfArg(inst, 0, 1));
 	}
 	const auto result = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpGroupNonUniformUMin, TypeU32(state), result,
-	                          ConstantU32(state, spv::ScopeSubgroup), spv::GroupOperationReduce, value);
+	state.builder.AddFunction(Reduce, TypeU32(state), result,
+	                          ConstantU32(state, spv::ScopeSubgroup), spv::GroupOperationReduce,
+	                          value);
 	return result;
 }
 
+uint32_t EmitSubgroupUMinU32(ValueEmitContext& ctx, const IR::Inst& inst) {
+	return EmitSubgroupReduceU32<spv::OpGroupNonUniformUMin, EmitUMin32>(ctx, inst);
+}
+
+uint32_t EmitSubgroupUMaxU32(ValueEmitContext& ctx, const IR::Inst& inst) {
+	return EmitSubgroupReduceU32<spv::OpGroupNonUniformUMax, EmitUMax32>(ctx, inst);
+}
+
 uint32_t EmitWriteLane(ValueEmitContext& ctx, const IR::Inst& inst) {
 	auto&      state = ctx.state;
 	const auto hit   = state.builder.AllocateId();
@@ -816,27 +869,23 @@ uint32_t EmitWriteLane(ValueEmitContext& ctx, const IR::Inst& inst) {
 }
 
 uint32_t EmitPermlane16U32(ValueEmitContext& ctx, const IR::Inst& inst) {
-	auto&      state     = ctx.state;
-	const auto flags     = inst.Flags<IR::PermlaneFlags>();
-	const auto subid     = EmitSubgroupLocalInvocationId(state);
-	const auto row       = state.builder.AllocateId();
-	const auto row_value = state.builder.AllocateId();
-	const auto lane      = state.builder.AllocateId();
-	const auto lane8     = state.builder.AllocateId();
-	const auto shift     = state.builder.AllocateId();
-	const auto upper     = state.builder.AllocateId();
-	const auto selected  = state.builder.AllocateId();
-	const auto shifted   = state.builder.AllocateId();
-	const auto index     = state.builder.AllocateId();
-	const auto target    = state.builder.AllocateId();
+	auto&      state    = ctx.state;
+	const auto flags    = inst.Flags<IR::PermlaneFlags>();
+	const auto subid    = EmitSubgroupLocalInvocationId(state);
+	const auto row      = state.builder.AllocateId();
+	const auto lane     = state.builder.AllocateId();
+	const auto lane8    = state.builder.AllocateId();
+	const auto shift    = state.builder.AllocateId();
+	const auto upper    = state.builder.AllocateId();
+	const auto selected = state.builder.AllocateId();
+	const auto shifted  = state.builder.AllocateId();
+	const auto index    = state.builder.AllocateId();
+	const auto target   = state.builder.AllocateId();
 	state.builder.AddFunction(spv::OpBitwiseAnd, TypeU32(state), row, subid,
 	                          ConstantU32(state, 0xfffffff0u));
-	if (flags.x16) {
-		state.builder.AddFunction(spv::OpBitwiseXor, TypeU32(state), row_value, row,
-		                          ConstantU32(state, 16));
-	} else {
-		state.builder.AddFunction(spv::OpCopyObject, TypeU32(state), row_value, row);
-	}
+	const auto row_value =
+	    flags.x16 ? Binary(state, spv::OpBitwiseXor, TypeU32(state), row, ConstantU32(state, 16))
+		          : row;
 	state.builder.AddFunction(spv::OpBitwiseAnd, TypeU32(state), lane, subid,
 	                          ConstantU32(state, 15));
 	state.builder.AddFunction(spv::OpBitwiseAnd, TypeU32(state), lane8, lane,
@@ -852,14 +901,11 @@ uint32_t EmitPermlane16U32(ValueEmitContext& ctx, const IR::Inst& inst) {
 	                          ConstantU32(state, 15));
 	state.builder.AddFunction(spv::OpBitwiseOr, TypeU32(state), target, row_value, index);
 	const auto shuffled = ctx.Shuffle(inst, 0, target);
-	uint32_t   result   = shuffled;
-	if (!flags.fetch_inactive) {
-		const auto source_exec = ctx.Shuffle(inst, 3, target);
-		result                 = state.builder.AllocateId();
-		state.builder.AddFunction(spv::OpSelect, TypeU32(state), result, source_exec, shuffled,
-		                          ConstantU32(state, 0));
+	if (flags.fetch_inactive) {
+		return shuffled;
 	}
-	return result;
+	return Select(state, TypeU32(state), EmitSourceLaneActive(ctx, ctx.Arg(inst, 3), target),
+	              shuffled, ctx.Arg(inst, 4));
 }
 
 uint32_t EmitGetAttribute(ValueEmitContext& ctx, const IR::Inst& inst) {

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

@@ -38,7 +38,8 @@ bool HasFlag(const IR::MemoryInfo& mem, uint32_t flag) {
 
 uint32_t AddressU32(ValueEmitContext& ctx, const IR::MemoryInfo& mem, const IR::Inst& address,
                     uint32_t component) {
-	const auto layout = Decoder::ImageAddressComponentLayout(mem.image_sample_flags, component);
+	const auto layout = Decoder::ImageAddressComponentLayout(mem.image_sample_flags,
+	                                                         mem.image_dimension, component);
 	const auto packed = layout.bit_offset / 32u;
 	if (packed >= address.NumArgs()) return ConstantU32(ctx.state, 0);
 	auto value = ctx.Def(address.Arg(packed));
@@ -55,10 +56,11 @@ uint32_t AddressU32(ValueEmitContext& ctx, const IR::MemoryInfo& mem, const IR::
 
 uint32_t AddressF32(ValueEmitContext& ctx, const IR::MemoryInfo& mem, const IR::Inst& address,
                     uint32_t component) {
-	const auto value = AddressU32(ctx, mem, address, component);
-	return Decoder::ImageAddressComponentLayout(mem.image_sample_flags, component).bit_width == 16u
-	           ? EmitF16BitsToF32(ctx.state, value)
-	           : Unary(ctx.state, spv::OpBitcast, TypeF32(ctx.state), value);
+	const auto value  = AddressU32(ctx, mem, address, component);
+	const auto layout = Decoder::ImageAddressComponentLayout(mem.image_sample_flags,
+	                                                         mem.image_dimension, component);
+	return layout.bit_width == 16u ? EmitF16BitsToF32(ctx.state, value)
+	                               : Unary(ctx.state, spv::OpBitcast, TypeF32(ctx.state), value);
 }
 
 ImageSampleLayout Layout(const IR::MemoryInfo& mem) {
@@ -77,6 +79,7 @@ ImageSampleLayout Layout(const IR::MemoryInfo& mem) {
 	layout.coord = cursor;
 	cursor += info.coordinate_components;
 	if (HasFlag(mem, Decoder::ImageSampleFlagLod)) layout.lod = cursor++;
+	if (HasFlag(mem, Decoder::ImageSampleFlagLodClamp)) layout.clamp = cursor++;
 	return layout;
 }
 
@@ -810,6 +813,25 @@ uint32_t EmitImageResult(ValueEmitContext& ctx, const IR::Inst& inst, const Emit
 	return mixed ? value : result_bits(mem.resource, value);
 }
 
+// Ordinals outside the table select candidates[0]. Past ordinal 0, a consecutive table is an
+// offset of ordinal - 1, which wraps at 0 so that one unsigned compare checks both ends.
+uint32_t SamplerArrayIndex(EmitterState& state, uint32_t ordinal,
+                           const std::vector<uint32_t>& candidates) {
+	const auto count = static_cast<uint32_t>(candidates.size());
+	const auto gap = std::adjacent_find(candidates.begin() + 1, candidates.end(),
+	    [](uint32_t lhs, uint32_t rhs) { return rhs != lhs + 1u; });
+	if (count == 1u || gap != candidates.end()) {
+		return EmitIndexSwitch(state, ordinal, count, TypeU32(state),
+		                       [&](uint32_t item) { return ConstantU32(state, candidates[item]); });
+	}
+	const auto offset = Binary(state, spv::OpISub, TypeU32(state), ordinal, ConstantU32(state, 1u));
+	const auto in_table = Binary(state, spv::OpULessThan, TypeBool(state), offset,
+	                             ConstantU32(state, count - 1u));
+	const auto index = Binary(state, spv::OpIAdd, TypeU32(state), offset,
+	                          ConstantU32(state, candidates[1]));
+	return Select(state, TypeU32(state), in_table, index, ConstantU32(state, candidates[0]));
+}
+
 struct ImageSampler {
 	uint32_t id = 0;
 	bool dynamic = false;
@@ -845,9 +867,7 @@ std::array<ImageSampler, 2> LoadImageSamplers(ValueEmitContext& ctx, const IR::I
 				selected = EmitIndirectResourceIndex(state, selected,
 				    samplers[index].indirect_mapping_offset,
 				    samplers[index].indirect_search_iterations, 0u);
-			array_index = EmitIndexSwitch(
-			    state, selected, static_cast<uint32_t>(candidates.size()),
-			    TypeU32(state), [&](uint32_t ordinal) { return ConstantU32(state, candidates[ordinal]); });
+			array_index = SamplerArrayIndex(state, selected, candidates);
 		}
 		sampler = {LoadSamplerDescriptor(state, index, array_index), array_index != 0u};
 	};
@@ -1069,7 +1089,12 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 				dref_value = AddressF32(ctx, mem, *address, layout.dref);
 			}
 		}
-		std::array<uint32_t, 4> operands {};
+		// _cl clamps the computed LOD from below.
+		uint32_t min_lod = 0;
+		if (layout.clamp != NoImageComponent) {
+			min_lod = AddressF32(ctx, mem, *address, layout.clamp);
+		}
+		std::array<uint32_t, 5> operands {};
 		size_t operand_count = 0;
 		if (HasFlag(mem, Decoder::ImageSampleFlagDerivative)) {
 			operands[0] = spv::ImageOperandsGradMask;
@@ -1081,6 +1106,11 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 			if (HasFlag(mem, Decoder::ImageSampleFlagLod) && layout.lod != NoImageComponent) {
 				lod = AddressF32(ctx, mem, *address, layout.lod);
 			}
+			// SPIR-V takes MinLod only with implicit LOD or Grad.
+			if (min_lod != 0) {
+				lod     = EmitGlsl<GLSLstd450FMax, IR::Type::F32>(state, lod, min_lod);
+				min_lod = 0;
+			}
 			operands[0] = spv::ImageOperandsLodMask;
 			operands[1] = lod;
 			operand_count = 2;
@@ -1104,6 +1134,13 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 				sample_operands[sample_operand_count++] = PackedOffset(
 				    ctx, offset_value, candidate.dimension, sample_operands[0]);
 			}
+			// MinLod follows the offset in operand order.
+			if (min_lod != 0) {
+				state.builder.RequireCapability(spv::CapabilityMinLod);
+				sample_operand_count = std::max(size_t{1}, sample_operand_count);
+				sample_operands[0] |= spv::ImageOperandsMinLodMask;
+				sample_operands[sample_operand_count++] = min_lod;
+			}
 			const auto coord =
 			    CoordF32(ctx, mem, *address, layout.coord,
 			             ImageDimensionInfoFor(candidate.dimension).coordinate_components,

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

@@ -56,6 +56,7 @@ EMIT_NATIVE(BitFieldSExtract, OpBitFieldSExtract, U32, uint32_t, uint32_t, uint3
 EMIT_NATIVE(SelectU1, OpSelect, U1, uint32_t, uint32_t, uint32_t)
 EMIT_NATIVE(SelectU32, OpSelect, U32, uint32_t, uint32_t, uint32_t)
 EMIT_NATIVE(SelectF32, OpSelect, F32, uint32_t, uint32_t, uint32_t)
+EMIT_NATIVE(SelectF64, OpSelect, F64, uint32_t, uint32_t, uint32_t)
 EMIT_NATIVE(IAdd32, OpIAdd, U32, uint32_t, uint32_t)
 EMIT_NATIVE(ISub32, OpISub, U32, uint32_t, uint32_t)
 EMIT_NATIVE(IMul32, OpIMul, U32, uint32_t, uint32_t)
@@ -229,10 +230,12 @@ uint32_t              EmitDppUpdateU32(ValueEmitContext& ctx, const IR::Inst& in
 uint32_t              EmitWqmU64(EmitterState& state, uint32_t value);
 uint32_t              EmitLaneId(EmitterState& state);
 uint32_t              EmitBallot(ValueEmitContext& ctx, IR::Value predicate);
+uint32_t              EmitLiveMask(EmitterState& state);
 uint32_t              EmitConditionRef(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitReadFirstLane(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitReadLane(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitSubgroupUMinU32(ValueEmitContext& ctx, const IR::Inst& inst);
+uint32_t              EmitSubgroupUMaxU32(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitWriteLane(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitPermlane16U32(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitGetAttribute(ValueEmitContext& ctx, const IR::Inst& inst);
@@ -288,6 +291,7 @@ inline constexpr auto EmitBufferAtomicIAdd32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicISub32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicSMin32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicUMin32    = EmitAtomic32;
+inline constexpr auto EmitBufferAtomicUMin64    = EmitBufferAtomic64;
 inline constexpr auto EmitBufferAtomicSMax32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicUMax32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicAnd32     = EmitAtomic32;

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

@@ -112,6 +112,7 @@ struct EmitterState {
 	uint32_t                                         lane_half               = 0;
 	std::array<BufferDefinition, 4>                  storage_buffers {};
 	std::vector<uint32_t>                           memory_byte_offsets;
+	std::vector<uint32_t>                           memory_byte_sizes;
 	uint32_t                                         bda_pagetable_variable  = 0;
 	uint32_t                                         fault_buffer_variable   = 0;
 	uint32_t                                         bda_pointer_function    = 0;
@@ -145,6 +146,7 @@ struct EmitterState {
 	const IR::Block*           current_block                         = nullptr;
 	uint32_t                   pixel_valid_mask_variable             = 0;
 	uint32_t                   subgroup_local_invocation_id_variable = 0;
+	uint32_t                   helper_invocation_variable            = 0;
 	uint32_t                   per_vertex_variable                   = 0;
 	uint32_t                   point_size_variable                   = 0;
 	uint32_t                   clip_distance_variable                = 0;
@@ -282,6 +284,7 @@ struct ImageSampleLayout {
 	uint32_t bias   = NoImageComponent;
 	uint32_t coord  = 0;
 	uint32_t lod    = NoImageComponent;
+	uint32_t clamp  = NoImageComponent;
 	uint32_t grad_x = NoImageComponent;
 	uint32_t grad_y = NoImageComponent;
 };
@@ -305,6 +308,7 @@ uint32_t VertexParameterScalarType(EmitterState& state, VertexInputScalarKind ki
 
 
 uint32_t OutputVariableForExport(const EmitterState& state, const IR::ExportInfo& exp);
+uint32_t MrtOutputType(EmitterState& state, uint32_t index);
 
 uint32_t ConstantU32(EmitterState& state, uint32_t value);
 
@@ -414,7 +418,7 @@ Prospero::BufferFormat StorageBufferFormat(const EmitterState& state, const IR::
 
 uint32_t SharedIndirectBufferResource(const EmitterState& state, uint32_t resource);
 
-void EmitMemoryOffsets(EmitterState& state);
+void EmitMemoryBindings(EmitterState& state);
 
 uint32_t LdsDwordCount(const EmitterState& state);
 void EnsureLdsStorage(EmitterState& state);
@@ -531,6 +535,7 @@ inline constexpr auto EmitFNegateValue = EmitNative<spv::OpFNegate, IR::Type::F3
 inline constexpr auto EmitFAbsValue = EmitGlsl<GLSLstd450FAbs, IR::Type::F32, uint32_t>;
 
 uint32_t EmitF16BitsToF32(EmitterState& state, uint32_t bits);
+uint32_t EmitF32BitsToF16RtzBits(EmitterState& state, uint32_t bits);
 
 void EmitProgram(EmitterState& state);
 

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

@@ -1541,16 +1541,16 @@ uint32_t EmitAppendConsume(ValueEmitContext& ctx, const IR::Inst& inst) {
 	const auto index =
 	    Binary(state, spv::OpShiftRightLogical, TypeU32(state), address, ConstantU32(state, 2));
 	const auto access = PrepareMemoryResourceAccess(state, mem);
-	const auto exec   = ctx.Arg(inst, 1);
-	const auto ballot = ctx.Ballot(inst.Arg(1));
-	const auto low    = state.builder.AllocateId();
-	const auto high   = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(state), low, ballot, 0);
-	state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(state), high, ballot, 1);
+	// Elect and count from the guest EXEC words, which V_MBCNT also ranks lanes by.
+	const auto low   = ctx.Arg(inst, 1);
+	const auto high  = ctx.Arg(inst, 2);
 	const auto count = Binary(state, spv::OpIAdd, TypeU32(state),
 	                          Unary(state, spv::OpBitCount, TypeU32(state), low),
 	                          Unary(state, spv::OpBitCount, TypeU32(state), high));
-	const auto first = ctx.FirstLane(ballot);
+	const auto mask  = state.builder.AllocateId();
+	state.builder.AddFunction(spv::OpCompositeConstruct, TypeU32Vector(state, 4), mask, low, high,
+	                          ConstantU32(state, 0), ConstantU32(state, 0));
+	const auto first = ctx.FirstLane(mask);
 	const auto source_lane =
 	    state.lane_count == 2
 	        ? Binary(state, spv::OpBitwiseAnd, TypeU32(state), first, ConstantU32(state, 31))
@@ -1564,9 +1564,7 @@ uint32_t EmitAppendConsume(ValueEmitContext& ctx, const IR::Inst& inst) {
 	const auto condition = AndCondition(
 	    state, is_first,
 	    AndCondition(state,
-	                 state.lane_count == 2 ? Binary(state, spv::OpINotEqual, TypeBool(state), count,
-	                                                ConstantU32(state, 0))
-	                                       : exec,
+	                 Binary(state, spv::OpINotEqual, TypeBool(state), count, ConstantU32(state, 0)),
 	                 bounds));
 	const auto atomic = EmitValueOrZeroIfCondition(state, condition, [&]() {
 		const auto value = state.builder.AllocateId();

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

@@ -154,9 +154,10 @@ uint32_t SharedIndirectBufferResource(const EmitterState& state, uint32_t resour
 	return representative;
 }
 
-void EmitMemoryOffsets(EmitterState& state) {
+void EmitMemoryBindings(EmitterState& state) {
 	const auto count = state.program.bindings.descriptor_counts[static_cast<size_t>(IR::DescriptorBindingKind::Buffers)];
 	state.memory_byte_offsets.resize(count);
+	state.memory_byte_sizes.resize(count);
 	std::vector<bool> dynamic_offsets(count);
 	for (uint32_t resource = 0; resource < state.program.info.buffers.size(); ++resource) {
 		if (SharedIndirectBufferResource(state, resource) == UINT32_MAX) continue;
@@ -174,6 +175,8 @@ void EmitMemoryOffsets(EmitterState& state) {
 			if (dynamic_offsets[i + lane]) continue;
 			state.memory_byte_offsets[i + lane] = EmitNative<spv::OpBitFieldUExtract, IR::Type::U32>(
 			    state, word, ConstantU32(state, lane * 8u), ConstantU32(state, 8u));
+			state.memory_byte_sizes[i + lane] =
+			    EmitShaderDataDwordLoad(state, state.program.bindings.MemorySizeDword() + i + lane);
 		}
 	}
 }
@@ -257,22 +260,28 @@ MemoryResourceAccess PrepareStorageBufferResourceAccess(EmitterState& state,
 		state.builder.RequireCapability(spv::CapabilityShaderNonUniform);
 		state.builder.RequireCapability(spv::CapabilityStorageBufferArrayNonUniformIndexing);
 		state.builder.AddAnnotation(spv::OpDecorate, access.object_pointer, spv::DecorationNonUniform);
-		const auto offsets_start = state.program.bindings.memory_offset_dword +
-		    (state.program.bindings.UsesPushData() ? state.program.bindings.push_data_start_dword : 0u);
+		const auto& bindings   = state.program.bindings;
+		const auto  data_start = bindings.UsesPushData() ? bindings.push_data_start_dword : 0u;
 		const auto word = EmitShaderDataDwordLoadAt(state, EmitAddU32(
-		    state, ConstantU32(state, offsets_start),
+		    state, ConstantU32(state, bindings.memory_offset_dword + data_start),
 		    Binary(state, spv::OpShiftRightLogical, TypeU32(state), array_index, ConstantU32(state, 2))));
 		const auto shift = Binary(state, spv::OpShiftLeftLogical, TypeU32(state),
 		    Binary(state, spv::OpBitwiseAnd, TypeU32(state), array_index, ConstantU32(state, 3)),
 		    ConstantU32(state, 3));
 		access.byte_offset = EmitNative<spv::OpBitFieldUExtract, IR::Type::U32>(
 		    state, word, shift, ConstantU32(state, 8));
+		access.length = EmitShaderDataDwordLoadAt(state, EmitAddU32(
+		    state, ConstantU32(state, bindings.MemorySizeDword() + data_start), array_index));
 	} else {
 		access.byte_offset = state.memory_byte_offsets[static_index];
+		access.length      = state.memory_byte_sizes[static_index];
+	}
+	// storage_buffers[i] views the binding as 1 << i byte elements.
+	const auto shift = static_cast<uint32_t>(&buffer - state.storage_buffers.data());
+	if (shift != 0u) {
+		access.length = Binary(state, spv::OpShiftRightLogical, TypeU32(state), access.length,
+		                       ConstantU32(state, shift));
 	}
-	access.length      = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpArrayLength, TypeU32(state), access.length,
-	                          access.object_pointer, 0);
 	return access;
 }
 
@@ -472,6 +481,17 @@ uint32_t EncodeFormattedStoreComponent(EmitterState& state, const Format::Buffer
                                        UnormRounding rounding) {
 	EXIT_IF(rounding != UnormRounding::RoundEven && info.type != Format::ComponentType::Unorm);
 	const auto bits = info.component_bits[component];
+	if (info.type == Format::ComponentType::Float && bits < 16u) {
+		// Truncate to f16, whose bias the unsigned 11/10-bit floats share, and drop mantissa
+		// bits; negatives and NaN, whose f16 bits lie above +inf, store zero.
+		const auto half = data_bits == 16u ? EmitAndConstant(state, data, 0xffffu)
+		                                   : EmitF32BitsToF16RtzBits(state, data);
+		return Select(state, TypeU32(state),
+		              EmitCompareU32Constant(state, spv::OpUGreaterThan, half, 0x7c00u),
+		              ConstantU32(state, 0),
+		              EmitBinaryU32(state, spv::OpShiftRightLogical, half,
+		                            ConstantU32(state, 15u - bits)));
+	}
 	if (data_bits == 16u) {
 		if (info.type == Format::ComponentType::Uint) {
 			data = EmitBitFieldUExtract(state, data, ConstantU32(state, 0), ConstantU32(state, 16));
@@ -565,6 +585,7 @@ spv::Op SpirvAtomicOpcode(IR::ValueOpcode opcode) {
 		case IR::ValueOpcode::ImageAtomicUMin32:
 		case IR::ValueOpcode::ImageAtomicUMin64:
 		case IR::ValueOpcode::BufferAtomicUMin32:
+		case IR::ValueOpcode::BufferAtomicUMin64:
 		case IR::ValueOpcode::SharedAtomicUMin32: return spv::OpAtomicUMin;
 		case IR::ValueOpcode::ImageAtomicSMax32:
 		case IR::ValueOpcode::BufferAtomicSMax32:

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

@@ -476,6 +476,8 @@ uint32_t BuiltInForInput(IR::StageInputKind kind) {
 		case IR::StageInputKind::LocalInvocationId: return spv::BuiltInLocalInvocationId;
 		case IR::StageInputKind::LocalInvocationIndex: return spv::BuiltInLocalInvocationIndex;
 		case IR::StageInputKind::GlobalInvocationId: return spv::BuiltInGlobalInvocationId;
+		case IR::StageInputKind::BaseVertex: return spv::BuiltInBaseVertex;
+		case IR::StageInputKind::BaseInstance: return spv::BuiltInBaseInstance;
 		default: return UINT32_MAX;
 	}
 }
@@ -526,7 +528,9 @@ void DefineInputs(EmitterState& state) {
 			case IR::StageInputKind::PrimitiveId:
 			case IR::StageInputKind::InstanceIndex:
 			case IR::StageInputKind::Layer:
-			case IR::StageInputKind::SampleId: type = TypeI32(state); break;
+			case IR::StageInputKind::SampleId:
+			case IR::StageInputKind::BaseVertex:
+			case IR::StageInputKind::BaseInstance: type = TypeI32(state); break;
 			case IR::StageInputKind::WorkgroupId:
 			case IR::StageInputKind::NumWorkgroups:
 			case IR::StageInputKind::LocalInvocationId:
@@ -595,6 +599,13 @@ void DefineInputs(EmitterState& state) {
 			state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationFlat);
 		}
 	}
+	if (state.program.info.helper_invocation) {
+		const auto variable = DefineInterfaceVariable(state, TypeBool(state), spv::StorageClassInput,
+		                                              "gl_HelperInvocation");
+		state.helper_invocation_variable = variable;
+		state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationBuiltIn,
+		                            spv::BuiltInHelperInvocation);
+	}
 }
 
 void DefineOutputs(EmitterState& state) {
@@ -691,13 +702,15 @@ void DefineOutputs(EmitterState& state) {
 				break;
 			case IR::StageOutputKind::Parameter:
 			case IR::StageOutputKind::Mrt: {
-				const bool uint_output =
-				    binding.kind == IR::StageOutputKind::Mrt &&
-				    state.program.stage == ShaderType::Pixel &&
-				    binding.index < std::size(state.input_info.pixel->target_output_mode) &&
-				    (state.input_info.pixel->target_output_mode[binding.index] == 7u ||
-				     (state.input_info.pixel->target_packed_unorm_mask & (1u << binding.index)) != 0u);
-				const auto type = uint_output ? TypeU32Vector(state, 4) : TypeF32Vector(state, 4);
+				auto type = TypeF32Vector(state, 4);
+				if (binding.kind == IR::StageOutputKind::Mrt &&
+				    state.program.stage == ShaderType::Pixel && binding.index < 8u) {
+					// Packed UNorm stores its encoded bits.
+					const auto bit = 1u << binding.index;
+					type = (state.input_info.pixel->target_packed_unorm_mask & bit) != 0u
+					           ? TypeU32Vector(state, 4)
+					           : MrtOutputType(state, binding.index);
+				}
 				binding.variable_id = DefineInterfaceVariable(state, type, spv::StorageClassOutput,
 				                                              binding.debug_name.c_str());
 				const bool dual_source = binding.kind == IR::StageOutputKind::Mrt &&
@@ -793,6 +806,10 @@ void DefineModule(EmitterState& state) {
 	if (InputVariableForKind(state, IR::StageInputKind::SampleId) != 0) {
 		state.builder.RequireCapability(spv::CapabilitySampleRateShading);
 	}
+	if (InputVariableForKind(state, IR::StageInputKind::BaseVertex) != 0 ||
+	    InputVariableForKind(state, IR::StageInputKind::BaseInstance) != 0) {
+		state.builder.RequireCapability(spv::CapabilityDrawParameters);
+	}
 	if (state.program.info.image_gather_extended) {
 		state.builder.RequireCapability(spv::CapabilityImageGatherExtended);
 	}

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

@@ -235,7 +235,7 @@ void EmitDirectInstruction(ValueEmitContext& ctx, const IR::Inst& inst) {
 	const auto op = inst.GetOpcode();
 	if (ctx.half != 0 &&
 	    (op == Op::Ballot || op == Op::ReadFirstLane || op == Op::SubgroupUMinU32 ||
-	     op == Op::ReadConstBuffer || op == Op::ReadRealtimeClock ||
+	     op == Op::SubgroupUMaxU32 || op == Op::ReadConstBuffer || op == Op::ReadRealtimeClock ||
 	     (op == Op::LoadAddressU32 && ctx.Memory(inst).kind == IR::ResourceKind::ScalarAddress))) {
 		// Whole-wave results, clocks and SMEM loads (SGPR operands, no EXEC) match in both halves.
 		ctx.Define(inst, ctx.other_half->Result(inst));
@@ -746,7 +746,7 @@ void EmitProgram(EmitterState& state) {
 		state.lds_base_dwords = EmitBinaryU32(state, spv::OpIMul, index,
 		                                     ConstantU32(state, LdsDwordCount(state)));
 	}
-	EmitMemoryOffsets(state);
+	EmitMemoryBindings(state);
 	if (program.blocks.empty()) {
 		EmitReturn(ctx);
 	} else if (state.program.dispatcher_fallback) {

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

@@ -1710,21 +1710,25 @@ Graph Structurize(const Graph& graph) {
 	return GotoStructurizer(graph).Run();
 }
 
-bool MayWriteScalarRegister(const Decoder::Instruction& inst, uint32_t code) {
-	// Vector comparisons and carry results can write a scalar pair in wave64.
+bool MayWriteScalarRegister(const Decoder::Instruction& inst, uint32_t code, uint32_t wave_size) {
+	// A vector lane mask (comparison or carry) spans wave_size / 32 SGPRs; a lane read writes one.
 	if (inst.family == Decoder::Family::VOP1 || inst.family == Decoder::Family::VOP2 ||
 	    inst.family == Decoder::Family::VOPC || inst.family == Decoder::Family::VOP3 ||
 	    inst.family == Decoder::Family::VOP3P) {
+		const bool lane_read =
+		    inst.opcode == Opcode::V_READLANE_B32 || inst.opcode == Opcode::V_READFIRSTLANE_B32;
+		const uint32_t size = lane_read ? 1u : wave_size / 32u;
 		uint32_t destination = 0;
 		for (const auto* operand: {&inst.dst, &inst.dst2})
-			if (ScalarOperandCode(*operand, destination) && code >= destination && code <= destination + 1u)
+			if (ScalarOperandCode(*operand, destination) && code >= destination && code < destination + size)
 				return true;
 	}
 	return InstructionWritesScalarCode(inst, code);
 }
 
 uint32_t FindScalarDefinition(const Decoder::Program& program, const Graph& graph,
-                              uint32_t before, uint32_t code, std::vector<uint32_t>& crossed_calls) {
+                              uint32_t before, uint32_t code, uint32_t wave_size,
+                              std::vector<uint32_t>& crossed_calls) {
 	EXIT_IF(before >= program.instructions.size());
 	const auto found = std::ranges::find_if(graph.blocks, [before](const auto& block) {
 		return block.inst_begin <= before && before < block.inst_end;
@@ -1743,11 +1747,11 @@ uint32_t FindScalarDefinition(const Decoder::Program& program, const Graph& grap
 		for (uint32_t i = point.second; i > block->inst_begin;) {
 			const auto& inst = program.instructions[--i];
 			if (inst.opcode == Opcode::S_SWAPPC_B64) {
-				EXIT_NOT_IMPLEMENTED(MayWriteScalarRegister(inst, code));
+				EXIT_NOT_IMPLEMENTED(MayWriteScalarRegister(inst, code, wave_size));
 				crossed_calls.push_back(i);
 				continue;
 			}
-			if (MayWriteScalarRegister(inst, code)) { index = i; break; }
+			if (MayWriteScalarRegister(inst, code, wave_size)) { index = i; break; }
 		}
 		if (index != UINT32_MAX || block->id == graph.entry_block) {
 			EXIT_IF(definition && *definition != index);

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

@@ -113,9 +113,10 @@ struct Graph {
 };
 
 Graph       BuildGraph(const Decoder::Program& program);
-bool MayWriteScalarRegister(const Decoder::Instruction& inst, uint32_t code);
+bool MayWriteScalarRegister(const Decoder::Instruction& inst, uint32_t code, uint32_t wave_size);
 uint32_t FindScalarDefinition(const Decoder::Program& program, const Graph& graph,
-                              uint32_t before, uint32_t code, std::vector<uint32_t>& crossed_calls);
+                              uint32_t before, uint32_t code, uint32_t wave_size,
+                              std::vector<uint32_t>& crossed_calls);
 // Returns structured control flow or failure diagnostics without changing the native graph.
 // On failure, failure_block is an original block ID or UINT32_MAX.
 Graph       Structurize(const Graph& graph);

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

@@ -14,12 +14,6 @@ struct MimgSampleInfo {
 	uint32_t    flags    = 0;
 };
 
-struct MimgGatherInfo {
-	uint32_t encoding = 0;
-	Opcode   decoded  = Opcode::UNSUPPORTED;
-	uint32_t flags    = 0;
-};
-
 struct MimgAtomicInfo {
 	uint32_t encoding    = 0;
 	Opcode   decoded     = Opcode::UNSUPPORTED;
@@ -73,7 +67,7 @@ constexpr uint32_t ImageSampleAddressComponents(uint32_t flags, ImageDimension d
 	if ((flags & ImageSampleFlagBias) != 0) {
 		components++;
 	}
-	if ((flags & ImageSampleFlagLod) != 0) {
+	if ((flags & (ImageSampleFlagLod | ImageSampleFlagLodClamp)) != 0) {
 		components++;
 	}
 	if ((flags & ImageSampleFlagDerivative) != 0) {
@@ -181,23 +175,72 @@ constexpr MimgSampleInfo MIMG_SAMPLE_OPCODE_LIST[] = {
          ImageSampleFlagOffset | ImageSampleFlagAdjust},
 };
 
-constexpr MimgGatherInfo MIMG_GATHER_OPCODE_LIST[] = {
-    {0x44u, Opcode::IMAGE_GATHER4_L, ImageSampleFlagLod},
-    {0x47u, Opcode::IMAGE_GATHER4_LZ, ImageSampleFlagLevelZero},
-    {0x48u, Opcode::IMAGE_GATHER4_C, ImageSampleFlagCompare},
-    {0x4cu, Opcode::IMAGE_GATHER4_C_L, ImageSampleFlagCompare | ImageSampleFlagLod},
-    {0x4fu, Opcode::IMAGE_GATHER4_C_LZ,
-     ImageSampleFlagCompare | ImageSampleFlagLevelZero},
-    {0x57u, Opcode::IMAGE_GATHER4_LZ_O,
-     ImageSampleFlagLevelZero | ImageSampleFlagOffset},
-    {0x58u, Opcode::IMAGE_GATHER4_C_O,
-     ImageSampleFlagCompare | ImageSampleFlagOffset},
-    {0x5fu, Opcode::IMAGE_GATHER4_C_LZ_O,
+constexpr MimgSampleInfo MIMG_GATHER_OPCODE_LIST[] = {
+    {0x40u, "image_gather4", 0},
+    {0x41u, "image_gather4_cl", ImageSampleFlagLodClamp},
+    {0x44u, "image_gather4_l", ImageSampleFlagLod},
+    {0x45u, "image_gather4_b", ImageSampleFlagBias},
+    {0x46u, "image_gather4_b_cl", ImageSampleFlagBias | ImageSampleFlagLodClamp},
+    {0x47u, "image_gather4_lz", ImageSampleFlagLevelZero},
+    {0x48u, "image_gather4_c", ImageSampleFlagCompare},
+    {0x49u, "image_gather4_c_cl", ImageSampleFlagCompare | ImageSampleFlagLodClamp},
+    {0x4cu, "image_gather4_c_l", ImageSampleFlagCompare | ImageSampleFlagLod},
+    {0x4du, "image_gather4_c_b", ImageSampleFlagCompare | ImageSampleFlagBias},
+    {0x4eu, "image_gather4_c_b_cl",
+     ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagLodClamp},
+    {0x4fu, "image_gather4_c_lz", ImageSampleFlagCompare | ImageSampleFlagLevelZero},
+    {0x50u, "image_gather4_o", ImageSampleFlagOffset},
+    {0x51u, "image_gather4_cl_o", ImageSampleFlagLodClamp | ImageSampleFlagOffset},
+    {0x54u, "image_gather4_l_o", ImageSampleFlagLod | ImageSampleFlagOffset},
+    {0x55u, "image_gather4_b_o", ImageSampleFlagBias | ImageSampleFlagOffset},
+    {0x56u, "image_gather4_b_cl_o",
+     ImageSampleFlagBias | ImageSampleFlagLodClamp | ImageSampleFlagOffset},
+    {0x57u, "image_gather4_lz_o", ImageSampleFlagLevelZero | ImageSampleFlagOffset},
+    {0x58u, "image_gather4_c_o", ImageSampleFlagCompare | ImageSampleFlagOffset},
+    {0x59u, "image_gather4_c_cl_o",
+     ImageSampleFlagCompare | ImageSampleFlagLodClamp | ImageSampleFlagOffset},
+    {0x5cu, "image_gather4_c_l_o",
+     ImageSampleFlagCompare | ImageSampleFlagLod | ImageSampleFlagOffset},
+    {0x5du, "image_gather4_c_b_o",
+     ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagOffset},
+    {0x5eu, "image_gather4_c_b_cl_o",
+     ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagLodClamp |
+         ImageSampleFlagOffset},
+    {0x5fu, "image_gather4_c_lz_o",
      ImageSampleFlagCompare | ImageSampleFlagLevelZero | ImageSampleFlagOffset},
-    {0x61u, Opcode::IMAGE_GATHER4H, ImageSampleFlagGatherHorizontal},
-    {0xc8u, Opcode::IMAGE_GATHER4_C, ImageSampleFlagCompare | ImageSampleFlagAdjust},
-    {0xd8u, Opcode::IMAGE_GATHER4_C_O,
+    {0x61u, "image_gather4h", ImageSampleFlagGatherHorizontal},
+    {0xc0u, "image_gather4_a", ImageSampleFlagAdjust},
+    {0xc1u, "image_gather4_cl_a", ImageSampleFlagLodClamp | ImageSampleFlagAdjust},
+    {0xc5u, "image_gather4_b_a", ImageSampleFlagBias | ImageSampleFlagAdjust},
+    {0xc6u, "image_gather4_b_cl_a",
+     ImageSampleFlagBias | ImageSampleFlagLodClamp | ImageSampleFlagAdjust},
+    {0xc8u, "image_gather4_c_a", ImageSampleFlagCompare | ImageSampleFlagAdjust},
+    {0xc9u, "image_gather4_c_cl_a",
+     ImageSampleFlagCompare | ImageSampleFlagLodClamp | ImageSampleFlagAdjust},
+    {0xcdu, "image_gather4_c_b_a",
+     ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagAdjust},
+    {0xceu, "image_gather4_c_b_cl_a",
+     ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagLodClamp |
+         ImageSampleFlagAdjust},
+    {0xd0u, "image_gather4_o_a", ImageSampleFlagOffset | ImageSampleFlagAdjust},
+    {0xd1u, "image_gather4_cl_o_a",
+     ImageSampleFlagLodClamp | ImageSampleFlagOffset | ImageSampleFlagAdjust},
+    {0xd5u, "image_gather4_b_o_a",
+     ImageSampleFlagBias | ImageSampleFlagOffset | ImageSampleFlagAdjust},
+    {0xd6u, "image_gather4_b_cl_o_a",
+     ImageSampleFlagBias | ImageSampleFlagLodClamp | ImageSampleFlagOffset |
+         ImageSampleFlagAdjust},
+    {0xd8u, "image_gather4_c_o_a",
      ImageSampleFlagCompare | ImageSampleFlagOffset | ImageSampleFlagAdjust},
+    {0xd9u, "image_gather4_c_cl_o_a",
+     ImageSampleFlagCompare | ImageSampleFlagLodClamp | ImageSampleFlagOffset |
+         ImageSampleFlagAdjust},
+    {0xddu, "image_gather4_c_b_o_a",
+     ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagOffset |
+         ImageSampleFlagAdjust},
+    {0xdeu, "image_gather4_c_b_cl_o_a",
+     ImageSampleFlagCompare | ImageSampleFlagBias | ImageSampleFlagLodClamp |
+         ImageSampleFlagOffset | ImageSampleFlagAdjust},
 };
 
 constexpr MimgAtomicInfo MIMG_ATOMIC_OPCODE_LIST[] = {
@@ -219,13 +262,13 @@ constexpr auto MIMG_SAMPLE_OPS = Detail::MakeOpcodeTable<0x100>(MIMG_SAMPLE_OPCO
 constexpr auto MIMG_GATHER_OPS = Detail::MakeOpcodeTable<0x100>(MIMG_GATHER_OPCODE_LIST);
 constexpr auto MIMG_ATOMIC_OPS = Detail::MakeOpcodeTable<0x100>(MIMG_ATOMIC_OPCODE_LIST);
 
-Opcode DecodeMimgOpcode(uint32_t opcode, const MimgSampleInfo* sample, const MimgGatherInfo* gather,
+Opcode DecodeMimgOpcode(uint32_t opcode, const MimgSampleInfo* sample, const MimgSampleInfo* gather,
                         const MimgAtomicInfo* atomic) {
 	if (sample != nullptr) {
 		return Opcode::IMAGE_SAMPLE;
 	}
 	if (gather != nullptr) {
-		return gather->decoded;
+		return Opcode::IMAGE_GATHER4;
 	}
 	if (atomic != nullptr) {
 		return atomic->decoded;
@@ -244,24 +287,10 @@ Opcode DecodeMimgOpcode(uint32_t opcode, const MimgSampleInfo* sample, const Mim
 	}
 }
 
-uint32_t DecodeMimgSampleFlags(const MimgSampleInfo* sample, const MimgGatherInfo* gather) {
-	if (sample != nullptr) {
-		return sample->flags;
-	}
-	if (gather != nullptr) {
-		return gather->flags;
-	}
-	return 0;
-}
-
 uint32_t DecodeMimgAddressComponents(uint32_t opcode, ImageDimension dimension,
-                                     const MimgSampleInfo* sample, const MimgGatherInfo* gather,
-                                     const MimgAtomicInfo* atomic) {
-	if (sample != nullptr) {
-		return ImageSampleAddressComponents(sample->flags, dimension);
-	}
-	if (gather != nullptr) {
-		return ImageSampleAddressComponents(gather->flags, dimension);
+                                     const MimgSampleInfo* sampled, const MimgAtomicInfo* atomic) {
+	if (sampled != nullptr) {
+		return ImageSampleAddressComponents(sampled->flags, dimension);
 	}
 	if (atomic != nullptr) {
 		return ImageCoordComponents(dimension);
@@ -290,33 +319,31 @@ uint32_t CountDmaskComponents(uint32_t dmask) {
 
 } // namespace
 
-ImageAddressComponent ImageAddressComponentLayout(uint32_t flags, uint32_t component) {
-	const auto width = [flags](uint32_t index) {
-		if ((flags & ImageSampleFlagA16) == 0u) return 32u;
-		uint32_t cursor = 0;
-		if ((flags & ImageSampleFlagOffset) != 0u) {
-			if (index == cursor++) return 32u;
-		}
-		if ((flags & ImageSampleFlagBias) != 0u) {
-			if (index == cursor++) return 16u;
-		}
-		if ((flags & ImageSampleFlagCompare) != 0u && index == cursor) return 32u;
-		return 16u;
-	};
-	uint32_t offset = 0;
-	for (uint32_t index = 0; index < component; index++) {
-		const auto bits = width(index);
-		if (bits == 32u) offset = (offset + 31u) & ~31u;
-		offset += bits;
+ImageAddressComponent ImageAddressComponentLayout(uint32_t flags, ImageDimension dimension,
+                                                  uint32_t component) {
+	if ((flags & ImageSampleFlagA16) == 0u) return {component * 32u, 32u};
+	// A16 offsets, bias and PCF reference each take a whole VGPR, the bias as a half in its low
+	// 16 bits. The other components pack two per VGPR, except that each 1D derivative and the last
+	// two 3D ones take the low half of their own VGPR.
+	const bool     has_bias    = (flags & ImageSampleFlagBias) != 0u;
+	const uint32_t bias_index  = (flags & ImageSampleFlagOffset) != 0u ? 1u : 0u;
+	const uint32_t whole_vgprs = std::popcount(
+	    flags & (ImageSampleFlagOffset | ImageSampleFlagBias | ImageSampleFlagCompare));
+	if (component < whole_vgprs) {
+		return {component * 32u, has_bias && component == bias_index ? 16u : 32u};
 	}
-	const auto bits = width(component);
-	if (bits == 32u) offset = (offset + 31u) & ~31u;
-	return {offset, bits};
+	const uint32_t axes =
+	    (flags & ImageSampleFlagDerivative) != 0u ? ImageGradientComponents(dimension) : 0u;
+	const uint32_t unpaired = axes % 2u * 2u;
+	const uint32_t paired   = axes * 2u - unpaired;
+	const uint32_t index    = component - whole_vgprs;
+	const uint32_t half     = index + std::clamp(index, paired, paired + unpaired) - paired;
+	return {whole_vgprs * 32u + half * 16u, 16u};
 }
 
-uint32_t ImageAddressDwordCount(uint32_t flags, uint32_t components) {
+uint32_t ImageAddressDwordCount(uint32_t flags, ImageDimension dimension, uint32_t components) {
 	if (components == 0u) return 0u;
-	const auto last = ImageAddressComponentLayout(flags, components - 1u);
+	const auto last = ImageAddressComponentLayout(flags, dimension, components - 1u);
 	return (last.bit_offset + last.bit_width + 31u) / 32u;
 }
 
@@ -329,16 +356,17 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	const auto     dimension  = DecodeImageDimension((word0 >> 3u) & 0x7u);
 	const uint32_t word_count = 2u + nsa_dwords;
 
-	const uint32_t vdata  = (word1 >> 8u) & 0xffu;
-	const uint32_t vaddr  = word1 & 0xffu;
-	const uint32_t srsrc  = (word1 >> 16u) & 0x1fu;
-	const uint32_t ssamp  = (word1 >> 21u) & 0x1fu;
-	const bool     r128   = ((word0 >> 15u) & 0x1u) != 0u;
-	const bool     a16    = ((word1 >> 30u) & 0x1u) != 0u;
-	const bool     d16    = ((word1 >> 31u) & 0x1u) != 0u;
-	const auto*    sample = Detail::FindOpcode(MIMG_SAMPLE_OPS, opcode);
-	const auto*    gather = Detail::FindOpcode(MIMG_GATHER_OPS, opcode);
-	const auto*    atomic = Detail::FindOpcode(MIMG_ATOMIC_OPS, opcode);
+	const uint32_t vdata   = (word1 >> 8u) & 0xffu;
+	const uint32_t vaddr   = word1 & 0xffu;
+	const uint32_t srsrc   = (word1 >> 16u) & 0x1fu;
+	const uint32_t ssamp   = (word1 >> 21u) & 0x1fu;
+	const bool     r128    = ((word0 >> 15u) & 0x1u) != 0u;
+	const bool     a16     = ((word1 >> 30u) & 0x1u) != 0u;
+	const bool     d16     = ((word1 >> 31u) & 0x1u) != 0u;
+	const auto*    sample  = Detail::FindOpcode(MIMG_SAMPLE_OPS, opcode);
+	const auto*    gather  = Detail::FindOpcode(MIMG_GATHER_OPS, opcode);
+	const auto*    atomic  = Detail::FindOpcode(MIMG_ATOMIC_OPS, opcode);
+	const auto*    sampled = sample != nullptr ? sample : gather;
 
 	inst.pc                 = pc;
 	inst.word_count         = word_count;
@@ -351,11 +379,11 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	inst.data_dwords        = d16 ? (inst.data_components + 1u) / 2u : inst.data_components;
 	inst.glc                = ((word0 >> 13u) & 1u) != 0;
 	inst.slc                = ((word0 >> 25u) & 1u) != 0;
-	inst.image_sample_flags = DecodeMimgSampleFlags(sample, gather);
+	inst.image_sample_flags = sampled != nullptr ? sampled->flags : 0u;
 	if (a16) {
 		inst.image_sample_flags |= ImageSampleFlagA16;
 	}
-	if ((sample != nullptr || gather != nullptr) && (word0 & (1u << 12u)) != 0u) {
+	if (sampled != nullptr && (word0 & (1u << 12u)) != 0u) {
 		inst.image_sample_flags |= ImageSampleFlagUnnormalized;
 	}
 	inst.image_dimension  = dimension;
@@ -364,8 +392,7 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	for (uint32_t i = 0; i < nsa_dwords * 4u; i++) {
 		inst.image_nsa_addr[i] = (code[word_index + 2u + i / 4u] >> ((i % 4u) * 8u)) & 0xffu;
 	}
-	inst.image_address_components =
-	    DecodeMimgAddressComponents(opcode, dimension, sample, gather, atomic);
+	inst.image_address_components = DecodeMimgAddressComponents(opcode, dimension, sampled, atomic);
 	SetRawWords(inst, code, word_index, word_count);
 
 	if (inst.opcode == Opcode::UNSUPPORTED) {
@@ -389,8 +416,8 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 			}
 		}
 	}
-	const bool supports_d16 = sample != nullptr || gather != nullptr || opcode == 0x00u ||
-	                          opcode == 0x01u || opcode == 0x08u || opcode == 0x09u;
+	const bool supports_d16 = sampled != nullptr || opcode == 0x00u || opcode == 0x01u ||
+	                          opcode == 0x08u || opcode == 0x09u;
 	if (d16 && !supports_d16) {
 		SetUnsupported(inst, Family::MIMG, opcode, "MIMG opcode does not support D16 data");
 	}
@@ -407,9 +434,10 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	inst.src_count = 3;
 }
 
-const char* MimgSampleOpcodeName(uint32_t opcode) {
-	const auto* sample = Detail::FindOpcode(MIMG_SAMPLE_OPS, opcode);
-	return sample != nullptr ? sample->name : nullptr;
+const char* MimgSampledOpcodeName(uint32_t opcode) {
+	const auto* sample  = Detail::FindOpcode(MIMG_SAMPLE_OPS, opcode);
+	const auto* sampled = sample != nullptr ? sample : Detail::FindOpcode(MIMG_GATHER_OPS, opcode);
+	return sampled != nullptr ? sampled->name : nullptr;
 }
 
 } // namespace Libs::Graphics::ShaderRecompiler::Decoder

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

@@ -8,15 +8,16 @@ namespace Libs::Graphics::ShaderRecompiler::Decoder {
 void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index,
                 Instruction& inst);
 
-const char* MimgSampleOpcodeName(uint32_t opcode);
+const char* MimgSampledOpcodeName(uint32_t opcode);
 
 struct ImageAddressComponent {
 	uint32_t bit_offset;
 	uint32_t bit_width;
 };
 
-ImageAddressComponent ImageAddressComponentLayout(uint32_t flags, uint32_t component);
-uint32_t ImageAddressDwordCount(uint32_t flags, uint32_t components);
+ImageAddressComponent ImageAddressComponentLayout(uint32_t flags, ImageDimension dimension,
+                                                  uint32_t component);
+uint32_t ImageAddressDwordCount(uint32_t flags, ImageDimension dimension, uint32_t components);
 
 } // namespace Libs::Graphics::ShaderRecompiler::Decoder
 

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

@@ -22,7 +22,7 @@ constexpr MemoryOpcodeInfo SMEM_OPCODE_LIST[] = {
     {0x04u, Opcode::S_LOAD_DWORDX16, 16, 32},      {0x08u, Opcode::S_BUFFER_LOAD_DWORD, 1, 32},
     {0x09u, Opcode::S_BUFFER_LOAD_DWORDX2, 2, 32}, {0x0au, Opcode::S_BUFFER_LOAD_DWORDX4, 4, 32},
     {0x0bu, Opcode::S_BUFFER_LOAD_DWORDX8, 8, 32}, {0x0cu, Opcode::S_BUFFER_LOAD_DWORDX16, 16, 32},
-    {0x25u, Opcode::S_MEMREALTIME, 2, 32},
+    {0x20u, Opcode::S_DCACHE_INV, 0, 0},           {0x25u, Opcode::S_MEMREALTIME, 2, 32},
 };
 
 constexpr MemoryOpcodeInfo MUBUF_OPCODE_LIST[] = {
@@ -69,6 +69,7 @@ constexpr MemoryOpcodeInfo MUBUF_OPCODE_LIST[] = {
     {0x40u, Opcode::BUFFER_ATOMIC_FMAX, 1, 32},
     {0x50u, Opcode::BUFFER_ATOMIC_SWAP_X2, 2, 32},
     {0x52u, Opcode::BUFFER_ATOMIC_ADD_X2, 2, 32},
+    {0x56u, Opcode::BUFFER_ATOMIC_UMIN_X2, 2, 32},
     {0x59u, Opcode::BUFFER_ATOMIC_AND_X2, 2, 32},
     {0x5au, Opcode::BUFFER_ATOMIC_OR_X2, 2, 32},
     {0x71u, Opcode::BUFFER_GL0_INV, 0, 0},
@@ -95,14 +96,14 @@ constexpr MemoryOpcodeInfo MTBUF_OPCODE_LIST[] = {
 };
 
 constexpr MemoryOpcodeInfo FLAT_OPCODE_LIST[] = {
-    {0x08u, Opcode::FLAT_LOAD_UBYTE, 1, 8},     {0x09u, Opcode::FLAT_LOAD_SBYTE, 1, 8, true},
-    {0x0au, Opcode::FLAT_LOAD_USHORT, 1, 16},   {0x0bu, Opcode::FLAT_LOAD_SSHORT, 1, 16, true},
-    {0x0cu, Opcode::FLAT_LOAD_DWORD, 1, 32},    {0x0du, Opcode::FLAT_LOAD_DWORDX2, 2, 32},
-    {0x0eu, Opcode::FLAT_LOAD_DWORDX4, 4, 32},  {0x0fu, Opcode::FLAT_LOAD_DWORDX3, 3, 32},
-    {0x18u, Opcode::FLAT_STORE_BYTE, 1, 8},     {0x1au, Opcode::FLAT_STORE_SHORT, 1, 16},
-    {0x1cu, Opcode::FLAT_STORE_DWORD, 1, 32},   {0x1du, Opcode::FLAT_STORE_DWORDX2, 2, 32},
-    {0x1eu, Opcode::FLAT_STORE_DWORDX4, 4, 32}, {0x1fu, Opcode::FLAT_STORE_DWORDX3, 3, 32},
-    {0x24u, Opcode::FLAT_LOAD_SHORT_D16, 1, 16},
+    {0x08u, Opcode::FLAT_LOAD_UBYTE, 1, 8},      {0x09u, Opcode::FLAT_LOAD_SBYTE, 1, 8, true},
+    {0x0au, Opcode::FLAT_LOAD_USHORT, 1, 16},    {0x0bu, Opcode::FLAT_LOAD_SSHORT, 1, 16, true},
+    {0x0cu, Opcode::FLAT_LOAD_DWORD, 1, 32},     {0x0du, Opcode::FLAT_LOAD_DWORDX2, 2, 32},
+    {0x0eu, Opcode::FLAT_LOAD_DWORDX4, 4, 32},   {0x0fu, Opcode::FLAT_LOAD_DWORDX3, 3, 32},
+    {0x18u, Opcode::FLAT_STORE_BYTE, 1, 8},      {0x1au, Opcode::FLAT_STORE_SHORT, 1, 16},
+    {0x1cu, Opcode::FLAT_STORE_DWORD, 1, 32},    {0x1du, Opcode::FLAT_STORE_DWORDX2, 2, 32},
+    {0x1eu, Opcode::FLAT_STORE_DWORDX4, 4, 32},  {0x1fu, Opcode::FLAT_STORE_DWORDX3, 3, 32},
+    {0x24u, Opcode::FLAT_LOAD_SHORT_D16, 1, 16}, {0x25u, Opcode::FLAT_LOAD_SHORT_D16_HI, 1, 16},
 };
 
 constexpr MemoryOpcodeInfo DS_OPCODE_LIST[] = {
@@ -285,6 +286,9 @@ void DecodeSmem(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	if (inst.opcode == Opcode::UNSUPPORTED) {
 		SetUnsupported(inst, Family::SMEM, opcode, "SMEM opcode is not implemented");
 	}
+	if (inst.opcode == Opcode::S_DCACHE_INV) {
+		return;
+	}
 
 	DecodeScalarDestination(sdst, pc, inst.dst);
 	if (inst.opcode == Opcode::S_MEMREALTIME) {
@@ -419,9 +423,10 @@ void DecodeFlat(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	}
 
 	DecodeVectorGpr(IsFlatStoreOpcode(inst.opcode) ? data : vdst, inst.dst);
-	if (inst.opcode == Opcode::FLAT_LOAD_SHORT_D16) {
-		// D16 loads reuse partial destinations to preserve the untouched high half.
-		inst.dst.sdwa_sel = 4u;
+	if (inst.opcode == Opcode::FLAT_LOAD_SHORT_D16 ||
+	    inst.opcode == Opcode::FLAT_LOAD_SHORT_D16_HI) {
+		// D16 loads reuse partial destinations to preserve the untouched half.
+		inst.dst.sdwa_sel = inst.opcode == Opcode::FLAT_LOAD_SHORT_D16_HI ? 5u : 4u;
 	}
 	DecodeVectorGpr(addr, inst.src0);
 	inst.src_count = 1;

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

@@ -123,10 +123,9 @@ std::string ImageSampleFlagsToString(uint32_t flags) {
 }
 
 std::string FormatMimg(const Instruction& inst) {
-	const char* sample_name =
-	    inst.opcode == Opcode::IMAGE_SAMPLE ? MimgSampleOpcodeName(inst.opcode_id) : nullptr;
+	const char*            sampled_name = MimgSampledOpcodeName(inst.opcode_id);
 	const std::string_view name =
-	    sample_name != nullptr ? sample_name : magic_enum::enum_name(inst.opcode);
+	    sampled_name != nullptr ? sampled_name : magic_enum::enum_name(inst.opcode);
 	std::string text =
 	    fmt::format("0x{:08x}: {} {}, {}, {}, {} ; dmask=0x{:x} image_dim={}", inst.pc, name,
 	                OperandToString(inst.dst).c_str(), OperandToString(inst.src0).c_str(),
@@ -138,22 +137,10 @@ std::string FormatMimg(const Instruction& inst) {
 	if (inst.image_r128) {
 		text += " r128=1";
 	}
-	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_L:
-		case Opcode::IMAGE_GATHER4_C_LZ:
-		case Opcode::IMAGE_GATHER4_LZ_O:
-		case Opcode::IMAGE_GATHER4_C_O:
-		case Opcode::IMAGE_GATHER4_C_LZ_O:
-		case Opcode::IMAGE_GATHER4H:
-			text += fmt::format(" sample_flags={} addr_components={}",
-			                    ImageSampleFlagsToString(inst.image_sample_flags).c_str(),
-			                    inst.image_address_components);
-			break;
-		default: break;
+	if (sampled_name != nullptr) {
+		text += fmt::format(" sample_flags={} addr_components={}",
+		                    ImageSampleFlagsToString(inst.image_sample_flags).c_str(),
+		                    inst.image_address_components);
 	}
 	if (inst.image_nsa_dwords != 0) {
 		text += fmt::format(" nsa_dwords={} nsa_addr=", inst.image_nsa_dwords);
@@ -258,7 +245,9 @@ void DecodeScalarSource(uint32_t code, uint32_t pc, Operand& operand) {
 		case 126u: operand.kind = OperandKind::ExecLo; return;
 		case 127u: operand.kind = OperandKind::ExecHi; return;
 		case 235u: operand.kind = OperandKind::SharedBase; return;
+		case 236u: operand.kind = OperandKind::SharedLimit; return;
 		case 237u: operand.kind = OperandKind::PrivateBase; return;
+		case 238u: operand.kind = OperandKind::PrivateLimit; return;
 		case 239u: operand.kind = OperandKind::PopsExitingWaveId; return;
 		case 248u:
 			operand.kind      = OperandKind::FloatInlineConstant;
@@ -459,7 +448,9 @@ std::string OperandToString(const Operand& operand) {
 		case OperandKind::M0: text = "m0"; break;
 		case OperandKind::PopsExitingWaveId: text = "pops_exiting_wave_id"; break;
 		case OperandKind::SharedBase: text = "shared_base"; break;
+		case OperandKind::SharedLimit: text = "shared_limit"; break;
 		case OperandKind::PrivateBase: text = "private_base"; break;
+		case OperandKind::PrivateLimit: text = "private_limit"; break;
 		case OperandKind::Null: text = "null"; break;
 		default: text = "unknown"; break;
 	}
@@ -566,6 +557,7 @@ std::string InstructionToString(const Instruction& inst) {
 			                                               OperandToString(inst.src0).c_str()));
 		case Opcode::S_BARRIER:
 			return WithUnsupportedReason(inst, fmt::format("0x{:08x}: s_barrier", inst.pc));
+		case Opcode::S_DCACHE_INV:
 		case Opcode::BUFFER_GL0_INV:
 		case Opcode::BUFFER_GL1_INV:
 			return fmt::format("0x{:08x}: {}", inst.pc, magic_enum::enum_name(inst.opcode));
@@ -612,15 +604,7 @@ 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_L:
-		case Opcode::IMAGE_GATHER4_C_LZ:
-		case Opcode::IMAGE_GATHER4_LZ_O:
-		case Opcode::IMAGE_GATHER4_C_O:
-		case Opcode::IMAGE_GATHER4_C_LZ_O:
-		case Opcode::IMAGE_GATHER4H: return WithUnsupportedReason(inst, FormatMimg(inst));
+		case Opcode::IMAGE_GATHER4: return WithUnsupportedReason(inst, FormatMimg(inst));
 		case Opcode::S_LOAD_DWORD:
 		case Opcode::S_LOAD_DWORDX2:
 		case Opcode::S_LOAD_DWORDX4:
@@ -681,6 +665,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::BUFFER_ATOMIC_SUB:
 		case Opcode::BUFFER_ATOMIC_SMIN:
 		case Opcode::BUFFER_ATOMIC_UMIN:
+		case Opcode::BUFFER_ATOMIC_UMIN_X2:
 		case Opcode::BUFFER_ATOMIC_SMAX:
 		case Opcode::BUFFER_ATOMIC_UMAX:
 		case Opcode::BUFFER_ATOMIC_AND:
@@ -697,6 +682,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::FLAT_LOAD_USHORT:
 		case Opcode::FLAT_LOAD_SSHORT:
 		case Opcode::FLAT_LOAD_SHORT_D16:
+		case Opcode::FLAT_LOAD_SHORT_D16_HI:
 		case Opcode::FLAT_LOAD_DWORD:
 		case Opcode::FLAT_LOAD_DWORDX2:
 		case Opcode::FLAT_LOAD_DWORDX3:

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

@@ -544,6 +544,7 @@ enum class Opcode {
 	S_BUFFER_LOAD_DWORDX8,
 	S_BUFFER_LOAD_DWORDX16,
 	S_MEMREALTIME,
+	S_DCACHE_INV,
 	BUFFER_LOAD_FORMAT_X,
 	BUFFER_LOAD_FORMAT_XY,
 	BUFFER_LOAD_FORMAT_XYZ,
@@ -596,6 +597,7 @@ enum class Opcode {
 	BUFFER_ATOMIC_SUB,
 	BUFFER_ATOMIC_SMIN,
 	BUFFER_ATOMIC_UMIN,
+	BUFFER_ATOMIC_UMIN_X2,
 	BUFFER_ATOMIC_SMAX,
 	BUFFER_ATOMIC_UMAX,
 	BUFFER_ATOMIC_AND,
@@ -612,6 +614,7 @@ enum class Opcode {
 	FLAT_LOAD_USHORT,
 	FLAT_LOAD_SSHORT,
 	FLAT_LOAD_SHORT_D16,
+	FLAT_LOAD_SHORT_D16_HI,
 	FLAT_LOAD_DWORD,
 	FLAT_LOAD_DWORDX2,
 	FLAT_LOAD_DWORDX3,
@@ -707,15 +710,7 @@ enum class Opcode {
 	IMAGE_ATOMIC_FMIN,
 	IMAGE_ATOMIC_FMAX,
 	IMAGE_SAMPLE,
-	IMAGE_GATHER4_L,
-	IMAGE_GATHER4_LZ,
-	IMAGE_GATHER4_C,
-	IMAGE_GATHER4_C_L,
-	IMAGE_GATHER4_C_LZ,
-	IMAGE_GATHER4_LZ_O,
-	IMAGE_GATHER4_C_O,
-	IMAGE_GATHER4_C_LZ_O,
-	IMAGE_GATHER4H,
+	IMAGE_GATHER4,
 	V_INTERP_P1_F32,
 	V_INTERP_P2_F32,
 	V_INTERP_MOV_F32,
@@ -766,7 +761,9 @@ enum class OperandKind {
 	M0,
 	PopsExitingWaveId,
 	SharedBase,
+	SharedLimit,
 	PrivateBase,
+	PrivateLimit,
 	Null,
 	Vgpr,
 };

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

@@ -5,30 +5,20 @@
 
 namespace Libs::Graphics::ShaderRecompiler::Frontend {
 
-void Translator::EmitCompareResult(const Decoder::Instruction& inst, IR::U1 value, bool scalar,
-                                   bool cmpx) {
+void Translator::EmitCompareResult(const Decoder::Instruction& inst, IR::U1 value, bool scalar) {
 	if (scalar) {
 		ir.SetScc(value);
 		return;
 	}
-	const auto masked = ir.LogicalAnd(ir.GetExec(), value);
-	if (cmpx) {
-		const auto mask = BallotMask(masked);
-		ir.SetExec(masked);
-		ir.SetExecLo(mask[0]);
-		ir.SetExecHi(mask[1]);
-		return;
-	}
-	WriteMask(inst.dst, masked);
+	WriteVectorMask(inst.dst, value);
 }
 
-void Translator::EmitCompareConstant(const Decoder::Instruction& inst, bool value, bool scalar,
-                                     bool cmpx) {
-	EmitCompareResult(inst, IR::U1(IR::Value(value)), scalar, cmpx);
+void Translator::EmitCompareConstant(const Decoder::Instruction& inst, bool value, bool scalar) {
+	EmitCompareResult(inst, IR::U1(IR::Value(value)), scalar);
 }
 
 void Translator::EmitIntegerCompare(const Decoder::Instruction& inst, IR::ValueOpcode opcode,
-                                    IR::Type type, bool scalar, bool cmpx, bool swap_operands) {
+                                    IR::Type type, bool scalar, bool swap_operands) {
 	// The equality compares share their IR opcode with the unsigned ones.
 	const bool signed_64 = opcode == IR::ValueOpcode::SLessThan64 ||
 	                       opcode == IR::ValueOpcode::SLessThanEqual64 ||
@@ -49,18 +39,18 @@ void Translator::EmitIntegerCompare(const Decoder::Instruction& inst, IR::ValueO
 	if (swap_operands) {
 		std::swap(lhs, rhs);
 	}
-	EmitCompareResult(inst, IR::U1(ir.Emit(opcode, {lhs, rhs})), scalar, cmpx);
+	EmitCompareResult(inst, IR::U1(ir.Emit(opcode, {lhs, rhs})), scalar);
 }
 
 void Translator::EmitInteger16Compare(const Decoder::Instruction& inst, IR::ValueOpcode opcode,
-                                      bool signed_value, bool cmpx) {
+                                      bool signed_value) {
 	const auto lhs = ReadU16AsU32(inst.src0, signed_value);
 	const auto rhs = ReadU16AsU32(inst.src1, signed_value);
-	EmitCompareResult(inst, IR::U1(ir.Emit(opcode, {lhs, rhs})), false, cmpx);
+	EmitCompareResult(inst, IR::U1(ir.Emit(opcode, {lhs, rhs})), false);
 }
 
 void Translator::EmitFloatCompare(const Decoder::Instruction& inst, IR::ValueOpcode opcode,
-                                  bool half, bool cmpx) {
+                                  bool half) {
 	const auto type = IR::ArgTypeOf(opcode, 0);
 	const auto lhs =
 	    half ? IR::Value(ReadF16AsF32(inst.src0)) : ReadOperand(inst.src0, type);
@@ -68,10 +58,10 @@ void Translator::EmitFloatCompare(const Decoder::Instruction& inst, IR::ValueOpc
 	    half ? IR::Value(ReadF16AsF32(inst.src1)) : ReadOperand(inst.src1, type);
 	const IR::FPCompareFlags flags{
 	    .flush_input_denorms = !half && type == IR::Type::F32 && flush_f32_inputs};
-	EmitCompareResult(inst, IR::U1(ir.Emit(opcode, {lhs, rhs}, flags)), false, cmpx);
+	EmitCompareResult(inst, IR::U1(ir.Emit(opcode, {lhs, rhs}, flags)), false);
 }
 
-void Translator::EmitFloatOrderedCompare(const Decoder::Instruction& inst, bool ordered, bool cmpx,
+void Translator::EmitFloatOrderedCompare(const Decoder::Instruction& inst, bool ordered,
                                          bool is64) {
 	const auto type      = is64 ? IR::Type::F64 : IR::Type::F32;
 	const auto is_nan_op = is64 ? IR::ValueOpcode::FPIsNan64 : IR::ValueOpcode::FPIsNan32;
@@ -79,16 +69,16 @@ void Translator::EmitFloatOrderedCompare(const Decoder::Instruction& inst, bool
 	const auto rhs       = ReadOperand(inst.src1, type);
 	const auto unordered = ir.LogicalOr(IR::U1(ir.Emit(is_nan_op, {lhs})),
 	                                    IR::U1(ir.Emit(is_nan_op, {rhs})));
-	EmitCompareResult(inst, ordered ? ir.LogicalNot(unordered) : unordered, false, cmpx);
+	EmitCompareResult(inst, ordered ? ir.LogicalNot(unordered) : unordered, false);
 }
 
-void Translator::EmitFloatClassCompare(const Decoder::Instruction& inst, bool cmpx, bool half) {
+void Translator::EmitFloatClassCompare(const Decoder::Instruction& inst, bool half) {
 	const auto value = half ? IR::Value(Read16LaneBits(inst.src0, false))
 	                        : ReadOperand(inst.src0, IR::Type::F32);
 	const auto mask = half ? IR::Value(Read16LaneBits(inst.src1, false))
 	                      : ReadOperand(inst.src1, IR::Type::U32);
 	const auto opcode = half ? IR::ValueOpcode::FPCmpClass16 : IR::ValueOpcode::FPCmpClass32;
-	EmitCompareResult(inst, IR::U1(ir.Emit(opcode, {value, mask})), false, cmpx);
+	EmitCompareResult(inst, IR::U1(ir.Emit(opcode, {value, mask})), false);
 }
 
 } // namespace Libs::Graphics::ShaderRecompiler::Frontend

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

@@ -51,33 +51,33 @@ void Translator::S_SUBVECTOR_LOOP(const Decoder::Instruction& inst, bool begin)
 
 void Translator::S_SAVEEXEC(const Decoder::Instruction& inst, IR::ValueOpcode operation,
                             bool negate_exec, bool negate_source, bool write_64) {
-	if (!write_64) {
-		// Read the encoded scalar word and preserve EXEC_HI, including in wave32.
-		const auto old = ir.GetExecLo();
-		const auto src = ReadU32(inst.src0);
+	const auto combine = [&](IR::U32 old, IR::U32 src) {
 		const auto lhs = negate_exec ? ir.BitwiseNot(old) : old;
 		const auto rhs = negate_source ? ir.BitwiseNot(src) : src;
-		IR::U32 result;
-		switch (operation) {
-			case IR::ValueOpcode::LogicalAnd: result = ir.BitwiseAnd(lhs, rhs); break;
-			case IR::ValueOpcode::LogicalOr: result = ir.BitwiseOr(lhs, rhs); break;
-			default: EXIT("unsupported SAVEEXEC operation");
-		}
+		return operation == IR::ValueOpcode::LogicalAnd ? ir.BitwiseAnd(lhs, rhs)
+		                                                : ir.BitwiseOr(lhs, rhs);
+	};
+	if (!write_64) {
+		// Read the encoded scalar word and preserve EXEC_HI, including in wave32.
+		const auto old    = ir.GetExecLo();
+		const auto result = combine(old, ReadU32(inst.src0));
 		WriteRawU32(inst.dst, old);
 		WriteRawU32(ConditionOperand(Decoder::OperandKind::ExecLo), result);
 		ir.SetScc(ir.INotEqual(result, IR::U32(IR::Value(0u))));
 		return;
 	}
-	const auto old    = ir.GetExec();
-	const auto src    = ReadMask(inst.src0);
-	const auto lhs    = negate_exec ? ir.LogicalNot(old) : old;
-	const auto rhs    = negate_source ? ir.LogicalNot(src) : src;
-	const auto result = IR::U1(ir.Emit(operation, {lhs, rhs}));
-	WriteMask(inst.dst, old, true);
-	const auto mask = BallotMask(result);
-	ir.SetExec(result);
-	ir.SetExecLo(mask[0]);
-	ir.SetExecHi(mask[1]);
+	// EXEC words combine bitwise; the per-lane EXEC combines the per-lane masks, so helper lanes
+	// stay enabled.
+	const auto old      = ir.GetExec();
+	const auto src      = ReadMask(inst.src0);
+	const auto lhs      = negate_exec ? ir.LogicalNot(old) : old;
+	const auto rhs      = negate_source ? ir.LogicalNot(src) : src;
+	const auto src_mask = ReadU32Pair(inst.src0);
+	const std::array old_mask {ir.GetExecLo(), ir.GetExecHi()};
+	const std::array mask {combine(old_mask[0], src_mask[0]), combine(old_mask[1], src_mask[1])};
+	WriteMask(inst.dst, old, old_mask, true);
+	WriteMask(ConditionOperand(Decoder::OperandKind::ExecLo), IR::U1(ir.Emit(operation, {lhs, rhs})),
+	          mask);
 	ir.SetScc(ir.INotEqual(ir.BitwiseOr(mask[0], mask[1]), IR::U32(IR::Value(0u))));
 }
 
@@ -95,7 +95,7 @@ void Translator::ADD_U32(const Decoder::Instruction& inst, bool vector, bool use
 	}
 	WriteOperand(DestinationOperand(inst), result);
 	if (vector) {
-		WriteMask(inst.dst2, ir.LogicalAnd(ir.GetExec(), carry));
+		WriteVectorMask(inst.dst2, carry);
 	} else {
 		ir.SetScc(carry);
 	}
@@ -108,7 +108,7 @@ void Translator::SUB_U32(const Decoder::Instruction& inst, bool vector, bool rev
 	const auto borrow = ir.UGreaterThan(rhs, lhs);
 	WriteOperand(DestinationOperand(inst), result);
 	if (vector) {
-		WriteMask(inst.dst2, ir.LogicalAnd(ir.GetExec(), borrow));
+		WriteVectorMask(inst.dst2, borrow);
 	} else {
 		ir.SetScc(borrow);
 	}
@@ -126,7 +126,7 @@ void Translator::SUBB_U32(const Decoder::Instruction& inst, bool vector, bool re
 	const auto borrow    = ir.LogicalOr(borrow0, borrow1);
 	WriteOperand(DestinationOperand(inst), result);
 	if (vector) {
-		WriteMask(inst.dst2, ir.LogicalAnd(ir.GetExec(), borrow));
+		WriteVectorMask(inst.dst2, borrow);
 	} else {
 		ir.SetScc(borrow);
 	}
@@ -465,9 +465,13 @@ void Translator::V_PERMLANE16_B32(const Decoder::Instruction& inst, bool x16) {
 	    .fetch_inactive = inst.dst.op_sel,
 	    .bound_control  = inst.dst.op_sel_hi,
 	};
-	const auto result =
-	    ir.Emit(IR::ValueOpcode::Permlane16U32,
-	            {ReadU32(inst.src0), ReadU32(inst.src1), ReadU32(inst.src2), ir.GetExec()}, flags);
+	// An EXEC-disabled source lane without FI disables the write unless BOUND_CTRL reads zero.
+	const auto inactive = flags.fetch_inactive || flags.bound_control
+	                          ? IR::U32(IR::Value(0u))
+	                          : ir.GetVectorReg(static_cast<IR::VectorReg>(inst.dst.reg));
+	const auto result   = ir.Emit(
+	    IR::ValueOpcode::Permlane16U32,
+	    {ReadU32(inst.src0), ReadU32(inst.src1), ReadU32(inst.src2), ir.GetExec(), inactive}, flags);
 	auto dst      = DestinationOperand(inst);
 	dst.op_sel    = false;
 	dst.op_sel_hi = false;

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

@@ -49,18 +49,11 @@ void Translator::V_CVT_F16_16(const Decoder::Instruction& inst, bool signed_valu
 }
 
 void Translator::V_CVT_16_F16(const Decoder::Instruction& inst, bool signed_value) {
-	const auto value = ReadF16AsF32(inst.src0);
 	// The 32-bit conversion already saturates and maps NaN to zero.
-	if (signed_value) {
-		const auto converted = ir.Emit(IR::ValueOpcode::ConvertS32F32, {value});
-		const auto low = ir.Emit(IR::ValueOpcode::SMax32, {converted, IR::Value(0xffff8000u)});
-		Write16Bits(DestinationOperand(inst),
-		            IR::U32(ir.Emit(IR::ValueOpcode::SMin32, {low, IR::Value(0x7fffu)})));
-		return;
-	}
-	const auto converted = ir.Emit(IR::ValueOpcode::ConvertU32F32, {value});
-	Write16Bits(DestinationOperand(inst),
-	            IR::U32(ir.Emit(IR::ValueOpcode::UMin32, {converted, IR::Value(0xffffu)})));
+	const auto converted = IR::U32(
+	    ir.Emit(signed_value ? IR::ValueOpcode::ConvertS32F32 : IR::ValueOpcode::ConvertU32F32,
+	            {ReadF16AsF32(inst.src0)}));
+	Write16Bits(DestinationOperand(inst), Saturate16(converted, signed_value));
 }
 
 void Translator::V_CVT_RPI_I32_F32(const Decoder::Instruction& inst) {
@@ -130,14 +123,30 @@ void Translator::V_CVT_PK_U8_F32(const Decoder::Instruction& inst) {
 	                     {ReadU32(SourceAt(inst, 2)), byte, shift, IR::Value(8u)}));
 }
 
+void Translator::V_CVT_PK_16_32(const Decoder::Instruction& inst, bool signed_value) {
+	WriteOperand(DestinationOperand(inst),
+	             PackU16Lanes(Saturate16(ReadU32(inst.src0), signed_value),
+	                          Saturate16(ReadU32(inst.src1), signed_value)));
+}
+
 void Translator::V_PACK_B32_F16(const Decoder::Instruction& inst) {
 	WriteOperand(DestinationOperand(inst),
 	             PackU16Lanes(Read16LaneBits(inst.src0, false), Read16LaneBits(inst.src1, false)));
 }
 
 IR::U32 Translator::PackU16Lanes(IR::U32 low, IR::U32 high) {
+	// A zero high lane zero-extends; descriptor tracking bounds a mask, not a bitfield insert.
+	if (high == IR::Value(0u)) return ir.BitwiseAnd(low, IR::U32(IR::Value(0xffffu)));
 	return IR::U32(
 	    ir.Emit(IR::ValueOpcode::BitFieldInsert, {low, high, IR::Value(16u), IR::Value(16u)}));
 }
 
+IR::U32 Translator::Saturate16(IR::U32 value, bool signed_value) {
+	if (!signed_value) {
+		return IR::U32(ir.Emit(IR::ValueOpcode::UMin32, {value, IR::Value(0xffffu)}));
+	}
+	const auto low = ir.Emit(IR::ValueOpcode::SMax32, {value, IR::Value(0xffff8000u)});
+	return IR::U32(ir.Emit(IR::ValueOpcode::SMin32, {low, IR::Value(0x7fffu)}));
+}
+
 } // namespace Libs::Graphics::ShaderRecompiler::Frontend

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

@@ -1,6 +1,7 @@
 #include "graphics/shader/recompiler/frontend/translate/Translator.h"
 
 #include <array>
+#include <bit>
 #include <utility>
 
 namespace Libs::Graphics::ShaderRecompiler::Frontend {
@@ -118,6 +119,33 @@ void Translator::FloatUnary(const Decoder::Instruction& inst, IR::ValueOpcode op
 	WriteOperand(DestinationOperand(inst), ir.Emit(opcode, {ReadOperand(inst.src0, type)}));
 }
 
+void Translator::V_FRACT(const Decoder::Instruction& inst, IR::Type type) {
+	// x - floor(x) rounds to 1.0 for a tiny negative x; v_fract returns the largest value below
+	// one instead. The ordered compare lets NaN through.
+	const auto destination = DestinationOperand(inst);
+	if (type == IR::Type::F64) {
+		const auto result    = ir.Emit(IR::ValueOpcode::FPFract64, {ReadOperand(inst.src0, type)});
+		const auto below_one = ir.Emit(IR::ValueOpcode::BitCastF64U64,
+		                               {IR::Value(uint64_t {0x3fefffffffffffff})});
+		const auto above     = ir.Emit(IR::ValueOpcode::FPOrdGreaterThan64, {result, below_one});
+		WriteOperand(destination, ir.Emit(IR::ValueOpcode::SelectF64, {above, below_one, result}));
+		return;
+	}
+	const bool half   = type == IR::Type::F16;
+	const auto source = half ? ReadF16AsF32(inst.src0) : IR::F32(ReadOperand(inst.src0, type));
+	const auto result = IR::F32(ir.Emit(IR::ValueOpcode::FPFract32, {source}));
+	// Clamp in the destination precision: 0x3f7fe000 is the half 0x3bff.
+	const auto below_one =
+	    IR::F32(IR::Value::F32(std::bit_cast<float>(half ? 0x3f7fe000u : 0x3f7fffffu)));
+	const auto above   = IR::U1(ir.Emit(IR::ValueOpcode::FPOrdGreaterThan32, {result, below_one}));
+	const auto clamped = SelectF32(above, below_one, result);
+	if (half) {
+		WriteF16(destination, clamped);
+	} else {
+		WriteOperand(destination, clamped);
+	}
+}
+
 void Translator::FloatBinary(const Decoder::Instruction& inst, IR::ValueOpcode opcode,
                              bool reverse) {
 	std::array<IR::Value, 2> args;

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

@@ -25,8 +25,7 @@ void Translator::V_MAD_I16(const Decoder::Instruction& inst) {
 	const auto rhs    = ReadU16AsU32(inst.src1, true);
 	auto       result = ir.IAdd(ir.IMul(lhs, rhs), ReadU16AsU32(inst.src2, true));
 	if (inst.dst.clamp) {
-		result = IR::U32(ir.Emit(IR::ValueOpcode::SMax32, {result, IR::Value(0xffff8000u)}));
-		result = IR::U32(ir.Emit(IR::ValueOpcode::SMin32, {result, IR::Value(0x7fffu)}));
+		result = Saturate16(result, true);
 	}
 	Write16Bits(DestinationOperand(inst), result);
 }
@@ -108,19 +107,7 @@ void Translator::S_U64_MASK(const Decoder::Instruction& inst, IR::ValueOpcode lo
 			default: return false;
 		}
 	};
-	if (is_exec_or_vcc(inst.dst) || is_exec_or_vcc(inst.src0) ||
-	    (inst.src_count > 1u && is_exec_or_vcc(inst.src1))) {
-		const auto mask = WriteMask(inst.dst, invocation_result, true);
-		ir.SetScc(ir.INotEqual(ir.BitwiseOr(mask[0], mask[1]), IR::U32(IR::Value(0u))));
-		return;
-	}
-
-	// Descriptor tracking reasons about each scalar word independently.
-	const auto lhs        = ReadU32Pair(inst.src0);
-	auto       mask_valid = ReadMaskValid(inst.src0);
-	if (inst.src_count > 1u) {
-		mask_valid = ir.LogicalAnd(mask_valid, ReadMaskValid(inst.src1));
-	}
+	const auto lhs = ReadU32Pair(inst.src0);
 	std::array<IR::U32, 2> result;
 	if (unary) {
 		result = {ir.BitwiseNot(lhs[0]), ir.BitwiseNot(lhs[1])};
@@ -133,11 +120,21 @@ void Translator::S_U64_MASK(const Decoder::Instruction& inst, IR::ValueOpcode lo
 			result[component] = negate_result ? ir.BitwiseNot(IR::U32(value)) : IR::U32(value);
 		}
 	}
-	WriteU32Pair(inst.dst, result);
-	if (inst.dst.kind == Decoder::OperandKind::Sgpr) {
-		const auto dst = static_cast<IR::ScalarReg>(inst.dst.reg);
-		ir.SetThreadBitScalarReg(dst, invocation_result);
-		ir.SetScalarMaskTag(dst, mask_valid);
+	if (is_exec_or_vcc(inst.dst) || is_exec_or_vcc(inst.src0) ||
+	    (inst.src_count > 1u && is_exec_or_vcc(inst.src1))) {
+		WriteMask(inst.dst, invocation_result, result, true);
+	} else {
+		// Descriptor tracking reasons about each scalar word independently.
+		auto mask_valid = ReadMaskValid(inst.src0);
+		if (inst.src_count > 1u) {
+			mask_valid = ir.LogicalAnd(mask_valid, ReadMaskValid(inst.src1));
+		}
+		WriteU32Pair(inst.dst, result);
+		if (inst.dst.kind == Decoder::OperandKind::Sgpr) {
+			const auto dst = static_cast<IR::ScalarReg>(inst.dst.reg);
+			ir.SetThreadBitScalarReg(dst, invocation_result);
+			ir.SetScalarMaskTag(dst, mask_valid);
+		}
 	}
 	ir.SetScc(ir.INotEqual(ir.BitwiseOr(result[0], result[1]), IR::U32(IR::Value(0u))));
 }
@@ -293,7 +290,8 @@ void Translator::V_MAD_U64_U32(const Decoder::Instruction& inst) {
 	if (inst.dst2.kind != Decoder::OperandKind::Null &&
 	    inst.dst2.kind != Decoder::OperandKind::Unknown) {
 		// The 32x32 product fits in 64 bits, so only the addition can carry.
-		WriteMask(inst.dst2, IR::U1(ir.Emit(IR::ValueOpcode::ULessThan64, {result, product})));
+		const auto carry = IR::U1(ir.Emit(IR::ValueOpcode::ULessThan64, {result, product}));
+		WriteVectorMask(inst.dst2, carry);
 	}
 }
 
@@ -473,7 +471,7 @@ void Translator::S_BITCMP_B32(const Decoder::Instruction& inst, bool expected) {
 	const auto offset = ir.BitwiseAnd(ReadU32(inst.src1), IR::U32(IR::Value(31u)));
 	const auto bit =
 	    IR::U32(ir.Emit(IR::ValueOpcode::BitFieldUExtract, {value, offset, IR::Value(1u)}));
-	WriteCompareResult(inst.dst, ir.IEqual(bit, IR::U32(IR::Value(expected ? 1u : 0u))));
+	ir.SetScc(ir.IEqual(bit, IR::U32(IR::Value(expected ? 1u : 0u))));
 }
 
 void Translator::S_BITCMP_B64(const Decoder::Instruction& inst, bool expected) {
@@ -484,7 +482,7 @@ void Translator::S_BITCMP_B64(const Decoder::Instruction& inst, bool expected) {
 	    ir.Select(ir.ULessThan(offset, IR::U32(IR::Value(32u))), value[0], value[1]);
 	const auto bit =
 	    IR::U32(ir.Emit(IR::ValueOpcode::BitFieldUExtract, {word, word_bit, IR::Value(1u)}));
-	WriteCompareResult(inst.dst, ir.IEqual(bit, IR::U32(IR::Value(expected ? 1u : 0u))));
+	ir.SetScc(ir.IEqual(bit, IR::U32(IR::Value(expected ? 1u : 0u))));
 }
 
 void Translator::V_ALIGN_B32(const Decoder::Instruction& inst, bool byte_offset) {

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

@@ -322,8 +322,8 @@ IR::Value Translator::MakeImageAddress(const Decoder::Instruction& inst,
                                        const Decoder::Operand&     base) {
 	std::array<IR::Value, 13> components {};
 	components.fill(IR::Value(0u));
-	const auto count =
-	    Decoder::ImageAddressDwordCount(inst.image_sample_flags, inst.image_address_components);
+	const auto count = Decoder::ImageAddressDwordCount(
+	    inst.image_sample_flags, inst.image_dimension, inst.image_address_components);
 	EXIT_IF(count > components.size());
 	const auto nsa_components =
 	    std::min(inst.image_nsa_dwords * 4u, Decoder::MaxImageNsaAddressComponents);
@@ -863,9 +863,10 @@ void Translator::DS_WRITE2(const Decoder::Instruction& inst) {
 
 void Translator::DS_APPEND_CONSUME(const Decoder::Instruction& inst, IR::ValueOpcode opcode) {
 	const auto memory = MemoryInfoFromDecoded(inst);
-	WriteOperand(inst.dst, ir.Emit(opcode,
-	                               {ReadU32(MemorySourceAt(inst, 0)), ir.GetExec()},
-	                               AddMemoryInfo(memory, inst.pc)));
+	const auto exec   = ExecMask();
+	WriteOperand(inst.dst,
+	             ir.Emit(opcode, {ReadU32(MemorySourceAt(inst, 0)), exec[0], exec[1], ir.GetExec()},
+	                     AddMemoryInfo(memory, inst.pc)));
 }
 
 void Translator::DS_ADDTID(const Decoder::Instruction& inst, bool write) {
@@ -901,10 +902,7 @@ void Translator::DS_ORDERED_COUNT(const Decoder::Instruction& inst) {
 		const auto packer = ir.BitwiseAnd(m0, IR::U32(IR::Value(0xfu)));
 		address = ir.IAdd(address, ir.ShiftLeftLogical(packer, IR::U32(IR::Value(2u))));
 	}
-	const auto exec   = ir.GetExec();
-	const auto ballot = ir.Emit(IR::ValueOpcode::Ballot, {exec});
-	const auto low = ir.CompositeExtract(ballot, 0u);
-	const auto high = ir.CompositeExtract(ballot, 1u);
+	const auto [low, high] = ExecMask();
 	const auto active = ir.INotEqual(ir.BitwiseOr(low, high), IR::U32(IR::Value(0u)));
 	const auto first = ir.Select(
 	    ir.INotEqual(low, IR::U32(IR::Value(0u))),
@@ -943,6 +941,8 @@ void Translator::DS_PERMUTE(const Decoder::Instruction& inst, bool backward) {
 
 void Translator::EmitMemory(const Decoder::Instruction& inst) {
 	switch (inst.opcode) {
+		// Translated scalar loads have no cache of their own to invalidate.
+		case Decoder::Opcode::S_DCACHE_INV: return;
 		case Decoder::Opcode::BUFFER_GL0_INV:
 		case Decoder::Opcode::BUFFER_GL1_INV:
 			ir.Emit(IR::ValueOpcode::DataCacheInvalidate);
@@ -1031,6 +1031,8 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 			return BUFFER_ATOMIC(inst, IR::ValueOpcode::BufferAtomicSMin32);
 		case Decoder::Opcode::BUFFER_ATOMIC_UMIN:
 			return BUFFER_ATOMIC(inst, IR::ValueOpcode::BufferAtomicUMin32);
+		case Decoder::Opcode::BUFFER_ATOMIC_UMIN_X2:
+			return BUFFER_ATOMIC(inst, IR::ValueOpcode::BufferAtomicUMin64);
 		case Decoder::Opcode::BUFFER_ATOMIC_SMAX:
 			return BUFFER_ATOMIC(inst, IR::ValueOpcode::BufferAtomicSMax32);
 		case Decoder::Opcode::BUFFER_ATOMIC_UMAX:
@@ -1148,6 +1150,7 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::FLAT_LOAD_USHORT:
 		case Decoder::Opcode::FLAT_LOAD_SSHORT:
 		case Decoder::Opcode::FLAT_LOAD_SHORT_D16:
+		case Decoder::Opcode::FLAT_LOAD_SHORT_D16_HI:
 		case Decoder::Opcode::FLAT_LOAD_DWORD:
 		case Decoder::Opcode::FLAT_LOAD_DWORDX2:
 		case Decoder::Opcode::FLAT_LOAD_DWORDX3:
@@ -1167,15 +1170,7 @@ void 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_L:
-		case Decoder::Opcode::IMAGE_GATHER4_C_LZ:
-		case Decoder::Opcode::IMAGE_GATHER4_LZ_O:
-		case Decoder::Opcode::IMAGE_GATHER4_C_O:
-		case Decoder::Opcode::IMAGE_GATHER4_C_LZ_O:
-		case Decoder::Opcode::IMAGE_GATHER4H: return IMAGE_GATHER(inst);
+		case Decoder::Opcode::IMAGE_GATHER4: return IMAGE_GATHER(inst);
 
 		case Decoder::Opcode::DS_MIN_F32:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicFMin32, false);

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

@@ -73,43 +73,41 @@ void Translator::EmitScalar(const Decoder::Instruction& inst) {
 
 		case O::S_CMP_EQ_U32:
 		case O::S_CMP_EQ_I32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::IEqual32, IR::Type::U32, true, false);
+			EmitIntegerCompare(inst, IR::ValueOpcode::IEqual32, IR::Type::U32, true);
 			return;
 		case O::S_CMP_LG_U32:
 		case O::S_CMP_LG_I32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::INotEqual32, IR::Type::U32, true, false);
+			EmitIntegerCompare(inst, IR::ValueOpcode::INotEqual32, IR::Type::U32, true);
 			return;
 		case O::S_CMP_GT_U32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::UGreaterThan32, IR::Type::U32, true, false);
+			EmitIntegerCompare(inst, IR::ValueOpcode::UGreaterThan32, IR::Type::U32, true);
 			return;
 		case O::S_CMP_GE_U32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::UGreaterThanEqual32, IR::Type::U32, true,
-			                   false);
+			EmitIntegerCompare(inst, IR::ValueOpcode::UGreaterThanEqual32, IR::Type::U32, true);
 			return;
 		case O::S_CMP_LT_U32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThan32, IR::Type::U32, true, false);
+			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThan32, IR::Type::U32, true);
 			return;
 		case O::S_CMP_LE_U32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThanEqual32, IR::Type::U32, true, false);
+			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThanEqual32, IR::Type::U32, true);
 			return;
 		case O::S_CMP_GT_I32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::SGreaterThan32, IR::Type::U32, true, false);
+			EmitIntegerCompare(inst, IR::ValueOpcode::SGreaterThan32, IR::Type::U32, true);
 			return;
 		case O::S_CMP_GE_I32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::SGreaterThanEqual32, IR::Type::U32, true,
-			                   false);
+			EmitIntegerCompare(inst, IR::ValueOpcode::SGreaterThanEqual32, IR::Type::U32, true);
 			return;
 		case O::S_CMP_LT_I32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThan32, IR::Type::U32, true, false);
+			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThan32, IR::Type::U32, true);
 			return;
 		case O::S_CMP_LE_I32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThanEqual32, IR::Type::U32, true, false);
+			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThanEqual32, IR::Type::U32, true);
 			return;
 		case O::S_CMP_EQ_U64:
-			EmitIntegerCompare(inst, IR::ValueOpcode::IEqual64, IR::Type::U64, true, false);
+			EmitIntegerCompare(inst, IR::ValueOpcode::IEqual64, IR::Type::U64, true);
 			return;
 		case O::S_CMP_LG_U64:
-			EmitIntegerCompare(inst, IR::ValueOpcode::INotEqual64, IR::Type::U64, true, false);
+			EmitIntegerCompare(inst, IR::ValueOpcode::INotEqual64, IR::Type::U64, true);
 			return;
 
 		case O::S_AND_B64:

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

@@ -129,6 +129,20 @@ std::array<IR::U32, 2> Translator::BallotMask(IR::U1 value) {
 	        program.wave_size == 64u ? ir.CompositeExtract(mask, 1) : IR::U32(IR::Value(0u))};
 }
 
+std::array<IR::U32, 2> Translator::ExecMask() {
+	return {ir.GetExecLo(), program.wave_size == 64u ? ir.GetExecHi() : IR::U32(IR::Value(0u))};
+}
+
+// Pixel EXEC words exclude the helper invocations that the per-lane EXEC keeps enabled.
+std::array<IR::U32, 2> Translator::VectorMask(IR::U1 value) {
+	auto mask = BallotMask(value);
+	if (program.stage == ShaderType::Pixel) {
+		const auto exec = ExecMask();
+		mask = {ir.BitwiseAnd(mask[0], exec[0]), ir.BitwiseAnd(mask[1], exec[1])};
+	}
+	return mask;
+}
+
 IR::U32 Translator::ReadRawU32(const Decoder::Operand& operand) {
 	switch (operand.kind) {
 		case Decoder::OperandKind::LiteralConstant:
@@ -138,6 +152,8 @@ IR::U32 Translator::ReadRawU32(const Decoder::Operand& operand) {
 		case Decoder::OperandKind::SharedBase:
 		case Decoder::OperandKind::PrivateBase:
 		case Decoder::OperandKind::PopsExitingWaveId: return IR::U32(IR::Value(0u));
+		case Decoder::OperandKind::SharedLimit:
+		case Decoder::OperandKind::PrivateLimit: return IR::U32(IR::Value(0xffffffffu));
 		case Decoder::OperandKind::Sgpr:
 			return ir.GetScalarReg(static_cast<IR::ScalarReg>(operand.reg));
 		case Decoder::OperandKind::Vgpr:
@@ -388,21 +404,9 @@ void Translator::WriteOperand(const Decoder::Operand& operand, IR::Value value)
 		switch (operand.kind) {
 			case Decoder::OperandKind::Scc: ir.SetScc(IR::U1(value)); return;
 			case Decoder::OperandKind::ExecLo:
-			case Decoder::OperandKind::ExecHi: {
-				const auto mask = BallotMask(IR::U1(value));
-				ir.SetExec(IR::U1(value));
-				ir.SetExecLo(mask[0]);
-				ir.SetExecHi(mask[1]);
-				return;
-			}
+			case Decoder::OperandKind::ExecHi:
 			case Decoder::OperandKind::VccLo:
-			case Decoder::OperandKind::VccHi: {
-				const auto mask = BallotMask(IR::U1(value));
-				ir.SetVcc(IR::U1(value));
-				ir.SetVccLo(mask[0]);
-				ir.SetVccHi(mask[1]);
-				return;
-			}
+			case Decoder::OperandKind::VccHi: EXIT("lane masks are written through WriteMask");
 			default:
 				WriteRawU32(operand, ir.Select(IR::U1(value), IR::U32(IR::Value(1u)),
 				                               IR::U32(IR::Value(0u))));
@@ -471,17 +475,17 @@ IR::U32 Translator::ReadU32(const Decoder::Operand& operand) {
 }
 
 std::array<IR::U32, 2> Translator::ReadU32Pair(const Decoder::Operand& operand) {
-	if (operand.kind == Decoder::OperandKind::PrivateBase ||
-	    operand.kind == Decoder::OperandKind::SharedBase) {
-		return {IR::U32(IR::Value(0u)), IR::U32(IR::Value(
-		    operand.kind == Decoder::OperandKind::PrivateBase ? Decoder::PrivateApertureHigh
-		                                                      : Decoder::SharedApertureHigh))};
-	}
-	if (operand.kind == Decoder::OperandKind::ExecLo) {
-		return {ir.GetExecLo(), ir.GetExecHi()};
-	}
-	if (operand.kind == Decoder::OperandKind::VccLo) {
-		return {ir.GetVccLo(), ir.GetVccHi()};
+	// Each aperture limit is its base plus 0xffffffff.
+	switch (operand.kind) {
+		case Decoder::OperandKind::SharedBase:
+		case Decoder::OperandKind::SharedLimit:
+			return {ReadRawU32(operand), IR::U32(IR::Value(Decoder::SharedApertureHigh))};
+		case Decoder::OperandKind::PrivateBase:
+		case Decoder::OperandKind::PrivateLimit:
+			return {ReadRawU32(operand), IR::U32(IR::Value(Decoder::PrivateApertureHigh))};
+		case Decoder::OperandKind::ExecLo: return {ir.GetExecLo(), ir.GetExecHi()};
+		case Decoder::OperandKind::VccLo: return {ir.GetVccLo(), ir.GetVccHi()};
+		default: break;
 	}
 	if (operand.kind == Decoder::OperandKind::FloatInlineConstant) {
 		const auto bits =
@@ -639,7 +643,7 @@ IR::U1 Translator::ReadMask(const Decoder::Operand& operand) {
 		}
 		case Decoder::OperandKind::ExecLo:
 		case Decoder::OperandKind::ExecHi: return ir.GetExec();
-		case Decoder::OperandKind::VccLo:
+		case Decoder::OperandKind::VccLo: return ir.GetVcc();
 		case Decoder::OperandKind::VccHi:
 			return program.wave_size == 32u
 			           ? ThreadBit({ReadRawU32(operand), IR::U32(IR::Value(0u))})
@@ -677,9 +681,8 @@ IR::U1 Translator::ReadMaskValid(const Decoder::Operand& operand) {
 	}
 }
 
-std::array<IR::U32, 2> Translator::WriteMask(const Decoder::Operand& operand, IR::U1 value,
-                                             bool write_64) {
-	const auto mask = BallotMask(value);
+void Translator::WriteMask(const Decoder::Operand& operand, IR::U1 value,
+                           const std::array<IR::U32, 2>& mask, bool write_64) {
 	switch (operand.kind) {
 		case Decoder::OperandKind::Sgpr: {
 			const auto reg  = static_cast<IR::ScalarReg>(operand.reg);
@@ -700,43 +703,43 @@ std::array<IR::U32, 2> Translator::WriteMask(const Decoder::Operand& operand, IR
 				ir.SetThreadBitScalarReg(high, IR::U1(IR::Value(false)));
 				ir.SetScalarMaskTag(high, IR::U1(IR::Value(false)));
 			}
-			return mask;
+			return;
 		}
 		case Decoder::OperandKind::ExecLo:
 		case Decoder::OperandKind::ExecHi: {
 			ir.SetExec(value);
 			ir.SetExecLo(mask[0]);
 			ir.SetExecHi(mask[1]);
-			return mask;
+			return;
 		}
 		case Decoder::OperandKind::VccLo:
 		case Decoder::OperandKind::VccHi: {
-			if (!write_64 && program.wave_size == 32u) {
+			if (operand.kind == Decoder::OperandKind::VccHi && !write_64 &&
+			    program.wave_size == 32u) {
 				WriteRawU32(operand, mask[0]);
-				return mask;
+				return;
 			}
 			ir.SetVcc(value);
 			ir.SetVccLo(mask[0]);
-			ir.SetVccHi(mask[1]);
-			return mask;
+			if (write_64 || program.wave_size == 64u) {
+				ir.SetVccHi(mask[1]);
+			}
+			return;
 		}
-		case Decoder::OperandKind::Scc: ir.SetScc(value); return mask;
+		case Decoder::OperandKind::Scc: ir.SetScc(value); return;
 		default:
 			WriteRawU32(operand, mask[0]);
 			WriteRawU32(OffsetOperand(operand, 1), mask[1]);
-			return mask;
+			return;
 	}
 }
 
-void Translator::WriteCompareResult(const Decoder::Operand& operand, IR::U1 value) {
-	if (operand.kind == Decoder::OperandKind::Scc) {
-		WriteOperand(operand, value);
-		return;
-	}
+void Translator::WriteVectorMask(const Decoder::Operand& operand, IR::U1 value) {
 	// VALU instructions only write lanes enabled by the incoming EXEC mask. This is
 	// especially important for V_CMPX: replacing EXEC with the ungated comparison would
 	// reactivate lanes disabled by an enclosing divergent region.
-	WriteMask(operand, ir.LogicalAnd(ir.GetExec(), value));
+	const auto active = ir.LogicalAnd(ir.GetExec(), value);
+	WriteMask(operand, active, VectorMask(active));
 }
 
 void Translator::AddBranchCondition(const CFG::Graph& graph, const CFG::BasicBlock& source,
@@ -1153,7 +1156,10 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 			}
 		}
 		entry_ir.SetExec(initial_exec);
-		const auto initial_mask = entry_ir.Emit(IR::ValueOpcode::Ballot, {initial_exec});
+		// Pixel EXEC words start with the live (non-helper) lanes; helpers keep the per-lane EXEC.
+		const auto initial_mask = options.stage == ShaderType::Pixel
+		                              ? entry_ir.Emit(IR::ValueOpcode::LiveMask)
+		                              : entry_ir.Emit(IR::ValueOpcode::Ballot, {initial_exec});
 		const auto initial_lo = entry_ir.CompositeExtract(initial_mask, 0);
 		const auto initial_hi = options.wave_size == 64u ? entry_ir.CompositeExtract(initial_mask, 1)
 		                                               : IR::U32(IR::Value(0u));
@@ -1238,12 +1244,12 @@ 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)));
+			// Instance IDs exclude the start instance, which shaders add from user data.
 			if (mesh.fast_launch) {
 				// Fast launch broadcasts the subgroup's base vertex and instance to every lane.
 				entry_ir.SetVectorReg(static_cast<IR::VectorReg>(5), entry_ir.IAdd(draw(1), group));
-				entry_ir.SetVectorReg(
-				    static_cast<IR::VectorReg>(6),
-				    entry_ir.IAdd(draw(2), builtin(IR::StageInputKind::WorkgroupId, 1)));
+				entry_ir.SetVectorReg(static_cast<IR::VectorReg>(6),
+				                      builtin(IR::StageInputKind::WorkgroupId, 1));
 			} else {
 				// GS adjacency addresses local ES records in LDS. Fans retain the draw's
 				// center in every subgroup; strip winding follows the global primitive.
@@ -1271,14 +1277,14 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 				                                         entry_ir.ShiftLeftLogical(second, u32(18))));
 				entry_ir.SetVectorReg(static_cast<IR::VectorReg>(1),
 				                      entry_ir.ShiftLeftLogical(third, u32(2)));
-				const auto index_bytes  = draw(3);
+				const auto index_bytes  = draw(2);
 				const auto indexed      = entry_ir.INotEqual(index_bytes, u32(0));
-				const auto index_low    = draw(4);
+				const auto index_low    = draw(3);
 				const auto byte_offset  = entry_ir.IAdd(entry_ir.BitwiseAnd(index_low, u32(3)),
 				                                           entry_ir.IMul(input_vertex, index_bytes));
 				const auto index_resource = entry_ir.Emit(
 				    IR::ValueOpcode::GetAddressResource,
-				    {entry_ir.BitwiseAnd(index_low, u32(~3u)), draw(5)});
+				    {entry_ir.BitwiseAnd(index_low, u32(~3u)), draw(4)});
 				const auto memory_index = static_cast<uint32_t>(result.memory_info.size());
 				result.memory_info.push_back({.kind = IR::ResourceKind::Global});
 				const auto packed_index = entry_ir.Emit(
@@ -1292,9 +1298,8 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 				     entry_ir.IMul(index_bytes, u32(8))}));
 				entry_ir.SetVectorReg(static_cast<IR::VectorReg>(5),
 				                      entry_ir.IAdd(draw(1), entry_ir.Select(indexed, index, input_vertex)));
-				entry_ir.SetVectorReg(
-				    static_cast<IR::VectorReg>(8),
-				    entry_ir.IAdd(draw(2), builtin(IR::StageInputKind::WorkgroupId, 1)));
+				entry_ir.SetVectorReg(static_cast<IR::VectorReg>(8),
+				                      builtin(IR::StageInputKind::WorkgroupId, 1));
 			}
 		} else if (options.stage == ShaderType::Local) {
 			entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(3), IR::U32(IR::Value(64u)));
@@ -1384,10 +1389,27 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 			                      IR::U32(IR::Value(options.wave_size << 12u)));
 			entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(3),
 			                      entry_ir.BitwiseOr(vertex_count, IR::U32(IR::Value(1u << 28u))));
+			// Vulkan adds an indirect record's bases to the indices; the hardware passes them
+			// only in the user SGPRs the packet patches.
+			const auto& offsets = options.input_info.vertex->indirect_offsets;
+			const auto  index   = [&](uint8_t slot, IR::StageInputKind kind,
+			                          IR::StageInputKind base_kind) {
+				if (offsets == ShaderIndirectOffsets {}) {
+					return builtin(kind);
+				}
+				const auto base = builtin(base_kind);
+				if (slot < options.user_data_count) {
+					entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(options.user_data_base + slot),
+					                      base);
+				}
+				return entry_ir.ISub(builtin(kind), base);
+			};
 			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(5),
-			                      builtin(IR::StageInputKind::VertexIndex));
+			                      index(offsets.vertex, IR::StageInputKind::VertexIndex,
+			                            IR::StageInputKind::BaseVertex));
 			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(8),
-			                      builtin(IR::StageInputKind::InstanceIndex));
+			                      index(offsets.instance, IR::StageInputKind::InstanceIndex,
+			                            IR::StageInputKind::BaseInstance));
 		}
 	}
 	const bool flush_f32_inputs = options.stage == ShaderType::Compute &&

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

@@ -27,6 +27,8 @@ private:
 	Decoder::Operand ScalarDestinationOperand(const Decoder::Operand& operand, uint32_t offset);
 	Decoder::Operand PlainOperand(const Decoder::Operand& operand);
 	std::array<IR::U32, 2> BallotMask(IR::U1 value);
+	std::array<IR::U32, 2> ExecMask();
+	std::array<IR::U32, 2> VectorMask(IR::U1 value);
 	IR::U32                ReadRawU32(const Decoder::Operand& operand);
 	IR::U32                ReadScalarCode(uint32_t code);
 	IR::U32                ApplyBitSourceModifiers(const Decoder::Operand& operand, IR::U32 value);
@@ -52,9 +54,9 @@ private:
 	IR::U32 ConditionBit(const Decoder::Operand& operand);
 	IR::U1  ReadMask(const Decoder::Operand& operand);
 	IR::U1  ReadMaskValid(const Decoder::Operand& operand);
-	std::array<IR::U32, 2> WriteMask(const Decoder::Operand& operand, IR::U1 value,
-	                                 bool write_64 = false);
-	void    WriteCompareResult(const Decoder::Operand& operand, IR::U1 value);
+	void    WriteMask(const Decoder::Operand& operand, IR::U1 value,
+	                  const std::array<IR::U32, 2>& mask, bool write_64 = false);
+	void    WriteVectorMask(const Decoder::Operand& operand, IR::U1 value);
 
 	IR::MemoryFlags AddMemoryInfo(const IR::MemoryInfo& memory, uint32_t pc);
 	IR::ExportFlags AddExportInfo(const Decoder::Instruction& inst);
@@ -114,17 +116,16 @@ private:
 
 	IR::F32 SelectF32(IR::U1 condition, IR::F32 true_value, IR::F32 false_value);
 	IR::U32 PackU16Lanes(IR::U32 low, IR::U32 high);
-	void EmitCompareResult(const Decoder::Instruction& inst, IR::U1 value, bool scalar, bool cmpx);
-	void EmitCompareConstant(const Decoder::Instruction& inst, bool value, bool scalar, bool cmpx);
+	IR::U32 Saturate16(IR::U32 value, bool signed_value);
+	void EmitCompareResult(const Decoder::Instruction& inst, IR::U1 value, bool scalar);
+	void EmitCompareConstant(const Decoder::Instruction& inst, bool value, bool scalar);
 	void EmitIntegerCompare(const Decoder::Instruction& inst, IR::ValueOpcode opcode, IR::Type type,
-	                        bool scalar, bool cmpx, bool swap_operands = false);
+	                        bool scalar, bool swap_operands = false);
 	void EmitInteger16Compare(const Decoder::Instruction& inst, IR::ValueOpcode opcode,
-	                          bool signed_value, bool cmpx);
-	void EmitFloatCompare(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool half,
-	                      bool cmpx);
-	void EmitFloatOrderedCompare(const Decoder::Instruction& inst, bool ordered, bool cmpx,
-	                             bool is64 = false);
-	void EmitFloatClassCompare(const Decoder::Instruction& inst, bool cmpx, bool half = false);
+	                          bool signed_value);
+	void EmitFloatCompare(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool half);
+	void EmitFloatOrderedCompare(const Decoder::Instruction& inst, bool ordered, bool is64 = false);
+	void EmitFloatClassCompare(const Decoder::Instruction& inst, bool half = false);
 	void V_CVT_F32_UBYTE(const Decoder::Instruction& inst, uint32_t byte_index);
 	void V_CVT_F32_U32(const Decoder::Instruction& inst);
 	void V_CVT_F32_I32(const Decoder::Instruction& inst);
@@ -141,6 +142,7 @@ private:
 	void V_CVT_PKRTZ_F16_F32(const Decoder::Instruction& inst);
 	void V_CVT_PKNORM_F32(const Decoder::Instruction& inst, IR::ValueOpcode opcode);
 	void V_CVT_PK_U8_F32(const Decoder::Instruction& inst);
+	void V_CVT_PK_16_32(const Decoder::Instruction& inst, bool signed_value);
 	void V_PACK_B32_F16(const Decoder::Instruction& inst);
 	void PackedFloat16(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool accumulator,
 	                   bool quiet_snan);
@@ -151,6 +153,7 @@ private:
 	void Float16Ternary(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool accumulator,
 	                    bool mix);
 	void FloatUnary(const Decoder::Instruction& inst, IR::ValueOpcode opcode);
+	void V_FRACT(const Decoder::Instruction& inst, IR::Type type);
 	void FloatBinary(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool reverse);
 	void FloatTernary(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool accumulator,
 	                  bool mix);

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

@@ -25,444 +25,310 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_PERMLANEX16_B32: V_PERMLANE16_B32(inst, true); return;
 
 		case O::V_CMP_F_I32:
-		case O::V_CMP_F_U32: EmitCompareConstant(inst, false, false, false); return;
+		case O::V_CMP_F_U32: EmitCompareConstant(inst, false, false); return;
 		case O::V_CMP_T_I32:
-		case O::V_CMP_T_U32: EmitCompareConstant(inst, true, false, false); return;
+		case O::V_CMP_T_U32: EmitCompareConstant(inst, true, false); return;
 		case O::V_CMP_EQ_U32:
 		case O::V_CMP_EQ_I32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::IEqual32, IR::Type::U32, false, false);
-			return;
 		case O::V_CMPX_EQ_U32:
 		case O::V_CMPX_EQ_I32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::IEqual32, IR::Type::U32, false, true);
+			EmitIntegerCompare(inst, IR::ValueOpcode::IEqual32, IR::Type::U32, false);
 			return;
 		case O::V_CMP_NE_U32:
 		case O::V_CMP_NE_I32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::INotEqual32, IR::Type::U32, false, false);
-			return;
 		case O::V_CMPX_NE_U32:
 		case O::V_CMPX_NE_I32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::INotEqual32, IR::Type::U32, false, true);
+			EmitIntegerCompare(inst, IR::ValueOpcode::INotEqual32, IR::Type::U32, false);
 			return;
 		case O::V_CMP_GT_U32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::UGreaterThan32, IR::Type::U32, false, false);
-			return;
 		case O::V_CMPX_GT_U32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::UGreaterThan32, IR::Type::U32, false, true);
+			EmitIntegerCompare(inst, IR::ValueOpcode::UGreaterThan32, IR::Type::U32, false);
 			return;
 		case O::V_CMP_GE_U32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::UGreaterThanEqual32, IR::Type::U32, false,
-			                   false);
-			return;
 		case O::V_CMPX_GE_U32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::UGreaterThanEqual32, IR::Type::U32, false,
-			                   true);
+			EmitIntegerCompare(inst, IR::ValueOpcode::UGreaterThanEqual32, IR::Type::U32, false);
 			return;
 		case O::V_CMP_LT_U32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThan32, IR::Type::U32, false, false);
-			return;
 		case O::V_CMPX_LT_U32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThan32, IR::Type::U32, false, true);
+			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThan32, IR::Type::U32, false);
 			return;
 		case O::V_CMP_LE_U32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThanEqual32, IR::Type::U32, false,
-			                   false);
-			return;
 		case O::V_CMPX_LE_U32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThanEqual32, IR::Type::U32, false, true);
+			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThanEqual32, IR::Type::U32, false);
 			return;
 		case O::V_CMP_GT_I32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::SGreaterThan32, IR::Type::U32, false, false);
-			return;
 		case O::V_CMPX_GT_I32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::SGreaterThan32, IR::Type::U32, false, true);
+			EmitIntegerCompare(inst, IR::ValueOpcode::SGreaterThan32, IR::Type::U32, false);
 			return;
 		case O::V_CMP_GE_I32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::SGreaterThanEqual32, IR::Type::U32, false,
-			                   false);
-			return;
 		case O::V_CMPX_GE_I32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::SGreaterThanEqual32, IR::Type::U32, false,
-			                   true);
+			EmitIntegerCompare(inst, IR::ValueOpcode::SGreaterThanEqual32, IR::Type::U32, false);
 			return;
 		case O::V_CMP_LT_I32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThan32, IR::Type::U32, false, false);
-			return;
 		case O::V_CMPX_LT_I32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThan32, IR::Type::U32, false, true);
+			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThan32, IR::Type::U32, false);
 			return;
 		case O::V_CMP_LE_I32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThanEqual32, IR::Type::U32, false,
-			                   false);
-			return;
 		case O::V_CMPX_LE_I32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThanEqual32, IR::Type::U32, false, true);
+			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThanEqual32, IR::Type::U32, false);
 			return;
 		case O::V_CMP_F_I64:
-		case O::V_CMP_F_U64: EmitCompareConstant(inst, false, false, false); return;
-		case O::V_CMP_T_I64:
-		case O::V_CMP_T_U64: EmitCompareConstant(inst, true, false, false); return;
+		case O::V_CMP_F_U64:
 		case O::V_CMPX_F_I64:
-		case O::V_CMPX_F_U64: EmitCompareConstant(inst, false, false, true); return;
+		case O::V_CMPX_F_U64: EmitCompareConstant(inst, false, false); return;
+		case O::V_CMP_T_I64:
+		case O::V_CMP_T_U64:
 		case O::V_CMPX_T_I64:
-		case O::V_CMPX_T_U64: EmitCompareConstant(inst, true, false, true); return;
+		case O::V_CMPX_T_U64: EmitCompareConstant(inst, true, false); return;
 		case O::V_CMP_EQ_I64:
 		case O::V_CMP_EQ_U64:
-			EmitIntegerCompare(inst, IR::ValueOpcode::IEqual64, IR::Type::U64, false, false);
-			return;
 		case O::V_CMPX_EQ_I64:
 		case O::V_CMPX_EQ_U64:
-			EmitIntegerCompare(inst, IR::ValueOpcode::IEqual64, IR::Type::U64, false, true);
+			EmitIntegerCompare(inst, IR::ValueOpcode::IEqual64, IR::Type::U64, false);
 			return;
 		case O::V_CMP_LT_U64:
-			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThan64, IR::Type::U64, false, false);
-			return;
 		case O::V_CMPX_LT_U64:
-			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThan64, IR::Type::U64, false, true);
+			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThan64, IR::Type::U64, false);
 			return;
 		case O::V_CMP_GT_U64:
-			EmitIntegerCompare(inst, IR::ValueOpcode::UGreaterThan64, IR::Type::U64, false, false);
-			return;
 		case O::V_CMPX_GT_U64:
-			EmitIntegerCompare(inst, IR::ValueOpcode::UGreaterThan64, IR::Type::U64, false, true);
+			EmitIntegerCompare(inst, IR::ValueOpcode::UGreaterThan64, IR::Type::U64, false);
 			return;
 		case O::V_CMP_NE_I64:
 		case O::V_CMP_NE_U64:
-			EmitIntegerCompare(inst, IR::ValueOpcode::INotEqual64, IR::Type::U64, false, false);
-			return;
 		case O::V_CMPX_NE_I64:
 		case O::V_CMPX_NE_U64:
-			EmitIntegerCompare(inst, IR::ValueOpcode::INotEqual64, IR::Type::U64, false, true);
+			EmitIntegerCompare(inst, IR::ValueOpcode::INotEqual64, IR::Type::U64, false);
 			return;
 
 		case O::V_CMP_LT_I64:
-			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThan64, IR::Type::U64, false, false);
-			return;
 		case O::V_CMPX_LT_I64:
-			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThan64, IR::Type::U64, false, true);
+			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThan64, IR::Type::U64, false);
 			return;
 		case O::V_CMP_LE_I64:
-			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThanEqual64, IR::Type::U64, false, false);
-			return;
 		case O::V_CMPX_LE_I64:
-			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThanEqual64, IR::Type::U64, false, true);
+			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThanEqual64, IR::Type::U64, false);
 			return;
 		// The IR has no signed 64-bit greater-than: a > b is b < a, and a >= b is b <= a.
 		case O::V_CMP_GT_I64:
-			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThan64, IR::Type::U64, false, false,
-			                   true);
-			return;
 		case O::V_CMPX_GT_I64:
-			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThan64, IR::Type::U64, false, true,
-			                   true);
+			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThan64, IR::Type::U64, false, true);
 			return;
 		case O::V_CMP_GE_I64:
-			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThanEqual64, IR::Type::U64, false, false,
-			                   true);
-			return;
 		case O::V_CMPX_GE_I64:
-			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThanEqual64, IR::Type::U64, false, true,
-			                   true);
+			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThanEqual64, IR::Type::U64, false, true);
 			return;
 		case O::V_CMP_LE_U64:
-			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThanEqual64, IR::Type::U64, false, false);
-			return;
 		case O::V_CMPX_LE_U64:
-			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThanEqual64, IR::Type::U64, false, true);
+			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThanEqual64, IR::Type::U64, false);
 			return;
 		case O::V_CMP_GE_U64:
-			EmitIntegerCompare(inst, IR::ValueOpcode::UGreaterThanEqual64, IR::Type::U64, false, false);
-			return;
 		case O::V_CMPX_GE_U64:
-			EmitIntegerCompare(inst, IR::ValueOpcode::UGreaterThanEqual64, IR::Type::U64, false, true);
+			EmitIntegerCompare(inst, IR::ValueOpcode::UGreaterThanEqual64, IR::Type::U64, false);
 			return;
 
 		case O::V_CMP_EQ_U16:
-			EmitInteger16Compare(inst, IR::ValueOpcode::IEqual32, false, false);
+		case O::V_CMPX_EQ_U16:
+			EmitInteger16Compare(inst, IR::ValueOpcode::IEqual32, false);
 			return;
 		case O::V_CMP_EQ_I16:
-			EmitInteger16Compare(inst, IR::ValueOpcode::IEqual32, true, false);
-			return;
 		case O::V_CMPX_EQ_I16:
-			EmitInteger16Compare(inst, IR::ValueOpcode::IEqual32, true, true);
+			EmitInteger16Compare(inst, IR::ValueOpcode::IEqual32, true);
 			return;
 		case O::V_CMP_NE_U16:
-			EmitInteger16Compare(inst, IR::ValueOpcode::INotEqual32, false, false);
-			return;
 		case O::V_CMPX_NE_U16:
-			EmitInteger16Compare(inst, IR::ValueOpcode::INotEqual32, false, true);
+			EmitInteger16Compare(inst, IR::ValueOpcode::INotEqual32, false);
 			return;
 		case O::V_CMP_NE_I16:
-			EmitInteger16Compare(inst, IR::ValueOpcode::INotEqual32, true, false);
-			return;
 		case O::V_CMPX_NE_I16:
-			EmitInteger16Compare(inst, IR::ValueOpcode::INotEqual32, true, true);
+			EmitInteger16Compare(inst, IR::ValueOpcode::INotEqual32, true);
 			return;
 		case O::V_CMP_GT_U16:
-			EmitInteger16Compare(inst, IR::ValueOpcode::UGreaterThan32, false, false);
-			return;
 		case O::V_CMPX_GT_U16:
-			EmitInteger16Compare(inst, IR::ValueOpcode::UGreaterThan32, false, true);
+			EmitInteger16Compare(inst, IR::ValueOpcode::UGreaterThan32, false);
 			return;
 		case O::V_CMP_GE_U16:
-			EmitInteger16Compare(inst, IR::ValueOpcode::UGreaterThanEqual32, false, false);
+			EmitInteger16Compare(inst, IR::ValueOpcode::UGreaterThanEqual32, false);
 			return;
 		case O::V_CMP_LT_U16:
-			EmitInteger16Compare(inst, IR::ValueOpcode::ULessThan32, false, false);
-			return;
 		case O::V_CMPX_LT_U16:
-			EmitInteger16Compare(inst, IR::ValueOpcode::ULessThan32, false, true);
-			return;
-		case O::V_CMPX_EQ_U16:
-			EmitInteger16Compare(inst, IR::ValueOpcode::IEqual32, false, true);
+			EmitInteger16Compare(inst, IR::ValueOpcode::ULessThan32, false);
 			return;
 		case O::V_CMP_LE_U16:
-			EmitInteger16Compare(inst, IR::ValueOpcode::ULessThanEqual32, false, false);
+			EmitInteger16Compare(inst, IR::ValueOpcode::ULessThanEqual32, false);
 			return;
 		case O::V_CMP_GT_I16:
-			EmitInteger16Compare(inst, IR::ValueOpcode::SGreaterThan32, true, false);
-			return;
 		case O::V_CMPX_GT_I16:
-			EmitInteger16Compare(inst, IR::ValueOpcode::SGreaterThan32, true, true);
+			EmitInteger16Compare(inst, IR::ValueOpcode::SGreaterThan32, true);
 			return;
 		case O::V_CMP_GE_I16:
-			EmitInteger16Compare(inst, IR::ValueOpcode::SGreaterThanEqual32, true, false);
-			return;
 		case O::V_CMPX_GE_I16:
-			EmitInteger16Compare(inst, IR::ValueOpcode::SGreaterThanEqual32, true, true);
+			EmitInteger16Compare(inst, IR::ValueOpcode::SGreaterThanEqual32, true);
 			return;
 		case O::V_CMP_LT_I16:
-			EmitInteger16Compare(inst, IR::ValueOpcode::SLessThan32, true, false);
-			return;
 		case O::V_CMPX_LT_I16:
-			EmitInteger16Compare(inst, IR::ValueOpcode::SLessThan32, true, true);
+			EmitInteger16Compare(inst, IR::ValueOpcode::SLessThan32, true);
 			return;
 		case O::V_CMP_LE_I16:
-			EmitInteger16Compare(inst, IR::ValueOpcode::SLessThanEqual32, true, false);
-			return;
 		case O::V_CMPX_LE_I16:
-			EmitInteger16Compare(inst, IR::ValueOpcode::SLessThanEqual32, true, true);
+			EmitInteger16Compare(inst, IR::ValueOpcode::SLessThanEqual32, true);
 			return;
 
-		case O::V_CMP_F_F32: EmitCompareConstant(inst, false, false, false); return;
-		case O::V_CMP_TRU_F32: EmitCompareConstant(inst, true, false, false); return;
+		case O::V_CMP_F_F32: EmitCompareConstant(inst, false, false); return;
+		case O::V_CMP_TRU_F32: EmitCompareConstant(inst, true, false); return;
 		case O::V_CMP_EQ_F32:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdEqual32, false, false);
+		case O::V_CMPX_EQ_F32:
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdEqual32, false);
 			return;
 		case O::V_CMP_EQ_F64:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdEqual64, false, false);
+		case O::V_CMPX_EQ_F64:
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdEqual64, false);
 			return;
 		case O::V_CMP_LE_F64:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdLessThanEqual64, false, false);
-			return;
 		case O::V_CMPX_LE_F64:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdLessThanEqual64, false, true);
-			return;
-		case O::V_CMPX_GE_F64:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdGreaterThanEqual64, false, true);
-			return;
-		case O::V_CMP_F_F64: EmitCompareConstant(inst, false, false, false); return;
-		case O::V_CMP_TRU_F64: EmitCompareConstant(inst, true, false, false); return;
-		case O::V_CMPX_F_F64: EmitCompareConstant(inst, false, false, true); return;
-		case O::V_CMPX_TRU_F64: EmitCompareConstant(inst, true, false, true); return;
-		case O::V_CMP_LT_F64:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdLessThan64, false, false);
-			return;
-		case O::V_CMPX_LT_F64:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdLessThan64, false, true);
-			return;
-		case O::V_CMPX_EQ_F64:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdEqual64, false, true);
-			return;
-		case O::V_CMP_GT_F64:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdGreaterThan64, false, false);
-			return;
-		case O::V_CMPX_GT_F64:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdGreaterThan64, false, true);
-			return;
-		case O::V_CMP_LG_F64:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdNotEqual64, false, false);
-			return;
-		case O::V_CMPX_LG_F64:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdNotEqual64, false, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdLessThanEqual64, false);
 			return;
 		case O::V_CMP_GE_F64:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdGreaterThanEqual64, false, false);
+		case O::V_CMPX_GE_F64:
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdGreaterThanEqual64, false);
+			return;
+		case O::V_CMP_F_F64:
+		case O::V_CMPX_F_F64: EmitCompareConstant(inst, false, false); return;
+		case O::V_CMP_TRU_F64:
+		case O::V_CMPX_TRU_F64: EmitCompareConstant(inst, true, false); return;
+		case O::V_CMP_LT_F64:
+		case O::V_CMPX_LT_F64:
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdLessThan64, false);
+			return;
+		case O::V_CMP_GT_F64:
+		case O::V_CMPX_GT_F64:
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdGreaterThan64, false);
+			return;
+		case O::V_CMP_LG_F64:
+		case O::V_CMPX_LG_F64:
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdNotEqual64, false);
 			return;
 		case O::V_CMP_NLG_F64:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordEqual64, false, false);
-			return;
 		case O::V_CMPX_NLG_F64:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordEqual64, false, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordEqual64, false);
 			return;
 		case O::V_CMP_NEQ_F64:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordNotEqual64, false, false);
-			return;
 		case O::V_CMPX_NEQ_F64:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordNotEqual64, false, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordNotEqual64, false);
 			return;
 		case O::V_CMP_NLE_F64:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordGreaterThan64, false, false);
-			return;
 		case O::V_CMPX_NLE_F64:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordGreaterThan64, false, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordGreaterThan64, false);
 			return;
 		case O::V_CMP_NLT_F64:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordGreaterThanEqual64, false, false);
-			return;
 		case O::V_CMPX_NLT_F64:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordGreaterThanEqual64, false, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordGreaterThanEqual64, false);
 			return;
 		case O::V_CMP_NGE_F64:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordLessThan64, false, false);
-			return;
 		case O::V_CMPX_NGE_F64:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordLessThan64, false, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordLessThan64, false);
 			return;
 		case O::V_CMP_NGT_F64:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordLessThanEqual64, false, false);
-			return;
 		case O::V_CMPX_NGT_F64:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordLessThanEqual64, false, true);
-			return;
-		case O::V_CMP_O_F64: EmitFloatOrderedCompare(inst, true, false, true); return;
-		case O::V_CMP_U_F64: EmitFloatOrderedCompare(inst, false, false, true); return;
-		case O::V_CMPX_O_F64: EmitFloatOrderedCompare(inst, true, true, true); return;
-		case O::V_CMPX_U_F64: EmitFloatOrderedCompare(inst, false, true, true); return;
-		case O::V_CMPX_EQ_F32:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdEqual32, false, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordLessThanEqual64, false);
 			return;
+		case O::V_CMP_O_F64:
+		case O::V_CMPX_O_F64: EmitFloatOrderedCompare(inst, true, true); return;
+		case O::V_CMP_U_F64:
+		case O::V_CMPX_U_F64: EmitFloatOrderedCompare(inst, false, true); return;
 		case O::V_CMP_LG_F32:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdNotEqual32, false, false);
-			return;
 		case O::V_CMPX_LG_F32:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdNotEqual32, false, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdNotEqual32, false);
 			return;
 		case O::V_CMP_GT_F32:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdGreaterThan32, false, false);
-			return;
 		case O::V_CMPX_GT_F32:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdGreaterThan32, false, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdGreaterThan32, false);
 			return;
 		case O::V_CMP_GE_F32:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdGreaterThanEqual32, false, false);
-			return;
 		case O::V_CMPX_GE_F32:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdGreaterThanEqual32, false, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdGreaterThanEqual32, false);
 			return;
 		case O::V_CMP_LT_F32:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdLessThan32, false, false);
-			return;
 		case O::V_CMPX_LT_F32:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdLessThan32, false, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdLessThan32, false);
 			return;
 		case O::V_CMP_LE_F32:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdLessThanEqual32, false, false);
-			return;
 		case O::V_CMPX_LE_F32:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdLessThanEqual32, false, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdLessThanEqual32, false);
 			return;
 		case O::V_CMP_NLG_F32:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordEqual32, false, false);
-			return;
 		case O::V_CMPX_NLG_F32:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordEqual32, false, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordEqual32, false);
 			return;
 		case O::V_CMP_NEQ_F32:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordNotEqual32, false, false);
-			return;
 		case O::V_CMPX_NEQ_F32:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordNotEqual32, false, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordNotEqual32, false);
 			return;
 		case O::V_CMP_NLE_F32:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordGreaterThan32, false, false);
-			return;
 		case O::V_CMPX_NLE_F32:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordGreaterThan32, false, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordGreaterThan32, false);
 			return;
 		case O::V_CMP_NLT_F32:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordGreaterThanEqual32, false, false);
-			return;
 		case O::V_CMPX_NLT_F32:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordGreaterThanEqual32, false, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordGreaterThanEqual32, false);
 			return;
 		case O::V_CMP_NGE_F32:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordLessThan32, false, false);
-			return;
 		case O::V_CMPX_NGE_F32:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordLessThan32, false, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordLessThan32, false);
 			return;
 		case O::V_CMP_NGT_F32:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordLessThanEqual32, false, false);
-			return;
 		case O::V_CMPX_NGT_F32:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordLessThanEqual32, false, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordLessThanEqual32, false);
 			return;
 		case O::V_CMP_EQ_F16:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdEqual32, true, false);
-			return;
 		case O::V_CMPX_EQ_F16:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdEqual32, true, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdEqual32, true);
 			return;
 		case O::V_CMP_LG_F16:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdNotEqual32, true, false);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdNotEqual32, true);
 			return;
 		case O::V_CMP_GT_F16:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdGreaterThan32, true, false);
-			return;
 		case O::V_CMPX_GT_F16:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdGreaterThan32, true, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdGreaterThan32, true);
 			return;
 		case O::V_CMP_GE_F16:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdGreaterThanEqual32, true, false);
-			return;
 		case O::V_CMPX_GE_F16:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdGreaterThanEqual32, true, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdGreaterThanEqual32, true);
 			return;
 		case O::V_CMP_LT_F16:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdLessThan32, true, false);
-			return;
 		case O::V_CMPX_LT_F16:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdLessThan32, true, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdLessThan32, true);
 			return;
 		case O::V_CMP_LE_F16:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdLessThanEqual32, true, false);
-			return;
 		case O::V_CMPX_LE_F16:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdLessThanEqual32, true, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdLessThanEqual32, true);
 			return;
 		case O::V_CMP_NGE_F16:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordLessThan32, true, false);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordLessThan32, true);
 			return;
 		case O::V_CMP_NGT_F16:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordLessThanEqual32, true, false);
-			return;
 		case O::V_CMPX_NGT_F16:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordLessThanEqual32, true, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordLessThanEqual32, true);
 			return;
 		case O::V_CMPX_NLE_F16:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordGreaterThan32, true, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordGreaterThan32, true);
 			return;
 		case O::V_CMP_NEQ_F16:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordNotEqual32, true, false);
-			return;
 		case O::V_CMPX_NEQ_F16:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordNotEqual32, true, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordNotEqual32, true);
 			return;
 		case O::V_CMP_NLT_F16:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordGreaterThanEqual32, true, false);
-			return;
 		case O::V_CMPX_NLT_F16:
-			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordGreaterThanEqual32, true, true);
+			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordGreaterThanEqual32, true);
 			return;
-		case O::V_CMP_O_F32: EmitFloatOrderedCompare(inst, true, false); return;
-		case O::V_CMP_U_F32: EmitFloatOrderedCompare(inst, false, false); return;
-		case O::V_CMPX_O_F32: EmitFloatOrderedCompare(inst, true, true); return;
-		case O::V_CMPX_U_F32: EmitFloatOrderedCompare(inst, false, true); return;
-		case O::V_CMP_CLASS_F32: EmitFloatClassCompare(inst, false); return;
-		case O::V_CMPX_CLASS_F32: EmitFloatClassCompare(inst, true); return;
-		case O::V_CMP_CLASS_F16: EmitFloatClassCompare(inst, false, true); return;
-		case O::V_CMPX_CLASS_F16: EmitFloatClassCompare(inst, true, true); return;
+		case O::V_CMP_O_F32:
+		case O::V_CMPX_O_F32: EmitFloatOrderedCompare(inst, true); return;
+		case O::V_CMP_U_F32:
+		case O::V_CMPX_U_F32: EmitFloatOrderedCompare(inst, false); return;
+		case O::V_CMP_CLASS_F32:
+		case O::V_CMPX_CLASS_F32: EmitFloatClassCompare(inst); return;
+		case O::V_CMP_CLASS_F16:
+		case O::V_CMPX_CLASS_F16: EmitFloatClassCompare(inst, true); return;
 
 		case O::V_CVT_F32_UBYTE0: V_CVT_F32_UBYTE(inst, 0); return;
 		case O::V_CVT_F32_UBYTE1: V_CVT_F32_UBYTE(inst, 1); return;
@@ -495,8 +361,8 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 			return;
 		case O::V_CVT_PK_U8_F32: V_CVT_PK_U8_F32(inst); return;
 		case O::V_PACK_B32_F16: V_PACK_B32_F16(inst); return;
-		case O::V_CVT_PK_U16_U32:
-		case O::V_CVT_PK_I16_I32: return PackB16(inst, false, false);
+		case O::V_CVT_PK_U16_U32: return V_CVT_PK_16_32(inst, false);
+		case O::V_CVT_PK_I16_I32: return V_CVT_PK_16_32(inst, true);
 
 		case O::V_LSHLREV_B16:
 			return Integer16Shift(inst, IR::ValueOpcode::ShiftLeftLogical32, false);
@@ -563,7 +429,7 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_CEIL_F16: return Float16Unary(inst, IR::ValueOpcode::FPCeil32, false);
 		case O::V_TRUNC_F16: return Float16Unary(inst, IR::ValueOpcode::FPTrunc32, false);
 		case O::V_RNDNE_F16: return Float16Unary(inst, IR::ValueOpcode::FPRoundEven32, false);
-		case O::V_FRACT_F16: return Float16Unary(inst, IR::ValueOpcode::FPFract32, false);
+		case O::V_FRACT_F16: return V_FRACT(inst, IR::Type::F16);
 		case O::V_SIN_F16: return Float16Trig(inst, IR::ValueOpcode::FPSin);
 		case O::V_COS_F16: return Float16Trig(inst, IR::ValueOpcode::FPCos);
 		case O::V_MIN3_F16: return Float16Ternary(inst, IR::ValueOpcode::FPMinTri32, false, false);
@@ -574,8 +440,8 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_RCP_F32: return FloatUnary(inst, IR::ValueOpcode::FPRecip32);
 		case O::V_RCP_F64: return FloatUnary(inst, IR::ValueOpcode::FPRecip64);
 		case O::V_RCP_IFLAG_F32: return FloatUnary(inst, IR::ValueOpcode::FPRecipIFlag32);
-		case O::V_FRACT_F32: return FloatUnary(inst, IR::ValueOpcode::FPFract32);
-		case O::V_FRACT_F64: return FloatUnary(inst, IR::ValueOpcode::FPFract64);
+		case O::V_FRACT_F32: return V_FRACT(inst, IR::Type::F32);
+		case O::V_FRACT_F64: return V_FRACT(inst, IR::Type::F64);
 		case O::V_TRUNC_F32: return FloatUnary(inst, IR::ValueOpcode::FPTrunc32);
 		case O::V_TRUNC_F64: return FloatUnary(inst, IR::ValueOpcode::FPTrunc64);
 		case O::V_CEIL_F32: return FloatUnary(inst, IR::ValueOpcode::FPCeil32);

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

@@ -118,15 +118,14 @@ ResourcePlan& ResourcePlan::operator=(ResourcePlan&& other) noexcept {
 }
 
 Program::~Program() {
-	// Planning expressions can refer to block values but outlive block storage in the base class.
+	// Planning expressions can refer to block values but outlive block storage in the base class,
+	// and values cross block boundaries: detach everything before any storage is destroyed.
 	for (auto& inst: value_storage) {
-		inst.Invalidate();
+		inst.DropArgs();
 	}
-	// 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) {
 		for (auto& inst: *block) {
-			inst.Invalidate();
+			inst.DropArgs();
 		}
 	}
 }
@@ -169,6 +168,8 @@ bool EquivalentSelectorDomain(const ResourcePlan& program,
                               const DescriptorSource::IndirectDescriptor& right) {
 	return left.selector == right.selector && left.signed_count == right.signed_count &&
 	       left.workgroup_axis == right.workgroup_axis &&
+	       left.workgroup_multiplier == right.workgroup_multiplier &&
+	       left.workgroup_shift == right.workgroup_shift &&
 	       EquivalentValue(program, left.key_count, right.key_count) &&
 	       EquivalentValue(program, left.selector_first, right.selector_first) &&
 	       EquivalentValue(program, left.selector_mask, right.selector_mask);

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

@@ -0,0 +1,184 @@
+#ifndef EMULATOR_INCLUDE_EMULATOR_GRAPHICS_SHADER_RECOMPILER_SERIALIZE_H_
+#define EMULATOR_INCLUDE_EMULATOR_GRAPHICS_SHADER_RECOMPILER_SERIALIZE_H_
+
+#include "common/assert.h"
+#include "graphics/shader/recompiler/backend/spirv/SpirvEmitter.h"
+#include "graphics/shader/recompiler/ir/ShaderIR.h"
+#include "graphics/shader/recompiler/ir/passes/ResourceMaterialization.h"
+#include "graphics/shader/rectListShader.h"
+
+#include <algorithm>
+#include <bit>
+#include <optional>
+#include <span>
+#include <string>
+#include <tuple>
+#include <type_traits>
+#include <unordered_map>
+#include <vector>
+
+// Shader cache records. Io(io, value) appends value to a Writer or reads it back from a Reader
+// field by field, so padding never reaches the bytes; instructions travel as plan-local indices.
+namespace Libs::Graphics::ShaderRecompiler::IR {
+
+struct Writer {
+	std::vector<uint8_t>                      bytes;
+	std::unordered_map<const Inst*, uint32_t> index; // Instructions that values may refer to.
+
+	void Raw(const void* data, size_t size) {
+		const auto* begin = static_cast<const uint8_t*>(data);
+		bytes.insert(bytes.end(), begin, begin + size);
+	}
+};
+
+struct Reader {
+	std::span<const uint8_t> bytes;
+	std::vector<Inst*>       insts;     // Instructions by index.
+	bool                     ok = true; // False once a read ran past the end or was invalid.
+
+	void Raw(void* data, size_t size) {
+		ok = ok && size <= bytes.size();
+		if (!ok) return;
+		std::copy_n(bytes.data(), size, static_cast<uint8_t*>(data));
+		bytes = bytes.subspan(size);
+	}
+	// Every element takes at least one byte, which bounds counts read from damaged records.
+	uint32_t Count(uint32_t count) {
+		ok = ok && count <= bytes.size();
+		return ok ? count : 0u;
+	}
+};
+
+template <class S> constexpr bool Reading = std::is_same_v<S, Reader>;
+template <class T, template <class...> class Of> constexpr bool IsA = false;
+template <template <class...> class Of, class... A> constexpr bool IsA<Of<A...>, Of> = true;
+template <class T, class U> concept Is = std::is_same_v<std::remove_const_t<T>, U>;
+
+template <class S, class T> void Io(S& io, T& value);
+
+template <class S, class... T> void IoAll(S& io, T&&... values) {
+	(Io(io, values), ...);
+}
+
+// The value's Type, Opaque for an instruction, then the instruction index or the zero-extended
+// immediate.
+inline void IoValue(Writer& writer, const Value& value) {
+	auto     type = Type::Opaque;
+	uint64_t bits = 0;
+	if (const auto* inst = value.TryInstruction()) {
+		bits = writer.index.at(inst);
+	} else {
+		type = value.GetType();
+		switch (type) {
+			case Type::Void: break;
+			case Type::ScalarReg: bits = RegIndex(value.ScalarRegister()); break;
+			case Type::VectorReg: bits = RegIndex(value.VectorRegister()); break;
+			case Type::U1: bits = value.U1(); break;
+			case Type::U8: bits = value.U8(); break;
+			case Type::U16: bits = value.U16(); break;
+			case Type::U32: bits = value.U32(); break;
+			case Type::U64: bits = value.U64(); break;
+			case Type::F16: bits = value.F16Bits(); break;
+			case Type::F32: bits = std::bit_cast<uint32_t>(value.F32Value()); break;
+			default: EXIT("value type %u has no cache encoding\n", static_cast<uint32_t>(type));
+		}
+	}
+	IoAll(writer, type, bits);
+}
+
+inline void IoValue(Reader& reader, Value& value) {
+	auto     type = Type::Void;
+	uint64_t bits = 0;
+	IoAll(reader, type, bits);
+	switch (type) {
+		case Type::Void: value = {}; break;
+		case Type::Opaque:
+			reader.ok = reader.ok && bits < reader.insts.size();
+			if (reader.ok) value = Value(reader.insts[bits]);
+			break;
+		case Type::ScalarReg: value = Value(static_cast<ScalarReg>(bits)); break;
+		case Type::VectorReg: value = Value(static_cast<VectorReg>(bits)); break;
+		case Type::U1: value = Value(bits != 0u); break;
+		case Type::U8: value = Value(static_cast<uint8_t>(bits)); break;
+		case Type::U16: value = Value(static_cast<uint16_t>(bits)); break;
+		case Type::U32: value = Value(static_cast<uint32_t>(bits)); break;
+		case Type::U64: value = Value(bits); break;
+		case Type::F16: value = Value::F16(static_cast<uint16_t>(bits)); break;
+		case Type::F32: value = Value::F32(std::bit_cast<float>(static_cast<uint32_t>(bits))); break;
+		default: reader.ok = false; break;
+	}
+}
+
+// Scalars and standard containers; structs have their own overloads below.
+template <class S, class T> void Io(S& io, T& value) {
+	using U = std::remove_const_t<T>;
+	static_assert(!std::is_pointer_v<U>, "pointers do not persist");
+	if constexpr (std::is_arithmetic_v<U> || std::is_enum_v<U>) {
+		io.Raw(&value, sizeof(U));
+	} else if constexpr (std::is_same_v<U, Value>) {
+		IoValue(io, value);
+	} else if constexpr (IsA<U, std::vector> || IsA<U, std::basic_string>) {
+		auto count = static_cast<uint32_t>(value.size());
+		Io(io, count);
+		if constexpr (Reading<S>) value.resize(io.Count(count));
+		using E = typename U::value_type;
+		if constexpr (std::is_arithmetic_v<E> || std::is_enum_v<E>) {
+			io.Raw(value.data(), value.size() * sizeof(E));
+		} else {
+			for (auto& element: value) Io(io, element);
+		}
+	} else if constexpr (IsA<U, std::optional>) {
+		bool has = value.has_value();
+		Io(io, has);
+		if constexpr (Reading<S>) {
+			if (has) value = typename U::value_type {}; // Not emplace(): clang rejects nested aggregates.
+			else value.reset();
+		}
+		if (has) Io(io, *value);
+	} else {
+		static_assert(requires { std::tuple_size<U>::value; }, "struct without an Io overload");
+		std::apply([&](auto&... elements) { IoAll(io, elements...); }, value); // std::array, std::pair
+	}
+}
+
+// Binds every field, so that a field added later fails to compile until it is listed.
+#define KYTY_IO_FIELDS(Struct, ...)                                                                \
+	template <class S, Is<Struct> T> void Io(S& io, T& object) {                                   \
+		auto& [__VA_ARGS__] = object;                                                              \
+		IoAll(io, __VA_ARGS__);                                                                    \
+	}
+KYTY_IO_FIELDS(MemoryInfo, a, b, c, d, e, f, g, h, i, j, k, l, m, n, o, p, q, r, s, t, u, v, w, x, y)
+KYTY_IO_FIELDS(BufferResource, a, b, c, d, e, f, g, h, i, j, k, l, m, n, o, p, q, r, s)
+KYTY_IO_FIELDS(ImageResource, a, b, c, d, e, f, g, h, i, j, k, l, m, n, o, p, q, r, s, t, u, v, w)
+KYTY_IO_FIELDS(SamplerResource, a, b, c, d, e, f, g, h, i, j, k)
+KYTY_IO_FIELDS(SampledResourcePair, a, b, c)
+KYTY_IO_FIELDS(StageInput, a, b, c, d, e)
+KYTY_IO_FIELDS(StageOutput, a, b, c, d)
+KYTY_IO_FIELDS(ShaderInfo, a, b, c, d, e, f, g, h, i, j, k, l, m, n, o, p, q, r, s, t, u, v, w, x, y, z,
+               A, B, C, D, E, F, G, H, I, J, K)
+KYTY_IO_FIELDS(DescriptorSource::IndirectDescriptor::SelectorRead::CountedRows, a, b)
+KYTY_IO_FIELDS(DescriptorSource::IndirectDescriptor::SelectorRead::MaskRanks, a, b, c)
+KYTY_IO_FIELDS(DescriptorSource::IndirectDescriptor::SelectorRead, a, b, c, d, e, f, g, h, i, j, k,
+               l)
+KYTY_IO_FIELDS(DescriptorSource::IndirectDescriptor::PointerTable, a, b, c, d)
+KYTY_IO_FIELDS(DescriptorSource::IndirectDescriptor, a, b, c, d, e, f, g, h, i, j, k, l, m, n, o, p,
+               q, r)
+KYTY_IO_FIELDS(DescriptorSource, a, b, c)
+KYTY_IO_FIELDS(SrtRead, a, b)
+KYTY_IO_FIELDS(ResourceBlock, a, b, c, d, e)
+KYTY_IO_FIELDS(UniformFill, a, b, c, d, e, f)
+KYTY_IO_FIELDS(UniformFillPlan, a, b, c, d)
+KYTY_IO_FIELDS(ResourceSpecialization::Buffer, a, b, c, d, e, f, g, h)
+KYTY_IO_FIELDS(ResourceSpecialization::Image, a, b, c, d, e, f, g, h, i, j, k)
+KYTY_IO_FIELDS(ResourceSpecialization::Sampler, a, b, c)
+KYTY_IO_FIELDS(ResourceSpecialization, a, b, c)
+KYTY_IO_FIELDS(Bindings, a, b, c, d, e)
+KYTY_IO_FIELDS(CompiledShaderInfo, a, b, c, d, e, f, g, h, i, j)
+KYTY_IO_FIELDS(Spirv::Profile, a, b, c)
+KYTY_IO_FIELDS(ShaderTessellationInputInfo, a, b, c, d, e, f, g, h, i)
+KYTY_IO_FIELDS(RectListParameter, a, b, c)
+#undef KYTY_IO_FIELDS
+
+} // namespace Libs::Graphics::ShaderRecompiler::IR
+
+#endif /* EMULATOR_INCLUDE_EMULATOR_GRAPHICS_SHADER_RECOMPILER_SERIALIZE_H_ */

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

@@ -14,7 +14,6 @@
 
 #include <array>
 #include <bit>
-#include <deque>
 #include <list>
 #include <memory>
 #include <optional>
@@ -28,8 +27,6 @@ class CodeGenerator;
 
 namespace Libs::Graphics::ShaderRecompiler::IR {
 
-class SrtWalker;
-
 // Finite descriptor domains also bound the arithmetic used to prove their indices.
 constexpr uint32_t MaxIndirectDescriptorProbes = 65536u;
 
@@ -254,6 +251,8 @@ enum class StageInputKind {
 	LocalInvocationId,
 	LocalInvocationIndex,
 	GlobalInvocationId,
+	BaseVertex,
+	BaseInstance,
 	Parameter,
 };
 
@@ -368,7 +367,7 @@ static_assert(static_cast<uint32_t>(DescriptorBindingKind::Count) == 58u);
 
 struct PushData {
 	static constexpr uint32_t DwordCount = 32;
-	static constexpr uint32_t MeshDrawDwordCount = 6;
+	static constexpr uint32_t MeshDrawDwordCount = 5;
 	static constexpr uint32_t NoStart    = UINT32_MAX;
 	std::array<uint32_t, DwordCount> dwords {};
 
@@ -498,10 +497,16 @@ struct Bindings {
 	uint32_t dispatch_thread_dword = PushData::NoStart;
 	uint32_t memory_offset_dword = 0;
 
-	[[nodiscard]] uint32_t ShaderDataDwords() const {
+	// Packed 8-bit binding offsets, then each binding's bound range in bytes: OpArrayLength
+	// reads 0 for ranges of 2 GiB or more on NVIDIA.
+	[[nodiscard]] uint32_t MemorySizeDword() const {
 		return memory_offset_dword +
 		       (descriptor_counts[static_cast<size_t>(DescriptorBindingKind::Buffers)] + 3u) / 4u;
 	}
+	[[nodiscard]] uint32_t ShaderDataDwords() const {
+		return MemorySizeDword() +
+		       descriptor_counts[static_cast<size_t>(DescriptorBindingKind::Buffers)];
+	}
 	[[nodiscard]] bool UsesPushData() const {
 		return push_data_start_dword != PushData::NoStart;
 	}
@@ -544,6 +549,7 @@ struct ShaderInfo {
 	bool                            subgroup_shuffle              = false;
 	bool                            subgroup_arithmetic           = false;
 	bool                            subgroup_local_invocation_id  = false;
+	bool                            helper_invocation             = false;
 	bool                            compute_derivatives           = false;
 	bool                            image_gather_extended         = false;
 	bool                            function_lds                  = false;
@@ -568,6 +574,13 @@ struct DescriptorSource {
 
 				bool operator==(const CountedRows& other) const = default;
 			};
+			struct MaskRanks {
+				uint32_t source = UINT32_MAX;
+				uint32_t stride = 0; // Byte spacing of the mask words its load can address.
+				uint32_t offset = 0;
+
+				bool operator==(const MaskRanks& other) const = default;
+			};
 			uint32_t source = UINT32_MAX;
 			uint32_t stride = 0;
 			uint32_t offset = 0;
@@ -575,10 +588,12 @@ struct DescriptorSource {
 			uint32_t bit_offset = 0;
 			uint32_t bit_mask = UINT32_MAX;
 			bool global = false;
-			// Four DWORD bitmap words per record; set-bit ranks subtract from the last key.
+			// Four or eight DWORD bitmap words per record; set ranks subtract from the last key.
 			bool bitset = false;
 			std::optional<CountedRows> rows;
 			std::vector<std::pair<uint32_t, uint32_t>> byte_ranges;
+			std::optional<uint32_t> excluded; // A guard keeps this key from the read.
+			std::optional<MaskRanks> ranks;   // Records are the bits some mask word sets.
 
 			bool operator==(const SelectorRead& other) const = default;
 		};
@@ -602,8 +617,12 @@ struct DescriptorSource {
 		bool     indexed = false;
 		bool     signed_count = false;
 		uint32_t workgroup_axis  = UINT32_MAX;
+		// Nonzero shift: workgroup keys are (selector_first + ID) * multiplier >> shift.
+		uint32_t workgroup_multiplier = 0;
+		uint32_t workgroup_shift      = 0;
 		Value    key_count;
-		// First selector index, or first descriptor key for a direct or bitmap range.
+		// First selector index, workgroup key base, or first descriptor key for a direct or bitmap
+		// range.
 		Value                 selector_first;
 		Value    selector_mask;
 		std::vector<uint32_t> sources;
@@ -630,6 +649,7 @@ struct ResourceBlock {
 	Value                 condition;
 	std::vector<uint32_t> successors;
 	std::vector<uint32_t> sources;
+	std::vector<uint32_t> images;
 	std::vector<uint32_t> srt_reads;
 };
 
@@ -658,24 +678,10 @@ struct UniformFillPlan {
 // condition and fill values without translated blocks, plus reusable evaluation scratch.
 struct ResourcePlan {
 	struct SelectorKeys {
-		uint32_t              source = UINT32_MAX;
 		std::vector<uint32_t> values;
 		bool                  resolved = false;
 	};
 
-	struct EvaluationContext {
-		struct Entry {
-			uint64_t value      = 0;
-			uint64_t generation = 0;
-		};
-
-		std::vector<Entry> values;
-		uint64_t           generation = 0;
-		const SrtWalker*    session = nullptr;
-		Value              active_mask;
-		uint8_t            read_policy = 0;
-	};
-
 	ResourcePlan();
 	~ResourcePlan();
 
@@ -695,21 +701,18 @@ struct ResourcePlan {
 	std::vector<ResourceBlock>          control_flow;
 	std::unordered_map<const Inst*, std::vector<uint32_t>> control_phi_predecessors;
 	std::vector<SrtRead>                srt_reads;
-	bool                                requires_specialization_memory = false;
 	std::vector<std::pair<uint64_t, uint64_t>> source_reads;
 	bool                                capture_specialization_reads = false;
 	bool                                srt_plan_complete          = false;
 	bool                                resource_tracking_complete = false;
 	ShaderInfo                          info;
 	UniformFillPlan                     uniform_fill;
+	// x86 walker over the retained SRT, descriptor and condition values.
 	std::unique_ptr<Xbyak::CodeGenerator> srt_code;
-	// GPU-thread scratch for nested clean/EXEC memos, activity and material keys.
-	mutable std::deque<EvaluationContext> evaluation_contexts;
-	mutable uint32_t                       evaluation_value_count = 0;
-	mutable std::vector<uint8_t>            active_sources;
-	mutable std::vector<uint8_t>            visited_blocks;
-	mutable std::vector<uint8_t>            reachable_blocks;
-	mutable std::vector<uint32_t>           pending_blocks;
+	uint32_t                              srt_user_data = 0; // User-data DWORDs the walker reads.
+	uint32_t                              srt_captures  = 0; // Scratch offset of loaded addresses.
+	// GPU-thread scratch for walker results and material keys.
+	mutable std::vector<uint64_t>           srt_scratch;
 	mutable std::vector<uint32_t>           material_keys;
 	mutable std::vector<SelectorKeys>       selector_keys;
 };

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

@@ -25,22 +25,6 @@ Value Value::F32(float value) {
 	return Value(Type::F32, std::bit_cast<uint32_t>(value));
 }
 
-bool Value::IsEmpty() const {
-	return type == Type::Void;
-}
-
-bool Value::IsImmediate() const {
-	return type != Type::Opaque;
-}
-
-bool Value::IsIdentity() const {
-	return type == Type::Opaque && inst->GetOpcode() == ValueOpcode::Identity;
-}
-
-bool Value::IsPhi() const {
-	return type == Type::Opaque && inst->GetOpcode() == ValueOpcode::Phi;
-}
-
 Type Value::GetType() const {
 	if (IsPhi()) {
 		return inst->Flags<Type>();
@@ -56,19 +40,6 @@ Inst* Value::Instruction() const {
 	return inst;
 }
 
-Inst* Value::TryInstruction() const {
-	return type == Type::Opaque ? inst : nullptr;
-}
-
-Inst* Value::ResolveInstruction() const {
-	EXIT_IF(type != Type::Opaque);
-	return IsIdentity() ? inst->Arg(0).ResolveInstruction() : inst;
-}
-
-Value Value::Resolve() const {
-	return IsIdentity() ? inst->Arg(0).Resolve() : *this;
-}
-
 ScalarReg Value::ScalarRegister() const {
 	EXIT_IF(type != Type::ScalarReg);
 	return scalar_reg;
@@ -94,11 +65,6 @@ uint16_t Value::U16() const {
 	return imm_u16;
 }
 
-uint32_t Value::U32() const {
-	EXIT_IF(type != Type::U32);
-	return imm_u32;
-}
-
 uint64_t Value::U64() const {
 	EXIT_IF(type != Type::U64);
 	return imm_u64;
@@ -152,10 +118,6 @@ Inst::~Inst() {
 	ClearArgs();
 }
 
-ValueOpcode Inst::GetOpcode() const {
-	return opcode;
-}
-
 Type Inst::GetType() const {
 	if (opcode == ValueOpcode::Phi) {
 		return static_cast<Type>(flags);
@@ -178,51 +140,25 @@ size_t Inst::UseCount() const {
 	return uses.size();
 }
 
-size_t Inst::NumArgs() const {
-	return num_args == PhiArity ? phi_args.size() : num_args;
-}
-
 size_t Inst::NumPhiBlocks() const {
 	return num_args == PhiArity ? phi_args.size() : 0;
 }
 
-Value Inst::Arg(size_t index) const {
-	EXIT_IF(index >= NumArgs());
-	if (num_args <= InlineArity) {
-		return fixed_args[index];
-	}
-	return num_args == PhiArity ? phi_args[index].second : large_args[index];
-}
-
 Block* Inst::PhiBlock(size_t index) const {
 	EXIT_IF(opcode != ValueOpcode::Phi || index >= phi_args.size());
 	return phi_args[index].first;
 }
 
-Block* Inst::Parent() const {
-	return parent;
-}
-
-const std::vector<Use>& Inst::Uses() const {
-	return uses;
-}
-
 void Inst::SetParent(Block* block) {
 	parent = block;
 }
 
 void Inst::SetArg(size_t index, Value value) {
-	const auto old = Arg(index);
-	if (auto* old_inst = old.TryInstruction(); old_inst != nullptr) {
+	auto& arg = ArgSlot(*this, index);
+	if (auto* old_inst = arg.TryInstruction(); old_inst != nullptr) {
 		RemoveUse(old_inst, index);
 	}
-	if (num_args <= InlineArity) {
-		fixed_args[index] = value;
-	} else if (num_args == PhiArity) {
-		phi_args[index].second = value;
-	} else {
-		large_args[index] = value;
-	}
+	arg = value;
 	if (auto* new_inst = value.TryInstruction(); new_inst != nullptr) {
 		AddUse(new_inst, index);
 	}
@@ -238,11 +174,15 @@ void Inst::AddPhiOperand(Block* predecessor, Value value) {
 }
 
 void Inst::ReplaceUsesWith(Value replacement, bool preserve) {
-	const auto old_uses = uses;
-	for (const auto& use: old_uses) {
-		use.user->SetArg(use.operand, replacement);
-	}
+	// Detach the arguments first (a phi can use itself), then move every use over in order.
 	Invalidate();
+	const auto old_uses = std::exchange(uses, {});
+	for (const auto& use: old_uses) {
+		ArgSlot(*use.user, use.operand) = replacement;
+	}
+	if (auto* target = replacement.TryInstruction(); target != nullptr) {
+		target->uses.insert(target->uses.end(), old_uses.begin(), old_uses.end());
+	}
 	if (preserve) {
 		opcode = ValueOpcode::Identity;
 		num_args = 1;
@@ -256,9 +196,6 @@ void Inst::Invalidate() {
 }
 
 void Inst::AddUse(Inst* used, size_t operand) {
-	const auto found = std::ranges::find_if(
-	    used->uses, [&](const Use& use) { return use.user == this && use.operand == operand; });
-	EXIT_IF(found != used->uses.end());
 	used->uses.push_back({this, operand});
 }
 
@@ -275,6 +212,10 @@ void Inst::ClearArgs() {
 			RemoveUse(value_inst, index);
 		}
 	}
+	DropArgs();
+}
+
+void Inst::DropArgs() {
 	if (num_args == PhiArity) {
 		std::destroy_at(&phi_args);
 		std::construct_at(&fixed_args);

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

@@ -33,14 +33,14 @@ public:
 	static Value F16(uint16_t bits);
 	static Value F32(float value);
 
-	[[nodiscard]] bool IsEmpty() const;
-	[[nodiscard]] bool IsImmediate() const;
+	[[nodiscard]] bool IsEmpty() const { return type == Type::Void; }
+	[[nodiscard]] bool IsImmediate() const { return type != Type::Opaque; }
 	[[nodiscard]] bool IsIdentity() const;
 	[[nodiscard]] bool IsPhi() const;
 	[[nodiscard]] Type GetType() const;
 
 	[[nodiscard]] Inst*     Instruction() const;
-	[[nodiscard]] Inst*     TryInstruction() const;
+	[[nodiscard]] Inst*     TryInstruction() const { return type == Type::Opaque ? inst : nullptr; }
 	[[nodiscard]] Inst*     ResolveInstruction() const;
 	[[nodiscard]] Value     Resolve() const;
 	[[nodiscard]] ScalarReg ScalarRegister() const;
@@ -112,7 +112,7 @@ public:
 	Inst(Inst&&)                 = delete;
 	Inst& operator=(Inst&&)      = delete;
 
-	[[nodiscard]] ValueOpcode             GetOpcode() const;
+	[[nodiscard]] ValueOpcode             GetOpcode() const { return opcode; }
 	[[nodiscard]] Type                    GetType() const;
 	[[nodiscard]] bool                    MayHaveSideEffects() const;
 	[[nodiscard]] bool                    HasUses() const;
@@ -121,15 +121,11 @@ public:
 	[[nodiscard]] size_t                  NumPhiBlocks() const;
 	[[nodiscard]] Value                   Arg(size_t index) const;
 	[[nodiscard]] Block*                  PhiBlock(size_t index) const;
-	[[nodiscard]] Block*                  Parent() const;
-	[[nodiscard]] const std::vector<Use>& Uses() const;
-	// Runtime indices belong to the resource plan that owns this instruction.
-	[[nodiscard]] uint32_t EvaluationIndex(uint32_t& count) const {
-		if (evaluation_index == UINT32_MAX) {
-			evaluation_index = count++;
-		}
-		return evaluation_index;
-	}
+	[[nodiscard]] Block*                  Parent() const { return parent; }
+	[[nodiscard]] const std::vector<Use>& Uses() const { return uses; }
+	// SRT walker output index, assigned by the resource plan that owns this instruction.
+	[[nodiscard]] uint32_t WalkOutput() const { return walk_output; }
+	void SetWalkOutput(uint32_t index) const { walk_output = index; }
 	// Backend definitions belong to one emission. Wave64 emulation needs one per half.
 	[[nodiscard]] uint32_t Definition(uint32_t half) const { return definitions[half]; }
 	void SetDefinition(uint32_t half, uint32_t definition) const {
@@ -142,6 +138,8 @@ public:
 	void AddPhiOperand(Block* predecessor, Value value);
 	void ReplaceUsesWith(Value replacement, bool preserve = true);
 	void Invalidate();
+	// Drops the arguments without updating their use lists, for teardown that destroys both sides.
+	void DropArgs();
 
 	template <typename T>
 	requires(sizeof(T) <= sizeof(uint64_t) && std::is_trivially_copyable_v<T>)
@@ -161,6 +159,8 @@ public:
 private:
 	friend void EliminateDeadCode(const std::vector<Block*>& blocks);
 
+	template <typename Self>
+	static auto& ArgSlot(Self& self, size_t index);
 	void AddUse(Inst* used, size_t operand);
 	void RemoveUse(Inst* used, size_t operand);
 	void ClearArgs();
@@ -171,7 +171,7 @@ private:
 	ValueOpcode         opcode;
 	uint8_t             num_args;
 	bool                live = false;
-	mutable uint32_t    evaluation_index = UINT32_MAX;
+	mutable uint32_t    walk_output = UINT32_MAX;
 	mutable std::array<uint32_t, 2> definitions {};
 	uint64_t            flags;
 	Block*              parent = nullptr;
@@ -185,4 +185,44 @@ private:
 
 static_assert(sizeof(Inst) <= 120, "Inst operand and backend storage unintentionally increased");
 
+// Called per operand by the IR passes and the resource walker; inline, as the build has no LTO.
+inline bool Value::IsIdentity() const {
+	return type == Type::Opaque && inst->GetOpcode() == ValueOpcode::Identity;
+}
+
+inline bool Value::IsPhi() const {
+	return type == Type::Opaque && inst->GetOpcode() == ValueOpcode::Phi;
+}
+
+inline Inst* Value::ResolveInstruction() const {
+	EXIT_IF(type != Type::Opaque);
+	return IsIdentity() ? inst->Arg(0).ResolveInstruction() : inst;
+}
+
+inline Value Value::Resolve() const {
+	return IsIdentity() ? inst->Arg(0).Resolve() : *this;
+}
+
+inline uint32_t Value::U32() const {
+	EXIT_IF(type != Type::U32);
+	return imm_u32;
+}
+
+inline size_t Inst::NumArgs() const {
+	return num_args == PhiArity ? phi_args.size() : num_args;
+}
+
+template <typename Self>
+auto& Inst::ArgSlot(Self& self, size_t index) {
+	EXIT_IF(index >= self.NumArgs());
+	if (self.num_args <= InlineArity) {
+		return self.fixed_args[index];
+	}
+	return self.num_args == PhiArity ? self.phi_args[index].second : self.large_args[index];
+}
+
+inline Value Inst::Arg(size_t index) const {
+	return ArgSlot(*this, index);
+}
+
 } // namespace Libs::Graphics::ShaderRecompiler::IR

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

@@ -123,6 +123,7 @@ BufferAccess BufferAccessOf(ValueOpcode opcode) {
 		case ValueOpcode::BufferAtomicISub32:
 		case ValueOpcode::BufferAtomicSMin32:
 		case ValueOpcode::BufferAtomicUMin32:
+		case ValueOpcode::BufferAtomicUMin64:
 		case ValueOpcode::BufferAtomicSMax32:
 		case ValueOpcode::BufferAtomicUMax32:
 		case ValueOpcode::BufferAtomicAnd32:
@@ -140,6 +141,7 @@ uint32_t BufferComponentCount(ValueOpcode opcode) {
 	switch (opcode) {
 		case ValueOpcode::BufferAtomicSwap64:
 		case ValueOpcode::BufferAtomicIAdd64:
+		case ValueOpcode::BufferAtomicUMin64:
 		case ValueOpcode::BufferAtomicAnd64:
 		case ValueOpcode::BufferAtomicOr64:
 		case ValueOpcode::LoadBufferU32x2:

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

@@ -96,6 +96,7 @@ VALUE_OPCODE(WqmU64, U64, U64)
 // Core value operations used to expand architectural side effects.
 VALUE_OPCODE(SelectU1, U1, U1, U1, U1)
 VALUE_OPCODE(SelectF32, F32, U1, F32, F32)
+VALUE_OPCODE(SelectF64, F64, U1, F64, F64)
 VALUE_OPCODE(IAdd32, U32, U32, U32)
 VALUE_OPCODE(IAdd64, U64, U64, U64)
 VALUE_OPCODE(ISub32, U32, U32, U32)
@@ -218,11 +219,13 @@ VALUE_OPCODE(FPFract64, F64, F64)
 
 VALUE_OPCODE(LaneId, U32)
 VALUE_OPCODE(Ballot, U32x4, U1)
+VALUE_OPCODE(LiveMask, U32x4)
 VALUE_OPCODE(ReadFirstLane, U32, U32, U1)
 VALUE_OPCODE(ReadLane, U32, U32, U32)
 VALUE_OPCODE(SubgroupUMinU32, U32, U32)
+VALUE_OPCODE(SubgroupUMaxU32, U32, U32)
 VALUE_OPCODE(WriteLane, U32, U32, U32, U32)
-VALUE_OPCODE(Permlane16U32, U32, U32, U32, U32, U1)
+VALUE_OPCODE(Permlane16U32, U32, U32, U32, U32, U1, U32)
 VALUE_OPCODE(PermuteU32, U32, U32, U32, U1)
 VALUE_OPCODE(BpermuteU32, U32, U32, U32, U1)
 
@@ -266,6 +269,7 @@ VALUE_OPCODE(BufferAtomicIAdd64, U64, BufferResource, U32, U32, U32, U64, U1)
 VALUE_OPCODE(BufferAtomicISub32, U32, BufferResource, U32, U32, U32, U32, U1)
 VALUE_OPCODE(BufferAtomicSMin32, U32, BufferResource, U32, U32, U32, U32, U1)
 VALUE_OPCODE(BufferAtomicUMin32, U32, BufferResource, U32, U32, U32, U32, U1)
+VALUE_OPCODE(BufferAtomicUMin64, U64, BufferResource, U32, U32, U32, U64, U1)
 VALUE_OPCODE(BufferAtomicSMax32, U32, BufferResource, U32, U32, U32, U32, U1)
 VALUE_OPCODE(BufferAtomicUMax32, U32, BufferResource, U32, U32, U32, U32, U1)
 VALUE_OPCODE(BufferAtomicAnd32, U32, BufferResource, U32, U32, U32, U32, U1)
@@ -308,8 +312,8 @@ VALUE_OPCODE(SharedAtomicOr32, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicOr64, Void, U32, U64, U1)
 VALUE_OPCODE(SharedAtomicXor32, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicMaskedOr32, Void, U32, U32, U32, U1)
-VALUE_OPCODE(DataAppend, U32, U32, U1)
-VALUE_OPCODE(DataConsume, U32, U32, U1)
+VALUE_OPCODE(DataAppend, U32, U32, U32, U32, U1)
+VALUE_OPCODE(DataConsume, U32, U32, U32, U32, U1)
 VALUE_OPCODE(SwizzleU32, U32, U32, U32, U1)
 
 VALUE_OPCODE(ImageQueryDimensions, U32x4, ImageResource, ImageAddress)

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

@@ -167,6 +167,16 @@ bool ReplaceBinaryIdentity(Inst& inst, Type type, uint64_t identity) {
 	return false;
 }
 
+// Whether one operand is the other's negation.
+bool IsComplementPair(const Inst& inst, ValueOpcode negation) {
+	for (uint32_t arg = 0; arg < 2u; ++arg) {
+		const auto* negate = Arg(inst, arg).TryInstruction();
+		if (negate != nullptr && negate->GetOpcode() == negation &&
+		    Arg(*negate, 0) == Arg(inst, arg ^ 1u)) return true;
+	}
+	return false;
+}
+
 bool FoldSelect(Inst& inst) {
 	const auto condition   = Arg(inst, 0);
 	const auto true_value  = Arg(inst, 1);
@@ -626,8 +636,9 @@ void FoldInstruction(Block& block, Block::iterator instruction,
 		case ValueOpcode::BitwiseXor32:
 			if (!FoldU32(inst, [opcode = inst.GetOpcode()](uint32_t a, uint32_t b) {
 				    return opcode == ValueOpcode::BitwiseOr32 ? a | b : a ^ b;
-			    })) {
-				ReplaceBinaryIdentity(inst, Type::U32, 0u);
+			    }) && !ReplaceBinaryIdentity(inst, Type::U32, 0u) &&
+			    IsComplementPair(inst, ValueOpcode::BitwiseNot32)) {
+				Replace(inst, Value(UINT32_MAX));
 			}
 			return;
 		case ValueOpcode::BitwiseNot32: {
@@ -802,6 +813,8 @@ void FoldInstruction(Block& block, Block::iterator instruction,
 					Replace(inst, lhs.U1() ? lhs : rhs);
 				} else if (IsImmediate(rhs, Type::U1)) {
 					Replace(inst, rhs.U1() ? rhs : lhs);
+				} else if (IsComplementPair(inst, ValueOpcode::LogicalNot)) {
+					Replace(inst, Value(true));
 				} else {
 					const auto* left = lhs.TryInstruction();
 					const auto* right = rhs.TryInstruction();
@@ -892,8 +905,10 @@ private:
 		if (inst.GetOpcode() != ValueOpcode::CompositeExtractU32x4 ||
 		    !Immediate(Arg(inst, 1), 0u)) return {};
 		const auto* source = Arg(inst, 0).TryInstruction();
-		return source != nullptr && source->GetOpcode() == ValueOpcode::Ballot
-		           ? Arg(*source, 0) : Value {};
+		if (source == nullptr) return {};
+		// Helper lanes keep the per-lane EXEC, so the live mask projects to true.
+		if (source->GetOpcode() == ValueOpcode::LiveMask) return Value(true);
+		return source->GetOpcode() == ValueOpcode::Ballot ? Arg(*source, 0) : Value {};
 	}
 
 	bool CanProject(Value value) {

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

@@ -99,89 +99,66 @@ Value GetRealValue(PhiMap& phi_map, Value source, uint32_t lane, uint32_t wave_s
 	return Value(&phi);
 }
 
-// A full-EXEC prefix scan with both half-wave endpoint reads is a reduction.
+// A full-EXEC prefix scan with both half-wave endpoint reads is a reduction. Wave64 first joins
+// the two rows of each half; a wave32 half is one row.
 // Fragment subgroups may have no physical invocation at either endpoint.
-class UMinReductionMatcher {
+class WaveReductionMatcher {
 public:
-	explicit UMinReductionMatcher(Block* block): block(block) {}
+	WaveReductionMatcher(const Block* block, ValueOpcode operation)
+	    : block(block), operation(operation) {}
 
-	Value Match(const Inst& minimum) {
-		const auto* low  = Get(minimum.Arg(0), ValueOpcode::ReadLane);
-		const auto* high = Get(minimum.Arg(1), ValueOpcode::ReadLane);
+	Value Match(const Inst& reduction, uint32_t wave_size) const {
+		const auto* low  = Get(reduction.Arg(0), ValueOpcode::ReadLane);
+		const auto* high = Get(reduction.Arg(1), ValueOpcode::ReadLane);
 		if (low == nullptr || high == nullptr || low->Arg(0).Resolve() != high->Arg(0).Resolve()) {
 			return {};
 		}
-		const auto lane = [](Value value, uint32_t expected) {
+		const auto lane = [&](Value value, uint32_t expected) {
 			value = value.Resolve();
 			return value.IsImmediate() && value.GetType() == Type::U32 &&
-			       value.U32() % 64u == expected;
+			       value.U32() % wave_size == expected;
 		};
-		if (!((lane(low->Arg(1), 31) && lane(high->Arg(1), 63)) ||
-		      (lane(low->Arg(1), 63) && lane(high->Arg(1), 31)))) {
+		const auto half = wave_size / 2u;
+		if (!((lane(low->Arg(1), half - 1u) && lane(high->Arg(1), wave_size - 1u)) ||
+		      (lane(low->Arg(1), wave_size - 1u) && lane(high->Arg(1), half - 1u)))) {
 			return {};
 		}
-		const auto* joined = Get(Unwrap(low->Arg(0)), ValueOpcode::UMin32);
+		if (wave_size == 32u) return MatchScan(low->Arg(0));
+		const auto* joined = Get(low->Arg(0), operation);
 		if (joined == nullptr) return {};
 		for (size_t index = 0; index < 2; index++) {
-			auto attempt = *this;
-			const auto* cross = attempt.Get(attempt.Unwrap(joined->Arg(index)), ValueOpcode::Permlane16U32);
+			const auto* cross = Get(joined->Arg(index), ValueOpcode::Permlane16U32);
 			if (cross == nullptr) continue;
-			const auto flags = cross->Flags<PermlaneFlags>();
-			auto prefix = attempt.Unwrap(joined->Arg(1 - index));
+			const auto flags  = cross->Flags<PermlaneFlags>();
+			const auto prefix = joined->Arg(1 - index).Resolve();
 			if (!flags.x16 || flags.fetch_inactive || !flags.bound_control ||
 			    cross->Arg(1).Resolve() != Value(UINT32_MAX) ||
-			    cross->Arg(2).Resolve() != Value(UINT32_MAX) ||
-			    !attempt.FullExec(cross->Arg(3)) || attempt.Unwrap(cross->Arg(0)) != prefix) {
+			    cross->Arg(2).Resolve() != Value(UINT32_MAX) || !FullExec(cross->Arg(3)) ||
+			    cross->Arg(0).Resolve() != prefix) {
 				continue;
 			}
-			for (const uint32_t shift: {8u, 4u, 2u, 1u}) {
-				prefix = attempt.MatchRow(prefix, shift);
-			}
-			prefix = attempt.Unwrap(prefix);
-			if (prefix.IsEmpty() || attempt.saved_exec.IsEmpty()) return prefix;
-			const auto* seed = attempt.Get(prefix, ValueOpcode::SelectU32);
-			if (seed != nullptr && seed->Arg(2).Resolve() == Value(UINT32_MAX)) {
-				const auto predicate = seed->Arg(0).Resolve();
-				const auto* subset = attempt.Get(predicate, ValueOpcode::LogicalAnd);
-				if (predicate == attempt.saved_exec ||
-				    (subset != nullptr && (subset->Arg(0).Resolve() == attempt.saved_exec ||
-				                           subset->Arg(1).Resolve() == attempt.saved_exec))) {
-					return prefix;
-				}
-			}
+			if (const auto seed = MatchScan(prefix); !seed.IsEmpty()) return seed;
 		}
 		return {};
 	}
 
 private:
+	// The row scan's input.
+	Value MatchScan(Value prefix) const {
+		for (const uint32_t shift: {8u, 4u, 2u, 1u}) {
+			prefix = MatchRow(prefix, shift);
+		}
+		return prefix;
+	}
+
 	Inst* Get(Value value, ValueOpcode opcode) const {
 		auto* inst = value.Resolve().TryInstruction();
 		return inst != nullptr && inst->Parent() == block && inst->GetOpcode() == opcode
 		           ? inst : nullptr;
 	}
 
-	bool FullExec(Value predicate) {
-		predicate = predicate.Resolve();
-		if (predicate == Value(true)) return true;
-		const auto* either = Get(predicate, ValueOpcode::LogicalOr);
-		if (either == nullptr) return false;
-		for (size_t index = 0; index < 2; index++) {
-			const auto* negate = Get(either->Arg(index), ValueOpcode::LogicalNot);
-			if (negate == nullptr || negate->Arg(0).Resolve() != either->Arg(1 - index).Resolve()) continue;
-			const auto original = negate->Arg(0).Resolve();
-			if (saved_exec.IsEmpty()) saved_exec = original;
-			return saved_exec == original;
-		}
-		return false;
-	}
-
-	Value Unwrap(Value value) {
-		while (const auto* select = Get(value, ValueOpcode::SelectU32)) {
-			if (!FullExec(select->Arg(0))) break;
-			value = select->Arg(1);
-		}
-		return value.Resolve();
-	}
+	// Constant propagation folds the x | ~x EXEC of s_orn2_saveexec to true.
+	static bool FullExec(Value predicate) { return predicate.Resolve() == Value(true); }
 
 	static bool IsRowShift(const Inst& inst, uint32_t shift) {
 		const auto flags = inst.Flags<DppMoveFlags>();
@@ -189,22 +166,24 @@ private:
 		       flags.bank_mask == 0xf && !flags.fetch_inactive && !flags.bound_control && !flags.dpp8;
 	}
 
-	Value MatchRow(Value value, uint32_t shift) {
-		const auto* update = Get(Unwrap(value), ValueOpcode::DppUpdateU32);
+	Value MatchRow(Value value, uint32_t shift) const {
+		const auto* update = Get(value, ValueOpcode::DppUpdateU32);
 		if (update == nullptr || !IsRowShift(*update, shift) || !FullExec(update->Arg(2))) return {};
-		const auto* minimum = Get(update->Arg(0), ValueOpcode::UMin32);
-		if (minimum == nullptr) return {};
+		const auto* combined = Get(update->Arg(0), operation);
+		if (combined == nullptr) return {};
 		for (size_t index = 0; index < 2; index++) {
-			const auto* move = Get(minimum->Arg(index), ValueOpcode::DppMoveU32);
+			const auto* move = Get(combined->Arg(index), ValueOpcode::DppMoveU32);
 			if (move == nullptr || !IsRowShift(*move, shift) || !FullExec(move->Arg(1))) continue;
-			const auto previous = Unwrap(minimum->Arg(1 - index));
-			if (Unwrap(move->Arg(0)) == previous && Unwrap(update->Arg(1)) == previous) return previous;
+			const auto previous = combined->Arg(1 - index).Resolve();
+			if (move->Arg(0).Resolve() == previous && update->Arg(1).Resolve() == previous) {
+				return previous;
+			}
 		}
 		return {};
 	}
 
-	Block* block;
-	Value  saved_exec;
+	const Block* block;
+	ValueOpcode  operation;
 };
 
 } // namespace
@@ -217,18 +196,22 @@ ReadLaneStats EliminateReadLane(Program& program, uint32_t wave_size) {
 
 	for (auto* block: program.blocks) {
 		for (auto& inst: *block) {
-			if (wave_size == 64u && inst.GetOpcode() == ValueOpcode::UMin32) {
-				const auto seed = UMinReductionMatcher(block).Match(inst);
+			const auto opcode = inst.GetOpcode();
+			if (opcode == ValueOpcode::UMin32 || opcode == ValueOpcode::UMax32) {
+				const auto seed = WaveReductionMatcher(block, opcode).Match(inst, wave_size);
 				if (!seed.IsEmpty()) {
 					const auto position = std::find_if(block->begin(), block->end(),
 					                                  [&](const Inst& candidate) { return &candidate == &inst; });
-					auto& reduction = *block->PrependNewInst(position, ValueOpcode::SubgroupUMinU32, {seed});
+					const auto reduce    = opcode == ValueOpcode::UMin32
+					                           ? ValueOpcode::SubgroupUMinU32
+					                           : ValueOpcode::SubgroupUMaxU32;
+					auto&      reduction = *block->PrependNewInst(position, reduce, {seed});
 					inst.ReplaceUsesWith(Value(&reduction));
 					stats.rewritten_reductions++;
 				}
 				continue;
 			}
-			if (inst.GetOpcode() != ValueOpcode::ReadLane) {
+			if (opcode != ValueOpcode::ReadLane) {
 				continue;
 			}
 			const auto selector = inst.Arg(1).Resolve();

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

@@ -3,6 +3,7 @@
 #include "common/assert.h"
 #include "graphics/guest_gpu/gpu_format.h"
 #include "graphics/shader/recompiler/BufferFormat.h"
+#include "graphics/shader/recompiler/ir/Serialize.h"
 #include "graphics/shader/recompiler/ir/ShaderIR.h"
 #include "graphics/shader/shaderBindings.h"
 
@@ -186,35 +187,6 @@ const DescriptorSource* Source(const ResourcePlan& program, uint32_t source) {
 	return &program.descriptor_sources[source];
 }
 
-void MarkCleanReads(const std::unordered_set<const Inst*>& planned_reads,
-                    const DescriptorSource* source, Value extra = {}) {
-	if (source == nullptr && extra.IsEmpty()) {
-		return;
-	}
-	std::vector<Value> pending;
-	if (source != nullptr) {
-		pending.assign(source->dwords.begin(), source->dwords.begin() + source->dword_count);
-	}
-	if (!extra.IsEmpty()) pending.push_back(extra);
-	std::vector<const Inst*> visited;
-	while (!pending.empty()) {
-		auto* inst = pending.back().Resolve().TryInstruction();
-		pending.pop_back();
-		if (inst == nullptr || std::ranges::find(visited, inst) != visited.end()) {
-			continue;
-		}
-		visited.push_back(inst);
-		if (planned_reads.contains(inst)) {
-			auto flags = inst->Flags<SrtReadFlags>();
-			flags.clean = 1u;
-			inst->SetFlags(flags);
-		}
-		for (size_t arg = 0; arg < inst->NumArgs(); arg++) {
-			pending.push_back(inst->Arg(arg));
-		}
-	}
-}
-
 bool ReadTable(uint64_t base, uint64_t size, uint64_t offset,
                const SrtRuntime& runtime, std::span<uint32_t> words) {
 	const auto count = std::min<uint64_t>(words.size(), offset < size ? (size - offset) / 4u : 0u);
@@ -309,22 +281,22 @@ bool CanReadIndexedSelector(const ShaderBufferResource& material, uint32_t offse
 	           uint64_t {UINT32_MAX} + 1u;
 }
 
-bool EvaluateIndirectCount(const DescriptorSource::IndirectDescriptor& indirect,
-                           const SrtRuntime& runtime, SrtWalker& clean, uint32_t& count) {
+bool EvaluateIndirectCount(const ResourcePlan& program, const DescriptorSource::IndirectDescriptor& indirect,
+                           const SrtRuntime& runtime, uint32_t& count) {
 	if (indirect.workgroup_axis != UINT32_MAX) {
 		if (runtime.workgroup_counts.size() != 3u || indirect.workgroup_axis >= 3u) return false;
 		count = std::ranges::find(runtime.workgroup_counts, 0u) != runtime.workgroup_counts.end()
 		    ? 0u : runtime.workgroup_counts[indirect.workgroup_axis];
 		return true;
 	}
-	if (!clean.Evaluate(indirect.key_count, count)) return false;
+	if (!WalkedValue(program, indirect.key_count, count)) return false;
 	if (indirect.signed_count && std::bit_cast<int32_t>(count) < 0) count = 0;
 	return true;
 }
 
-bool ResolveSelectorKeys(const DescriptorSource::IndirectDescriptor& indirect,
+bool ResolveSelectorKeys(const ResourcePlan& program, const DescriptorSource::IndirectDescriptor& indirect,
                          const DescriptorValue& material_value, const SrtRuntime& runtime,
-                         SrtWalker& clean, std::vector<uint32_t>& keys) {
+                         std::vector<uint32_t>& keys) {
 	const auto* selector = &*indirect.selector;
 	keys.clear();
 	const auto read_selector = [&](uint64_t base, uint64_t size, uint64_t offset,
@@ -359,11 +331,23 @@ bool ResolveSelectorKeys(const DescriptorSource::IndirectDescriptor& indirect,
 		keys.resize(count);
 		return true;
 	};
+	// Reads the key of each record whose rank is set.
+	const auto read_ranks = [&](uint64_t base, uint64_t size, uint64_t first, uint64_t step,
+	                            uint32_t ranks) {
+		keys.reserve(std::popcount(ranks));
+		for (; ranks != 0u; ranks &= ranks - 1u) {
+			const auto offset = first + std::countr_zero(ranks) * step;
+			uint32_t key = 0;
+			if (offset > UINT32_MAX || !read_selector(base, size, offset, {&key, 1})) return false;
+			keys.push_back(key);
+		}
+		return true;
+	};
 	if (selector->rows) {
 		const auto& rows = *selector->rows;
 		DescriptorValue count_value;
 		ShaderBufferResource counts, material;
-		if (!clean.EvaluateDescriptor(rows.source, count_value) ||
+		if (!WalkedDescriptor(program, rows.source, count_value) ||
 		    !DecodeBufferDescriptor(count_value, counts) ||
 		    !DecodeBufferDescriptor(material_value, material) || selector->stride != 4u ||
 		    selector->offset != 0u || rows.stride == 0u || (rows.stride & 3u) != 0u ||
@@ -392,7 +376,7 @@ bool ResolveSelectorKeys(const DescriptorSource::IndirectDescriptor& indirect,
 				if (candidate_end - first > MaxIndirectDescriptorProbes) break;
 				end = candidate_end;
 			}
-			// Coalesce small gaps, but interpret only selected prefixes. The strict
+			// Coalesce small gaps, but interpret only selected prefixes. The specialization
 			// reader captures the entire fetched span for writable-alias rejection.
 			batch.resize(end - first);
 			if (!ReadScalarTable(material.Base48(), material.GetSize(), first * 4u, runtime, batch))
@@ -405,19 +389,21 @@ bool ResolveSelectorKeys(const DescriptorSource::IndirectDescriptor& indirect,
 		}
 	} else if (selector->bitset) {
 		ShaderBufferResource material;
-		std::array<uint32_t, 4> masks {};
+		std::array<uint32_t, 8> storage {};
+		const auto record_words = std::min<size_t>(selector->stride / 4u, storage.size());
+		const auto masks = std::span(storage).first(record_words);
 		uint32_t first = 0, count = 0;
 		if (!DecodeBufferDescriptor(material_value, material) ||
-		    !clean.Evaluate(indirect.selector_first, first) ||
-		    !EvaluateIndirectCount(indirect, runtime, clean, count) || count != masks.size() * 32u ||
-		    selector->stride != sizeof(masks) || selector->offset != 0u ||
-		    material.Type() != 0u || material.Stride() != sizeof(masks) ||
+		    !WalkedValue(program, indirect.selector_first, first) ||
+		    !EvaluateIndirectCount(program, indirect, runtime, count) || count != masks.size() * 32u ||
+		    selector->stride != masks.size_bytes() || selector->offset != 0u ||
+		    material.Type() != 0u || material.Stride() != selector->stride ||
 		    material.SwizzleEnabled() || material.AddTid() || material.OutOfBounds() != 0u ||
 		    (material.Base48() & 3u) != 0u) return false;
 		if (material.Format() != Prospero::BufferFormat::kInvalid) {
 			const auto words = material.GetSize() / sizeof(uint32_t);
-			// Bitmap storage can grow with resolution while its candidate domain stays
-			// at 128 keys. Bound scratch memory, preserving each coherent read's capture.
+			// Bitmap storage can grow with resolution while its candidate domain stays at 32 keys
+			// per record word. Bound scratch memory, preserving each coherent read's capture.
 			for (uint64_t first_word = 0; first_word < words; first_word += keys.size()) {
 				keys.resize(std::min<uint64_t>(words - first_word, MaxIndirectDescriptorProbes));
 				if (!ReadScalarTable(material.Base48(), material.GetSize(), first_word * 4u,
@@ -462,8 +448,8 @@ bool ResolveSelectorKeys(const DescriptorSource::IndirectDescriptor& indirect,
 			base = material.Base48();
 			size = material.GetSize();
 		}
-		if (!clean.Evaluate(indirect.selector_first, first) ||
-		    !EvaluateIndirectCount(indirect, runtime, clean, count) || count > MaxIndirectDescriptorProbes)
+		if (!WalkedValue(program, indirect.selector_first, first) ||
+		    !EvaluateIndirectCount(program, indirect, runtime, count) || count > MaxIndirectDescriptorProbes)
 			return false;
 		// SMEM uses its instruction's byte stride; V# stride only supplies its bounds.
 		const auto step = selector->stride != 0u ? selector->stride : material.Stride();
@@ -473,35 +459,48 @@ bool ResolveSelectorKeys(const DescriptorSource::IndirectDescriptor& indirect,
 		     uint64_t {first} + count > material.NumRecords())) return false;
 		const auto last = count != 0u ? uint64_t {first} + count - 1u : 0u;
 		if (last > UINT32_MAX) return false;
+		// Out-of-bounds scalar reads return zero.
 		const auto end = count != 0u ? last * step + selector->offset + 4u : 0u;
-		if (end > uint64_t {UINT32_MAX} + 1u ||
-		    (indirect.workgroup_axis == UINT32_MAX && end > size))
-			return false;
-		if (!read_keys(base, size, uint64_t {first} * step + selector->offset, step, count))
+		if (end > uint64_t {UINT32_MAX} + 1u) return false;
+		const auto start = uint64_t {first} * step + selector->offset;
+		if (selector->ranks) {
+			// A scan selects the ranks some word of its mask's address class sets.
+			const auto& mask = *selector->ranks;
+			DescriptorValue mask_value;
+			ShaderBufferResource masks;
+			if (count > 32u || !WalkedDescriptor(program, mask.source, mask_value) ||
+			    !DecodeBufferDescriptor(mask_value, masks)) return false;
+			// Each chunk holds whole strides of the power-of-two class, or one of its words.
+			const auto words = mask.stride / 4u;
+			const auto chunk =
+			    words <= MaxIndirectDescriptorProbes ? MaxIndirectDescriptorProbes : 1u;
+			uint32_t ranks = 0;
+			for (auto at = uint64_t {mask.offset}; at + 4u <= masks.GetSize();
+			     at += uint64_t {std::max(chunk, words)} * 4u) {
+				keys.resize(std::min<uint64_t>((masks.GetSize() - at) / 4u, chunk));
+				if (!ReadScalarTable(masks.Base48(), masks.GetSize(), at, runtime, keys))
+					return false;
+				for (size_t word = 0; word < keys.size(); word += words) ranks |= keys[word];
+			}
+			keys.clear();
+			if (count < 32u) ranks &= (1u << count) - 1u;
+			if (!read_ranks(base, size, start, step, ranks)) return false;
+		} else if (!read_keys(base, size, start, step, count)) {
 			return false;
+		}
 	} else if (!indirect.selector_mask.IsEmpty()) {
 		uint32_t mask = 0;
 		uint32_t count = 0;
 		if (material_value.dword_count != 2u ||
-		    !clean.Evaluate(indirect.selector_mask, mask) ||
-		    !EvaluateIndirectCount(indirect, runtime, clean, count) || count == 0u || count > 32u) {
+		    !WalkedValue(program, indirect.selector_mask, mask) ||
+		    !EvaluateIndirectCount(program, indirect, runtime, count) || count == 0u || count > 32u) {
 			return false;
 		}
 		if (count < 32u) mask &= (1u << count) - 1u;
 		const auto material_base =
 		    (static_cast<uint64_t>(material_value.dwords[1]) << 32u) | material_value.dwords[0];
-		keys.reserve(std::popcount(mask));
-		while (mask != 0u) {
-			const auto index = std::countr_zero(mask);
-			const auto offset = static_cast<uint64_t>(selector->offset) +
-			                    static_cast<uint64_t>(index) * selector->stride;
-			if (offset > UINT32_MAX) return false;
-			uint32_t key = 0;
-			if (!read_selector(material_base, UINT64_MAX, static_cast<uint32_t>(offset),
-			                   {&key, 1})) return false;
-			keys.push_back(key);
-			mask &= mask - 1u;
-		}
+		if (!read_ranks(material_base, UINT64_MAX, selector->offset, selector->stride, mask))
+			return false;
 	} else if (selector->stride == 0u) {
 		ShaderBufferResource material;
 		if (!DecodeBufferDescriptor(material_value, material) ||
@@ -531,6 +530,7 @@ bool ResolveSelectorKeys(const DescriptorSource::IndirectDescriptor& indirect,
 		if (selector->bit_offset != 0u || selector->bit_mask != UINT32_MAX || selector->maximum != UINT32_MAX)
 			for (auto& key: keys)
 				key = std::min((key >> selector->bit_offset) & selector->bit_mask, selector->maximum);
+		if (selector->excluded) std::erase(keys, *selector->excluded);
 		std::ranges::sort(keys);
 		keys.erase(std::unique(keys.begin(), keys.end()), keys.end());
 	}
@@ -541,8 +541,7 @@ template <typename Specialization, typename Normalize>
 bool MaterializeIndirectDescriptor(const ResourcePlan&                         program,
                                    uint32_t source_index,
                                    uint32_t resource_index, uint32_t dword_count, uint32_t read_dwords,
-                                   const SrtRuntime& runtime, SrtWalker& clean,
-                                   ResourceSnapshot&             snapshot,
+                                   const SrtRuntime& runtime, ResourceSnapshot& snapshot,
                                    std::vector<DescriptorValue>& descriptors,
                                    std::vector<Specialization>&  specializations,
                                    Normalize&& normalize) {
@@ -605,8 +604,8 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 		if (selector == nullptr) keys.clear();
 		DescriptorValue material_value;
 		if ((selector != nullptr && (byte_keys || !selector_keys->resolved) &&
-		     !clean.EvaluateDescriptor(selector->source, material_value)) ||
-		    !clean.EvaluateDescriptor(indirect.table_source, table_value)) {
+		     !WalkedDescriptor(program, selector->source, material_value)) ||
+		    !WalkedDescriptor(program, indirect.table_source, table_value)) {
 			return false;
 		}
 		if (table_value.dword_count == 2u) {
@@ -659,7 +658,7 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 			                                       : uint64_t {1} << 32u;
 			if (!indirect.key_count.IsEmpty()) {
 				uint32_t key_count = 0u;
-				if (!EvaluateIndirectCount(indirect, runtime, clean, key_count)) return false;
+				if (!EvaluateIndirectCount(program, indirect, runtime, key_count)) return false;
 				domain = std::min(domain, uint64_t {key_count});
 			}
 			if (indexed_pointer &&
@@ -771,8 +770,8 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 		if (selector == nullptr) {
 			uint32_t key_count = 0;
 			uint32_t first_key = 0;
-			if (!EvaluateIndirectCount(indirect, runtime, clean, key_count)) return false;
-			if (!indirect.selector_first.IsEmpty() && !clean.Evaluate(indirect.selector_first, first_key))
+			if (!EvaluateIndirectCount(program, indirect, runtime, key_count)) return false;
+			if (!indirect.selector_first.IsEmpty() && !WalkedValue(program, indirect.selector_first, first_key))
 				return false;
 			if (indirect.table_stride == 0u || key_count > MaxIndirectDescriptorProbes ||
 			    (indirect.selector_first.IsEmpty() && key_count != 0u &&
@@ -785,10 +784,16 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 			std::iota(keys.begin(), keys.end(), first_key);
 			if (!keys.empty() && keys.back() < keys.front())
 				std::rotate(keys.begin(), keys.begin() + (UINT32_MAX - first_key + 1ull), keys.end());
+			if (indirect.workgroup_shift != 0u) {
+				for (auto& key: keys)
+					key = static_cast<uint32_t>(uint64_t {key} * indirect.workgroup_multiplier >>
+					                            indirect.workgroup_shift);
+				keys.erase(std::unique(keys.begin(), keys.end()), keys.end());
+			}
 		} else {
 			if (!indirect.selector_mask.IsEmpty() && table_value.dword_count != 2u) return false;
 			if (!selector_keys->resolved) {
-				if (!ResolveSelectorKeys(indirect, material_value, runtime, clean, keys)) return false;
+				if (!ResolveSelectorKeys(program, indirect, material_value, runtime, keys)) return false;
 				selector_keys->resolved = true;
 			}
 		}
@@ -827,7 +832,7 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 			                     std::span(candidate.dwords).first(read_dwords))) {
 				return read_failed("table descriptor read", key, 0u, table_base, table_offset, descriptor_bytes);
 			}
-		} else if (!clean.EvaluateDescriptor(sources[entry], candidate)) {
+		} else if (!WalkedDescriptor(program, sources[entry], candidate)) {
 			return false;
 		}
 		if (!normalize(candidate))
@@ -1070,6 +1075,26 @@ static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSna
 			return dimension == Decoder::ImageDimension::Dim1D ||
 			       dimension == Decoder::ImageDimension::Dim2D;
 		};
+		// The descriptor's type selects the address layout. A 2D or 3D candidate reads the leading
+		// coordinates of an equal or larger dimension, but an array with more than one slice reads
+		// the next as its slice. Derivative, LOD, clamp and mip operands shift with the dimension.
+		const auto leading = [&](Decoder::ImageDimension dimension, const DescriptorValue& value) {
+			const bool sliced =
+			    static_cast<Prospero::ImageType>((value.dwords[3] >> 28u) & 0xfu) ==
+			        Prospero::ImageType::kColor2DArray &&
+			    (value.dwords[4] & 0x1fffu) != ((value.dwords[4] >> 16u) & 0x1fffu);
+			return !sliced &&
+			       (dimension == Decoder::ImageDimension::Dim2D || dimension == base.dimension);
+		};
+		const bool leading_coordinates = (base.dimension == Decoder::ImageDimension::Dim2D ||
+		                                  base.dimension == Decoder::ImageDimension::Dim3D) &&
+		    leading(image_class.dimension, snapshot.images[exemplar]) &&
+		    std::ranges::none_of(program.memory_info, [&](const MemoryInfo& memory) {
+			    return memory.kind == ResourceKind::Image && memory.resource == root_index &&
+			           (memory.image_has_mip || (memory.image_sample_flags &
+			               (Decoder::ImageSampleFlagLod | Decoder::ImageSampleFlagDerivative |
+			                Decoder::ImageSampleFlagLodClamp)) != 0u);
+		    });
 		for (uint32_t candidate = 0; candidate < specialization.images.size(); candidate++) {
 			auto& image = specialization.images[candidate];
 			if (image.indirect_root != root_index) {
@@ -1089,8 +1114,10 @@ static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSna
 			    base.dimension == Decoder::ImageDimension::Dim2D &&
 			    !image.cube && !image_class.cube && is_1d_or_2d(image.dimension) &&
 			    is_1d_or_2d(image_class.dimension);
+			const bool prefix_coordinates =
+			    leading_coordinates && leading(image.dimension, snapshot.images[candidate]);
 			if ((!same_coordinates && !(is_2d(image.dimension) && is_2d(image_class.dimension)) &&
-			     !unfiltered_coordinates) ||
+			     !unfiltered_coordinates && !prefix_coordinates) ||
 			    image.conversion_format != image_class.conversion_format ||
 			    (!storage_swizzle && image.shader_swizzle != image_class.shader_swizzle)) {
 				return SpecializationFail(
@@ -1193,14 +1220,16 @@ static std::vector<ResourceBlock> ResourceControlFlow(const Program& program, co
 				block.sources.push_back(program.info.buffers.at(memory.resource).source);
 			} else {
 				block.sources.push_back(program.info.images.at(memory.resource).source);
+				block.images.push_back(memory.resource);
 				if (image.needs_sampler) {
 					block.sources.push_back(program.info.samplers.at(memory.sampler).source);
 				}
 			}
 		}
-		std::ranges::sort(block.sources);
-		block.sources.erase(std::unique(block.sources.begin(), block.sources.end()),
-		                    block.sources.end());
+		for (auto* list: {&block.sources, &block.images}) {
+			std::ranges::sort(*list);
+			list->erase(std::unique(list->begin(), list->end()), list->end());
+		}
 	}
 	for (auto& block: blocks) block.condition = predicate(block.condition);
 	if (std::ranges::none_of(
@@ -1341,6 +1370,17 @@ static UniformFillPlan AnalyzeUniformFill(const Program& program) {
 	return result;
 }
 
+// State an extracted or loaded plan derives from its fields: one key set per selector group, which
+// extraction numbers densely in order of first use (false when they are not), and the walker.
+static bool DerivePlanState(ResourcePlan& plan) {
+	for (const auto group: plan.selector_groups) {
+		if (group == plan.selector_keys.size()) plan.selector_keys.emplace_back();
+		if (group != UINT32_MAX && group >= plan.selector_keys.size()) return false;
+	}
+	if (plan.srt_plan_complete && plan.resource_tracking_complete) CompileSrtWalker(plan);
+	return true;
+}
+
 ResourcePlan ExtractResourcePlan(const Program& program) {
 	ResourcePlan plan;
 	plan.stage                      = program.stage;
@@ -1380,16 +1420,13 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 		if (const auto found = cloned.find(source); found != cloned.end()) {
 			return Value(found->second);
 		}
-		if (source->GetOpcode() == ValueOpcode::LoadBufferU32) {
-			plan.requires_specialization_memory = true;
-			plan.capture_specialization_reads   = true;
-		}
+		if (source->GetOpcode() == ValueOpcode::LoadBufferU32) plan.capture_specialization_reads = true;
 		auto& target =
 		    plan.value_storage.emplace_back(source->GetOpcode(), source->Flags<uint64_t>());
 		if (source->GetOpcode() == ValueOpcode::LoadAddressU32 ||
 		    source->GetOpcode() == ValueOpcode::ReadConstBuffer ||
 		    source->GetOpcode() == ValueOpcode::LoadBufferU32) {
-			target.SetFlags(SrtReadFlags {.index = source->Flags<MemoryFlags>().index});
+			target.SetFlags(MemoryFlags {.index = source->Flags<MemoryFlags>().index});
 		}
 		cloned.emplace(source, &target);
 		if (source->GetOpcode() == ValueOpcode::Phi) {
@@ -1411,7 +1448,6 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 		target.dword_count    = source.dword_count;
 		target.indirect_descriptor = source.indirect_descriptor;
 		if (target.indirect_descriptor.has_value()) {
-			plan.requires_specialization_memory = true;
 			capture_indirect_reads = true;
 			target.indirect_descriptor->key_count = Clone(target.indirect_descriptor->key_count);
 			target.indirect_descriptor->selector_first =
@@ -1425,17 +1461,18 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 	}
 	// Destination tables share keys when their complete selector domains agree.
 	// Resolve those domains lazily at dispatch time, without repeated guest reads.
+	std::vector<uint32_t> group_sources; // The first source of each group.
 	plan.selector_groups.resize(plan.descriptor_sources.size(), UINT32_MAX);
 	for (uint32_t source = 0; source < plan.descriptor_sources.size(); ++source) {
 		const auto& descriptor = program.descriptor_sources[source].indirect_descriptor;
 		if (!descriptor || !descriptor->selector) continue;
 		const auto& a = *descriptor;
 		uint32_t group = 0;
-		for (; group < plan.selector_keys.size(); ++group) {
-			const auto& b = *program.descriptor_sources[plan.selector_keys[group].source].indirect_descriptor;
+		for (; group < group_sources.size(); ++group) {
+			const auto& b = *program.descriptor_sources[group_sources[group]].indirect_descriptor;
 			if (EquivalentSelectorDomain(program, a, b)) break;
 		}
-		if (group == plan.selector_keys.size()) plan.selector_keys.push_back({.source = source});
+		if (group == group_sources.size()) group_sources.push_back(source);
 		plan.selector_groups[source] = group;
 	}
 	plan.srt_reads.reserve(program.srt_reads.size());
@@ -1523,57 +1560,16 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 	plan.uniform_fill.element_count = Clone(plan.uniform_fill.element_count);
 	plan.capture_specialization_reads |=
 	    capture_indirect_reads || capture_control_reads || !plan.source_reads.empty();
-	if (plan.capture_specialization_reads) {
-		// Every former flat alias now targets its retained raw read. Mark those reads
-		// on the normalized graph, so discarded Phi/EXEC dependencies stay discarded.
-		std::unordered_set<const Inst*> planned_reads;
-		for (const auto& read: plan.srt_reads) {
-			auto* inst = read.value.ResolveInstruction();
-			EXIT_IF(inst == nullptr || (inst->GetOpcode() != ValueOpcode::LoadAddressU32 &&
-			                           inst->GetOpcode() != ValueOpcode::ReadConstBuffer));
-			planned_reads.insert(inst);
-		}
-		for (const auto& source: plan.descriptor_sources) {
-			if (!source.indirect_descriptor.has_value()) continue;
-			const auto& indirect = *source.indirect_descriptor;
-			if (indirect.selector) {
-				MarkCleanReads(planned_reads, Source(plan, indirect.selector->source));
-				if (indirect.selector->rows)
-					MarkCleanReads(planned_reads, Source(plan, indirect.selector->rows->source));
-			}
-			MarkCleanReads(planned_reads, nullptr, indirect.selector_mask);
-			MarkCleanReads(planned_reads, nullptr, indirect.selector_first);
-			MarkCleanReads(planned_reads, nullptr, indirect.key_count);
-			if (indirect.sources.empty()) {
-				MarkCleanReads(planned_reads, Source(plan, indirect.table_source));
-			} else {
-				for (const auto candidate: indirect.sources) {
-					MarkCleanReads(planned_reads, Source(plan, candidate));
-				}
-			}
-		}
-		// Clean writable addresses prove that shader writes cannot overlap captured reads.
-		for (const auto& buffer: plan.info.buffers) {
-			if (buffer.written) MarkCleanReads(planned_reads, Source(plan, buffer.source));
-		}
-		for (const auto& image: plan.info.images) {
-			if (image.written) MarkCleanReads(planned_reads, Source(plan, image.source));
-		}
-	}
 	if (capture_indirect_reads || !plan.source_reads.empty())
 		plan.resource_tracking_complete &= !program.has_address_writes;
-	CompileSrtWalker(plan);
+	DerivePlanState(plan);
 	return plan;
 }
 
 bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime,
                           ResourceSnapshot& snapshot, ResourceSpecialization& specialization) {
 	for (auto& keys: program.selector_keys) keys.resolved = false;
-	if (!program.resource_tracking_complete ||
-	    ((program.requires_specialization_memory || !program.source_reads.empty()) &&
-	     runtime.read_specialization_memory == nullptr)) {
-		return false;
-	}
+	if (!program.resource_tracking_complete) return false;
 	const auto uses_workgroup_count = [](const DescriptorSource& source) {
 		return source.indirect_descriptor && source.indirect_descriptor->workgroup_axis != UINT32_MAX;
 	};
@@ -1582,20 +1578,16 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 	     std::ranges::any_of(program.descriptor_sources, uses_workgroup_count));
 	auto& reads = snapshot.specialization_reads;
 	reads = program.source_reads;
-	SrtReadCapture capture {runtime, reads};
-	auto observed = capture_reads ? capture.ObservedRuntime() : runtime;
-	SrtWalker clean(program, CleanRuntime(observed));
-	SrtWalker walker(program, observed,
-	                 capture_reads || program.requires_specialization_memory ? &clean : nullptr);
-	if (!walker.RefreshFlatBuffer(snapshot.flattened_srt)) {
+	if (!WalkSrt(program, runtime, snapshot.flattened_srt, capture_reads ? &reads : nullptr)) {
 		return SpecializationFail(fmt::format("SRT refresh complete={} reads={} user_data={}",
 		    program.srt_plan_complete, program.srt_reads.size(), runtime.user_data.size()));
 	}
-	const auto active = std::span<const uint8_t>(program.active_sources);
+	SrtReadCapture capture {runtime, reads};
+	auto observed = capture_reads ? capture.ObservedRuntime() : runtime;
 	std::array<uint32_t, 3> workgroup_counts {};
 	if (runtime.indirect_args_address != 0u) {
 		for (uint32_t source = 0; source < program.descriptor_sources.size(); ++source) {
-			if ((!active.empty() && !active[source]) ||
+			if (!WalkedActive(program, source) ||
 			    !uses_workgroup_count(program.descriptor_sources[source])) continue;
 			if ((runtime.indirect_args_address & 3u) != 0u ||
 			    !ReadTable(runtime.indirect_args_address, sizeof(workgroup_counts), 0u,
@@ -1610,24 +1602,24 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 	std::array<uint32_t, 4> stored {};
 	bool uniform_fill = words != 0;
 	if (uniform_fill && !fill.pattern_size.IsEmpty())
-		uniform_fill = clean.Evaluate(fill.pattern_size, words) && words >= 1 && words <= 4;
+		uniform_fill = WalkedValue(program, fill.pattern_size, words) && words >= 1 && words <= 4;
 	uint32_t elements = 0;
 	if (uniform_fill && !fill.element_count.IsEmpty())
-		uniform_fill = clean.Evaluate(fill.element_count, elements) && elements != 0;
+		uniform_fill = WalkedValue(program, fill.element_count, elements) && elements != 0;
 	for (uint32_t i = 0; i < words && uniform_fill; ++i) {
-		uniform_fill = clean.Evaluate(fill.values[i], stored[i]) && stored[i] == stored[0];
+		uniform_fill = WalkedValue(program, fill.values[i], stored[i]) && stored[i] == stored[0];
 	}
 	if (uniform_fill) {
 		snapshot.uniform_fill = fill.fill;
 		snapshot.uniform_fill.value = stored[0];
 		snapshot.uniform_fill.bounded_elements = elements;
 	}
-	const auto evaluate = [&](uint32_t source, DescriptorValue& value, bool written = false) {
+	const auto evaluate = [&](uint32_t source, DescriptorValue& value) {
 		if (source >= program.descriptor_sources.size()) {
 			return false;
 		}
-		if (active.empty() || active[source]) {
-			return (capture_reads && written ? clean : walker).EvaluateDescriptor(source, value) ||
+		if (WalkedActive(program, source)) {
+			return WalkedDescriptor(program, source, value) ||
 			    SpecializationFail(fmt::format("descriptor evaluation source={}", source));
 		}
 		value = {};
@@ -1651,9 +1643,9 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 			if (source == nullptr) return false;
 			if (source->indirect_descriptor.has_value()) {
 				snapshot.buffers[i] = {.dword_count = 4u};
-				if (active.empty() || active[base.source]) {
+				if (WalkedActive(program, base.source)) {
 					if (!MaterializeIndirectDescriptor(
-					        program, base.source, i, 4u, 4u, observed, clean, snapshot,
+					        program, base.source, i, 4u, 4u, observed, snapshot,
 					        snapshot.buffers, specialization.buffers,
 					        [&](DescriptorValue& value) {
 						        return NormalizeIndirectBuffer(
@@ -1661,7 +1653,7 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 					        }))
 						return resource_failed("buffer", i, base.source, snapshot.buffers.size());
 				}
-			} else if (!evaluate(base.source, snapshot.buffers[i], base.written)) {
+			} else if (!evaluate(base.source, snapshot.buffers[i])) {
 				return false;
 			}
 		}
@@ -1711,13 +1703,13 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 		}
 		if (source->indirect_descriptor.has_value()) {
 			snapshot.images[i] = {.dword_count = 8u};
-			if (!active.empty() && !active[image.source]) {
+			if (!WalkedActive(program, image.source) || !WalkedImageActive(program, i)) {
 				continue;
 			}
 			const auto& indirect = *source->indirect_descriptor;
 			const bool bounded_table = IsBoundedDescriptorTable(program, indirect);
 			if (!MaterializeIndirectDescriptor(
-			        program, image.source, i, 8u, image.r128 ? 4u : 8u, observed, clean, snapshot, snapshot.images,
+			        program, image.source, i, 8u, image.r128 ? 4u : 8u, observed, snapshot, snapshot.images,
 			        specialization.images, [&](DescriptorValue& value) {
 				        // A broad heap also contains resources for other typed image operations.
 				        if (NullImageDescriptor(value) || !ValidImageDescriptor(value, image.r128) ||
@@ -1728,10 +1720,12 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 				return resource_failed("image", i, image.source, snapshot.images.size());
 			}
 		} else {
-			if (!evaluate(image.source, snapshot.images[i], image.written)) {
+			if (!evaluate(image.source, snapshot.images[i])) {
 				return false;
 			}
-			if (!ValidImageDescriptor(snapshot.images[i], image.r128)) {
+			// An image listed only by unreached blocks reads zero, not its source's descriptor.
+			if (!WalkedImageActive(program, i) ||
+			    !ValidImageDescriptor(snapshot.images[i], image.r128)) {
 				snapshot.images[i].dwords.fill(0);
 			}
 		}
@@ -1749,8 +1743,8 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 			const auto* indirect = descriptor->indirect_descriptor ? &*descriptor->indirect_descriptor : nullptr;
 			if (indirect != nullptr && indirect->sources.empty()) {
 				snapshot.samplers[i] = {.dword_count = 4u};
-				if ((active.empty() || active[source]) && !MaterializeIndirectDescriptor(
-				        program, source, i, 4u, 4u, observed, clean, snapshot, snapshot.samplers,
+				if (WalkedActive(program, source) && !MaterializeIndirectDescriptor(
+				        program, source, i, 4u, 4u, observed, snapshot, snapshot.samplers,
 				        specialization.samplers,
 				        [](DescriptorValue&) { return true; }))
 					return resource_failed("sampler", i, source, snapshot.samplers.size());
@@ -1777,8 +1771,9 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 }
 
 uint32_t ImageHeapKind(const ImageResource& root, DescriptorValue& value) {
-	// Mirrors BuildResourceSpecialization for sampled roots: entries it would reject, and
-	// entries that need shader-side conversion, read as null instead of failing the draw.
+	// Mirrors BuildResourceSpecialization for sampled roots: entries it would reject, entries of
+	// another dimension and entries that need shader-side conversion read as null instead of
+	// failing the draw.
 	const auto format = static_cast<Prospero::BufferFormat>((value.dwords[1] >> 20u) & 0x1ffu);
 	const auto tile = static_cast<Prospero::TileMode>((value.dwords[3] >> 20u) & 0x1fu);
 	const auto numeric_class =
@@ -2065,4 +2060,86 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
 	image_remap.Apply(images);
 }
 
+namespace {
+
+using ControlPhis = decltype(ResourcePlan::control_phi_predecessors);
+
+// Instructions come first, so that any value can refer to them by index, then their arguments.
+// Plan Phis have no blocks; a control Phi lists the plan blocks of its arguments after them.
+void IoValues(Writer& writer, const std::list<Inst>& values, const ControlPhis& phis) {
+	IoAll(writer, static_cast<uint32_t>(values.size()));
+	for (const auto& inst: values) {
+		writer.index.emplace(&inst, static_cast<uint32_t>(writer.index.size()));
+		IoAll(writer, inst.GetOpcode(), inst.Flags<uint64_t>());
+	}
+	for (const auto& inst: values) {
+		if (inst.GetOpcode() != ValueOpcode::Phi) {
+			for (size_t index = 0; index < inst.NumArgs(); ++index) IoAll(writer, inst.Arg(index));
+			continue;
+		}
+		std::vector<Value> args;
+		for (size_t index = 0; index < inst.NumArgs(); ++index) {
+			EXIT_IF(inst.PhiBlock(index) != nullptr);
+			args.push_back(inst.Arg(index));
+		}
+		const auto found = phis.find(&inst);
+		IoAll(writer, args, found != phis.end() ? found->second : std::vector<uint32_t> {});
+	}
+}
+
+void IoValues(Reader& reader, std::list<Inst>& values, ControlPhis& phis) {
+	uint32_t count = 0;
+	Io(reader, count);
+	count = reader.Count(count);
+	for (uint32_t index = 0; index < count && reader.ok; ++index) {
+		ValueOpcode opcode {};
+		uint64_t    flags = 0;
+		IoAll(reader, opcode, flags);
+		reader.ok = reader.ok && opcode < ValueOpcode::Count;
+		if (reader.ok) reader.insts.push_back(&values.emplace_back(opcode, flags));
+	}
+	for (auto* inst: reader.insts) {
+		if (inst->GetOpcode() != ValueOpcode::Phi) {
+			for (size_t index = 0; index < inst->NumArgs(); ++index) {
+				Value arg;
+				Io(reader, arg);
+				inst->SetArg(index, arg);
+			}
+			continue;
+		}
+		std::vector<Value>    args;
+		std::vector<uint32_t> predecessors;
+		IoAll(reader, args, predecessors);
+		for (const auto arg: args) inst->AddPhiOperand(nullptr, arg);
+		if (!predecessors.empty()) phis.emplace(inst, std::move(predecessors));
+	}
+}
+
+// Binds every plan field: one added later fails to compile here until it is stored or derived.
+template <class S, class P> void IoPlan(S& io, P& plan) {
+	auto& [stage, shader_hash, user_data_base, user_data_count, values, memory_info, descriptor_sources,
+	       selector_groups, control_flow, control_phis, srt_reads, source_reads, capture_specialization_reads,
+	       srt_plan_complete, resource_tracking_complete, info, uniform_fill,
+	       // Derived on load.
+	       srt_code, srt_user_data, srt_captures, srt_scratch, material_keys, selector_keys] = plan;
+	IoValues(io, values, control_phis);
+	IoAll(io, stage, shader_hash, user_data_base, user_data_count, memory_info, descriptor_sources,
+	      selector_groups, control_flow, srt_reads, source_reads, capture_specialization_reads,
+	      srt_plan_complete, resource_tracking_complete, info, uniform_fill);
+}
+
+} // namespace
+
+std::vector<uint8_t> SaveResourcePlan(const ResourcePlan& plan) {
+	Writer writer;
+	IoPlan(writer, plan);
+	return std::move(writer.bytes);
+}
+
+bool LoadResourcePlan(std::span<const uint8_t> bytes, ResourcePlan& plan) {
+	Reader reader {bytes};
+	IoPlan(reader, plan);
+	return reader.ok && reader.bytes.empty() && DerivePlanState(plan);
+}
+
 } // namespace Libs::Graphics::ShaderRecompiler::IR

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

@@ -53,6 +53,11 @@ struct ResourceSpecialization {
 // its values and is independent of the translated shader CFG.
 ResourcePlan ExtractResourcePlan(const Program& program);
 
+// Shader cache record of a plan; loading it into a default-constructed plan rebuilds the walker
+// and scratch state like extraction. A failed load leaves a plan that must not be used.
+std::vector<uint8_t> SaveResourcePlan(const ResourcePlan& plan);
+bool                 LoadResourcePlan(std::span<const uint8_t> bytes, ResourcePlan& plan);
+
 // Refreshes cached resources and specialization in place. A failed refresh must not be used.
 bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime,
                           ResourceSnapshot& snapshot, ResourceSpecialization& specialization);

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

@@ -7,6 +7,8 @@
 
 #include <algorithm>
 #include <bit>
+#include <cmath>
+#include <cstdio>
 #include <fmt/format.h>
 #include <map>
 #include <numeric>
@@ -1014,7 +1016,9 @@ private:
 				    (!shared_image_heap && (a.table_offset != b.table_offset ||
 				                           a.table_immediate != b.table_immediate ||
 				                           a.table_stride != b.table_stride)) || a.indexed != b.indexed ||
-				    a.workgroup_axis != b.workgroup_axis || a.sources != b.sources ||
+				    a.workgroup_axis != b.workgroup_axis ||
+				    a.workgroup_multiplier != b.workgroup_multiplier ||
+				    a.workgroup_shift != b.workgroup_shift || a.sources != b.sources ||
 				    a.signed_count != b.signed_count ||
 				    !EquivalentValue(m_program, a.key_count, b.key_count) ||
 				    a.selector_first.IsEmpty() != b.selector_first.IsEmpty() ||
@@ -1136,31 +1140,63 @@ private:
 		LaneQuantifier lanes = LaneQuantifier::All;
 	};
 
+	// The immediate that edge proves value differs from on every lane, if any.
+	std::optional<uint32_t> ExcludedBy(const EdgePredicate& edge, Value value) const {
+		if (edge.lanes != LaneQuantifier::All) return {};
+		const auto* test = edge.condition.TryInstruction();
+		if (test != nullptr && test->NumArgs() == 2u &&
+		    ((test->GetOpcode() == ValueOpcode::INotEqual32 && edge.positive) ||
+		     (test->GetOpcode() == ValueOpcode::IEqual32 && !edge.positive))) {
+			for (uint32_t arg = 0; arg < 2u; ++arg) {
+				uint32_t immediate;
+				if (ImmediateU32(test->Arg(arg), immediate) &&
+				    EquivalentValue(m_program, test->Arg(arg ^ 1u), value)) return immediate;
+			}
+		}
+		return {};
+	}
+
 	bool NonzeroOnEntry(Value value, const Block* block) const {
 		if (block == nullptr) return false;
 		bool witnessed = false;
 		const bool guarded = GuardedOnEntry(block, [](const Block* current) {
 			return current->ImmPredecessors().size() != 1u;
 		}, [&](const EdgePredicate& edge) {
-			if (edge.lanes != LaneQuantifier::All) return false;
-			const auto* test = edge.condition.TryInstruction();
-			if (test != nullptr && test->NumArgs() == 2u &&
-			    ((test->GetOpcode() == ValueOpcode::INotEqual32 && edge.positive) ||
-			     (test->GetOpcode() == ValueOpcode::IEqual32 && !edge.positive))) {
-				for (uint32_t arg = 0; arg < 2u; ++arg) {
-					uint32_t immediate;
-					if (ImmediateU32(test->Arg(arg), immediate) && immediate == 0u &&
-					    EquivalentValue(m_program, test->Arg(arg ^ 1u), value)) return witnessed = true;
-				}
-			}
-			return false;
+			return ExcludedBy(edge, value) == 0u && (witnessed = true);
 		});
 		return guarded && witnessed;
 	}
 
-	uint32_t WorkgroupAxis(Value key) const {
+	// The key that every path from its definition to block rules out, if any.
+	std::optional<uint32_t> ExcludedKeyOnEntry(Value key, const Block* block) const {
+		const auto* definition = key.Resolve().TryInstruction();
+		if (definition == nullptr) return {};
+		std::optional<uint32_t> excluded;
+		const auto defined = [&](const Block* current) { return current == definition->Parent(); };
+		const bool guarded = GuardedOnEntry(block, defined, [&](const EdgePredicate& edge) {
+			const auto immediate = ExcludedBy(edge, key);
+			if (!immediate || (excluded && excluded != immediate)) return false;
+			excluded = immediate;
+			return true;
+		});
+		return guarded ? excluded : std::nullopt;
+	}
+
+	// The workgroup ID axis of key; with first, also of a runtime first + workgroup ID sum.
+	uint32_t WorkgroupAxis(Value key, Value* first = nullptr) const {
 		const auto* builtin = key.Resolve().TryInstruction();
 		uint32_t axis = UINT32_MAX;
+		if (first != nullptr && builtin != nullptr && builtin->GetOpcode() == ValueOpcode::IAdd32) {
+			for (uint32_t arm = 0; arm < 2u; ++arm) {
+				axis = WorkgroupAxis(builtin->Arg(arm));
+				const auto base = builtin->Arg(arm ^ 1u);
+				if (axis == UINT32_MAX ||
+				    !ValidateRuntimeValue(m_program, base, RuntimeValueType::Integer)) continue;
+				*first = base;
+				return axis;
+			}
+			return UINT32_MAX;
+		}
 		if (m_program.stage != ShaderType::Compute || builtin == nullptr ||
 		    builtin->GetOpcode() != ValueOpcode::GetBuiltin || builtin->NumArgs() != 2u ||
 		    builtin->Arg(0) != Value(static_cast<uint32_t>(StageInputKind::WorkgroupId)) ||
@@ -1168,6 +1204,27 @@ private:
 		return axis;
 	}
 
+	// Matches key = (first + workgroup ID) * multiplier >> shift with shift >= 32; first and the
+	// division are optional.
+	bool MatchWorkgroupKey(Value key, DescriptorSource::IndirectDescriptor& indirect) const {
+		uint32_t bits = 0, multiplier = 0;
+		const auto* shift = key.Resolve().TryInstruction();
+		if (shift != nullptr && shift->GetOpcode() == ValueOpcode::ShiftRightLogical32) {
+			const auto* high = shift->Arg(0).Resolve().TryInstruction();
+			if (high != nullptr && high->GetOpcode() == ValueOpcode::UMulHi &&
+			    ImmediateU32(shift->Arg(1), bits) && bits < 32u &&
+			    ImmediateU32(high->Arg(1), multiplier)) {
+				key = high->Arg(0);
+				bits += 32u;
+			}
+		}
+		indirect.workgroup_axis = WorkgroupAxis(key, &indirect.selector_first);
+		if (indirect.workgroup_axis == UINT32_MAX) return false;
+		indirect.workgroup_multiplier = multiplier;
+		indirect.workgroup_shift = bits;
+		return true;
+	}
+
 	bool MatchTableOffset(Value value, Value& key, uint32_t& offset, uint32_t& stride) const {
 		offset = 0;
 		for (;;) {
@@ -1672,6 +1729,8 @@ private:
 		uint32_t initial_arm;
 	};
 
+	static bool ZeroBased(uint32_t initial, uint32_t step) { return initial == 0u && step == 1u; }
+
 	std::optional<Induction> MatchInduction(Value key, const auto& accepts_step) const {
 		key = key.Resolve();
 		const auto* phi = key.TryInstruction();
@@ -1733,8 +1792,7 @@ private:
 	}
 
 	uint32_t PostTestLoopMaximum(Value key) const {
-		const auto induction = MatchInduction(key,
-		    [](uint32_t initial, uint32_t step) { return initial == 0u && step == 1u; });
+		const auto induction = MatchInduction(key, ZeroBased);
 		if (!induction) return UINT32_MAX;
 		const auto [phi, step, initial_arm] = *induction;
 		const auto* incoming = phi->PhiBlock(initial_arm ^ 1u);
@@ -1767,8 +1825,7 @@ private:
 
 	const Inst* BoundedLoop(Value key, const Block* use, const auto& accepts_bound,
 	                       LaneQuantifier lanes = LaneQuantifier::Any) const {
-		return BoundedInduction(key, use,
-		    [](uint32_t initial, uint32_t step) { return initial == 0u && step == 1u; },
+		return BoundedInduction(key, use, ZeroBased,
 		    [&](const Inst& compare, uint32_t initial) {
 			    return (compare.GetOpcode() == ValueOpcode::SLessThan32 ||
 			            compare.GetOpcode() == ValueOpcode::ULessThan32) &&
@@ -1793,6 +1850,29 @@ private:
 		});
 	}
 
+	// A zero-based key that only per-lane data bounds takes the count all uniform zero-based
+	// loops in the program share, assuming each lane's count stays within it.
+	bool ProgramLoopCount(Value key, DescriptorSource::IndirectDescriptor& indirect) const {
+		if (!MatchInduction(key, ZeroBased) || BoundedLoop(key, nullptr) != nullptr) return false;
+		const Inst* count = nullptr;
+		for (auto* block: m_program.blocks) {
+			for (auto& phi: *block) {
+				if (phi.GetOpcode() != ValueOpcode::Phi) break;
+				const auto* compare = BoundedLoop(Value(&phi), nullptr);
+				if (compare == nullptr) continue;
+				if (count != nullptr && !EquivalentValue(m_program, compare->Arg(1), count->Arg(1)))
+					return false;
+				count = compare;
+			}
+		}
+		if (count == nullptr) return false;
+		std::printf("Warning: shader 0x%016llx bounds a per-lane table loop by its uniform loops' "
+		            "count\n", static_cast<unsigned long long>(m_program.shader_hash));
+		indirect.key_count    = count->Arg(1);
+		indirect.signed_count = count->GetOpcode() == ValueOpcode::SLessThan32;
+		return true;
+	}
+
 	Value ActiveValue(Value value, Value guard) const {
 		value = value.Resolve();
 		while (const auto* select = value.TryInstruction()) {
@@ -1829,6 +1909,10 @@ private:
 	}
 
 	bool Dominates(const Block* initial, const Block* use, Value witness = {}) const {
+		// The proven local-zero invocation fails every all-lane bypass condition that excludes it,
+		// such as an all-inactive or an all-nonzero bypass.
+		const auto* negated = SimplifyGuard(witness).TryInstruction();
+		const bool nonzero = negated != nullptr && negated->GetOpcode() == ValueOpcode::LogicalNot;
 		std::vector<const Block*> pending {use};
 		for (size_t index = 0; index < pending.size(); ++index) {
 			const auto* current = pending[index];
@@ -1837,9 +1921,9 @@ private:
 			for (const auto* previous: current->ImmPredecessors()) {
 				if (!witness.IsEmpty()) {
 					const auto edge = ConditionalEdge(previous, current);
-					// The proven local-zero invocation makes an all-inactive bypass impossible.
-					if (edge && edge->lanes == LaneQuantifier::All && !edge->positive &&
-					    Implies(witness, edge->condition)) continue;
+					if (edge && edge->lanes == LaneQuantifier::All &&
+					    (edge->positive ? nonzero && Implies(edge->condition, negated->Arg(0))
+					                    : Implies(witness, edge->condition))) continue;
 				}
 				if (std::ranges::find(pending, previous) == pending.end()) pending.push_back(previous);
 			}
@@ -1948,7 +2032,10 @@ private:
 
 	bool LocalZeroWitness(Value guard) const {
 		const auto* equal = SimplifyGuard(guard).TryInstruction();
-		if (equal == nullptr || equal->GetOpcode() != ValueOpcode::IEqual32) return false;
+		const bool negated = equal != nullptr && equal->GetOpcode() == ValueOpcode::LogicalNot;
+		if (negated) equal = SimplifyGuard(equal->Arg(0)).TryInstruction();
+		const auto expected = negated ? ValueOpcode::INotEqual32 : ValueOpcode::IEqual32;
+		if (equal == nullptr || equal->GetOpcode() != expected) return false;
 		for (uint32_t arg = 0; arg < 2u; ++arg) {
 			if (equal->Arg(arg).Resolve() != Value(0u)) continue;
 			Value induction;
@@ -1982,6 +2069,10 @@ private:
 		    EquivalentValue(m_program, test->Arg(0), value)) return test->Arg(1);
 		if ((positive ? op == ValueOpcode::UGreaterThan32 : op == ValueOpcode::ULessThanEqual32) &&
 		    EquivalentValue(m_program, test->Arg(1), value)) return test->Arg(0);
+		uint32_t limit = 0;
+		if ((positive ? op == ValueOpcode::ULessThanEqual32 : op == ValueOpcode::UGreaterThan32) &&
+		    ImmediateU32(test->Arg(1), limit) && limit != UINT32_MAX &&
+		    EquivalentValue(m_program, test->Arg(0), value)) return Value(limit + 1u);
 		const auto* scan = value.Resolve().TryInstruction();
 		if (finite_bitscan && scan != nullptr && scan->GetOpcode() == ValueOpcode::FindILsb32 &&
 		    (positive ? op == ValueOpcode::INotEqual32 : op == ValueOpcode::IEqual32)) {
@@ -2056,7 +2147,7 @@ private:
 		const auto* read = ActiveValue(value, active).TryInstruction();
 		if (read == nullptr || read->GetOpcode() != ValueOpcode::LoadSharedU32 ||
 		    m_program.stage != ShaderType::Compute ||
-		    read->Arg(1).Resolve() != Value(true)) return {};
+		    !Implies(active, read->Arg(1))) return {};
 		uint64_t group_size = 1u;
 		for (const auto dimension: m_input.compute->threads_num) {
 			if (dimension == 0u || dimension > m_input.compute->wave_size) return {};
@@ -2087,10 +2178,16 @@ private:
 		const auto* limit = BoundedLoop(read_offset.index, read->Parent(),
 		    [](const Inst& compare, uint32_t) { return compare.GetOpcode() == ValueOpcode::ULessThan32; });
 		if (limit == nullptr) return {};
+		auto counter_active = active;
 		const auto* counter_read = ActiveValue(limit->Arg(1), active).TryInstruction();
+		// Every lane reads the same counter word, so a selected lane carries it.
+		if (counter_read != nullptr && counter_read->GetOpcode() == ValueOpcode::ReadFirstLane) {
+			counter_active = counter_read->Arg(1);
+			counter_read = ActiveValue(counter_read->Arg(0), counter_active).TryInstruction();
+		}
 		uint32_t consumed_counter = 0;
 		if (counter_read == nullptr || counter_read->GetOpcode() != ValueOpcode::LoadSharedU32 ||
-		    counter_read->Arg(1).Resolve() != Value(true) ||
+		    !Implies(counter_active, counter_read->Arg(1)) ||
 		    !constant_address(*counter_read, consumed_counter)) return {};
 
 		struct Initializer { const Inst* inst; uint32_t address; uint32_t words; };
@@ -2304,25 +2401,23 @@ private:
 		    ((offset | stride | memory->offset) & 3u) == 0u &&
 		    uint64_t {offset} + memory->offset <= UINT32_MAX;
 		if (affine) {
-			workgroup_axis = WorkgroupAxis(index);
-			if (workgroup_axis == UINT32_MAX) {
-				const auto* sum = index.Resolve().TryInstruction();
-				if (sum != nullptr && sum->GetOpcode() == ValueOpcode::IAdd32) {
-					for (uint32_t arm = 0; arm < 2u; ++arm) {
-						const auto axis = WorkgroupAxis(sum->Arg(arm));
-						const auto base = sum->Arg(arm ^ 1u);
-						if (axis == UINT32_MAX ||
-						    !ValidateRuntimeValue(m_program, base, RuntimeValueType::Integer)) continue;
-						workgroup_axis = axis;
-						first = base;
-						break;
-					}
-				}
-			}
+			workgroup_axis = WorkgroupAxis(index, &first);
 			const auto* bound = BoundedLoop(index, read->Parent());
 			if (bound != nullptr && bound->GetOpcode() == ValueOpcode::ULessThan32) {
 				count = bound->Arg(1);
+			} else if (workgroup_axis == UINT32_MAX && !require_bound) {
+				// Other bounded indices, such as bit scans, read the records up to their maximum;
+				// a bit scan of a nonzero mask load reads only the ranks its words set.
+				const auto maximum = SelectorMaximum(index);
+				if (maximum == 31u) selector.ranks = BitScanRanks(index);
+				if (maximum < MaxIndirectDescriptorProbes) count = Value(maximum + 1u);
 			}
+			// A record selected from a shared append list is below the list's bound.
+			const auto* lane = index.Resolve().TryInstruction();
+			if (count.IsEmpty() && lane != nullptr &&
+			    lane->GetOpcode() == ValueOpcode::ReadFirstLane &&
+			    HasActiveLane(lane->Arg(1), lane->Parent()))
+				count = SharedLoadBound(lane->Arg(0), lane->Arg(1));
 		}
 		// Raw pointers have no V# bounds for enumerating an unbounded selector domain.
 		if (memory->kind == ResourceKind::ScalarAddress && count.IsEmpty() &&
@@ -2498,7 +2593,8 @@ private:
 	// Each equation is an inductive subset obligation. Phis require every incoming
 	// value; AND needs either operand. Grounding rejects cycles without an initial
 	// value, then the greatest fixed point checks every initial and carried edge.
-	bool BitsetMaskSubset(Value mask, Value active, Value word, const Inst& read) const {
+	bool BitsetMaskSubset(Value mask, Value active, Value word,
+	                      std::span<const Inst* const> reads) const {
 		struct State {
 			Value value;
 			Value guard;
@@ -2510,7 +2606,10 @@ private:
 			bool all = true;
 			std::vector<size_t> terms;
 		};
-		const auto load_guard = read.Arg(4).Resolve();
+		// A scalar load reads one word without EXEC; four-word loads take their guard.
+		const bool scalar = reads[0]->GetOpcode() == ValueOpcode::ReadConstBuffer;
+		const auto load_guard = scalar ? Value(true) : reads[0]->Arg(4).Resolve();
+		const auto words = static_cast<uint32_t>(reads.size()) * (scalar ? 1u : 4u);
 		std::vector<State> states(2);
 		std::vector<Equation> equations(2);
 		const auto add = [&](State state) {
@@ -2521,9 +2620,9 @@ private:
 			equations.emplace_back();
 			return states.size() - 1u;
 		};
-		std::array<size_t, 5> roots {add({.value = active})};
-		for (uint32_t component = 0; component < 4u; ++component)
-			roots[component + 1u] = add({mask, active, component, component});
+		std::vector<size_t> roots {add({.value = active})};
+		for (uint32_t component = 0; component < words; ++component)
+			roots.push_back(add({mask, active, component, component}));
 		for (size_t index = 2; index < states.size(); ++index) {
 			const auto state = states[index];
 			const auto* inst = state.value.TryInstruction();
@@ -2573,11 +2672,14 @@ private:
 				switch (inst->GetOpcode()) {
 					case ValueOpcode::CompositeExtractU32x4: {
 						uint32_t component;
-						if (inst->Arg(0).Resolve().TryInstruction() == &read &&
-						    ImmediateU32(inst->Arg(1), component) && component == state.component)
+						if (inst->Arg(0).Resolve().TryInstruction() == reads[state.component / 4u] &&
+						    ImmediateU32(inst->Arg(1), component) && component == state.component % 4u)
 							equation.terms = {1u};
 						break;
 					}
+					case ValueOpcode::ReadConstBuffer:
+						if (scalar && inst == reads[state.component]) equation.terms = {1u};
+						break;
 					case ValueOpcode::Phi:
 						// The predecessor may be an earlier outer iteration. Establish
 						// this register's invariant without the current word or EXEC.
@@ -2598,17 +2700,17 @@ private:
 								const auto* masked = key.TryInstruction();
 								uint32_t bound = 0;
 								if (masked != nullptr && masked->GetOpcode() == ValueOpcode::BitwiseAnd32 &&
-								    DescendingLoop(word, masked->Parent(), bound) && bound == 3u) {
+								    DescendingLoop(word, masked->Parent(), bound) && bound == words - 1u) {
 									for (uint32_t side = 0; side < 2u; ++side) {
 										uint32_t bits;
-										if (ImmediateU32(masked->Arg(side), bits) && (bits & 3u) == 3u)
+										if (ImmediateU32(masked->Arg(side), bits) && (bits & bound) == bound)
 											key = masked->Arg(side ^ 1u).Resolve();
 									}
 								}
 								return EquivalentValue(m_program, key, word);
 							});
 							uint32_t literal;
-							if (ImmediateU32(equality, literal) && literal < 4u) selected_word = literal;
+							if (ImmediateU32(equality, literal) && literal < words) selected_word = literal;
 							equation.terms.push_back(term(inst->Arg(1), inst->Arg(0), selected_word));
 						}
 						if (!selected || !*selected)
@@ -2646,7 +2748,53 @@ private:
 		return std::ranges::all_of(roots, [&](size_t root) { return valid[root]; });
 	}
 
-	bool MatchBitsetSelector(Value rank, Value active, SelectorBounds bounds,
+	// The mask words whose set bits are every record a bit scan of index selects: the scanned
+	// value keeps only bits of one scalar load, which reads a word its address can reach.
+	std::optional<DescriptorSource::IndirectDescriptor::SelectorRead::MaskRanks>
+	BitScanRanks(Value index) {
+		const auto* scan = index.Resolve().TryInstruction();
+		if (scan == nullptr || scan->GetOpcode() != ValueOpcode::FindILsb32) return {};
+		const auto* masked = scan->Arg(0).Resolve().TryInstruction();
+		if (masked == nullptr) return {};
+		std::vector<const Inst*> values {masked};
+		const Inst* load = nullptr;
+		for (size_t i = 0; i < values.size() && load == nullptr; ++i) {
+			const auto* value = values[i];
+			if (value->GetOpcode() == ValueOpcode::ReadConstBuffer) load = value;
+			if (value->GetOpcode() != ValueOpcode::Phi &&
+			    value->GetOpcode() != ValueOpcode::BitwiseAnd32) continue;
+			for (size_t arg = 0; arg < value->NumArgs(); ++arg) {
+				const auto* next = value->Arg(arg).Resolve().TryInstruction();
+				if (next != nullptr && std::ranges::find(values, next) == values.end())
+					values.push_back(next);
+			}
+		}
+		uint32_t memory_index = 0;
+		const auto* memory = load != nullptr ? ScalarReadMemory(*load, memory_index) : nullptr;
+		if (memory == nullptr || memory->offset > INT32_MAX ||
+		    !MemoryIndexBelongsTo(memory_index, *load) ||
+		    !BitsetMaskSubset(scan->Arg(0), Value(true), Value(0u), std::array {load})) return {};
+		auto address = load->Arg(1);
+		auto active = Value(true);
+		const auto* lane = address.Resolve().TryInstruction();
+		if (lane != nullptr && lane->GetOpcode() == ValueOpcode::ReadFirstLane) {
+			if (!HasActiveLane(lane->Arg(1), lane->Parent())) return {};
+			active = lane->Arg(1);
+			address = lane->Arg(0);
+		}
+		// A left shift keeps the address a multiple of its scale; other addresses read any word.
+		const auto* shift = ActiveValue(address, active).TryInstruction();
+		uint32_t bits = 0;
+		if (shift == nullptr || shift->GetOpcode() != ValueOpcode::ShiftLeftLogical32 ||
+		    !ImmediateU32(shift->Arg(1), bits) || bits >= 32u) bits = 0;
+		DescriptorSource source;
+		if (!MakeRuntimeTableSource(*load, source)) return {};
+		return DescriptorSource::IndirectDescriptor::SelectorRead::MaskRanks {
+		    .source = InternSource(source), .stride = std::max(1u << bits, 4u),
+		    .offset = memory->offset & ~3u};
+	}
+
+	bool MatchBitsetSelector(Value rank, Value active, SelectorBounds bounds, uint32_t words,
 	                         DescriptorSource::IndirectDescriptor& indirect) {
 		const auto selected_value = [&](Value value) {
 			for (;;) {
@@ -2670,7 +2818,7 @@ private:
 		}
 		const auto* word_inst = word.TryInstruction();
 		if (word_inst == nullptr || std::ranges::none_of(bounds, [&](const auto& bound) {
-			return bound.first == word_inst && bound.second == 3u;
+			return bound.first == word_inst && bound.second == words - 1u;
 		})) return false;
 		// FFBH's leading-zero count is inverted again to recover the set-bit rank.
 		for (uint32_t inverse = 0; inverse < 2u; ++inverse) {
@@ -2682,21 +2830,34 @@ private:
 		const auto* scan = bit.TryInstruction();
 		if (scan == nullptr || scan->GetOpcode() != ValueOpcode::FindUMsb32 ||
 		    !ImpliesNonzero(active, scan->Arg(0))) return false;
+		const auto bitmap_load = [&](const Inst& read, uint32_t offset) {
+			const auto index = read.Flags<MemoryFlags>().index;
+			if (read.GetOpcode() != ValueOpcode::LoadBufferU32x4 || InCycle(read.Parent()) ||
+			    index >= m_program.memory_info.size()) return false;
+			const auto& memory = m_program.memory_info[index];
+			return memory.kind == ResourceKind::Buffer && !memory.typed && !memory.formatted &&
+			       memory.idxen && !memory.offen && memory.offset == offset &&
+			       memory.data_bits == 32u && memory.data_dwords == 4u &&
+			       read.Arg(2).Resolve() == Value(0u) && read.Arg(3).Resolve() == Value(0u);
+		};
 		for (const auto* block: m_program.blocks) {
 			for (const auto& read: *block) {
-				if (read.GetOpcode() != ValueOpcode::LoadBufferU32x4 || InCycle(block)) continue;
-				const auto index = read.Flags<MemoryFlags>().index;
-				if (index >= m_program.memory_info.size()) continue;
-				const auto& memory = m_program.memory_info[index];
-				if (memory.kind != ResourceKind::Buffer || memory.typed || memory.formatted ||
-				    !memory.idxen || memory.offen || memory.offset != 0u || memory.data_bits != 32u ||
-				    memory.data_dwords != 4u || read.Arg(2).Resolve() != Value(0u) ||
-				    read.Arg(3).Resolve() != Value(0u) ||
-				    !BitsetMaskSubset(scan->Arg(0), active, word, read)) continue;
+				if (!bitmap_load(read, 0u)) continue;
+				// An eight-word record continues in a second four-word load of the same buffer and EXEC.
+				std::array<const Inst*, 2> reads {&read};
+				if (words == 8u)
+					for (const auto* other: m_program.blocks)
+						for (const auto& load: *other)
+							if (reads[1] == nullptr && bitmap_load(load, 16u) &&
+							    EquivalentValue(m_program, load.Arg(0), read.Arg(0)) &&
+							    EquivalentValue(m_program, load.Arg(4), read.Arg(4))) reads[1] = &load;
+				const auto parts = std::span(reads).first(words / 4u);
 				DescriptorSource source;
-				if (!MakeRuntimeTableSource(read, source)) continue;
+				if (std::ranges::find(parts, nullptr) != parts.end() ||
+				    !BitsetMaskSubset(scan->Arg(0), active, word, parts) ||
+				    !MakeRuntimeTableSource(read, source)) continue;
 				indirect.selector.emplace(DescriptorSource::IndirectDescriptor::SelectorRead {
-				    .source = InternSource(source), .stride = 16u, .offset = 0u, .bitset = true});
+				    .source = InternSource(source), .stride = words * 4u, .offset = 0u, .bitset = true});
 				return true;
 			}
 		}
@@ -2726,7 +2887,8 @@ private:
 		first.SetArg(1, Value(maximum));
 		indirect.selector_first = Value(&first);
 		indirect.key_count = Value(maximum + 1u);
-		if (maximum == 127u) MatchBitsetSelector(lane->Arg(0), lane->Arg(1), bounds, indirect);
+		if (maximum == 127u || maximum == 255u)
+			MatchBitsetSelector(lane->Arg(0), lane->Arg(1), bounds, (maximum + 1u) / 32u, indirect);
 		return true;
 	}
 
@@ -2993,6 +3155,47 @@ private:
 		return true;
 	}
 
+	// The EXEC predicate of the image access through use; empty for any other use.
+	Value AccessPredicate(const Use& use, bool sampler) const {
+		const auto& access = *use.user;
+		if (use.operand != (sampler ? 1u : 0u) ||
+		    ImageOpcodeInfoOf(access.GetOpcode()).access == ImageAccess::None) return {};
+		if (sampler && (access.GetOpcode() != ValueOpcode::ImageSampleRaw ||
+		    (m_program.memory_info[access.Flags<MemoryFlags>().index].image_sample_flags &
+		     Decoder::ImageSampleFlagAdjust) != 0u)) return {};
+		const auto mask = access.Arg(access.NumArgs() - 1u);
+		return mask.GetType() == Type::U1 ? mask : Value {};
+	}
+
+	// A key no bound covers takes its maximum from a signed clamp to a constant, on the key or on
+	// the key plus a constant in an earlier-tracked sibling read of the same table. Negative keys
+	// and keys past that maximum read the first row.
+	uint32_t ClampedTableMaximum(Value key, const DescriptorSource& table,
+	                             const DescriptorSource::IndirectDescriptor& indirect) const {
+		uint32_t row = 0, offset = 0;
+		const auto clamped = [&](Value value) {
+			const auto* min = value.Resolve().TryInstruction();
+			if (min != nullptr && min->GetOpcode() == ValueOpcode::SMin32)
+				for (uint32_t arg = 0; arg < 2u; ++arg)
+					if (ImmediateU32(min->Arg(arg), row)) return min->Arg(arg ^ 1u);
+			return Value {};
+		};
+		if (!clamped(key).IsEmpty()) return row;
+		for (const auto& sibling: m_indirect_descriptors) {
+			const auto& read = *m_sources[sibling.source].indirect_descriptor;
+			const auto* add = clamped(sibling.key).Resolve().TryInstruction();
+			if (add == nullptr || add->GetOpcode() != ValueOpcode::IAdd32 ||
+			    m_sources[read.table_source] != table ||
+			    read.table_offset != indirect.table_offset ||
+			    read.table_stride != indirect.table_stride ||
+			    read.table_immediate != indirect.table_immediate) continue;
+			for (uint32_t arg = 0; arg < 2u; ++arg)
+				if (ImmediateU32(add->Arg(arg), offset) &&
+				    EquivalentValue(m_program, add->Arg(arg ^ 1u), key)) return row;
+		}
+		return UINT32_MAX;
+	}
+
 	bool TryMakeTableDescriptor(Inst& handle, const DescriptorSource& descriptor,
 	                            IndirectDescriptorPlan& plan) {
 		const bool sampler = handle.GetOpcode() == ValueOpcode::GetSamplerResource;
@@ -3019,6 +3222,7 @@ private:
 		} else if (!plan.table_indexed &&
 		           MatchScalarMaterialKey(key, indirect, material_source,
 		                                  !sampler && table_source.dword_count == 4u, sampler)) {
+			indirect.selector->excluded = ExcludedKeyOnEntry(key, handle.Parent());
 			// Keep live descriptor payloads, but project descriptor-only SMEM onto
 			// the GPU key using the same plan as other indirect image reads.
 			plan.retain_reads |= table_source.dword_count == 4u;
@@ -3080,36 +3284,42 @@ private:
 			if (indirect.key_count.IsEmpty() &&
 			    !MatchUniformizedMaterialKey(key, handle, indirect, material_source, false) &&
 			    !MatchUniformizedMaterialKey(key, handle, indirect, material_source, true) &&
-			    !MatchReversedImageRange(key, indirect)) {
-				const auto maximum = SelectorMaximum(key);
-				if (maximum >= MaxIndirectDescriptorProbes) return false;
-				indirect.key_count = Value(maximum + 1u);
+			    !MatchReversedImageRange(key, indirect) && !MatchWorkgroupKey(key, indirect)) {
+				auto maximum = SelectorMaximum(key);
+				if (maximum >= MaxIndirectDescriptorProbes) {
+					// Each image access bounds the key on the lanes it executes.
+					maximum = 0u;
+					for (const auto& use: handle.Uses()) {
+						const auto active = AccessPredicate(use, false);
+						if (active.IsEmpty()) {
+							maximum = UINT32_MAX;
+							break;
+						}
+						maximum = std::max(maximum, SelectorMaximum(key, {}, 0, active));
+					}
+				}
+				if (maximum >= MaxIndirectDescriptorProbes)
+					maximum = ClampedTableMaximum(key, table_source, indirect);
+				if (maximum < MaxIndirectDescriptorProbes)
+					indirect.key_count = Value(maximum + 1u);
+				else if (!ProgramLoopCount(key, indirect))
+					return false;
 			}
 			if ((table_offset & 3u) != 0u) return false;
 		} else {
 			// Keep the GPU selector and descriptor reads, and preserve the proven scalar
 			// offset domain within the bounded image table.
 			if (sampler || plan.split_offsets) return false;
-			const auto active = [&](const Use& use, bool sampler) -> Value {
-				const auto& access = *use.user;
-				if (use.operand != (sampler ? 1u : 0u) ||
-				    ImageOpcodeInfoOf(access.GetOpcode()).access == ImageAccess::None) return {};
-				if (sampler && (access.GetOpcode() != ValueOpcode::ImageSampleRaw ||
-				    (m_program.memory_info[access.Flags<MemoryFlags>().index].image_sample_flags &
-				     Decoder::ImageSampleFlagAdjust) != 0u)) return {};
-				const auto mask = access.Arg(access.NumArgs() - 1u);
-				return mask.GetType() == Type::U1 ? mask : Value {};
-			};
 			bool shared_selector = false;
 			if (!plan.table_indexed) {
 				for (const auto& owner: m_indirect_descriptors) {
 					const auto& selected = *m_sources[owner.source].indirect_descriptor;
 					if (!selected.selector || !EquivalentValue(m_program, key, owner.key) ||
 					    !std::ranges::all_of(handle.Uses(), [&](const Use& target) {
-						    const auto target_active = active(target, false);
+						    const auto target_active = AccessPredicate(target, false);
 						    return !target_active.IsEmpty() &&
 						        std::ranges::any_of(owner.handle->Uses(), [&](const Use& witness) {
-							        const auto owner_active = active(witness,
+							        const auto owner_active = AccessPredicate(witness,
 							            owner.handle->GetOpcode() == ValueOpcode::GetSamplerResource);
 							        return !owner_active.IsEmpty() && AvailableBefore(*witness.user, *target.user) &&
 							            Implies(target_active, owner_active);
@@ -3430,6 +3640,42 @@ private:
 		return result;
 	}
 
+	// Bounds of an F32 value on the lanes where guard holds; NaN when unknown.
+	std::pair<float, float> FloatRange(Value value, Value guard, uint32_t depth = 0) const {
+		constexpr std::pair unknown {NAN, NAN};
+		value = ActiveValue(value, guard);
+		if (value.IsImmediate() && value.GetType() == Type::F32)
+			return {value.F32Value(), value.F32Value()};
+		const auto* inst = value.TryInstruction();
+		// Each multiply-add evaluates three operands; keep the search shallow.
+		if (depth > 12u || inst == nullptr) return unknown;
+		const auto range = [&](uint32_t arg) {
+			return FloatRange(inst->Arg(arg), guard, depth + 1u);
+		};
+		switch (inst->GetOpcode()) {
+			case ValueOpcode::BitCastF32U32:
+			case ValueOpcode::BitCastU32F32: return range(0);
+			// Clamping maps NaN to zero.
+			case ValueOpcode::FPSaturate32: return {0.0f, 1.0f};
+			case ValueOpcode::FPFloor32: {
+				const auto [low, high] = range(0);
+				return {std::floor(low), std::floor(high)};
+			}
+			case ValueOpcode::FPMad32: {
+				// The rounded multiply and add are monotonic, so the bounds lie at the corners.
+				const auto a = range(0), b = range(1), c = range(2);
+				if (!std::isfinite(a.first) || !std::isfinite(a.second) ||
+				    !std::isfinite(b.first) || !std::isfinite(b.second)) return unknown;
+				const auto [low, high] = std::minmax({a.first * b.first, a.first * b.second,
+				                                      a.second * b.first, a.second * b.second});
+				const std::pair bounds {low + c.first, high + c.second};
+				if (!std::isfinite(bounds.first) || !std::isfinite(bounds.second)) return unknown;
+				return bounds;
+			}
+			default: return unknown;
+		}
+	}
+
 	uint32_t SelectorMaximum(Value value, SelectorBounds bounds = {}, uint32_t depth = 0,
 	                         Value guard = {}) const {
 		if (depth > 32u) return UINT32_MAX;
@@ -3445,9 +3691,13 @@ private:
 		switch (inst->GetOpcode()) {
 			case ValueOpcode::ReadFirstLane:
 				return SelectedSelectorDomain(*inst, guard).maximum;
-			case ValueOpcode::ReadLane:
-				// ReadLane ignores EXEC; only context-free bounds on every source lane apply.
-				return PossibleU32Bits(inst->Arg(0));
+			case ValueOpcode::ReadLane: {
+				// ReadLane ignores EXEC; only context-free bounds on every source lane apply,
+				// unless guard proves a waterfall key equal to the executing lanes' local value.
+				const auto local = EqualLocalKey(guard, value);
+				return local.IsEmpty() ? PossibleU32Bits(inst->Arg(0))
+				                       : SelectorMaximum(local, bounds, depth + 1u, guard);
+			}
 			case ValueOpcode::Phi: {
 				const auto initial = InitialCandidateMask(value);
 				if (!initial.IsEmpty()) return SelectorMaximum(initial, bounds, depth + 1u, guard);
@@ -3474,6 +3724,18 @@ private:
 				               NonzeroOnEntry(inst->Arg(0), inst->Parent()) ? 31u : UINT32_MAX;
 			case ValueOpcode::UMin32:
 			case ValueOpcode::BitwiseAnd32: return std::min(maximum(0), maximum(1));
+			case ValueOpcode::SMin32: {
+				// Signed and unsigned minimums agree on nonnegative operands.
+				const auto left = maximum(0), right = maximum(1);
+				if (std::max(left, right) <= INT32_MAX) return std::min(left, right);
+				break;
+			}
+			case ValueOpcode::ConvertS32F32: {
+				// The conversion truncates toward zero.
+				const auto [low, high] = FloatRange(inst->Arg(0), guard);
+				if (low >= 0.0f && high < 0x1p31f) return static_cast<uint32_t>(high);
+				break;
+			}
 			case ValueOpcode::SelectU32:
 				if (!guard.IsEmpty() && Implies(guard, inst->Arg(0))) return maximum(1);
 				return SelectorPredicateFalse(inst->Arg(0), bounds, depth + 1u, guard)

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

@@ -175,6 +175,8 @@ void AddBuiltinInput(ShaderInfo& info, StageInputKind kind) {
 		case StageInputKind::GlobalInvocationId:
 			AddInput(info, kind, 0, 3, "gl_GlobalInvocationID");
 			break;
+		case StageInputKind::BaseVertex: AddInput(info, kind, 0, 1, "gl_BaseVertex"); break;
+		case StageInputKind::BaseInstance: AddInput(info, kind, 0, 1, "gl_BaseInstance"); break;
 		case StageInputKind::PackedAncillary:
 		case StageInputKind::Parameter: break;
 	}
@@ -366,7 +368,6 @@ void Visit(Program& program, ShaderStageInputInfo input_info, InputUsage& inputs
 			info.function_lds |= kind == ResourceKind::Lds &&
 			                     program.stage != ShaderType::Compute && program.stage != ShaderType::Mesh;
 			if (shared_access == SharedAccess::Append || shared_access == SharedAccess::Consume) {
-				info.subgroup_ballot = true;
 				info.subgroup_shuffle = true;
 				info.subgroup_local_invocation_id = true;
 			}
@@ -445,6 +446,8 @@ void Visit(Program& program, ShaderStageInputInfo input_info, InputUsage& inputs
 				case StageInputKind::InstanceIndex:
 				case StageInputKind::FrontFacing:
 				case StageInputKind::LocalInvocationIndex:
+				case StageInputKind::BaseVertex:
+				case StageInputKind::BaseInstance:
 					if (component != 0u) {
 						return Fail("typed scalar builtin component is out of range");
 					}
@@ -517,7 +520,9 @@ void Visit(Program& program, ShaderStageInputInfo input_info, InputUsage& inputs
 			    inst.Flags<CFG::BranchCondition>() != CFG::BranchCondition::ScalarInstruction;
 			break;
 		case ValueOpcode::Ballot: info.subgroup_ballot = true; break;
-		case ValueOpcode::SubgroupUMinU32: info.subgroup_arithmetic = true; break;
+		case ValueOpcode::LiveMask: info.subgroup_ballot = info.helper_invocation = true; break;
+		case ValueOpcode::SubgroupUMinU32:
+		case ValueOpcode::SubgroupUMaxU32: info.subgroup_arithmetic = true; break;
 		case ValueOpcode::DppMoveU32:
 		case ValueOpcode::ReadFirstLane:
 		case ValueOpcode::ReadLane: {
@@ -576,7 +581,7 @@ void CollectShaderInfo(Program& program, ShaderStageInputInfo input_info) {
 	info.has_bitwise_xor = info.uses_lds = info.uses_gds = info.uses_flattened_srt =
 	    info.uses_dispatch_threads = info.uses_swizzle = false;
 	info.bvh = info.subgroup_ballot = info.subgroup_vote = info.subgroup_barrier =
-	    info.subgroup_shuffle = info.subgroup_arithmetic =
+	    info.subgroup_shuffle = info.subgroup_arithmetic = info.helper_invocation =
 	    info.subgroup_local_invocation_id = info.compute_derivatives = info.image_gather_extended =
 	    info.function_lds = info.function_scratch = info.pixel_valid_mask = info.buffer_int64_atomics =
 	    info.buffer_u8 = info.buffer_u16 = info.shared_int64_atomics = info.coherent_buffers = info.float64 = false;

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

@@ -8,8 +8,6 @@
 
 namespace Libs::Graphics::ShaderRecompiler::IR {
 
-class Value;
-
 using SrtMemoryReader = bool (*)(void* userdata, uint64_t address, std::span<uint32_t> values);
 
 // Image descriptors at base + first + k * step, k < count, within size bytes that a sampled
@@ -42,7 +40,6 @@ protected:
 struct SrtRuntime {
 	std::span<const uint32_t> user_data;
 	uint64_t                  shader_base                = 0;
-	SrtMemoryReader           read_memory                = nullptr;
 	void*                     userdata                   = nullptr;
 	SrtMemoryReader           read_specialization_memory = nullptr;
 	std::span<const uint32_t> workgroup_counts;
@@ -57,20 +54,12 @@ public:
 	SrtRuntime ObservedRuntime();
 
 private:
-	static bool ReadStrict(void* userdata, uint64_t address, std::span<uint32_t> values);
-	static bool ReadOrdinary(void* userdata, uint64_t address, std::span<uint32_t> values);
+	static bool Read(void* userdata, uint64_t address, std::span<uint32_t> values);
 
 	SrtRuntime                                  m_source;
 	std::vector<std::pair<uint64_t, uint64_t>>& m_ranges;
 };
 
-// Retained reads no longer need the guest instruction PC. A clean read evaluates
-// its address and bounds through the strict reader as well as the final DWORD.
-struct SrtReadFlags {
-	uint32_t index = 0;
-	uint32_t clean = 0;
-};
-
 enum class RuntimeValueType { Any, Integer, ImmutableInteger };
 
 bool ValidateRuntimeValue(const ResourcePlan& program, Value value,
@@ -78,9 +67,21 @@ bool ValidateRuntimeValue(const ResourcePlan& program, Value value,
 // Uniform across a compute wave, including values derived from its workgroup ID.
 // This does not imply that the host can evaluate the value.
 bool ValidateWaveUniformValue(const ResourcePlan& program, Value value);
-void CompileSrtWalker(ResourcePlan& program);
-// Uses the strict reader for values that affect shader specialization.
-SrtRuntime CleanRuntime(SrtRuntime runtime);
+// Compiles the plan's x86 walker. Without values it walks the plan's resources; otherwise it
+// only evaluates the values, with every block reachable.
+void CompileSrtWalker(ResourcePlan& plan, std::span<const Value> values = {});
+// Runs the walker. flat receives the scalar reads of reachable blocks and reads, when non-null,
+// one DWORD range per guest load. False when a reachable read is unknown.
+bool WalkSrt(const ResourcePlan& plan, const SrtRuntime& runtime, std::vector<uint32_t>& flat,
+             std::vector<std::pair<uint64_t, uint64_t>>* reads);
+// Results of the last walk; false when unknown.
+bool WalkedValue(const ResourcePlan& plan, Value value, uint32_t& result);
+bool WalkedDescriptor(const ResourcePlan& plan, uint32_t source, DescriptorValue& result);
+bool WalkedActive(const ResourcePlan& plan, uint32_t source);
+bool WalkedImageActive(const ResourcePlan& plan, uint32_t image);
+// Length of the guest load at rip if it belongs to the walker running on this thread, else 0.
+// The exception handler completes such a load as a zero read.
+uint32_t SrtWalkerLoadLength(uint64_t rip);
 
 // Nonnegative affine coefficients: element zero is constant, remaining elements are
 // caller-selected leaves. Runtime range proofs must additionally exclude U32 wrap.
@@ -124,43 +125,6 @@ std::optional<std::array<uint64_t, Size>> AffineIndex(
 	return left;
 }
 
-// One memoized evaluation session shared by the entire shader resource refresh.
-class SrtWalker {
-public:
-	SrtWalker(const ResourcePlan& program, const SrtRuntime& runtime,
-	          SrtWalker* clean_evaluator = nullptr);
-	~SrtWalker();
-	SrtWalker(const SrtWalker&)            = delete;
-	SrtWalker& operator=(const SrtWalker&) = delete;
-
-	bool Evaluate(Value value, uint32_t& result);
-	bool EvaluateDescriptor(uint32_t source, DescriptorValue& result);
-	// Refreshes reachable scalar reads and active descriptor sources in one walk.
-	bool RefreshFlatBuffer(std::vector<uint32_t>& flat);
-
-private:
-	SrtWalker(SrtWalker& parent, Value active_mask, bool clean = false,
-	          SrtWalker* clean_evaluator = nullptr);
-	ResourcePlan::EvaluationContext& AcquireContext();
-	void ResetUnknowns();
-	static float Float32(uint64_t bits);
-	bool EvaluateWide(Value value, uint64_t& result);
-	bool Arg(const Inst& inst, size_t index, uint64_t& result);
-	bool EvaluatePhi(const Inst& inst, uint64_t& result);
-	bool EvaluateExtract(const Inst& inst, uint64_t& result);
-	bool EvaluateRawRead(const Inst& inst, uint64_t& result);
-	bool EvaluateInst(const Inst& inst, uint64_t& result);
-
-	const ResourcePlan&              m_program;
-	SrtRuntime                      m_runtime;
-	SrtWalker*                      m_clean_evaluator = nullptr;
-	Value                           m_active_mask;
-	const std::vector<uint8_t>*      m_reachable_blocks = nullptr;
-	const SrtWalker*                m_session;
-	uint8_t                         m_read_policy;
-	ResourcePlan::EvaluationContext& m_context;
-};
-
 } // namespace Libs::Graphics::ShaderRecompiler::IR
 
 #endif /* EMULATOR_INCLUDE_EMULATOR_GRAPHICS_SHADER_RECOMPILER_SRTWALKER_H_ */

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

@@ -348,14 +348,20 @@ void VisitInstruction(Pass& pass, Block* block, Inst& inst) {
 } // namespace
 
 void RewriteToSsa(const BlockList& blocks) {
-	Pass pass;
+	Pass                               pass;
+	std::unordered_map<Block*, size_t> filled_predecessors;
 	for (auto* block: blocks) {
+		if (block->ImmPredecessors().empty()) {
+			pass.Seal(block);
+		}
 		for (auto& inst: *block) {
 			VisitInstruction(pass, block, inst);
 		}
-	}
-	for (auto* block: blocks) {
-		pass.Seal(block);
+		for (auto* successor: block->ImmSuccessors()) {
+			if (++filled_predecessors[successor] == successor->ImmPredecessors().size()) {
+				pass.Seal(successor);
+			}
+		}
 	}
 }
 

src/graphics/shader/rectListShader.cpp

@@ -18,45 +18,9 @@ using ShaderRecompiler::Spirv::Builder;
 
 constexpr uint32_t SpirvVersion15 = 0x00010500u;
 
-struct Parameter {
-	uint32_t input_location  = 0;
-	uint32_t output_location = 0;
-	bool     flat            = false;
-};
-
-std::vector<Parameter> GetParameters(const ShaderVertexInputInfo& vertex_info,
-                                     const ShaderPixelInputInfo*  pixel_info) {
-	if (pixel_info == nullptr) {
-		return {};
-	}
-	EXIT_IF(pixel_info->input_num > ShaderVertexInputInfo::RES_MAX);
-	EXIT_IF(pixel_info->stage.program == nullptr);
-
-	std::vector<uint32_t> active_inputs;
-	for (const auto& input: pixel_info->stage.program->info.inputs) {
-		if (input.kind == ShaderRecompiler::IR::StageInputKind::Parameter) {
-			active_inputs.push_back(input.location);
-		}
-	}
-
-	std::vector<Parameter> parameters;
-	std::array<bool, 32> output_locations {};
-	for (const auto input: active_inputs) {
-		const auto input_location = ShaderPixelParameterMappedLocation(*pixel_info, 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;
-}
-
 class RectListEmitter {
 public:
-	RectListEmitter(const std::vector<Parameter>& parameters_, spv::ExecutionModel model)
+	RectListEmitter(const std::vector<RectListParameter>& parameters_, spv::ExecutionModel model)
 	    : parameters(parameters_) {
 		builder.AddMemoryModel(spv::AddressingModelLogical, spv::MemoryModelGLSL450);
 
@@ -315,42 +279,68 @@ private:
 		              Result(spv::OpFAdd, vec4_float_type, p1, p2));
 	}
 
-	Builder                       builder {SpirvVersion15};
-	const std::vector<Parameter>& parameters;
-	std::vector<uint32_t>         interfaces;
-	std::vector<uint32_t>         inputs;
-	std::vector<uint32_t>         outputs;
-	uint32_t                      main                  = 0;
-	uint32_t                      void_type             = 0;
-	uint32_t                      bool_type             = 0;
-	uint32_t                      uint_type             = 0;
-	uint32_t                      int_type              = 0;
-	uint32_t                      float_type            = 0;
-	uint32_t                      vec2_bool_type        = 0;
-	uint32_t                      vec2_float_type       = 0;
-	uint32_t                      vec3_float_type       = 0;
-	uint32_t                      vec4_float_type       = 0;
-	uint32_t                      function_type         = 0;
-	uint32_t                      per_vertex_type       = 0;
-	uint32_t                      ptr_input_float       = 0;
-	uint32_t                      ptr_input_vec4_float  = 0;
-	uint32_t                      ptr_output_float      = 0;
-	uint32_t                      ptr_output_vec4_float = 0;
-	uint32_t                      gl_in                 = 0;
-	uint32_t                      gl_out                = 0;
-	uint32_t                      tess_inner            = 0;
-	uint32_t                      tess_outer            = 0;
-	uint32_t                      tess_coord            = 0;
-	uint32_t                      invocation_id         = 0;
+	Builder                               builder {SpirvVersion15};
+	const std::vector<RectListParameter>& parameters;
+	std::vector<uint32_t>                 interfaces;
+	std::vector<uint32_t>                 inputs;
+	std::vector<uint32_t>                 outputs;
+	uint32_t                              main                  = 0;
+	uint32_t                              void_type             = 0;
+	uint32_t                              bool_type             = 0;
+	uint32_t                              uint_type             = 0;
+	uint32_t                              int_type              = 0;
+	uint32_t                              float_type            = 0;
+	uint32_t                              vec2_bool_type        = 0;
+	uint32_t                              vec2_float_type       = 0;
+	uint32_t                              vec3_float_type       = 0;
+	uint32_t                              vec4_float_type       = 0;
+	uint32_t                              function_type         = 0;
+	uint32_t                              per_vertex_type       = 0;
+	uint32_t                              ptr_input_float       = 0;
+	uint32_t                              ptr_input_vec4_float  = 0;
+	uint32_t                              ptr_output_float      = 0;
+	uint32_t                              ptr_output_vec4_float = 0;
+	uint32_t                              gl_in                 = 0;
+	uint32_t                              gl_out                = 0;
+	uint32_t                              tess_inner            = 0;
+	uint32_t                              tess_outer            = 0;
+	uint32_t                              tess_coord            = 0;
+	uint32_t                              invocation_id         = 0;
 };
 
 } // namespace
 
-RectListShaders BuildRectListShaders(const ShaderVertexInputInfo& vertex_info,
-                                     const ShaderPixelInputInfo*  pixel_info) {
-	const auto      parameters = GetParameters(vertex_info, pixel_info);
-	RectListEmitter control(parameters, spv::ExecutionModelTessellationControl);
-	RectListEmitter evaluation(parameters, spv::ExecutionModelTessellationEvaluation);
+std::vector<RectListParameter> RectListParameters(
+    const ShaderPixelInputInfo& pixel_info, const ShaderRecompiler::IR::CompiledShaderInfo& program) {
+	EXIT_IF(pixel_info.input_num > ShaderVertexInputInfo::RES_MAX);
+	std::vector<uint32_t> active_inputs;
+	for (const auto& input: program.info.inputs) {
+		if (input.kind == ShaderRecompiler::IR::StageInputKind::Parameter) {
+			active_inputs.push_back(input.location);
+		}
+	}
+	std::vector<RectListParameter> parameters;
+	for (const auto input: active_inputs) {
+		parameters.push_back({ShaderPixelParameterMappedLocation(pixel_info, input),
+		                      ShaderPixelParameterLocation(pixel_info, active_inputs, input),
+		                      ShaderPixelParameterIsFlat(pixel_info, input)});
+	}
+	return parameters;
+}
+
+RectListShaders BuildRectListShaders(std::span<const RectListParameter> parameters,
+                                     uint32_t                           param_export_mask) {
+	std::vector<RectListParameter> exported;
+	std::array<bool, 32>           output_locations {};
+	for (const auto& parameter: parameters) {
+		if ((param_export_mask & (1u << parameter.input_location)) != 0 &&
+		    !output_locations[parameter.output_location]) {
+			exported.push_back(parameter);
+			output_locations[parameter.output_location] = true;
+		}
+	}
+	RectListEmitter control(exported, spv::ExecutionModelTessellationControl);
+	RectListEmitter evaluation(exported, spv::ExecutionModelTessellationEvaluation);
 	return {control.EmitControl(), evaluation.EmitEvaluation()};
 }
 

src/graphics/shader/rectListShader.h

@@ -2,20 +2,35 @@
 #define EMULATOR_SRC_GRAPHICS_SHADER_RECTLISTSHADER_H_
 
 #include <cstdint>
+#include <span>
 #include <vector>
 
 namespace Libs::Graphics {
 
 struct ShaderPixelInputInfo;
-struct ShaderVertexInputInfo;
+
+namespace ShaderRecompiler::IR {
+struct CompiledShaderInfo;
+} // namespace ShaderRecompiler::IR
+
+// A pixel parameter the generated stages forward from vertex export input_location.
+struct RectListParameter {
+	uint32_t input_location  = 0;
+	uint32_t output_location = 0;
+	bool     flat            = false;
+};
 
 struct RectListShaders {
 	std::vector<uint32_t> control;
 	std::vector<uint32_t> evaluation;
 };
 
-RectListShaders BuildRectListShaders(const ShaderVertexInputInfo& vertex_info,
-                                     const ShaderPixelInputInfo*  pixel_info);
+// Every parameter a rectangle-list pixel program reads.
+std::vector<RectListParameter> RectListParameters(
+    const ShaderPixelInputInfo& pixel_info, const ShaderRecompiler::IR::CompiledShaderInfo& program);
+// Forwards the parameters the vertex stage exports, one per output location.
+RectListShaders BuildRectListShaders(std::span<const RectListParameter> parameters,
+                                     uint32_t                           param_export_mask);
 
 } // namespace Libs::Graphics
 

src/graphics/shader/shader.cpp

@@ -11,7 +11,6 @@
 #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"
@@ -40,15 +39,19 @@
 
 namespace Libs::Graphics {
 
+// A mapped shader and its vertex metadata, parsed once when it is mapped.
+struct ShaderMapData: ShaderMappedData {
+	ShaderVertexMetadata vertex;
+	const char*          vertex_error = nullptr;
+};
+
 struct ShaderMapEntry {
-	ShaderMappedData data;
-	uint64_t hash;
+	ShaderMapData data;
+	uint64_t      hash;
 };
 
 static std::unique_ptr<std::unordered_map<uint64_t, ShaderMapEntry>> g_shader_map;
 static std::mutex                                                      g_shader_map_mutex;
-static std::unordered_map<uint64_t, ShaderTessellationInputInfo>       g_tess_layouts;
-static std::mutex                                                      g_tess_layouts_mutex;
 
 void ShaderInit() {
 	EXIT_IF(g_shader_map != nullptr);
@@ -64,8 +67,11 @@ void ShaderMapUserData(uint64_t addr, const ShaderMappedData& data) {
 		     addr, data.code_size_bytes);
 	}
 	const auto hash = XXH3_64bits(reinterpret_cast<const void*>(addr), data.code_size_bytes);
+	ShaderMapData mapped {data};
+	mapped.vertex_error =
+	    ShaderReadVertexMetadata(data, HW::UserSgprInfo::SGPRS_MAX, mapped.vertex);
 	std::scoped_lock lock(g_shader_map_mutex);
-	(*g_shader_map)[addr] = {data, hash};
+	(*g_shader_map)[addr] = {mapped, hash};
 }
 
 static ShaderMapEntry ShaderGetMappedData(uint64_t addr, const char* label) {
@@ -261,8 +267,8 @@ static uint32_t ShaderCalcPsSystemInputBase(const HW::ShaderRegisters& regs) {
 
 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) {
+                                           const ShaderMapData&   data,
+                                           ShaderVertexInputInfo& info) {
 	KYTY_PROFILER_FUNCTION();
 
 	info = {};
@@ -277,13 +283,12 @@ static bool ShaderGetStaticVertexInputInfo(uint64_t shader_addr, const HW::UserS
 		     " 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)) {
+	if (data.vertex_error != nullptr) {
 		LOGF("ShaderGetVertexInputInfo(): invalid AGC metadata shader=0x%016" PRIx64 ": %s\n",
-		     shader_addr, metadata_error.c_str());
+		     shader_addr, data.vertex_error);
 		return false;
 	}
+	const auto& metadata = data.vertex;
 
 	if (metadata.vertex_attrib_reg >= 0) {
 		info.fetch_external   = false;
@@ -377,7 +382,6 @@ static void ShaderGetStaticInputInfoPS(
 }
 
 static void ShaderGetStaticInputInfoCS(const HW::ComputeShaderInfo& regs,
-                                       const HW::ShaderRegisters& /*sh*/,
                                        const ShaderMappedData& data, ShaderComputeInputInfo& info) {
 	const bool async_compute             = info.async_compute;
 	const bool dispatch_thread_dimensions = info.dispatch_thread_dimensions;
@@ -418,6 +422,7 @@ void BuildStageStaticKey(const ShaderVertexInputInfo& info, std::vector<uint32_t
 	key.push_back(info.lds_size_dwords);
 	key.push_back(info.pa_cl_vs_out_cntl);
 	key.push_back(info.user_clip_planes);
+	key.push_back(info.indirect_offsets.vertex | (uint32_t {info.indirect_offsets.instance} << 8u));
 	key.push_back(static_cast<uint32_t>(info.clip_space.enabled));
 	if (info.clip_space.enabled) {
 		for (const float value: info.clip_space.scale) {
@@ -486,7 +491,9 @@ void BuildStageStaticKey(const ShaderPixelInputInfo& info, std::vector<uint32_t>
 	key.push_back(static_cast<uint32_t>(info.object_primitive_id));
 	key.push_back(static_cast<uint32_t>(info.alpha_blend_source));
 	key.insert(key.end(), std::begin(info.target_output_mode), std::end(info.target_output_mode));
-	key.push_back(info.target_packed_unorm_mask | (uint32_t {info.target_round_mode} << 8u));
+	key.push_back(info.target_packed_unorm_mask | (uint32_t {info.target_round_mode} << 8u) |
+	              (uint32_t {info.target_uint_mask} << 16u) |
+	              (uint32_t {info.target_sint_mask} << 24u));
 	const auto mask = info.target_shader_mask;
 	key.push_back((mask | (mask >> 1u) | (mask >> 2u) | (mask >> 3u)) & 0x11111111u);
 	for (uint32_t base = 0; base < info.target_export_mapping.size(); base += 4u) {
@@ -655,18 +662,6 @@ PrepareTessellationPrograms(const HW::VertexShaderInfo& regs, const HW::Context&
 	};
 	EXIT_IF(tess.input_control_points == 0 || tess.input_control_points > 32 ||
 	        tess.output_control_points == 0 || tess.output_control_points > 32);
-	// Reflection decodes both programs; the key also covers the VGT fields cached in tess.
-	const uint64_t key[] = {local_hash, control_hash,
-	                        uint64_t {sh.m_vgtLsHsConfig & 0xfff00u} << 32u | sh.m_vgtTfParam};
-	{
-		std::scoped_lock lock(g_tess_layouts_mutex);
-		auto [entry, inserted] = g_tess_layouts.try_emplace(XXH3_64bits(key, sizeof(key)), tess);
-		if (inserted) {
-			ShaderRecompiler::AnalyzeTessellationPrograms(params[0].code, params[1].code,
-			                                              entry->second);
-		}
-		tess = entry->second;
-	}
 	for (auto& stage: input_info) {
 		stage.tess = tess;
 	}
@@ -684,10 +679,9 @@ ShaderParams PrepareProgram(
 	    std::span<const uint32_t>(regs.ps_user_sgpr.value, regs.ps_regs.rsrc2.user_sgpr), data);
 }
 
-ShaderParams PrepareProgram(const HW::ComputeShaderInfo& regs, const HW::ShaderRegisters& sh,
-                            ShaderComputeInputInfo& info) {
+ShaderParams PrepareProgram(const HW::ComputeShaderInfo& regs, ShaderComputeInputInfo& info) {
 	const auto [data, hash] = ShaderGetMappedData(regs.cs_regs.data_addr, "ShaderGetInputInfoCS():");
-	ShaderGetStaticInputInfoCS(regs, sh, data, info);
+	ShaderGetStaticInputInfoCS(regs, data, info);
 	return GetShaderParams(
 	    regs.cs_regs.data_addr, hash,
 	    std::span<const uint32_t>(regs.cs_user_sgpr.value, regs.cs_regs.user_sgpr), data);

src/graphics/shader/shader.h

@@ -15,17 +15,6 @@
 
 namespace Libs::Graphics {
 
-namespace ShaderError {
-
-[[nodiscard]] inline bool Fail(std::string* error, std::string_view message) {
-	if (error != nullptr) {
-		error->assign(message.data(), message.size());
-	}
-	return false;
-}
-
-} // namespace ShaderError
-
 namespace HW {
 struct VertexShaderInfo;
 struct PixelShaderInfo;
@@ -45,16 +34,14 @@ enum class ShaderType {
 	TessellationEvaluation
 };
 
-namespace ShaderRecompiler::IR {
-struct CompiledShaderInfo;
-} // namespace ShaderRecompiler::IR
+struct ShaderModule;
 
 struct ShaderStageRuntime {
-	const ShaderRecompiler::IR::CompiledShaderInfo* program = nullptr;
-	const ShaderRecompiler::IR::ResourceSnapshot*   resources = nullptr;
+	const ShaderModule*                           module    = nullptr;
+	const ShaderRecompiler::IR::ResourceSnapshot* resources = nullptr;
 
 	[[nodiscard]] explicit operator bool() const {
-		return program != nullptr && resources != nullptr;
+		return module != nullptr && resources != nullptr;
 	}
 };
 
@@ -137,6 +124,18 @@ struct ShaderTessellationInputInfo {
 	uint32_t output_topology       = 0;
 };
 
+// GS user-data slots an indirect draw packet writes its record's base vertex and first instance
+// to. Like the hardware index VGPRs, the vertex shader's indices exclude both bases once either
+// slot is set.
+struct ShaderIndirectOffsets {
+	static constexpr uint8_t None = UINT8_MAX;
+
+	uint8_t vertex   = None;
+	uint8_t instance = None;
+
+	bool operator==(const ShaderIndirectOffsets&) const = default;
+};
+
 struct ShaderVertexInputInfo {
 	static constexpr int RES_MAX = 32;
 
@@ -159,6 +158,7 @@ struct ShaderVertexInputInfo {
 	ShaderClipSpaceTransform clip_space;
 	ShaderMeshInputInfo      mesh;
 	ShaderTessellationInputInfo tess;
+	ShaderIndirectOffsets   indirect_offsets;
 	bool                    fetch_external      = false;
 	bool                    fetch_embedded      = false;
 };
@@ -194,6 +194,9 @@ struct ShaderPixelInputInfo {
 	uint32_t                                       target_shader_mask           = UINT32_MAX;
 	std::array<Prospero::ColorComponentMapping, 8> target_export_mapping        = {};
 	uint8_t                                        target_packed_unorm_mask      = 0;
+	// Exports to UINT and SINT targets, which take integer outputs of the same signedness.
+	uint8_t                                        target_uint_mask             = 0;
+	uint8_t                                        target_sint_mask             = 0;
 	// CB round mode per converted export: zero adds half an LSB, one truncates.
 	uint8_t                                        target_round_mode            = 0;
 	uint32_t                                       scratch_size_dwords          = 0;

src/graphics/shader/shaderCompiler.h

@@ -32,6 +32,7 @@ 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);
+// Each stage's tess holds the VGT fields; the caller adds the layout reflected from the programs.
 std::array<ShaderParams, 3>
 PrepareTessellationPrograms(const HW::VertexShaderInfo& regs, const HW::Context& context,
                             std::array<ShaderVertexInputInfo, 3>& input_info);
@@ -39,8 +40,7 @@ ShaderParams PrepareProgram(
     const HW::PixelShaderInfo& regs, const HW::ShaderRegisters& sh,
     std::span<const Prospero::ColorComponentMapping, 8> target_export_mapping,
     ShaderPixelInputInfo&                               input_info);
-ShaderParams PrepareProgram(const HW::ComputeShaderInfo& regs, const HW::ShaderRegisters& sh,
-                            ShaderComputeInputInfo& input_info);
+ShaderParams PrepareProgram(const HW::ComputeShaderInfo& regs, ShaderComputeInputInfo& input_info);
 
 } // namespace Libs::Graphics
 

src/graphics/shader/shaderVertexMetadata.cpp

@@ -5,10 +5,10 @@
 
 namespace Libs::Graphics {
 
-bool ShaderReadVertexMetadata(const ShaderMappedData& data, uint32_t max_user_sgprs,
-                              ShaderVertexMetadata& metadata, std::string* error) {
+const char* ShaderReadVertexMetadata(const ShaderMappedData& data, uint32_t max_user_sgprs,
+                                     ShaderVertexMetadata& metadata) {
 	if (data.user_data == nullptr) {
-		return ShaderError::Fail(error, "missing AGC user-data header");
+		return "missing AGC user-data header";
 	}
 
 	ShaderUserData user_data {};
@@ -17,8 +17,7 @@ bool ShaderReadVertexMetadata(const ShaderMappedData& data, uint32_t max_user_sg
 	constexpr uint32_t DirectResourceCount =
 	    static_cast<uint32_t>(AgcDirectResourceType::Last) + 1u;
 	if (user_data.direct_resource_count > DirectResourceCount) {
-		return ShaderError::Fail(error,
-		                         "AGC direct-resource count exceeds the known resource domain");
+		return "AGC direct-resource count exceeds the known resource domain";
 	}
 
 	std::array<uint16_t, DirectResourceCount> direct_offsets {};
@@ -26,7 +25,7 @@ bool ShaderReadVertexMetadata(const ShaderMappedData& data, uint32_t max_user_sg
 	    static_cast<uint64_t>(user_data.direct_resource_count) * sizeof(uint16_t);
 	if (direct_size != 0) {
 		if (user_data.direct_resource_offset == nullptr) {
-			return ShaderError::Fail(error, "missing AGC direct-resource offsets");
+			return "missing AGC direct-resource offsets";
 		}
 		std::memcpy(direct_offsets.data(), user_data.direct_resource_offset, direct_size);
 	}
@@ -47,33 +46,33 @@ bool ShaderReadVertexMetadata(const ShaderMappedData& data, uint32_t max_user_sg
 	}
 
 	if (next.vertex_attrib_reg >= 0 && next.vertex_buffer_reg < 0) {
-		return ShaderError::Fail(error, "vertex attribute table requires a vertex buffer table");
+		return "vertex attribute table requires a vertex buffer table";
 	}
 	if (next.vertex_buffer_reg < 0) {
 		metadata = next;
-		return true;
+		return nullptr;
 	}
 	if (static_cast<uint32_t>(next.vertex_buffer_reg) + 1u >= max_user_sgprs) {
-		return ShaderError::Fail(error, "vertex table pointer exceeds the user-SGPR domain");
+		return "vertex table pointer exceeds the user-SGPR domain";
 	}
 	if (next.vertex_attrib_reg < 0) {
 		metadata = next;
-		return true;
+		return nullptr;
 	}
 	if (static_cast<uint32_t>(next.vertex_attrib_reg) + 1u >= max_user_sgprs) {
-		return ShaderError::Fail(error, "vertex table pointer exceeds the user-SGPR domain");
+		return "vertex table pointer exceeds the user-SGPR domain";
 	}
 	if (data.num_input_semantics == 0 ||
 	    data.num_input_semantics > ShaderVertexInputInfo::RES_MAX) {
-		return ShaderError::Fail(error, "vertex semantic count is outside the supported domain");
+		return "vertex semantic count is outside the supported domain";
 	}
 
 	if (data.input_semantics == nullptr) {
-		return ShaderError::Fail(error, "missing vertex input semantics");
+		return "missing vertex input semantics";
 	}
 	next.input_semantics = {data.input_semantics, data.num_input_semantics};
 	metadata                   = next;
-	return true;
+	return nullptr;
 }
 
 } // namespace Libs::Graphics

src/graphics/shader/shaderVertexMetadata.h

@@ -4,7 +4,6 @@
 #include "graphics/shader/shader.h"
 
 #include <span>
-#include <string>
 
 namespace Libs::Graphics {
 
@@ -14,9 +13,10 @@ struct ShaderVertexMetadata {
 	std::span<const ShaderSemantic> input_semantics;
 };
 
-// Validates the vertex metadata and returns a view into the shader header.
-bool ShaderReadVertexMetadata(const ShaderMappedData& data, uint32_t max_user_sgprs,
-                              ShaderVertexMetadata& metadata, std::string* error);
+// Reads the vertex metadata as a view into the shader header. Returns nullptr if it is
+// valid, otherwise why it was rejected.
+const char* ShaderReadVertexMetadata(const ShaderMappedData& data, uint32_t max_user_sgprs,
+                                     ShaderVertexMetadata& metadata);
 
 } // namespace Libs::Graphics
 

src/kernel/memory.cpp

@@ -9,6 +9,7 @@
 #include "graphics/host_gpu/renderer/renderContext.h"
 #include "libs/errno.h"
 #include "libs/libs.h"
+#include "loader/systemContent.h"
 
 #include <algorithm>
 #include <array>
@@ -63,15 +64,21 @@ constexpr int PROT_GPU_WRITE = 0x20;
 
 enum class GpuAccessMode { NoAccess, Read, Write, ReadWrite };
 
-constexpr uint64_t PAGE_TABLE_POOL_SIZE   = 4ull * 1024ull * 1024ull * 1024ull;
-constexpr uint64_t PAGE_TABLE_GRANULARITY = 2ull * 1024ull * 1024ull;
-constexpr int      PAGE_TABLE_POOL_ENTRIES =
-    static_cast<int>(PAGE_TABLE_POOL_SIZE / PAGE_TABLE_GRANULARITY);
-constexpr uint64_t DEFAULT_FLEXIBLE_MEMORY_SIZE = 1ull * 1024ull * 1024ull * 1024ull;
+constexpr uint64_t PAGE_TABLE_GRANULARITY       = 2ull * 1024ull * 1024ull;
+constexpr uint64_t DEFAULT_FLEXIBLE_MEMORY_SIZE = 448ull * 1024ull * 1024ull;
+// A game gets 12.5 GiB. A fixed 64 MiB cache and mapping table memory past the 64 MiB included
+// for each of the CPU, GPU and AMM tables come off the top; flexible and direct memory share what
+// is left. One 4 KiB table maps 2 MiB.
+constexpr uint64_t GUEST_MEMORY_BUDGET   = 12800ull * 1024ull * 1024ull;
+constexpr uint64_t RESERVED_CACHE_SIZE   = 64ull * 1024ull * 1024ull;
+constexpr uint64_t INCLUDED_TABLE_MEMORY = 64ull * 1024ull * 1024ull;
+constexpr uint64_t MAPPING_TABLE_SIZE    = 4ull * 1024ull;
 
-static uint64_t                      g_flexible_memory_size        = DEFAULT_FLEXIBLE_MEMORY_SIZE;
-static bool                          g_flexible_memory_size_frozen = false;
-static Graphics::RenderContext*       g_gpu_resources               = nullptr;
+static uint64_t                 g_guest_memory_size    = GUEST_MEMORY_BUDGET - RESERVED_CACHE_SIZE;
+static uint64_t                 g_flexible_memory_size = DEFAULT_FLEXIBLE_MEMORY_SIZE;
+static uint64_t                 g_cpu_table_memory     = INCLUDED_TABLE_MEMORY;
+static uint64_t                 g_gpu_table_memory     = INCLUDED_TABLE_MEMORY;
+static Graphics::RenderContext* g_gpu_resources        = nullptr;
 #if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
 static void (*g_test_gpu_range)(bool, uint64_t, size_t) = nullptr;
 #endif
@@ -664,11 +671,8 @@ public:
 
 	KYTY_CLASS_NO_COPY(PhysicalMemory);
 
-	static constexpr uint64_t TotalSize() { return static_cast<uint64_t>(13824) * 1024 * 1024; }
-	static uint64_t           Size() {
-		EXIT_IF(g_flexible_memory_size >= TotalSize());
-		return TotalSize() - g_flexible_memory_size;
-	}
+	static uint64_t TotalSize() { return g_guest_memory_size; }
+	static uint64_t Size() { return TotalSize() - g_flexible_memory_size; }
 
 	bool Alloc(uint64_t search_start, uint64_t search_end, size_t len, size_t alignment,
 	           uint64_t* phys_addr_out, int memory_type,
@@ -738,6 +742,8 @@ public:
 
 	static uint64_t Size() { return g_flexible_memory_size; }
 	uint64_t        Available();
+	void            Charge(uint64_t size);
+	void            Release(uint64_t size);
 
 	bool Map(uint64_t vaddr, size_t len, int prot, VirtualMemory::Mode mode, GpuAccessMode gpu_mode,
 	         const char* name);
@@ -996,7 +1002,22 @@ static bool SelfTestSub64SharedPlaceholderAlias() {
 static bool ReplaceFixedRangeWithReserved(uint64_t start, uint64_t size);
 
 void Initialize() {
-	g_flexible_memory_size_frozen = true;
+	const auto params = Loader::SystemContentGetMemoryParams();
+	if (params.flexible_size != 0) {
+		g_flexible_memory_size = params.flexible_size;
+	}
+	g_cpu_table_memory = std::max(params.cpu_table_memory, INCLUDED_TABLE_MEMORY);
+	g_gpu_table_memory = std::max(params.gpu_table_memory, INCLUDED_TABLE_MEMORY);
+
+	const auto table_overhead = g_cpu_table_memory + g_gpu_table_memory +
+	                            std::max(params.amm_table_memory, INCLUDED_TABLE_MEMORY) -
+	                            3 * INCLUDED_TABLE_MEMORY;
+	EXIT_IF(g_flexible_memory_size + table_overhead >= GUEST_MEMORY_BUDGET - RESERVED_CACHE_SIZE);
+	g_guest_memory_size = GUEST_MEMORY_BUDGET - RESERVED_CACHE_SIZE - table_overhead;
+	LOGF("\t flexible memory size = %" PRIu64 " MiB, direct memory size = %" PRIu64 " MiB\n",
+	     g_flexible_memory_size / (1024ull * 1024ull),
+	     (g_guest_memory_size - g_flexible_memory_size) / (1024ull * 1024ull));
+
 	VirtualMemory::Init();
 	g_guest_address_space = std::make_unique<GuestAddressSpace>(PhysicalMemory::TotalSize());
 	g_physical_memory     = std::make_unique<PhysicalMemory>();
@@ -1064,15 +1085,6 @@ void RegisterCallbacks(callback_func_t alloc_func, callback_func_t free_func) {
 	}
 }
 
-void SetFlexibleMemorySize(uint64_t size) {
-	constexpr uint64_t GuestPageSize = 0x4000;
-	EXIT_IF(g_flexible_memory_size_frozen || g_guest_address_space != nullptr);
-	EXIT_IF(size == 0 || (size & (GuestPageSize - 1u)) != 0 || size >= PhysicalMemory::TotalSize());
-	g_flexible_memory_size = size;
-	LOGF("\t flexible memory size = 0x%016" PRIx64 " (%" PRIu64 " MiB)\n", size,
-	     size / (1024ull * 1024ull));
-}
-
 bool PhysicalMemory::Alloc(uint64_t search_start, uint64_t search_end, size_t len, size_t alignment,
                            uint64_t* phys_addr_out, int memory_type, AllocationKind kind) {
 	if (phys_addr_out == nullptr) {
@@ -1967,6 +1979,19 @@ uint64_t FlexibleMemory::Available() {
 	return (Size() >= m_allocated_total ? Size() - m_allocated_total : 0);
 }
 
+void FlexibleMemory::Charge(uint64_t size) {
+	Common::LockGuard lock(m_mutex);
+
+	m_allocated_total += size;
+}
+
+void FlexibleMemory::Release(uint64_t size) {
+	Common::LockGuard lock(m_mutex);
+
+	EXIT_IF(size > m_allocated_total);
+	m_allocated_total -= size;
+}
+
 void PooledMemory::AddFreeUnlocked(uint64_t start, uint64_t size) {
 	if (size == 0) {
 		return;
@@ -2640,12 +2665,12 @@ int KYTY_SYSV_ABI KernelGetPageTableStats(int* cpu_total, int* cpu_available, in
 	const auto gpu_used =
 	    (g_virtual_ranges != nullptr ? g_virtual_ranges->CountPageTableEntries(true) : 0);
 
-	*cpu_total     = PAGE_TABLE_POOL_ENTRIES;
-	*gpu_total     = PAGE_TABLE_POOL_ENTRIES;
-	*cpu_available = PAGE_TABLE_POOL_ENTRIES -
-	                 static_cast<int>(std::min<uint64_t>(cpu_used, PAGE_TABLE_POOL_ENTRIES));
-	*gpu_available = PAGE_TABLE_POOL_ENTRIES -
-	                 static_cast<int>(std::min<uint64_t>(gpu_used, PAGE_TABLE_POOL_ENTRIES));
+	const auto cpu_tables = g_cpu_table_memory / MAPPING_TABLE_SIZE;
+	const auto gpu_tables = g_gpu_table_memory / MAPPING_TABLE_SIZE;
+	*cpu_total            = static_cast<int>(cpu_tables);
+	*gpu_total            = static_cast<int>(gpu_tables);
+	*cpu_available        = static_cast<int>(cpu_tables - std::min(cpu_used, cpu_tables));
+	*gpu_available        = static_cast<int>(gpu_tables - std::min(gpu_used, gpu_tables));
 
 	LOGF_COLOR(Log::Color::Green,
 	           "\t cpu_total     = %d\n"
@@ -3770,6 +3795,14 @@ uint64_t AllocateProgramMemory(uint64_t search_addr, uint64_t size, VirtualMemor
 	return AllocateGuestRuntimeMemory(search_addr, size, mode, name, VirtualRangeType::Code, false);
 }
 
+void ChargeFlexibleMemory(uint64_t size) {
+	g_flexible_memory->Charge(size);
+}
+
+void ReleaseFlexibleMemory(uint64_t size) {
+	g_flexible_memory->Release(size);
+}
+
 void SetProgramMemoryProtection(uint64_t vaddr, uint64_t size, VirtualMemory::Mode mode) {
 	std::lock_guard<std::recursive_mutex> memory_operation_lock(g_memory_operation_mutex);
 
@@ -4358,7 +4391,8 @@ DebugStats GetDebugStats() {
 	stats.direct_total       = PhysicalMemory::Size();
 	stats.flexible_total     = FlexibleMemory::Size();
 	stats.pool_committed     = g_memory_pool_committed.load(std::memory_order_relaxed);
-	stats.page_entries_total = PAGE_TABLE_POOL_ENTRIES;
+	stats.cpu_page_entries_total = g_cpu_table_memory / MAPPING_TABLE_SIZE;
+	stats.gpu_page_entries_total = g_gpu_table_memory / MAPPING_TABLE_SIZE;
 	if (g_physical_memory != nullptr) {
 		Common::LockGuard lock(g_physical_memory->GetMutex());
 		for (const auto& [start, block]: g_physical_memory->GetPhysicalBlocks()) {

src/kernel/memory.h

@@ -26,9 +26,10 @@ struct DebugStats {
 	uint64_t flexible_total      = 0;
 	uint64_t flexible_used       = 0;
 	uint64_t pool_committed      = 0;
-	uint64_t cpu_page_entries    = 0; // 2 MiB page-table entries in use.
-	uint64_t gpu_page_entries    = 0;
-	uint64_t page_entries_total  = 0; // Per side.
+	uint64_t cpu_page_entries       = 0; // 2 MiB page-table entries in use.
+	uint64_t gpu_page_entries       = 0;
+	uint64_t cpu_page_entries_total = 0;
+	uint64_t gpu_page_entries_total = 0;
 };
 [[nodiscard]] DebugStats GetDebugStats();
 
@@ -125,7 +126,6 @@ static_assert(sizeof(KernelMemoryPoolBlockStats) == 16,
               "KernelMemoryPoolBlockStats struct size is incorrect");
 
 void                   RegisterCallbacks(callback_func_t alloc_func, callback_func_t free_func);
-void                   SetFlexibleMemorySize(uint64_t size);
 int AllocateDirectMemory(int64_t search_start, int64_t search_end, size_t size, size_t alignment,
                          int memory_type, int64_t* phys_addr_out, bool automatic = false);
 int MapAutomaticMemory(uint64_t vaddr, size_t size, int type, int prot);
@@ -197,6 +197,9 @@ int KYTY_SYSV_ABI KernelMemoryPoolGetBlockStats(KernelMemoryPoolBlockStats* outp
 
 uint64_t AllocateProgramMemory(uint64_t search_addr, uint64_t size,
                                Common::VirtualMemory::Mode mode, const char* name);
+// Loaded modules' segments use flexible memory without taking its backing.
+void ChargeFlexibleMemory(uint64_t size);
+void ReleaseFlexibleMemory(uint64_t size);
 void SetProgramMemoryProtection(uint64_t vaddr, uint64_t size, Common::VirtualMemory::Mode mode);
 uint64_t AllocateRuntimeMemory(uint64_t search_addr, uint64_t size,
                                Common::VirtualMemory::Mode mode, const char* name,

src/kernel/memoryAddressSpace.inc

@@ -820,12 +820,12 @@ public:
 		return ProtectMappedUnlocked(vaddr, size, mode);
 	}
 
+	// Tracker protections only read m_mapped; the page tracker orders them per page.
 	bool ProtectTransient(uint64_t vaddr, uint64_t size, VirtualMemory::Mode mode) {
 		if (size == 0 || UINT64_MAX - vaddr < size) {
 			return false;
 		}
-
-		std::lock_guard lock(m_mutex);
+		std::shared_lock lock(m_mutex);
 		return ProtectMappedUnlocked(vaddr, size, mode);
 	}