graphics: cache the macOS thread id in TrackingSpinLock (#1169)

Every lock and unlock called pthread_mach_thread_np. Cache it per thread like the Linux branch does.

CMakeLists.txt

@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.18...4.3)
+cmake_minimum_required(VERSION 3.19...4.3)
 
 if(CMAKE_VERSION VERSION_GREATER_EQUAL 4.0)
 	set(CMAKE_POLICY_VERSION_MINIMUM 3.10 CACHE STRING "Minimum policy version for bundled dependencies")
@@ -413,7 +413,16 @@ add_executable(archive_file_tests EXCLUDE_FROM_ALL
 target_link_libraries(archive_file_tests common ZArchive::zarchive)
 target_include_directories(archive_file_tests PRIVATE ${inc_headers})
 
-add_executable(avplayer_file_tests EXCLUDE_FROM_ALL "${KYTY_TESTS_DIR}/AvPlayerFileTests.cpp")
+add_executable(trophy_system_tests EXCLUDE_FROM_ALL
+	"${KYTY_TESTS_DIR}/TrophySystemTests.cpp"
+)
+target_link_libraries(trophy_system_tests common)
+target_include_directories(trophy_system_tests PRIVATE ${inc_headers})
+
+add_executable(avplayer_file_tests EXCLUDE_FROM_ALL
+	"${KYTY_TESTS_DIR}/AvPlayerFileTests.cpp"
+	"${KYTY_SOURCE_DIR}/loader/timer.cpp"
+)
 target_link_libraries(avplayer_file_tests common FFmpeg::ffmpeg ZArchive::zarchive)
 target_include_directories(avplayer_file_tests PRIVATE ${inc_headers})
 
@@ -606,6 +615,7 @@ if(BUILD_TESTING)
 	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 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>)
 	add_test(NAME bit_array COMMAND $<TARGET_FILE:bit_array_tests>)
 	add_test(NAME lru_cache COMMAND $<TARGET_FILE:lru_cache_tests>)
@@ -664,15 +674,12 @@ if(BUILD_TESTING)
 	add_test(NAME color_render_target_1d
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --rt1d-only)
 
-	if(WIN32)
-		# These tests still depend on the Windows multisample-depth path.
-		add_test(NAME texture_cache_image_overlap
-			COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --image-overlap-only)
-		add_test(NAME texture_cache_htile_clear
-			COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --htile-clear-only)
-		add_test(NAME buffer_cache_ranges
-			COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --buffer-cache-range-only)
-	endif()
+	add_test(NAME texture_cache_image_overlap
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --image-overlap-only)
+	add_test(NAME texture_cache_htile_clear
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --htile-clear-only)
+	add_test(NAME buffer_cache_ranges
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --buffer-cache-range-only)
 
 	add_custom_target(kyty_tests DEPENDS
 		ime_dialog_tests
@@ -682,6 +689,7 @@ if(BUILD_TESTING)
 		memory_tracker_tests
 		page_manager_tests
 		archive_file_tests
+		trophy_system_tests
 		avplayer_file_tests
 		bit_array_tests
 		lru_cache_tests
@@ -712,8 +720,9 @@ if(APPLE AND (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR
 	add_executable(macos_sse4a_tests EXCLUDE_FROM_ALL
 		"${KYTY_TESTS_DIR}/macosSse4aTests.cpp"
 		"${KYTY_SOURCE_DIR}/loader/x64InstructionEmulator.cpp"
+		"${KYTY_SOURCE_DIR}/loader/guestInstructionPatcher.cpp"
 	)
-	target_link_libraries(macos_sse4a_tests common fmt::fmt Zydis::Zydis)
+	target_link_libraries(macos_sse4a_tests common fmt::fmt Zydis::Zydis xbyak::xbyak)
 	target_include_directories(macos_sse4a_tests PRIVATE ${inc_headers})
 	if(BUILD_TESTING)
 		add_test(NAME macos_sse4a COMMAND $<TARGET_FILE:macos_sse4a_tests>)
@@ -725,6 +734,15 @@ endif()
 add_executable(kyty_emulator "${KYTY_SOURCE_DIR}/main.cpp" ${kyty_emulator_src})
 configure_macos_guest_address_space(kyty_emulator)
 
+set(KYTY_TROPHY_UNLOCK_SOUND "${CMAKE_CURRENT_SOURCE_DIR}/assets/sounds/trophy-unlock.wav")
+add_custom_target(kyty_trophy_sound
+	COMMAND ${CMAKE_COMMAND} -E make_directory "$<TARGET_FILE_DIR:kyty_emulator>/assets/sounds"
+	COMMAND ${CMAKE_COMMAND} -E copy_if_different "${KYTY_TROPHY_UNLOCK_SOUND}"
+		"$<TARGET_FILE_DIR:kyty_emulator>/assets/sounds/trophy-unlock.wav"
+	VERBATIM)
+add_dependencies(kyty_emulator kyty_trophy_sound)
+install(FILES "${KYTY_TROPHY_UNLOCK_SOUND}" DESTINATION assets/sounds)
+
 target_link_libraries(kyty_emulator ${kyty_emulator_link_libraries})
 if (WIN32)
 	target_link_libraries(kyty_emulator opengl32)

src/common/CMakeLists.txt

@@ -27,6 +27,7 @@ target_link_libraries(common
 		kyty_git_version
 		spdlog::spdlog
 		ZArchive::zarchive
+		nlohmann_json::nlohmann_json
 )
 
 get_property(inc_headers TARGET common PROPERTY INCLUDE_DIRECTORIES)

src/common/emulatorConfig.cpp

@@ -152,14 +152,26 @@ bool ReadbackLinearImagesEnabled() {
 	return g_config->readback_linear_images;
 }
 
+bool SyncRawImageBuffersEnabled() {
+	return g_config->sync_raw_image_buffers;
+}
+
 bool TessellationEnabled() {
 	return g_config->tessellation_enabled;
 }
 
+bool TrophyEnabled() {
+	return g_config->trophy_enabled;
+}
+
 bool PlayGoHackEnabled() {
 	return g_config->playgo_hack_enabled;
 }
 
+bool SkipNoticeScreen() {
+	return g_config->skip_notice_screen;
+}
+
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 bool RedZoneProtectionEnabled() {
 	return g_config->red_zone_protection_enabled;

src/common/emulatorConfig.h

@@ -73,8 +73,11 @@ struct ConfigOptions {
 	bool                   gpu_assisted_validation_enabled = false;
 	bool                   renderdoc_enabled           = false;
 	bool                   readback_linear_images      = false;
+	bool                   sync_raw_image_buffers      = false;
 	bool                   tessellation_enabled        = false;
+	bool                   trophy_enabled              = true;
 	bool                   playgo_hack_enabled         = false;
+	bool                   skip_notice_screen          = false;
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 	bool red_zone_protection_enabled = false;
 #endif
@@ -122,8 +125,11 @@ bool GpuAssistedValidationEnabled();
 
 bool RenderDocEnabled();
 bool ReadbackLinearImagesEnabled();
+bool SyncRawImageBuffersEnabled();
 bool TessellationEnabled();
+bool TrophyEnabled();
 bool PlayGoHackEnabled();
+bool SkipNoticeScreen();
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 bool RedZoneProtectionEnabled();
 #endif

src/common/trophies.cpp

@@ -0,0 +1,453 @@
+#include "common/trophies.h"
+
+#include "common/file.h"
+
+#include <algorithm>
+#include <array>
+#include <charconv>
+#include <fmt/format.h>
+#include <nlohmann/json.hpp>
+#include <span>
+
+namespace Common::Trophies {
+namespace {
+
+using Json  = nlohmann::json;
+using Files = std::map<std::string_view, std::span<const std::byte>>;
+
+constexpr uint64_t MaxUnlocksSize = uint64_t {1} << 20u;
+constexpr uint64_t MaxPackageSize = uint64_t {1} << 30u;
+
+uint64_t ReadBigEndian(const std::byte* data, size_t size) {
+	uint64_t value = 0;
+	for (size_t i = 0; i < size; ++i) {
+		value = (value << 8u) | std::to_integer<uint8_t>(data[i]);
+	}
+	return value;
+}
+
+int ReadId(const Json& value) {
+	if (value.is_number_integer()) {
+		const auto number = value.get<int64_t>();
+		return number >= 0 && number <= INT32_MAX ? static_cast<int>(number) : -1;
+	}
+	if (!value.is_string()) {
+		return -1;
+	}
+	const auto& text        = value.get_ref<const std::string&>();
+	int         id          = -1;
+	const auto [end, error] = std::from_chars(text.data(), text.data() + text.size(), id);
+	return error == std::errc {} && end == text.data() + text.size() && id >= 0 ? id : -1;
+}
+
+std::optional<uint64_t> ReadUnsigned(const Json& value) {
+	if (value.is_number_unsigned()) {
+		return value.get<uint64_t>();
+	}
+	if (value.is_number_integer()) {
+		const auto number = value.get<int64_t>();
+		return number >= 0 ? std::optional<uint64_t>(static_cast<uint64_t>(number)) : std::nullopt;
+	}
+	if (!value.is_string()) {
+		return std::nullopt;
+	}
+	const auto& text        = value.get_ref<const std::string&>();
+	uint64_t    number      = 0;
+	const auto [end, error] = std::from_chars(text.data(), text.data() + text.size(), number);
+	return error == std::errc {} && end == text.data() + text.size()
+	           ? std::optional<uint64_t>(number)
+	           : std::nullopt;
+}
+
+Json ReadJson(const Files& files, std::string_view name) {
+	const auto entry = files.find(name);
+	if (entry == files.end()) {
+		return {};
+	}
+	return Json::parse(entry->second.begin(), entry->second.end(), nullptr, false);
+}
+
+bool ReadFiles(std::span<const std::byte> data, Files& files, uint64_t& size) {
+	if (data.size() < 0x40 || data.size() > MaxPackageSize ||
+	    ReadBigEndian(data.data(), 4) != 0xb228c60a || ReadBigEndian(data.data() + 4, 4) != 1) {
+		return false;
+	}
+	size             = ReadBigEndian(data.data() + 8, 8);
+	const auto count = ReadBigEndian(data.data() + 0x10, 4);
+	const auto toc   = ReadBigEndian(data.data() + 0x14, 4);
+	if (size < 0x40 || size > data.size() || count > 4096 || toc > size ||
+	    0x20 + count * 0x40 > size - toc) {
+		return false;
+	}
+	for (uint64_t i = 0; i < count; ++i) {
+		const auto*            entry = data.data() + toc + 0x20 + i * 0x40;
+		const auto*            name  = reinterpret_cast<const char*>(entry);
+		const std::string_view filename(name, std::find(name, name + 0x20, '\0') - name);
+		const auto             offset = ReadBigEndian(entry + 0x20, 8);
+		const auto             length = ReadBigEndian(entry + 0x28, 8);
+		if (offset > size || length > size - offset) {
+			return false;
+		}
+		files.emplace(filename, data.subspan(offset, length));
+	}
+	return true;
+}
+
+Package ParsePackage(std::span<const std::byte> data, int console_language) {
+	Files    files;
+	uint64_t size = 0;
+	if (!ReadFiles(data, files, size)) {
+		return {};
+	}
+	const auto conf = ReadJson(files, "tropconf.json");
+	if (!conf.is_object() || !conf.contains("trophies") || !conf["trophies"].is_array() ||
+	    conf["trophies"].size() > 1000 || !conf.contains("defaultLanguage") ||
+	    !conf["defaultLanguage"].is_string()) {
+		return {};
+	}
+	static constexpr std::array<std::string_view, 30> locales = {
+	    "ja-JP", "en-US", "fr-FR",   "es-ES",   "de-DE",  "it-IT", "nl-NL", "pt-PT",
+	    "ru-RU", "ko-KR", "zh-Hant", "zh-Hans", "fi-FI",  "sv-SE", "da-DK", "no-NO",
+	    "pl-PL", "pt-BR", "en-GB",   "tr-TR",   "es-419", "ar-AE", "fr-CA", "cs-CZ",
+	    "hu-HU", "el-GR", "ro-RO",   "th-TH",   "vi-VN",  "id-ID"};
+	const auto locale =
+	    locales[console_language >= 0 && console_language < locales.size() ? console_language : 1];
+	auto meta = ReadJson(files, fmt::format("tropmeta_{}.json", locale));
+	if (meta.is_null()) {
+		meta = ReadJson(files, fmt::format("tropmeta_{}.json",
+		                                   conf["defaultLanguage"].get_ref<const std::string&>()));
+	}
+	if (!meta.is_object() || !meta.contains("metadata") || !meta["metadata"].is_object()) {
+		return {};
+	}
+	const auto& texts = meta["metadata"];
+	if (!texts.contains("titleMetadata") || !texts["titleMetadata"].is_object() ||
+	    !texts["titleMetadata"].contains("name") || !texts["titleMetadata"]["name"].is_string() ||
+	    !texts.contains("trophyMetadata") || !texts["trophyMetadata"].is_array()) {
+		return {};
+	}
+	Package package;
+	package.title = texts["titleMetadata"]["name"].get<std::string>();
+	package.groups.emplace(-1, package.title);
+	uint64_t icon_bytes = 0;
+	for (const auto& definition: conf["trophies"]) {
+		if (!definition.is_object() || !definition.contains("id") ||
+		    !definition.contains("grade") || !definition["grade"].is_string()) {
+			return {};
+		}
+		Trophy trophy;
+		trophy.id              = ReadId(definition["id"]);
+		const auto grade       = definition["grade"].get<std::string>();
+		const auto grade_index = std::string_view("PGSB").find(grade);
+		if (trophy.id < 0 || grade.size() != 1 || grade_index == std::string_view::npos) {
+			return {};
+		}
+		trophy.grade = static_cast<int>(grade_index) + 1;
+		if (definition.contains("groupId")) {
+			trophy.group_id = ReadId(definition["groupId"]);
+			if (trophy.group_id < 0) {
+				return {};
+			}
+		}
+		if (definition.contains("platinumTrophyId")) {
+			trophy.platinum_id = ReadId(definition["platinumTrophyId"]);
+		}
+		trophy.hidden     = definition.contains("hidden") && definition["hidden"] == true;
+		trophy.has_reward = definition.contains("hasReward") && definition["hasReward"] == true;
+		if (definition.contains("unlockCondition")) {
+			const auto& condition = definition["unlockCondition"];
+			if (condition.is_object()) {
+				trophy.progressive =
+				    condition.contains("progressive") && condition["progressive"] == true;
+				if (condition.contains("udsStatId")) {
+					trophy.uds_stat_id = ReadUnsigned(condition["udsStatId"]);
+					if (!trophy.uds_stat_id) {
+						return {};
+					}
+				}
+				if (condition.contains("targetValue")) {
+					trophy.target = ReadUnsigned(condition["targetValue"]);
+				}
+				const auto comparison = condition.find("comparator");
+				if (comparison != condition.end()) {
+					if (*comparison == "ge")
+						trophy.comparison = Comparison::GreaterEqual;
+					else if (*comparison == "gt")
+						trophy.comparison = Comparison::Greater;
+					else if (*comparison == "le")
+						trophy.comparison = Comparison::LessEqual;
+					else if (*comparison == "lt")
+						trophy.comparison = Comparison::Less;
+				}
+			}
+		}
+		const auto icon = files.find(fmt::format("trop{:04}.png", trophy.id));
+		if (icon != files.end()) {
+			if (icon->second.size() > size - icon_bytes) {
+				return {};
+			}
+			trophy.icon_png.assign(icon->second.begin(), icon->second.end());
+			icon_bytes += icon->second.size();
+		}
+		package.groups.try_emplace(trophy.group_id);
+		if (package.groups.size() > 50 ||
+		    !package.trophies.emplace(trophy.id, std::move(trophy)).second) {
+			return {};
+		}
+	}
+	if (texts.contains("groupMetadata") && texts["groupMetadata"].is_array()) {
+		for (const auto& group: texts["groupMetadata"]) {
+			if (group.is_object() && group.contains("id") && group.contains("name") &&
+			    group["name"].is_string()) {
+				const auto found = package.groups.find(ReadId(group["id"]));
+				if (found != package.groups.end()) {
+					found->second = group["name"].get<std::string>();
+				}
+			}
+		}
+	}
+	for (const auto& text: texts["trophyMetadata"]) {
+		if (!text.is_object() || !text.contains("id")) {
+			continue;
+		}
+		const auto found = package.trophies.find(ReadId(text["id"]));
+		if (found == package.trophies.end()) {
+			continue;
+		}
+		for (const auto& [key, field]: {std::pair {"name", &found->second.name},
+		                                {"detail", &found->second.description},
+		                                {"reward", &found->second.reward}}) {
+			if (text.contains(key) && text[key].is_string()) {
+				*field = text[key].get<std::string>();
+			}
+		}
+	}
+	return package;
+}
+
+} // namespace
+
+Package LoadPackage(const std::filesystem::path& path, int console_language) {
+	File file(path, File::Mode::Read);
+	if (file.IsInvalid() || file.Size() > MaxPackageSize) {
+		return {};
+	}
+	return ParsePackage(file.ReadWholeBuffer(), console_language);
+}
+
+std::filesystem::path PackagePath(uint32_t service_label) {
+	return std::filesystem::path(PackageDirectory) / fmt::format("trophy{:02}.ucp", service_label);
+}
+
+std::filesystem::path UnlocksPath(const std::filesystem::path& root, std::string_view title_id,
+                                  int user_id, uint32_t service_label) {
+	if (title_id.empty() || !std::all_of(title_id.begin(), title_id.end(), [](char c) {
+		    return (c >= 'A' && c <= 'Z') || (c >= 'a' && c <= 'z') || (c >= '0' && c <= '9') ||
+		           c == '_' || c == '-';
+	    })) {
+		return {};
+	}
+	return root / "_Trophies" / title_id /
+	       fmt::format("trophies_{}_{}.json", user_id, service_label);
+}
+
+UnlockData LoadUnlockData(const std::filesystem::path& path) {
+	File file(path, File::Mode::Read);
+	if (file.IsInvalid() || file.Size() > MaxUnlocksSize) {
+		return {};
+	}
+	const auto bytes = file.ReadWholeBuffer();
+	const auto json  = Json::parse(bytes.begin(), bytes.end(), nullptr, false);
+	if (!json.is_object() || !json.contains("unlockedTrophies") ||
+	    !json["unlockedTrophies"].is_array()) {
+		return {};
+	}
+	UnlockData unlocks;
+	for (const auto& value: json["unlockedTrophies"]) {
+		const auto id = ReadId(value);
+		if (id >= 0) {
+			unlocks.unlocked.insert(id);
+		}
+	}
+	if (json.contains("unlockedAt") && json["unlockedAt"].is_object()) {
+		for (auto it = json["unlockedAt"].begin(); it != json["unlockedAt"].end(); ++it) {
+			const auto id = ReadId(Json(it.key()));
+			if (id >= 0 && unlocks.unlocked.contains(id)) {
+				const auto tick = ReadUnsigned(it.value());
+				if (tick && *tick >= UnixEpochTick && *tick <= 315537897599999999ULL) {
+					unlocks.timestamps.emplace(id, *tick);
+				}
+			}
+		}
+	}
+	return unlocks;
+}
+
+bool SaveUnlockData(const std::filesystem::path& path, const UnlockData& unlocks) {
+	if (path.empty() || !File::CreateDirectories(path.parent_path())) {
+		return false;
+	}
+	Json timestamps = Json::object();
+	for (const auto& [id, tick]: unlocks.timestamps) {
+		if (unlocks.unlocked.contains(id)) {
+			timestamps[std::to_string(id)] = tick;
+		}
+	}
+	const auto text =
+	    Json {{"unlockedTrophies", unlocks.unlocked}, {"unlockedAt", timestamps}}.dump();
+	auto temporary = path;
+	temporary += ".tmp";
+	File file;
+	if (text.size() > MaxUnlocksSize || !file.Create(temporary)) {
+		return false;
+	}
+	uint32_t written = 0;
+	file.Write(text.data(), static_cast<uint32_t>(text.size()), &written);
+	const bool complete = written == text.size() && file.Flush();
+	file.Close();
+	std::error_code error;
+	if (complete) {
+		std::filesystem::rename(temporary, path, error);
+		if (!error) {
+			return true;
+		}
+	}
+	std::filesystem::remove(temporary, error);
+	return false;
+}
+
+UdsRules LoadUdsRules(const std::filesystem::path& path) {
+	File file(path, File::Mode::Read);
+	if (file.IsInvalid() || file.Size() > MaxPackageSize) {
+		return {};
+	}
+	const auto bytes = file.ReadWholeBuffer();
+	Files      files;
+	uint64_t   size = 0;
+	if (!ReadFiles(bytes, files, size)) {
+		return {};
+	}
+	const auto definitions = ReadJson(files, "stats_definition.json");
+	const auto extraction  = ReadJson(files, "stats_extraction.json");
+	if (!definitions.is_object() || !definitions.contains("statDefinitionArray") ||
+	    !definitions["statDefinitionArray"].is_array() || !extraction.is_object() ||
+	    !extraction.contains("statsExtractionRuleArray") ||
+	    !extraction["statsExtractionRuleArray"].is_array()) {
+		return {};
+	}
+	// Only direct unsigned latest-value stats are evaluated here. Aggregated stats
+	// and filtered/nested extraction require UDS state and event types we do not implement.
+	std::map<uint64_t, bool> stats;
+	for (const auto& definition: definitions["statDefinitionArray"]) {
+		if (!definition.is_object() || !definition.contains("statId") ||
+		    !definition.contains("aggregation") || definition["aggregation"] != "latest" ||
+		    !definition.contains("dataType")) {
+			continue;
+		}
+		const auto  id   = ReadUnsigned(definition["statId"]);
+		const auto& type = definition["dataType"];
+		if (!id || (type != "uint64" && type != "uint32")) {
+			continue;
+		}
+		const uint64_t maximum = type == "uint32" ? UINT32_MAX : UINT64_MAX;
+		if ((definition.contains("minValue") && ReadUnsigned(definition["minValue"]) != 0) ||
+		    (definition.contains("maxValue") && ReadUnsigned(definition["maxValue"]) != maximum)) {
+			continue;
+		}
+		stats.emplace(*id, type == "uint32");
+	}
+	UdsRules events;
+	for (const auto& rule: extraction["statsExtractionRuleArray"]) {
+		if (!rule.is_object() || !rule.contains("condition") || !rule.contains("action")) {
+			continue;
+		}
+		const auto& condition = rule["condition"];
+		const auto& action    = rule["action"];
+		if (!condition.is_object() || condition.size() != 1 || !condition.contains("eventName") ||
+		    !condition["eventName"].is_string() || !action.is_object() ||
+		    !action.contains("input") || !action["input"].is_string() ||
+		    !action.contains("output") || !action["output"].is_object() ||
+		    !action["output"].contains("statId")) {
+			continue;
+		}
+		const auto  id    = ReadUnsigned(action["output"]["statId"]);
+		const auto& input = action["input"].get_ref<const std::string&>();
+		if (!id || !stats.contains(*id) || !input.starts_with("$.") || input.size() == 2 ||
+		    input.find_first_not_of(
+		        "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789_", 2) !=
+		        std::string::npos) {
+			continue;
+		}
+		events[condition["eventName"].get<std::string>()].push_back(
+		    {*id, input.substr(2), stats.at(*id)});
+	}
+	return events;
+}
+
+std::filesystem::path UdsPackagePath(uint32_t service_label) {
+	return std::filesystem::path("sce_sys/uds") / fmt::format("uds{:02}.ucp", service_label);
+}
+
+std::vector<int> FindUdsTrophies(const Package& package, std::string_view event_name,
+                                 const std::map<std::string, UdsInteger>& properties) {
+	const auto rules = package.event_rules.find(std::string(event_name));
+	if (rules == package.event_rules.end()) {
+		return {};
+	}
+	std::vector<int> ids;
+	for (const auto& [id, trophy]: package.trophies) {
+		if (trophy.grade == 1 || !trophy.uds_stat_id || !trophy.target) {
+			continue;
+		}
+		for (const auto& rule: rules->second) {
+			const auto value = properties.find(rule.input);
+			if (rule.stat_id != *trophy.uds_stat_id || value == properties.end()) {
+				continue;
+			}
+			std::optional<uint64_t> number;
+			if (rule.is_uint32) {
+				if (const auto* input = std::get_if<uint32_t>(&value->second)) number = *input;
+			} else if (const auto* input = std::get_if<uint64_t>(&value->second)) {
+				number = *input;
+			}
+			if (!number) continue;
+			bool matched = false;
+			switch (trophy.comparison) {
+				case Comparison::GreaterEqual: matched = *number >= *trophy.target; break;
+				case Comparison::Greater: matched = *number > *trophy.target; break;
+				case Comparison::LessEqual: matched = *number <= *trophy.target; break;
+				case Comparison::Less: matched = *number < *trophy.target; break;
+				case Comparison::None: break;
+			}
+			if (matched) {
+				ids.push_back(id);
+				break;
+			}
+		}
+	}
+	return ids;
+}
+
+uint32_t Progress::Percentage() const {
+	const auto points        = total_grade[2] * 6 + total_grade[3] * 2 + total_grade[4];
+	const auto earned_points = earned_grade[2] * 6 + earned_grade[3] * 2 + earned_grade[4];
+	return points == 0 ? 0 : earned_points * 100 / points;
+}
+
+Progress GetProgress(const Package& package, const UnlockData& unlocks, std::optional<int> group) {
+	Progress progress;
+	for (const auto& [id, trophy]: package.trophies) {
+		if (group && trophy.group_id != *group) {
+			continue;
+		}
+		++progress.total;
+		++progress.total_grade[trophy.grade];
+		if (unlocks.unlocked.contains(id)) {
+			++progress.earned;
+			++progress.earned_grade[trophy.grade];
+		}
+	}
+	return progress;
+}
+
+} // namespace Common::Trophies

src/common/trophies.h

@@ -0,0 +1,87 @@
+#ifndef KYTY_COMMON_TROPHIES_H_
+#define KYTY_COMMON_TROPHIES_H_
+
+#include <array>
+#include <cstddef>
+#include <cstdint>
+#include <filesystem>
+#include <map>
+#include <optional>
+#include <set>
+#include <string>
+#include <string_view>
+#include <variant>
+#include <vector>
+
+namespace Common::Trophies {
+
+inline constexpr char     PackageDirectory[] = "sce_sys/trophy2";
+inline constexpr uint64_t UnixEpochTick      = 62135596800000000ULL;
+
+enum class Comparison { None, GreaterEqual, Greater, LessEqual, Less };
+
+struct Trophy {
+	int                     id          = 0;
+	int                     group_id    = -1;
+	int                     platinum_id = -1;
+	int                     grade       = 0;
+	std::optional<uint64_t> target;
+	std::optional<uint64_t> uds_stat_id;
+	Comparison              comparison  = Comparison::None;
+	bool                    progressive = false;
+	std::string             name;
+	std::string             description;
+	std::string             reward;
+	std::vector<std::byte>  icon_png;
+	bool                    hidden     = false;
+	bool                    has_reward = false;
+};
+
+struct UdsRule {
+	uint64_t    stat_id;
+	std::string input;
+	bool        is_uint32;
+};
+
+using UdsRules   = std::map<std::string, std::vector<UdsRule>>;
+using UdsInteger = std::variant<int32_t, uint32_t, uint64_t>;
+
+struct Package {
+	std::string                title;
+	std::map<int, std::string> groups;
+	std::map<int, Trophy>      trophies;
+	UdsRules                   event_rules;
+};
+
+struct UnlockData {
+	std::set<int>           unlocked;
+	std::map<int, uint64_t> timestamps;
+};
+
+struct Progress {
+	std::array<uint32_t, 5> total_grade {};
+	std::array<uint32_t, 5> earned_grade {};
+	uint32_t                total  = 0;
+	uint32_t                earned = 0;
+
+	[[nodiscard]] uint32_t Percentage() const;
+};
+
+[[nodiscard]] Progress GetProgress(const Package& package, const UnlockData& unlocks,
+                                   std::optional<int> group = std::nullopt);
+
+[[nodiscard]] Package LoadPackage(const std::filesystem::path& path, int console_language);
+[[nodiscard]] std::filesystem::path PackagePath(uint32_t service_label);
+[[nodiscard]] std::filesystem::path UdsPackagePath(uint32_t service_label);
+[[nodiscard]] UdsRules              LoadUdsRules(const std::filesystem::path& path);
+[[nodiscard]] std::vector<int> FindUdsTrophies(const Package& package, std::string_view event_name,
+                                               const std::map<std::string, UdsInteger>& properties);
+[[nodiscard]] std::filesystem::path UnlocksPath(const std::filesystem::path& root,
+                                                std::string_view title_id, int user_id,
+                                                uint32_t service_label);
+[[nodiscard]] UnlockData            LoadUnlockData(const std::filesystem::path& path);
+[[nodiscard]] bool SaveUnlockData(const std::filesystem::path& path, const UnlockData& unlocks);
+
+} // namespace Common::Trophies
+
+#endif // KYTY_COMMON_TROPHIES_H_

src/graphics/guest_gpu/command_processor/commandProcessor.h

@@ -23,6 +23,12 @@ enum class ContextStateOperation : uint32_t {
 	PushClear = 3,
 };
 
+struct IndirectDrawRegisters {
+	uint32_t vertex_offset;
+	uint32_t instance_offset;
+	uint32_t index_offset;
+};
+
 class Pm4Execution {
 public:
 	[[nodiscard]] bool MadeProgress() const noexcept { return m_made_progress; }
@@ -79,10 +85,11 @@ public:
 	void DrawIndex(DrawIndexArgs args);
 	void DrawIndexOffset(uint32_t index_offset, uint32_t index_count);
 	void DrawIndexAuto(DrawAutoArgs args);
-	void DrawIndirect(uint32_t data_offset, uint32_t draw_initiator, bool indexed);
+	void DrawIndirect(uint32_t data_offset, IndirectDrawRegisters registers,
+	                  uint32_t draw_initiator, bool indexed);
 	void DrawIndirectMulti(uint32_t data_offset, uint32_t max_count_or_count,
 	                       const volatile uint32_t* count_addr, uint32_t stride_in_bytes,
-	                       uint32_t draw_initiator, bool indexed);
+	                       IndirectDrawRegisters registers, uint32_t draw_initiator, bool indexed);
 	void WriteAtEndOfPipe32(uint32_t cache_policy, uint32_t event_write_dest,
 	                        uint32_t eop_event_type, uint32_t cache_action, uint32_t event_index,
 	                        uint32_t event_write_source, void* dst_gpu_addr, uint32_t value,

src/graphics/guest_gpu/command_processor/pm4Handlers.cpp

@@ -184,7 +184,6 @@ static HW::RenderControl DecodeRenderControl(uint32_t value) {
 	    KYTY_PM4_GET(value, DB_RENDER_CONTROL, COPY_STENCIL_TO_COLOR) != 0;
 	r.copy_centroid          = KYTY_PM4_GET(value, DB_RENDER_CONTROL, COPY_CENTROID) != 0;
 	r.copy_sample            = KYTY_PM4_GET(value, DB_RENDER_CONTROL, COPY_SAMPLE);
-	EXIT_NOT_IMPLEMENTED(r.copy_depth_to_color || r.copy_stencil_to_color);
 
 	return r;
 }
@@ -1642,8 +1641,12 @@ KYTY_CP_OP_PARSER(CpOpDrawIndirect) {
 	const auto data_offset    = buffer[0];
 	const auto draw_initiator = buffer[3];
 	const bool indexed        = (cmd_id == 0xc0032500);
+	// 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};
 
-	cp.DrawIndirect(data_offset, draw_initiator, indexed);
+	cp.DrawIndirect(data_offset, registers, draw_initiator, indexed);
 
 	return 4;
 }
@@ -1661,13 +1664,17 @@ 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.
+	const IndirectDrawRegisters registers {
+	    buffer[1] & 0xffffu, buffer[2] & 0xffffu,
+	    indexed && (buffer[3] & (1u << 28u)) != 0 ? buffer[1] >> 16u : Pm4::SH_NOP};
 
 	if (count_indirect == 0) {
 		count_addr = nullptr;
 	}
 
 	cp.DrawIndirectMulti(data_offset, max_count_or_count, count_addr, stride_in_bytes,
-	                     draw_initiator, indexed);
+	                     registers, draw_initiator, indexed);
 
 	return 9;
 }

src/graphics/guest_gpu/graphicsRun.cpp

@@ -875,65 +875,33 @@ void CommandProcessor::DrawIndexOffset(uint32_t index_offset, uint32_t index_cou
 	DrawIndex({.index_count = index_count, .index_addr = index_addr});
 }
 
-void CommandProcessor::DrawIndirect(uint32_t data_offset, uint32_t draw_initiator, bool indexed) {
-	EXIT_NOT_IMPLEMENTED((draw_initiator & ~0x20u) != 2u);
-	EXIT_NOT_IMPLEMENTED(m_draw_indirect_args_base_addr == 0);
-
-	const auto* args_addr =
-	    reinterpret_cast<const void*>(m_draw_indirect_args_base_addr + data_offset);
-
-	if (!indexed) {
-		DrawIndirectArgs args {};
-		std::memcpy(&args, args_addr, sizeof(args));
-		m_num_instances = args.instance_count;
-		DrawIndexAuto({.vertex_count   = args.vertex_count_per_instance,
-		               .instance_count = args.instance_count,
-		               .first_vertex   = args.start_vertex_location,
-		               .first_instance = args.start_instance_location,
-		               .offset_source  = DrawOffsetSource::IndirectArgs});
-		return;
-	}
-
-	DrawIndexedIndirectArgs args {};
-	std::memcpy(&args, args_addr, sizeof(args));
-
-	uint64_t index_size = 0;
-	switch (m_index_type_and_size) {
-		case 0: index_size = 2; break;
-		case 1: index_size = 4; break;
-		case 2: index_size = 1; break;
-		default: EXIT("unknown index_type_and_size: %u\n", m_index_type_and_size);
-	}
-
-	auto* index_addr = reinterpret_cast<const void*>(
-	    m_index_base_addr + static_cast<uint64_t>(args.start_index_location) * index_size);
-
-	const uint32_t index_count =
-	    (m_index_buffer_size != 0 ? std::min(args.index_count_per_instance, m_index_buffer_size)
-	                              : args.index_count_per_instance);
-	if (GraphicsRunDebugDumpEnabled() && index_count != args.index_count_per_instance) {
-		static std::atomic<uint32_t> log_count {0};
-		if (log_count.fetch_add(1, std::memory_order_relaxed) < 64) {
-			LOGF("\t DrawIndexIndirect: clamped index_count from %" PRIu32 " to %" PRIu32
-			     " using INDEX_BUFFER_SIZE\n",
-			     args.index_count_per_instance, index_count);
+static void PatchIndirectDrawOffsets(CommandProcessor& cp, IndirectDrawRegisters registers,
+                                     uint32_t vertex_offset, uint32_t instance_offset,
+                                     uint32_t first_index) {
+	const auto write = [&](uint32_t location, uint32_t value) {
+		if (location == Pm4::SH_NOP) {
+			return;
 		}
-	}
+		EXIT_NOT_IMPLEMENTED(location >= Pm4::SH_NUM ||
+		                     g_hw_sh_indirect_func[location] == nullptr);
+		g_hw_sh_indirect_func[location](cp, location, value);
+	};
+	write(registers.vertex_offset, vertex_offset);
+	write(registers.instance_offset, instance_offset);
+	write(registers.index_offset, first_index);
+}
 
-	m_num_instances = args.instance_count;
-	DrawIndex({.index_count    = index_count,
-	           .index_addr     = index_addr,
-	           .instance_count = args.instance_count,
-	           .base_vertex    = static_cast<int32_t>(args.base_vertex_location),
-	           .first_instance = args.start_instance_location,
-	           .offset_source  = DrawOffsetSource::IndirectArgs});
+void CommandProcessor::DrawIndirect(uint32_t data_offset, IndirectDrawRegisters registers,
+                                    uint32_t draw_initiator, bool indexed) {
+	const auto args_size = indexed ? sizeof(DrawIndexedIndirectArgs) : sizeof(DrawIndirectArgs);
+	DrawIndirectMulti(data_offset, 1, nullptr, args_size, registers, draw_initiator, indexed);
 }
 
 void CommandProcessor::DrawIndirectMulti(uint32_t data_offset, uint32_t max_count_or_count,
                                          const volatile uint32_t* count_addr,
-                                         uint32_t stride_in_bytes, uint32_t draw_initiator,
-                                         bool indexed) {
-	EXIT_NOT_IMPLEMENTED((draw_initiator & ~0x20u) != 2u);
+                                         uint32_t stride_in_bytes, IndirectDrawRegisters registers,
+                                         uint32_t draw_initiator, bool indexed) {
+	EXIT_NOT_IMPLEMENTED((draw_initiator & ~0x20u) != (indexed ? 0u : 2u));
 	EXIT_NOT_IMPLEMENTED(m_draw_indirect_args_base_addr == 0);
 
 	uint32_t draw_count = max_count_or_count;
@@ -967,6 +935,8 @@ void CommandProcessor::DrawIndirectMulti(uint32_t data_offset, uint32_t max_coun
 
 		if (!indexed) {
 			auto* args = reinterpret_cast<const DrawIndirectArgs*>(args_addr);
+			PatchIndirectDrawOffsets(*this, registers, args->start_vertex_location,
+			                         args->start_instance_location, 0);
 			m_num_instances = args->instance_count;
 			DrawIndexAuto({.vertex_count   = args->vertex_count_per_instance,
 			               .instance_count = args->instance_count,
@@ -977,6 +947,8 @@ void CommandProcessor::DrawIndirectMulti(uint32_t data_offset, uint32_t max_coun
 		}
 
 		auto* args = reinterpret_cast<const DrawIndexedIndirectArgs*>(args_addr);
+		PatchIndirectDrawOffsets(*this, registers, args->base_vertex_location,
+		                         args->start_instance_location, args->start_index_location);
 
 		auto* index_addr = reinterpret_cast<const void*>(
 		    m_index_base_addr + static_cast<uint64_t>(args->start_index_location) * index_size);

src/graphics/host_gpu/memoryTracker.h

@@ -82,9 +82,8 @@ public:
 		static_assert(std::is_nothrow_invocable_v<RangeFunc&, uint64_t, uint64_t>);
 		static_assert(std::is_nothrow_invocable_v<UploadFunc&>);
 		CheckNotInUploadCallback();
-		Iterate<true>(vaddr, size, [](RegionManager*, uint64_t, uint64_t) {});
 		const auto* previous_upload_owner = std::exchange(s_upload_owner, this);
-		Iterate<false>(vaddr, size, [&](RegionManager* manager, uint64_t offset, uint64_t bytes) {
+		Iterate<true>(vaddr, size, [&](RegionManager* manager, uint64_t offset, uint64_t bytes) {
 			manager->lock.lock();
 			manager->ForEachModifiedRange<DirtySource::Cpu, true>(manager->GetCpuAddr() + offset,
 			                                                      bytes, range_func);

src/graphics/host_gpu/pageManager.cpp

@@ -15,6 +15,12 @@
 #include <mutex>
 #include <vector>
 
+#if defined(_MSC_VER)
+#include <intrin.h>
+#elif defined(__x86_64__)
+#include <xmmintrin.h>
+#endif
+
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 #ifndef NOMINMAX
 #define NOMINMAX
@@ -62,7 +68,13 @@ class SpinGuard final {
 public:
 	explicit SpinGuard(std::atomic_flag& lock): m_lock(lock) {
 		while (m_lock.test_and_set(std::memory_order_acquire)) {
-			std::atomic_signal_fence(std::memory_order_seq_cst);
+#if defined(__x86_64__) || defined(_M_X64)
+			_mm_pause();
+#elif defined(_M_ARM64)
+			__yield();
+#elif defined(__aarch64__)
+			asm volatile("yield");
+#endif
 		}
 	}
 	~SpinGuard() { m_lock.clear(std::memory_order_release); }

src/graphics/host_gpu/regionManager.h

@@ -9,6 +9,12 @@
 #include <mutex>
 #include <utility>
 
+#if defined(_MSC_VER)
+#include <intrin.h>
+#elif defined(__x86_64__)
+#include <xmmintrin.h>
+#endif
+
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 #ifndef NOMINMAX
 #define NOMINMAX
@@ -32,7 +38,13 @@ public:
 		const auto thread = CurrentThread();
 		while (m_lock.test_and_set(std::memory_order_acquire)) {
 			EXIT_NOT_IMPLEMENTED(m_owner.load(std::memory_order_relaxed) == thread);
-			std::atomic_signal_fence(std::memory_order_seq_cst);
+#if defined(__x86_64__) || defined(_M_X64)
+			_mm_pause();
+#elif defined(_M_ARM64)
+			__yield();
+#elif defined(__aarch64__)
+			asm volatile("yield");
+#endif
 		}
 		m_owner.store(thread, std::memory_order_relaxed);
 	}
@@ -48,7 +60,9 @@ private:
 		return GetCurrentThreadId();
 #elif defined(__APPLE__)
 		// mach thread port is a nonzero per-thread id (0 is the "no owner" sentinel).
-		return static_cast<uint32_t>(pthread_mach_thread_np(pthread_self()));
+		static thread_local const uint32_t tid =
+		    static_cast<uint32_t>(pthread_mach_thread_np(pthread_self()));
+		return tid;
 #elif defined(__linux__)
 		static thread_local const uint32_t tid = static_cast<uint32_t>(::syscall(SYS_gettid));
 		return tid;

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

@@ -34,9 +34,36 @@ void BufferCache::WriteDataBuffer(Buffer& buffer, uint64_t address, const void*
 	auto* bytes = static_cast<const uint8_t*>(source);
 	while (size != 0) {
 		const auto chunk  = std::min(size, m_staging_buffer.Size());
+		// Copy waits for ring reuse and flushes the fresh host data before submission.
 		const auto offset = m_staging_buffer.Copy(bytes, chunk, 4);
-		buffer.CopyFrom(m_scheduler.Current(), m_staging_buffer, offset, buffer.Offset(address),
-		                chunk, vk::AccessFlagBits::eHostWrite);
+		const auto destination_offset = buffer.Offset(address);
+		EXIT_IF(destination_offset > buffer.Size() || chunk > buffer.Size() - destination_offset);
+		m_scheduler.EndRendering();
+		const auto command = m_scheduler.Current().Handle();
+		vk::BufferMemoryBarrier2 before {};
+		before.srcStageMask  = vk::PipelineStageFlagBits2::eAllCommands;
+		before.srcAccessMask = vk::AccessFlagBits2::eMemoryRead | vk::AccessFlagBits2::eMemoryWrite;
+		before.dstStageMask  = vk::PipelineStageFlagBits2::eTransfer;
+		before.dstAccessMask = vk::AccessFlagBits2::eTransferWrite;
+		before.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
+		before.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
+		before.buffer        = buffer.Handle();
+		before.offset        = destination_offset;
+		before.size          = chunk;
+		vk::DependencyInfo dependency {};
+		dependency.dependencyFlags          = vk::DependencyFlagBits::eByRegion;
+		dependency.bufferMemoryBarrierCount = 1;
+		dependency.pBufferMemoryBarriers    = &before;
+		command.pipelineBarrier2(dependency);
+		const vk::BufferCopy copy {offset, destination_offset, chunk};
+		command.copyBuffer(m_staging_buffer.Handle(), buffer.Handle(), 1, &copy);
+		auto after          = before;
+		after.srcStageMask  = vk::PipelineStageFlagBits2::eTransfer;
+		after.srcAccessMask = vk::AccessFlagBits2::eTransferWrite;
+		after.dstStageMask  = vk::PipelineStageFlagBits2::eAllCommands;
+		after.dstAccessMask = vk::AccessFlagBits2::eMemoryRead | vk::AccessFlagBits2::eMemoryWrite;
+		dependency.pBufferMemoryBarriers = &after;
+		command.pipelineBarrier2(dependency);
 		bytes += chunk;
 		address += chunk;
 		size -= chunk;

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

@@ -694,7 +694,9 @@ ImageId TextureCache::ResolveDepthOverlap(const ImageInfo& requested, BindingTyp
 	    requested.type == cached.info.type && requested.pitch == cached.info.pitch &&
 	    !requested.HasStencil() && !cached.info.HasStencil() && !requested.HasMetadata() &&
 	    !cached.info.HasMetadata();
-	// PPSA04264
+	// PPSA04264, PPSA04288
+	// A partial view retains the entire matching array layout. HTile belongs to
+	// the depth source, independently of the data slices copied into color storage.
 	const bool retain_cached_layout =
 	    requested.samples == 1 && cached.info.samples == 1 && cached.backing.samples == 1 &&
 	    requested.bytes_per_block == cached.info.bytes_per_block &&
@@ -713,7 +715,8 @@ ImageId TextureCache::ResolveDepthOverlap(const ImageInfo& requested, BindingTyp
 	    requested.data.size / requested.resources.layers ==
 	        cached.info.data.size / cached.info.resources.layers &&
 	    !requested.HasStencil() && !cached.info.HasStencil() && !requested.HasMetadata() &&
-	    !cached.info.HasMetadata();
+	    (cached.info.metadata.kind == ImageMetadataKind::None ||
+	     cached.info.metadata.kind == ImageMetadataKind::Htile);
 	bool recreate = cached.info.resources < requested.resources ||
 	                requested.IsVolume() != cached.info.IsVolume();
 	switch (binding) {
@@ -738,8 +741,6 @@ ImageId TextureCache::ResolveDepthOverlap(const ImageInfo& requested, BindingTyp
 		info.data       = cached.info.data;
 		info.resources  = cached.info.resources;
 		info.mip_layout = cached.info.mip_layout;
-	} else {
-		info.resources = std::max(requested.resources, cached.info.resources);
 	}
 	info.htile_clear_mask     = 0;
 	const auto replacement_id = InsertImage(info);
@@ -1791,14 +1792,12 @@ bool BufferCache::SynchronizeBufferFromImage(Buffer& buffer, uint64_t vaddr, uin
 	const auto available  = buffer.Size() - buf_offset;
 	uint32_t   levels     = 0;
 	uint64_t   copy_size  = 0;
-	if (image.info.IsVolume()) {
-		// Volume mips contain strided block slices, so a mip's linear span cannot prove that
-		// every retained slice fits. Keep volume synchronization whole-image only.
-		if (!buffer.IsInBounds(image.info.data.address, image.info.data.size)) {
-			return false;
-		}
+	if (buffer.IsInBounds(image.info.data.address, image.info.data.size)) {
 		levels    = image.info.resources.levels;
 		copy_size = image.info.data.size;
+	} else if (image.info.IsVolume() || image.info.resources.layers > 1) {
+		// Array and volume slices are strided across the full mip chain.
+		return false;
 	} else {
 		for (; levels < image.info.resources.levels; ++levels) {
 			const auto& mip = image.info.mip_layout[levels];
@@ -1962,6 +1961,11 @@ bool TextureCache::TouchMeta(uint64_t address, uint32_t slice, bool is_clear) {
 	return true;
 }
 
+bool TextureCache::IsRegionRegistered(uint64_t address, uint64_t size) {
+	std::scoped_lock lock {m_lock};
+	return !FindImagesInRegion(address, size, false).empty();
+}
+
 void TextureCache::UnmapMemory(uint64_t address, uint64_t size) {
 	if (!GuestRange {address, size}.Valid()) {
 		EXIT("TextureCache: invalid unmap range\n");

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

@@ -61,6 +61,7 @@ public:
 	                                        uint32_t packed_clear);
 	void               InvalidateMemory(uint64_t address, uint64_t size);
 	void               InvalidateMemoryFromGPU(uint64_t address, uint64_t size);
+	[[nodiscard]] bool IsRegionRegistered(uint64_t address, uint64_t size);
 
 	[[nodiscard]] bool IsMeta(uint64_t address);
 	[[nodiscard]] bool IsMetaCleared(uint64_t address, uint32_t slice);

src/graphics/host_gpu/renderer/commandScheduler.cpp

@@ -226,43 +226,37 @@ void CommandScheduler::PopPendingOperations() {
 }
 
 void CommandScheduler::DeferOperation(Common::UniqueFunction<void>&& operation) {
+	QueueOperation(std::move(operation), false);
+}
+
+void CommandScheduler::DeferPriorityOperation(Common::UniqueFunction<void>&& operation) {
+	QueueOperation(std::move(operation), true);
+}
+
+void CommandScheduler::QueueOperation(Common::UniqueFunction<void>&& operation, bool priority) {
 	CheckActive();
 	EXIT_IF(!operation);
 	std::unique_lock lock(m_operation_mutex);
 	if (m_operation_state == OperationState::Open) {
-		m_pending_operations.push({std::move(operation), CurrentTick()});
-		return;
-	}
-	if (g_deferred_callback_scheduler == this) {
+		auto& queue = priority ? m_priority_operations : m_pending_operations;
+		queue.push({std::move(operation), CurrentTick()});
 		lock.unlock();
-		operation();
+		if (priority) {
+			m_operation_available.notify_one();
+		}
 		return;
 	}
-	m_operation_available.wait(lock,
-	                           [this] { return m_operation_state == OperationState::Closed; });
+	if (g_deferred_callback_scheduler != this) {
+		m_operation_available.wait(lock,
+		                           [this] { return m_operation_state == OperationState::Closed; });
+	}
 	lock.unlock();
 	operation();
 }
 
-void CommandScheduler::DeferPriorityOperation(Common::UniqueFunction<void>&& operation) {
-	CheckActive();
-	EXIT_IF(!operation);
-	std::unique_lock lock(m_operation_mutex);
-	if (m_operation_state == OperationState::Open) {
-		m_priority_operations.push({std::move(operation), CurrentTick()});
-		lock.unlock();
-		m_operation_available.notify_one();
-		return;
-	}
-	if (g_deferred_callback_scheduler == this) {
-		lock.unlock();
-		operation();
-		return;
-	}
-	m_operation_available.wait(lock,
-	                           [this] { return m_operation_state == OperationState::Closed; });
-	lock.unlock();
-	operation();
+bool CommandScheduler::HasPendingPriorityOperations() {
+	std::lock_guard lock(m_operation_mutex);
+	return !m_priority_operations.empty() || m_priority_active;
 }
 
 void CommandScheduler::PriorityOperationsThread(std::stop_token stop) {

src/graphics/host_gpu/renderer/commandScheduler.h

@@ -38,8 +38,10 @@ public:
 	void                      PopPendingOperations();
 	void                      DrainPriorityOperations();
 	void                      WaitPriorityOperations(uint64_t tick);
+	// Guest-memory completions use the priority queue; normal callbacks maintain GPU resources.
 	void                      DeferOperation(Common::UniqueFunction<void>&& operation);
 	void                      DeferPriorityOperation(Common::UniqueFunction<void>&& operation);
+	[[nodiscard]] bool        HasPendingPriorityOperations();
 	[[nodiscard]] static bool InDeferredOperation() noexcept;
 
 	[[nodiscard]] bool Active() const noexcept { return m_command.m_registers != nullptr; }
@@ -82,6 +84,7 @@ private:
 
 	void BeginNext();
 	void PriorityOperationsThread(std::stop_token stop);
+	void QueueOperation(Common::UniqueFunction<void>&& operation, bool priority);
 	void RunOperation(Common::UniqueFunction<void>&& operation);
 
 	MasterSemaphore              m_master;

src/graphics/host_gpu/renderer/depthRenderTarget.cpp

@@ -315,8 +315,6 @@ void RenderExecutor::ResolveRenderDepthTarget(CommandBuffer& buffer, RenderDepth
 	r.depth_compare_op        = static_cast<vk::CompareOp>(dc.zfunc);
 
 	r.depth_bounds_test_enable = dc.depth_bounds_enable;
-	r.depth_min_bounds         = hw.GetDepthBoundsMin();
-	r.depth_max_bounds         = hw.GetDepthBoundsMax();
 
 	r.stencil_clear_enable =
 	    has_stencil && rc.stencil_clear_enable && !z.depth_view.stencil_write_disable;

src/graphics/host_gpu/renderer/depthRenderTarget.h

@@ -29,8 +29,6 @@ struct RenderDepthInfo {
 	bool                        depth_write_enable       = false;
 	vk::CompareOp               depth_compare_op         = vk::CompareOp::eNever;
 	bool                        depth_bounds_test_enable = false;
-	float                       depth_min_bounds         = 0.0f;
-	float                       depth_max_bounds         = 0.0f;
 	bool                        stencil_clear_enable     = false;
 	uint8_t                     stencil_clear_value      = 0;
 	bool                        stencil_test_enable      = false;

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

@@ -390,7 +390,28 @@ inline constexpr std::array<VideoOutFormatPolicy, 7> 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 float16 = [](uint32_t value) {
+		const auto sign = (value & 0x8000u) << 16u;
+		const auto exponent = (value >> 10u) & 0x1fu;
+		const auto mantissa = value & 0x3ffu;
+		if (exponent == 0) {
+			return std::copysign(static_cast<float>(mantissa) * 0x1p-24f,
+			                     std::bit_cast<float>(sign));
+		}
+		return std::bit_cast<float>(sign | ((exponent == 31 ? 255u : exponent + 112u) << 23u) |
+		                            (mantissa << 13u));
+	};
 	switch (format) {
+		case vk::Format::eR16G16Sfloat:
+			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;
+			[[fallthrough]];
+		case vk::Format::eR16Unorm:
+			next.float32[0] = static_cast<float>(packed & 0xffffu) / 65535.0f;
+			break;
 		// A single-plane float target carries its clear as raw float bits, the same encoding the
 		// depth decoder below uses. Without this the clear is discarded and the target keeps stale
 		// contents.
@@ -443,6 +464,10 @@ inline constexpr std::array<VideoOutFormatPolicy, 7> 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) &&
+	    (packed & 0xffffu) != (packed >> 16u)) {
+		return false;
+	}
 	if (format == vk::Format::eR16G16B16A16Sfloat) {
 		if (packed != 0) {
 			return false;

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

@@ -3,6 +3,7 @@
 #include "common/alignment.h"
 #include "common/assert.h"
 #include "common/common.h"
+#include "common/emulatorConfig.h"
 #include "common/file.h"
 #include "common/logging/log.h"
 #include "common/profiler.h"
@@ -126,8 +127,9 @@ NativeStorageBuffer(RenderContext& context, const PreparedBindings::BufferSource
 	if (size > graphics.GetPhysicalDeviceProperties().limits.maxStorageBufferRange) {
 		EXIT("storage buffer range is unsupported\n");
 	}
-	auto [buffer, offset] = context.GetBufferCache().ObtainBuffer(address, size, resource.written,
-	                                                              resource.formatted, id);
+	auto [buffer, offset] = context.GetBufferCache().ObtainBuffer(
+	    address, size, resource.written,
+	    resource.formatted || (resource.read && Config::SyncRawImageBuffersEnabled()), id);
 	const auto aligned_offset = Common::AlignDown(offset, alignment);
 	const auto adjustment     = offset - aligned_offset;
 	const auto max_range      = graphics.GetPhysicalDeviceProperties().limits.maxStorageBufferRange;
@@ -807,30 +809,52 @@ void RenderExecutor::PrepareBindings(const ShaderStageRuntime& runtime,
 	}
 }
 
-void RenderExecutor::FindBuffers(PreparedBindings& prepared) {
+void RenderExecutor::FindBuffers(std::span<PreparedBindings* const> stages) {
 	KYTY_PROFILER_FUNCTION();
-	EXIT_IF(prepared.runtime == nullptr || !*prepared.runtime);
-	const auto& program  = *prepared.runtime->program;
-	const auto& snapshot = *prepared.runtime->resources;
-	auto&       cache    = m_context.GetBufferCache();
-
-	prepared.buffer_sources.clear();
-	const auto& layout = program.bindings;
-	if (layout.memory_offset_count == 0) {
-		return;
-	}
-	const auto& resources = layout.descriptors.front().resources;
-	prepared.buffer_sources.reserve(resources.size());
-	for (const auto resource: resources) {
-		auto descriptor = DecodeNativeDescriptor<ShaderBufferResource>(snapshot.buffers[resource]);
-		const auto address = descriptor.Base48();
-		const auto requested_size = descriptor.GetSize();
-		if (address == 0 || requested_size == 0) {
-			prepared.buffer_sources.push_back({});
+	auto& cache = m_context.GetBufferCache();
+	for (auto* stage: stages) {
+		auto& prepared = *stage;
+		EXIT_IF(prepared.runtime == nullptr || !*prepared.runtime);
+		const auto& program  = *prepared.runtime->program;
+		const auto& snapshot = *prepared.runtime->resources;
+		prepared.buffer_sources.clear();
+		const auto& layout = program.bindings;
+		if (layout.memory_offset_count == 0) {
 			continue;
 		}
-		const auto size = Libs::LibKernel::Memory::ClampRangeSize(address, requested_size);
-		prepared.buffer_sources.push_back({address, size, cache.FindBuffer(address, size)});
+		const auto& resources = layout.descriptors.front().resources;
+		prepared.buffer_sources.reserve(resources.size());
+		for (const auto resource: resources) {
+			const auto descriptor = DecodeNativeDescriptor<ShaderBufferResource>(snapshot.buffers[resource]);
+			const auto address = descriptor.Base48();
+			auto size = descriptor.GetSize();
+			if (address == 0 || size == 0) {
+				prepared.buffer_sources.push_back({});
+				continue;
+			}
+			if (descriptor.NumRecords() == UINT32_MAX) {
+				if (program.info.buffers[resource].written) {
+					// Sentinel write ranges end before tables captured by any bound stage.
+					for (const auto* reader: stages) {
+						for (const auto [read_address, read_size]: reader->runtime->resources->specialization_reads) {
+							if (read_size == 0) continue;
+							if (read_address > address) {
+								size = std::min(size, read_address - address);
+							} else if (address - read_address < read_size) {
+								EXIT("scalar resource reads overlap a shader buffer write\n");
+							}
+						}
+					}
+				} else {
+					const auto& graphics = m_context.GetGraphics();
+					const auto limit = uint64_t {graphics.GetPhysicalDeviceProperties().limits.maxStorageBufferRange} -
+					                   (graphics.StorageMinAlignment() - 1);
+					size = std::min({size, uint64_t {256} * 1024 * 1024, limit});
+				}
+			}
+			size = Libs::LibKernel::Memory::ClampRangeSize(address, size);
+			prepared.buffer_sources.push_back({address, size, cache.FindBuffer(address, size)});
+		}
 	}
 }
 
@@ -919,8 +943,8 @@ void RenderExecutor::RebindImages(PreparedBindings& prepared) {
 void RenderExecutor::PrepareGraphicsBindings(std::span<PreparedBindings* const> stages,
                                              std::span<RenderColorInfo> colors) {
 	bool uses_dma = false;
+	FindBuffers(stages);
 	for (auto* stage: stages) {
-		FindBuffers(*stage);
 		uses_dma |= stage->runtime->program->info.uses_dma;
 	}
 	if (uses_dma) {

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

@@ -831,9 +831,6 @@ PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(
 	if (static_params.sample_shading_enable && !m_graphics.sample_rate_shading_enabled) {
 		EXIT("Pipeline: sample-rate shading is required but unsupported by the host\n");
 	}
-	static_params.depth_bounds_test_enable = depth.depth_bounds_test_enable;
-	static_params.depth_min_bounds         = depth.depth_min_bounds;
-	static_params.depth_max_bounds         = depth.depth_max_bounds;
 	const bool rect_list = Prospero::IsRectList(command.GetUserConfig().GetPrimType());
 	static_params.cull_back  = !rect_list && mc.cull_back;
 	static_params.cull_front = !rect_list && mc.cull_front;

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

@@ -38,9 +38,6 @@ struct PipelineStaticParameters {
 	bool                       primitive_restart_enable = false;
 	uint32_t                   samples                  = 1;
 	bool                       sample_shading_enable    = false;
-	bool                       depth_bounds_test_enable = false;
-	float                      depth_min_bounds         = 0.0f;
-	float                      depth_max_bounds         = 0.0f;
 	uint32_t                   color_mask[RENDER_COLOR_ATTACHMENTS_MAX]           = {};
 	bool                       cull_front                                         = false;
 	bool                       cull_back                                          = false;
@@ -64,7 +61,7 @@ struct PipelineStaticParameters {
 static_assert(std::is_trivially_copyable_v<PipelineStaticParameters>);
 static_assert(std::is_standard_layout_v<PipelineStaticParameters>);
 static_assert(alignof(PipelineStaticParameters) == 1);
-static_assert(sizeof(PipelineStaticParameters) == 125);
+static_assert(sizeof(PipelineStaticParameters) == 116);
 
 struct PipelineRenderingState {
 	std::array<vk::Format, RENDER_COLOR_ATTACHMENTS_MAX> color_formats {};

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

@@ -471,14 +471,6 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	EXIT_NOT_IMPLEMENTED(pipeline.pipeline_layout == nullptr);
 
 	vk::PipelineDepthStencilStateCreateInfo depth_stencil_info {};
-	depth_stencil_info.depthBoundsTestEnable =
-#if defined(__APPLE__)
-	    VK_FALSE; // MoltenVK lacks the depthBounds feature; depth-bounds testing is disabled
-#else
-	    (static_params.depth_bounds_test_enable ? VK_TRUE : VK_FALSE);
-#endif
-	depth_stencil_info.minDepthBounds    = static_params.depth_min_bounds;
-	depth_stencil_info.maxDepthBounds    = static_params.depth_max_bounds;
 
 	std::vector<vk::DynamicState> dynamic_states {
 	    vk::DynamicState::eViewportWithCount,
@@ -497,6 +489,8 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	    vk::DynamicState::eBlendConstants,
 	};
 #if !defined(__APPLE__)
+	dynamic_states.push_back(vk::DynamicState::eDepthBoundsTestEnable);
+	dynamic_states.push_back(vk::DynamicState::eDepthBounds);
 	if (rendering.color_count != 0) {
 		dynamic_states.push_back(vk::DynamicState::eColorWriteEnableEXT);
 	}

src/graphics/host_gpu/renderer/render.h

@@ -161,7 +161,7 @@ public:
 	                      uint32_t mode);
 
 	void PrepareBindings(const ShaderStageRuntime& runtime, PreparedBindings& prepared);
-	void                           FindBuffers(PreparedBindings& bindings);
+	void                           FindBuffers(std::span<PreparedBindings* const> stages);
 	void                           RebindBuffers(PreparedBindings& bindings);
 	void                           RebindImages(PreparedBindings& bindings);
 	void CommitBindings(CommandBuffer& buffer, vk::PipelineBindPoint pipeline_bind_point,

src/graphics/host_gpu/renderer/renderCompute.cpp

@@ -342,7 +342,8 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 		std::copy(std::begin(input_info.dispatch_threads_num), std::end(input_info.dispatch_threads_num),
 		          bindings.shader_data.begin() + program.bindings.dispatch_thread_dword);
 	}
-	FindBuffers(bindings);
+	PreparedBindings* descriptor_stage = &bindings;
+	FindBuffers(std::span {&descriptor_stage, 1u});
 	if (program.info.uses_dma) {
 		m_context.PrepareBda();
 	}
@@ -351,7 +352,6 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 	RebindBuffers(bindings);
 
 	auto              vk_buffer        = buffer.Handle();
-	PreparedBindings* descriptor_stage = &bindings;
 	CommitBindings(buffer, vk::PipelineBindPoint::eCompute, pipeline,
 	               std::span {&descriptor_stage, 1u});
 	bool has_storage_writes = bindings.shared_memory.buffer != nullptr ||
@@ -457,7 +457,8 @@ void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
 	auto& pipeline = m_context.GetPipelineCache().GetComputePipeline(input_info, compute_program);
 	auto& bindings = m_compute_bindings;
 	PrepareBindings(input_info.stage, bindings);
-	FindBuffers(bindings);
+	PreparedBindings* descriptor_stage = &bindings;
+	FindBuffers(std::span {&descriptor_stage, 1u});
 	const auto& program = *input_info.stage.program;
 	if (program.info.uses_dma) {
 		m_context.PrepareBda();
@@ -469,7 +470,6 @@ void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
 	    args_addr, sizeof(vk::DispatchIndirectCommand), false);
 	EXIT_IF(args_buffer == nullptr || (args_offset & 3u) != 0);
 	RebindBuffers(bindings);
-	PreparedBindings* descriptor_stage = &bindings;
 	CommitBindings(buffer, vk::PipelineBindPoint::eCompute, pipeline,
 	               std::span {&descriptor_stage, 1u});
 	const auto vk_buffer = buffer.Handle();

src/graphics/host_gpu/renderer/renderContext.cpp

@@ -99,7 +99,12 @@ void RenderContext::UnmapMemory(uint64_t vaddr, uint64_t size) {
 		     vaddr, size);
 	}
 	const auto unmap = [this, vaddr, size] {
-		if (m_command_scheduler.Active()) {
+		// Check cache ownership on the GPU thread. Guest-memory callbacks can still
+		// access a range with no cached data, so they must finish before it is unmapped.
+		if (m_command_scheduler.Active() &&
+		    (m_buffer_cache.IsRegionRegistered(vaddr, size) ||
+		     m_texture_cache.IsRegionRegistered(vaddr, size) ||
+		     m_command_scheduler.HasPendingPriorityOperations())) {
 			const auto tick = m_command_scheduler.CurrentTick();
 			m_command_scheduler.Finish();
 			m_command_scheduler.WaitPriorityOperations(tick);

src/graphics/host_gpu/renderer/renderDraw.cpp

@@ -379,6 +379,12 @@ static void SetGraphicsDynamicParams(const CommandBuffer& buffer, vk::CommandBuf
 	vk_buffer.setDepthTestEnable(depth.depth_test_enable ? VK_TRUE : VK_FALSE);
 	vk_buffer.setDepthWriteEnable(depth.depth_write_enable ? VK_TRUE : VK_FALSE);
 	vk_buffer.setDepthCompareOp(depth.depth_compare_op);
+#if !defined(__APPLE__)
+	vk_buffer.setDepthBoundsTestEnable(depth.depth_bounds_test_enable ? VK_TRUE : VK_FALSE);
+	if (depth.depth_bounds_test_enable) {
+		vk_buffer.setDepthBounds(ctx.GetDepthBoundsMin(), ctx.GetDepthBoundsMax());
+	}
+#endif
 
 	const auto& mode              = ctx.GetModeControl();
 	const auto& poly_offset       = ctx.GetPolyOffset();
@@ -583,8 +589,12 @@ 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 = 0;
-	for (uint32_t slot = 0; slot < RENDER_COLOR_ATTACHMENTS_MAX; slot++) {
-		if (sh_regs.target_output_mode[slot] != 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;
 		}
@@ -896,7 +906,7 @@ bool RenderExecutor::PrepareDrawRenderState(CommandBuffer& buffer, const DrawCal
 	if (state.ps_active) {
 		for (const auto& output: state.ps_input_info.stage.program->info.outputs) {
 			if (output.kind == ShaderRecompiler::IR::StageOutputKind::Mrt) {
-				mrt_mask |= 1u << output.index;
+				mrt_mask |= 1u << output.location;
 			}
 		}
 	}
@@ -1274,8 +1284,10 @@ void RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
 	             : ResolveDrawOffsets(ucfg.GetIndexOffset(), state.vertex_info[0]);
 
 	DrawEmitInfo emit {};
-	emit.vertex_offset  = vertex_offset + args.base_vertex;
-	emit.first_instance = instance_offset;
+	// Native fetches consume the patched SGPRs; rewritten fetches need Vulkan offsets.
+	const bool native_indirect = indirect && !state.vertex_info[0].fetch_embedded;
+	emit.vertex_offset  = native_indirect ? 0 : vertex_offset + args.base_vertex;
+	emit.first_instance = native_indirect ? 0 : instance_offset;
 
 	ExecutePreparedDraw(submit_id, buffer, draw, state, topology, emit, index_source,
 	                    primitive_restart);
@@ -1363,8 +1375,10 @@ void RenderExecutor::DrawAuto(uint64_t submit_id, CommandBuffer& buffer, const D
 	    indirect ? std::pair<int32_t, uint32_t> {0, args.first_instance}
 	             : ResolveDrawOffsets(ucfg.GetIndexOffset(), state.vertex_info[0]);
 	DrawEmitInfo emit {};
-	emit.first_vertex = static_cast<uint32_t>(vertex_offset + static_cast<int32_t>(args.first_vertex));
-	emit.first_instance = instance_offset;
+	const bool native_indirect = indirect && !state.vertex_info[0].fetch_embedded;
+	emit.first_vertex = native_indirect ? 0 :
+	    static_cast<uint32_t>(vertex_offset + static_cast<int32_t>(args.first_vertex));
+	emit.first_instance = native_indirect ? 0 : instance_offset;
 
 	DrawIndexBufferSource index_source {};
 	ExecutePreparedDraw(submit_id, buffer, draw, state, topology, emit, index_source, false);

src/graphics/presentation/systemOverlay.cpp

@@ -3,14 +3,17 @@
 #include <SDL3/SDL.h>
 
 #include "common/assert.h"
+#include "common/logging/log.h"
 #include "common/stringUtils.h"
 #include "graphics/host_gpu/graphicContext.h"
 #include "imgui.h"
+#include "imgui_internal.h"
 #include "imgui_impl_vulkan.h"
 #include "libs/controller.h"
 #include "libs/dialog.h"
 #include "libs/ime.h"
 #include "libs/imeDialog.h"
+#include "stb_image.h"
 
 #include <algorithm>
 #include <array>
@@ -18,12 +21,17 @@
 #include <chrono>
 #include <cmath>
 #include <cstdint>
+#include <cstring>
 #include <deque>
+#include <limits>
+#include <memory>
 #include <mutex>
+#include <optional>
 #include <span>
 #include <string>
 #include <string_view>
 #include <utility>
+#include <vector>
 
 namespace Libs::Graphics {
 
@@ -178,6 +186,45 @@ bool                         g_input_lifecycle_active     = false;
 bool                         g_controller_captured        = false;
 OverlaySession               g_session;
 SDL_Window*                  g_input_window               = nullptr;
+struct TrophyNotification {
+	std::string            name;
+	int32_t                grade;
+	std::vector<std::byte> icon_png;
+};
+std::mutex                     g_trophy_notification_mutex;
+std::deque<TrophyNotification> g_trophy_notifications;
+std::atomic<bool>              g_trophy_notification_active {false};
+
+std::vector<std::string_view> WrapTrophyTitle(std::string_view text, float font_size, float width) {
+	std::vector<std::string_view> lines;
+	auto* const                  font = ImGui::GetFont();
+	for (;;) {
+		const auto newline   = text.find('\n');
+		const auto paragraph = text.substr(0, newline);
+		auto       line      = paragraph.substr(0, 0);
+		for (auto start = paragraph.find_first_not_of(" \t\r"); start != std::string_view::npos;) {
+			const auto end  = std::min(paragraph.find_first_of(" \t\r", start), paragraph.size());
+			const auto word = paragraph.substr(start, end - start);
+			const std::string_view candidate(line.empty() ? word.begin() : line.begin(),
+			                                 word.end());
+			// Keep an oversized word intact; the toast's clip rectangle handles overflow.
+			if (!line.empty() && font->CalcTextSizeA(font_size, std::numeric_limits<float>::max(),
+			                                       0.0f, candidate.data(),
+			                                       candidate.data() + candidate.size()).x > width) {
+				lines.push_back(line);
+				line = word;
+			} else {
+				line = candidate;
+			}
+			start = paragraph.find_first_not_of(" \t\r", end);
+		}
+		lines.push_back(line);
+		if (newline == std::string_view::npos) {
+			return lines;
+		}
+		text.remove_prefix(newline + 1);
+	}
+}
 
 void ClearInputEvents() {
 	std::scoped_lock lock(g_input_mutex);
@@ -467,8 +514,9 @@ SystemOverlayVisualState GetSystemOverlayVisualState() noexcept {
 	const auto core   = CoreIme::GetVisualState();
 	const auto dialog = DialogIme::GetVisualState();
 	const auto system = SystemDialog::GetVisualState();
-	return {core.active || dialog.active || system.active,
-	        core.revision + dialog.revision + system.revision};
+	const bool trophy_active = g_trophy_notification_active.load(std::memory_order_acquire);
+	return {core.active || dialog.active || system.active || trophy_active,
+	        core.revision + dialog.revision + system.revision + (trophy_active ? 1u : 0u)};
 }
 
 bool ProcessSystemOverlayInput(const SDL_Event& event) {
@@ -599,12 +647,34 @@ bool ProcessSystemOverlayInput(const SDL_Event& event) {
 	}
 }
 
+void NotifyTrophyUnlocked(std::string_view name, int32_t grade,
+                          std::span<const std::byte> icon_png) {
+	std::scoped_lock lock(g_trophy_notification_mutex);
+	g_trophy_notifications.push_back(
+	    {std::string(name), grade, std::vector<std::byte>(icon_png.begin(), icon_png.end())});
+	g_trophy_notification_active.store(true, std::memory_order_release);
+}
+
 struct SystemOverlay::Impl {
+	using Clock = std::chrono::steady_clock;
+
 	explicit Impl(GraphicContext& context): graphics(context) {}
 
 	~Impl() {
+		SDL_DestroyAudioStream(trophy_sound_stream);
+		SDL_free(trophy_sound_data);
+		if (trophy_sound_initialized) {
+			SDL_QuitSubSystem(SDL_INIT_AUDIO);
+		}
 		ReleaseVulkan();
 		if (imgui_context != nullptr) {
+			ImGui::SetCurrentContext(imgui_context);
+			if (trophy_notification_image) {
+				ImGui::UnregisterUserTexture(trophy_notification_image.get());
+			}
+			for (const auto& texture: retired_trophy_images) {
+				ImGui::UnregisterUserTexture(texture.get());
+			}
 			ImGui::DestroyContext(imgui_context);
 		}
 	}
@@ -631,6 +701,75 @@ struct SystemOverlay::Impl {
 		style.ItemSpacing    = {8.0f, 8.0f};
 	}
 
+	void PlayTrophySound() {
+		if (!trophy_sound_attempted) {
+			trophy_sound_attempted   = true;
+			trophy_sound_initialized = SDL_InitSubSystem(SDL_INIT_AUDIO);
+			if (!trophy_sound_initialized) {
+				LOGF("[Trophy] could not initialize audio: %s\n", SDL_GetError());
+				return;
+			}
+			const char* base_path = SDL_GetBasePath();
+			if (base_path == nullptr) {
+				LOGF("[Trophy] could not locate sound: %s\n", SDL_GetError());
+				return;
+			}
+			const std::string path = std::string(base_path) + "assets/sounds/trophy-unlock.wav";
+			SDL_AudioSpec     spec {};
+			if (!SDL_LoadWAV(path.c_str(), &spec, &trophy_sound_data, &trophy_sound_size)) {
+				LOGF("[Trophy] could not load %s: %s\n", path.c_str(), SDL_GetError());
+				return;
+			}
+			if (trophy_sound_size > static_cast<uint32_t>(std::numeric_limits<int>::max())) {
+				LOGF("[Trophy] sound is too large: %s\n", path.c_str());
+				return;
+			}
+			trophy_sound_stream = SDL_OpenAudioDeviceStream(SDL_AUDIO_DEVICE_DEFAULT_PLAYBACK,
+			                                                &spec, nullptr, nullptr);
+			if (trophy_sound_stream == nullptr) {
+				LOGF("[Trophy] could not open audio stream: %s\n", SDL_GetError());
+				return;
+			}
+		}
+		if (trophy_sound_stream != nullptr &&
+		    (!SDL_PutAudioStreamData(trophy_sound_stream, trophy_sound_data,
+		                             static_cast<int>(trophy_sound_size)) ||
+		     !SDL_ResumeAudioStreamDevice(trophy_sound_stream))) {
+			LOGF("[Trophy] could not play sound: %s\n", SDL_GetError());
+		}
+	}
+
+	std::unique_ptr<ImTextureData> LoadTrophyIcon(std::span<const std::byte> encoded) {
+		if (encoded.empty() ||
+		    encoded.size() > static_cast<std::size_t>(std::numeric_limits<int>::max())) {
+			return nullptr;
+		}
+
+		const auto* data   = reinterpret_cast<const stbi_uc*>(encoded.data());
+		const int   size   = static_cast<int>(encoded.size());
+		int         width  = 0;
+		int         height = 0;
+		if (!stbi_info_from_memory(data, size, &width, &height, nullptr) || width <= 0 ||
+		    height <= 0 || width > 512 || height > 512) {
+			LOGF("[Trophy] invalid icon or dimensions exceed 512x512\n");
+			return nullptr;
+		}
+		auto* pixels = stbi_load_from_memory(data, size, &width, &height, nullptr, 4);
+		if (pixels == nullptr) {
+			LOGF("[Trophy] could not decode trophy icon: %s\n", stbi_failure_reason());
+			return nullptr;
+		}
+
+		auto texture = std::make_unique<ImTextureData>();
+		texture->Create(ImTextureFormat_RGBA32, width, height);
+		std::memcpy(texture->Pixels, pixels,
+		            static_cast<std::size_t>(width) * static_cast<std::size_t>(height) * 4);
+		texture->UseColors = true;
+		stbi_image_free(pixels);
+		ImGui::RegisterUserTexture(texture.get());
+		return texture;
+	}
+
 	void EnsureVulkan(vk::Format format, uint32_t image_count) {
 		EnsureContext();
 		if (vulkan_initialized) {
@@ -952,9 +1091,131 @@ struct SystemOverlay::Impl {
 		}
 	}
 
+	void DrawTrophyNotification(vk::Extent2D frame_extent, Clock::time_point now) {
+		std::erase_if(retired_trophy_images, [](const auto& texture) {
+			if (texture->Status != ImTextureStatus_Destroyed) {
+				return false;
+			}
+			ImGui::UnregisterUserTexture(texture.get());
+			return true;
+		});
+		constexpr float TOAST_LIFETIME = 5.0f;
+		if (trophy_notification &&
+		    std::chrono::duration<float>(now - trophy_notification_born).count() >=
+		        TOAST_LIFETIME) {
+			trophy_notification.reset();
+			if (trophy_notification_image) {
+				// Keep the texture registered until the backend retires in-flight frames.
+				trophy_notification_image->WantDestroyNextFrame = true;
+				retired_trophy_images.push_back(std::move(trophy_notification_image));
+			}
+		}
+		if (!trophy_notification) {
+			{
+				std::scoped_lock lock(g_trophy_notification_mutex);
+				if (g_trophy_notifications.empty()) {
+					g_trophy_notification_active.store(!retired_trophy_images.empty(),
+					                                   std::memory_order_release);
+					return;
+				}
+				trophy_notification = std::move(g_trophy_notifications.front());
+				g_trophy_notifications.pop_front();
+			}
+			trophy_notification_born  = now;
+			trophy_notification_image = LoadTrophyIcon(trophy_notification->icon_png);
+			PlayTrophySound();
+		}
+		const float age = std::chrono::duration<float>(now - trophy_notification_born).count();
+		const float screen_scale =
+		    std::clamp(std::min(static_cast<float>(frame_extent.width) / 1920.0f,
+		                        static_cast<float>(frame_extent.height) / 1080.0f),
+		               0.65f, 1.25f);
+		const float width = std::max(
+		    0.0f, std::min(520.0f * screen_scale, static_cast<float>(frame_extent.width) - 32.0f));
+		constexpr std::array GRADE_COLORS {
+		    IM_COL32(161, 225, 245, 255), IM_COL32(255, 207, 82, 255), IM_COL32(211, 220, 232, 255),
+		    IM_COL32(211, 142, 91, 255)};
+		const int   grade        = trophy_notification->grade;
+		const bool  has_grade    = grade > 0 && grade <= static_cast<int>(GRADE_COLORS.size());
+		const float pad          = 10.0f * screen_scale;
+		const float base_height  = 104.0f * screen_scale;
+		const float icon_size    = base_height - 2.0f * pad;
+		const float text_offset  = trophy_notification_image != nullptr
+		                               ? pad + icon_size + 18.0f * screen_scale
+		                               : 20.0f * screen_scale;
+		const float title_offset = text_offset + (has_grade ? 30.0f * screen_scale : 0.0f);
+		const float title_size   = 26.0f * screen_scale;
+		const float line_height  = 30.0f * screen_scale;
+		const float text_padding = 16.0f * screen_scale;
+		const auto  title_lines  = WrapTrophyTitle(trophy_notification->name, title_size,
+		                                           width - title_offset - text_padding);
+		const float block_height =
+		    static_cast<float>(title_lines.size()) * line_height + 28.0f * screen_scale;
+		const float  height   = std::max(base_height, block_height + 2.0f * text_padding);
+		const float  rounding = 14.0f * screen_scale;
+		const float  fade_in  = std::clamp(age / 0.22f, 0.0f, 1.0f);
+		const float  fade_out = std::clamp((TOAST_LIFETIME - age) / 0.4f, 0.0f, 1.0f);
+		const float  alpha    = std::min(fade_in, fade_out);
+		const float  slide    = (1.0f - fade_in) * (width + 24.0f);
+		const ImVec2 top_left {static_cast<float>(frame_extent.width) - 32.0f - width + slide,
+		                       32.0f};
+		const ImVec2 bottom_right {top_left.x + width, top_left.y + height};
+		auto*        draw = ImGui::GetForegroundDrawList();
+		const auto   fade = [alpha](int r, int g, int b, int a) {
+            return IM_COL32(r, g, b, static_cast<int>(static_cast<float>(a) * alpha));
+		};
+		draw->AddRectFilled({top_left.x + 3.0f, top_left.y + 5.0f},
+		                    {bottom_right.x + 3.0f, bottom_right.y + 5.0f}, fade(0, 0, 0, 90),
+		                    rounding);
+		draw->AddRectFilled(top_left, bottom_right, fade(24, 26, 32, 235), rounding);
+		draw->AddRect(top_left, bottom_right, fade(255, 255, 255, 28), rounding,
+		              ImDrawFlags_RoundCornersAll, 1.0f);
+
+		const float text_left = top_left.x + text_offset;
+		if (trophy_notification_image != nullptr) {
+			const ImVec2 icon_min {top_left.x + pad, top_left.y + (height - icon_size) * 0.5f};
+			const ImVec2 icon_max {icon_min.x + icon_size, icon_min.y + icon_size};
+			draw->AddImageRounded(trophy_notification_image->GetTexRef(), icon_min, icon_max,
+			                      {0, 0}, {1, 1}, fade(255, 255, 255, 255), 8.0f * screen_scale);
+			draw->AddRect(icon_min, icon_max, fade(255, 255, 255, 40), 8.0f * screen_scale,
+			              ImDrawFlags_RoundCornersAll, 1.0f);
+		}
+
+		const float title_y = top_left.y + (height - block_height) * 0.5f;
+		const float title_x = top_left.x + title_offset;
+		if (has_grade) {
+			// Small cup glyph tinted with the trophy grade.
+			const ImU32 cup = ImGui::GetColorU32(GRADE_COLORS[grade - 1], alpha);
+			const float u   = screen_scale;
+			const float cx  = text_left + 9.0f * u;
+			const float cy  = title_y + 15.0f * u;
+			draw->AddRectFilled({cx - 8.0f * u, cy - 11.0f * u}, {cx + 8.0f * u, cy + 2.0f * u},
+			                    cup, 2.0f * u, ImDrawFlags_RoundCornersBottom);
+			draw->AddRect({cx - 11.0f * u, cy - 9.0f * u}, {cx - 6.0f * u, cy - 3.0f * u}, cup,
+			              2.0f * u, ImDrawFlags_RoundCornersLeft, 1.8f * u);
+			draw->AddRect({cx + 6.0f * u, cy - 9.0f * u}, {cx + 11.0f * u, cy - 3.0f * u}, cup,
+			              2.0f * u, ImDrawFlags_RoundCornersRight, 1.8f * u);
+			draw->AddRectFilled({cx - 1.5f * u, cy + 2.0f * u}, {cx + 1.5f * u, cy + 7.0f * u}, cup);
+			draw->AddRectFilled({cx - 5.0f * u, cy + 7.0f * u}, {cx + 5.0f * u, cy + 10.0f * u}, cup,
+			                    1.5f * u);
+		}
+		draw->PushClipRect(top_left, {bottom_right.x - text_padding, bottom_right.y}, true);
+		float line_y = title_y;
+		for (const auto line : title_lines) {
+			draw->AddText(nullptr, title_size, {title_x, line_y}, fade(255, 255, 255, 255),
+			              line.data(), line.data() + line.size());
+			line_y += line_height;
+		}
+		draw->AddText(nullptr, 20.0f * screen_scale, {text_left, line_y + 8.0f * screen_scale},
+		              fade(176, 180, 192, 255), "Trophy earned!");
+		draw->PopClipRect();
+	}
+
 	bool PrepareFrame(vk::Extent2D frame_extent, vk::Format format, uint32_t image_count) {
 		OverlaySnapshot snapshot;
-		if (!GetOverlaySnapshot(&snapshot)) {
+		const bool      has_overlay = GetOverlaySnapshot(&snapshot);
+		const auto now = Clock::now();
+		if (!has_overlay && !g_trophy_notification_active.load(std::memory_order_acquire)) {
 			return false;
 		}
 		const auto prepared_session = snapshot.session;
@@ -977,7 +1238,6 @@ struct SystemOverlay::Impl {
 		auto& io       = ImGui::GetIO();
 		io.DisplaySize = {static_cast<float>(frame_extent.width),
 		                  static_cast<float>(frame_extent.height)};
-		const auto now = std::chrono::steady_clock::now();
 		io.DeltaTime   = last_frame == std::chrono::steady_clock::time_point {}
 		                     ? 1.0f / 60.0f
 		                     : std::clamp(std::chrono::duration<float>(now - last_frame).count(),
@@ -985,15 +1245,19 @@ struct SystemOverlay::Impl {
 		last_frame     = now;
 		ImGui_ImplVulkan_NewFrame();
 		ImGui::NewFrame();
-		if (!GetOverlaySnapshot(&snapshot) || snapshot.session != prepared_session) {
+		GetOverlaySnapshot(&snapshot);
+		if (snapshot.session != prepared_session) {
 			ImGui::EndFrame();
 			return false;
 		}
-		if (snapshot.session.kind == OverlayKind::Dialog) {
-			DrawDialog(snapshot.dialog, frame_extent);
-		} else {
-			DrawIme(snapshot.ime, frame_extent);
+		if (has_overlay) {
+			if (snapshot.session.kind == OverlayKind::Dialog) {
+				DrawDialog(snapshot.dialog, frame_extent);
+			} else {
+				DrawIme(snapshot.ime, frame_extent);
+			}
 		}
+		DrawTrophyNotification(frame_extent, now);
 		ImGui::Render();
 		extent = frame_extent;
 		return true;
@@ -1040,9 +1304,18 @@ struct SystemOverlay::Impl {
 	float                                 button_height      = 42.0f;
 	ImVec2                                panel_offset {};
 	ImVec2                                right_stick {};
-	OverlaySession                        session;
-	vk::Extent2D                          extent {};
-	std::chrono::steady_clock::time_point last_frame;
+	OverlaySession                              session;
+	std::unique_ptr<ImTextureData>              trophy_notification_image;
+	std::optional<TrophyNotification>           trophy_notification;
+	Clock::time_point                           trophy_notification_born;
+	std::vector<std::unique_ptr<ImTextureData>> retired_trophy_images;
+	SDL_AudioStream*                            trophy_sound_stream      = nullptr;
+	uint8_t*                                    trophy_sound_data        = nullptr;
+	uint32_t                                    trophy_sound_size        = 0;
+	bool                                        trophy_sound_attempted   = false;
+	bool                                        trophy_sound_initialized = false;
+	vk::Extent2D                                extent {};
+	std::chrono::steady_clock::time_point       last_frame;
 };
 
 SystemOverlay::SystemOverlay(GraphicContext& graphics): m_impl(std::make_unique<Impl>(graphics)) {}

src/graphics/presentation/systemOverlay.h

@@ -4,7 +4,10 @@
 #include "common/common.h"
 #include "graphics/host_gpu/vulkanCommon.h"
 
+#include <cstddef>
 #include <memory>
+#include <span>
+#include <string_view>
 
 union SDL_Event;
 struct SDL_Window;
@@ -22,6 +25,8 @@ void                     InitializeSystemOverlayInput(SDL_Window* window);
 void                     ShutdownSystemOverlayInput();
 bool                     ProcessSystemOverlayInput(const SDL_Event& event);
 SystemOverlayVisualState GetSystemOverlayVisualState() noexcept;
+void                     NotifyTrophyUnlocked(std::string_view name, int32_t grade,
+                                              std::span<const std::byte> icon_png);
 
 class SystemOverlay final {
 public:

src/graphics/presentation/window/hostInput.cpp

@@ -454,7 +454,7 @@ void HostInputToggleMouseToJoystick() {
 	LOGF("Mouse to right stick: enabled (F7 to release)\n");
 }
 
-bool HostInputWaitEvent(SDL_Event* event) {
+bool HostInputWaitEvent(SDL_Event* event, int max_wait_ms) {
 	int timeout = -1;
 	if (!g_mouse.enabled || SDL_GetKeyboardFocus() != g_mouse_window) {
 		g_mouse.next_poll = 0;
@@ -473,6 +473,7 @@ bool HostInputWaitEvent(SDL_Event* event) {
 		    now_ms < g_cursor_hide_at ? static_cast<int>(g_cursor_hide_at - now_ms) : 0;
 		timeout = timeout < 0 ? cursor_timeout : std::min(timeout, cursor_timeout);
 	}
+	timeout              = timeout < 0 ? max_wait_ms : std::min(timeout, max_wait_ms);
 	const bool has_event = SDL_WaitEventTimeout(event, timeout);
 
 	if (Config::HideCursorEnabled()) {

src/graphics/presentation/window/hostInput.h

@@ -13,7 +13,7 @@ void               HostInputShutdown();
 void               HostInputKey(int key_code, bool down);
 void               HostInputMouseButton(uint8_t mouse_button, bool down);
 void               HostInputToggleMouseToJoystick();
-[[nodiscard]] bool HostInputWaitEvent(SDL_Event* event);
+[[nodiscard]] bool HostInputWaitEvent(SDL_Event* event, int max_wait_ms);
 
 } // namespace Libs::Graphics
 

src/graphics/presentation/window/swapchain.cpp

@@ -998,7 +998,7 @@ void Presenter::Impl::Present() {
 		}
 
 		presented_overlay_revision.store(overlay_visual.revision, std::memory_order_release);
-		window.UpdateTitle();
+		window.loop.presented_frames.fetch_add(1, std::memory_order_relaxed);
 		return;
 	}
 	LOGF("Vulkan presentation retry exhausted; dropping frame\n");

src/graphics/presentation/window/window.cpp

@@ -722,6 +722,10 @@ void WindowContext::Run() {
 	loop.need_exit = false;
 	loop.paused.store(false, std::memory_order_release);
 
+	constexpr uint64_t title_interval_ms = 1000;
+	auto               title_time        = SDL_GetTicks();
+	auto               title_frames      = loop.presented_frames.load(std::memory_order_relaxed);
+
 	while (!loop.need_exit) {
 		if (loop.paused.load(std::memory_order_acquire)) {
 			if (!timer.IsPaused()) {
@@ -731,7 +735,20 @@ void WindowContext::Run() {
 			timer.Resume();
 		}
 
-		if (!HostInputWaitEvent(&loop.event)) {
+		// Refresh the title on the main thread without making presentation wait for it.
+		const auto now     = SDL_GetTicks();
+		const auto elapsed = now - title_time;
+		if (elapsed >= title_interval_ms) {
+			const auto frames = loop.presented_frames.load(std::memory_order_relaxed);
+			if (frames != 0) {
+				UpdateTitle(frames, static_cast<double>(frames - title_frames) * 1000.0 /
+				                        static_cast<double>(elapsed));
+			}
+			title_time   = now;
+			title_frames = frames;
+		}
+		const auto wait_ms = static_cast<int>(title_interval_ms - (now - title_time));
+		if (!HostInputWaitEvent(&loop.event, wait_ms)) {
 			continue;
 		}
 		ProcessEvent(timer.GetTimeS());
@@ -913,7 +930,7 @@ void WindowContext::UpdateIcon() {
 	}
 }
 
-void WindowContext::UpdateTitle() {
+void WindowContext::UpdateTitle(uint64_t frame_num, double current_fps) {
 	static char title[128];
 	static char title_id[12];
 	static char app_ver[12];
@@ -923,10 +940,6 @@ void WindowContext::UpdateTitle() {
 	static bool has_app_ver =
 	    Loader::SystemContentParamSfoGetString("APP_VER", app_ver, sizeof(app_ver));
 	static const std::string processor_name = Common::GetSystemInfo().ProcessorName;
-	static uint64_t fps_start   = Common::Timer::QueryPerformanceCounter();
-	static uint64_t frame_num   = 0;
-	static uint64_t fps_frames  = 0;
-	static double   current_fps = 0.0;
 
 #if KYTY_BUILD == KYTY_BUILD_DEBUG
 	static constexpr auto build_type = "Debug";
@@ -936,17 +949,6 @@ void WindowContext::UpdateTitle() {
 	static constexpr auto build_type = "Unknown";
 #endif
 
-	const auto now       = Common::Timer::QueryPerformanceCounter();
-	const auto frequency = Common::Timer::QueryPerformanceFrequency();
-	frame_num++;
-	fps_frames++;
-	if (now - fps_start >= frequency) {
-		current_fps = static_cast<double>(fps_frames) * static_cast<double>(frequency) /
-		              static_cast<double>(now - fps_start);
-		fps_start   = now;
-		fps_frames  = 0;
-	}
-
 	const auto* device_name = graphic_ctx.GetPhysicalDeviceProperties().deviceName.data();
 	auto text = fmt::format(
 	    "[{} | {}] {}{}{}{}{}{}[{}] [{}], frame: {}, fps: {:.0f}", KYTY_BUILD_LABEL, build_type,
@@ -954,16 +956,7 @@ void WindowContext::UpdateTitle() {
 	    (has_title_id ? ", " : ""), (has_app_ver ? app_ver : ""), (has_app_ver ? " " : ""),
 	    device_name, processor_name, frame_num, current_fps);
 
-	struct TitleUpdate {
-		SDL_Window*  window;
-		std::string* text;
-	} update {window, &text};
-	EXIT_IF(!SDL_RunOnMainThread(
-	    [](void* data) {
-		    auto& title = *static_cast<TitleUpdate*>(data);
-		    SDL_SetWindowTitle(title.window, title.text->c_str());
-	    },
-	    &update, true));
+	SDL_SetWindowTitle(window, text.c_str());
 }
 
 } // namespace Libs::Graphics

src/graphics/presentation/window/windowInternal.h

@@ -27,6 +27,7 @@ struct WindowLoopState {
 	SDL_Event        event {};
 	bool             need_exit = false;
 	std::atomic_bool paused    = false;
+	std::atomic<uint64_t> presented_frames {0};
 };
 
 struct WindowContext {
@@ -42,7 +43,7 @@ struct WindowContext {
 	void                                                    RecreateSurface();
 	void                                                    RefreshSurfaceCapabilities();
 	void                                                    UpdateIcon();
-	void                                                    UpdateTitle();
+	void UpdateTitle(uint64_t frame_num, double current_fps);
 	void                                                    Resize(int width, int height);
 	void ProcessWindowEvent(const SDL_WindowEvent& event);
 	void ProcessDisplayEvent(const SDL_DisplayEvent& event);

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

@@ -205,7 +205,8 @@ Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program
 				if (memory_index >= program.memory_info.size()) {
 					Fail(program, "address operation has invalid memory metadata");
 				}
-				const auto kind = program.memory_info[memory_index].kind;
+				const auto& memory = program.memory_info[memory_index];
+				const auto kind = memory.kind;
 				if (kind == IR::ResourceKind::Scratch || kind == IR::ResourceKind::FlatLocal) {
 					if (program.scratch_dwords == 0) {
 						Fail(program, "scratch operation has no per-thread storage");
@@ -217,6 +218,9 @@ Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program
 					}
 				} else if (address_access == IR::AddressAccess::Write) {
 					Fail(program, "writable FLAT/GLOBAL addresses require GPU ownership tracking");
+				} else {
+					// Descriptor stores can alias coherent physical-address loads.
+					requirements.coherent_buffers |= memory.coherent;
 				}
 			}
 			if (IR::BufferAccessOf(inst.GetOpcode()) != IR::BufferAccess::None) {
@@ -276,6 +280,10 @@ Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program
 			switch (inst.GetOpcode()) {
 				case IR::ValueOpcode::StoreCompletion: requirements.subgroup_barrier = true; break;
 				case IR::ValueOpcode::BvhIntersect: requirements.bvh = true; break;
+				case IR::ValueOpcode::ConditionRef:
+					requirements.subgroup_ballot |=
+					    inst.Flags<CFG::BranchCondition>() != CFG::BranchCondition::ScalarInstruction;
+					break;
 				case IR::ValueOpcode::Ballot: requirements.subgroup_ballot = true; break;
 				case IR::ValueOpcode::DppMoveU32:
 				case IR::ValueOpcode::ReadFirstLane:

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

@@ -17,21 +17,8 @@ Pair ExtractPair(EmitterState& state, uint32_t value) {
 	return result;
 }
 
-uint32_t MakePair(EmitterState& state, uint32_t low, uint32_t high) {
-	const auto result = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpCompositeConstruct, TypeU64(state), result, low, high);
-	return result;
-}
-
-uint32_t CompareEqual64(EmitterState& state, uint32_t lhs_value, uint32_t rhs_value,
-                        bool not_equal) {
-	const auto compare = Binary(state, not_equal ? spv::OpINotEqual : spv::OpIEqual,
-	                            TypeBoolVector(state, 2), lhs_value, rhs_value);
-	return Unary(state, not_equal ? spv::OpAny : spv::OpAll, TypeBool(state), compare);
-}
-
 uint32_t EmitMinMaxF64(EmitterState& state, uint32_t lhs, uint32_t rhs, bool max_value) {
-	const auto bits_type = TypeU64(state);
+	const auto bits_type = TypeU32Vector(state, 2);
 	const auto lhs_bits = Unary(state, spv::OpBitcast, bits_type, lhs);
 	const auto rhs_bits = Unary(state, spv::OpBitcast, bits_type, rhs);
 	const auto ordered = Binary(state, max_value ? spv::OpFOrdGreaterThan : spv::OpFOrdLessThan,
@@ -52,17 +39,6 @@ uint32_t EmitMinMaxF64(EmitterState& state, uint32_t lhs, uint32_t rhs, bool max
 	              rhs, result);
 }
 
-uint32_t CompareOrdered64(EmitterState& state, uint32_t lhs_value, uint32_t rhs_value,
-                          spv::Op high_compare, spv::Op low_compare) {
-	const auto lhs         = ExtractPair(state, lhs_value);
-	const auto rhs         = ExtractPair(state, rhs_value);
-	const auto high_equal  = Binary(state, spv::OpIEqual, TypeBool(state), lhs.high, rhs.high);
-	const auto high_result = Binary(state, high_compare, TypeBool(state), lhs.high, rhs.high);
-	const auto low_result  = Binary(state, low_compare, TypeBool(state), lhs.low, rhs.low);
-	const auto low_path = Binary(state, spv::OpLogicalAnd, TypeBool(state), high_equal, low_result);
-	return Binary(state, spv::OpLogicalOr, TypeBool(state), high_result, low_path);
-}
-
 uint32_t EmitMulHigh(EmitterState& state, uint32_t lhs, uint32_t rhs, bool signed_value) {
 	const auto operand_type = signed_value ? TypeI32(state) : TypeU32(state);
 	const auto pair_type    = signed_value ? TypeI32Pair(state) : TypeU32Pair(state);
@@ -80,86 +56,6 @@ uint32_t EmitMulHigh(EmitterState& state, uint32_t lhs, uint32_t rhs, bool signe
 	return signed_value ? Unary(state, spv::OpBitcast, TypeU32(state), high) : high;
 }
 
-uint32_t EmitShift64(EmitterState& state, spv::Op opcode, uint32_t value, uint32_t shift) {
-	const auto pair       = ExtractPair(state, value);
-	const auto amount     = EmitAndConstant(state, shift, 63u);
-	const auto word_shift = EmitAndConstant(state, amount, 31u);
-	const auto at_least_32 =
-	    Binary(state, spv::OpUGreaterThanEqual, TypeBool(state), amount, ConstantU32(state, 32));
-	const auto nonzero =
-	    Binary(state, spv::OpINotEqual, TypeBool(state), amount, ConstantU32(state, 0));
-	const auto carry_count = EmitAndConstant(
-	    state, Binary(state, spv::OpISub, TypeU32(state), ConstantU32(state, 32), word_shift), 31u);
-	if (opcode == spv::OpShiftLeftLogical) {
-		const auto low = Binary(state, opcode, TypeU32(state), pair.low, word_shift);
-		const auto carry =
-		    Select(state, TypeU32(state), nonzero,
-		           Binary(state, spv::OpShiftRightLogical, TypeU32(state), pair.low, carry_count),
-		           ConstantU32(state, 0));
-		const auto high =
-		    Binary(state, spv::OpBitwiseOr, TypeU32(state),
-		           Binary(state, opcode, TypeU32(state), pair.high, word_shift), carry);
-		return MakePair(state,
-		                Select(state, TypeU32(state), at_least_32, ConstantU32(state, 0), low),
-		                Select(state, TypeU32(state), at_least_32, low, high));
-	}
-	const auto high = Binary(state, opcode, TypeU32(state), pair.high, word_shift);
-	const auto carry =
-	    Select(state, TypeU32(state), nonzero,
-	           Binary(state, spv::OpShiftLeftLogical, TypeU32(state), pair.high, carry_count),
-	           ConstantU32(state, 0));
-	const auto low = Binary(
-	    state, spv::OpBitwiseOr, TypeU32(state),
-	    Binary(state, spv::OpShiftRightLogical, TypeU32(state), pair.low, word_shift), carry);
-	const auto fill = opcode == spv::OpShiftRightArithmetic
-	                      ? Binary(state, opcode, TypeU32(state), pair.high, ConstantU32(state, 31))
-	                      : ConstantU32(state, 0);
-	return MakePair(state, Select(state, TypeU32(state), at_least_32, high, low),
-	                Select(state, TypeU32(state), at_least_32, fill, high));
-}
-
-uint32_t EmitConstantShift64(EmitterState& state, spv::Op opcode, uint32_t value, uint32_t shift) {
-	shift &= 63u;
-	if (shift == 0u) {
-		return value;
-	}
-	const auto pair = ExtractPair(state, value);
-	if (opcode == spv::OpShiftLeftLogical) {
-		if (shift < 32u) {
-			const auto low  = Binary(state, spv::OpShiftLeftLogical, TypeU32(state), pair.low,
-			                         ConstantU32(state, shift));
-			const auto high = Binary(state, spv::OpBitwiseOr, TypeU32(state),
-			                         Binary(state, spv::OpShiftLeftLogical, TypeU32(state),
-			                                pair.high, ConstantU32(state, shift)),
-			                         Binary(state, spv::OpShiftRightLogical, TypeU32(state),
-			                                pair.low, ConstantU32(state, 32u - shift)));
-			return MakePair(state, low, high);
-		}
-		return MakePair(state, ConstantU32(state, 0),
-		                shift == 32u ? pair.low
-		                             : Binary(state, spv::OpShiftLeftLogical, TypeU32(state),
-		                                      pair.low, ConstantU32(state, shift - 32u)));
-	}
-	if (shift < 32u) {
-		const auto low = Binary(state, spv::OpBitwiseOr, TypeU32(state),
-		                        Binary(state, spv::OpShiftRightLogical, TypeU32(state), pair.low,
-		                               ConstantU32(state, shift)),
-		                        Binary(state, spv::OpShiftLeftLogical, TypeU32(state), pair.high,
-		                               ConstantU32(state, 32u - shift)));
-		const auto high =
-		    Binary(state, opcode, TypeU32(state), pair.high, ConstantU32(state, shift));
-		return MakePair(state, low, high);
-	}
-	const auto high = opcode == spv::OpShiftRightArithmetic
-	                      ? Binary(state, spv::OpShiftRightArithmetic, TypeU32(state), pair.high,
-	                               ConstantU32(state, 31u))
-	                      : ConstantU32(state, 0);
-	const auto low  = shift == 32u ? pair.high
-	                               : Binary(state, opcode, TypeU32(state), pair.high,
-	                                        ConstantU32(state, shift - 32u));
-	return MakePair(state, low, high);
-}
-
 uint32_t EmitMinMax3(EmitterState& state, uint32_t a, uint32_t b, uint32_t c, bool signed_value,
                      bool max_value) {
 	const auto ab = signed_value ? EmitMinMaxI32Value(state, a, b, max_value)
@@ -223,6 +119,28 @@ uint32_t EmitF32ToU32(EmitterState& state, uint32_t src, bool signed_value) {
 }
 
 } // 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);
+	return result;
+}
+
+uint32_t EmitFPMad32(EmitterState& state, uint32_t a, uint32_t b, uint32_t c) {
+	// Explicit denormal checks typically drop 3DMiniGolf menu FPS from 28-29 to 21-24
+	// on NVIDIA RTX 5080 Laptop GPU, so leave them disabled for this experiment.
+	// a = EmitFlushF32DenormToSignedZero(state, a);
+	// b = EmitFlushF32DenormToSignedZero(state, b);
+	// c = EmitFlushF32DenormToSignedZero(state, c);
+	const auto product = EmitFPMul32(state, a, b);
+	state.builder.AddAnnotation(spv::OpDecorate, product, spv::DecorationNoContraction);
+	// const auto sum = EmitFPAdd32(state, EmitFlushF32DenormToSignedZero(state, product), c);
+	const auto sum = EmitFPAdd32(state, product, c);
+	state.builder.AddAnnotation(spv::OpDecorate, sum, spv::DecorationNoContraction);
+	// return EmitFlushF32DenormToSignedZero(state, sum);
+	return sum;
+}
+
 uint32_t EmitFPMedTri32(EmitterState& state, uint32_t a, uint32_t b, uint32_t c) {
 	const auto min_ab   = EmitMinMaxF32Value(state, a, b, false);
 	const auto min3     = EmitMinMaxF32Value(state, min_ab, c, false);
@@ -237,57 +155,15 @@ uint32_t EmitFPMedTri32(EmitterState& state, uint32_t a, uint32_t b, uint32_t c)
 }
 
 uint32_t EmitFindUMsb64(EmitterState& state, uint32_t value) {
-	const auto pair   = ExtractPair(state, value);
-	const auto high_i = state.builder.AllocateId();
-	const auto low_i  = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpExtInst, TypeI32(state), high_i, GlslStd450(state),
-	                          GLSLstd450FindUMsb, pair.high);
-	state.builder.AddFunction(spv::OpExtInst, TypeI32(state), low_i, GlslStd450(state),
-	                          GLSLstd450FindUMsb, pair.low);
-	const auto high = Unary(state, spv::OpBitcast, TypeU32(state), high_i);
-	const auto low  = Unary(state, spv::OpBitcast, TypeU32(state), low_i);
-	const auto high_nonzero =
-	    Binary(state, spv::OpINotEqual, TypeBool(state), pair.high, ConstantU32(state, 0));
-	return Select(state, TypeU32(state), high_nonzero,
-	              Binary(state, spv::OpIAdd, TypeU32(state), high, ConstantU32(state, 32)), low);
-}
-
-uint32_t EmitIMul64(EmitterState& state, uint32_t lhs_value, uint32_t rhs_value) {
-	const auto lhs   = ExtractPair(state, lhs_value);
-	const auto rhs   = ExtractPair(state, rhs_value);
-	const auto low   = Binary(state, spv::OpIMul, TypeU32(state), lhs.low, rhs.low);
-	const auto high0 = EmitMulHigh(state, lhs.low, rhs.low, false);
-	const auto high1 = Binary(state, spv::OpIMul, TypeU32(state), lhs.low, rhs.high);
-	const auto high2 = Binary(state, spv::OpIMul, TypeU32(state), lhs.high, rhs.low);
-	return MakePair(state, low,
-	                Binary(state, spv::OpIAdd, TypeU32(state),
-	                       Binary(state, spv::OpIAdd, TypeU32(state), high0, high1), high2));
-}
-
-uint32_t EmitISub64(EmitterState& state, uint32_t lhs_value, uint32_t rhs_value) {
-	const auto lhs    = ExtractPair(state, lhs_value);
-	const auto rhs    = ExtractPair(state, rhs_value);
-	const auto low    = Binary(state, spv::OpISub, TypeU32(state), lhs.low, rhs.low);
-	const auto borrow = Binary(state, spv::OpULessThan, TypeBool(state), lhs.low, rhs.low);
-	const auto borrow_u32 =
-	    Select(state, TypeU32(state), borrow, ConstantU32(state, 1), ConstantU32(state, 0));
-	const auto high0 = Binary(state, spv::OpISub, TypeU32(state), lhs.high, rhs.high);
-	const auto high  = Binary(state, spv::OpISub, TypeU32(state), high0, borrow_u32);
-	return MakePair(state, low, high);
-}
-
-uint32_t EmitIAdd64(EmitterState& state, uint32_t lhs_value, uint32_t rhs_value) {
-	const auto lhs      = ExtractPair(state, lhs_value);
-	const auto rhs      = ExtractPair(state, rhs_value);
-	const auto low_pair = state.builder.AllocateId();
-	const auto low      = state.builder.AllocateId();
-	const auto carry    = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpIAddCarry, TypeU32Pair(state), low_pair, lhs.low, rhs.low);
-	state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(state), low, low_pair, 0);
-	state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(state), carry, low_pair, 1);
-	const auto high0 = Binary(state, spv::OpIAdd, TypeU32(state), lhs.high, rhs.high);
-	const auto high  = Binary(state, spv::OpIAdd, TypeU32(state), high0, carry);
-	return MakePair(state, low, high);
+	// GLSL FindUMsb is restricted to 32-bit components.
+	const auto unpacked = Unary(state, spv::OpBitcast, TypeU32Vector(state, 2), value);
+	const auto pair =
+	    ExtractPair(state, EmitExt(state, TypeU32Vector(state, 2), GLSLstd450FindUMsb, {unpacked}));
+	const auto high_found =
+	    Binary(state, spv::OpINotEqual, TypeBool(state), pair.high, ConstantU32(state, UINT32_MAX));
+	return Select(state, TypeU32(state), high_found,
+	              Binary(state, spv::OpIAdd, TypeU32(state), pair.high, ConstantU32(state, 32)),
+	              pair.low);
 }
 
 uint32_t EmitConvertU16U32(EmitterState& state, uint32_t arg0) {
@@ -316,7 +192,7 @@ uint32_t EmitConvertU32F32(EmitterState& state, uint32_t arg0) {
 uint32_t EmitConvertF32F64(EmitterState& state, uint32_t arg0) {
 	const auto converted = Unary(state, spv::OpFConvert, TypeF32(state), arg0);
 	const auto source    = EmitNative<spv::OpCompositeExtract, IR::Type::U32>(
-	    state, Unary(state, spv::OpBitcast, TypeU64(state), arg0), 1u);
+	    state, Unary(state, spv::OpBitcast, TypeU32Vector(state, 2), arg0), 1u);
 	const auto exponent = EmitAndConstant(state, source, 0x7ff00000u);
 	const auto overflow =
 	    Binary(state, spv::OpLogicalAnd, TypeBool(state),
@@ -334,7 +210,21 @@ uint32_t EmitConvertF64F32(EmitterState& state, uint32_t arg0) {
 	                                                  EmitFlushF32DenormToSignedZero(state, arg0));
 }
 
+uint32_t EmitCompositeConstructU64(ValueEmitContext& ctx, uint32_t arg0, IR::Value arg1) {
+	arg1 = arg1.Resolve();
+	return arg1.IsImmediate() && arg1.U32() == 0u
+	           ? Unary(ctx.state, spv::OpUConvert, TypeU64(ctx.state), arg0)
+	           : PackU64(ctx.state, arg0, ctx.Def(arg1));
+}
+
 uint32_t EmitCompositeExtractU64(EmitterState& state, uint32_t arg0, IR::Value arg1) {
+	const auto value = arg1.U32() == 0u ? arg0
+	                                    : Binary(state, spv::OpShiftRightLogical, TypeU64(state),
+	                                             arg0, ConstantU32(state, 32));
+	return Unary(state, spv::OpUConvert, TypeU32(state), value);
+}
+
+uint32_t EmitCompositeExtractU32x2(EmitterState& state, uint32_t arg0, IR::Value arg1) {
 	return EmitNative<spv::OpCompositeExtract, IR::Type::U32>(state, arg0, arg1.U32());
 }
 
@@ -366,33 +256,11 @@ uint32_t EmitIAbs32(EmitterState& state, uint32_t arg0) {
 	return Select(state, TypeU32(state), negative, neg, value);
 }
 
-uint32_t EmitShiftLeftLogical64(ValueEmitContext& ctx, uint32_t arg0, IR::Value arg1) {
-	auto& state = ctx.state;
-	return arg1.Resolve().IsImmediate()
-	           ? EmitConstantShift64(state, spv::OpShiftLeftLogical, arg0, arg1.Resolve().U32())
-	           : EmitShift64(state, spv::OpShiftLeftLogical, arg0, ctx.Def(arg1));
-}
-
-uint32_t EmitShiftRightLogical64(ValueEmitContext& ctx, uint32_t arg0, IR::Value arg1) {
-	auto& state = ctx.state;
-	return arg1.Resolve().IsImmediate()
-	           ? EmitConstantShift64(state, spv::OpShiftRightLogical, arg0, arg1.Resolve().U32())
-	           : EmitShift64(state, spv::OpShiftRightLogical, arg0, ctx.Def(arg1));
-}
-
-uint32_t EmitShiftRightArithmetic64(ValueEmitContext& ctx, uint32_t arg0, IR::Value arg1) {
-	auto& state = ctx.state;
-	return arg1.Resolve().IsImmediate()
-	           ? EmitConstantShift64(state, spv::OpShiftRightArithmetic, arg0, arg1.Resolve().U32())
-	           : EmitShift64(state, spv::OpShiftRightArithmetic, arg0, ctx.Def(arg1));
-}
-
-uint32_t EmitBitwiseAnd64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
-	return Binary(state, spv::OpBitwiseAnd, TypeU64(state), arg0, arg1);
-}
-
 uint32_t EmitBitCount64(EmitterState& state, uint32_t arg0) {
-	const auto pair = ExtractPair(state, Unary(state, spv::OpBitCount, TypeU64(state), arg0));
+	// Vulkan bit-count operands remain 32-bit, including when counting a U64.
+	const auto unpacked = Unary(state, spv::OpBitcast, TypeU32Vector(state, 2), arg0);
+	const auto pair =
+	    ExtractPair(state, Unary(state, spv::OpBitCount, TypeU32Vector(state, 2), unpacked));
 	return Binary(state, spv::OpIAdd, TypeU32(state), pair.low, pair.high);
 }
 
@@ -446,38 +314,6 @@ uint32_t EmitUMedTri32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32
 	return EmitMed3(state, arg0, arg1, arg2, false);
 }
 
-uint32_t EmitIEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
-	return CompareEqual64(state, arg0, arg1, false);
-}
-
-uint32_t EmitINotEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
-	return CompareEqual64(state, arg0, arg1, true);
-}
-
-uint32_t EmitULessThan64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
-	return CompareOrdered64(state, arg0, arg1, spv::OpULessThan, spv::OpULessThan);
-}
-
-uint32_t EmitSLessThan64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
-	return CompareOrdered64(state, arg0, arg1, spv::OpSLessThan, spv::OpULessThan);
-}
-
-uint32_t EmitUGreaterThan64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
-	return CompareOrdered64(state, arg0, arg1, spv::OpUGreaterThan, spv::OpUGreaterThan);
-}
-
-uint32_t EmitSLessThanEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
-	return CompareOrdered64(state, arg0, arg1, spv::OpSLessThan, spv::OpULessThanEqual);
-}
-
-uint32_t EmitULessThanEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
-	return CompareOrdered64(state, arg0, arg1, spv::OpULessThan, spv::OpULessThanEqual);
-}
-
-uint32_t EmitUGreaterThanEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
-	return CompareOrdered64(state, arg0, arg1, spv::OpUGreaterThan, spv::OpUGreaterThanEqual);
-}
-
 uint32_t EmitFPIsNan32(EmitterState& state, uint32_t arg0) {
 	return EmitNative<spv::OpFUnordNotEqual, IR::Type::U1>(state, arg0, arg0);
 }

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

@@ -125,20 +125,23 @@ uint32_t EmitClassMaskBitMatch(EmitterState& state, uint32_t mask, uint32_t bit,
 	return EmitLogicalAndBool(state, selected, class_match);
 }
 
-uint32_t EmitClassMaskF32(EmitterState& state, uint32_t value, uint32_t mask) {
-	const auto bits          = EmitBitcastF32ToU32(state, value);
-	const auto sign_bits     = EmitAndConstant(state, bits, 0x80000000u);
-	const auto abs_bits      = EmitAndConstant(state, bits, 0x7fffffffu);
-	const auto exponent_bits = EmitAndConstant(state, abs_bits, 0x7f800000u);
-	const auto mantissa_bits = EmitAndConstant(state, abs_bits, 0x007fffffu);
-	const auto quiet_bits    = EmitAndConstant(state, mantissa_bits, 0x00400000u);
+static uint32_t EmitClassMaskBits(EmitterState& state, uint32_t bits, uint32_t mask, bool half) {
+	const uint32_t sign_mask     = half ? 0x8000u : 0x80000000u;
+	const uint32_t exponent_mask = half ? 0x7c00u : 0x7f800000u;
+	const uint32_t mantissa_mask = half ? 0x03ffu : 0x007fffffu;
+	const uint32_t quiet_mask    = half ? 0x0200u : 0x00400000u;
+	const auto     sign_bits     = EmitAndConstant(state, bits, sign_mask);
+	const auto     abs_bits      = EmitAndConstant(state, bits, sign_mask - 1u);
+	const auto     exponent_bits = EmitAndConstant(state, abs_bits, exponent_mask);
+	const auto     mantissa_bits = EmitAndConstant(state, abs_bits, mantissa_mask);
+	const auto     quiet_bits    = EmitAndConstant(state, mantissa_bits, quiet_mask);
 
 	const auto sign             = EmitCompareU32Constant(state, spv::OpINotEqual, sign_bits, 0);
 	const auto positive         = EmitLogicalNotBool(state, sign);
 	const auto exponent_zero    = EmitCompareU32Constant(state, spv::OpIEqual, exponent_bits, 0);
 	const auto exponent_nonzero = EmitLogicalNotBool(state, exponent_zero);
 	const auto exponent_inf =
-	    EmitCompareU32Constant(state, spv::OpIEqual, exponent_bits, 0x7f800000u);
+	    EmitCompareU32Constant(state, spv::OpIEqual, exponent_bits, exponent_mask);
 	const auto finite_exponent  = EmitLogicalNotBool(state, exponent_inf);
 	const auto mantissa_zero    = EmitCompareU32Constant(state, spv::OpIEqual, mantissa_bits, 0);
 	const auto mantissa_nonzero = EmitLogicalNotBool(state, mantissa_zero);
@@ -178,6 +181,14 @@ uint32_t EmitClassMaskF32(EmitterState& state, uint32_t value, uint32_t mask) {
 	    EmitClassMaskBitMatch(state, mask, 9, EmitLogicalAndBool(state, inf, positive)));
 }
 
+uint32_t EmitClassMaskF32(EmitterState& state, uint32_t value, uint32_t mask) {
+	return EmitClassMaskBits(state, EmitBitcastF32ToU32(state, value), mask, false);
+}
+
+uint32_t EmitClassMaskF16(EmitterState& state, uint32_t bits, uint32_t mask) {
+	return EmitClassMaskBits(state, bits, mask, true);
+}
+
 uint32_t EmitMinMaxF32Value(EmitterState& state, uint32_t lhs, uint32_t rhs, bool max_value) {
 	const auto lhs_class = EmitClassifyF32(state, lhs);
 	const auto rhs_class = EmitClassifyF32(state, rhs);

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

@@ -136,8 +136,8 @@ uint32_t ImageType(EmitterState& state, const IR::ImageResource& image) {
 		EXIT("invalid image resource class");
 	}
 	const auto& info = ImageDimensionInfoFor(image.dimension);
-	const auto scalar_type = image.atomic64 ? TypeScalarU64(state)
-	                                        : ImageScalarType(state, image.numeric_class);
+	const auto  scalar_type =
+	    image.atomic64 ? TypeU64(state) : ImageScalarType(state, image.numeric_class);
 	return state.builder.Type(spv::OpTypeImage, scalar_type,
 	                          info.spirv_dimension, image.depth_compare ? 1u : 0u, info.arrayed,
 	                          info.multisampled, sampled, format);

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

@@ -19,8 +19,8 @@ uint32_t Vector(EmitterState& s, uint32_t type, std::span<const uint32_t> compon
 
 uint32_t NodeWord(EmitterState& s, uint32_t block, uint32_t word) {
 	const auto offset = Binary(s, spv::OpShiftLeftLogical, TypeU32(s), word, ConstantU32(s, 2));
-	const auto address = Binary(s, spv::OpIAdd, TypeScalarU64(s), block,
-	                            Unary(s, spv::OpUConvert, TypeScalarU64(s), offset));
+	const auto address =
+	    Binary(s, spv::OpIAdd, TypeU64(s), block, Unary(s, spv::OpUConvert, TypeU64(s), offset));
 	const auto pointer = Unary(s, spv::OpConvertUToPtr, TypePhysicalU32Pointer(s), address);
 	const auto result = s.builder.AllocateId();
 	s.builder.AddFunction(spv::OpLoad, TypeU32(s), result, pointer, spv::MemoryAccessAlignedMask, 4u);
@@ -224,7 +224,7 @@ uint32_t Boxes(EmitterState& s, uint32_t first, uint32_t second, uint32_t half,
 
 void DefineBvhIntersect(EmitterState& s) {
 	if (!s.requirements.bvh) return;
-	const auto u = TypeU32(s), f = TypeF32(s), b = TypeBool(s), wide = TypeScalarU64(s);
+	const auto u = TypeU32(s), f = TypeF32(s), b = TypeBool(s), wide = TypeU64(s);
 	const auto vec4 = TypeU32Vector(s, 4), vec3 = TypeF32Vector(s, 3);
 	const auto signature = s.builder.Type(spv::OpTypeFunction, vec4, vec4, wide, f, vec3, vec3, vec3);
 	s.bvh_intersect_function = s.builder.AllocateId();
@@ -260,22 +260,24 @@ void DefineBvhIntersect(EmitterState& s) {
 	const auto bary = Binary(s, spv::OpINotEqual, b, and_bits(words[3], 0x01000000), ConstantU32(s, 0));
 	const auto sort = Binary(s, spv::OpINotEqual, b, and_bits(words[1], 0x80000000), ConstantU32(s, 0));
 	const auto grow = and_bits(shr(words[1], 23), 0xff);
-	const auto base = Binary(s, spv::OpShiftLeftLogical, wide,
-	    DeviceAddressFromWords(s, words[0], and_bits(words[1], 0xff)), ConstantDeviceAddress(s, 8));
-	const auto index = Binary(s, spv::OpShiftRightLogical, wide, node, ConstantDeviceAddress(s, 3));
+	const auto base  = Binary(s, spv::OpShiftLeftLogical, wide,
+	                          PackU64(s, words[0], and_bits(words[1], 0xff)), ConstantU32(s, 8));
+	const auto index = Binary(s, spv::OpShiftRightLogical, wide, node, ConstantU32(s, 3));
 	const auto last_index = Binary(s, spv::OpIAdd, wide, index,
-	    Select(s, wide, full, ConstantDeviceAddress(s, 1), ConstantDeviceAddress(s, 0)));
+	                               Select(s, wide, full, ConstantU64(s, 1), ConstantU64(s, 0)));
 	const auto address = Binary(s, spv::OpIAdd, wide, base,
-	    Binary(s, spv::OpShiftLeftLogical, wide, index, ConstantDeviceAddress(s, 6)));
-	const auto last_address = Binary(s, spv::OpIAdd, wide, base,
-	    Binary(s, spv::OpShiftLeftLogical, wide, last_index, ConstantDeviceAddress(s, 6)));
+	                            Binary(s, spv::OpShiftLeftLogical, wide, index, ConstantU32(s, 6)));
+	const auto last_address =
+	    Binary(s, spv::OpIAdd, wide, base,
+	           Binary(s, spv::OpShiftLeftLogical, wide, last_index, ConstantU32(s, 6)));
 	auto valid = Binary(s, spv::OpIEqual, b, shr(words[3], 28), ConstantU32(s, 8));
 	const auto require = [&](uint32_t condition) {
 		valid = Binary(s, spv::OpLogicalAnd, b, valid, condition);
 	};
 	require(Binary(s, spv::OpULessThanEqual, b, kind, ConstantU32(s, 5)));
-	require(Binary(s, spv::OpULessThanEqual, b, last_index, DeviceAddressFromWords(s, words[2], and_bits(words[3], 0x3ff))));
-	require(Binary(s, spv::OpULessThan, b, last_address, ConstantDeviceAddress(s, uint64_t{1} << 48)));
+	require(Binary(s, spv::OpULessThanEqual, b, last_index,
+	               PackU64(s, words[2], and_bits(words[3], 0x3ff))));
+	require(Binary(s, spv::OpULessThan, b, last_address, ConstantU64(s, uint64_t {1} << 48)));
 	require(Unary(s, spv::OpLogicalNot, b, Unary(s, spv::OpIsNan, b, extent)));
 	for (uint32_t i = 0; i < 3; ++i) {
 		for (const auto value: {origin[i], direction[i], inverse[i]})
@@ -292,13 +294,13 @@ void DefineBvhIntersect(EmitterState& s) {
 	const auto invalid = Vector(s, vec4, invalid_words);
 	const auto result = EmitValueOrDefaultIfCondition(s, valid, vec4, invalid, [&] {
 		const auto first = GetBdaPointer(s, address);
-		const auto second = EmitValueOrDefaultIfCondition(s, full, wide, ConstantDeviceAddress(s, 0), [&] {
-			return GetBdaPointer(s, Binary(s, spv::OpIAdd, wide, address, ConstantDeviceAddress(s, 64)));
+		const auto second  = EmitValueOrDefaultIfCondition(s, full, wide, ConstantU64(s, 0), [&] {
+			return GetBdaPointer(s, Binary(s, spv::OpIAdd, wide, address, ConstantU64(s, 64)));
 		});
-		const auto present = Binary(s, spv::OpLogicalAnd, b,
-		    Binary(s, spv::OpINotEqual, b, first, ConstantDeviceAddress(s, 0)),
+		const auto present = Binary(
+		    s, spv::OpLogicalAnd, b, Binary(s, spv::OpINotEqual, b, first, ConstantU64(s, 0)),
 		    Binary(s, spv::OpLogicalOr, b, Unary(s, spv::OpLogicalNot, b, full),
-		        Binary(s, spv::OpINotEqual, b, second, ConstantDeviceAddress(s, 0))));
+		           Binary(s, spv::OpINotEqual, b, second, ConstantU64(s, 0))));
 		return EmitValueOrDefaultIfCondition(s, present, vec4, invalid, [&] {
 			const auto tri_label = s.builder.AllocateId(), box_label = s.builder.AllocateId();
 			const auto merge = s.builder.AllocateId();
@@ -336,9 +338,8 @@ uint32_t EmitBvhIntersect(ValueEmitContext& ctx, const IR::Inst& inst) {
 		return Vector(s, TypeF32Vector(s, 3), values);
 	};
 	const auto descriptor = ctx.Arg(inst, 0);
-	const auto node = node_words == 2u
-	                      ? DeviceAddressFromWords(s, ctx.Arg(*ray, 0), ctx.Arg(*ray, 1))
-	                      : Unary(s, spv::OpUConvert, TypeScalarU64(s), ctx.Arg(*ray, 0));
+	const auto node   = node_words == 2u ? PackU64(s, ctx.Arg(*ray, 0), ctx.Arg(*ray, 1))
+	                                     : Unary(s, spv::OpUConvert, TypeU64(s), ctx.Arg(*ray, 0));
 	const auto extent = Unary(s, spv::OpBitcast, TypeF32(s), ctx.Arg(*ray, node_words));
 	const auto origin = vector(node_words + 1u), direction = vector(node_words + 4u),
 	           inverse = vector(node_words + 7u);

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

@@ -42,8 +42,9 @@ uint32_t EmitBuiltinU32(EmitterState& state, IR::StageInputKind kind, uint32_t c
 		                  EmitBinaryU32(state, spv::OpIMul, group, ConstantU32(state, size)));
 	}
 	const bool centroid = kind == IR::StageInputKind::BaryCoordSmoothCentroid;
+	const bool sample = kind == IR::StageInputKind::BaryCoordSmoothSample;
 	const auto variable = InputVariableForKind(
-	    state, centroid ? IR::StageInputKind::BaryCoordSmooth : kind);
+	    state, centroid || sample ? IR::StageInputKind::BaryCoordSmooth : kind);
 	if (variable == 0) {
 		return ConstantU32(state, 0);
 	}
@@ -76,15 +77,24 @@ uint32_t EmitBuiltinU32(EmitterState& state, IR::StageInputKind kind, uint32_t c
 		state.builder.AddFunction(spv::OpBitcast, TypeU32(state), bits, value);
 		return bits;
 	}
-	if (centroid || kind == IR::StageInputKind::BaryCoordSmooth ||
+	if (centroid || sample || kind == IR::StageInputKind::BaryCoordSmooth ||
 	    kind == IR::StageInputKind::BaryCoordNoPerspective) {
 		const auto value   = state.builder.AllocateId();
 		const auto bits    = state.builder.AllocateId();
-		if (centroid) {
+		if (centroid || sample) {
 			const auto coordinates = state.builder.AllocateId();
 			state.builder.RequireCapability(spv::CapabilityInterpolationFunction);
-			state.builder.AddFunction(spv::OpExtInst, TypeF32Vector(state, 3), coordinates,
-			                          GlslStd450(state), GLSLstd450InterpolateAtCentroid, variable);
+			if (sample) {
+				const auto sample_id = state.builder.AllocateId();
+				state.builder.AddFunction(spv::OpLoad, TypeI32(state), sample_id,
+				                          InputVariableForKind(state, IR::StageInputKind::SampleId));
+				state.builder.AddFunction(spv::OpExtInst, TypeF32Vector(state, 3), coordinates,
+				                          GlslStd450(state), GLSLstd450InterpolateAtSample,
+				                          variable, sample_id);
+			} else {
+				state.builder.AddFunction(spv::OpExtInst, TypeF32Vector(state, 3), coordinates,
+				                          GlslStd450(state), GLSLstd450InterpolateAtCentroid, variable);
+			}
 			state.builder.AddFunction(spv::OpCompositeExtract, TypeF32(state), value,
 			                          coordinates, component + 1u);
 		} else {
@@ -252,8 +262,12 @@ uint32_t ExportVector(ValueEmitContext& ctx, uint32_t data, const IR::ExportInfo
                       bool uint_output) {
 	auto& state = ctx.state;
 	if (exp.compr && !uint_output) {
-		const auto unpack =
-		    MrtOutputMode(state, exp) == 5u ? GLSLstd450UnpackUnorm2x16 : GLSLstd450UnpackHalf2x16;
+		auto unpack = GLSLstd450UnpackHalf2x16;
+		switch (MrtOutputMode(state, exp)) {
+			case 5: unpack = GLSLstd450UnpackUnorm2x16; break;
+			case 6: unpack = GLSLstd450UnpackSnorm2x16; break;
+			default: break;
+		}
 		uint32_t f32[4] = {ConstantF32(state, 0), ConstantF32(state, 0), ConstantF32(state, 0),
 		                   ConstantF32(state, 0x3f800000u)};
 		for (uint32_t pair = 0; pair < 2u; pair++) {
@@ -430,16 +444,16 @@ uint32_t EmitWqmU64(EmitterState& state, uint32_t value) {
 	const auto quad_bits   = state.builder.AllocateId();
 	const auto result      = state.builder.AllocateId();
 	state.builder.AddFunction(spv::OpShiftRightLogical, TypeU64(state), shifted_one, value,
-	                          ConstantU64(state, 0x0000000100000001ull));
+	                          ConstantU32(state, 1));
 	state.builder.AddFunction(spv::OpBitwiseOr, TypeU64(state), merged_one, value, shifted_one);
 	state.builder.AddFunction(spv::OpShiftRightLogical, TypeU64(state), shifted_two, merged_one,
-	                          ConstantU64(state, 0x0000000200000002ull));
+	                          ConstantU32(state, 2));
 	state.builder.AddFunction(spv::OpBitwiseOr, TypeU64(state), merged_two, merged_one,
 	                          shifted_two);
 	state.builder.AddFunction(spv::OpBitwiseAnd, TypeU64(state), quad_bits, merged_two,
 	                          ConstantU64(state, 0x1111111111111111ull));
 	state.builder.AddFunction(spv::OpIMul, TypeU64(state), result, quad_bits,
-	                          ConstantU64(state, 0x0000000f0000000full));
+	                          ConstantU64(state, 15));
 	return result;
 }
 
@@ -699,26 +713,18 @@ uint32_t EmitDppUpdateU32(ValueEmitContext& ctx, const IR::Inst& inst) {
 }
 
 uint32_t EmitConditionRef(ValueEmitContext& ctx, const IR::Inst& inst) {
-	if (ctx.other_half == nullptr) return ctx.Arg(inst, 0);
-	// A native scalar branch makes one decision for both emulated wave halves.
-	if (ctx.half != 0) return ctx.other_half->Def(IR::Value(&inst));
+	// Scalar branches include lanes disabled by EXEC, which the branch body may restore.
+	if (ctx.other_half != nullptr && ctx.half != 0) return ctx.other_half->Def(IR::Value(&inst));
 	const auto kind = inst.Flags<CFG::BranchCondition>();
 	if (kind == CFG::BranchCondition::ScalarInstruction) return ctx.Arg(inst, 0);
-	const auto ballot = ctx.Ballot(inst.Arg(0));
-	const auto low = ctx.state.builder.AllocateId();
-	const auto high = ctx.state.builder.AllocateId();
-	const auto combined = ctx.state.builder.AllocateId();
-	const auto result = ctx.state.builder.AllocateId();
-	ctx.state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(ctx.state), low, ballot, 0);
-	ctx.state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(ctx.state), high, ballot, 1);
 	const bool zero = kind == CFG::BranchCondition::ExecZero ||
 	                  kind == CFG::BranchCondition::VccZero || kind == CFG::BranchCondition::SccZero;
-	ctx.state.builder.AddFunction(zero ? spv::OpBitwiseAnd : spv::OpBitwiseOr,
-	                              TypeU32(ctx.state), combined, low, high);
-	ctx.state.builder.AddFunction(zero ? spv::OpIEqual : spv::OpINotEqual,
-	                              TypeBool(ctx.state), result, combined,
-	                              ConstantU32(ctx.state, zero ? ~0u : 0u));
-	return result;
+	const auto ballot = ctx.Ballot(inst.Arg(0));
+	const auto reference = zero ? ctx.Ballot(IR::Value(true))
+	                            : ConstantU32CompositeZero(ctx.state, 4);
+	const auto compare = Binary(ctx.state, zero ? spv::OpIEqual : spv::OpINotEqual,
+	                            TypeBoolVector(ctx.state, 4), ballot, reference);
+	return Unary(ctx.state, zero ? spv::OpAll : spv::OpAny, TypeBool(ctx.state), compare);
 }
 
 uint32_t EmitBallot(ValueEmitContext& ctx, IR::Value predicate) {

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

@@ -849,74 +849,76 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 			return;
 		}
 		const auto key = ctx.Def(handle->Arg(0));
-		if (state.flattened_srt_variable == 0 || image.indirect_search_iterations == 0u ||
+		if (state.flattened_srt_variable == 0 ||
 		    image.indirect_resources.size() < 2u) {
 			ctx.Fail(inst, "has no indirect image runtime mapping");
 			return;
 		}
 		const auto selected = EmitIndirectResourceIndex(
 		    state, key, image.indirect_mapping_offset, image.indirect_search_iterations, 0u);
-		const auto kind = *IR::DescriptorBindingForImage(image);
-		const auto& binding = *IR::FindBinding(state.program.bindings, kind);
-		uint32_t first_child = 0;
-		bool homogeneous = true;
+		struct SampleRun {
+			uint32_t first;
+			uint32_t count;
+			uint32_t resource;
+			uint32_t slot_bias;
+		};
+		std::vector<SampleRun> runs;
 		for (uint32_t ordinal = 0; ordinal < image.indirect_resources.size(); ++ordinal) {
 			const auto resource = image.indirect_resources[ordinal];
 			const auto& candidate = state.program.info.images[resource];
-			if (candidate.dimension != image.dimension || candidate.cube != image.cube ||
-			    IR::DescriptorBindingForImage(candidate) != kind ||
-			    candidate.mip_count != 1u) {
-				homogeneous = false;
-				break;
+			const auto kind = *IR::DescriptorBindingForImage(candidate);
+			const auto& binding = *IR::FindBinding(state.program.bindings, kind);
+			if (!runs.empty()) {
+				auto& run = runs.back();
+				const auto& first = state.program.info.images[run.resource];
+				const auto next_slot = run.slot_bias + ordinal;
+				if (candidate.dimension == first.dimension && candidate.cube == first.cube &&
+				    IR::DescriptorBindingForImage(first) == kind &&
+				    candidate.mip_count == 1u && first.mip_count == 1u &&
+				    (ordinal == 1u || (next_slot < binding.resources.size() &&
+				                      binding.resources[next_slot] == resource))) {
+					// Native roots precede appended children; only the root may have a slot gap.
+					if (ordinal == 1u)
+						run.slot_bias = ResourceForDescriptor(state, kind, resource) - ordinal;
+					++run.count;
+					continue;
+				}
 			}
-			if (ordinal == 1u) first_child = ResourceForDescriptor(state, kind, resource);
-			const auto child_slot = first_child + ordinal - 1u;
-			if (ordinal > 1u && (child_slot >= binding.resources.size() ||
-			                     binding.resources[child_slot] != resource)) {
-				homogeneous = false;
-				break;
+			runs.push_back({ordinal, 1u, resource,
+			                ResourceForDescriptor(state, kind, resource) - ordinal});
+		}
+		const auto EmitRun = [&](uint32_t index) {
+			const auto& run = runs[index];
+			if (run.count == 1u) return EmitSample(run.resource);
+			auto slot = Binary(state, spv::OpIAdd, TypeU32(state), selected,
+			                   ConstantU32(state, run.slot_bias));
+			if (run.first == 0u) {
+				const auto kind = *IR::DescriptorBindingForImage(image);
+				const auto root_slot = ResourceForDescriptor(state, kind, mem.resource);
+				if (root_slot != run.slot_bias) {
+					const auto is_root = Binary(state, spv::OpIEqual, TypeBool(state), selected,
+					                            ConstantU32(state, 0u));
+					const auto root_or_child = state.builder.AllocateId();
+					state.builder.AddFunction(spv::OpSelect, TypeU32(state), root_or_child,
+					                          is_root, ConstantU32(state, root_slot), slot);
+					slot = root_or_child;
+				}
 			}
+			return EmitSample(run.resource, slot);
+		};
+		auto run_index = ConstantU32(state, 0u);
+		for (uint32_t index = 1; index < runs.size(); ++index) {
+			const auto at_or_after = Binary(state, spv::OpUGreaterThanEqual, TypeBool(state),
+			                               selected, ConstantU32(state, runs[index].first));
+			const auto next = state.builder.AllocateId();
+			state.builder.AddFunction(spv::OpSelect, TypeU32(state), next, at_or_after,
+			                          ConstantU32(state, index), run_index);
+			run_index = next;
 		}
-		if (homogeneous) {
-			const auto child_index = Binary(state, spv::OpIAdd, TypeU32(state), selected,
-			                                ConstantU32(state, first_child - 1u));
-			const auto is_root = Binary(state, spv::OpIEqual, TypeBool(state), selected,
-			                            ConstantU32(state, 0u));
-			const auto array_index = state.builder.AllocateId();
-			state.builder.AddFunction(spv::OpSelect, TypeU32(state), array_index, is_root,
-			                          ConstantU32(state, ResourceForDescriptor(state, kind, mem.resource)),
-			                          child_index);
-			const auto sample = EmitSample(mem.resource, array_index);
-			const auto result = dref ? sample : UnpackImageTexel(ctx, mem, sample);
-			ctx.Define(inst, ResultVector(ctx, result, numeric_class, dref, mem));
-			return;
-		}
-		const auto            default_label = state.builder.AllocateId();
-		const auto            merge_label   = state.builder.AllocateId();
-		std::vector<uint32_t> labels(image.indirect_resources.size() - 1u);
-		std::vector<uint32_t> switch_words {spv::OpSwitch, selected, default_label};
-		for (uint32_t candidate = 1; candidate < image.indirect_resources.size(); candidate++) {
-			labels[candidate - 1u] = state.builder.AllocateId();
-			switch_words.push_back(candidate);
-			switch_words.push_back(labels[candidate - 1u]);
-		}
-		state.builder.AddFunction(spv::OpSelectionMerge, merge_label,
-		                          spv::SelectionControlMaskNone);
-		state.builder.AddFunction(switch_words);
-		std::vector<uint32_t> phi_words {spv::OpPhi, result_type, state.builder.AllocateId()};
-		EmitLabel(state, default_label);
-		phi_words.push_back(EmitSample(image.indirect_resources[0]));
-		phi_words.push_back(default_label);
-		state.builder.AddFunction(spv::OpBranch, merge_label);
-		for (uint32_t candidate = 1; candidate < image.indirect_resources.size(); candidate++) {
-			EmitLabel(state, labels[candidate - 1u]);
-			phi_words.push_back(EmitSample(image.indirect_resources[candidate]));
-			phi_words.push_back(labels[candidate - 1u]);
-			state.builder.AddFunction(spv::OpBranch, merge_label);
-		}
-		EmitLabel(state, merge_label);
-		state.builder.AddFunction(phi_words);
-		auto result = phi_words[2];
+		auto result = runs.size() == 1u
+		                  ? EmitRun(0u)
+		                  : EmitIndexSwitch(state, run_index, static_cast<uint32_t>(runs.size()),
+		                                    result_type, EmitRun);
 		if (!dref) {
 			result = UnpackImageTexel(ctx, mem, result);
 		}
@@ -930,9 +932,9 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 		ctx.Define(inst, EmitValueOrDefaultIfCondition(
 		                     state, ctx.Arg(inst, inst.NumArgs() - 1), result_type, zero, [&]() {
 			           const auto pointer      = state.builder.AllocateId();
-			           const auto pointer_type = state.builder.Type(
-			               spv::OpTypePointer, spv::StorageClassImage,
-			               image.atomic64 ? TypeScalarU64(state) : TypeU32(state));
+			           const auto pointer_type =
+			               state.builder.Type(spv::OpTypePointer, spv::StorageClassImage,
+			                                  image.atomic64 ? TypeU64(state) : TypeU32(state));
 			           state.builder.AddFunction(spv::OpImageTexelPointer, pointer_type, pointer,
 			                                     ImageDescriptorPointer(state, mem.resource),
 			                                     CoordU32(ctx, mem, *address, dimension),

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

@@ -34,16 +34,16 @@ uint32_t              EmitConvertF32F64(EmitterState& state, uint32_t arg0);
 uint32_t              EmitConvertF64F32(EmitterState& state, uint32_t arg0);
 EMIT_NATIVE(ConvertF32U32, OpConvertUToF, F32, uint32_t)
 EMIT_NATIVE(ConvertF64U32, OpConvertUToF, F64, uint32_t)
-EMIT_NATIVE(CompositeConstructU64, OpCompositeConstruct, U64, uint32_t, uint32_t)
+uint32_t EmitCompositeConstructU64(ValueEmitContext& ctx, uint32_t arg0, IR::Value arg1);
 EMIT_NATIVE(CompositeConstructU32x2, OpCompositeConstruct, U32x2, uint32_t, uint32_t)
 EMIT_NATIVE(CompositeConstructU32x3, OpCompositeConstruct, U32x3, uint32_t, uint32_t, uint32_t)
 EMIT_NATIVE(CompositeConstructF32x2, OpCompositeConstruct, F32x2, uint32_t, uint32_t)
 EMIT_NATIVE(CompositeConstructU32x4, OpCompositeConstruct, U32x4, uint32_t, uint32_t, uint32_t,
             uint32_t)
 uint32_t              EmitCompositeExtractU64(EmitterState& state, uint32_t arg0, IR::Value arg1);
-inline constexpr auto EmitCompositeExtractU32x2 = EmitCompositeExtractU64;
-inline constexpr auto EmitCompositeExtractU32x3 = EmitCompositeExtractU64;
-inline constexpr auto EmitCompositeExtractU32x4 = EmitCompositeExtractU64;
+uint32_t              EmitCompositeExtractU32x2(EmitterState& state, uint32_t arg0, IR::Value arg1);
+inline constexpr auto EmitCompositeExtractU32x3 = EmitCompositeExtractU32x2;
+inline constexpr auto EmitCompositeExtractU32x4 = EmitCompositeExtractU32x2;
 inline constexpr auto EmitPackHalf2x16 = EmitGlsl<GLSLstd450PackHalf2x16, IR::Type::U32, uint32_t>;
 inline constexpr auto EmitPackSnorm2x16 =
     EmitGlsl<GLSLstd450PackSnorm2x16, IR::Type::U32, uint32_t>;
@@ -63,9 +63,9 @@ 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)
 EMIT_NATIVE(UDiv32, OpUDiv, U32, uint32_t, uint32_t)
-uint32_t EmitIAdd64(EmitterState& state, uint32_t arg0, uint32_t arg1);
-uint32_t EmitISub64(EmitterState& state, uint32_t arg0, uint32_t arg1);
-uint32_t EmitIMul64(EmitterState& state, uint32_t arg0, uint32_t arg1);
+EMIT_NATIVE(IAdd64, OpIAdd, U64, uint32_t, uint32_t)
+EMIT_NATIVE(ISub64, OpISub, U64, uint32_t, uint32_t)
+EMIT_NATIVE(IMul64, OpIMul, U64, uint32_t, uint32_t)
 EMIT_NATIVE(IAddCarry32, OpIAddCarry, U32x2, uint32_t, uint32_t)
 uint32_t EmitSMulHi(EmitterState& state, uint32_t arg0, uint32_t arg1);
 uint32_t EmitUMulHi(EmitterState& state, uint32_t arg0, uint32_t arg1);
@@ -73,14 +73,26 @@ uint32_t EmitIAbs32(EmitterState& state, uint32_t arg0);
 EMIT_NATIVE(ShiftLeftLogical32, OpShiftLeftLogical, U32, uint32_t, uint32_t)
 EMIT_NATIVE(ShiftRightLogical32, OpShiftRightLogical, U32, uint32_t, uint32_t)
 EMIT_NATIVE(ShiftRightArithmetic32, OpShiftRightArithmetic, U32, uint32_t, uint32_t)
-uint32_t EmitShiftLeftLogical64(ValueEmitContext& ctx, uint32_t arg0, IR::Value arg1);
-uint32_t EmitShiftRightLogical64(ValueEmitContext& ctx, uint32_t arg0, IR::Value arg1);
-uint32_t EmitShiftRightArithmetic64(ValueEmitContext& ctx, uint32_t arg0, IR::Value arg1);
+template <spv::Op opcode>
+uint32_t EmitShift64(ValueEmitContext& ctx, uint32_t value, IR::Value shift) {
+	shift = shift.Resolve();
+	if (shift.IsImmediate()) {
+		const auto amount = shift.U32() & 63u;
+		return amount == 0u ? value
+		                    : Binary(ctx.state, opcode, TypeU64(ctx.state), value,
+		                             ConstantU32(ctx.state, amount));
+	}
+	return Binary(ctx.state, opcode, TypeU64(ctx.state), value,
+	              EmitAndConstant(ctx.state, ctx.Def(shift), 63u));
+}
+inline constexpr auto EmitShiftLeftLogical64     = EmitShift64<spv::OpShiftLeftLogical>;
+inline constexpr auto EmitShiftRightLogical64    = EmitShift64<spv::OpShiftRightLogical>;
+inline constexpr auto EmitShiftRightArithmetic64 = EmitShift64<spv::OpShiftRightArithmetic>;
 EMIT_NATIVE(BitwiseAnd32, OpBitwiseAnd, U32, uint32_t, uint32_t)
 EMIT_NATIVE(BitwiseOr32, OpBitwiseOr, U32, uint32_t, uint32_t)
 EMIT_NATIVE(BitwiseXor32, OpBitwiseXor, U32, uint32_t, uint32_t)
 EMIT_NATIVE(BitwiseNot32, OpNot, U32, uint32_t)
-uint32_t EmitBitwiseAnd64(EmitterState& state, uint32_t arg0, uint32_t arg1);
+EMIT_NATIVE(BitwiseAnd64, OpBitwiseAnd, U64, uint32_t, uint32_t)
 EMIT_NATIVE(BitReverse32, OpBitReverse, U32, uint32_t)
 EMIT_NATIVE(BitCount32, OpBitCount, U32, uint32_t)
 uint32_t EmitBitCount64(EmitterState& state, uint32_t arg0);
@@ -107,14 +119,14 @@ EMIT_NATIVE(UGreaterThan32, OpUGreaterThan, U1, uint32_t, uint32_t)
 EMIT_NATIVE(INotEqual32, OpINotEqual, U1, uint32_t, uint32_t)
 EMIT_NATIVE(SGreaterThanEqual32, OpSGreaterThanEqual, U1, uint32_t, uint32_t)
 EMIT_NATIVE(UGreaterThanEqual32, OpUGreaterThanEqual, U1, uint32_t, uint32_t)
-uint32_t EmitIEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1);
-uint32_t EmitINotEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1);
-uint32_t EmitULessThan64(EmitterState& state, uint32_t arg0, uint32_t arg1);
-uint32_t EmitSLessThan64(EmitterState& state, uint32_t arg0, uint32_t arg1);
-uint32_t EmitSLessThanEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1);
-uint32_t EmitULessThanEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1);
-uint32_t EmitUGreaterThanEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1);
-uint32_t EmitUGreaterThan64(EmitterState& state, uint32_t arg0, uint32_t arg1);
+EMIT_NATIVE(IEqual64, OpIEqual, U1, uint32_t, uint32_t)
+EMIT_NATIVE(INotEqual64, OpINotEqual, U1, uint32_t, uint32_t)
+EMIT_NATIVE(ULessThan64, OpULessThan, U1, uint32_t, uint32_t)
+EMIT_NATIVE(SLessThan64, OpSLessThan, U1, uint32_t, uint32_t)
+EMIT_NATIVE(SLessThanEqual64, OpSLessThanEqual, U1, uint32_t, uint32_t)
+EMIT_NATIVE(ULessThanEqual64, OpULessThanEqual, U1, uint32_t, uint32_t)
+EMIT_NATIVE(UGreaterThanEqual64, OpUGreaterThanEqual, U1, uint32_t, uint32_t)
+EMIT_NATIVE(UGreaterThan64, OpUGreaterThan, U1, uint32_t, uint32_t)
 EMIT_NATIVE(LogicalOr, OpLogicalOr, U1, uint32_t, uint32_t)
 EMIT_NATIVE(LogicalAnd, OpLogicalAnd, U1, uint32_t, uint32_t)
 EMIT_NATIVE(LogicalXor, OpLogicalNotEqual, U1, uint32_t, uint32_t)
@@ -152,6 +164,7 @@ inline constexpr auto EmitFPOrdGreaterThanEqual32 = EmitFloatCompare32<spv::OpFO
 inline constexpr auto EmitFPUnordGreaterThanEqual32 = EmitFloatCompare32<spv::OpFUnordGreaterThanEqual>;
 uint32_t              EmitFPIsNan32(EmitterState& state, uint32_t arg0);
 inline constexpr auto EmitFPCmpClass32 = EmitClassMaskF32;
+inline constexpr auto EmitFPCmpClass16 = EmitClassMaskF16;
 EMIT_NATIVE(FPAdd32, OpFAdd, F32, uint32_t, uint32_t)
 EMIT_NATIVE(FPSub32, OpFSub, F32, uint32_t, uint32_t)
 EMIT_NATIVE(FPMul32, OpFMul, F32, uint32_t, uint32_t)
@@ -160,8 +173,8 @@ EMIT_NATIVE(FPMul64, OpFMul, F64, uint32_t, uint32_t)
 inline constexpr auto EmitFPFma64 =
     EmitGlsl<GLSLstd450Fma, IR::Type::F64, uint32_t, uint32_t, uint32_t>;
 uint32_t              EmitFPRecip64(EmitterState& state, uint32_t arg0);
-inline constexpr auto EmitFPFma32 =
-    EmitGlsl<GLSLstd450Fma, IR::Type::F32, uint32_t, uint32_t, uint32_t>;
+uint32_t EmitFPFma32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32_t arg2);
+uint32_t EmitFPMad32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32_t arg2);
 uint32_t EmitFPMin32(EmitterState& state, uint32_t arg0, uint32_t arg1);
 uint32_t EmitFPMax32(EmitterState& state, uint32_t arg0, uint32_t arg1);
 uint32_t EmitFPMin64(EmitterState& state, uint32_t arg0, uint32_t arg1);
@@ -321,15 +334,21 @@ inline constexpr auto EmitImageSampleRaw       = EmitImage;
 inline constexpr auto EmitImageGatherRaw       = EmitImage;
 inline constexpr auto EmitImageAtomicCompareSwap32 = EmitImage;
 inline constexpr auto EmitImageAtomicSwap32    = EmitImage;
+inline constexpr auto EmitImageAtomicSwap64    = EmitImage;
 inline constexpr auto EmitImageAtomicIAdd32    = EmitImage;
+inline constexpr auto EmitImageAtomicIAdd64    = EmitImage;
 inline constexpr auto EmitImageAtomicSMin32    = EmitImage;
 inline constexpr auto EmitImageAtomicUMin32    = EmitImage;
+inline constexpr auto EmitImageAtomicUMin64    = EmitImage;
 inline constexpr auto EmitImageAtomicSMax32    = EmitImage;
 inline constexpr auto EmitImageAtomicUMax32    = EmitImage;
 inline constexpr auto EmitImageAtomicUMax64    = EmitImage;
 inline constexpr auto EmitImageAtomicAnd32     = EmitImage;
+inline constexpr auto EmitImageAtomicAnd64     = EmitImage;
 inline constexpr auto EmitImageAtomicOr32      = EmitImage;
+inline constexpr auto EmitImageAtomicOr64      = EmitImage;
 inline constexpr auto EmitImageAtomicXor32     = EmitImage;
+inline constexpr auto EmitImageAtomicXor64     = EmitImage;
 inline constexpr auto EmitImageAtomicFMin32    = EmitImage;
 inline constexpr auto EmitImageAtomicFMax32    = EmitImage;
 void                  EmitUnreachable(ValueEmitContext& ctx, const IR::Inst& inst);

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

@@ -98,7 +98,7 @@ struct EmitterState {
 	uint32_t                                        void_type = 0;
 	uint32_t                                        bool_type = 0;
 	uint32_t                                        u32_type = 0;
-	uint32_t                                        native_u64_type = 0;
+	uint32_t                                        u64_type = 0;
 	uint32_t                                        i32_type = 0;
 	uint32_t                                        f32_type = 0;
 	uint32_t                                        f64_type = 0;
@@ -168,7 +168,6 @@ uint32_t TypeBool(EmitterState& state);
 uint32_t TypeBoolVector(EmitterState& state, uint32_t components);
 uint32_t TypeU32(EmitterState& state);
 uint32_t TypeU64(EmitterState& state);
-uint32_t TypeScalarU64(EmitterState& state);
 uint32_t TypeU32Pair(EmitterState& state);
 uint32_t TypeI32(EmitterState& state);
 uint32_t TypeI32Pair(EmitterState& state);
@@ -221,6 +220,12 @@ inline uint32_t Unary(EmitterState& state, spv::Op opcode, uint32_t type, uint32
 	return result;
 }
 
+inline uint32_t PackU64(EmitterState& state, uint32_t low, uint32_t high) {
+	const auto pair = state.builder.AllocateId();
+	state.builder.AddFunction(spv::OpCompositeConstruct, TypeU32Vector(state, 2), pair, low, high);
+	return Unary(state, spv::OpBitcast, TypeU64(state), pair);
+}
+
 inline uint32_t Binary(EmitterState& state, spv::Op opcode, uint32_t type, uint32_t lhs,
                        uint32_t rhs) {
 	const auto result = state.builder.AllocateId();
@@ -463,6 +468,9 @@ void EmitAtomicMemoryBarrier(EmitterState& state, IR::ResourceKind kind);
 uint32_t EmitFloatAtomicReplacement(EmitterState& state, uint32_t old, uint32_t source,
                                     bool max_value);
 
+uint32_t EmitDsFloatAtomicReplacement(EmitterState& state, uint32_t old, uint32_t source,
+                                      bool max_value);
+
 uint32_t EmitDsSwizzleTargetLane(EmitterState& state, uint32_t subid, uint32_t control);
 
 inline constexpr auto EmitSelectValueU32 =
@@ -507,6 +515,7 @@ uint32_t EmitClassMaskBitMatch(EmitterState& state, uint32_t mask, uint32_t bit,
                                uint32_t class_match);
 
 uint32_t EmitClassMaskF32(EmitterState& state, uint32_t value, uint32_t mask);
+uint32_t EmitClassMaskF16(EmitterState& state, uint32_t bits, uint32_t mask);
 
 uint32_t EmitMinMaxF32Value(EmitterState& state, uint32_t lhs, uint32_t rhs, bool max_value);
 
@@ -527,8 +536,6 @@ void EmitProgram(EmitterState& state);
 void DefineGetBdaPointer(EmitterState& state);
 void DefineBvhIntersect(EmitterState& state);
 uint32_t GetBdaPointer(EmitterState& state, uint32_t address);
-uint32_t ConstantDeviceAddress(EmitterState& state, uint64_t value);
-uint32_t DeviceAddressFromWords(EmitterState& state, uint32_t low, uint32_t high);
 
 // These templates accept local lambdas from several emitter translation units.
 template <typename Fn>

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

@@ -133,22 +133,22 @@ uint32_t GuestAddress(ValueEmitContext& ctx, const IR::Inst& inst, const IR::Mem
 	}
 	uint32_t address = 0;
 	if (mem.address_is_full) {
-		address = DeviceAddressFromWords(state, low, ctx.Arg(inst, 2));
+		address = PackU64(state, low, ctx.Arg(inst, 2));
 	} else {
 		const auto* handle = inst.Arg(0).Resolve().TryInstruction();
 		if (handle == nullptr || handle->GetOpcode() != IR::ValueOpcode::GetAddressResource ||
 		    handle->NumArgs() != 2) {
 			ctx.Fail(inst, "has no address base pair");
-			return ConstantDeviceAddress(state, 0);
+			return ConstantU64(state, 0);
 		}
 		auto base_low = ctx.Arg(*handle, 0);
 		if (mem.kind == IR::ResourceKind::ScalarAddress) {
 			base_low = Binary(state, spv::OpBitwiseAnd, TypeU32(state), base_low,
 			                  ConstantU32(state, ~3u));
 		}
-		const auto base = DeviceAddressFromWords(state, base_low, ctx.Arg(*handle, 1));
-		address         = Binary(state, spv::OpIAdd, TypeScalarU64(state), base,
-		                         Unary(state, spv::OpUConvert, TypeScalarU64(state), low));
+		const auto base = PackU64(state, base_low, ctx.Arg(*handle, 1));
+		address         = Binary(state, spv::OpIAdd, TypeU64(state), base,
+		                         Unary(state, spv::OpUConvert, TypeU64(state), low));
 	}
 	auto immediate = static_cast<int32_t>(mem.offset);
 	if (mem.kind == IR::ResourceKind::ScalarAddress) {
@@ -156,9 +156,8 @@ uint32_t GuestAddress(ValueEmitContext& ctx, const IR::Inst& inst, const IR::Mem
 	}
 	return immediate == 0
 	           ? address
-	           : Binary(state, spv::OpIAdd, TypeScalarU64(state), address,
-	                    ConstantDeviceAddress(
-	                        state, static_cast<uint64_t>(static_cast<int64_t>(immediate))));
+	           : Binary(state, spv::OpIAdd, TypeU64(state), address,
+	                    ConstantU64(state, static_cast<uint64_t>(static_cast<int64_t>(immediate))));
 }
 
 uint32_t FaultElementPointer(EmitterState& state, uint32_t index) {
@@ -181,29 +180,46 @@ void RecordBdaFault(EmitterState& state, uint32_t page) {
 	                          Binary(state, spv::OpBitwiseOr, TypeU32(state), value, bit));
 }
 
-uint32_t LoadBdaDword(ValueEmitContext& ctx, uint32_t address) {
+uint32_t LoadBdaDword(ValueEmitContext& ctx, uint32_t address, bool coherent = false) {
 	auto&      state   = ctx.state;
 	const auto bda     = GetBdaPointer(state, address);
 	const auto present =
-	    Binary(state, spv::OpINotEqual, TypeBool(state), bda, ConstantDeviceAddress(state, 0));
+	    Binary(state, spv::OpINotEqual, TypeBool(state), bda, ConstantU64(state, 0));
 	return EmitValueOrZeroIfCondition(state, present, [&]() {
-		const auto pointer = state.builder.AllocateId();
-		state.builder.AddFunction(spv::OpConvertUToPtr, TypePhysicalU32Pointer(state), pointer,
-		                          bda);
+		auto pointer = state.builder.AllocateId();
+		if (coherent) {
+			const auto block = state.builder.DecoratedType(
+			    spv::OpTypeStruct, {{spv::OpMemberDecorate, {0, spv::DecorationOffset, 0}},
+			                        {spv::OpMemberDecorate, {0, spv::DecorationCoherent}},
+			                        {spv::OpDecorate, {spv::DecorationBlock}}}, TypeU32(state));
+			state.builder.AddFunction(
+			    spv::OpConvertUToPtr, TypePointer(state, spv::StorageClassPhysicalStorageBuffer, block),
+			    pointer, bda);
+			const auto base = pointer;
+			pointer = state.builder.AllocateId();
+			state.builder.AddFunction(spv::OpAccessChain, TypePhysicalU32Pointer(state), pointer,
+			                          base, ConstantU32(state, 0));
+		} else {
+			state.builder.AddFunction(spv::OpConvertUToPtr, TypePhysicalU32Pointer(state), pointer,
+			                          bda);
+		}
 		const auto         value     = state.builder.AllocateId();
 		constexpr uint32_t alignment = sizeof(uint32_t);
 		state.builder.AddFunction(spv::OpLoad, TypeU32(state), value, pointer,
-		                          spv::MemoryAccessAlignedMask, alignment);
+		                          spv::MemoryAccessAlignedMask |
+		                              (coherent ? spv::MemoryAccessVolatileMask : spv::MemoryAccessMaskNone),
+		                          alignment);
 		return value;
 	});
 }
 
-uint32_t LoadBda(ValueEmitContext& ctx, uint32_t address, uint32_t active, uint32_t bits) {
+uint32_t LoadBda(ValueEmitContext& ctx, uint32_t address, uint32_t active, uint32_t bits,
+                 bool coherent = false) {
 	auto& state = ctx.state;
 	return EmitValueOrZeroIfCondition(state, active, [&]() {
-		const auto aligned = Binary(state, spv::OpBitwiseAnd, TypeScalarU64(state), address,
-		                            ConstantDeviceAddress(state, ~uint64_t {3}));
-		const auto first   = LoadBdaDword(ctx, aligned);
+		const auto aligned = Binary(state, spv::OpBitwiseAnd, TypeU64(state), address,
+		                            ConstantU64(state, ~uint64_t {3}));
+		const auto first   = LoadBdaDword(ctx, aligned, coherent);
 		const auto byte =
 		    Binary(state, spv::OpBitwiseAnd, TypeU32(state),
 		           Unary(state, spv::OpUConvert, TypeU32(state), address), ConstantU32(state, 3));
@@ -212,8 +228,8 @@ uint32_t LoadBda(ValueEmitContext& ctx, uint32_t address, uint32_t active, uint3
 		               : Binary(state, bits == 16u ? spv::OpUGreaterThan : spv::OpINotEqual,
 		                        TypeBool(state), byte, ConstantU32(state, bits == 16u ? 2u : 0u));
 		const auto second = EmitValueOrZeroIfCondition(state, crosses, [&]() {
-			return LoadBdaDword(ctx, Binary(state, spv::OpIAdd, TypeScalarU64(state), aligned,
-			                                ConstantDeviceAddress(state, sizeof(uint32_t))));
+			return LoadBdaDword(ctx, Binary(state, spv::OpIAdd, TypeU64(state), aligned,
+			                                ConstantU64(state, sizeof(uint32_t))), coherent);
 		});
 		const auto shift =
 		    Binary(state, spv::OpShiftLeftLogical, TypeU32(state), byte, ConstantU32(state, 3));
@@ -700,34 +716,67 @@ uint32_t FormattedLoad(ValueEmitContext& ctx, const IR::Inst& inst, const IR::Me
 			return LoadWordPrepared(ctx, inst, RebaseRawComponent(mem, 0u), resource);
 		}
 		const auto plan = PrepareFormattedMemory(ctx, inst, mem, resource, info);
-		return EmitValueOrDefaultIfCondition(
+		const auto value = EmitValueOrDefaultIfCondition(
 		    ctx.state, plan.in_bounds, TypeU32(ctx.state),
 		    FormattedOutOfBoundsValue(ctx, mem, plan, 1u),
 		    [&]() { return LoadFormattedInBounds(ctx, mem, plan, 0u); });
+		if (mem.data_bits != 16u) {
+			return value;
+		}
+		if (info.type == Format::ComponentType::Uint) {
+			return EmitUMin32(ctx.state, value, ConstantU32(ctx.state, 0xffffu));
+		}
+		if (info.type == Format::ComponentType::Sint) {
+			const auto lower = EmitSMax32(ctx.state, value, ConstantU32(ctx.state, 0xffff8000u));
+			return EmitSMin32(ctx.state, lower, ConstantU32(ctx.state, 0x7fffu));
+		}
+		return EmitConvertF16F32(ctx.state, EmitBitCastF32U32(ctx.state, value));
 	});
 }
 
 uint32_t EncodeFormattedStoreComponent(ValueEmitContext& ctx,
                                        const Format::BufferFormatInfo& info,
-                                       uint32_t component, uint32_t data) {
+                                       uint32_t component, uint32_t data, uint32_t data_bits) {
+	auto& state = ctx.state;
 	const auto bits = info.component_bits[component];
-	if (info.type == Format::ComponentType::Unorm) {
-		const auto value = EmitBitCastF32U32(ctx.state, data);
-		const auto clamped = EmitFPMin32(
-		    ctx.state, EmitFPMax32(ctx.state, value, ConstantF32Value(ctx.state, 0.0f)),
-		    ConstantF32Value(ctx.state, 1.0f));
-		const auto scaled = EmitFPMul32(
-		    ctx.state, clamped, ConstantF32Value(ctx.state, float((1u << bits) - 1u)));
-		return Unary(ctx.state, spv::OpConvertFToU, TypeU32(ctx.state),
-		             EmitFPRoundEven32(ctx.state, scaled));
+	if (data_bits == 16u) {
+		if (info.type == Format::ComponentType::Uint) {
+			data = EmitBitFieldUExtract(state, data, ConstantU32(state, 0), ConstantU32(state, 16));
+		} else if (info.type == Format::ComponentType::Sint) {
+			data = EmitBitFieldSExtract(state, data, ConstantU32(state, 0), ConstantU32(state, 16));
+		} else {
+			data = EmitBitCastU32F32(state, EmitF16BitsToF32(state, data));
+		}
 	}
-	if (bits == 16u && (info.type == Format::ComponentType::Snorm ||
-	                    info.type == Format::ComponentType::Float)) {
-		const auto value = EmitBitCastF32U32(ctx.state, data);
-		const auto pair = EmitCompositeConstructF32x2(ctx.state, value,
-		                                               ConstantF32Value(ctx.state, 0.0f));
-		return info.type == Format::ComponentType::Float ? EmitPackHalf2x16(ctx.state, pair)
-		                                               : EmitPackSnorm2x16(ctx.state, pair);
+	const bool normalized = info.type == Format::ComponentType::Unorm ||
+	                        info.type == Format::ComponentType::Snorm;
+	const bool scaled = info.type == Format::ComponentType::Uscaled ||
+	                    info.type == Format::ComponentType::Sscaled;
+	if (normalized || scaled) {
+		// Two-bit packed channels are unsigned even in signed formats.
+		const bool is_signed = bits != 2u && (info.type == Format::ComponentType::Snorm ||
+		                                      info.type == Format::ComponentType::Sscaled);
+		const auto max_value = static_cast<float>((1u << (bits - (is_signed ? 1u : 0u))) - 1u);
+		const auto lower = normalized ? (is_signed ? -1.0f : 0.0f)
+		                              : (is_signed ? -max_value - 1.0f : 0.0f);
+		const auto upper = normalized ? 1.0f : max_value;
+		auto value = Select(state, TypeF32(state), EmitClassifyF32Bits(state, data).nan,
+		                    ConstantF32Value(state, 0.0f), EmitBitCastF32U32(state, data));
+		value = EmitGlsl<GLSLstd450FClamp, IR::Type::F32>(
+		    state, value, ConstantF32Value(state, lower), ConstantF32Value(state, upper));
+		if (normalized) {
+			value = EmitFPRoundEven32(
+			    state, EmitFPMul32(state, value, ConstantF32Value(state, max_value)));
+		}
+		// Float-to-integer conversion truncates scaled values toward zero.
+		value = Unary(state, is_signed ? spv::OpConvertFToS : spv::OpConvertFToU,
+		              is_signed ? TypeI32(state) : TypeU32(state), value);
+		return is_signed ? Unary(state, spv::OpBitcast, TypeU32(state), value) : value;
+	}
+	if (bits == 16u && info.type == Format::ComponentType::Float) {
+		const auto pair = EmitCompositeConstructF32x2(
+		    state, EmitBitCastF32U32(state, data), ConstantF32Value(state, 0.0f));
+		return EmitPackHalf2x16(state, pair);
 	}
 	return data;
 }
@@ -750,7 +799,7 @@ void StoreFormattedPrepared(ValueEmitContext& ctx, const IR::Inst& inst,
 					                              value, data, component);
 				}
 			}
-			value = EncodeFormattedStoreComponent(ctx, info, component, value);
+			value = EncodeFormattedStoreComponent(ctx, info, component, value, mem.data_bits);
 			const auto bits = info.component_bits[component];
 			if (info.packed_bitfield) {
 				packed = EmitBitFieldInsert(
@@ -792,26 +841,25 @@ uint32_t LoadIndirectScalarBuffer(ValueEmitContext& ctx, const IR::Inst& inst) {
 	const auto  word1  = ctx.Arg(handle, 1);
 	const auto  stride = EmitBitFieldUExtract(state, word1, ConstantU32(state, 16),
 	                                         ConstantU32(state, 14));
-	const auto size = Binary(
-	    state, spv::OpIMul, TypeScalarU64(state),
-	    Unary(state, spv::OpUConvert, TypeScalarU64(state),
-	          EmitUMax32(state, stride, ConstantU32(state, 1))),
-	    Unary(state, spv::OpUConvert, TypeScalarU64(state), ctx.Arg(handle, 2)));
-	const auto offset = Binary(
-	    state, spv::OpIAdd, TypeScalarU64(state),
-	    Unary(state, spv::OpUConvert, TypeScalarU64(state),
-	          Binary(state, spv::OpBitwiseAnd, TypeU32(state), ctx.Arg(inst, 1),
-	                 ConstantU32(state, ~3u))),
-	    ConstantDeviceAddress(state, ctx.Memory(inst).offset & ~3u));
-	const auto end = Binary(state, spv::OpIAdd, TypeScalarU64(state), offset,
-	                        ConstantDeviceAddress(state, sizeof(uint32_t)));
+	const auto  size   = Binary(state, spv::OpIMul, TypeU64(state),
+	                            Unary(state, spv::OpUConvert, TypeU64(state),
+	                                  EmitUMax32(state, stride, ConstantU32(state, 1))),
+	                            Unary(state, spv::OpUConvert, TypeU64(state), ctx.Arg(handle, 2)));
+	const auto  offset = Binary(state, spv::OpIAdd, TypeU64(state),
+	                            Unary(state, spv::OpUConvert, TypeU64(state),
+	                                  Binary(state, spv::OpBitwiseAnd, TypeU32(state),
+	                                         ctx.Arg(inst, 1), ConstantU32(state, ~3u))),
+	                            ConstantU64(state, ctx.Memory(inst).offset & ~3u));
+	const auto  end =
+	    Binary(state, spv::OpIAdd, TypeU64(state), offset, ConstantU64(state, sizeof(uint32_t)));
 	return EmitValueOrZeroIfCondition(
 	    state, Binary(state, spv::OpULessThanEqual, TypeBool(state), end, size), [&]() {
-		    const auto base = DeviceAddressFromWords(
-		        state, Binary(state, spv::OpBitwiseAnd, TypeU32(state), ctx.Arg(handle, 0),
-		                      ConstantU32(state, ~3u)),
-		        Binary(state, spv::OpBitwiseAnd, TypeU32(state), word1, ConstantU32(state, 0xffffu)));
-		    return LoadBdaDword(ctx, Binary(state, spv::OpIAdd, TypeScalarU64(state), base, offset));
+		    const auto base = PackU64(state,
+		                              Binary(state, spv::OpBitwiseAnd, TypeU32(state),
+		                                     ctx.Arg(handle, 0), ConstantU32(state, ~3u)),
+		                              Binary(state, spv::OpBitwiseAnd, TypeU32(state), word1,
+		                                     ConstantU32(state, 0xffffu)));
+		    return LoadBdaDword(ctx, Binary(state, spv::OpIAdd, TypeU64(state), base, offset));
 	    });
 }
 
@@ -852,21 +900,21 @@ IndirectBufferAccess PrepareIndirectBuffer(ValueEmitContext& ctx, const IR::Inst
 	const auto address = CalculateBufferAddress(state, index, ctx.Arg(inst, 2), soffset,
 	                                            ctx.Memory(inst).offset, stride, swizzle,
 	                                            field(word3, 21, 2));
-	const auto base = DeviceAddressFromWords(state, ctx.Arg(handle, 0), field(word1, 0, 16));
+	const auto base        = PackU64(state, ctx.Arg(handle, 0), field(word1, 0, 16));
 	const auto raw_records = Binary(state, spv::OpISub, TypeU32(state), records, soffset);
 	return {
-	    .address = Binary(state, spv::OpIAdd, TypeScalarU64(state), base,
-	                      Unary(state, spv::OpUConvert, TypeScalarU64(state), address.byte)),
-	    .offset = address.offset,
-	    .stride = stride,
-	    .swizzle = swizzle,
-	    .format = field(word3, 12, 7),
-	    .selector = ctx.Memory(inst).formatted ? field(word3, 0, 3) : 0u,
-	    .mode = field(word3, 28, 2),
-	    .records = records,
-	    .index_in_bounds = Binary(state, spv::OpULessThan, TypeBool(state), index, records),
+	    .address          = Binary(state, spv::OpIAdd, TypeU64(state), base,
+	                               Unary(state, spv::OpUConvert, TypeU64(state), address.byte)),
+	    .offset           = address.offset,
+	    .stride           = stride,
+	    .swizzle          = swizzle,
+	    .format           = field(word3, 12, 7),
+	    .selector         = ctx.Memory(inst).formatted ? field(word3, 0, 3) : 0u,
+	    .mode             = field(word3, 28, 2),
+	    .records          = records,
+	    .index_in_bounds  = Binary(state, spv::OpULessThan, TypeBool(state), index, records),
 	    .scalar_in_bounds = Binary(state, spv::OpULessThanEqual, TypeBool(state), soffset, records),
-	    .raw_records = raw_records,
+	    .raw_records      = raw_records,
 	    .raw_index_in_bounds = Binary(state, spv::OpULessThan, TypeBool(state), index, raw_records),
 	};
 }
@@ -912,13 +960,13 @@ uint32_t LoadIndirectBuffer(ValueEmitContext& ctx, const IR::Inst& inst, uint32_
 	const auto buffer = PrepareIndirectBuffer(ctx, inst);
 	const auto valid_format = Binary(state, spv::OpINotEqual, TypeBool(state), buffer.format,
 	                                  ConstantU32(state, 0));
-	const auto base = Binary(state, spv::OpBitwiseAnd, TypeScalarU64(state), buffer.address,
-	                         ConstantDeviceAddress(state, ~uint64_t {3}));
+	const auto base         = Binary(state, spv::OpBitwiseAnd, TypeU64(state), buffer.address,
+	                                 ConstantU64(state, ~uint64_t {3}));
 	std::array<uint32_t, 4> values {};
 	for (uint32_t component = 0; component < components; ++component) {
-		const auto address = component == 0u ? base : Binary(
-		    state, spv::OpIAdd, TypeScalarU64(state), base,
-		    ConstantDeviceAddress(state, component * 4u));
+		const auto address = component == 0u ? base
+		                                     : Binary(state, spv::OpIAdd, TypeU64(state), base,
+		                                              ConstantU64(state, component * 4u));
 		values[component] = EmitValueOrZeroIfCondition(
 		    state, AndCondition(state, valid_format,
 		                        IndirectBufferInBounds(state, buffer, component * 4u, 4u, false)),
@@ -946,13 +994,14 @@ uint32_t LoadIndirectFormattedX(ValueEmitContext& ctx, const IR::Inst& inst) {
 			        TypeU32(state), constant, [&]() {
 				        const auto in_bounds = IndirectBufferInBounds(state, buffer, 0u, info.byte_size, true);
 				        const auto base = Binary(
-				            state, spv::OpBitwiseAnd, TypeScalarU64(state), buffer.address,
-				            ConstantDeviceAddress(state, ~(uint64_t(std::min(4u, info.byte_size)) - 1u)));
+				            state, spv::OpBitwiseAnd, TypeU64(state), buffer.address,
+				            ConstantU64(state, ~(uint64_t(std::min(4u, info.byte_size)) - 1u)));
 				        const auto load = [&](uint32_t component, uint32_t bits) {
 					        const auto offset = Format::GetFormatComponentByteOffset(info, component);
-					        const auto address = offset == 0u ? base : Binary(
-					            state, spv::OpIAdd, TypeScalarU64(state), base,
-					            ConstantDeviceAddress(state, offset));
+					        const auto address = offset == 0u
+					                                 ? base
+					                                 : Binary(state, spv::OpIAdd, TypeU64(state),
+					                                          base, ConstantU64(state, offset));
 					        return LoadBda(ctx, address, in_bounds, bits);
 				        };
 				        const auto emit_component = [&](uint32_t component) {
@@ -1072,23 +1121,9 @@ void StoreWideShared(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t compo
 
 } // namespace
 
-uint32_t ConstantDeviceAddress(EmitterState& state, uint64_t value) {
-	return state.builder.Constant(spv::OpConstant, TypeScalarU64(state),
-	                              static_cast<uint32_t>(value),
-	                              static_cast<uint32_t>(value >> 32u));
-}
-
-uint32_t DeviceAddressFromWords(EmitterState& state, uint32_t low, uint32_t high) {
-	const auto low64  = Unary(state, spv::OpUConvert, TypeScalarU64(state), low);
-	const auto high64 = Binary(state, spv::OpShiftLeftLogical, TypeScalarU64(state),
-	                           Unary(state, spv::OpUConvert, TypeScalarU64(state), high),
-	                           ConstantDeviceAddress(state, 32));
-	return Binary(state, spv::OpBitwiseOr, TypeScalarU64(state), low64, high64);
-}
-
 uint32_t GetBdaPointer(EmitterState& state, uint32_t address) {
 	const auto result = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpFunctionCall, TypeScalarU64(state), result,
+	state.builder.AddFunction(spv::OpFunctionCall, TypeU64(state), result,
 	                          state.bda_pointer_function, address);
 	return result;
 }
@@ -1097,7 +1132,7 @@ void DefineGetBdaPointer(EmitterState& state) {
 	if (!state.program.info.uses_dma) {
 		return;
 	}
-	const auto type            = TypeScalarU64(state);
+	const auto type            = TypeU64(state);
 	const auto function_type   = state.builder.Type(spv::OpTypeFunction, type, type);
 	state.bda_pointer_function = state.builder.AllocateId();
 	const auto address         = state.builder.AllocateId();
@@ -1109,14 +1144,14 @@ void DefineGetBdaPointer(EmitterState& state) {
 	EmitLabel(state, entry_label);
 
 	const auto extended = Binary(state, spv::OpUGreaterThanEqual, TypeBool(state), address,
-	                             ConstantDeviceAddress(state, LibKernel::Memory::kExtendedMemoryBase));
-	const auto packed = Select(state, type, extended,
-	                           Binary(state, spv::OpISub, type, address,
-	                                  ConstantDeviceAddress(state,
-	                                      LibKernel::Memory::kExtendedMemoryBase - LOWER_ADDRESS_SIZE)),
-	                           address);
+	                             ConstantU64(state, LibKernel::Memory::kExtendedMemoryBase));
+	const auto packed   = Select(
+	    state, type, extended,
+	    Binary(state, spv::OpISub, type, address,
+	           ConstantU64(state, LibKernel::Memory::kExtendedMemoryBase - LOWER_ADDRESS_SIZE)),
+	    address);
 	const auto page64        = Binary(state, spv::OpShiftRightLogical, type, packed,
-	                                  ConstantDeviceAddress(state, BufferCache::CACHING_PAGEBITS));
+	                                  ConstantU32(state, BufferCache::CACHING_PAGEBITS));
 	const auto page          = Unary(state, spv::OpUConvert, TypeU32(state), page64);
 	const auto entry_pointer = state.builder.AllocateId();
 	state.builder.AddFunction(spv::OpAccessChain, TypeStorageBufferElementPointer(state, 64),
@@ -1124,8 +1159,7 @@ void DefineGetBdaPointer(EmitterState& state) {
 	                          page);
 	const auto base = state.builder.AllocateId();
 	state.builder.AddFunction(spv::OpLoad, type, base, entry_pointer);
-	const auto missing =
-	    Binary(state, spv::OpIEqual, TypeBool(state), base, ConstantDeviceAddress(state, 0));
+	const auto missing = Binary(state, spv::OpIEqual, TypeBool(state), base, ConstantU64(state, 0));
 	const auto fault_label     = state.builder.AllocateId();
 	const auto available_label = state.builder.AllocateId();
 	const auto merge_label     = state.builder.AllocateId();
@@ -1138,14 +1172,14 @@ void DefineGetBdaPointer(EmitterState& state) {
 
 	EmitLabel(state, available_label);
 	const auto offset    = Binary(state, spv::OpBitwiseAnd, type, address,
-	                              ConstantDeviceAddress(state, BufferCache::CACHING_PAGESIZE - 1));
+	                              ConstantU64(state, BufferCache::CACHING_PAGESIZE - 1));
 	const auto available = Binary(state, spv::OpIAdd, type, base, offset);
 	state.builder.AddFunction(spv::OpBranch, merge_label);
 
 	EmitLabel(state, merge_label);
 	const auto result = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpPhi, type, result, ConstantDeviceAddress(state, 0),
-	                          fault_label, available, available_label);
+	state.builder.AddFunction(spv::OpPhi, type, result, ConstantU64(state, 0), fault_label,
+	                          available, available_label);
 	state.builder.AddFunction(spv::OpReturnValue, result);
 	state.builder.AddFunction(spv::OpFunctionEnd);
 }
@@ -1203,12 +1237,12 @@ void EmitSharedAtomic64(ValueEmitContext& ctx, const IR::Inst& inst) {
 				semantics = spv::MemorySemanticsUniformMemoryMask;
 			}
 			const auto pointer = state.builder.AllocateId();
-			state.builder.AddFunction(spv::OpAccessChain,
-			                          TypePointer(state, state.lds_storage_class, TypeScalarU64(state)),
-			                          pointer, state.lds_u64_variable, ConstantU32(state, 0), native_index);
-			const auto value = Unary(state, spv::OpBitcast, TypeScalarU64(state), ctx.Arg(inst, 1));
 			state.builder.AddFunction(
-			    SpirvAtomicOpcode(inst.GetOpcode()), TypeScalarU64(state), state.builder.AllocateId(),
+			    spv::OpAccessChain, TypePointer(state, state.lds_storage_class, TypeU64(state)),
+			    pointer, state.lds_u64_variable, ConstantU32(state, 0), native_index);
+			const auto value = ctx.Arg(inst, 1);
+			state.builder.AddFunction(
+			    SpirvAtomicOpcode(inst.GetOpcode()), TypeU64(state), state.builder.AllocateId(),
 			    pointer, ConstantU32(state, spv::ScopeWorkgroup),
 			    ConstantU32(state, spv::MemorySemanticsAcquireReleaseMask | semantics), value);
 		});
@@ -1227,34 +1261,10 @@ uint32_t EmitBufferFloatAtomic(ValueEmitContext& ctx, const IR::Inst& inst) {
 void EmitSharedFloatAtomic(ValueEmitContext& ctx, const IR::Inst& inst) {
 	const auto& mem       = ctx.Memory(inst);
 	const bool  max_value = inst.GetOpcode() == IR::ValueOpcode::SharedAtomicFMax32;
-	EmitIfCondition(ctx.state, ctx.Arg(inst, inst.NumArgs() - 1), [&]() {
-		const auto index = DwordIndex(ctx, inst, mem);
-		const auto resource = PrepareMemoryResourceAccess(ctx.state, mem);
-		EmitIfCondition(
-		    ctx.state, EmitMemoryElementInBounds(ctx.state, resource, index), [&]() {
-			    ctx.state.builder.AddFunction(spv::OpStore, ctx.scratch_u32_variable,
-			                                  ctx.Arg(inst, 1));
-			    const auto data = ctx.state.builder.AllocateId();
-			    ctx.state.builder.AddFunction(spv::OpLoad, TypeU32(ctx.state), data,
-			                                  ctx.scratch_u32_variable);
-			    AtomicUpdate(
-			        ctx.state, EmitMemoryElementPointer(ctx.state, resource, index),
-			        mem.kind, [&](uint32_t old) {
-				        const auto old_f =
-				            Unary(ctx.state, spv::OpBitcast, TypeF32(ctx.state), old);
-				        const auto compare_f =
-				            Unary(ctx.state, spv::OpBitcast, TypeF32(ctx.state), ctx.Arg(inst, 2));
-				        const auto data_f =
-				            Unary(ctx.state, spv::OpBitcast, TypeF32(ctx.state), data);
-				        const auto compare = Binary(
-				            ctx.state, max_value ? spv::OpFOrdGreaterThan : spv::OpFOrdLessThan,
-				            TypeBool(ctx.state), max_value ? old_f : compare_f,
-				            max_value ? compare_f : old_f);
-				        return Unary(ctx.state, spv::OpBitcast, TypeU32(ctx.state),
-				                     Select(ctx.state, TypeF32(ctx.state), compare, data_f, old_f));
-			        });
-		    });
-	});
+	EmitAtomicUpdate(ctx, inst, mem,
+	                 [max_value](EmitterState& state, uint32_t old, uint32_t value) {
+		                 return EmitDsFloatAtomicReplacement(state, old, value, max_value);
+	                 });
 }
 
 uint32_t EmitAppendConsume(ValueEmitContext& ctx, const IR::Inst& inst) {
@@ -1386,7 +1396,7 @@ void EmitLoadMemory(ValueEmitContext& ctx, const IR::Inst& inst) {
 	else if (address_info.access == IR::AddressAccess::Read &&
 	         mem.kind != IR::ResourceKind::Scratch)
 		value = LoadBda(ctx, GuestAddress(ctx, inst, mem), ctx.Arg(inst, inst.NumArgs() - 1),
-		                address_info.data_bits);
+		                address_info.data_bits, mem.coherent);
 	else if (op == IR::ValueOpcode::LoadBufferU32 && mem.formatted)
 		value = mem.kind == IR::ResourceKind::IndirectBuffer ? LoadIndirectFormattedX(ctx, inst)
 		                                                     : FormattedLoad(ctx, inst, mem);

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

@@ -44,6 +44,16 @@ uint32_t EmitIndirectResourceIndex(EmitterState& state, uint32_t key, uint32_t m
 	};
 	const auto mapping = ConstantU32(state, mapping_offset);
 	const auto count = LoadMapping(mapping);
+	if (search_iterations == 0u) {
+		const auto index = Binary(state, spv::OpShiftRightLogical, TypeU32(state), key,
+		                          ConstantU32(state, 2u));
+		const auto valid = Binary(state, spv::OpULessThan, TypeBool(state), index, count);
+		const auto safe = Select(state, TypeU32(state), valid, index, ConstantU32(state, 0u));
+		const auto entry = Binary(state, spv::OpIAdd, TypeU32(state), safe,
+		                          ConstantU32(state, mapping_offset + 1u));
+		return Select(state, TypeU32(state), valid, LoadMapping(entry),
+		              ConstantU32(state, default_resource));
+	}
 	const auto entry_at = [&](uint32_t index) {
 		return Binary(state, spv::OpIAdd, TypeU32(state), mapping,
 		              Binary(state, spv::OpIAdd, TypeU32(state),
@@ -138,7 +148,7 @@ void EnsureLdsStorage(EmitterState& state) {
 	};
 	if (state.requirements.shared_int64_atomics) {
 		state.lds_variable = define(TypeU32(state), 4u);
-		state.lds_u64_variable = define(TypeScalarU64(state), 8u);
+		state.lds_u64_variable = define(TypeU64(state), 8u);
 		state.builder.AddName(state.lds_u64_variable, "lds_qwords");
 	} else {
 		state.lds_variable = state.builder.DefineGlobalVariable(
@@ -361,10 +371,12 @@ spv::Op SpirvAtomicOpcode(IR::ValueOpcode opcode) {
 		case IR::ValueOpcode::ImageAtomicCompareSwap32:
 		case IR::ValueOpcode::BufferAtomicCmpSwap32: return spv::OpAtomicCompareExchange;
 		case IR::ValueOpcode::ImageAtomicSwap32:
+		case IR::ValueOpcode::ImageAtomicSwap64:
 		case IR::ValueOpcode::BufferAtomicSwap32:
 		case IR::ValueOpcode::BufferAtomicSwap64:
 		case IR::ValueOpcode::SharedAtomicSwap32: return spv::OpAtomicExchange;
 		case IR::ValueOpcode::ImageAtomicIAdd32:
+		case IR::ValueOpcode::ImageAtomicIAdd64:
 		case IR::ValueOpcode::BufferAtomicIAdd32:
 		case IR::ValueOpcode::SharedAtomicIAdd64:
 		case IR::ValueOpcode::SharedAtomicIAdd32: return spv::OpAtomicIAdd;
@@ -374,6 +386,7 @@ spv::Op SpirvAtomicOpcode(IR::ValueOpcode opcode) {
 		case IR::ValueOpcode::BufferAtomicSMin32:
 		case IR::ValueOpcode::SharedAtomicSMin32: return spv::OpAtomicSMin;
 		case IR::ValueOpcode::ImageAtomicUMin32:
+		case IR::ValueOpcode::ImageAtomicUMin64:
 		case IR::ValueOpcode::BufferAtomicUMin32:
 		case IR::ValueOpcode::SharedAtomicUMin32: return spv::OpAtomicUMin;
 		case IR::ValueOpcode::ImageAtomicSMax32:
@@ -384,15 +397,18 @@ spv::Op SpirvAtomicOpcode(IR::ValueOpcode opcode) {
 		case IR::ValueOpcode::BufferAtomicUMax32:
 		case IR::ValueOpcode::SharedAtomicUMax32: return spv::OpAtomicUMax;
 		case IR::ValueOpcode::ImageAtomicAnd32:
+		case IR::ValueOpcode::ImageAtomicAnd64:
 		case IR::ValueOpcode::BufferAtomicAnd32:
 		case IR::ValueOpcode::BufferAtomicAnd64:
 		case IR::ValueOpcode::SharedAtomicAnd32: return spv::OpAtomicAnd;
 		case IR::ValueOpcode::ImageAtomicOr32:
+		case IR::ValueOpcode::ImageAtomicOr64:
 		case IR::ValueOpcode::BufferAtomicOr32:
 		case IR::ValueOpcode::BufferAtomicOr64:
 		case IR::ValueOpcode::SharedAtomicOr64:
 		case IR::ValueOpcode::SharedAtomicOr32: return spv::OpAtomicOr;
 		case IR::ValueOpcode::ImageAtomicXor32:
+		case IR::ValueOpcode::ImageAtomicXor64:
 		case IR::ValueOpcode::BufferAtomicXor32:
 		case IR::ValueOpcode::SharedAtomicXor32: return spv::OpAtomicXor;
 		default: return spv::OpNop;
@@ -404,7 +420,7 @@ uint32_t EmitAtomicOperation(ValueEmitContext& ctx, const IR::Inst& inst, uint32
 	const auto opcode = SpirvAtomicOpcode(inst.GetOpcode());
 	const auto old    = ctx.state.builder.AllocateId();
 	const bool wide   = inst.GetType() == IR::Type::U64;
-	const auto type   = wide ? TypeScalarU64(ctx.state) : TypeU32(ctx.state);
+	const auto type   = wide ? TypeU64(ctx.state) : TypeU32(ctx.state);
 	if (opcode == spv::OpAtomicCompareExchange) {
 		const auto desired    = ctx.Arg(inst, inst.NumArgs() - 3);
 		const auto comparator = ctx.Arg(inst, inst.NumArgs() - 2);
@@ -413,13 +429,12 @@ uint32_t EmitAtomicOperation(ValueEmitContext& ctx, const IR::Inst& inst, uint32
 		    ConstantU32(ctx.state, scope), ConstantU32(ctx.state, spv::MemorySemanticsMaskNone),
 		    ConstantU32(ctx.state, spv::MemorySemanticsMaskNone), desired, comparator);
 	} else {
-		auto value = ctx.Arg(inst, inst.NumArgs() - 2);
-		if (wide) value = Unary(ctx.state, spv::OpBitcast, type, value);
+		const auto value = ctx.Arg(inst, inst.NumArgs() - 2);
 		ctx.state.builder.AddFunction(opcode, type, old, pointer,
 		                              ConstantU32(ctx.state, scope),
 		                              ConstantU32(ctx.state, spv::MemorySemanticsMaskNone), value);
 	}
-	return wide ? Unary(ctx.state, spv::OpBitcast, TypeU64(ctx.state), old) : old;
+	return old;
 }
 
 void EmitAtomicMemoryBarrier(EmitterState& state, IR::ResourceKind kind) {
@@ -438,6 +453,20 @@ void EmitAtomicMemoryBarrier(EmitterState& state, IR::ResourceKind kind) {
 	                          ConstantU32(state, spv::MemorySemanticsAcquireReleaseMask | memory));
 }
 
+namespace {
+
+uint32_t FloatOrderKey(EmitterState& state, uint32_t bits) {
+	const auto negative = EmitCompareU32Constant(state, spv::OpINotEqual,
+	                                             EmitAndConstant(state, bits, 0x80000000u), 0u);
+	const auto negative_key = state.builder.AllocateId();
+	state.builder.AddFunction(spv::OpNot, TypeU32(state), negative_key, bits);
+	const auto positive_key =
+	    EmitBinaryU32(state, spv::OpBitwiseXor, bits, ConstantU32(state, 0x80000000u));
+	return EmitSelectValueU32(state, negative, negative_key, positive_key);
+}
+
+} // namespace
+
 uint32_t EmitFloatAtomicReplacement(EmitterState& state, uint32_t old, uint32_t source,
                                     bool max_value) {
 	struct OrderedBits {
@@ -447,15 +476,7 @@ uint32_t EmitFloatAtomicReplacement(EmitterState& state, uint32_t old, uint32_t
 	};
 	const auto classify = [&](uint32_t bits) {
 		const auto cls = EmitClassifyF32Bits(state, bits);
-		const auto negative = EmitCompareU32Constant(state, spv::OpINotEqual,
-		                                             EmitAndConstant(state, bits, 0x80000000u), 0u);
-		const auto negative_key = state.builder.AllocateId();
-		state.builder.AddFunction(spv::OpNot, TypeU32(state), negative_key, bits);
-		const auto positive_key =
-		    EmitBinaryU32(state, spv::OpBitwiseXor, bits, ConstantU32(state, 0x80000000u));
-		return OrderedBits {
-		    cls.nan, cls.zero,
-		    EmitSelectValueU32(state, negative, negative_key, positive_key)};
+		return OrderedBits {cls.nan, cls.zero, FloatOrderKey(state, bits)};
 	};
 	const auto source_class = classify(source);
 	const auto old_class    = classify(old);
@@ -470,6 +491,45 @@ uint32_t EmitFloatAtomicReplacement(EmitterState& state, uint32_t old, uint32_t
 	    old);
 }
 
+uint32_t EmitDsFloatAtomicReplacement(EmitterState& state, uint32_t old, uint32_t source,
+                                      bool max_value) {
+	struct OrderedBits {
+		uint32_t nan;
+		uint32_t signaling_nan;
+		uint32_t key;
+	};
+	const auto classify = [&](uint32_t bits) {
+		const auto abs = EmitAndConstant(state, bits, 0x7fffffffu);
+		const auto nan = EmitCompareU32Constant(state, spv::OpUGreaterThan, abs, 0x7f800000u);
+		const auto signaling_nan = EmitLogicalAndBool(
+		    state, nan, EmitCompareU32Constant(state, spv::OpIEqual,
+		                                      EmitAndConstant(state, bits, 0x00400000u), 0u));
+		const auto small = EmitCompareU32Constant(state, spv::OpULessThan, abs, 0x00800000u);
+		const auto sign = EmitAndConstant(state, bits, 0x80000000u);
+		const auto positive_denorm = EmitLogicalAndBool(
+		    state, EmitCompareU32Constant(state, spv::OpIEqual, sign, 0u),
+		    EmitCompareU32Constant(state, spv::OpINotEqual, abs, 0u));
+		// Negative denorms tie with -0; positive denorms tie only with each other.
+		const auto small_bits = EmitSelectValueU32(
+		    state, positive_denorm, ConstantU32(state, 1u), sign);
+		const auto key_bits = EmitSelectValueU32(state, small, small_bits, bits);
+		return OrderedBits {nan, signaling_nan, FloatOrderKey(state, key_bits)};
+	};
+	const auto lhs = classify(old);
+	const auto rhs = classify(source);
+	const auto choose_source = state.builder.AllocateId();
+	state.builder.AddFunction(max_value ? spv::OpUGreaterThan : spv::OpULessThanEqual,
+	                          TypeBool(state), choose_source, rhs.key, lhs.key);
+	// Keep original bits: ordering denorms does not flush the selected operand.
+	auto result = EmitSelectValueU32(state, choose_source, source, old);
+	result = EmitSelectValueU32(state, lhs.nan, source, result);
+	result = EmitSelectValueU32(state, rhs.nan, old, result);
+	result = EmitSelectValueU32(state, rhs.signaling_nan,
+	                            EmitOrU32(state, source, ConstantU32(state, 0x00400000u)), result);
+	return EmitSelectValueU32(state, lhs.signaling_nan,
+	                          EmitOrU32(state, old, ConstantU32(state, 0x00400000u)), result);
+}
+
 uint32_t EmitDsSwizzleTargetLane(EmitterState& state, uint32_t subid, uint32_t control) {
 	if ((control & 0xc000u) == 0xc000u) {
 		const uint32_t mask         = control & 0x1fu;

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

@@ -35,14 +35,11 @@ uint32_t TypeU32(EmitterState& state) {
 }
 
 uint32_t TypeU64(EmitterState& state) {
-	return TypeU32Vector(state, 2);
-}
-
-uint32_t TypeScalarU64(EmitterState& state) {
-	if (state.native_u64_type == 0) {
-		state.native_u64_type = state.builder.Type(spv::OpTypeInt, 64, 0);
+	if (state.u64_type == 0) {
+		state.builder.RequireCapability(spv::CapabilityInt64);
+		state.u64_type = state.builder.Type(spv::OpTypeInt, 64, 0);
 	}
-	return state.native_u64_type;
+	return state.u64_type;
 }
 
 uint32_t TypeU32Pair(EmitterState& state) {
@@ -124,8 +121,9 @@ uint32_t TypeFunction(EmitterState& state) {
 }
 
 uint32_t TypeStorageBufferElement(EmitterState& state, uint32_t bits) {
-	return bits == 32u ? TypeU32(state) : bits == 64u ? TypeScalarU64(state)
-	                                                   : state.builder.Type(spv::OpTypeInt, bits, 0);
+	return bits == 32u   ? TypeU32(state)
+	       : bits == 64u ? TypeU64(state)
+	                     : state.builder.Type(spv::OpTypeInt, bits, 0);
 }
 
 uint32_t StorageRuntimeArrayType(EmitterState& state, uint32_t bits) {
@@ -331,9 +329,8 @@ uint32_t ConstantBool(EmitterState& state, bool value) {
 }
 
 uint32_t ConstantU64(EmitterState& state, uint64_t value) {
-	return state.builder.Constant(spv::OpConstantComposite, TypeU64(state),
-	                              ConstantU32(state, static_cast<uint32_t>(value)),
-	                              ConstantU32(state, static_cast<uint32_t>(value >> 32u)));
+	return state.builder.Constant(spv::OpConstant, TypeU64(state), static_cast<uint32_t>(value),
+	                              static_cast<uint32_t>(value >> 32u));
 }
 
 uint32_t ConstantU32CompositeZero(EmitterState& state, uint32_t components) {

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

@@ -689,9 +689,7 @@ void EmitProgram(EmitterState& state) {
 	DefineBvhIntersect(state);
 	for (const auto* block: program.blocks) {
 		if (std::ranges::any_of(*block, [](const IR::Inst& inst) {
-			    return inst.GetOpcode() == IR::ValueOpcode::SwizzleU32 ||
-			           inst.GetOpcode() == IR::ValueOpcode::SharedAtomicFMin32 ||
-			           inst.GetOpcode() == IR::ValueOpcode::SharedAtomicFMax32;
+			    return inst.GetOpcode() == IR::ValueOpcode::SwizzleU32;
 		    })) {
 			ctx.scratch_u32_variable = state.builder.AllocateId();
 			if (state.lane_count == 2) {

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

@@ -20,6 +20,12 @@ struct MimgGatherInfo {
 	uint32_t flags    = 0;
 };
 
+struct MimgAtomicInfo {
+	uint32_t encoding    = 0;
+	Opcode   decoded     = Opcode::UNSUPPORTED;
+	bool     supports_64 = false;
+};
+
 constexpr ImageDimension DecodeImageDimension(uint32_t dim) {
 	switch (dim) {
 		case 0u: return ImageDimension::Dim1D;
@@ -190,17 +196,17 @@ constexpr MimgGatherInfo MIMG_GATHER_OPCODE_LIST[] = {
     {0x61u, Opcode::IMAGE_GATHER4H, ImageSampleFlagGatherHorizontal},
 };
 
-constexpr Detail::OpcodeMap MIMG_ATOMIC_OPCODE_LIST[] = {
-    {0x0fu, Opcode::IMAGE_ATOMIC_SWAP},
+constexpr MimgAtomicInfo MIMG_ATOMIC_OPCODE_LIST[] = {
+    {0x0fu, Opcode::IMAGE_ATOMIC_SWAP, true},
     {0x10u, Opcode::IMAGE_ATOMIC_CMPSWAP},
-    {0x11u, Opcode::IMAGE_ATOMIC_ADD},
+    {0x11u, Opcode::IMAGE_ATOMIC_ADD, true},
     {0x14u, Opcode::IMAGE_ATOMIC_SMIN},
-    {0x15u, Opcode::IMAGE_ATOMIC_UMIN},
+    {0x15u, Opcode::IMAGE_ATOMIC_UMIN, true},
     {0x16u, Opcode::IMAGE_ATOMIC_SMAX},
-    {0x17u, Opcode::IMAGE_ATOMIC_UMAX},
-    {0x18u, Opcode::IMAGE_ATOMIC_AND},
-    {0x19u, Opcode::IMAGE_ATOMIC_OR},
-    {0x1au, Opcode::IMAGE_ATOMIC_XOR},
+    {0x17u, Opcode::IMAGE_ATOMIC_UMAX, true},
+    {0x18u, Opcode::IMAGE_ATOMIC_AND, true},
+    {0x19u, Opcode::IMAGE_ATOMIC_OR, true},
+    {0x1au, Opcode::IMAGE_ATOMIC_XOR, true},
     {0x1eu, Opcode::IMAGE_ATOMIC_FMIN},
     {0x1fu, Opcode::IMAGE_ATOMIC_FMAX},
 };
@@ -210,7 +216,7 @@ constexpr auto MIMG_GATHER_OPS = Detail::MakeOpcodeTable<0x100>(MIMG_GATHER_OPCO
 constexpr auto MIMG_ATOMIC_OPS = Detail::MakeOpcodeTable<0x100>(MIMG_ATOMIC_OPCODE_LIST);
 
 Opcode DecodeMimgOpcode(uint32_t opcode, const MimgSampleInfo* sample, const MimgGatherInfo* gather,
-                        const Detail::OpcodeMap* atomic) {
+                        const MimgAtomicInfo* atomic) {
 	if (sample != nullptr) {
 		return Opcode::IMAGE_SAMPLE;
 	}
@@ -246,7 +252,7 @@ uint32_t DecodeMimgSampleFlags(const MimgSampleInfo* sample, const MimgGatherInf
 
 uint32_t DecodeMimgAddressComponents(uint32_t opcode, ImageDimension dimension,
                                      const MimgSampleInfo* sample, const MimgGatherInfo* gather,
-                                     const Detail::OpcodeMap* atomic) {
+                                     const MimgAtomicInfo* atomic) {
 	if (sample != nullptr) {
 		return ImageSampleAddressComponents(sample->flags, dimension);
 	}
@@ -370,7 +376,7 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 			SetUnsupported(inst, Family::MIMG, opcode, "MIMG image atomic has invalid DMASK");
 		} else if (inst.dmask == mask64) {
 			inst.data_bits = 64u;
-			if (inst.opcode != Opcode::IMAGE_ATOMIC_UMAX) {
+			if (!atomic->supports_64) {
 				SetUnsupported(inst, Family::MIMG, opcode,
 				               "MIMG 64-bit image atomic opcode is not implemented");
 			}

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

@@ -63,6 +63,8 @@ constexpr MemoryOpcodeInfo MUBUF_OPCODE_LIST[] = {
     {0x50u, Opcode::BUFFER_ATOMIC_SWAP_X2, 2, 32},
     {0x59u, Opcode::BUFFER_ATOMIC_AND_X2, 2, 32},
     {0x5au, Opcode::BUFFER_ATOMIC_OR_X2, 2, 32},
+    {0x80u, Opcode::BUFFER_LOAD_FORMAT_D16_X, 1, 16, false, false, true},
+    {0x84u, Opcode::BUFFER_STORE_FORMAT_D16_X, 1, 16, false, false, true},
 };
 
 constexpr MemoryOpcodeInfo MTBUF_OPCODE_LIST[] = {
@@ -84,6 +86,7 @@ constexpr MemoryOpcodeInfo FLAT_OPCODE_LIST[] = {
     {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},
 };
 
 constexpr MemoryOpcodeInfo DS_OPCODE_LIST[] = {
@@ -108,6 +111,7 @@ constexpr MemoryOpcodeInfo DS_OPCODE_LIST[] = {
     {0x39u, Opcode::DS_READ_I8, 1, 8, true},     {0x3au, Opcode::DS_READ_U8, 1, 8},
     {0x3bu, Opcode::DS_READ_I16, 1, 16, true},   {0x3cu, Opcode::DS_READ_U16, 1, 16},
     {0x3du, Opcode::DS_CONSUME, 1, 32},          {0x3eu, Opcode::DS_APPEND, 1, 32},
+    {0x3fu, Opcode::DS_ORDERED_COUNT, 1, 32},
     {0x40u, Opcode::DS_ADD_U64, 2, 32},         {0x4au, Opcode::DS_OR_B64, 2, 32},
     {0x4du, Opcode::DS_WRITE_B64, 2, 32},        {0x4eu, Opcode::DS_WRITE2_B64, 4, 32},
     {0x4fu, Opcode::DS_WRITE2ST64_B64, 4, 32},   {0x76u, Opcode::DS_READ_B64, 2, 32},
@@ -205,14 +209,15 @@ uint32_t DsSourceCount(Opcode opcode) {
 		case Opcode::DS_WRITE2ST64_B32:
 		case Opcode::DS_WRITE2_B64:
 		case Opcode::DS_WRITE2ST64_B64:
-		case Opcode::DS_MSKOR_B32:
+		case Opcode::DS_MSKOR_B32: return 3u;
 		case Opcode::DS_MIN_F32:
-		case Opcode::DS_MAX_F32: return 3u;
+		case Opcode::DS_MAX_F32:
 		case Opcode::DS_PERMUTE_B32:
 		case Opcode::DS_BPERMUTE_B32: return 2u;
 		case Opcode::DS_READ_ADDTID_B32:
 		case Opcode::DS_CONSUME:
 		case Opcode::DS_APPEND: return 0u;
+		case Opcode::DS_ORDERED_COUNT: return 1u;
 		default: return IsDsWriteOpcode(opcode) || IsDsAtomicOpcode(opcode) ? 2u : 1u;
 	}
 }
@@ -293,6 +298,9 @@ void DecodeMubuf(uint32_t pc, std::span<const uint32_t> code, uint32_t word_inde
 	}
 
 	DecodeVectorGpr(vdata, inst.dst);
+	if (inst.opcode == Opcode::BUFFER_LOAD_FORMAT_D16_X) {
+		inst.dst.sdwa_sel = 4u;
+	}
 	DecodeVectorGpr(vaddr, inst.src0);
 	DecodeScalarSource(srsrc * 4u, pc, inst.src1);
 	DecodeScalarSource(soffset, pc, inst.src2);
@@ -342,7 +350,6 @@ void DecodeFlat(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	const uint32_t word0  = code[word_index];
 	const uint32_t word1  = code[word_index + 1u];
 	const uint32_t offset = word0 & 0xfffu;
-	const uint32_t dlc    = (word0 >> 12u) & 1u;
 	const uint32_t lds    = (word0 >> 13u) & 1u;
 	const uint32_t seg    = (word0 >> 14u) & 0x3u;
 	const uint32_t opcode = (word0 >> 18u) & 0x7fu;
@@ -354,6 +361,7 @@ void DecodeFlat(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	inst.pc             = pc;
 	inst.word_count     = 2;
 	inst.offset         = seg == 0u ? (offset & 0x7ffu) : SignExtendU32(offset, 12u);
+	inst.dlc            = ((word0 >> 12u) & 1u) != 0;
 	inst.glc            = ((word0 >> 16u) & 1u) != 0;
 	inst.slc            = ((word0 >> 17u) & 1u) != 0;
 	inst.family         = Family::FLAT;
@@ -363,7 +371,7 @@ void DecodeFlat(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	ApplyMemoryInfo(inst, info);
 	SetRawWords(inst, code, word_index, 2);
 
-	if (dlc != 0 || lds != 0 || inst.glc || inst.slc || seg == 3u) {
+	if (lds != 0 || inst.glc || inst.slc || seg == 3u) {
 		SetUnsupported(inst, Family::FLAT, opcode, "FLAT modifiers or segment are not implemented");
 		return;
 	}
@@ -373,6 +381,10 @@ 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;
+	}
 	DecodeVectorGpr(addr, inst.src0);
 	inst.src_count = 1;
 	if (seg == 0u || saddr == 0x7du || saddr == 0x7fu) {
@@ -407,6 +419,14 @@ void DecodeDs(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index,
 	if (inst.opcode == Opcode::UNSUPPORTED) {
 		SetUnsupported(inst, Family::DS, opcode, "DS opcode is not implemented");
 	}
+	if (inst.opcode == Opcode::DS_ORDERED_COUNT) {
+		// Keep OFFSET1's wave type as well as its add/swap selector for GDS addressing.
+		inst.offset           = offset0 & 0xfcu;
+		inst.secondary_offset = offset1;
+		if (!inst.gds || ((offset1 >> 4u) & 3u) > 1u || ((offset1 >> 6u) & 3u) != 0u) {
+			SetUnsupported(inst, Family::DS, opcode, "DS ordered count variant is not implemented");
+		}
+	}
 	if (inst.opcode == Opcode::DS_SWIZZLE_B32 && inst.offset >= 0xe000u) {
 		SetUnsupported(inst, Family::DS, opcode, "DS swizzle FFT mode is not implemented");
 	}

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

@@ -53,6 +53,8 @@ constexpr OpcodeMap SOP1_OPCODE_LIST[] = {
     {0x14u, Opcode::S_FF1_I32_B64},
     {0x15u, Opcode::S_FLBIT_I32_B32},
     {0x16u, Opcode::S_FLBIT_I32_B64},
+    {0x19u, Opcode::S_SEXT_I32_I8},
+    {0x1au, Opcode::S_SEXT_I32_I16},
     {0x1bu, Opcode::S_BITSET0_B32},
     {0x1cu, Opcode::S_BITSET0_B64},
     {0x1du, Opcode::S_BITSET1_B32},

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

@@ -503,6 +503,8 @@ std::string InstructionToString(const Instruction& inst) {
 			                                               OperandToString(inst.dst).c_str(),
 			                                               OperandToString(inst.src0).c_str()));
 		case Opcode::S_ABS_I32:
+		case Opcode::S_SEXT_I32_I8:
+		case Opcode::S_SEXT_I32_I16:
 		case Opcode::S_BREV_B32:
 		case Opcode::S_BREV_B64:
 		case Opcode::S_BCNT1_I32_B32:
@@ -617,10 +619,12 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::BUFFER_LOAD_FORMAT_XY:
 		case Opcode::BUFFER_LOAD_FORMAT_XYZ:
 		case Opcode::BUFFER_LOAD_FORMAT_XYZW:
+		case Opcode::BUFFER_LOAD_FORMAT_D16_X:
 		case Opcode::BUFFER_STORE_FORMAT_X:
 		case Opcode::BUFFER_STORE_FORMAT_XY:
 		case Opcode::BUFFER_STORE_FORMAT_XYZ:
 		case Opcode::BUFFER_STORE_FORMAT_XYZW:
+		case Opcode::BUFFER_STORE_FORMAT_D16_X:
 		case Opcode::BUFFER_LOAD_UBYTE:
 		case Opcode::BUFFER_LOAD_USHORT:
 		case Opcode::BUFFER_LOAD_DWORD:
@@ -663,6 +667,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::FLAT_LOAD_SBYTE:
 		case Opcode::FLAT_LOAD_USHORT:
 		case Opcode::FLAT_LOAD_SSHORT:
+		case Opcode::FLAT_LOAD_SHORT_D16:
 		case Opcode::FLAT_LOAD_DWORD:
 		case Opcode::FLAT_LOAD_DWORDX2:
 		case Opcode::FLAT_LOAD_DWORDX3:

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

@@ -45,6 +45,8 @@ enum class Opcode {
 	S_CMOV_B64,
 	S_MOVK_I32,
 	S_ABS_I32,
+	S_SEXT_I32_I8,
+	S_SEXT_I32_I16,
 	S_ABSDIFF_I32,
 	S_BREV_B32,
 	S_BREV_B64,
@@ -224,6 +226,9 @@ enum class Opcode {
 	V_MAC_F32,
 	V_MADMK_F32,
 	V_MADAK_F32,
+	V_FMAC_F32,
+	V_FMAMK_F32,
+	V_FMAAK_F32,
 	V_MIN_F32,
 	V_MAX_F32,
 	V_ADD_F16,
@@ -265,6 +270,7 @@ enum class Opcode {
 	V_MIN3_I32,
 	V_MIN3_U32,
 	V_MIN3_F16,
+	V_MIN3_U16,
 	V_MAX3_F32,
 	V_MAX3_I32,
 	V_MAX3_U32,
@@ -378,6 +384,7 @@ enum class Opcode {
 	V_CMPX_LG_F32,
 	V_CMPX_GE_F32,
 	V_CMPX_O_F32,
+	V_CMPX_U_F32,
 	V_CMPX_NGE_F32,
 	V_CMPX_NLG_F32,
 	V_CMPX_NGT_F32,
@@ -394,6 +401,8 @@ enum class Opcode {
 	V_CMP_T_I32,
 	V_CMP_CLASS_F32,
 	V_CMPX_CLASS_F32,
+	V_CMP_CLASS_F16,
+	V_CMPX_CLASS_F16,
 	V_CMP_LT_I16,
 	V_CMP_EQ_I16,
 	V_CMP_LE_I16,
@@ -438,6 +447,7 @@ enum class Opcode {
 	V_CMPX_LT_U16,
 	V_CMPX_EQ_U16,
 	V_CMPX_GT_U16,
+	V_CMPX_NE_U16,
 	V_CMP_NE_U16,
 	V_CMP_GE_U16,
 	V_CMP_F_U32,
@@ -448,19 +458,34 @@ enum class Opcode {
 	V_CMP_NE_U32,
 	V_CMP_GE_U32,
 	V_CMP_T_U32,
+	V_CMP_F_I64,
 	V_CMP_EQ_I64,
 	V_CMP_LT_I64,
 	V_CMP_LE_I64,
 	V_CMP_NE_I64,
+	V_CMP_T_I64,
+	V_CMP_F_U64,
 	V_CMP_LT_U64,
 	V_CMP_EQ_U64,
 	V_CMP_LE_U64,
 	V_CMP_GT_U64,
 	V_CMP_NE_U64,
 	V_CMP_GE_U64,
+	V_CMP_T_U64,
+	V_CMPX_F_I64,
+	V_CMPX_LT_I64,
+	V_CMPX_EQ_I64,
+	V_CMPX_LE_I64,
 	V_CMPX_NE_I64,
+	V_CMPX_T_I64,
+	V_CMPX_F_U64,
+	V_CMPX_LT_U64,
+	V_CMPX_EQ_U64,
 	V_CMPX_LE_U64,
+	V_CMPX_GT_U64,
 	V_CMPX_NE_U64,
+	V_CMPX_GE_U64,
+	V_CMPX_T_U64,
 	V_CMPX_LT_U32,
 	V_CMPX_EQ_U32,
 	V_CMPX_LE_U32,
@@ -483,10 +508,12 @@ enum class Opcode {
 	BUFFER_LOAD_FORMAT_XY,
 	BUFFER_LOAD_FORMAT_XYZ,
 	BUFFER_LOAD_FORMAT_XYZW,
+	BUFFER_LOAD_FORMAT_D16_X,
 	BUFFER_STORE_FORMAT_X,
 	BUFFER_STORE_FORMAT_XY,
 	BUFFER_STORE_FORMAT_XYZ,
 	BUFFER_STORE_FORMAT_XYZW,
+	BUFFER_STORE_FORMAT_D16_X,
 	BUFFER_LOAD_UBYTE,
 	BUFFER_LOAD_SBYTE,
 	BUFFER_LOAD_USHORT,
@@ -529,6 +556,7 @@ enum class Opcode {
 	FLAT_LOAD_SBYTE,
 	FLAT_LOAD_USHORT,
 	FLAT_LOAD_SSHORT,
+	FLAT_LOAD_SHORT_D16,
 	FLAT_LOAD_DWORD,
 	FLAT_LOAD_DWORDX2,
 	FLAT_LOAD_DWORDX3,
@@ -572,6 +600,7 @@ enum class Opcode {
 	DS_BPERMUTE_B32,
 	DS_CONSUME,
 	DS_APPEND,
+	DS_ORDERED_COUNT,
 	DS_READ_I8,
 	DS_READ_U8,
 	DS_READ_I16,

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

@@ -35,8 +35,8 @@ constexpr Vop2OpcodeInfo VOP2_OPCODE_LIST[] = {
     {0x04u, Opcode::V_SUB_F32, Vop2SdwaProfile::Float32},
     {0x05u, Opcode::V_SUBREV_F32},
     {0x08u, Opcode::V_MUL_F32, Vop2SdwaProfile::Float32},
-    {0x09u, Opcode::V_MUL_I32_I24, Vop2SdwaProfile::IntegerFullDestination},
-    {0x0bu, Opcode::V_MUL_U32_U24, Vop2SdwaProfile::IntegerFullDestination},
+    {0x09u, Opcode::V_MUL_I32_I24, Vop2SdwaProfile::IntegerPartialDestination},
+    {0x0bu, Opcode::V_MUL_U32_U24, Vop2SdwaProfile::IntegerPartialDestination},
     {0x0fu, Opcode::V_MIN_F32},
     {0x10u, Opcode::V_MAX_F32},
     {0x11u, Opcode::V_MIN_I32},
@@ -65,9 +65,9 @@ constexpr Vop2OpcodeInfo VOP2_OPCODE_LIST[] = {
     {0x2au, Opcode::V_SUBREV_CO_CI_U32},
     {0x26u, Opcode::V_SUB_NC_U32, Vop2SdwaProfile::IntegerPartialDestination},
     {0x27u, Opcode::V_SUBREV_NC_U32, Vop2SdwaProfile::IntegerFullDestination},
-    {0x2bu, Opcode::V_MAC_F32},
-    {0x2cu, Opcode::V_MADMK_F32},
-    {0x2du, Opcode::V_MADAK_F32},
+    {0x2bu, Opcode::V_FMAC_F32},
+    {0x2cu, Opcode::V_FMAMK_F32},
+    {0x2du, Opcode::V_FMAAK_F32},
     {0x2fu, Opcode::V_CVT_PKRTZ_F16_F32, Vop2SdwaProfile::PackedFloat16},
     {0x32u, Opcode::V_ADD_F16, Vop2SdwaProfile::Float16},
     {0x33u, Opcode::V_SUB_F16, Vop2SdwaProfile::Float16},
@@ -232,7 +232,7 @@ constexpr VopcOpcodeInfo VOPC_OPCODE_LIST[] = {
     {0x11u, Opcode::V_CMPX_LT_F32},        {0x12u, Opcode::V_CMPX_EQ_F32},
     {0x13u, Opcode::V_CMPX_LE_F32},        {0x14u, Opcode::V_CMPX_GT_F32},
     {0x15u, Opcode::V_CMPX_LG_F32},        {0x16u, Opcode::V_CMPX_GE_F32},
-    {0x17u, Opcode::V_CMPX_O_F32},
+    {0x17u, Opcode::V_CMPX_O_F32},         {0x18u, Opcode::V_CMPX_U_F32},
     {0x19u, Opcode::V_CMPX_NGE_F32},       {0x1au, Opcode::V_CMPX_NLG_F32},
     {0x1bu, Opcode::V_CMPX_NGT_F32},       {0x1cu, Opcode::V_CMPX_NLE_F32},
     {0x1du, Opcode::V_CMPX_NEQ_F32},       {0x1eu, Opcode::V_CMPX_NLT_F32},
@@ -245,28 +245,39 @@ constexpr VopcOpcodeInfo VOPC_OPCODE_LIST[] = {
     {0x88u, Opcode::V_CMP_CLASS_F32},      {0x89u, Opcode::V_CMP_LT_I16},
     {0x8au, Opcode::V_CMP_EQ_I16},         {0x8bu, Opcode::V_CMP_LE_I16},
     {0x8cu, Opcode::V_CMP_GT_I16},         {0x8du, Opcode::V_CMP_NE_I16},
-    {0x8eu, Opcode::V_CMP_GE_I16},         {0x91u, Opcode::V_CMPX_LT_I32},
+    {0x8eu, Opcode::V_CMP_GE_I16},         {0x8fu, Opcode::V_CMP_CLASS_F16, false},
+    {0x91u, Opcode::V_CMPX_LT_I32},
     {0x92u, Opcode::V_CMPX_EQ_I32},        {0x93u, Opcode::V_CMPX_LE_I32},
     {0x94u, Opcode::V_CMPX_GT_I32},        {0x95u, Opcode::V_CMPX_NE_I32},
     {0x96u, Opcode::V_CMPX_GE_I32},        {0x98u, Opcode::V_CMPX_CLASS_F32},
     {0x99u, Opcode::V_CMPX_LT_I16, false}, {0x9au, Opcode::V_CMPX_EQ_I16, false},
     {0x9bu, Opcode::V_CMPX_LE_I16, false}, {0x9cu, Opcode::V_CMPX_GT_I16, false},
     {0x9du, Opcode::V_CMPX_NE_I16, false}, {0x9eu, Opcode::V_CMPX_GE_I16, false},
+    {0x9fu, Opcode::V_CMPX_CLASS_F16, false},
     {0xa9u, Opcode::V_CMP_LT_U16},         {0xaau, Opcode::V_CMP_EQ_U16},
     {0xabu, Opcode::V_CMP_LE_U16},         {0xacu, Opcode::V_CMP_GT_U16},
     {0xadu, Opcode::V_CMP_NE_U16},         {0xaeu, Opcode::V_CMP_GE_U16},
     {0xb9u, Opcode::V_CMPX_LT_U16, false}, {0xbau, Opcode::V_CMPX_EQ_U16, false},
-    {0xbcu, Opcode::V_CMPX_GT_U16},        {0xc0u, Opcode::V_CMP_F_U32},
+    {0xbcu, Opcode::V_CMPX_GT_U16},        {0xbdu, Opcode::V_CMPX_NE_U16, false},
+    {0xc0u, Opcode::V_CMP_F_U32},
     {0xc1u, Opcode::V_CMP_LT_U32},         {0xc2u, Opcode::V_CMP_EQ_U32},
     {0xc3u, Opcode::V_CMP_LE_U32},         {0xc4u, Opcode::V_CMP_GT_U32},
     {0xc5u, Opcode::V_CMP_NE_U32},         {0xc6u, Opcode::V_CMP_GE_U32},
     {0xc7u, Opcode::V_CMP_T_U32},          {0xa2u, Opcode::V_CMP_EQ_I64, false},
     {0xa1u, Opcode::V_CMP_LT_I64, false},  {0xa3u, Opcode::V_CMP_LE_I64, false},
     {0xa5u, Opcode::V_CMP_NE_I64, false},
+    {0xa0u, Opcode::V_CMP_F_I64, false},   {0xa7u, Opcode::V_CMP_T_I64, false},
+    {0xb0u, Opcode::V_CMPX_F_I64, false},  {0xb1u, Opcode::V_CMPX_LT_I64, false},
+    {0xb2u, Opcode::V_CMPX_EQ_I64, false}, {0xb3u, Opcode::V_CMPX_LE_I64, false},
+    {0xb7u, Opcode::V_CMPX_T_I64, false},
     {0xb5u, Opcode::V_CMPX_NE_I64, false}, {0xd1u, Opcode::V_CMPX_LT_U32},
     {0xd2u, Opcode::V_CMPX_EQ_U32},        {0xd3u, Opcode::V_CMPX_LE_U32},
     {0xd4u, Opcode::V_CMPX_GT_U32},        {0xd5u, Opcode::V_CMPX_NE_U32},
     {0xd6u, Opcode::V_CMPX_GE_U32},        {0xe1u, Opcode::V_CMP_LT_U64, false},
+    {0xe0u, Opcode::V_CMP_F_U64, false},   {0xe7u, Opcode::V_CMP_T_U64, false},
+    {0xf0u, Opcode::V_CMPX_F_U64, false},  {0xf1u, Opcode::V_CMPX_LT_U64, false},
+    {0xf2u, Opcode::V_CMPX_EQ_U64, false}, {0xf4u, Opcode::V_CMPX_GT_U64, false},
+    {0xf6u, Opcode::V_CMPX_GE_U64, false}, {0xf7u, Opcode::V_CMPX_T_U64, false},
     {0xe2u, Opcode::V_CMP_EQ_U64, false},  {0xe3u, Opcode::V_CMP_LE_U64, false},
     {0xe4u, Opcode::V_CMP_GT_U64, false},  {0xe5u, Opcode::V_CMP_NE_U64, false},
     {0xe6u, Opcode::V_CMP_GE_U64, false},  {0xf3u, Opcode::V_CMPX_LE_U64, false},
@@ -308,6 +319,7 @@ constexpr OpcodeMap VOP3_OPCODE_LIST[] = {
     {0x152u, Opcode::V_MIN3_I32},
     {0x153u, Opcode::V_MIN3_U32},
     {0x351u, Opcode::V_MIN3_F16},
+    {0x353u, Opcode::V_MIN3_U16},
     {0x154u, Opcode::V_MAX3_F32},
     {0x155u, Opcode::V_MAX3_I32},
     {0x156u, Opcode::V_MAX3_U32},
@@ -458,8 +470,9 @@ bool IsNativeVop3F16TernaryOpcode(Opcode opcode) {
 	       opcode == Opcode::V_MED3_F16 || opcode == Opcode::V_FMA_F16;
 }
 
-bool IsNativeVop3I16TernaryOpcode(Opcode opcode) {
-	return opcode == Opcode::V_MED3_I16 || opcode == Opcode::V_MAD_I16;
+bool IsNativeVop3B16TernaryOpcode(Opcode opcode) {
+	return opcode == Opcode::V_MIN3_U16 || opcode == Opcode::V_MED3_I16 ||
+	       opcode == Opcode::V_MAD_I16;
 }
 
 bool IsNativeVop3B16BinaryOpcode(Opcode opcode) {
@@ -566,7 +579,7 @@ struct Vop1SdwaRule {
 
 constexpr Vop1SdwaRule VOP1_SDWA_RULES[] = {
     {Opcode::V_MOV_B32, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), SdwaSelBytes() | SdwaSelWords(),
-     SdwaSelWords() | SdwaSelFull(), false},
+     SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), false},
     {Opcode::V_CVT_F32_U32, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), 0, 0, false},
     {Opcode::V_CVT_F32_I32, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), 0, 0, false},
     {Opcode::V_CVT_F32_UBYTE0, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), 0, 0, false},
@@ -800,6 +813,9 @@ bool IsVop2FloatOpcode(Opcode opcode) {
 		case Opcode::V_MAC_F32:
 		case Opcode::V_MADMK_F32:
 		case Opcode::V_MADAK_F32:
+		case Opcode::V_FMAC_F32:
+		case Opcode::V_FMAMK_F32:
+		case Opcode::V_FMAAK_F32:
 		case Opcode::V_CVT_PKRTZ_F16_F32:
 		case Opcode::V_ADD_F16:
 		case Opcode::V_SUB_F16:
@@ -887,6 +903,7 @@ bool IsVopcFloatCompareOpcode(Opcode opcode) {
 		case Opcode::V_CMPX_LG_F32:
 		case Opcode::V_CMPX_GE_F32:
 		case Opcode::V_CMPX_O_F32:
+		case Opcode::V_CMPX_U_F32:
 		case Opcode::V_CMPX_NGE_F32:
 		case Opcode::V_CMPX_NLG_F32:
 		case Opcode::V_CMPX_NGT_F32:
@@ -913,6 +930,8 @@ bool IsVopcFloatCompareOpcode(Opcode opcode) {
 		case Opcode::V_CMPX_NEQ_F16:
 		case Opcode::V_CMPX_NLT_F16:
 		case Opcode::V_CMP_CLASS_F32:
+		case Opcode::V_CMP_CLASS_F16:
+		case Opcode::V_CMPX_CLASS_F16:
 		case Opcode::V_CMPX_CLASS_F32: return true;
 		default: return false;
 	}
@@ -1057,6 +1076,7 @@ void FinalizeVop2Instruction(std::span<const uint32_t> code, uint32_t word_index
                              Instruction& inst) {
 	switch (inst.opcode) {
 		case Opcode::V_MADMK_F32:
+		case Opcode::V_FMAMK_F32:
 		case Opcode::V_FMAMK_F16:
 			inst.src2      = inst.src1;
 			inst.src1      = {};
@@ -1064,6 +1084,7 @@ void FinalizeVop2Instruction(std::span<const uint32_t> code, uint32_t word_index
 			inst.src_count = 3;
 			break;
 		case Opcode::V_MADAK_F32:
+		case Opcode::V_FMAAK_F32:
 		case Opcode::V_FMAAK_F16:
 			inst.src2      = {};
 			inst.src2.kind = OperandKind::LiteralConstant;
@@ -1358,7 +1379,7 @@ void ApplyNativeVop3TernaryModifiers(Instruction& inst, uint32_t op_sel, uint32_
 	inst.dst.sdwa_sel = ((op_sel & 0x8u) != 0) ? 5u : 4u;
 }
 
-void ApplyNativeVop3I16TernarySelectors(Instruction& inst, uint32_t op_sel) {
+void ApplyNativeVop3B16TernarySelectors(Instruction& inst, uint32_t op_sel) {
 	Operand* sources[] = {&inst.src0, &inst.src1, &inst.src2};
 	for (uint32_t i = 0; i < 3u; i++) {
 		sources[i]->op_sel = ((op_sel >> i) & 1u) != 0;
@@ -1401,6 +1422,7 @@ bool SupportsNativeVop3SourceModifiers(Opcode opcode) {
 		case Opcode::V_MIN_F32:
 		case Opcode::V_MAX_F32:
 		case Opcode::V_MAC_F32:
+		case Opcode::V_FMAC_F32:
 		case Opcode::V_MAD_F32:
 		case Opcode::V_ADD_F64:
 		case Opcode::V_MUL_F64:
@@ -1436,9 +1458,13 @@ bool SupportsNativeVop3ResultModifiers(Opcode opcode) {
 		case Opcode::V_MIN_F32:
 		case Opcode::V_MAX_F32:
 		case Opcode::V_MAC_F32:
+		case Opcode::V_FMAC_F32:
 		case Opcode::V_MAD_F32:
 		case Opcode::V_FMA_F32:
 		case Opcode::V_FMA_F16:
+		case Opcode::V_MIN3_F16:
+		case Opcode::V_MAX3_F16:
+		case Opcode::V_MED3_F16:
 		case Opcode::V_CUBEID_F32:
 		case Opcode::V_CVT_PKRTZ_F16_F32:
 		case Opcode::V_LDEXP_F32:
@@ -1469,9 +1495,10 @@ bool HasUnsupportedNativeVop3Modifiers(Opcode opcode, bool permlane, bool mad_mi
 		return clamp != 0u || omod != 0u;
 	}
 	if (IsNativeVop3F16TernaryOpcode(opcode)) {
-		return opcode != Opcode::V_FMA_F16 && (clamp != 0u || omod != 0u);
+		// Clamp is applied by WriteF16 for every f16 ternary; omod only for V_FMA_F16.
+		return opcode != Opcode::V_FMA_F16 && omod != 0u;
 	}
-	if (IsNativeVop3I16TernaryOpcode(opcode)) {
+	if (IsNativeVop3B16TernaryOpcode(opcode)) {
 		return abs != 0u || (clamp != 0u && !clamp_modifier) || omod != 0u || neg != 0u;
 	}
 	if (IsNativeVop3B16BinaryOpcode(opcode)) {
@@ -1519,6 +1546,7 @@ bool IsVopcCompareExec(Opcode opcode) {
 		case Opcode::V_CMPX_LG_F32:
 		case Opcode::V_CMPX_GE_F32:
 		case Opcode::V_CMPX_O_F32:
+		case Opcode::V_CMPX_U_F32:
 		case Opcode::V_CMPX_NGE_F32:
 		case Opcode::V_CMPX_NLG_F32:
 		case Opcode::V_CMPX_NGT_F32:
@@ -1532,15 +1560,27 @@ bool IsVopcCompareExec(Opcode opcode) {
 		case Opcode::V_CMPX_NE_I32:
 		case Opcode::V_CMPX_GE_I32:
 		case Opcode::V_CMPX_CLASS_F32:
+		case Opcode::V_CMPX_CLASS_F16:
 		case Opcode::V_CMPX_LT_U32:
 		case Opcode::V_CMPX_EQ_U32:
 		case Opcode::V_CMPX_LE_U32:
 		case Opcode::V_CMPX_GT_U32:
 		case Opcode::V_CMPX_NE_U32:
 		case Opcode::V_CMPX_GE_U32:
+		case Opcode::V_CMPX_F_I64:
+		case Opcode::V_CMPX_LT_I64:
+		case Opcode::V_CMPX_EQ_I64:
+		case Opcode::V_CMPX_LE_I64:
 		case Opcode::V_CMPX_NE_I64:
+		case Opcode::V_CMPX_T_I64:
+		case Opcode::V_CMPX_F_U64:
+		case Opcode::V_CMPX_LT_U64:
+		case Opcode::V_CMPX_EQ_U64:
 		case Opcode::V_CMPX_LE_U64:
+		case Opcode::V_CMPX_GT_U64:
 		case Opcode::V_CMPX_NE_U64:
+		case Opcode::V_CMPX_GE_U64:
+		case Opcode::V_CMPX_T_U64:
 		case Opcode::V_CMPX_LT_I16:
 		case Opcode::V_CMPX_EQ_I16:
 		case Opcode::V_CMPX_LE_I16:
@@ -1550,6 +1590,7 @@ bool IsVopcCompareExec(Opcode opcode) {
 		case Opcode::V_CMPX_LT_U16:
 		case Opcode::V_CMPX_EQ_U16:
 		case Opcode::V_CMPX_GT_U16:
+		case Opcode::V_CMPX_NE_U16:
 		case Opcode::V_CMPX_LT_F16:
 		case Opcode::V_CMPX_EQ_F16:
 		case Opcode::V_CMPX_LE_F16:
@@ -1694,7 +1735,7 @@ void DecodeVop3(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	const bool vop3b_uses_sdst = carry_in_out || vop3b_carry_out || vop3b_mad_u64;
 	const bool mad_mix         = false;
 	const bool f16_ternary     = IsNativeVop3F16TernaryOpcode(inst.opcode);
-	const bool i16_ternary     = IsNativeVop3I16TernaryOpcode(inst.opcode);
+	const bool b16_ternary     = IsNativeVop3B16TernaryOpcode(inst.opcode);
 	const bool b16_binary      = IsNativeVop3B16BinaryOpcode(inst.opcode);
 	const bool pack_b32_f16    = inst.opcode == Opcode::V_PACK_B32_F16;
 	const bool permlane        = IsPermlaneOpcode(inst.opcode);
@@ -1805,8 +1846,8 @@ void DecodeVop3(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 		ApplyNativeVop3TernaryModifiers(inst, op_sel, abs, neg);
 	} else if (f16_ternary) {
 		ApplyNativeVop3TernaryModifiers(inst, op_sel, abs, neg);
-	} else if (i16_ternary) {
-		ApplyNativeVop3I16TernarySelectors(inst, op_sel);
+	} else if (b16_ternary) {
+		ApplyNativeVop3B16TernarySelectors(inst, op_sel);
 	} else if (b16_binary) {
 		ApplyNativeVop3B16BinaryModifiers(inst, op_sel);
 	} else if (pack_b32_f16) {

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

@@ -31,6 +31,7 @@ void Translator::EmitIntegerCompare(const Decoder::Instruction& inst, IR::ValueO
 	const bool signed_64 = opcode == IR::ValueOpcode::SLessThan64 ||
 	                       opcode == IR::ValueOpcode::SLessThanEqual64 ||
 	                       inst.opcode == Decoder::Opcode::V_CMP_EQ_I64 ||
+	                       inst.opcode == Decoder::Opcode::V_CMPX_EQ_I64 ||
 	                       inst.opcode == Decoder::Opcode::V_CMP_NE_I64 ||
 	                       inst.opcode == Decoder::Opcode::V_CMPX_NE_I64;
 	const auto read = [&](const Decoder::Operand& operand) {
@@ -72,11 +73,13 @@ void Translator::EmitFloatOrderedCompare(const Decoder::Instruction& inst, bool
 	EmitCompareResult(inst, ordered ? ir.LogicalNot(unordered) : unordered, false, cmpx);
 }
 
-void Translator::EmitFloatClassCompare(const Decoder::Instruction& inst, bool cmpx) {
-	const auto value = ReadOperand(inst.src0, IR::Type::F32);
-	const auto mask  = ReadOperand(inst.src1, IR::Type::U32);
-	EmitCompareResult(inst, IR::U1(ir.Emit(IR::ValueOpcode::FPCmpClass32, {value, mask})), false,
-	                  cmpx);
+void Translator::EmitFloatClassCompare(const Decoder::Instruction& inst, bool cmpx, 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);
 }
 
 } // namespace Libs::Graphics::ShaderRecompiler::Frontend

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

@@ -248,6 +248,7 @@ void Translator::S_CSELECT_B32(const Decoder::Instruction& inst) {
 
 void Translator::ScalarSelect64(const Decoder::Instruction& inst,
                                  const Decoder::Operand& false_source) {
+	// Preserve per-word expressions for descriptor tracking and mask provenance.
 	const auto condition     = ir.GetScc();
 	const auto lhs           = ReadU32Pair(inst.src0);
 	const auto rhs           = ReadU32Pair(false_source);

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

@@ -8,13 +8,13 @@ namespace Libs::Graphics::ShaderRecompiler::Frontend {
 void Translator::PackedFloat16(const Decoder::Instruction& inst, IR::ValueOpcode opcode,
                                bool accumulator, bool quiet_snan) {
 	const auto translate_lane = [&](bool high) {
-		const auto lhs = ReadF16LaneAsF32(inst.src0, high, true);
-		const auto rhs = ReadF16LaneAsF32(inst.src1, high, true);
+		const auto lhs = ReadF16LaneAsF32(inst.src0, high);
+		const auto rhs = ReadF16LaneAsF32(inst.src1, high);
 		IR::F32    result;
 		if (accumulator) {
-			result = IR::F32(ir.Emit(opcode, {lhs, rhs, ReadF16LaneAsF32(inst.dst, high, true)}));
+			result = IR::F32(ir.Emit(opcode, {lhs, rhs, ReadF16LaneAsF32(inst.dst, high)}));
 		} else if (inst.src_count == 3u) {
-			result = IR::F32(ir.Emit(opcode, {lhs, rhs, ReadF16LaneAsF32(inst.src2, high, true)}));
+			result = IR::F32(ir.Emit(opcode, {lhs, rhs, ReadF16LaneAsF32(inst.src2, high)}));
 		} else {
 			result = IR::F32(ir.Emit(opcode, {lhs, rhs}));
 		}

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

@@ -31,11 +31,12 @@ void Translator::V_MAD_I16(const Decoder::Instruction& inst) {
 	Write16Bits(DestinationOperand(inst), ir.BitwiseAnd(result, IR::U32(IR::Value(0xffffu))));
 }
 
-void Translator::V_MED3_I16(const Decoder::Instruction& inst) {
+void Translator::Integer16Ternary(const Decoder::Instruction& inst, IR::ValueOpcode opcode,
+                                  bool sign) {
 	const auto result = IR::U32(ir.Emit(
-	    IR::ValueOpcode::SMedTri32, {ReadU16AsU32(inst.src0, true), ReadU16AsU32(inst.src1, true),
-	                                 ReadU16AsU32(inst.src2, true)}));
-	Write16Bits(DestinationOperand(inst), ir.BitwiseAnd(result, IR::U32(IR::Value(0xffffu))));
+	    opcode, {ReadU16AsU32(inst.src0, sign), ReadU16AsU32(inst.src1, sign),
+	             ReadU16AsU32(inst.src2, sign)}));
+	Write16Bits(DestinationOperand(inst), result);
 }
 
 void Translator::PackedInteger16Shift(const Decoder::Instruction& inst, IR::ValueOpcode opcode,
@@ -114,6 +115,7 @@ void Translator::S_U64_MASK(const Decoder::Instruction& inst, IR::ValueOpcode lo
 		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) {
@@ -269,22 +271,18 @@ void Translator::Integer24(const Decoder::Instruction& inst, bool sign, bool add
 }
 
 void Translator::V_MAD_U64_U32(const Decoder::Instruction& inst) {
-	const auto lhs       = ReadU32(inst.src0);
-	const auto rhs       = ReadU32(inst.src1);
-	const auto add       = ExtractU64(ReadU64(inst.src2));
-	const auto mul_low   = ir.IMul(lhs, rhs);
-	const auto mul_high  = IR::U32(ir.Emit(IR::ValueOpcode::UMulHi, {lhs, rhs}));
-	const auto low       = ir.IAdd(mul_low, add[0]);
-	const auto carry_low = ir.ULessThan(low, mul_low);
-	const auto high0     = ir.IAdd(mul_high, add[1]);
-	const auto carry0    = ir.ULessThan(high0, mul_high);
-	const auto high =
-	    ir.IAdd(high0, ir.Select(carry_low, IR::U32(IR::Value(1u)), IR::U32(IR::Value(0u))));
-	const auto carry1 = ir.ULessThan(high, high0);
-	WriteOperand(DestinationOperand(inst), ir.ConstructU64(low, high));
+	const auto lhs = ReadU32(inst.src0);
+	const auto rhs = ReadU32(inst.src1);
+	// Keep the 32x32 product in words: native widening IMul64 fails NVIDIA compilation
+	// in large vertex shaders (Astro's Playroom). The 64-bit addition stays native.
+	const auto product = ir.ConstructU64(
+	    ir.IMul(lhs, rhs), IR::U32(ir.Emit(IR::ValueOpcode::UMulHi, {lhs, rhs})));
+	const auto result  = ir.Emit(IR::ValueOpcode::IAdd64, {product, ReadU64(inst.src2)});
+	WriteOperand(DestinationOperand(inst), result);
 	if (inst.dst2.kind != Decoder::OperandKind::Null &&
 	    inst.dst2.kind != Decoder::OperandKind::Unknown) {
-		WriteMask(inst.dst2, ir.LogicalOr(carry0, carry1));
+		// 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})));
 	}
 }
 
@@ -407,9 +405,12 @@ IR::U64 Translator::RightMask64(IR::U32 count) {
 }
 
 void Translator::S_BFM_B64(const Decoder::Instruction& inst) {
-	const auto count  = ir.BitwiseAnd(ReadU32(inst.src0), IR::U32(IR::Value(63u)));
-	const auto offset = ir.BitwiseAnd(ReadU32(inst.src1), IR::U32(IR::Value(63u)));
-	const auto result = ir.Emit(IR::ValueOpcode::ShiftLeftLogical64, {RightMask64(count), offset});
+	const auto count  = ReadU32(inst.src0);
+	const auto offset = ReadU32(inst.src1);
+	const auto one    = IR::Value(uint64_t {1});
+	const auto limit  = ir.Emit(IR::ValueOpcode::ShiftLeftLogical64, {one, count});
+	const auto mask   = ir.Emit(IR::ValueOpcode::ISub64, {limit, one});
+	const auto result = ir.Emit(IR::ValueOpcode::ShiftLeftLogical64, {mask, offset});
 	WriteOperand(DestinationOperand(inst), result);
 }
 

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

@@ -129,8 +129,9 @@ IR::MemoryInfo MemoryInfoFromDecoded(const Decoder::Instruction& decoded) {
 	// Vector loads use GLC/DLC to bypass L0/GL1; atomics use GLC only to return data.
 	const bool buffer_atomic = decoded.opcode >= Decoder::Opcode::BUFFER_ATOMIC_SWAP &&
 	                           decoded.opcode <= Decoder::Opcode::BUFFER_ATOMIC_FMAX;
-	memory.coherent = memory.kind == ResourceKind::Buffer && !buffer_atomic &&
-	                  (decoded.glc || decoded.dlc);
+	memory.coherent = (memory.kind == ResourceKind::Buffer || memory.kind == ResourceKind::Flat ||
+	                   memory.kind == ResourceKind::Global) &&
+	                  !buffer_atomic && (decoded.glc || decoded.dlc);
 	memory.resource      = ResourceIndexFromOperand(decoded.src1);
 	memory.sampler       = ResourceIndexFromOperand(decoded.src2);
 	if (memory.kind == ResourceKind::ScalarBuffer) {
@@ -186,7 +187,7 @@ Decoder::Operand MemorySourceAt(const Decoder::Instruction& decoded, uint32_t in
 	if (decoded.family == Decoder::Family::MUBUF || decoded.family == Decoder::Family::MTBUF) {
 		const bool store_or_atomic =
 		    (decoded.opcode >= Decoder::Opcode::BUFFER_STORE_FORMAT_X &&
-		     decoded.opcode <= Decoder::Opcode::BUFFER_STORE_FORMAT_XYZW) ||
+		     decoded.opcode <= Decoder::Opcode::BUFFER_STORE_FORMAT_D16_X) ||
 		    (decoded.opcode >= Decoder::Opcode::BUFFER_STORE_BYTE &&
 		     decoded.opcode <= Decoder::Opcode::BUFFER_STORE_DWORDX4) ||
 		    (decoded.opcode >= Decoder::Opcode::TBUFFER_STORE_FORMAT_X &&
@@ -426,7 +427,7 @@ void Translator::S_LOAD(const Decoder::Instruction& inst, bool raw) {
 void Translator::BUFFER_LOAD(const Decoder::Instruction& inst) {
 	const auto      memory = MemoryInfoFromDecoded(inst);
 	IR::ValueOpcode opcode;
-	const auto      bits = memory.data_bits;
+	const auto      bits = memory.formatted ? 32u : memory.data_bits;
 	const auto      sign = memory.data_signed;
 	switch (bits) {
 		case 8u: opcode = IR::ValueOpcode::LoadBufferU8; break;
@@ -472,7 +473,8 @@ void Translator::BUFFER_STORE(const Decoder::Instruction& inst) {
 	const auto      data     = ReadU32(data_src);
 	IR::ValueOpcode opcode;
 	IR::Value       value;
-	switch (memory.data_bits) {
+	// Formatted stores carry packed VGPR bits; the format determines their conversion.
+	switch (memory.formatted ? 32u : memory.data_bits) {
 		case 8u:
 			opcode = IR::ValueOpcode::StoreBufferU8;
 			value  = NarrowSubdword(data, 8u);
@@ -546,7 +548,10 @@ void Translator::BUFFER_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode
 	}
 }
 
-void Translator::IMAGE_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode opcode) {
+void Translator::IMAGE_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode opcode32,
+                               IR::ValueOpcode opcode64) {
+	const auto opcode = inst.data_bits == 64u ? opcode64 : opcode32;
+	EXIT_IF(opcode == IR::ValueOpcode::Count);
 	const auto memory   = MemoryInfoFromDecoded(inst);
 	const auto resource = GetImageResource(memory);
 	const auto address  = MakeImageAddress(inst, MemorySourceAt(inst, 1));
@@ -879,6 +884,50 @@ void Translator::DS_SWIZZLE_B32(const Decoder::Instruction& inst) {
 	                                ReadU32(MemorySourceAt(inst, 1)), ir.GetExec()}));
 }
 
+void Translator::DS_ORDERED_COUNT(const Decoder::Instruction& inst) {
+	// Approximate ordered counting with one GDS atomic per wave. Wave-launch ordering and
+	// release/done synchronization are not emulated.
+	const auto memory = MemoryInfoFromDecoded(inst);
+	const auto m0 = ir.GetM0();
+	auto address = ir.BitwiseAnd(ir.ShiftRightLogical(m0, IR::U32(IR::Value(16u))),
+	                             IR::U32(IR::Value(0xfffcu)));
+	if (((inst.secondary_offset >> 2u) & 3u) == 1u) {
+		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 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))),
+	    IR::U32(ir.Emit(IR::ValueOpcode::FindILsb32, {low})),
+	    ir.IAdd(IR::U32(IR::Value(32u)), IR::U32(ir.Emit(IR::ValueOpcode::FindILsb32, {high}))));
+	// An empty EXEC still returns the counter. Elect lane zero and perform an add of zero.
+	const auto selected = ir.Select(active, first, IR::U32(IR::Value(0u)));
+	const auto lane = IR::U32(ir.Emit(IR::ValueOpcode::LaneId));
+	const auto is_first = ir.IEqual(lane, selected);
+	const auto source = IR::U32(ir.Emit(IR::ValueOpcode::ReadFirstLane,
+	                                   {ReadU32(inst.src0), is_first}));
+	const auto value = ir.Select(active, source, IR::U32(IR::Value(0u)));
+	const auto flags = AddMemoryInfo(memory, inst.pc);
+	IR::U32 result;
+	if (((inst.secondary_offset >> 4u) & 3u) == 1u) {
+		const auto swapped = IR::U32(ir.Emit(IR::ValueOpcode::SharedAtomicSwap32,
+		    {address, value, ir.LogicalAnd(is_first, active)}, flags));
+		const auto unchanged = IR::U32(ir.Emit(IR::ValueOpcode::SharedAtomicIAdd32,
+		    {address, IR::Value(0u), ir.LogicalAnd(is_first, ir.LogicalNot(active))}, flags));
+		result = ir.Select(active, swapped, unchanged);
+	} else {
+		result = IR::U32(ir.Emit(IR::ValueOpcode::SharedAtomicIAdd32,
+		                        {address, value, is_first}, flags));
+	}
+	// DS_ORDERED_COUNT writes every destination lane, regardless of EXEC.
+	ir.SetVectorReg(static_cast<IR::VectorReg>(inst.dst.reg),
+	    IR::U32(ir.Emit(IR::ValueOpcode::ReadFirstLane, {result, is_first})));
+}
+
 void Translator::DS_PERMUTE(const Decoder::Instruction& inst, bool backward) {
 	const auto address = ir.IAdd(ReadU32(inst.src0), IR::U32(IR::Value(inst.offset)));
 	WriteOperand(inst.dst, ir.Emit(backward ? IR::ValueOpcode::BpermuteU32
@@ -932,6 +981,7 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::BUFFER_LOAD_FORMAT_XY:
 		case Decoder::Opcode::BUFFER_LOAD_FORMAT_XYZ:
 		case Decoder::Opcode::BUFFER_LOAD_FORMAT_XYZW:
+		case Decoder::Opcode::BUFFER_LOAD_FORMAT_D16_X:
 		case Decoder::Opcode::TBUFFER_LOAD_FORMAT_X:
 		case Decoder::Opcode::TBUFFER_LOAD_FORMAT_XY:
 		case Decoder::Opcode::TBUFFER_LOAD_FORMAT_XYZ:
@@ -947,6 +997,7 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::BUFFER_STORE_FORMAT_XY:
 		case Decoder::Opcode::BUFFER_STORE_FORMAT_XYZ:
 		case Decoder::Opcode::BUFFER_STORE_FORMAT_XYZW:
+		case Decoder::Opcode::BUFFER_STORE_FORMAT_D16_X:
 		case Decoder::Opcode::TBUFFER_STORE_FORMAT_X:
 		case Decoder::Opcode::TBUFFER_STORE_FORMAT_XY:
 		case Decoder::Opcode::TBUFFER_STORE_FORMAT_XYZ:
@@ -1041,24 +1092,30 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::IMAGE_ATOMIC_CMPSWAP:
 			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicCompareSwap32);
 		case Decoder::Opcode::IMAGE_ATOMIC_SWAP:
-			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicSwap32);
+			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicSwap32,
+			                    IR::ValueOpcode::ImageAtomicSwap64);
 		case Decoder::Opcode::IMAGE_ATOMIC_ADD:
-			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicIAdd32);
+			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicIAdd32,
+			                    IR::ValueOpcode::ImageAtomicIAdd64);
 		case Decoder::Opcode::IMAGE_ATOMIC_SMIN:
 			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicSMin32);
 		case Decoder::Opcode::IMAGE_ATOMIC_UMIN:
-			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicUMin32);
+			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicUMin32,
+			                    IR::ValueOpcode::ImageAtomicUMin64);
 		case Decoder::Opcode::IMAGE_ATOMIC_SMAX:
 			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicSMax32);
 		case Decoder::Opcode::IMAGE_ATOMIC_UMAX:
-			return IMAGE_ATOMIC(inst, inst.data_bits == 64u ? IR::ValueOpcode::ImageAtomicUMax64
-			                                                : IR::ValueOpcode::ImageAtomicUMax32);
+			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicUMax32,
+			                    IR::ValueOpcode::ImageAtomicUMax64);
 		case Decoder::Opcode::IMAGE_ATOMIC_AND:
-			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicAnd32);
+			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicAnd32,
+			                    IR::ValueOpcode::ImageAtomicAnd64);
 		case Decoder::Opcode::IMAGE_ATOMIC_OR:
-			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicOr32);
+			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicOr32,
+			                    IR::ValueOpcode::ImageAtomicOr64);
 		case Decoder::Opcode::IMAGE_ATOMIC_XOR:
-			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicXor32);
+			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicXor32,
+			                    IR::ValueOpcode::ImageAtomicXor64);
 		case Decoder::Opcode::IMAGE_ATOMIC_FMIN:
 			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicFMin32);
 		case Decoder::Opcode::IMAGE_ATOMIC_FMAX:
@@ -1068,6 +1125,7 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::FLAT_LOAD_SBYTE:
 		case Decoder::Opcode::FLAT_LOAD_USHORT:
 		case Decoder::Opcode::FLAT_LOAD_SSHORT:
+		case Decoder::Opcode::FLAT_LOAD_SHORT_D16:
 		case Decoder::Opcode::FLAT_LOAD_DWORD:
 		case Decoder::Opcode::FLAT_LOAD_DWORDX2:
 		case Decoder::Opcode::FLAT_LOAD_DWORDX3:
@@ -1103,6 +1161,7 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::DS_SWIZZLE_B32: return DS_SWIZZLE_B32(inst);
 		case Decoder::Opcode::DS_PERMUTE_B32: return DS_PERMUTE(inst, false);
 		case Decoder::Opcode::DS_BPERMUTE_B32: return DS_PERMUTE(inst, true);
+		case Decoder::Opcode::DS_ORDERED_COUNT: return DS_ORDERED_COUNT(inst);
 		case Decoder::Opcode::DS_CONSUME:
 			return DS_APPEND_CONSUME(inst, IR::ValueOpcode::DataConsume);
 		case Decoder::Opcode::DS_APPEND:

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

@@ -134,6 +134,14 @@ void Translator::EmitScalar(const Decoder::Instruction& inst) {
 
 		case O::S_ABS_I32:
 			return SimpleInteger(inst, IR::ValueOpcode::IAbs32, IR::Type::U32, false, false, true);
+		case O::S_SEXT_I32_I8:
+		case O::S_SEXT_I32_I16: {
+			// Sign extension leaves SCC unchanged.
+			const auto bits = inst.opcode == O::S_SEXT_I32_I8 ? 8u : 16u;
+			WriteOperand(inst.dst, IR::U32(ir.Emit(IR::ValueOpcode::BitFieldSExtract,
+			    {ReadU32(inst.src0), IR::Value(0u), IR::Value(bits)})));
+			return;
+		}
 		case O::S_MUL_I32:
 		case O::S_MULK_I32:
 			return SimpleInteger(inst, IR::ValueOpcode::IMul32, IR::Type::U32, false, false, false);
@@ -165,9 +173,16 @@ void Translator::EmitScalar(const Decoder::Instruction& inst) {
 		case O::S_BCNT1_I32_B32:
 			return SimpleInteger(inst, IR::ValueOpcode::BitCount32, IR::Type::U32, false, false,
 			                     true);
-		case O::S_BCNT1_I32_B64:
-			return SimpleInteger(inst, IR::ValueOpcode::BitCount64, IR::Type::U64, false, false,
-			                     true);
+		case O::S_BCNT1_I32_B64: {
+			// Vulkan bit counts operate on 32-bit words; avoid packing only to split again.
+			const auto value  = ReadU32Pair(inst.src0);
+			const auto low    = IR::U32(ir.Emit(IR::ValueOpcode::BitCount32, {value[0]}));
+			const auto high   = IR::U32(ir.Emit(IR::ValueOpcode::BitCount32, {value[1]}));
+			const auto result = ir.IAdd(low, high);
+			WriteOperand(inst.dst, result);
+			ir.SetScc(ir.INotEqual(result, IR::U32(IR::Value(0u))));
+			return;
+		}
 		case O::S_FF1_I32_B32:
 			return SimpleInteger(inst, IR::ValueOpcode::FindILsb32, IR::Type::U32, false, false,
 			                     false);

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

@@ -503,41 +503,11 @@ IR::U64 Translator::ReadU64(const Decoder::Operand& operand) {
 	return IR::U64(ReadOperand(operand, IR::Type::U64));
 }
 
-IR::F32 Translator::ReadF16LaneAsF32(const Decoder::Operand& operand, bool high_lane, bool packed) {
-	if (operand.kind == Decoder::OperandKind::FloatInlineConstant) {
-		const bool use_zero = packed && (high_lane ? operand.op_sel_hi : operand.op_sel);
-		auto       value    = use_zero ? IR::F32(IR::Value::F32(0.0f))
-		                               : ir.BitCastF32(IR::U32(IR::Value(operand.value)));
-		const auto half     = IR::F16(ir.Emit(IR::ValueOpcode::ConvertF16F32, {value}));
-		value               = IR::F32(ir.Emit(IR::ValueOpcode::ConvertF32F16, {half}));
-		if (operand.absolute) {
-			value = IR::F32(ir.Emit(IR::ValueOpcode::FPAbs32, {value}));
-		}
-		if (high_lane ? operand.negate_hi : operand.negate) {
-			value = IR::F32(ir.Emit(IR::ValueOpcode::FPNeg32, {value}));
-		}
-		return value;
-	}
-	auto raw_operand      = operand;
-	raw_operand.sdwa_sel  = 6;
-	raw_operand.sdwa_sext = false;
-	const auto bits       = ApplyBitSourceModifiers(raw_operand, ReadRawU32(operand));
-	uint32_t   offset     = (high_lane ? operand.op_sel_hi : operand.op_sel) ? 16u : 0u;
-	if (operand.sdwa_sel == 4u || operand.sdwa_sel == 5u) {
-		offset = operand.sdwa_sel == 5u ? 16u : 0u;
-	}
-	const auto half_u32 = IR::U32(
-	    ir.Emit(IR::ValueOpcode::BitFieldUExtract, {bits, IR::Value(offset), IR::Value(16u)}));
-	const auto half_u16 = IR::U16(ir.Emit(IR::ValueOpcode::ConvertU16U32, {half_u32}));
+IR::F32 Translator::ReadF16LaneAsF32(const Decoder::Operand& operand, bool high_lane) {
+	const auto bits     = Read16LaneBits(operand, high_lane);
+	const auto half_u16 = IR::U16(ir.Emit(IR::ValueOpcode::ConvertU16U32, {bits}));
 	const auto half     = IR::F16(ir.Emit(IR::ValueOpcode::BitCastF16U16, {half_u16}));
-	auto       value    = IR::F32(ir.Emit(IR::ValueOpcode::ConvertF32F16, {half}));
-	if (operand.absolute) {
-		value = IR::F32(ir.Emit(IR::ValueOpcode::FPAbs32, {value}));
-	}
-	if (high_lane ? operand.negate_hi : operand.negate) {
-		value = IR::F32(ir.Emit(IR::ValueOpcode::FPNeg32, {value}));
-	}
-	return value;
+	return IR::F32(ir.Emit(IR::ValueOpcode::ConvertF32F16, {half}));
 }
 
 IR::F32 Translator::ReadF16AsF32(const Decoder::Operand& operand) {
@@ -555,29 +525,6 @@ IR::F32 Translator::ReadMixF32(const Decoder::Operand& operand) {
 	return IR::F32(ReadOperand(value_operand, IR::Type::F32));
 }
 
-IR::U32 Translator::ReadU16LaneRaw(const Decoder::Operand& operand, bool high_lane) {
-	auto raw_operand      = operand;
-	raw_operand.sdwa_sel  = 6;
-	raw_operand.sdwa_sext = false;
-	const auto bits       = ApplyBitSourceModifiers(raw_operand, ReadRawU32(operand));
-	uint32_t   offset     = (high_lane ? operand.op_sel_hi : operand.op_sel) ? 16u : 0u;
-	uint32_t   width      = 16u;
-	if (operand.sdwa_sel <= 3u) {
-		offset = operand.sdwa_sel * 8u;
-		width  = 8u;
-	} else if (operand.sdwa_sel == 4u || operand.sdwa_sel == 5u) {
-		offset = operand.sdwa_sel == 5u ? 16u : 0u;
-	}
-	if (width == 8u && operand.sdwa_sext) {
-		// Extend the selected byte into the 16-bit operand before halfword modifiers.
-		const auto value = IR::U32(
-		    ir.Emit(IR::ValueOpcode::BitFieldSExtract, {bits, IR::Value(offset), IR::Value(width)}));
-		return ir.BitwiseAnd(value, IR::U32(IR::Value(0xffffu)));
-	}
-	return IR::U32(
-	    ir.Emit(IR::ValueOpcode::BitFieldUExtract, {bits, IR::Value(offset), IR::Value(width)}));
-}
-
 IR::U32 Translator::ReadU16LaneAsU32(const Decoder::Operand& operand, bool high_lane,
                                      bool sign_extend) {
 	auto value = Read16LaneBits(operand, high_lane);
@@ -593,12 +540,36 @@ IR::U32 Translator::ReadU16AsU32(const Decoder::Operand& operand, bool sign_exte
 }
 
 IR::U32 Translator::Read16LaneBits(const Decoder::Operand& operand, bool high_lane) {
-	auto value = ReadU16LaneRaw(operand, high_lane);
+	auto raw_operand      = operand;
+	raw_operand.sdwa_sel  = 6;
+	raw_operand.sdwa_sext = false;
+	auto bits             = ReadRawU32(operand);
+	if (operand.kind == Decoder::OperandKind::FloatInlineConstant) {
+		// Inline floats encode the operand's width before SDWA or OP_SEL selects bits.
+		const auto half      = IR::F16(ir.Emit(IR::ValueOpcode::ConvertF16F32, {ir.BitCastF32(bits)}));
+		const auto half_bits = IR::U16(ir.Emit(IR::ValueOpcode::BitCastU16F16, {half}));
+		bits                = IR::U32(ir.Emit(IR::ValueOpcode::ConvertU32U16, {half_bits}));
+	}
+	bits = ApplyBitSourceModifiers(raw_operand, bits);
+	uint32_t offset = (high_lane ? operand.op_sel_hi : operand.op_sel) ? 16u : 0u;
+	uint32_t width  = 16u;
+	if (operand.sdwa_sel <= 3u) {
+		offset = operand.sdwa_sel * 8u;
+		width  = 8u;
+	} else if (operand.sdwa_sel == 4u || operand.sdwa_sel == 5u) {
+		offset = operand.sdwa_sel == 5u ? 16u : 0u;
+	}
+	const auto extract = width == 8u && operand.sdwa_sext
+	                         ? IR::ValueOpcode::BitFieldSExtract
+	                         : IR::ValueOpcode::BitFieldUExtract;
+	auto value = IR::U32(ir.Emit(extract, {bits, IR::Value(offset), IR::Value(width)}));
+	if (width == 8u && operand.sdwa_sext) {
+		value = ir.BitwiseAnd(value, IR::U32(IR::Value(0xffffu)));
+	}
 	if (operand.absolute) {
 		value = ir.BitwiseAnd(value, IR::U32(IR::Value(0x7fffu)));
 	}
 	if (high_lane ? operand.negate_hi : operand.negate) {
-		// RDNA2 source NEG flips the sign bit, including packed integer operands.
 		value = ir.BitwiseXor(value, IR::U32(IR::Value(0x8000u)));
 	}
 	return value;
@@ -918,9 +889,16 @@ void IncludeInstructionVectorRegisters(const Decoder::Instruction& inst, uint32_
 		case Decoder::Opcode::V_CMP_GT_U64:
 		case Decoder::Opcode::V_CMP_NE_U64:
 		case Decoder::Opcode::V_CMP_GE_U64:
+		case Decoder::Opcode::V_CMPX_LT_I64:
+		case Decoder::Opcode::V_CMPX_EQ_I64:
+		case Decoder::Opcode::V_CMPX_LE_I64:
 		case Decoder::Opcode::V_CMPX_NE_I64:
+		case Decoder::Opcode::V_CMPX_LT_U64:
+		case Decoder::Opcode::V_CMPX_EQ_U64:
 		case Decoder::Opcode::V_CMPX_LE_U64:
+		case Decoder::Opcode::V_CMPX_GT_U64:
 		case Decoder::Opcode::V_CMPX_NE_U64:
+		case Decoder::Opcode::V_CMPX_GE_U64:
 			include_vector(inst.src0, 2u);
 			include_vector(inst.src1, 2u);
 			break;
@@ -1159,7 +1137,8 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 			}
 		} else if (options.stage == ShaderType::Mesh) {
 			const auto& mesh = options.input_info.vertex->mesh;
-			EXIT_NOT_IMPLEMENTED(mesh.primitives_per_group == 0u || mesh.vertices_per_group > 64u ||
+			EXIT_NOT_IMPLEMENTED(mesh.primitives_per_group == 0u || mesh.vertices_per_group > 0x1ffu ||
+			                     mesh.primitives_per_group > 0x1ffu ||
 			                     total_threads > 15u * options.wave_size);
 			const auto u32  = [](uint32_t value) { return IR::U32(IR::Value(value)); };
 			const auto draw = [&](uint32_t index) {
@@ -1195,6 +1174,10 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 			const auto wave_info = entry_ir.BitwiseOr(entry_ir.ShiftLeftLogical(wave, u32(24)),
 			                                          u32(((total_threads + options.wave_size - 1u) /
 			                                               options.wave_size) << 28u));
+			entry_ir.SetScalarReg(
+			    static_cast<IR::ScalarReg>(2),
+			    entry_ir.BitwiseOr(entry_ir.ShiftLeftLogical(vertices, u32(12)),
+			                       entry_ir.ShiftLeftLogical(primitives, u32(22))));
 			entry_ir.SetScalarReg(
 			    static_cast<IR::ScalarReg>(3),
 			    entry_ir.BitwiseOr(wave_info, entry_ir.BitwiseOr(entry_ir.ShiftLeftLogical(
@@ -1303,6 +1286,8 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 				}
 			};
 			barycentric_pair(ps->ps_perspective_center_vgpr, IR::StageInputKind::BaryCoordSmooth);
+			barycentric_pair(ps->ps_perspective_sample_vgpr,
+			                 IR::StageInputKind::BaryCoordSmoothSample);
 			barycentric_pair(ps->ps_perspective_centroid_vgpr,
 			                 IR::StageInputKind::BaryCoordSmoothCentroid);
 			uint32_t reg = ps->ps_system_input_base;

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

@@ -40,10 +40,9 @@ private:
 	IR::U32                ReadU32(const Decoder::Operand& operand);
 	std::array<IR::U32, 2> ReadU32Pair(const Decoder::Operand& operand);
 	IR::U64                ReadU64(const Decoder::Operand& operand);
-	IR::F32 ReadF16LaneAsF32(const Decoder::Operand& operand, bool high_lane, bool packed = false);
+	IR::F32 ReadF16LaneAsF32(const Decoder::Operand& operand, bool high_lane);
 	IR::F32 ReadF16AsF32(const Decoder::Operand& operand);
 	IR::F32 ReadMixF32(const Decoder::Operand& operand);
-	IR::U32 ReadU16LaneRaw(const Decoder::Operand& operand, bool high_lane);
 	IR::U32 ReadU16LaneAsU32(const Decoder::Operand& operand, bool high_lane, bool sign_extend);
 	IR::U32 ReadU16AsU32(const Decoder::Operand& operand, bool sign_extend);
 	IR::U32 Read16LaneBits(const Decoder::Operand& operand, bool high_lane);
@@ -86,7 +85,8 @@ private:
 	void          BUFFER_LOAD(const Decoder::Instruction& inst);
 	void          BUFFER_STORE(const Decoder::Instruction& inst);
 	void          BUFFER_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode opcode);
-	void          IMAGE_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode opcode);
+	void          IMAGE_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode opcode32,
+	                           IR::ValueOpcode opcode64 = IR::ValueOpcode::Count);
 	void DS_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool returns_value);
 	void FLAT_LOAD(const Decoder::Instruction& inst);
 	void FLAT_STORE(const Decoder::Instruction& inst);
@@ -109,6 +109,7 @@ private:
 	void DS_ADDTID(const Decoder::Instruction& inst, bool write);
 	void DS_SWIZZLE_B32(const Decoder::Instruction& inst);
 	void DS_PERMUTE(const Decoder::Instruction& inst, bool backward);
+	void DS_ORDERED_COUNT(const Decoder::Instruction& inst);
 
 	IR::F32 SelectF32(IR::U1 condition, IR::F32 true_value, IR::F32 false_value);
 	IR::U32 ConvertF32ToU32Saturated(IR::F32 value, float upper_bound, float safe_upper,
@@ -126,7 +127,7 @@ private:
 	void EmitFloatCompare(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool half,
 	                      bool cmpx);
 	void EmitFloatOrderedCompare(const Decoder::Instruction& inst, bool ordered, bool cmpx);
-	void EmitFloatClassCompare(const Decoder::Instruction& inst, bool cmpx);
+	void EmitFloatClassCompare(const Decoder::Instruction& inst, bool cmpx, 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);
@@ -165,8 +166,8 @@ private:
 	void FloatCube(const Decoder::Instruction& inst, uint32_t result_kind);
 	void Integer16Shift(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool arithmetic);
 	void Integer16Binary(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool sign);
+	void Integer16Ternary(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool sign);
 	void V_MAD_I16(const Decoder::Instruction& inst);
-	void V_MED3_I16(const Decoder::Instruction& inst);
 	void PackedInteger16Shift(const Decoder::Instruction& inst, IR::ValueOpcode opcode,
 	                          bool arithmetic);
 	void PackedInteger16Binary(const Decoder::Instruction& inst, IR::ValueOpcode opcode);

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

@@ -96,16 +96,34 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_CMPX_LE_I32:
 			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThanEqual32, IR::Type::U32, false, true);
 			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_CMPX_F_I64:
+		case O::V_CMPX_F_U64: EmitCompareConstant(inst, false, false, true); return;
+		case O::V_CMPX_T_I64:
+		case O::V_CMPX_T_U64: EmitCompareConstant(inst, true, false, true); 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);
+			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);
+			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);
+			return;
 		case O::V_CMP_NE_I64:
 		case O::V_CMP_NE_U64:
 			EmitIntegerCompare(inst, IR::ValueOpcode::INotEqual64, IR::Type::U64, false, false);
@@ -118,9 +136,15 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		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);
+			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);
+			return;
 		case O::V_CMP_LE_U64:
 			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThanEqual64, IR::Type::U64, false, false);
 			return;
@@ -130,6 +154,9 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		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);
+			return;
 
 		case O::V_CMP_EQ_U16:
 			EmitInteger16Compare(inst, IR::ValueOpcode::IEqual32, false, false);
@@ -143,6 +170,9 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		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);
+			return;
 		case O::V_CMP_NE_I16:
 			EmitInteger16Compare(inst, IR::ValueOpcode::INotEqual32, true, false);
 			return;
@@ -341,8 +371,11 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		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_CVT_F32_UBYTE0: V_CVT_F32_UBYTE(inst, 0); return;
 		case O::V_CVT_F32_UBYTE1: V_CVT_F32_UBYTE(inst, 1); return;
@@ -390,7 +423,8 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_SUB_NC_I16: return Integer16Binary(inst, IR::ValueOpcode::ISub32, false);
 		case O::V_MUL_LO_U16: return Integer16Binary(inst, IR::ValueOpcode::IMul32, false);
 		case O::V_MAD_I16: return V_MAD_I16(inst);
-		case O::V_MED3_I16: return V_MED3_I16(inst);
+		case O::V_MED3_I16: return Integer16Ternary(inst, IR::ValueOpcode::SMedTri32, true);
+		case O::V_MIN3_U16: return Integer16Ternary(inst, IR::ValueOpcode::UMinTri32, false);
 		case O::V_MIN_I16: return Integer16Binary(inst, IR::ValueOpcode::SMin32, true);
 		case O::V_MAX_I16: return Integer16Binary(inst, IR::ValueOpcode::SMax32, true);
 		case O::V_MIN_U16: return Integer16Binary(inst, IR::ValueOpcode::UMin32, false);
@@ -479,10 +513,13 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_MIN_F32: return FloatBinary(inst, IR::ValueOpcode::FPMin32, false);
 		case O::V_MAX_F32: return FloatBinary(inst, IR::ValueOpcode::FPMax32, false);
 		case O::V_LDEXP_F32: return FloatBinary(inst, IR::ValueOpcode::FPLdexp, false);
-		case O::V_MAC_F32: return FloatTernary(inst, IR::ValueOpcode::FPFma32, true, true);
+		case O::V_MAC_F32: return FloatTernary(inst, IR::ValueOpcode::FPMad32, true, true);
 		case O::V_MADMK_F32:
 		case O::V_MADAK_F32:
-		case O::V_MAD_F32:
+		case O::V_MAD_F32: return FloatTernary(inst, IR::ValueOpcode::FPMad32, false, true);
+		case O::V_FMAC_F32: return FloatTernary(inst, IR::ValueOpcode::FPFma32, true, true);
+		case O::V_FMAMK_F32:
+		case O::V_FMAAK_F32:
 		case O::V_FMA_F32: return FloatTernary(inst, IR::ValueOpcode::FPFma32, false, true);
 		case O::V_FMA_F64: return FloatTernary(inst, IR::ValueOpcode::FPFma64, false, false);
 		case O::V_MIN3_F32: return FloatTernary(inst, IR::ValueOpcode::FPMinTri32, false, false);

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

@@ -196,6 +196,7 @@ enum class StageInputKind {
 	Layer,
 	SampleId,
 	BaryCoordSmooth,
+	BaryCoordSmoothSample,
 	BaryCoordSmoothCentroid,
 	BaryCoordNoPerspective,
 	WorkgroupId,
@@ -485,14 +486,21 @@ struct BlockInfo {
 
 struct DescriptorSource {
 	struct IndirectDescriptor {
-		uint32_t material_source = UINT32_MAX;
+		struct SelectorRead {
+			uint32_t source = UINT32_MAX;
+			uint32_t stride = 0;
+			uint32_t offset = 0;
+
+			bool operator==(const SelectorRead& other) const = default;
+		};
+		std::optional<SelectorRead> selector;
 		uint32_t table_source    = 0;
-		uint32_t selector_stride = 0;
-		uint32_t selector_offset = 0;
 		uint32_t table_offset    = 0;
+		uint32_t table_immediate = 0;
 		uint32_t table_stride    = 0;
+		uint32_t table_record_bytes = 0;
+		bool     table_scalar = false;
 		uint32_t workgroup_axis  = UINT32_MAX;
-		uint32_t selector_shift  = 0;
 		Value    key_count;
 		Value                 selector_first;
 		Value    selector_mask;

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

@@ -224,15 +224,21 @@ ImageOpcodeInfo ImageOpcodeInfoOf(ValueOpcode opcode) {
 			return {ImageAccess::Write, ImageResourceClass::Storage, false};
 		case ValueOpcode::ImageAtomicCompareSwap32:
 		case ValueOpcode::ImageAtomicSwap32:
+		case ValueOpcode::ImageAtomicSwap64:
 		case ValueOpcode::ImageAtomicIAdd32:
+		case ValueOpcode::ImageAtomicIAdd64:
 		case ValueOpcode::ImageAtomicSMin32:
 		case ValueOpcode::ImageAtomicUMin32:
+		case ValueOpcode::ImageAtomicUMin64:
 		case ValueOpcode::ImageAtomicSMax32:
 		case ValueOpcode::ImageAtomicUMax32:
 		case ValueOpcode::ImageAtomicUMax64:
 		case ValueOpcode::ImageAtomicAnd32:
+		case ValueOpcode::ImageAtomicAnd64:
 		case ValueOpcode::ImageAtomicOr32:
+		case ValueOpcode::ImageAtomicOr64:
 		case ValueOpcode::ImageAtomicXor32:
+		case ValueOpcode::ImageAtomicXor64:
 		case ValueOpcode::ImageAtomicFMin32:
 		case ValueOpcode::ImageAtomicFMax32:
 			return {ImageAccess::Atomic, ImageResourceClass::Storage, false};

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

@@ -175,10 +175,12 @@ VALUE_OPCODE(FPOrdGreaterThanEqual32, U1, F32, F32)
 VALUE_OPCODE(FPUnordGreaterThanEqual32, U1, F32, F32)
 VALUE_OPCODE(FPIsNan32, U1, F32)
 VALUE_OPCODE(FPCmpClass32, U1, F32, U32)
+VALUE_OPCODE(FPCmpClass16, U1, U32, U32)
 
 VALUE_OPCODE(FPAdd32, F32, F32, F32)
 VALUE_OPCODE(FPSub32, F32, F32, F32)
 VALUE_OPCODE(FPFma32, F32, F32, F32, F32)
+VALUE_OPCODE(FPMad32, F32, F32, F32, F32)
 VALUE_OPCODE(FPMul32, F32, F32, F32)
 VALUE_OPCODE(FPAdd64, F64, F64, F64)
 VALUE_OPCODE(FPMul64, F64, F64, F64)
@@ -279,8 +281,8 @@ VALUE_OPCODE(WriteSharedU32, Void, U32, U32, U1)
 VALUE_OPCODE(WriteSharedU32x2, Void, U32, U32, U32, U1)
 VALUE_OPCODE(WriteSharedU32x3, Void, U32, U32, U32, U32, U1)
 VALUE_OPCODE(WriteSharedU32x4, Void, U32, U32, U32, U32, U32, U1)
-VALUE_OPCODE(SharedAtomicFMin32, Void, U32, U32, U32, U1)
-VALUE_OPCODE(SharedAtomicFMax32, Void, U32, U32, U32, U1)
+VALUE_OPCODE(SharedAtomicFMin32, Void, U32, U32, U1)
+VALUE_OPCODE(SharedAtomicFMax32, Void, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicSwap32, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicIAdd32, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicIAdd64, Void, U32, U64, U1)
@@ -309,15 +311,21 @@ VALUE_OPCODE(ImageSampleRaw, U32x4, ImageResource, SamplerResource, ImageAddress
 VALUE_OPCODE(ImageGatherRaw, U32x4, ImageResource, SamplerResource, ImageAddress, U32, U32, U32, U32)
 VALUE_OPCODE(ImageAtomicCompareSwap32, U32, ImageResource, ImageAddress, U32, U32, U1)
 VALUE_OPCODE(ImageAtomicSwap32, U32, ImageResource, ImageAddress, U32, U1)
+VALUE_OPCODE(ImageAtomicSwap64, U64, ImageResource, ImageAddress, U64, U1)
 VALUE_OPCODE(ImageAtomicIAdd32, U32, ImageResource, ImageAddress, U32, U1)
+VALUE_OPCODE(ImageAtomicIAdd64, U64, ImageResource, ImageAddress, U64, U1)
 VALUE_OPCODE(ImageAtomicSMin32, U32, ImageResource, ImageAddress, U32, U1)
 VALUE_OPCODE(ImageAtomicUMin32, U32, ImageResource, ImageAddress, U32, U1)
+VALUE_OPCODE(ImageAtomicUMin64, U64, ImageResource, ImageAddress, U64, U1)
 VALUE_OPCODE(ImageAtomicSMax32, U32, ImageResource, ImageAddress, U32, U1)
 VALUE_OPCODE(ImageAtomicUMax32, U32, ImageResource, ImageAddress, U32, U1)
 VALUE_OPCODE(ImageAtomicUMax64, U64, ImageResource, ImageAddress, U64, U1)
 VALUE_OPCODE(ImageAtomicAnd32, U32, ImageResource, ImageAddress, U32, U1)
+VALUE_OPCODE(ImageAtomicAnd64, U64, ImageResource, ImageAddress, U64, U1)
 VALUE_OPCODE(ImageAtomicOr32, U32, ImageResource, ImageAddress, U32, U1)
+VALUE_OPCODE(ImageAtomicOr64, U64, ImageResource, ImageAddress, U64, U1)
 VALUE_OPCODE(ImageAtomicXor32, U32, ImageResource, ImageAddress, U32, U1)
+VALUE_OPCODE(ImageAtomicXor64, U64, ImageResource, ImageAddress, U64, U1)
 VALUE_OPCODE(ImageAtomicFMin32, U32, ImageResource, ImageAddress, U32, U1)
 VALUE_OPCODE(ImageAtomicFMax32, U32, ImageResource, ImageAddress, U32, U1)
 VALUE_OPCODE(GetAttribute, U32, U32, U32)

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

@@ -101,7 +101,7 @@ SharedMemoryResources CollectMemoryResources(const Program& program, std::vector
 
 bool UsesFlattenedSrt(const Program& program) {
 	const auto uses_mapping = [](const auto& resource) {
-		return resource.indirect_search_iterations != 0u;
+		return resource.indirect_root != UINT32_MAX;
 	};
 	return std::ranges::any_of(program.blocks, [](const Block* block) {
 		return std::ranges::any_of(*block, [](const Inst& inst) {

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

@@ -347,6 +347,17 @@ void FoldInstruction(Block& block, Block::iterator instruction,
 			if (IsImmediate(low, Type::U32) && IsImmediate(high, Type::U32)) {
 				Replace(inst, Value(static_cast<uint64_t>(low.U32()) |
 				                    (static_cast<uint64_t>(high.U32()) << 32u)));
+			} else {
+				const auto* low_extract  = low.TryInstruction();
+				const auto* high_extract = high.TryInstruction();
+				if (low_extract != nullptr && high_extract != nullptr &&
+				    low_extract->GetOpcode() == ValueOpcode::CompositeExtractU64 &&
+				    high_extract->GetOpcode() == ValueOpcode::CompositeExtractU64 &&
+				    Arg(*low_extract, 1) == Value(0u) && Arg(*high_extract, 1) == Value(1u) &&
+				    Arg(*low_extract, 0) == Arg(*high_extract, 0)) {
+					// Guest register pairs need not unpack and repack between wide operations.
+					Replace(inst, Arg(*low_extract, 0));
+				}
 			}
 			return;
 		}

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

@@ -238,6 +238,13 @@ bool NormalizeIndirectStoreBuffer(DescriptorValue& value) {
 	return true;
 }
 
+bool IsBoundedDescriptorTable(const ResourcePlan& program,
+                              const DescriptorSource::IndirectDescriptor& indirect) {
+	const auto* table = Source(program, indirect.table_source);
+	return indirect.sources.empty() && !indirect.selector &&
+	       indirect.workgroup_axis == UINT32_MAX && table != nullptr && table->dword_count == 4u;
+}
+
 template <typename Specialization, typename Normalize>
 bool MaterializeIndirectDescriptor(const ResourcePlan&                         program,
                                    const DescriptorSource::IndirectDescriptor& indirect,
@@ -249,7 +256,24 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
                                    uint32_t maximum_resources, Normalize&& normalize) {
 	const auto  descriptor_bytes = dword_count * sizeof(uint32_t);
 	const auto& sources = indirect.sources;
+	const auto* selector = indirect.selector ? &*indirect.selector : nullptr;
 	auto& keys = program.material_keys;
+	const auto children_begin = descriptors.size();
+	const auto root_resource = specializations[resource_index];
+	const auto intern_candidate = [&](const DescriptorValue& candidate, uint32_t& ordinal) {
+		ordinal = 0u;
+		if (candidate == descriptors[resource_index]) return true;
+		const auto found = std::find(descriptors.begin() + children_begin, descriptors.end(), candidate);
+		ordinal = static_cast<uint32_t>(found - descriptors.begin() - children_begin + 1u);
+		if (found == descriptors.end()) {
+			if (descriptors.size() >= maximum_resources) return false;
+			descriptors.push_back(candidate);
+			auto child = root_resource;
+			child.indirect_root = resource_index;
+			specializations.push_back(child);
+		}
+		return true;
+	};
 	const auto read_keys = [&](const ShaderBufferResource& material, uint64_t first,
 	                           uint64_t step, uint64_t count) {
 		keys.resize(count);
@@ -269,8 +293,8 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 		keys.clear();
 		DescriptorValue material_value;
 		DescriptorValue table_value;
-		if ((indirect.material_source != UINT32_MAX &&
-		     !clean.EvaluateDescriptor(indirect.material_source, material_value)) ||
+		if ((selector != nullptr &&
+		     !clean.EvaluateDescriptor(selector->source, material_value)) ||
 		    !clean.EvaluateDescriptor(indirect.table_source, table_value)) {
 			return false;
 		}
@@ -283,7 +307,48 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 		} else {
 			return false;
 		}
-		if (indirect.material_source == UINT32_MAX) {
+		if (IsBoundedDescriptorTable(program, indirect)) {
+			if (!indirect.table_scalar && table.Format() == Prospero::BufferFormat::kInvalid) {
+				descriptors[resource_index] = {.dword_count = dword_count};
+				return true;
+			}
+			if (table_size > UINT32_MAX ||
+			    (indirect.table_record_bytes != 0u &&
+			     (table.Type() != 0u || table.SwizzleEnabled() || table.AddTid() ||
+			      table.OutOfBounds() != 0u || (table_base & 3u) != 0u ||
+			      (table.Stride() & 3u) != 0u || table.Stride() < indirect.table_record_bytes)))
+				return false;
+			// Descriptor bytes are refreshed together. The GPU selects the DWORD address;
+			// no material selector or unused table field is interpreted on the CPU.
+			keys.resize(table_size / 4u);
+			if (!ReadScalarTable(table_base, table_size, 0u, runtime, keys)) return false;
+			const auto mapping_offset = snapshot.flattened_srt.size();
+			descriptors[resource_index] = {.dword_count = dword_count};
+			snapshot.flattened_srt.resize(mapping_offset + 2u, 0u);
+			for (uint32_t word = 0; word < keys.size(); ++word) {
+				DescriptorValue candidate {.dword_count = dword_count};
+				std::copy_n(keys.begin() + word, std::min<size_t>(dword_count, keys.size() - word),
+				            candidate.dwords.begin());
+				if (!normalize(candidate)) return false;
+				uint32_t ordinal;
+				if (!intern_candidate(candidate, ordinal)) return false;
+				if (ordinal == 0u) continue;
+				snapshot.flattened_srt.resize(mapping_offset + 2u + word, 0u);
+				snapshot.flattened_srt[mapping_offset + 1u + word] = ordinal;
+			}
+			if (descriptors.size() == children_begin) {
+				snapshot.flattened_srt.resize(mapping_offset);
+			} else {
+				snapshot.flattened_srt[mapping_offset] =
+				    static_cast<uint32_t>(snapshot.flattened_srt.size() - mapping_offset - 1u);
+				auto& root = specializations[resource_index];
+				root.indirect_root = resource_index;
+				root.indirect_mapping_offset = static_cast<uint32_t>(mapping_offset);
+				root.indirect_search_iterations = 0u;
+			}
+			return true;
+		}
+		if (selector == nullptr) {
 			uint32_t key_count = 0;
 			if (indirect.workgroup_axis != UINT32_MAX) {
 				if (indirect.workgroup_axis >= runtime.workgroup_counts.size()) return false;
@@ -295,7 +360,7 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 				if (std::bit_cast<int32_t>(key_count) <= 0) key_count = 0;
 			}
 			if (indirect.table_stride == 0u || key_count > MaxIndirectDescriptorProbes ||
-			    (key_count != 0u && uint64_t {indirect.table_offset} +
+			    (key_count != 0u && uint64_t {indirect.table_offset} + indirect.table_immediate +
 			         uint64_t {key_count - 1u} * indirect.table_stride + descriptor_bytes >
 			             UINT32_MAX + 1ull)) {
 				return false;
@@ -306,19 +371,18 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 			ShaderBufferResource material;
 			uint32_t             first = 0, count = 0;
 			if (!DecodeBufferDescriptor(material_value, material) || material.Type() != 0u ||
-			    (indirect.selector_shift != 0u && (material.Base48() & 3u) != 0u) ||
 			    material.SwizzleEnabled() ||
 			    material.AddTid() || material.OutOfBounds() != 0u ||
-			    uint64_t {indirect.selector_offset} + 4u > material.Stride() ||
+			    uint64_t {selector->offset} + 4u > material.Stride() ||
 			    !clean.Evaluate(indirect.selector_first, first) ||
 			    !clean.Evaluate(indirect.key_count, count) || count > MaxIndirectDescriptorProbes ||
 			    uint64_t {first} + count > material.NumRecords())
 				return false;
 			if (count != 0u && (uint64_t {first} + count - 1u) * material.Stride() +
-			                           indirect.selector_offset + 4u >
+			                           selector->offset + 4u >
 			                       uint64_t {UINT32_MAX} + 1u)
 				return false;
-			if (!read_keys(material, uint64_t {first} * material.Stride() + indirect.selector_offset,
+			if (!read_keys(material, uint64_t {first} * material.Stride() + selector->offset,
 			               material.Stride(), count)) return false;
 		} else if (!indirect.selector_mask.IsEmpty()) {
 			uint32_t mask = 0;
@@ -334,8 +398,8 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 			keys.reserve(std::popcount(mask));
 			while (mask != 0u) {
 				const auto index = std::countr_zero(mask);
-				const auto offset = static_cast<uint64_t>(indirect.selector_offset) +
-				                    static_cast<uint64_t>(index) * indirect.selector_stride;
+				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 (!ReadScalarTable(material_base, UINT64_MAX, static_cast<uint32_t>(offset),
@@ -343,37 +407,14 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 				keys.push_back(key);
 				mask &= mask - 1u;
 			}
-		} else {
-			ShaderBufferResource material;
-			if (!DecodeBufferDescriptor(material_value, material) || table_value.dword_count != 4u) {
-				return false;
-			}
-			// The first aligned offset includes the immediate added after shader U32 arithmetic.
-			const auto step = std::max<uint64_t>(4u,
-			    std::gcd<uint64_t>(indirect.selector_stride, uint64_t {1} << 32u));
-			const uint64_t first = indirect.selector_offset;
-			const auto size = material.GetSize();
-			const auto limit = std::min(first + (uint64_t {1} << 32u) - step, size >= 4u ? size - 4u : 0u);
-			const auto probe_count = size >= 4u && first <= limit ? (limit - first) / step + 1u : 0u;
-			if (probe_count > MaxIndirectDescriptorProbes) {
-				return false;
-			}
-			if (!read_keys(material, first, step, probe_count)) return false;
-			if (indirect.selector_shift == 0u) keys.push_back(0u);
-		}
-		if (indirect.material_source != UINT32_MAX) {
-			if (indirect.selector_shift != 0u) {
-				for (auto& key: keys) key >>= indirect.selector_shift;
-				keys.push_back(0u); // An out-of-range material load returns zero.
-			}
+		} else return false;
+		if (selector != nullptr) {
 			std::ranges::sort(keys);
 			keys.erase(std::unique(keys.begin(), keys.end()), keys.end());
 		}
 	}
 
-	const auto children_begin = descriptors.size();
 	const auto mapping_offset = snapshot.flattened_srt.size();
-	const auto root_resource  = specializations[resource_index];
 	const auto key_count = sources.empty() ? keys.size() : sources.size();
 	snapshot.flattened_srt.resize(mapping_offset + 1u + key_count * 2u);
 	snapshot.flattened_srt[mapping_offset] = static_cast<uint32_t>(key_count);
@@ -382,8 +423,9 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 		DescriptorValue candidate;
 		candidate.dword_count = dword_count;
 		if (sources.empty()) {
-			const auto table_offset =
-			    static_cast<uint32_t>(key * indirect.table_stride) + indirect.table_offset;
+			const uint64_t table_offset =
+			    uint64_t {(key * indirect.table_stride + indirect.table_offset) & ~3u} +
+			    (indirect.table_immediate & ~3u);
 			if (!ReadScalarTable(table_base, table_size, table_offset, runtime,
 			                     std::span(candidate.dwords).first(dword_count))) {
 				return false;
@@ -395,19 +437,8 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 		uint32_t ordinal = 0;
 		if (entry == 0) {
 			descriptors[resource_index] = candidate;
-		} else if (descriptors[resource_index] != candidate) {
-			const auto found =
-			    std::find(descriptors.begin() + children_begin, descriptors.end(), candidate);
-			ordinal = static_cast<uint32_t>(found - descriptors.begin() - children_begin + 1u);
-			if (found == descriptors.end()) {
-				if (descriptors.size() >= maximum_resources) {
-					return false;
-				}
-				descriptors.push_back(candidate);
-				auto child          = root_resource;
-				child.indirect_root = resource_index;
-				specializations.push_back(child);
-			}
+		} else if (!intern_candidate(candidate, ordinal)) {
+			return false;
 		}
 		snapshot.flattened_srt[mapping_offset + 1u + entry * 2u] = key;
 		snapshot.flattened_srt[mapping_offset + 2u + entry * 2u] = ordinal;
@@ -565,9 +596,9 @@ static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSna
 		const auto key_count = root.indirect_mapping_offset < snapshot.flattened_srt.size()
 		                           ? snapshot.flattened_srt[root.indirect_mapping_offset]
 		                           : 0u;
-		if (root.indirect_search_iterations == 0u || key_count < 2u ||
+		if (key_count == 0u ||
 		    static_cast<size_t>(root.indirect_mapping_offset) + 1u +
-		            static_cast<size_t>(key_count) * 2u >
+		            static_cast<size_t>(key_count) * (root.indirect_search_iterations == 0u ? 1u : 2u) >
 		        snapshot.flattened_srt.size()) {
 			return SpecializationFail("indirect image specialization has an invalid key mapping");
 		}
@@ -663,7 +694,6 @@ static std::vector<ResourceBlock> ResourceControlFlow(const Program& program, co
 		}
 	}
 	std::vector<ResourceBlock> blocks(program.blocks.size());
-	std::vector<std::vector<uint32_t>> predecessors(blocks.size());
 	for (uint32_t i = 0; i < blocks.size(); i++) {
 		auto&                 block      = blocks[i];
 		const auto&           info       = program.block_info[i];
@@ -687,7 +717,6 @@ static std::vector<ResourceBlock> ResourceControlFlow(const Program& program, co
 				return {};
 			}
 			block.successors.push_back(found->second);
-			predecessors[found->second].push_back(i);
 		}
 		for (const auto& inst: *program.blocks[i]) {
 			const auto op     = inst.GetOpcode();
@@ -717,30 +746,7 @@ static std::vector<ResourceBlock> ResourceControlFlow(const Program& program, co
 		block.sources.erase(std::unique(block.sources.begin(), block.sources.end()),
 		                    block.sources.end());
 	}
-	// Host branches only matter when they guard a later resource access. Leave
-	// data-only branches (for example a mutable flag checked before a trap) on the GPU.
-	std::vector<uint8_t> resource_reachable(blocks.size());
-	std::vector<uint32_t> pending;
-	for (uint32_t i = 0; i < blocks.size(); ++i) {
-		if (!blocks[i].sources.empty() || !blocks[i].srt_reads.empty()) {
-			resource_reachable[i] = 1;
-			pending.push_back(i);
-		}
-	}
-	while (!pending.empty()) {
-		const auto index = pending.back();
-		pending.pop_back();
-		for (const auto predecessor: predecessors[index]) {
-			if (resource_reachable[predecessor]) continue;
-			resource_reachable[predecessor] = 1;
-			pending.push_back(predecessor);
-		}
-	}
-	for (auto& block: blocks) {
-		block.condition = std::ranges::any_of(block.successors, [&](uint32_t successor) {
-			return resource_reachable[successor] != 0;
-		}) ? predicate(block.condition) : Value {};
-	}
+	for (auto& block: blocks) block.condition = predicate(block.condition);
 	if (std::ranges::none_of(
 	        blocks, [](const ResourceBlock& block) { return !block.condition.IsEmpty(); })) {
 		return {};
@@ -1034,9 +1040,10 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 		if (!source.indirect_descriptor.has_value()) continue;
 		const auto& indirect                = *source.indirect_descriptor;
 		plan.requires_specialization_memory = true;
-		capture_indirect_reads |= indirect.material_source != UINT32_MAX || !indirect.sources.empty();
-		MarkCleanFlatSlots(plan, Source(plan, indirect.material_source), plan.clean_flat_slots,
-		                   indirect.selector_mask);
+		capture_indirect_reads = true;
+		if (indirect.selector)
+			MarkCleanFlatSlots(plan, Source(plan, indirect.selector->source), plan.clean_flat_slots);
+		MarkCleanFlatSlots(plan, nullptr, plan.clean_flat_slots, indirect.selector_mask);
 		MarkCleanFlatSlots(plan, nullptr, plan.clean_flat_slots, indirect.selector_first);
 		MarkCleanFlatSlots(plan, nullptr, plan.clean_flat_slots, indirect.key_count);
 		if (indirect.sources.empty()) {
@@ -1180,10 +1187,13 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 				continue;
 			}
 			const auto& indirect = *source->indirect_descriptor;
+			const bool bounded_table = IsBoundedDescriptorTable(program, indirect);
 			if (!MaterializeIndirectDescriptor(
 			        program, indirect, i, 8u, observed, clean, snapshot, snapshot.images,
 			        specialization.images, UINT32_MAX, [&](DescriptorValue& value) {
-				        if (NullImageDescriptor(value) || !ValidImageDescriptor(value, image.r128))
+				        // A broad heap also contains resources for other typed image operations.
+				        if (NullImageDescriptor(value) || !ValidImageDescriptor(value, image.r128) ||
+				            (bounded_table && DescriptorDimension(value, image.dimension) != image.dimension))
 					        value.dwords.fill(0u);
 				        return true;
 			        })) {

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

@@ -8,7 +8,6 @@
 #include <algorithm>
 #include <fmt/format.h>
 #include <map>
-#include <numeric>
 #include <optional>
 #include <span>
 #include <tuple>
@@ -374,11 +373,68 @@ public:
 			}
 		}
 		for (const auto& plan: m_indirect_descriptors) {
+			auto key = plan.key;
+			if (plan.table_read != nullptr) {
+				const auto* anchor = plan.table_indexed ? plan.handle : plan.table_read;
+				auto* block = anchor->Parent();
+				const auto where = std::ranges::find_if(block->Instructions(),
+				    [&](const Inst& inst) { return &inst == anchor; });
+				const auto emit = [&](ValueOpcode op, std::initializer_list<Value> args) {
+					return Value(&*block->PrependNewInst(where, op, args));
+				};
+				const auto& read = *plan.table_read;
+				const auto& memory = m_program.memory_info[read.Flags<MemoryFlags>().index];
+				Value offset;
+				Value valid(true);
+				if (plan.table_indexed) {
+					const auto* table = read.Arg(0).ResolveInstruction();
+					const auto stride = emit(ValueOpcode::BitFieldUExtract,
+					    {table->Arg(1), Value(16u), Value(14u)});
+					const auto format = emit(ValueOpcode::BitFieldUExtract,
+					    {table->Arg(3), Value(12u), Value(7u)});
+					offset = emit(ValueOpcode::IAdd32,
+					    {emit(ValueOpcode::IMul32, {read.Arg(1), stride}), Value(memory.offset)});
+					valid = emit(ValueOpcode::LogicalAnd,
+					    {emit(ValueOpcode::ULessThan32, {read.Arg(1), table->Arg(2)}),
+					     emit(ValueOpcode::INotEqual32, {format, Value(0u)})});
+				} else {
+					const auto dynamic = emit(ValueOpcode::BitwiseAnd32, {read.Arg(1), Value(~3u)});
+					offset = emit(ValueOpcode::IAdd32, {dynamic, Value(memory.offset & ~3u)});
+					valid = emit(ValueOpcode::ULessThanEqual32, {dynamic, offset});
+				}
+				key = emit(ValueOpcode::SelectU32, {valid, offset, Value(UINT32_MAX)});
+				if (!plan.table_indexed) {
+					// NativeDescriptorSource proved this read reaches the image. Preserve its
+					// SSA Phi edges while retaining descriptor DWORDs for other consumers.
+					std::map<const Inst*, Value> projected;
+					const auto project = [&](auto&& self, Value value) -> Value {
+						value = value.Resolve();
+						const auto* phi = value.TryInstruction();
+						if (phi == plan.table_read) return key;
+						if (phi == nullptr || phi->GetOpcode() != ValueOpcode::Phi) return Value(UINT32_MAX);
+						if (const auto found = projected.find(phi); found != projected.end()) return found->second;
+						auto* parent = phi->Parent();
+						const auto position = std::ranges::find_if(parent->Instructions(),
+						    [&](const Inst& inst) { return &inst == phi; });
+						auto& selected = *parent->PrependNewInst(position, ValueOpcode::Phi);
+						selected.SetFlags(Type::U32);
+						projected.emplace(phi, Value(&selected));
+						for (uint32_t arm = 0; arm < phi->NumArgs(); ++arm)
+							selected.AddPhiOperand(phi->PhiBlock(arm), self(self, phi->Arg(arm)));
+						return Value(&selected);
+					};
+					key = project(project, plan.handle->Arg(0));
+				}
+				plan.handle->SetArg(0, key);
+				for (uint32_t word = 1; word < plan.handle->NumArgs(); ++word)
+					plan.handle->SetArg(word, Value(0u));
+				continue;
+			}
 			if (plan.handle->GetOpcode() == ValueOpcode::GetBufferResource) {
 				plan.handle->SetArg(0, plan.key);
 				for (uint32_t word = 1; word < 4u; ++word) plan.handle->SetArg(word, Value(0u));
 			} else {
-				if (plan.reads[0] == nullptr) plan.handle->SetArg(0, plan.key);
+				if (plan.reads[0] == nullptr) plan.handle->SetArg(0, key);
 				for (uint32_t dword = 0; dword < 4u; dword++) {
 					plan.handle->SetArg(dword + 1u, plan.roots[dword + 4u]);
 				}
@@ -437,6 +493,10 @@ private:
 		std::array<Value, 8>       roots {};
 		std::array<uint32_t, 8>    memory {};
 		std::array<const Inst*, 8> reads {};
+		bool split_offsets = false;
+		bool retain_reads = false;
+		bool table_indexed = false;
+		const Inst* table_read = nullptr;
 	};
 
 	[[noreturn]] void Fail(uint32_t pc, const std::string& reason) const {
@@ -742,6 +802,15 @@ private:
 				}
 			}
 		}
+		for (auto* block: m_program.blocks) {
+			for (auto& inst: *block) {
+				uint32_t index = 0;
+				if (inst.GetOpcode() == ValueOpcode::LoadAddressU32 &&
+				    ScalarReadMemory(inst, index) != nullptr && inst.Arg(1).Resolve().IsImmediate() &&
+				    ValidateRuntimeValue(m_program, Value(&inst)))
+					CollectScalarRead(Value(&inst), inst.Flags<MemoryFlags>().pc);
+			}
+		}
 		for (auto* read: m_scalar_reads) {
 			const auto flags = read->Flags<MemoryFlags>();
 			auto& memory = m_program.memory_info[flags.index];
@@ -825,7 +894,7 @@ private:
 
 	uint32_t InternSource(const DescriptorSource& descriptor) {
 		for (uint32_t candidate = 0; candidate < m_sources.size(); candidate++) {
-			const auto& current = m_sources[candidate];
+			auto& current = m_sources[candidate];
 			if (current.dword_count != descriptor.dword_count ||
 			    current.indirect_descriptor.has_value() != descriptor.indirect_descriptor.has_value()) {
 				continue;
@@ -833,10 +902,9 @@ private:
 			if (current.indirect_descriptor.has_value()) {
 				const auto& a = *current.indirect_descriptor;
 				const auto& b = *descriptor.indirect_descriptor;
-				if (a.material_source != b.material_source || a.table_source != b.table_source ||
-				    a.selector_stride != b.selector_stride || a.selector_offset != b.selector_offset ||
-				    a.selector_shift != b.selector_shift ||
-				    a.table_offset != b.table_offset || a.table_stride != b.table_stride ||
+				if (a.selector != b.selector || a.table_source != b.table_source ||
+				    a.table_offset != b.table_offset || a.table_immediate != b.table_immediate ||
+				    a.table_stride != b.table_stride ||
 				    a.workgroup_axis != b.workgroup_axis || a.sources != b.sources ||
 				    !EquivalentValue(m_program, a.key_count, b.key_count) ||
 				    a.selector_first.IsEmpty() != b.selector_first.IsEmpty() ||
@@ -851,6 +919,12 @@ private:
 				same = same && EquivalentValue(m_program, current.dwords[i], descriptor.dwords[i]);
 			}
 			if (same) {
+				if (current.indirect_descriptor) {
+					current.indirect_descriptor->table_scalar |= descriptor.indirect_descriptor->table_scalar;
+					current.indirect_descriptor->table_record_bytes = std::max(
+					    current.indirect_descriptor->table_record_bytes,
+					    descriptor.indirect_descriptor->table_record_bytes);
+				}
 				return candidate;
 			}
 		}
@@ -940,41 +1014,6 @@ private:
 		return ValidateSource(descriptor, bad_dword);
 	}
 
-	bool MatchMaterialOffset(Value value, uint32_t& stride, uint32_t& offset) const {
-		stride = 1u;
-		offset = 0u;
-		value = value.Resolve();
-		while (const auto* inst = value.TryInstruction()) {
-			if (inst->NumArgs() != 2u) break;
-			uint32_t constant = 0;
-			Value remaining;
-			if (inst->GetOpcode() == ValueOpcode::ShiftLeftLogical32 &&
-			    ImmediateU32(inst->Arg(1), constant) && constant < 32u) {
-				stride *= 1u << constant;
-				remaining = inst->Arg(0);
-			} else if (inst->GetOpcode() == ValueOpcode::IAdd32 ||
-			           inst->GetOpcode() == ValueOpcode::IMul32) {
-				if (ImmediateU32(inst->Arg(0), constant)) remaining = inst->Arg(1);
-				else if (ImmediateU32(inst->Arg(1), constant)) remaining = inst->Arg(0);
-				else break;
-				if (inst->GetOpcode() == ValueOpcode::IAdd32) offset += stride * constant;
-				else stride *= constant;
-			} else {
-				break;
-			}
-			value = remaining.Resolve();
-		}
-		// Treat the remaining GPU-derived scalar as any U32 value.
-		return stride != 0u && value.GetType() == Type::U32 && !value.IsImmediate();
-	}
-
-	const Inst* MaterialKeyShift(Value key, uint32_t& amount) const {
-		const auto* shift = key.Resolve().TryInstruction();
-		return shift != nullptr && shift->GetOpcode() == ValueOpcode::ShiftRightLogical32 &&
-		               ImmediateU32(shift->Arg(1), amount) && amount != 0u && amount < 32u
-		           ? shift : nullptr;
-	}
-
 	enum class LaneQuantifier { Any, All };
 	struct EdgePredicate {
 		Value condition;
@@ -1421,9 +1460,9 @@ private:
 		        static_cast<uint64_t>(UINT32_MAX) + 1u) return false;
 		const auto mask = BoundedSetBitMask(offset.index, active);
 		if (mask.IsEmpty()) return false;
-		indirect.material_source = InternSource(material_source);
-		indirect.selector_stride = static_cast<uint32_t>(offset.stride);
-		indirect.selector_offset = static_cast<uint32_t>(offset.offset + memory.offset);
+		indirect.selector.emplace(DescriptorSource::IndirectDescriptor::SelectorRead{.source = InternSource(material_source),
+		                              .stride = static_cast<uint32_t>(offset.stride),
+		                              .offset = static_cast<uint32_t>(offset.offset + memory.offset)});
 		indirect.key_count = Value(32u);
 		indirect.selector_mask = mask;
 		return true;
@@ -1584,8 +1623,9 @@ private:
 	bool MatchDescriptorTable(Inst& handle, const DescriptorSource& descriptor,
 	                          IndirectDescriptorPlan& plan,
 	                          DescriptorSource& table_source, Value& key, uint32_t& table_offset,
-	                          uint32_t& table_stride) {
+	                          uint32_t& table_stride, uint32_t& table_immediate) {
 		Inst* table_handle = nullptr;
+		Value table_guard;
 		for (uint32_t dword = 0; dword < handle.NumArgs(); ++dword) {
 			auto* read = descriptor.dwords[dword].Resolve().TryInstruction();
 			if (read == nullptr) {
@@ -1593,7 +1633,31 @@ private:
 			}
 			uint32_t memory_index = 0;
 			const auto* memory = ScalarReadMemory(*read, memory_index);
-			if (memory == nullptr || memory->offset > INT32_MAX || (memory->offset & 3u) != 0u ||
+			uint32_t component = 0;
+			const bool indexed = memory == nullptr;
+			if (indexed) {
+				if (handle.NumArgs() != 8u) return false;
+				const auto* selected = UniformizedMaterialValue(descriptor.dwords[dword], handle);
+				if (selected == nullptr) return false;
+				const auto* extract = selected->Arg(1).Resolve().TryInstruction();
+				if (extract == nullptr || extract->GetOpcode() != ValueOpcode::CompositeExtractU32x4 ||
+				    !ImmediateU32(extract->Arg(1), component) || component >= 4u) return false;
+				read = extract->Arg(0).Resolve().TryInstruction();
+				if (read == nullptr || read->GetOpcode() != ValueOpcode::LoadBufferU32x4) return false;
+				memory_index = read->Flags<MemoryFlags>().index;
+				if (memory_index >= m_program.memory_info.size()) return false;
+				memory = &m_program.memory_info[memory_index];
+				uint32_t offset = 1u, scalar = 1u;
+				if (memory->kind != ResourceKind::Buffer || memory->typed || memory->formatted ||
+				    !memory->idxen || memory->offen || memory->data_bits != 32u ||
+				    memory->data_dwords != 4u ||
+				    !ImmediateU32(read->Arg(2), offset) || offset != 0u ||
+				    !ImmediateU32(read->Arg(3), scalar) || scalar != 0u ||
+				    !EquivalentValue(m_program, read->Arg(4), selected->Arg(0)) ||
+				    (dword != 0u && !EquivalentValue(m_program, table_guard, read->Arg(4)))) return false;
+				table_guard = read->Arg(4);
+			}
+			if (memory->offset > INT32_MAX || (memory->offset & 3u) != 0u ||
 			    !MemoryIndexBelongsTo(memory_index, *read)) {
 				return false;
 			}
@@ -1608,23 +1672,32 @@ private:
 			    (memory->kind == ResourceKind::ScalarAddress && read->Parent() != handle.Parent()) ||
 			    (table_handle != nullptr &&
 			     !EquivalentValue(m_program, Value(table_handle), Value(current_handle))) ||
-			    !MatchTableOffset(read->Arg(1), current_key, offset, stride) ||
-			    memory->offset > UINT32_MAX - offset) {
+			    (!indexed && !MatchTableOffset(read->Arg(1), current_key, offset, stride))) {
 				return false;
 			}
-			offset += memory->offset;
+			if (indexed) {
+				current_key = read->Arg(1);
+				offset = component * 4u;
+			}
 			if (dword == 0u) {
+				if (component != 0u) return false;
+				plan.table_indexed = indexed;
 				key = current_key;
 				table_offset = offset;
+				table_immediate = memory->offset;
 				table_stride = stride;
-			} else if (table_stride != stride || !EquivalentValue(m_program, key, current_key) ||
-			           static_cast<uint64_t>(table_offset) + dword * sizeof(uint32_t) != offset) {
+			} else if (plan.table_indexed != indexed || table_stride != stride ||
+			           !EquivalentValue(m_program, key, current_key) ||
+			           uint64_t {table_offset} + table_immediate + dword * sizeof(uint32_t) !=
+			               uint64_t {offset} + memory->offset) {
 				return false;
 			}
+			plan.split_offsets |= !indexed && offset != table_offset;
 			table_handle = current_handle;
-			if (handle.NumArgs() == 8u
-			        ? WorkgroupAxis(key) == UINT32_MAX && !UsesOnlyImageDescriptors(*read) :
-			    std::ranges::any_of(read->Uses(), [](const Use& use) {
+			if (handle.NumArgs() == 8u) {
+				if (memory->kind == ResourceKind::ScalarAddress)
+					plan.retain_reads |= !UsesOnlyImageDescriptors(*read);
+			} else if (std::ranges::any_of(read->Uses(), [](const Use& use) {
 				return use.user->GetOpcode() != ValueOpcode::GetBufferResource ||
 				       std::ranges::any_of(use.user->Uses(), [](const Use& consumer) {
 					       return consumer.user->GetOpcode() != ValueOpcode::StoreBufferU32;
@@ -1642,36 +1715,6 @@ private:
 		return true;
 	}
 
-	bool MatchUniformizedBufferKey(Value key, const Inst& image,
-	                               DescriptorSource::IndirectDescriptor& indirect,
-	                               DescriptorSource& material_source) {
-		const auto* selected = UniformizedMaterialValue(key, image);
-		if (selected == nullptr) return false;
-		uint32_t amount = 0;
-		const auto* shift = MaterialKeyShift(selected->Arg(1), amount);
-		if (shift == nullptr) return false;
-		const auto* loaded = shift->Arg(0).Resolve().TryInstruction();
-		if (loaded == nullptr || loaded->GetOpcode() != ValueOpcode::SelectU32 ||
-		    !EquivalentValue(m_program, selected->Arg(0), loaded->Arg(0))) return false;
-		const auto* read = loaded->Arg(1).Resolve().TryInstruction();
-		if (read == nullptr || read->GetOpcode() != ValueOpcode::LoadBufferU32) return false;
-		const auto& memory = m_program.memory_info[read->Flags<MemoryFlags>().index];
-		uint32_t offset = 1u, scalar = 1u;
-		if (memory.kind != ResourceKind::Buffer || memory.typed || memory.formatted ||
-		    !memory.idxen || memory.offen ||
-		    !ImmediateU32(read->Arg(2), offset) || offset != 0u ||
-		    !ImmediateU32(read->Arg(3), scalar) || scalar != 0u ||
-		    !EquivalentValue(m_program, selected->Arg(0), read->Arg(4)) ||
-		    !MakeRuntimeTableSource(*read, material_source)) return false;
-		// GPU index arithmetic may wrap. Every record in the bounded V# is a candidate.
-		indirect.material_source = InternSource(material_source);
-		indirect.selector_first = Value(0u);
-		indirect.key_count = material_source.dwords[2];
-		indirect.selector_offset = memory.offset;
-		indirect.selector_shift = amount;
-		return true;
-	}
-
 	bool TryMakeIndirectImage(Inst& handle, const DescriptorSource& descriptor,
 	                          IndirectDescriptorPlan& plan) {
 		if (handle.GetOpcode() != ValueOpcode::GetImageResource || handle.NumArgs() != 8u) return false;
@@ -1679,20 +1722,21 @@ private:
 		Value key;
 		uint32_t table_offset = 0;
 		uint32_t table_stride = 0;
+		uint32_t table_immediate = 0;
 		if (!MatchDescriptorTable(handle, descriptor, plan, table_source, key, table_offset,
-		                          table_stride)) return false;
+		                          table_stride, table_immediate)) return false;
 		DescriptorSource material_source;
 		DescriptorSource::IndirectDescriptor indirect;
 		indirect.table_offset = table_offset;
+		indirect.table_immediate = table_immediate;
 		indirect.table_stride = table_stride;
 		indirect.workgroup_axis = WorkgroupAxis(key);
-		if (indirect.workgroup_axis != UINT32_MAX) {
+		if (indirect.workgroup_axis != UINT32_MAX && !plan.table_indexed) {
 			// The descriptor also supplies dimensions to shader arithmetic. Keep its reads;
 			// only the image handle is projected onto the bounded workgroup key.
-			plan.reads.fill(nullptr);
-		} else if (table_stride != 32u) {
-			return false;
+			plan.retain_reads = true;
 		} else if (table_source.dword_count == 2u) {
+			if (table_stride != 32u) return false;
 			const auto* selector = key.Resolve().TryInstruction();
 			const bool bitscan = selector != nullptr && selector->GetOpcode() == ValueOpcode::FindILsb32 &&
 			    selector->NumArgs() == 1u && !m_shader_writes &&
@@ -1709,32 +1753,21 @@ private:
 			}
 			if ((table_offset & 3u) != 0u ||
 			    (bitscan && table_offset > UINT32_MAX - (32u * 32u - 1u))) return false;
-		} else if (!MatchUniformizedBufferKey(key, handle, indirect, material_source)) {
-			const auto* key_shift = MaterialKeyShift(key, indirect.selector_shift);
-			auto* material_read =
-			    (key_shift != nullptr ? key_shift->Arg(0) : key).Resolve().TryInstruction();
-			uint32_t material_memory_index = 0;
-			const auto* memory = material_read != nullptr
-			                         ? ScalarReadMemory(*material_read, material_memory_index) : nullptr;
-			if (table_offset != 0u || memory == nullptr || memory->kind != ResourceKind::ScalarBuffer ||
-			    memory->offset > INT32_MAX || !MemoryIndexBelongsTo(material_memory_index, *material_read)) {
-				return false;
-			}
-			if (!MatchMaterialOffset(material_read->Arg(1), indirect.selector_stride,
-			                         indirect.selector_offset)) {
-				return false;
-			}
-			const auto step = std::gcd<uint64_t>(indirect.selector_stride, uint64_t {1} << 32u);
-			indirect.selector_offset =
-			    (static_cast<uint32_t>(indirect.selector_offset % step) & ~3u) + (memory->offset & ~3u);
-			const auto* material_handle = material_read->Arg(0).Resolve().TryInstruction();
-			if (material_handle == nullptr ||
-			    material_handle->GetOpcode() != ValueOpcode::GetBufferResource ||
-			    !MakeRuntimeTableSource(*material_read, material_source)) {
-				return false;
-			}
-			indirect.material_source = InternSource(material_source);
+		} else {
+			// A bounded V# supplies the complete image table. Leave every GPU selector
+			// and descriptor read in the shader; the host only translates table bytes.
+			if (plan.split_offsets) return false;
+			indirect.table_stride = 0u;
+			indirect.table_offset = 0u;
+			indirect.table_immediate = 0u;
+			indirect.workgroup_axis = UINT32_MAX;
+			plan.table_read = plan.reads[0];
+			if (plan.table_indexed) indirect.table_record_bytes = table_immediate + 32u;
+			else indirect.table_scalar = true;
+			plan.retain_reads = true;
 		}
+
+		if (plan.retain_reads) plan.reads.fill(nullptr);
 		indirect.table_source = InternSource(table_source);
 		DescriptorSource image_source;
 		image_source.dword_count = 8u;
@@ -1768,9 +1801,10 @@ private:
 		Value key;
 		uint32_t table_offset = 0;
 		uint32_t table_stride = 0;
+		uint32_t table_immediate = 0;
 		if (!MatchDescriptorTable(handle, descriptor, plan, table_source, key, table_offset,
-		                          table_stride) || table_stride != 16u ||
-		    table_source.dword_count != 4u || table_offset != 0u) return false;
+		                          table_stride, table_immediate) || table_stride != 16u ||
+		    table_source.dword_count != 4u || table_offset != 0u || table_immediate != 0u) return false;
 		const auto* lane_read = key.Resolve().TryInstruction();
 		if (lane_read == nullptr || lane_read->GetOpcode() != ValueOpcode::ReadFirstLane)
 			return false;
@@ -1802,9 +1836,9 @@ private:
 		indirect.table_stride = table_stride;
 		if (!MakeRuntimeTableSource(*read, material_source) ||
 		    !BoundedSelectedIndex(read->Arg(1), read->Arg(4), indirect)) return false;
-		indirect.material_source = InternSource(material_source);
+		indirect.selector.emplace(DescriptorSource::IndirectDescriptor::SelectorRead{.source = InternSource(material_source),
+		                              .offset = memory.offset + lane * 4u});
 		indirect.table_source = InternSource(table_source);
-		indirect.selector_offset = memory.offset + lane * 4u;
 		DescriptorSource source;
 		source.dword_count = 4u;
 		source.dwords.fill(Value(0u));
@@ -2032,6 +2066,8 @@ private:
 				DescriptorSource descriptor;
 				MakeSource(*handle, inst.GetOpcode() == ValueOpcode::StoreBufferU32 ? 4u : 8u,
 				           false, false, memory.resource * 4u, descriptor, flags.pc);
+				uint32_t bad_dword = 0;
+				if (ValidateSource(descriptor, bad_dword)) continue;
 				IndirectDescriptorPlan plan;
 				if (TryMakeIndirectImage(*handle, descriptor, plan) || TryMakeFiniteImage(*handle, plan) ||
 				    TryMakeIndirectBuffer(*handle, descriptor, plan)) {

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

@@ -143,6 +143,7 @@ void ValidateValueReferences(const Program& program, ShaderStageInputInfo input_
 							}
 							break;
 						case StageInputKind::BaryCoordSmooth:
+						case StageInputKind::BaryCoordSmoothSample:
 						case StageInputKind::BaryCoordSmoothCentroid:
 						case StageInputKind::BaryCoordNoPerspective:
 							if (component >= 2u) {
@@ -299,6 +300,9 @@ void CollectBuiltinInputs(const Program& program, ShaderInfo& info) {
 					break;
 				case StageInputKind::Layer: AddInput(info, kind, 0, 1, "gl_Layer"); break;
 				case StageInputKind::SampleId: AddInput(info, kind, 0, 1, "gl_SampleID"); break;
+				case StageInputKind::BaryCoordSmoothSample:
+					AddInput(info, StageInputKind::SampleId, 0, 1, "gl_SampleID");
+					[[fallthrough]];
 				case StageInputKind::BaryCoordSmooth:
 				case StageInputKind::BaryCoordSmoothCentroid:
 					AddInput(info, StageInputKind::BaryCoordSmooth, 0, 3, "gl_BaryCoordKHR");
@@ -386,16 +390,23 @@ void CollectOutputs(const Program& program, ShaderStageInputInfo input_info, Sha
 					AddOutput(info, StageOutputKind::Parameter, export_info.index,
 					          export_info.index, fmt::format("out_param_{}", export_info.index));
 					break;
-				case ExportTargetKind::Mrt:
+				case ExportTargetKind::Mrt: {
 					if (alpha_remap && export_info.index != 0) {
 						break;
 					}
-					AddOutput(info, StageOutputKind::Mrt, export_info.index, export_info.index,
+					const auto slot = input_info.pixel->dual_source_blending
+					                      ? 0u : ShaderPixelExportTarget(input_info.pixel->target_shader_mask,
+					                                                    export_info.index);
+					if (slot >= 8) {
+						break;
+					}
+					AddOutput(info, StageOutputKind::Mrt, export_info.index, slot,
 					          fmt::format("out_mrt_{}", export_info.index));
 					if (alpha_remap) {
-						AddOutput(info, StageOutputKind::Mrt, 1, 1, "out_mrt_1");
+						AddOutput(info, StageOutputKind::Mrt, 1, 0, "out_mrt_1");
 					}
 					break;
+				}
 				default: break;
 			}
 		}

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

@@ -23,6 +23,9 @@ namespace {
 constexpr uint64_t AddressMask = 0x0000ffffffffffffull;
 
 bool AddSignedAddress(uint64_t base, int64_t offset, uint64_t& result) {
+	if (base > AddressMask) {
+		return false;
+	}
 	if (offset < 0) {
 		const auto magnitude = uint64_t {0} - static_cast<uint64_t>(offset);
 		if (magnitude > base) {
@@ -32,7 +35,7 @@ bool AddSignedAddress(uint64_t base, int64_t offset, uint64_t& result) {
 		return true;
 	}
 	const auto magnitude = static_cast<uint64_t>(offset);
-	if (magnitude > UINT64_MAX - base) {
+	if (magnitude > AddressMask - base) {
 		return false;
 	}
 	result = base + magnitude;
@@ -477,7 +480,7 @@ bool SrtWalker::EvaluateRawRead(const Inst& inst, uint64_t& result) {
 	if (!Arg(*handle, 0, low) || !Arg(*handle, 1, high) || !Arg(inst, 1, offset)) {
 		return false;
 	}
-	const auto base      = (high << 32u) | static_cast<uint32_t>(low);
+	const auto base      = ((high << 32u) | static_cast<uint32_t>(low)) & AddressMask;
 	const auto immediate = static_cast<int64_t>(static_cast<int32_t>(mem.offset));
 	uint64_t   address   = 0;
 	if (inst.GetOpcode() == ValueOpcode::ReadConstBuffer || vector) {
@@ -507,7 +510,7 @@ bool SrtWalker::EvaluateRawRead(const Inst& inst, uint64_t& result) {
 		if (byte_offset > size || size - byte_offset < sizeof(uint32_t)) {
 			return false;
 		}
-		address = (base & AddressMask & ~uint64_t {3}) + byte_offset;
+		address = (base & ~uint64_t {3}) + byte_offset;
 	} else {
 		const auto relative = (immediate & ~int64_t {3}) +
 		                      static_cast<int64_t>(static_cast<uint32_t>(offset) & ~3u);

src/graphics/shader/shader.cpp

@@ -525,6 +525,9 @@ static void ShaderGetStaticInputInfoPS(
 	EXIT_NOT_IMPLEMENTED(ps_info.input_num > std::size(ps_info.interpolator_settings));
 	ps_info.ps_system_input_base = ShaderCalcPsSystemInputBase(sh);
 	const uint32_t active_inputs = sh.ps_input_ena & sh.ps_input_addr;
+	if ((active_inputs & 0x00000001u) != 0) {
+		ps_info.ps_perspective_sample_vgpr = 0;
+	}
 	if ((active_inputs & 0x00000002u) != 0) {
 		ps_info.ps_perspective_center_vgpr = (active_inputs & 0x00000001u) != 0 ? 2u : 0u;
 	}
@@ -558,10 +561,12 @@ static void ShaderGetStaticInputInfoPS(
 		ps_info.interpolator_settings[i] = sh.ps_interpolator_settings[i];
 	}
 
+	ps_info.target_shader_mask = sh.m_cbShaderMask;
 	for (int i = 0; i < 8; i++) {
+		const auto slot = ShaderPixelExportTarget(ps_info.target_shader_mask, i);
 		ps_info.target_output_mode[i]    = sh.target_output_mode[i];
-		ps_info.target_export_mapping[i] = sh.target_output_mode[i] != 0
-		                                       ? target_export_mapping[i]
+		ps_info.target_export_mapping[i] = sh.target_output_mode[i] != 0 && slot < 8
+		                                       ? target_export_mapping[slot]
 		                                       : Prospero::ColorComponentMapping {};
 	}
 }
@@ -649,6 +654,7 @@ void BuildStageStaticKey(const ShaderPixelInputInfo& info, std::vector<uint32_t>
 	key.push_back(info.ps_system_input_base);
 	key.push_back(info.custom_interpolation_mask);
 	key.push_back(info.ps_perspective_center_vgpr);
+	key.push_back(info.ps_perspective_sample_vgpr);
 	key.push_back(info.ps_perspective_centroid_vgpr);
 	key.push_back(static_cast<uint32_t>(info.ps_pos_x));
 	key.push_back(static_cast<uint32_t>(info.ps_pos_y));
@@ -664,6 +670,8 @@ void BuildStageStaticKey(const ShaderPixelInputInfo& info, std::vector<uint32_t>
 	key.push_back(static_cast<uint32_t>(info.dual_source_blending));
 	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));
+	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) {
 		uint32_t packed = 0;
 		for (uint32_t i = 0; i < 4u; i++) {
@@ -697,11 +705,14 @@ ShaderParams PrepareProgram(const HW::VertexShaderInfo& regs, const HW::Context&
 	const auto& sh     = context.GetShaderRegisters();
 	const auto [data, hash] = ShaderGetMappedData(regs.es_regs.data_addr, "ShaderGetInputInfoVS():");
 	const bool merged = (context.GetShaderStages() & 0x20u) != 0;
+	// GS_EN controls amplification, not whether an NGG shader uses a workgroup.
+	const bool native_ngg = !merged && data.type == Prospero::ShaderBinaryType::kGs &&
+	                        regs.gs_regs.rsrc2.lds_size != 0;
 	auto        params = GetShaderParams(
 	    regs.es_regs.data_addr, hash,
 	    std::span<const uint32_t>(regs.gs_user_sgpr.value, regs.gs_regs.rsrc2.user_sgpr), data,
-	    merged ? 8u : 0u);
-	if (!merged) {
+	    merged || native_ngg ? 8u : 0u);
+	if (!merged && !native_ngg) {
 		if (!ShaderGetStaticVertexInputInfo(regs.es_regs.data_addr, regs.gs_user_sgpr,
 		                                    regs.gs_regs.rsrc2.user_sgpr, sh, data, info)) {
 			EXIT("failed to prepare vertex shader program\n");
@@ -746,7 +757,7 @@ ShaderParams PrepareProgram(const HW::VertexShaderInfo& regs, const HW::Context&
 		                     group.primitive_group_size != 1u || group.vertex_group_size != 1u ||
 		                     user_vgpr.vgpr1 || user_vgpr.vgpr2 || user_vgpr.vgpr3 ||
 		                     mesh.max_vertices != sh.m_vgtGsMaxVertOut);
-	} else {
+	} else if (merged) {
 		EXIT_NOT_IMPLEMENTED(regs.gs_regs.rsrc1.gs_vgpr_component_count != 3u ||
 		                     regs.gs_regs.rsrc2.es_vgpr_component_count != 3u);
 	}
@@ -755,7 +766,7 @@ ShaderParams PrepareProgram(const HW::VertexShaderInfo& regs, const HW::Context&
 	     user_config.GetPrimType() != Prospero::PrimitiveType::kTriFan &&
 	     user_config.GetPrimType() != Prospero::PrimitiveType::kTriStrip &&
 	     user_config.GetPrimType() != Prospero::PrimitiveType::kTriList) ||
-	    sh.m_vgtGsOutPrimType != 2u || sh.m_vgtGsMaxVertOut < 3u ||
+	    sh.m_vgtGsOutPrimType != 2u || (!native_ngg && sh.m_vgtGsMaxVertOut < 3u) ||
 	    group.vertex_group_size < mesh.InputPrimitiveSize() ||
 	    mesh.max_vertices == 0u) {
 		EXIT("unsupported GS assembly: input=%u output=%u vertices=%u GE=%u/%u max_output=%u\n",
@@ -763,14 +774,16 @@ ShaderParams PrepareProgram(const HW::VertexShaderInfo& regs, const HW::Context&
 		     group.primitive_group_size, group.vertex_group_size, mesh.max_vertices);
 	}
 	mesh.max_primitives = mesh.fast_launch ? mesh.max_vertices :
+	                      native_ngg ? mesh.InputPrimitiveCount(mesh.max_vertices) :
 	                      group.primitive_group_size * (sh.m_vgtGsMaxVertOut - 2u);
 	mesh.primitives_per_group = std::min({static_cast<uint32_t>(group.primitive_group_size),
 	                                      mesh.InputPrimitiveCount(group.vertex_group_size),
-	                                      mesh.max_vertices / sh.m_vgtGsMaxVertOut});
+	                                      native_ngg ? mesh.max_primitives :
+	                                          mesh.max_vertices / sh.m_vgtGsMaxVertOut});
 	EXIT_IF(mesh.primitives_per_group == 0u);
 	mesh.vertices_per_group = mesh.InputVertexCount(mesh.primitives_per_group);
-	mesh.threads_num[0] =
-	    ((mesh.max_vertices + mesh.wave_size - 1u) / mesh.wave_size) * mesh.wave_size;
+	const auto threads = native_ngg && !mesh.fast_launch ? mesh.vertices_per_group : mesh.max_vertices;
+	mesh.threads_num[0] = ((threads + mesh.wave_size - 1u) / mesh.wave_size) * mesh.wave_size;
 	mesh.threads_num[1] = mesh.threads_num[2] = 1u;
 	return params;
 }
@@ -910,6 +923,7 @@ void ShaderDbgDumpInputInfo(const ShaderPixelInputInfo& info) {
 	     "\t ps_system_input_base = %u\n"
 	     "\t custom_interpolation_mask = 0x%08" PRIx32 "\n"
 	     "\t ps_perspective_center_vgpr = %" PRIu32 "\n"
+	     "\t ps_perspective_sample_vgpr = %" PRIu32 "\n"
 	     "\t ps_perspective_centroid_vgpr = %" PRIu32 "\n"
 	     "\t ps_pos_x             = %s\n"
 	     "\t ps_pos_y             = %s\n"
@@ -923,7 +937,8 @@ void ShaderDbgDumpInputInfo(const ShaderPixelInputInfo& info) {
 	     "\t ps_early_z           = %s\n"
 	     "\t ps_execute_on_noop   = %s\n",
 	     info.input_num, info.ps_system_input_base, info.custom_interpolation_mask,
-	     info.ps_perspective_center_vgpr, info.ps_perspective_centroid_vgpr,
+	     info.ps_perspective_center_vgpr, info.ps_perspective_sample_vgpr,
+	     info.ps_perspective_centroid_vgpr,
 	     info.ps_pos_x ? "true" : "false",
 	     info.ps_pos_y ? "true" : "false", info.ps_pos_z ? "true" : "false",
 	     info.ps_pos_w ? "true" : "false", info.ps_front_face ? "true" : "false",

src/graphics/shader/shader.h

@@ -167,8 +167,10 @@ struct ShaderPixelInputInfo {
 	uint32_t                                       ps_system_input_base         = 0;
 	uint32_t                                       custom_interpolation_mask    = 0;
 	uint32_t                                       ps_perspective_center_vgpr   = UINT32_MAX;
+	uint32_t                                       ps_perspective_sample_vgpr   = UINT32_MAX;
 	uint32_t                                       ps_perspective_centroid_vgpr = UINT32_MAX;
 	uint8_t                                        target_output_mode[8]        = {};
+	uint32_t                                       target_shader_mask           = UINT32_MAX;
 	std::array<Prospero::ColorComponentMapping, 8> target_export_mapping        = {};
 	uint32_t                                       scratch_size_dwords          = 0;
 	bool                                           ps_pos_x                     = false;
@@ -309,6 +311,7 @@ struct ShaderMappedData {
 
 void ShaderInit();
 void ShaderMapUserData(uint64_t addr, const ShaderMappedData& data);
+uint32_t ShaderPixelExportTarget(uint32_t shader_mask, uint32_t export_index);
 
 void     ShaderDbgDumpInputInfo(const ShaderVertexInputInfo& info);
 void     ShaderDbgDumpInputInfo(const ShaderPixelInputInfo& info);

src/graphics/shader/shaderPixelParameter.cpp

@@ -11,6 +11,19 @@ constexpr uint32_t PsInputFlatShade  = 0x00000400u;
 
 } // namespace
 
+uint32_t ShaderPixelExportTarget(uint32_t shader_mask, uint32_t export_index) {
+	for (uint32_t target = 0; target < 8u; target++) {
+		if (((shader_mask >> (target * 4u)) & 0xfu) == 0u) {
+			continue;
+		}
+		if (export_index == 0u) {
+			return target;
+		}
+		export_index--;
+	}
+	return UINT32_MAX;
+}
+
 uint32_t ShaderPixelParameterMappedLocation(const ShaderPixelInputInfo& info, uint32_t input) {
 	return input < info.input_num ? info.interpolator_settings[input] & PsInputOffsetMask : input;
 }

src/kernel/eventQueue.cpp

@@ -121,6 +121,9 @@ int KernelEqueuePrivate::GetTriggeredEvents(KernelEvent* ev, int num) {
 				event.event = event.pending_events.front();
 				event.pending_events.pop_front();
 				event.triggered = true;
+			} else {
+				// A persistent level is delivered again by the next wait, not this batch.
+				break;
 			}
 
 			if (ret >= num) {

src/kernel/memory.cpp

@@ -21,6 +21,7 @@
 #include <map>
 #include <memory>
 #include <mutex>
+#include <shared_mutex>
 #include <vector>
 
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
@@ -168,6 +169,7 @@ static bool VirtualRangesOverlap(uint64_t left_start, uint64_t left_size, uint64
 #if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
 static uint32_t        g_test_backing_store_unmaps_before_failure = UINT32_MAX;
 static callback_func_t g_test_before_backing_map                   = nullptr;
+static callback_func_t g_test_backing_read                         = nullptr;
 #endif
 
 #include "memoryAddressSpace.inc"
@@ -359,12 +361,6 @@ public:
 		EditUnlocked(start, size, [protection](Range* r) { r->protection = protection; });
 	}
 
-	void SetMemoryType(uint64_t start, uint64_t size, int memory_type) {
-		Common::LockGuard lock(m_mutex);
-
-		EditUnlocked(start, size, [memory_type](Range* r) { r->memory_type = memory_type; });
-	}
-
 	bool Query(uint64_t addr, int flags, Range* out) {
 		EXIT_IF(out == nullptr);
 
@@ -702,7 +698,7 @@ public:
 	bool Release(uint64_t start, size_t len, uint64_t* vaddr, uint64_t* size,
 	             GpuAccessMode* gpu_mode);
 	bool Map(uint64_t vaddr, uint64_t phys_addr, size_t len, int prot, VirtualMemory::Mode mode,
-	         GpuAccessMode gpu_mode);
+	         GpuAccessMode gpu_mode, const char* name, int memory_type);
 	bool Unmap(uint64_t vaddr, uint64_t size, GpuAccessMode* gpu_mode,
 	           uint64_t* host_vaddr_to_release = nullptr);
 	bool Find(uint64_t vaddr, uint64_t* base_addr, size_t* len, int* prot,
@@ -715,7 +711,6 @@ public:
 	void ProtectMapping(uint64_t vaddr, uint64_t size, int prot, VirtualMemory::Mode mode,
 	                    GpuAccessMode gpu_mode);
 	void SetVirtualRangeName(uint64_t vaddr, uint64_t len, const char* name);
-	void SetVirtualRangeMemoryType(uint64_t vaddr, uint64_t len, int memory_type);
 
 	[[nodiscard]] Common::Mutex&                            GetMutex() { return m_mutex; }
 	[[nodiscard]] const std::map<uint64_t, AllocatedBlock>& GetPhysicalBlocks() const {
@@ -1368,7 +1363,8 @@ bool PhysicalMemory::Release(uint64_t start, size_t len, uint64_t* vaddr, uint64
 }
 
 bool PhysicalMemory::Map(uint64_t vaddr, uint64_t phys_addr, size_t len, int prot,
-                         VirtualMemory::Mode mode, GpuAccessMode gpu_mode) {
+                         VirtualMemory::Mode mode, GpuAccessMode gpu_mode, const char* name,
+                         int memory_type) {
 	Common::LockGuard lock(m_mutex);
 
 	if (len == 0 || UINT64_MAX - phys_addr < len) {
@@ -1413,6 +1409,8 @@ bool PhysicalMemory::Map(uint64_t vaddr, uint64_t phys_addr, size_t len, int pro
 	mapping.prot           = prot;
 	mapping.mode           = mode;
 	mapping.gpu_mode       = gpu_mode;
+	mapping.memory_type    = memory_type;
+	CopyVirtualRangeName(mapping.name, name);
 	m_mappings.push_back(mapping);
 	if (has_automatic) {
 		RemoveFreeRange(m_automatic_free, phys_addr, len);
@@ -2040,16 +2038,6 @@ void FlexibleMemory::SetVirtualRangeName(uint64_t vaddr, uint64_t len, const cha
 	}
 }
 
-void PhysicalMemory::SetVirtualRangeMemoryType(uint64_t vaddr, uint64_t len, int memory_type) {
-	Common::LockGuard lock(m_mutex);
-
-	for (auto& b: m_mappings) {
-		if (VirtualRangesOverlap(vaddr, len, b.map_vaddr, b.map_size)) {
-			b.memory_type = memory_type;
-		}
-	}
-}
-
 uint64_t FlexibleMemory::Available() {
 	Common::LockGuard lock(m_mutex);
 
@@ -2268,12 +2256,9 @@ static bool UnmapPooledBackingTransactional(const std::vector<PooledMemory::Mapp
 	return true;
 }
 
-int32_t KYTY_SYSV_ABI KernelMapNamedFlexibleMemory(void** addr_in_out, size_t len, int prot,
-                                                   int flags, const char* name) {
-	PRINT_NAME();
-
-	std::lock_guard<std::recursive_mutex> memory_operation_lock(g_memory_operation_mutex);
-
+// The ABI entry points and batch operations hold g_memory_operation_mutex.
+static int MapFlexibleMemory(void** addr_in_out, size_t len, int prot, int flags,
+                             const char* name) {
 	EXIT_NOT_IMPLEMENTED(addr_in_out == nullptr);
 
 	constexpr size_t   PAGE_SIZE                = 0x4000;
@@ -2398,8 +2383,17 @@ int32_t KYTY_SYSV_ABI KernelMapNamedFlexibleMemory(void** addr_in_out, size_t le
 	return OK;
 }
 
+int32_t KYTY_SYSV_ABI KernelMapNamedFlexibleMemory(void** addr_in_out, size_t len, int prot,
+                                                   int flags, const char* name) {
+	PRINT_NAME();
+	std::lock_guard<std::recursive_mutex> memory_operation_lock(g_memory_operation_mutex);
+	return MapFlexibleMemory(addr_in_out, len, prot, flags, name);
+}
+
 int KYTY_SYSV_ABI KernelMapFlexibleMemory(void** addr_in_out, size_t len, int prot, int flags) {
-	return KernelMapNamedFlexibleMemory(addr_in_out, len, prot, flags, "");
+	PRINT_NAME();
+	std::lock_guard<std::recursive_mutex> memory_operation_lock(g_memory_operation_mutex);
+	return MapFlexibleMemory(addr_in_out, len, prot, flags, "");
 }
 
 int KYTY_SYSV_ABI KernelSetPrtAperture(int index, void* addr, size_t len) {
@@ -2607,11 +2601,8 @@ static int UnmapMemoryRange(uint64_t vaddr, size_t len) {
 	return OK;
 }
 
-int KYTY_SYSV_ABI KernelMunmap(uint64_t vaddr, size_t len) {
-	PRINT_NAME();
-
-	std::lock_guard<std::recursive_mutex> memory_operation_lock(g_memory_operation_mutex);
-
+// The ABI entry point and batch operations hold g_memory_operation_mutex.
+static int UnmapMemory(uint64_t vaddr, size_t len) {
 	LOGF("\t start = 0x%016" PRIx64 "\n"
 	     "\t len   = 0x%016" PRIx64 "\n",
 	     vaddr, len);
@@ -2627,6 +2618,12 @@ int KYTY_SYSV_ABI KernelMunmap(uint64_t vaddr, size_t len) {
 	return UnmapMemoryRange(vaddr, len);
 }
 
+int KYTY_SYSV_ABI KernelMunmap(uint64_t vaddr, size_t len) {
+	PRINT_NAME();
+	std::lock_guard<std::recursive_mutex> memory_operation_lock(g_memory_operation_mutex);
+	return UnmapMemory(vaddr, len);
+}
+
 size_t KYTY_SYSV_ABI KernelGetDirectMemorySize() {
 	PRINT_NAME();
 
@@ -2909,7 +2906,8 @@ static int ReleaseDirectMemoryInternal(int64_t start, size_t len) {
 			for (const auto& removed: metadata_unmapped) {
 				EXIT_IF(!g_physical_memory->Map(removed.map_vaddr, removed.start_addr,
 				                                removed.map_size, removed.prot, removed.mode,
-				                                removed.gpu_mode));
+				                                removed.gpu_mode, removed.name,
+				                                removed.memory_type));
 			}
 			restore_owner_aliases(owner_unmapped);
 			restore_gpu_aliases();
@@ -2980,12 +2978,10 @@ int KYTY_SYSV_ABI KernelCheckedReleaseDirectMemory(int64_t start, size_t len) {
 	return result == KERNEL_ERROR_EACCES ? KERNEL_ERROR_ENOENT : result;
 }
 
-int KYTY_SYSV_ABI KernelMapDirectMemory(void** addr, size_t len, int prot, int flags,
-                                        int64_t direct_memory_start, size_t alignment) {
-	PRINT_NAME();
-
-	std::lock_guard<std::recursive_mutex> memory_operation_lock(g_memory_operation_mutex);
-
+// The ABI entry points and batch operations hold g_memory_operation_mutex.
+static int MapDirectMemory(void** addr, size_t len, int prot, int flags,
+                           int64_t direct_memory_start, size_t alignment, const char* name,
+                           int type = -1) {
 	if (addr == nullptr) {
 		return KERNEL_ERROR_EFAULT;
 	}
@@ -3036,10 +3032,8 @@ int KYTY_SYSV_ABI KernelMapDirectMemory(void** addr, size_t len, int prot, int f
 				                                           old.start_addr, old.mode));
 			}
 			EXIT_IF(!g_physical_memory->Map(old.map_vaddr, old.start_addr, old.map_size,
-			                               old.prot, old.mode, old.gpu_mode));
-			g_physical_memory->SetVirtualRangeName(old.map_vaddr, old.map_size, old.name);
-			g_physical_memory->SetVirtualRangeMemoryType(old.map_vaddr, old.map_size,
-			                                          old.memory_type);
+			                               old.prot, old.mode, old.gpu_mode, old.name,
+			                               old.memory_type));
 		}
 		MapGpuRange(in_addr, len);
 	};
@@ -3168,7 +3162,11 @@ int KYTY_SYSV_ABI KernelMapDirectMemory(void** addr, size_t len, int prot, int f
 		return KERNEL_ERROR_ENOMEM;
 	}
 
-	if (!g_physical_memory->Map(out_addr, direct_memory_start, len, prot, mode, gpu_mode)) {
+	PhysicalMemory::AllocatedBlock mapped_block {};
+	g_physical_memory->Find(direct_memory_start, false, &mapped_block);
+	const int memory_type = type == -1 ? mapped_block.memory_type : type;
+	if (!g_physical_memory->Map(out_addr, direct_memory_start, len, prot, mode, gpu_mode, name,
+	                            memory_type)) {
 		LOGF_COLOR(Log::Color::Red, "\t [Fail]\n");
 		if (replaced_mappings.empty()) {
 			EXIT_IF(!g_guest_address_space->UnmapBacking(out_addr, len));
@@ -3178,18 +3176,16 @@ int KYTY_SYSV_ABI KernelMapDirectMemory(void** addr, size_t len, int prot, int f
 		return KERNEL_ERROR_EBUSY;
 	}
 
-	PhysicalMemory::AllocatedBlock mapped_block {};
-	g_physical_memory->Find(direct_memory_start, false, &mapped_block);
 	const bool published = !replaced_mappings.empty() || reserved_target
 	                           ? g_virtual_ranges->ReplaceSpan(
 	                                 out_addr, len, replaced_mappings.empty()
 	                                                    ? VirtualRangeType::Reserved
 	                                                    : VirtualRangeType::Direct,
 	                                 direct_memory_start,
-	                                 prot, mapped_block.memory_type, VirtualRangeType::Direct, "")
+	                                 prot, memory_type, VirtualRangeType::Direct, name)
 	                           : g_virtual_ranges->Add(out_addr, len, direct_memory_start, prot,
-	                                                   mapped_block.memory_type,
-	                                                   VirtualRangeType::Direct, "");
+	                                                   memory_type,
+	                                                   VirtualRangeType::Direct, name);
 	if (!published) {
 		GpuAccessMode rollback_gpu_mode = GpuAccessMode::NoAccess;
 		EXIT_IF(!g_physical_memory->Unmap(out_addr, len, &rollback_gpu_mode));
@@ -3212,22 +3208,18 @@ int KYTY_SYSV_ABI KernelMapDirectMemory(void** addr, size_t len, int prot, int f
 	return OK;
 }
 
+int KYTY_SYSV_ABI KernelMapDirectMemory(void** addr, size_t len, int prot, int flags,
+                                        int64_t direct_memory_start, size_t alignment) {
+	PRINT_NAME();
+	std::lock_guard<std::recursive_mutex> memory_operation_lock(g_memory_operation_mutex);
+	return MapDirectMemory(addr, len, prot, flags, direct_memory_start, alignment, "");
+}
+
 int KYTY_SYSV_ABI KernelMapDirectMemory2(void** addr, size_t len, int type, int prot, int flags,
                                          int64_t direct_memory_start, size_t alignment) {
 	PRINT_NAME();
-
 	std::lock_guard<std::recursive_mutex> memory_operation_lock(g_memory_operation_mutex);
-
-	LOGF("\t type = %d\n", type);
-
-	auto ret = KernelMapDirectMemory(addr, len, prot, flags, direct_memory_start, alignment);
-	if (ret == OK && addr != nullptr && *addr != nullptr) {
-		const auto out_addr = reinterpret_cast<uint64_t>(*addr);
-		g_physical_memory->SetVirtualRangeMemoryType(out_addr, len, type);
-		g_virtual_ranges->SetMemoryType(out_addr, len, type);
-	}
-
-	return ret;
+	return MapDirectMemory(addr, len, prot, flags, direct_memory_start, alignment, "", type);
 }
 
 int MapAutomaticMemory(uint64_t vaddr, size_t size, int type, int prot) {
@@ -3246,11 +3238,11 @@ int MapAutomaticMemory(uint64_t vaddr, size_t size, int type, int prot) {
 	uint64_t mapped = 0;
 	for (const auto& [offset, length]: ranges) {
 		void*     address = reinterpret_cast<void*>(vaddr + mapped);
-		const int result  = KernelMapDirectMemory2(&address, length, type, prot, 0x10,
-		                                           static_cast<int64_t>(offset), page_size);
+		const int result  = MapDirectMemory(&address, length, prot, 0x10,
+		                                    static_cast<int64_t>(offset), page_size, "", type);
 		if (result != OK) {
 			if (mapped != 0) {
-				EXIT_IF(KernelMunmap(vaddr, mapped) != OK);
+				EXIT_IF(UnmapMemory(vaddr, mapped) != OK);
 			}
 			g_physical_memory->RestoreAutomatic(ranges);
 			return result;
@@ -3277,13 +3269,7 @@ int KYTY_SYSV_ABI KernelMapNamedDirectMemory(void** addr, size_t len, int prot,
 		return KERNEL_ERROR_ENAMETOOLONG;
 	}
 
-	auto ret = KernelMapDirectMemory(addr, len, prot, flags, direct_memory_start, alignment);
-	if (ret == OK && addr != nullptr) {
-		g_physical_memory->SetVirtualRangeName(reinterpret_cast<uint64_t>(*addr), len, name);
-		g_virtual_ranges->Rename(reinterpret_cast<uint64_t>(*addr), len, name);
-	}
-
-	return ret;
+	return MapDirectMemory(addr, len, prot, flags, direct_memory_start, alignment, name);
 }
 
 int KYTY_SYSV_ABI KernelIsAddressSanitizerEnabled() {
@@ -3389,7 +3375,8 @@ static bool ReplaceFixedRangeWithReserved(uint64_t start, uint64_t size) {
 				if (chunk.backend_unmapped) {
 					backend_restored = g_physical_memory->Map(
 					    chunk.range.start, chunk.range.offset, chunk.range.size,
-					    chunk.range.protection, chunk.mode, chunk.gpu_mode);
+					    chunk.range.protection, chunk.mode, chunk.gpu_mode, chunk.range.name,
+					    chunk.range.memory_type);
 				}
 			} else if (chunk.range.type == VirtualRangeType::Flexible && chunk.backend_unmapped) {
 				host_restored = backend_restored =
@@ -3584,6 +3571,10 @@ void TestBeforeNextBackingMap(callback_func_t callback) {
 	g_test_before_backing_map = callback;
 }
 
+void TestSetBackingReadCallback(callback_func_t callback) {
+	g_test_backing_read = callback;
+}
+
 void TestFailNextPhysicalMemoryUnmap() {
 	TestFailPhysicalMemoryUnmapAfter(0);
 }
@@ -3621,6 +3612,36 @@ uint64_t TestGuestBackingSize() {
 	return g_guest_address_space->GetBackingSize();
 }
 
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
+bool TestWindowsBackingViewModes() {
+	constexpr uint64_t size = 0x10000;
+	GuestBackingStore backing(size);
+	void* placeholder = VirtualAlloc2(GetCurrentProcess(), nullptr, size,
+	                                  MEM_RESERVE | MEM_RESERVE_PLACEHOLDER, PAGE_NOACCESS,
+	                                  nullptr, 0);
+	if (placeholder == nullptr) {
+		return false;
+	}
+	const auto address = reinterpret_cast<uint64_t>(placeholder);
+	bool valid = true;
+	for (auto [mode, protection]: std::array {
+	         std::pair {VirtualMemory::Mode::NoAccess, DWORD {PAGE_NOACCESS}},
+	         std::pair {VirtualMemory::Mode::Read, DWORD {PAGE_READONLY}},
+	         std::pair {VirtualMemory::Mode::ReadWrite, DWORD {PAGE_READWRITE}},
+	         std::pair {VirtualMemory::Mode::ExecuteReadWrite, DWORD {PAGE_EXECUTE_READWRITE}}}) {
+		if (!backing.MapFixed(address, size, 0, mode)) {
+			valid = false;
+			break;
+		}
+		MEMORY_BASIC_INFORMATION info {};
+		valid = VirtualQuery(placeholder, &info, sizeof(info)) != 0 &&
+		        info.Type == MEM_MAPPED && info.Protect == protection && valid;
+		EXIT_IF(!backing.Unmap(address, size));
+	}
+	return VirtualFree(placeholder, 0, MEM_RELEASE) != 0 && valid;
+}
+#endif
+
 bool TestGuestFreeRangeBounds() {
 	return GuestFreeRangeContains(0x10000, 0x20000, 0x18000, 0x4000) &&
 	       !GuestFreeRangeContains(0x10000, 0x20000, 0x40000, 0x4000) &&
@@ -3875,11 +3896,8 @@ bool FreeGuestMemory(uint64_t vaddr, uint64_t size) {
 	return FreeGuestMemoryOwner(vaddr, mapped_size);
 }
 
-int KYTY_SYSV_ABI KernelMprotect(const void* addr, size_t len, int prot) {
-	PRINT_NAME();
-
-	std::lock_guard<std::recursive_mutex> memory_operation_lock(g_memory_operation_mutex);
-
+// The ABI entry points and batch operations hold g_memory_operation_mutex.
+static int ProtectMemory(const void* addr, size_t len, int prot) {
 	auto vaddr = reinterpret_cast<uint64_t>(addr);
 
 	LOGF("\t addr = 0x%016" PRIx64 "\n"
@@ -3937,19 +3955,17 @@ int KYTY_SYSV_ABI KernelMprotect(const void* addr, size_t len, int prot) {
 	return OK;
 }
 
+int KYTY_SYSV_ABI KernelMprotect(const void* addr, size_t len, int prot) {
+	PRINT_NAME();
+	std::lock_guard<std::recursive_mutex> memory_operation_lock(g_memory_operation_mutex);
+	return ProtectMemory(addr, len, prot);
+}
+
 int KYTY_SYSV_ABI KernelMtypeprotect(const void* addr, size_t len, int type, int prot) {
 	PRINT_NAME();
-
 	std::lock_guard<std::recursive_mutex> memory_operation_lock(g_memory_operation_mutex);
-
-	LOGF("\t addr = 0x%016" PRIx64 "\n"
-	     "\t len  = 0x%016" PRIx64 "\n"
-	     "\t type = 0x%08" PRIx32 "\n"
-	     "\t prot = 0x%08" PRIx32 "\n",
-	     reinterpret_cast<uint64_t>(addr), static_cast<uint64_t>(len), static_cast<uint32_t>(type),
-	     static_cast<uint32_t>(prot));
-
-	return KernelMprotect(addr, len, prot);
+	LOGF("\t type = %d\n", type);
+	return ProtectMemory(addr, len, prot);
 }
 
 int KYTY_SYSV_ABI KernelBatchMap2(KernelBatchMapEntry* entries, int num_entries,
@@ -3996,23 +4012,19 @@ int KYTY_SYSV_ABI KernelBatchMap2(KernelBatchMapEntry* entries, int num_entries,
 		int result = OK;
 		switch (entry->operation) {
 			case MAP_OP_MAP_DIRECT:
-				result = KernelMapNamedDirectMemory(&entry->start, entry->length, entry->protection,
-				                                    flags, static_cast<int64_t>(entry->offset), 0,
-				                                    "anon");
+				result = MapDirectMemory(&entry->start, entry->length, entry->protection, flags,
+				                         static_cast<int64_t>(entry->offset), 0, "anon");
 				break;
 			case MAP_OP_UNMAP:
-				result = KernelMunmap(reinterpret_cast<uint64_t>(entry->start), entry->length);
+				result = UnmapMemory(reinterpret_cast<uint64_t>(entry->start), entry->length);
 				break;
 			case MAP_OP_PROTECT:
-				result = KernelMprotect(entry->start, entry->length, entry->protection);
+			case MAP_OP_TYPE_PROTECT:
+				result = ProtectMemory(entry->start, entry->length, entry->protection);
 				break;
 			case MAP_OP_MAP_FLEXIBLE:
-				result = KernelMapNamedFlexibleMemory(&entry->start, entry->length,
-				                                      entry->protection, flags, "anon");
-				break;
-			case MAP_OP_TYPE_PROTECT:
-				result =
-				    KernelMtypeprotect(entry->start, entry->length, entry->type, entry->protection);
+				result = MapFlexibleMemory(&entry->start, entry->length, entry->protection, flags,
+				                           "anon");
 				break;
 			default: result = KERNEL_ERROR_EINVAL; break;
 		}

src/kernel/memory.h

@@ -192,6 +192,7 @@ bool FreeGuestMemory(uint64_t vaddr, uint64_t size);
 
 #if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
 void     TestBeforeNextBackingMap(callback_func_t callback);
+void     TestSetBackingReadCallback(callback_func_t callback);
 void     TestFailNextPhysicalMemoryUnmap();
 void     TestFailPhysicalMemoryUnmapAfter(uint32_t successful_unmaps);
 void     TestFailGuestBackingStoreUnmapAfter(uint32_t successful_unmaps);

src/kernel/memoryAddressSpace.inc

@@ -161,30 +161,90 @@ public:
 		if (!IsAvailable() || !RangeOk(backing_offset, size)) {
 			return false;
 		}
+		std::lock_guard lock(m_mutex);
 		std::memset(m_backing_base + backing_offset, 0, size);
 		return true;
 	}
 
 	bool TryWriteBacking(uint64_t vaddr, const void* data, uint64_t size) {
-		if (data == nullptr) {
+		if (!IsAvailable() || data == nullptr || size == 0 || UINT64_MAX - vaddr < size) {
 			return false;
 		}
+
+		std::lock_guard lock(m_mutex);
+		const auto      end   = vaddr + size;
+		auto            first = m_maps.upper_bound(vaddr);
+		if (first == m_maps.begin()) {
+			return false;
+		}
+		--first;
+		auto current = vaddr;
+		for (auto mapping = first; current < end; ++mapping) {
+			if (mapping == m_maps.end() || current < mapping->second.vaddr ||
+			    current >= mapping->second.vaddr + mapping->second.size) {
+				return false;
+			}
+			const auto& entry = mapping->second;
+			const auto  bytes = std::min(end, entry.vaddr + entry.size) - current;
+			if (!RangeOk(entry.backing_offset + current - entry.vaddr, bytes)) {
+				return false;
+			}
+			current += bytes;
+		}
+
 		const auto* source = static_cast<const uint8_t*>(data);
-		return TryTransferBacking(vaddr, size,
-		                          [source](uint8_t* backing, uint64_t data_offset, size_t bytes) {
-			                          std::memcpy(backing, source + data_offset, bytes);
-		                          });
+		current = vaddr;
+		for (auto mapping = first; current < end; ++mapping) {
+			const auto& entry  = mapping->second;
+			const auto  bytes  = std::min(end, entry.vaddr + entry.size) - current;
+			const auto  offset = entry.backing_offset + current - entry.vaddr;
+			std::memcpy(m_backing_base + offset, source + (current - vaddr), bytes);
+			current += bytes;
+		}
+		return true;
 	}
 
 	bool TryReadBacking(uint64_t vaddr, void* data, uint64_t size) {
-		if (data == nullptr) {
+		if (!IsAvailable() || data == nullptr || size == 0 || UINT64_MAX - vaddr < size) {
 			return false;
 		}
+
+		std::shared_lock lock(m_mutex);
+#if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
+		if (g_test_backing_read != nullptr) {
+			g_test_backing_read(vaddr, size);
+		}
+#endif
+		const auto end   = vaddr + size;
+		auto       first = m_maps.upper_bound(vaddr);
+		if (first == m_maps.begin()) {
+			return false;
+		}
+		--first;
+		auto current = vaddr;
+		for (auto mapping = first; current < end; ++mapping) {
+			if (mapping == m_maps.end() || current < mapping->second.vaddr ||
+			    current >= mapping->second.vaddr + mapping->second.size) {
+				return false;
+			}
+			const auto& entry = mapping->second;
+			const auto  bytes = std::min(end, entry.vaddr + entry.size) - current;
+			if (!RangeOk(entry.backing_offset + current - entry.vaddr, bytes)) {
+				return false;
+			}
+			current += bytes;
+		}
+
 		auto* destination = static_cast<uint8_t*>(data);
-		return TryTransferBacking(
-		    vaddr, size, [destination](uint8_t* backing, uint64_t data_offset, size_t bytes) {
-			    std::memcpy(destination + data_offset, backing, bytes);
-		    });
+		current = vaddr;
+		for (auto mapping = first; current < end; ++mapping) {
+			const auto& entry  = mapping->second;
+			const auto  bytes  = std::min(end, entry.vaddr + entry.size) - current;
+			const auto  offset = entry.backing_offset + current - entry.vaddr;
+			std::memcpy(destination + (current - vaddr), m_backing_base + offset, bytes);
+			current += bytes;
+		}
+		return true;
 	}
 
 	bool TryReadSparseBacking(uint64_t vaddr, void* data, uint64_t size) {
@@ -192,29 +252,24 @@ public:
 			return false;
 		}
 
-		std::lock_guard lock(m_mutex);
-		const auto      end   = vaddr + size;
-		auto            first = m_maps.lower_bound(vaddr);
-		if (first != m_maps.begin()) {
-			auto previous = std::prev(first);
-			if (previous->second.vaddr + previous->second.size > vaddr) {
-				first = previous;
-			}
+		std::shared_lock lock(m_mutex);
+#if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
+		if (g_test_backing_read != nullptr) {
+			g_test_backing_read(vaddr, size);
 		}
-		for (auto mapping = first; mapping != m_maps.end() && mapping->second.vaddr < end;
-		     ++mapping) {
-			const auto& entry = mapping->second;
-			const auto  start = std::max(vaddr, entry.vaddr);
-			const auto  bytes = std::min(end, entry.vaddr + entry.size) - start;
-			if (!RangeOk(entry.backing_offset + start - entry.vaddr, bytes)) {
-				return false;
+#endif
+		const auto end     = vaddr + size;
+		auto       mapping = m_maps.lower_bound(vaddr);
+		if (mapping != m_maps.begin()) {
+			auto previous = std::prev(mapping);
+			if (previous->second.vaddr + previous->second.size > vaddr) {
+				mapping = previous;
 			}
 		}
 
 		auto* destination = static_cast<uint8_t*>(data);
 		auto  cursor      = vaddr;
-		for (auto mapping = first; mapping != m_maps.end() && mapping->second.vaddr < end;
-		     ++mapping) {
+		for (; mapping != m_maps.end() && mapping->second.vaddr < end; ++mapping) {
 			const auto& entry  = mapping->second;
 			const auto  start  = std::max(vaddr, entry.vaddr);
 			const auto  bytes  = std::min(end, entry.vaddr + entry.size) - start;
@@ -437,7 +492,7 @@ public:
 			return false;
 		}
 
-		std::lock_guard lock(m_mutex);
+		std::shared_lock lock(m_mutex);
 		const auto      end     = vaddr + size;
 		auto            current = vaddr;
 
@@ -457,39 +512,6 @@ public:
 	}
 
 private:
-	template <typename Transfer>
-	bool TryTransferBacking(uint64_t vaddr, uint64_t size, Transfer transfer) {
-		if (!IsAvailable() || size == 0 || UINT64_MAX - vaddr < size) {
-			return false;
-		}
-
-		std::lock_guard lock(m_mutex);
-		const auto      end     = vaddr + size;
-		auto            current = vaddr;
-		while (current < end) {
-			MapEntry entry {};
-			if (!FindContainingUnlocked(current, 1, &entry)) {
-				return false;
-			}
-			const auto bytes = std::min(end, entry.vaddr + entry.size) - current;
-			if (!RangeOk(entry.backing_offset + current - entry.vaddr, bytes)) {
-				return false;
-			}
-			current += bytes;
-		}
-
-		current = vaddr;
-		while (current < end) {
-			MapEntry entry {};
-			EXIT_IF(!FindContainingUnlocked(current, 1, &entry));
-			const auto bytes  = std::min(end, entry.vaddr + entry.size) - current;
-			const auto offset = entry.backing_offset + current - entry.vaddr;
-			transfer(m_backing_base + offset, current - vaddr, static_cast<size_t>(bytes));
-			current += bytes;
-		}
-		return true;
-	}
-
 	struct MapEntry {
 		uint64_t            vaddr          = 0;
 		uint64_t            size           = 0;
@@ -597,7 +619,7 @@ private:
 			return nullptr;
 		}
 
-		if (!ApplyProtection(ptr, size, mode)) {
+		if (mode == VirtualMemory::Mode::NoAccess && !ApplyProtection(ptr, size, mode)) {
 			UnmapViewOfFile2(GetCurrentProcess(), ptr, MEM_PRESERVE_PLACEHOLDER);
 			SetFailure(failure_reason, FailureReason::ProtectFailed);
 			return nullptr;
@@ -656,7 +678,7 @@ private:
 
 	uint8_t*                     m_backing_base = nullptr;
 	uint64_t                     m_size         = 0;
-	std::mutex                   m_mutex;
+	std::shared_mutex            m_mutex;
 	std::map<uint64_t, MapEntry> m_maps;
 };
 
@@ -727,7 +749,7 @@ public:
 		if (size == 0 || UINT64_MAX - vaddr < size) {
 			return false;
 		}
-		std::lock_guard lock(m_mutex);
+		std::shared_lock lock(m_mutex);
 		return std::any_of(m_owned.begin(), m_owned.end(), [vaddr, size](const auto& range) {
 			return vaddr >= range.first && vaddr + size <= range.first + range.second;
 		});
@@ -755,7 +777,7 @@ public:
 		if (!Owns(vaddr, size)) {
 			return false;
 		}
-		std::lock_guard lock(m_mutex);
+		std::shared_lock lock(m_mutex);
 		auto            mapping = m_mapped.upper_bound(vaddr);
 		if (mapping != m_mapped.begin()) {
 			--mapping;
@@ -1381,7 +1403,7 @@ private:
 #endif
 	}
 
-	std::mutex                                 m_mutex;
+	std::shared_mutex                          m_mutex;
 	std::map<uint64_t, uint64_t>               m_free;
 	std::map<uint64_t, MappedRegion>            m_mapped;
 	std::vector<std::pair<uint64_t, uint64_t>> m_owned;

src/kernel/pthread.cpp

@@ -3200,6 +3200,15 @@ static void* RunThread(void* arg) {
 	os_thread_id = GetHostThreadId();
 #endif
 	thread->host_thread_id = os_thread_id;
+#if KYTY_PLATFORM == KYTY_PLATFORM_LINUX
+	if (!thread->name.empty()) {
+#if defined(__APPLE__)
+		pthread_setname_np(thread->name.substr(0, 63).c_str());
+#else
+		pthread_setname_np(pthread_self(), thread->name.substr(0, 15).c_str());
+#endif
+	}
+#endif
 
 	LOGF("\tPthread run begin: %s, id = %d, os_thread_id = %" PRIu64 ", entry = 0x%016" PRIx64
 	     ", arg = 0x%016" PRIx64 ", stack_addr = 0x%016" PRIx64 ", stack_size = %" PRIu64 "\n",

src/launcher/CMakeLists.txt

@@ -96,6 +96,7 @@ install(TARGETS launcher
 
 if(APPLE)
 	install(TARGETS kyty_emulator DESTINATION "KytyPS5.app/Contents/MacOS")
+	install(FILES "${KYTY_TROPHY_UNLOCK_SOUND}" DESTINATION "KytyPS5.app/Contents/Resources/assets/sounds")
 	# Deploy and ad-hoc sign the Qt frameworks and plugins in the bundle.
 	find_program(KYTY_MACDEPLOYQT NAMES macdeployqt PATHS "${Qt6_DIR}/../../../bin" NO_DEFAULT_PATH)
 	if(NOT KYTY_MACDEPLOYQT)

src/launcher/forms/configuration_edit_dialog.ui

@@ -6,7 +6,7 @@
    <rect>
     <x>0</x>
     <y>0</y>
-    <width>760</width>
+    <width>900</width>
     <height>620</height>
    </rect>
   </property>
@@ -179,10 +179,37 @@
           <property name="text"><string>Auto-hide cursor</string></property>
          </widget>
         </item>
+        <item row="9" column="1">
+         <widget class="QCheckBox" name="checkBox_skip_notice_screen">
+          <property name="toolTip"><string>Skip startup logos and notices in supported games.</string></property>
+          <property name="text"><string>Skip boot logos</string></property>
+         </widget>
+        </item>
+        <item row="10" column="1">
+         <widget class="QCheckBox" name="checkBox_sync_raw_image_buffers">
+          <property name="toolTip"><string>Can fix missing textures and incorrect transparency in some games. May reduce performance.</string></property>
+          <property name="text"><string>Accurate texture synchronization</string></property>
+         </widget>
+        </item>
        </layout>
       </widget>
      </item>
-     <item row="3" column="0" colspan="2">
+     <item row="3" column="0">
+      <widget class="QGroupBox" name="notifications_group">
+       <property name="title"><string>Notifications</string></property>
+       <layout class="QVBoxLayout" name="notificationsLayout">
+        <item>
+         <widget class="QCheckBox" name="checkBox_trophy_notifications">
+          <property name="toolTip"><string>Show trophy unlock notifications and play their sound.
+To use your own sound, replace the emulator's assets/sounds/trophy-unlock.wav with a WAV file and restart the game.</string></property>
+          <property name="text"><string>Enable trophy notifications</string></property>
+          <property name="checked"><bool>true</bool></property>
+         </widget>
+        </item>
+       </layout>
+      </widget>
+     </item>
+     <item row="4" column="0" colspan="2">
       <widget class="QGroupBox" name="debug_group">
        <property name="title">
         <string>Debugging &amp;&amp; logs</string>
@@ -385,6 +412,13 @@
        <property name="toolTip"><string>Restore default settings</string></property>
       </widget>
      </item>
+     <item>
+      <widget class="QPushButton" name="settings_button">
+       <property name="minimumSize"><size><width>80</width><height>25</height></size></property>
+       <property name="text"><string>Game Configs</string></property>
+       <property name="toolTip"><string>Import or export game configs</string></property>
+      </widget>
+     </item>
      <item>
       <spacer name="buttonSpacer">
        <property name="orientation"><enum>Qt::Horizontal</enum></property>

src/launcher/forms/configuration_list_widget.ui

@@ -125,7 +125,7 @@
             </size>
            </property>
            <property name="toolTip">
-            <string>&lt;html&gt;&lt;head/&gt;&lt;body&gt;&lt;p&gt;Edit configuration&lt;/p&gt;&lt;/body&gt;&lt;/html&gt;</string>
+            <string>Edit game config</string>
            </property>
            <property name="text">
             <string/>
@@ -150,7 +150,7 @@
             </size>
            </property>
            <property name="toolTip">
-            <string>&lt;html&gt;&lt;head/&gt;&lt;body&gt;&lt;p&gt;Delete configuration&lt;/p&gt;&lt;/body&gt;&lt;/html&gt;</string>
+            <string>Clear game config</string>
            </property>
            <property name="styleSheet">
             <string notr="true"/>
@@ -169,6 +169,28 @@
            </property>
          </widget>
         </item>
+         <item>
+          <widget class="QToolButton" name="trophy_overview_button">
+           <property name="minimumSize">
+            <size>
+             <width>36</width>
+             <height>36</height>
+            </size>
+           </property>
+           <property name="toolTip">
+            <string>View saved trophies</string>
+           </property>
+           <property name="iconSize">
+            <size>
+             <width>32</width>
+             <height>32</height>
+            </size>
+           </property>
+           <property name="autoRaise">
+            <bool>true</bool>
+           </property>
+          </widget>
+         </item>
          <item>
           <spacer name="horizontalSpacer">
            <property name="orientation">

src/launcher/icons/refresh.svg

@@ -1 +1,6 @@
-<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" width="24" height="24"><path fill="#000" d="M17.65 6.35A7.96 7.96 0 0 0 12 4a8 8 0 1 0 7.73 10h-2.08A6 6 0 1 1 12 6c1.66 0 3.14.69 4.22 1.78L13 11h7V4l-2.35 2.35z"/></svg>
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 32 32">
+  <g fill="none" stroke="#000000" stroke-linecap="round" stroke-linejoin="round" stroke-width="2">
+    <path d="M26 16a10 10 0 1 1-2.93-7.07L27 13"/>
+    <path d="M27 5v8h-8"/>
+  </g>
+</svg>

src/launcher/icons/trophy.svg

@@ -0,0 +1,7 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 32 32">
+  <g fill="none" stroke="#000000" stroke-linecap="round" stroke-linejoin="round" stroke-width="2">
+    <path d="M9 5h14v8a7 7 0 0 1-14 0Z"/>
+    <path d="M9 8H5v3a4 4 0 0 0 4 4M23 8h4v3a4 4 0 0 1-4 4"/>
+    <path d="M16 20v4M12 28v-4h8v4M11 28h10"/>
+  </g>
+</svg>

src/launcher/include/cheatFile.h

@@ -0,0 +1,26 @@
+#ifndef CHEAT_FILE_H
+#define CHEAT_FILE_H
+
+#include <QByteArray>
+#include <QJsonDocument>
+#include <QList>
+#include <QString>
+
+namespace Cheats {
+
+[[nodiscard]] bool    IsSupportedTitleId(const QString& title_id);
+[[nodiscard]] QString PlanPath(const QString& title_id);
+
+// Local files retain their existing defaults; remote imports receive additional validation.
+[[nodiscard]] QString       ParseDocument(const QByteArray& data, QJsonDocument* document);
+[[nodiscard]] QString       ValidateImport(const QJsonDocument& document, const QString& title_id,
+                                           const QString& version, const QString& process);
+[[nodiscard]] QJsonDocument WithSelection(const QJsonDocument& document,
+                                          const QList<bool>&   enabled);
+// Refuse to replace a file changed since it was loaded; empty data also permits creation.
+[[nodiscard]] QString SaveDocument(const QString& path, const QJsonDocument& document,
+                                   const QByteArray& expected_data);
+
+} // namespace Cheats
+
+#endif // CHEAT_FILE_H

src/launcher/include/cheatRepository.h

@@ -0,0 +1,66 @@
+#ifndef CHEAT_REPOSITORY_H
+#define CHEAT_REPOSITORY_H
+
+#include <QByteArray>
+#include <QList>
+#include <QNetworkAccessManager>
+#include <QObject>
+#include <QPointer>
+#include <QString>
+#include <QStringList>
+#include <QUrl>
+
+#include <functional>
+
+class QNetworkReply;
+
+namespace Cheats {
+
+struct RemoteFile {
+	QString name;
+	QString title;
+	QString version;
+	QString format;
+
+	[[nodiscard]] bool IsSupported() const;
+	[[nodiscard]] QUrl Url() const;
+};
+
+struct CatalogResult {
+	QList<RemoteFile> files;
+	QString           error;
+};
+
+[[nodiscard]] CatalogResult ParseIndex(const QByteArray& data, const QString& format,
+                                       const QString& title_id);
+
+class CheatRepository final: public QObject {
+	Q_OBJECT
+
+public:
+	explicit CheatRepository(QObject* parent = nullptr);
+	void Load(const QString& title_id);
+	void Fetch(const RemoteFile& file);
+	void CancelFetch();
+
+signals:
+	void CatalogLoaded(QList<Cheats::RemoteFile> files, QString error);
+	void FileLoaded(QByteArray data, QString error);
+
+private:
+	QNetworkReply* Request(const QUrl& url, qint64 limit,
+	                       std::function<void(QByteArray, QString)> complete);
+
+	QNetworkAccessManager   m_network;
+	QList<QNetworkReply*>   m_catalog_replies;
+	QPointer<QNetworkReply> m_file_reply;
+	QList<RemoteFile>       m_files;
+	QStringList             m_errors;
+	int                     m_catalog_generation = 0;
+	int                     m_fetch_generation   = 0;
+	int                     m_pending            = 0;
+};
+
+} // namespace Cheats
+
+#endif // CHEAT_REPOSITORY_H

src/launcher/include/configuration.h

@@ -6,6 +6,7 @@
 #include <QByteArray>
 #include <QChar>
 #include <QColor>
+#include <QJsonObject>
 #include <QMetaEnum>
 #include <QMetaType>
 #include <QObject>
@@ -14,9 +15,6 @@
 #include <QStringList>
 #include <QVariant>
 
-#define KYTY_CFG_SET(n) s->setValue(#n, QVariant::fromValue(n).toString());
-#define KYTY_CFG_GET(n) n = s->value(#n).value<decltype(n)>();
-
 template <class T>
 inline QStringList EnumToList() {
 	QStringList ret;
@@ -121,11 +119,14 @@ public:
 	bool                   fullscreen_enabled          = false;
 	bool                   hide_cursor_enabled         = false;
 	bool                   readback_linear_images      = false;
+	bool                   sync_raw_image_buffers      = false;
 	bool                   tessellation_enabled        = false;
+	bool                   trophy_enabled              = true;
+	bool                   skip_notice_screen          = false;
 	int                    vblank_frequency            = 60;
 	int                    console_language            = DEFAULT_CONSOLE_LANGUAGE;
 	bool                   vulkan_validation_enabled   = false;
-	bool                   shader_validation_enabled   = true;
+	bool                   shader_validation_enabled   = false;
 	ShaderOptimizationType shader_optimization_type    = ShaderOptimizationType::Performance;
 	LogDirection           shader_log_direction        = LogDirection::Silent;
 	QString                shader_log_folder           = "_Shaders";
@@ -153,7 +154,10 @@ public:
 		fullscreen_enabled          = other.fullscreen_enabled;
 		hide_cursor_enabled         = other.hide_cursor_enabled;
 		readback_linear_images      = other.readback_linear_images;
+		sync_raw_image_buffers      = other.sync_raw_image_buffers;
 		tessellation_enabled        = other.tessellation_enabled;
+		trophy_enabled              = other.trophy_enabled;
+		skip_notice_screen          = other.skip_notice_screen;
 		vblank_frequency            = other.vblank_frequency;
 		console_language            = other.console_language;
 		vulkan_validation_enabled   = other.vulkan_validation_enabled;
@@ -186,88 +190,16 @@ public:
 		game_comment    = other.game_comment;
 	}
 
-	void WriteSettings(QSettings* s) const {
-		KYTY_CFG_SET(name);
-		KYTY_CFG_SET(basedir);
-		KYTY_CFG_SET(game_path);
-		KYTY_CFG_SET(custom_settings);
-		KYTY_CFG_SET(screen_resolution);
-		KYTY_CFG_SET(user_name);
-		KYTY_CFG_SET(user_id);
-		KYTY_CFG_SET(audio_input_device);
-		KYTY_CFG_SET(present_mode);
-		KYTY_CFG_SET(gpu_index);
-		KYTY_CFG_SET(fullscreen_enabled);
-		KYTY_CFG_SET(hide_cursor_enabled);
-		KYTY_CFG_SET(readback_linear_images);
-		KYTY_CFG_SET(tessellation_enabled);
-		KYTY_CFG_SET(vblank_frequency);
-		KYTY_CFG_SET(console_language);
-		KYTY_CFG_SET(vulkan_validation_enabled);
-		KYTY_CFG_SET(shader_validation_enabled);
-		KYTY_CFG_SET(shader_optimization_type);
-		KYTY_CFG_SET(shader_log_direction);
-		KYTY_CFG_SET(shader_log_folder);
-		KYTY_CFG_SET(command_buffer_dump_enabled);
-		KYTY_CFG_SET(command_buffer_dump_folder);
-		KYTY_CFG_SET(printf_direction);
-		KYTY_CFG_SET(printf_output_file);
-		KYTY_CFG_SET(profiler_enabled);
-		KYTY_CFG_SET(renderdoc_enabled);
-		KYTY_CFG_SET(amd_cpu_enabled);
-#if defined(_WIN32)
-		KYTY_CFG_SET(red_zone_protection_enabled);
-#endif
-		s->setValue("host_input_mapping", host_input_mapping);
-		KYTY_CFG_SET(elf);
-	}
+	void WriteSettings(QSettings* s) const;
+	void ReadSettings(QSettings* s);
 
-	void ReadSettings(QSettings* s) {
-		KYTY_CFG_GET(name);
-		KYTY_CFG_GET(basedir);
-		KYTY_CFG_GET(game_path);
-		KYTY_CFG_GET(custom_settings);
-		KYTY_CFG_GET(screen_resolution);
-		user_name          = s->value("user_name", user_name).toString();
-		bool user_id_ok    = false;
-		auto saved_user_id = s->value("user_id", user_id).toInt(&user_id_ok);
-		user_id            = user_id_ok && Config::IsConfiguredUserIdValid(saved_user_id)
-		                         ? saved_user_id
-		                         : Config::DEFAULT_USER_ID;
-		audio_input_device = s->value("audio_input_device", audio_input_device).toString();
-		KYTY_CFG_GET(present_mode);
-		gpu_index = s->value("gpu_index", -1).toInt();
-		if (EnumToText(present_mode).isEmpty()) {
-			present_mode = PresentMode::Mailbox;
-		}
-		KYTY_CFG_GET(fullscreen_enabled);
-		KYTY_CFG_GET(hide_cursor_enabled);
-		KYTY_CFG_GET(readback_linear_images);
-		KYTY_CFG_GET(tessellation_enabled);
-		vblank_frequency = s->value("vblank_frequency", vblank_frequency).toInt();
-		console_language = s->value("console_language", console_language).toInt();
-		if (console_language < 0 || console_language > MAX_CONSOLE_LANGUAGE) {
-			console_language = DEFAULT_CONSOLE_LANGUAGE;
-		}
-		KYTY_CFG_GET(vulkan_validation_enabled);
-		KYTY_CFG_GET(shader_validation_enabled);
-		KYTY_CFG_GET(shader_optimization_type);
-		KYTY_CFG_GET(shader_log_direction);
-		KYTY_CFG_GET(shader_log_folder);
-		KYTY_CFG_GET(command_buffer_dump_enabled);
-		KYTY_CFG_GET(command_buffer_dump_folder);
-		KYTY_CFG_GET(printf_direction);
-		KYTY_CFG_GET(printf_output_file);
-		KYTY_CFG_GET(profiler_enabled);
-		KYTY_CFG_GET(renderdoc_enabled);
-		amd_cpu_enabled = s->value("amd_cpu_enabled", false).toBool();
-#if defined(_WIN32)
-		red_zone_protection_enabled =
-		    s->value("red_zone_protection_enabled", red_zone_protection_enabled).toBool();
-#endif
-		host_input_mapping = s->value("host_input_mapping", host_input_mapping).toStringList();
-		elf                = s->value("elf", elf).toString();
-	}
+	[[nodiscard]] QVariantMap GameSettings() const;
+	// Call on a temporary configuration: validation can fail after loading values.
+	bool SetGameSettings(const QJsonObject& settings, QString& error);
+
+private:
+	template <class Settings>
+	void ReadGameSettingsValues(const Settings& s);
 };
 
 Q_DECLARE_METATYPE(Configuration*)

src/launcher/include/configurationEditDialog.h

@@ -33,6 +33,8 @@ public:
 
 signals:
 	void PreviewControllerColor(const QString& color);
+	void ImportGameSettings();
+	void ExportGameSettings();
 
 private:
 	Ui::ConfigurationEditDialog* m_ui = nullptr;

src/launcher/include/configurationListWidget.h

@@ -28,6 +28,7 @@ public:
 	~ConfigurationListWidget() override;
 
 	void SetRunEnabled(bool flag) { m_run_enabled = flag; }
+	void SetRuntimeDirectory(const QString& directory) { m_runtime_directory = directory; }
 
 	[[nodiscard]] const ConfigurationItem* GetSelectedItem() const { return m_selected_item; }
 	ConfigurationItem*                     GetSelectedItem() { return m_selected_item; }
@@ -42,6 +43,7 @@ public:
 	bool EnsureGameDirectory();
 	void ScanGameDirectory();
 	void ViewTrophies();
+	void ViewTrophyOverview();
 
 signals:
 
@@ -72,12 +74,15 @@ private:
 	void               SelectItem(QTreeWidgetItem* witem);
 	void               ApplyCompatibility();
 	void               UpdateToolbarIcons();
+	void               ImportGameSettings(QWidget* parent);
+	void               ExportGameSettings(QWidget* parent) const;
 	[[nodiscard]] bool HasValidGameDirectory() const;
 
 	ConfigurationItem*            m_selected_item = nullptr;
 	bool                          m_run_enabled   = true;
 	Ui::ConfigurationListWidget*  m_ui            = nullptr;
 	QString                       m_settings_file;
+	QString                       m_runtime_directory;
 	QStringList                   m_game_dirs;
 	Configuration                 m_global_info;
 	QMap<QString, Configuration*> m_custom_infos;

src/launcher/include/gameContent.h

@@ -12,7 +12,6 @@ namespace GameContent {
 
 inline constexpr uint64_t MaxMetadataSize      = uint64_t {1} << 20u;
 inline constexpr uint64_t MaxImageSize         = uint64_t {32} << 20u;
-inline constexpr uint64_t MaxTrophyPackageSize = uint64_t {128} << 20u;
 
 [[nodiscard]] std::filesystem::path ToPath(const QString& path);
 [[nodiscard]] QString               FromPath(const std::filesystem::path& path);

src/launcher/include/patchesDialog.h

@@ -1,30 +1,51 @@
 #ifndef PATCHES_DIALOG_H
 #define PATCHES_DIALOG_H
 
+#include "cheatRepository.h"
+
+#include <QByteArray>
 #include <QDialog>
+#include <QJsonDocument>
 #include <QString>
 
 class Configuration;
 class QLabel;
 class QListWidget;
 class QPushButton;
+class QTabWidget;
+class QTreeWidget;
 
 class PatchesDialog final: public QDialog {
 public:
 	explicit PatchesDialog(const Configuration& game, QWidget* parent = nullptr);
 	~PatchesDialog() override = default;
 
-	[[nodiscard]] static bool    IsSupportedTitleId(const QString& title_id);
-	[[nodiscard]] static QString PatchPlanPath(const QString& title_id);
-
 private:
 	void Load();
 	void Save();
+	void LoadRemote();
+	void PreviewRemote();
+	void ImportRemote();
 
-	QString      m_title_id;
-	QListWidget* m_patches = nullptr;
-	QLabel*      m_status  = nullptr;
-	QPushButton* m_apply   = nullptr;
+	QString                   m_title_id;
+	QString                   m_version;
+	QString                   m_process;
+	QByteArray                m_local_data;
+	QJsonDocument             m_local_document;
+	QJsonDocument             m_remote_document;
+	bool                      m_local_readable = false;
+	Cheats::CheatRepository*  m_repository     = nullptr;
+	QList<Cheats::RemoteFile> m_remote_files;
+	QTabWidget*               m_tabs           = nullptr;
+	QTreeWidget*              m_files          = nullptr;
+	QListWidget*              m_preview        = nullptr;
+	QLabel*                   m_remote_status  = nullptr;
+	QLabel*                   m_preview_status = nullptr;
+	QPushButton*              m_refresh        = nullptr;
+	QPushButton*              m_import         = nullptr;
+	QListWidget*              m_patches        = nullptr;
+	QLabel*                   m_status         = nullptr;
+	QPushButton*              m_apply          = nullptr;
 };
 
 #endif // PATCHES_DIALOG_H

src/launcher/include/trophyViewerDialog.h

@@ -3,6 +3,8 @@
 
 #include <QDialog>
 
+#include <vector>
+
 class QTabWidget;
 class QWidget;
 
@@ -12,10 +14,13 @@ public:
 	explicit TrophyViewerDialog(QWidget* parent = nullptr);
 
 	static bool HasTrophyData(const Configuration* info);
-	static void ShowForGame(const Configuration* info, QWidget* parent);
+	static void ShowForGame(const Configuration* info, const QString& runtime_directory,
+	                        QWidget* parent);
+	static void ShowOverview(const std::vector<const Configuration*>& games,
+	                         const QString& runtime_directory, QWidget* parent);
 
 private:
-	bool LoadGame(const Configuration& info, QString& error);
+	bool LoadGame(const Configuration& info, const QString& runtime_directory, QString& error);
 
 	QTabWidget* m_tabs = nullptr;
 };

src/launcher/resources.qrc

@@ -5,6 +5,8 @@
   <file alias="input-mapping.svg">icons/input-mapping.svg</file>
   <file alias="refresh.svg">icons/refresh.svg</file>
   <file alias="remove-configuration.svg">icons/remove-configuration.svg</file>
+  <file alias="trophy.svg">icons/trophy.svg</file>
+  <file alias="hidden-trophy.png">../../assets/images/hidden-trophy.png</file>
  </qresource>
  <qresource prefix="/styles">
   <file alias="launcher.qss">styles/launcher.qss</file>

src/launcher/src/cheatFile.cpp

@@ -0,0 +1,180 @@
+#include "cheatFile.h"
+
+#include <QCoreApplication>
+#include <QDir>
+#include <QFile>
+#include <QFileInfo>
+#include <QJsonArray>
+#include <QJsonObject>
+#include <QRegularExpression>
+#include <QSaveFile>
+
+namespace Cheats {
+
+namespace {
+
+QString Tr(const char* text) {
+	return QCoreApplication::translate("Cheats", text);
+}
+
+QString ValidateDocument(const QJsonDocument& document) {
+	if (!document.isObject()) {
+		return Tr("Invalid cheat JSON.");
+	}
+	const auto mods = document.object().value(QStringLiteral("mods"));
+	if (!mods.isArray()) {
+		return Tr("Cheat JSON has no \"mods\" array.");
+	}
+	for (const auto& value: mods.toArray()) {
+		const auto mod     = value.toObject();
+		const auto name    = mod.value(QStringLiteral("name"));
+		const auto enabled = mod.value(QStringLiteral("enabled"));
+		if (!value.isObject() || !name.isString() || name.toString().isEmpty() ||
+		    (!enabled.isUndefined() && !enabled.isBool())) {
+			return Tr("Invalid cheat entry.");
+		}
+	}
+	return {};
+}
+
+} // namespace
+
+bool IsSupportedTitleId(const QString& title_id) {
+	static const QRegularExpression pattern(QStringLiteral("\\APPSA[0-9]{5}\\z"));
+	return pattern.match(title_id.trimmed().toUpper()).hasMatch();
+}
+
+QString PlanPath(const QString& title_id) {
+	if (!IsSupportedTitleId(title_id)) {
+		return {};
+	}
+	return QDir(QCoreApplication::applicationDirPath())
+	    .filePath(QStringLiteral("_Patches/%1.json").arg(title_id.trimmed().toUpper()));
+}
+
+QString ParseDocument(const QByteArray& data, QJsonDocument* document) {
+	const auto parsed = QJsonDocument::fromJson(data);
+	const auto error  = ValidateDocument(parsed);
+	*document         = error.isEmpty() ? parsed : QJsonDocument();
+	return error;
+}
+
+QString ValidateImport(const QJsonDocument& document, const QString& title_id,
+                       const QString& version, const QString& process) {
+	const auto parse_error = ValidateDocument(document);
+	if (!parse_error.isEmpty()) {
+		return parse_error;
+	}
+	const auto root = document.object();
+	if (!IsSupportedTitleId(title_id) ||
+	    root.value(QStringLiteral("id"))
+	            .toString()
+	            .compare(title_id.trimmed(), Qt::CaseInsensitive) != 0) {
+		return Tr("The cheat file does not match this game's title ID.");
+	}
+	if (version.isEmpty() || root.value(QStringLiteral("version")).toString() != version) {
+		return Tr("The cheat file does not match this game's version (%1).").arg(version);
+	}
+	if (process.isEmpty() ||
+	    root.value(QStringLiteral("process")).toString().compare(process, Qt::CaseInsensitive) !=
+	        0) {
+		return Tr("The cheat file does not match this game's executable (%1).").arg(process);
+	}
+	static const QRegularExpression offset_pattern(
+	    QStringLiteral("\\A(?:0[xX])?[0-9a-fA-F]{1,16}\\z"));
+	static const QRegularExpression bytes_pattern(QStringLiteral("\\A(?:[0-9a-fA-F]{2})+\\z"));
+	const auto                      mods = root.value(QStringLiteral("mods")).toArray();
+	if (mods.isEmpty()) {
+		return Tr("The cheat file contains no cheats.");
+	}
+	if (root.contains(QStringLiteral("master"))) {
+		return Tr("This JSON uses master codes, which are not supported yet.");
+	}
+	for (const auto& value: mods) {
+		const auto mod    = value.toObject();
+		const auto memory = mod.value(QStringLiteral("memory"));
+		if (mod.contains(QStringLiteral("module_name"))) {
+			return Tr("This JSON uses module-specific cheats, which are not supported yet.");
+		}
+		if (!memory.isArray() || memory.toArray().isEmpty()) {
+			return Tr("Unsupported memory entries in cheat \"%1\".")
+			    .arg(mod.value(QStringLiteral("name")).toString());
+		}
+		for (const auto& entry: memory.toArray()) {
+			const auto write    = entry.toObject();
+			const auto offset   = write.value(QStringLiteral("offset")).toString();
+			const auto off      = write.value(QStringLiteral("off")).toString();
+			const auto on       = write.value(QStringLiteral("on")).toString();
+			const auto absolute = write.value(QStringLiteral("absolute"));
+			const auto expected = write.value(QStringLiteral("expected"));
+			if ((!absolute.isUndefined() && (!absolute.isBool() || absolute.toBool())) ||
+			    (!expected.isUndefined() &&
+			     (!expected.isString() ||
+			      expected.toString().compare(off, Qt::CaseInsensitive) != 0)) ||
+			    write.contains(QStringLiteral("section")) ||
+			    write.contains(QStringLiteral("sectionName")) ||
+			    write.contains(QStringLiteral("sectionProtection")) ||
+			    write.contains(QStringLiteral("logicalOffset")) ||
+			    write.contains(QStringLiteral("sectionOffset"))) {
+				return Tr("This JSON uses memory addressing features that are not supported yet.");
+			}
+			if (!entry.isObject() || !offset_pattern.match(offset).hasMatch() ||
+			    !bytes_pattern.match(off).hasMatch() || !bytes_pattern.match(on).hasMatch() ||
+			    off.size() != on.size()) {
+				return Tr("Unsupported memory entries in cheat \"%1\".")
+				    .arg(mod.value(QStringLiteral("name")).toString());
+			}
+		}
+	}
+	return {};
+}
+
+QJsonDocument WithSelection(const QJsonDocument& document, const QList<bool>& enabled) {
+	auto root = document.object();
+	auto mods = root.value(QStringLiteral("mods")).toArray();
+	if (mods.size() != enabled.size()) {
+		return {};
+	}
+	for (qsizetype index = 0; index < mods.size(); ++index) {
+		auto mod = mods[index].toObject();
+		mod.insert(QStringLiteral("enabled"), enabled[index]);
+		mods[index] = mod;
+	}
+	root.insert(QStringLiteral("mods"), mods);
+	return QJsonDocument(root);
+}
+
+QString SaveDocument(const QString& path, const QJsonDocument& document,
+                     const QByteArray& expected_data) {
+	if (path.isEmpty() || !document.isObject()) {
+		return Tr("Invalid cheat file.");
+	}
+	QFile input(path);
+	if (input.exists()) {
+		if (!input.open(QIODevice::ReadOnly)) {
+			return Tr("Could not read cheat file: %1").arg(input.errorString());
+		}
+		const auto current = input.readAll();
+		if (input.error() != QFile::NoError) {
+			return Tr("Could not read cheat file: %1").arg(input.errorString());
+		}
+		if (current != expected_data) {
+			return Tr("Cheat file changed; reload the dialog.");
+		}
+	} else if (!expected_data.isEmpty()) {
+		return Tr("Cheat file changed; reload the dialog.");
+	}
+	input.close();
+	if (!QDir().mkpath(QFileInfo(path).absolutePath())) {
+		return Tr("Could not create the local cheats directory.");
+	}
+	QSaveFile  output(path);
+	const auto data = document.toJson();
+	if (!output.open(QIODevice::WriteOnly) || output.write(data) != data.size() ||
+	    !output.commit()) {
+		return Tr("Could not save cheat file: %1").arg(output.errorString());
+	}
+	return {};
+}
+
+} // namespace Cheats

src/launcher/src/cheatRepository.cpp

@@ -0,0 +1,200 @@
+#include "cheatRepository.h"
+
+#include <QNetworkReply>
+#include <QNetworkRequest>
+#include <QRegularExpression>
+#include <QSet>
+#include <QTimer>
+
+#include <algorithm>
+#include <utility>
+
+namespace Cheats {
+
+namespace {
+
+constexpr qint64 kIndexLimit = 2 * 1024 * 1024;
+constexpr qint64 kFileLimit  = 8 * 1024 * 1024;
+
+const QString kBaseUrl = QStringLiteral(
+    "https://raw.githubusercontent.com/TeeKay87/HEN-Cheats-Collection/master/cheats/");
+
+bool IsFormat(const QString& format) {
+	return format == QStringLiteral("json") || format == QStringLiteral("mc4") ||
+	       format == QStringLiteral("shn");
+}
+
+QRegularExpressionMatch MatchName(const QString& name) {
+	static const QRegularExpression pattern(QStringLiteral(
+	    "\\A([A-Z]{4}[0-9]{5})_([0-9]+(?:\\.[0-9]+)+)_[A-Za-z0-9_.-]+\\.(json|mc4|shn)\\z"));
+	return pattern.match(name);
+}
+
+} // namespace
+
+bool RemoteFile::IsSupported() const {
+	return format == QStringLiteral("json") && !Url().isEmpty();
+}
+
+QUrl RemoteFile::Url() const {
+	const auto match = MatchName(name);
+	if (!IsFormat(format) || !match.hasMatch() || match.captured(3) != format) {
+		return {};
+	}
+	return QUrl(kBaseUrl + format + QLatin1Char('/') + name);
+}
+
+CatalogResult ParseIndex(const QByteArray& data, const QString& format, const QString& title_id) {
+	CatalogResult result;
+	if (!IsFormat(format) || data.isEmpty() || data.size() > kIndexLimit) {
+		result.error = CheatRepository::tr("Invalid cheat index.");
+		return result;
+	}
+
+	const auto    id = title_id.trimmed().toUpper();
+	QSet<QString> seen;
+	int           invalid = 0;
+	auto          text    = QString::fromUtf8(data);
+	if (text.startsWith(QChar(0xfeff))) {
+		text.remove(0, 1);
+	}
+	for (const auto& line: text.split(QLatin1Char('\n'))) {
+		const auto trimmed = line.trimmed();
+		if (trimmed.isEmpty()) {
+			continue;
+		}
+		const auto separator = trimmed.indexOf(QLatin1Char('='));
+		const auto name      = trimmed.left(separator);
+		const auto match     = MatchName(name);
+		if (separator < 0 || !match.hasMatch() || match.captured(3) != format) {
+			invalid++;
+			continue;
+		}
+		if (match.captured(1) != id || seen.contains(name)) {
+			continue;
+		}
+		seen.insert(name);
+		result.files.append(
+		    {name, trimmed.mid(separator + 1).trimmed(), match.captured(2), format});
+	}
+	if (invalid != 0) {
+		result.error = CheatRepository::tr("Skipped %1 invalid cheat index entries.").arg(invalid);
+	}
+	return result;
+}
+
+CheatRepository::CheatRepository(QObject* parent): QObject(parent), m_network(this) {}
+
+QNetworkReply* CheatRepository::Request(const QUrl& url, qint64 limit,
+                                        std::function<void(QByteArray, QString)> complete) {
+	QNetworkRequest request(url);
+	request.setAttribute(QNetworkRequest::RedirectPolicyAttribute,
+	                     QNetworkRequest::SameOriginRedirectPolicy);
+	request.setRawHeader("User-Agent", "Kyty-Launcher");
+	request.setTransferTimeout(15000);
+	auto* reply = m_network.get(request);
+	reply->setReadBufferSize(limit + 1);
+	connect(reply, &QNetworkReply::readyRead, this, [reply, limit]() {
+		if (reply->bytesAvailable() > limit) {
+			reply->setProperty("too_large", true);
+			reply->abort();
+		}
+	});
+	QTimer::singleShot(30000, reply, [reply]() {
+		if (!reply->isFinished()) {
+			reply->setProperty("timed_out", true);
+			reply->abort();
+		}
+	});
+	connect(reply, &QNetworkReply::finished, this,
+	        [this, reply, limit, complete = std::move(complete)]() {
+		        QString error;
+		        if (reply->property("too_large").toBool() || reply->bytesAvailable() > limit) {
+			        error = tr("Cheat download exceeds the size limit.");
+		        } else if (reply->property("timed_out").toBool()) {
+			        error = tr("Cheat download timed out.");
+		        } else if (reply->error() != QNetworkReply::NoError) {
+			        error = reply->errorString();
+		        } else if (reply->attribute(QNetworkRequest::HttpStatusCodeAttribute).toInt() !=
+		                   200) {
+			        error = tr("Unexpected response from the cheat repository.");
+		        }
+		        const auto data = error.isEmpty() ? reply->readAll() : QByteArray();
+		        m_catalog_replies.removeAll(reply);
+		        if (m_file_reply == reply) {
+			        m_file_reply.clear();
+		        }
+		        reply->deleteLater();
+		        complete(data, error);
+	        });
+	return reply;
+}
+
+void CheatRepository::Load(const QString& title_id) {
+	const auto generation = ++m_catalog_generation;
+	for (auto* reply: std::exchange(m_catalog_replies, {})) {
+		reply->abort();
+	}
+	CancelFetch();
+	m_files.clear();
+	m_errors.clear();
+	m_pending = 3;
+	for (const auto& format:
+	     {QStringLiteral("json"), QStringLiteral("mc4"), QStringLiteral("shn")}) {
+		auto* reply =
+		    Request(QUrl(kBaseUrl + format + QStringLiteral(".txt")), kIndexLimit,
+		            [this, generation, format, title_id](const QByteArray& data, QString error) {
+			            if (generation != m_catalog_generation) {
+				            return;
+			            }
+			            if (error.isEmpty()) {
+				            auto result = ParseIndex(data, format, title_id);
+				            m_files.append(result.files);
+				            error = result.error;
+			            }
+			            if (!error.isEmpty()) {
+				            m_errors.append(format.toUpper() + QStringLiteral(": ") + error);
+			            }
+			            if (--m_pending == 0) {
+				            std::sort(m_files.begin(), m_files.end(),
+				                      [](const RemoteFile& left, const RemoteFile& right) {
+					                      if (left.version != right.version) {
+						                      return left.version > right.version;
+					                      }
+					                      if (left.format != right.format) {
+						                      return left.format < right.format;
+					                      }
+					                      return left.name < right.name;
+				                      });
+				            emit CatalogLoaded(m_files, m_errors.join(QLatin1Char('\n')));
+			            }
+		            });
+		m_catalog_replies.append(reply);
+	}
+}
+
+void CheatRepository::Fetch(const RemoteFile& file) {
+	CancelFetch();
+	if (!file.IsSupported()) {
+		emit FileLoaded({}, tr("Only JSON cheat files are supported."));
+		return;
+	}
+	const auto generation = m_fetch_generation;
+	m_file_reply = Request(file.Url(), kFileLimit,
+	                       [this, generation](const QByteArray& data, const QString& error) {
+		                       if (generation == m_fetch_generation) {
+			                       emit FileLoaded(data, error);
+		                       }
+	                       });
+}
+
+void CheatRepository::CancelFetch() {
+	m_fetch_generation++;
+	if (m_file_reply != nullptr) {
+		auto* reply = m_file_reply.data();
+		m_file_reply.clear();
+		reply->abort();
+	}
+}
+
+} // namespace Cheats

src/launcher/src/configuration.cpp

@@ -0,0 +1,149 @@
+#include "configuration.h"
+
+QVariantMap Configuration::GameSettings() const {
+	QVariantMap values;
+#define KYTY_CFG_SET(n) values.insert(QStringLiteral(#n), QVariant::fromValue(n).toString())
+	KYTY_CFG_SET(screen_resolution);
+	KYTY_CFG_SET(user_name);
+	KYTY_CFG_SET(user_id);
+	KYTY_CFG_SET(audio_input_device);
+	KYTY_CFG_SET(present_mode);
+	KYTY_CFG_SET(gpu_index);
+	KYTY_CFG_SET(fullscreen_enabled);
+	KYTY_CFG_SET(hide_cursor_enabled);
+	KYTY_CFG_SET(readback_linear_images);
+	KYTY_CFG_SET(sync_raw_image_buffers);
+	KYTY_CFG_SET(tessellation_enabled);
+	KYTY_CFG_SET(trophy_enabled);
+	KYTY_CFG_SET(skip_notice_screen);
+	KYTY_CFG_SET(vblank_frequency);
+	KYTY_CFG_SET(console_language);
+	KYTY_CFG_SET(vulkan_validation_enabled);
+	KYTY_CFG_SET(shader_validation_enabled);
+	KYTY_CFG_SET(shader_optimization_type);
+	KYTY_CFG_SET(shader_log_direction);
+	KYTY_CFG_SET(shader_log_folder);
+	KYTY_CFG_SET(command_buffer_dump_enabled);
+	KYTY_CFG_SET(command_buffer_dump_folder);
+	KYTY_CFG_SET(printf_direction);
+	KYTY_CFG_SET(printf_output_file);
+	KYTY_CFG_SET(profiler_enabled);
+	KYTY_CFG_SET(renderdoc_enabled);
+	KYTY_CFG_SET(amd_cpu_enabled);
+#if defined(_WIN32)
+	KYTY_CFG_SET(red_zone_protection_enabled);
+#endif
+	KYTY_CFG_SET(elf);
+#undef KYTY_CFG_SET
+	return values;
+}
+
+template <class Settings>
+void Configuration::ReadGameSettingsValues(const Settings& s) {
+#define KYTY_CFG_GET(n) n = s.value(QStringLiteral(#n)).template value<decltype(n)>()
+	KYTY_CFG_GET(screen_resolution);
+	user_name          = s.value("user_name", user_name).toString();
+	bool user_id_ok    = false;
+	auto saved_user_id = s.value("user_id", user_id).toInt(&user_id_ok);
+	user_id            = user_id_ok && Config::IsConfiguredUserIdValid(saved_user_id)
+	                         ? saved_user_id
+	                         : Config::DEFAULT_USER_ID;
+	audio_input_device = s.value("audio_input_device", audio_input_device).toString();
+	KYTY_CFG_GET(present_mode);
+	gpu_index = s.value("gpu_index", -1).toInt();
+	if (EnumToText(present_mode).isEmpty()) {
+		present_mode = PresentMode::Mailbox;
+	}
+	KYTY_CFG_GET(fullscreen_enabled);
+	KYTY_CFG_GET(hide_cursor_enabled);
+	KYTY_CFG_GET(readback_linear_images);
+	KYTY_CFG_GET(sync_raw_image_buffers);
+	KYTY_CFG_GET(tessellation_enabled);
+	trophy_enabled   = s.value("trophy_enabled", trophy_enabled).toBool();
+	KYTY_CFG_GET(skip_notice_screen);
+	vblank_frequency = s.value("vblank_frequency", vblank_frequency).toInt();
+	console_language = s.value("console_language", console_language).toInt();
+	if (console_language < 0 || console_language > MAX_CONSOLE_LANGUAGE) {
+		console_language = DEFAULT_CONSOLE_LANGUAGE;
+	}
+	KYTY_CFG_GET(vulkan_validation_enabled);
+	KYTY_CFG_GET(shader_validation_enabled);
+	KYTY_CFG_GET(shader_optimization_type);
+	KYTY_CFG_GET(shader_log_direction);
+	KYTY_CFG_GET(shader_log_folder);
+	KYTY_CFG_GET(command_buffer_dump_enabled);
+	KYTY_CFG_GET(command_buffer_dump_folder);
+	KYTY_CFG_GET(printf_direction);
+	KYTY_CFG_GET(printf_output_file);
+	KYTY_CFG_GET(profiler_enabled);
+	KYTY_CFG_GET(renderdoc_enabled);
+	amd_cpu_enabled = s.value("amd_cpu_enabled", false).toBool();
+#if defined(_WIN32)
+	red_zone_protection_enabled =
+	    s.value("red_zone_protection_enabled", red_zone_protection_enabled).toBool();
+#endif
+	elf = s.value("elf", elf).toString();
+#undef KYTY_CFG_GET
+}
+
+void Configuration::WriteSettings(QSettings* s) const {
+	const auto values = GameSettings();
+	for (auto it = values.constBegin(); it != values.constEnd(); ++it) {
+		s->setValue(it.key(), it.value());
+	}
+	s->setValue("name", name);
+	s->setValue("basedir", basedir);
+	s->setValue("game_path", game_path);
+	s->setValue("custom_settings", QVariant::fromValue(custom_settings).toString());
+	s->setValue("host_input_mapping", host_input_mapping);
+}
+
+void Configuration::ReadSettings(QSettings* s) {
+	ReadGameSettingsValues(*s);
+	name               = s->value("name").toString();
+	basedir            = s->value("basedir").toString();
+	game_path          = s->value("game_path").toString();
+	custom_settings    = s->value("custom_settings").toBool();
+	host_input_mapping = s->value("host_input_mapping", host_input_mapping).toStringList();
+}
+
+bool Configuration::SetGameSettings(const QJsonObject& settings, QString& error) {
+	const auto fail = [&error](const QString& message) {
+		error = message;
+		return false;
+	};
+	auto values = GameSettings();
+	for (auto it = settings.constBegin(); it != settings.constEnd(); ++it) {
+		if (!values.contains(it.key())) {
+			return fail(tr("Unknown game setting: %1").arg(it.key()));
+		}
+		if (!it.value().isString()) {
+			return fail(tr("Game setting %1 must be a string.").arg(it.key()));
+		}
+		values.insert(it.key(), it.value().toString());
+	}
+
+	ReadGameSettingsValues(values);
+	const auto canonical = GameSettings();
+	for (auto it = settings.constBegin(); it != settings.constEnd(); ++it) {
+		if (canonical.value(it.key()) != values.value(it.key())) {
+			return fail(tr("Invalid value for game setting: %1").arg(it.key()));
+		}
+	}
+	if (user_name.trimmed().isEmpty() || user_name.toUtf8().size() > Config::MAX_USER_NAME_LENGTH) {
+		return fail(tr("User name must contain 1-16 UTF-8 bytes."));
+	}
+	if (gpu_index < -1) {
+		return fail(tr("GPU index must be -1 or greater."));
+	}
+	if (vblank_frequency < 30 || vblank_frequency > 360) {
+		return fail(tr("Vblank frequency must be between 30 and 360."));
+	}
+	if ((shader_log_direction == LogDirection::File && shader_log_folder.isEmpty()) ||
+	    (printf_direction == LogDirection::File && printf_output_file.isEmpty()) ||
+	    (command_buffer_dump_enabled && command_buffer_dump_folder.isEmpty())) {
+		return fail(tr("Enabled logging and command buffer dumps require an output path."));
+	}
+
+	return true;
+}

src/launcher/src/configurationEditDialog.cpp

@@ -19,6 +19,7 @@
 #include <QLineEdit>
 #include <QListView>
 #include <QListWidget>
+#include <QMenu>
 #include <QMessageBox>
 #include <QPushButton>
 #include <QResizeEvent>
@@ -123,8 +124,10 @@ ConfigurationEditDialog::ConfigurationEditDialog(Configuration& info, QWidget* p
     : QDialog(parent, Qt::WindowCloseButtonHint), m_ui(new Ui::ConfigurationEditDialog),
       m_info(info) {
 	m_ui->setupUi(this);
+	setMinimumWidth(width());
 	InitGameDirectories();
 	m_ui->controller_group->setVisible(false);
+	m_ui->settings_button->setVisible(false);
 
 	connect(m_ui->ok_button, &QPushButton::clicked, this, &ConfigurationEditDialog::save);
 	connect(m_ui->cancel_button, &QPushButton::clicked, this, &QDialog::reject);
@@ -164,9 +167,6 @@ ConfigurationEditDialog::ConfigurationEditDialog(Configuration& info, QWidget* p
 		        m_ui->lineEdit_printf_file->setEnabled(log == Configuration::LogDirection::File);
 	        });
 
-	// Keep the controls at a usable minimum while allowing the settings window
-	// and its expanding fields to use any additional space the user gives them.
-	layout()->setSizeConstraint(QLayout::SetMinimumSize);
 	setSizeGripEnabled(true);
 
 	restoreGeometry(g_last_geometry);
@@ -275,7 +275,10 @@ void ConfigurationEditDialog::Init(const Configuration& info) {
 	m_ui->checkBox_fullscreen->setChecked(info.fullscreen_enabled);
 	m_ui->checkBox_hide_cursor->setChecked(info.hide_cursor_enabled);
 	m_ui->checkBox_readback->setChecked(info.readback_linear_images);
+	m_ui->checkBox_sync_raw_image_buffers->setChecked(info.sync_raw_image_buffers);
 	m_ui->checkBox_tessellation->setChecked(info.tessellation_enabled);
+	m_ui->checkBox_skip_notice_screen->setChecked(info.skip_notice_screen);
+	m_ui->checkBox_trophy_notifications->setChecked(info.trophy_enabled);
 	m_ui->spinBox_vblank_frequency->setValue(info.vblank_frequency);
 	m_ui->comboBox_console_language->clear();
 	m_ui->comboBox_console_language->addItems(CONSOLE_LANGUAGE_NAMES);
@@ -355,6 +358,11 @@ void ConfigurationEditDialog::InitGameDirectories() {
 
 void ConfigurationEditDialog::SetGlobalSettings(const QStringList& dirs) {
 	m_global_settings = true;
+	auto* menu        = new QMenu(m_ui->settings_button);
+	menu->addAction(tr("Import..."), this, &ConfigurationEditDialog::ImportGameSettings);
+	menu->addAction(tr("Export..."), this, &ConfigurationEditDialog::ExportGameSettings);
+	m_ui->settings_button->setMenu(menu);
+	m_ui->settings_button->setVisible(true);
 	m_game_dirs_list->clear();
 	m_game_dirs_group->setMinimumWidth(GLOBAL_SETTINGS_GAME_DIRS_MIN_WIDTH);
 
@@ -428,7 +436,10 @@ static void UpdateInfo(Configuration& info, Ui::ConfigurationEditDialog& ui, boo
 	info.fullscreen_enabled        = ui.checkBox_fullscreen->isChecked();
 	info.hide_cursor_enabled       = ui.checkBox_hide_cursor->isChecked();
 	info.readback_linear_images    = ui.checkBox_readback->isChecked();
+	info.sync_raw_image_buffers    = ui.checkBox_sync_raw_image_buffers->isChecked();
 	info.tessellation_enabled      = ui.checkBox_tessellation->isChecked();
+	info.skip_notice_screen        = ui.checkBox_skip_notice_screen->isChecked();
+	info.trophy_enabled            = ui.checkBox_trophy_notifications->isChecked();
 	info.vblank_frequency          = ui.spinBox_vblank_frequency->value();
 	info.console_language          = ui.comboBox_console_language->currentIndex();
 	info.vulkan_validation_enabled = ui.checkBox_vulkan_validation->isChecked();