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

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

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

.github/workflows/build.yml

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

3rdparty/CMakeLists.txt

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

CMakeLists.txt

@@ -265,6 +265,7 @@ set(gpu_tiler_shader_names
 	promote_d16
 	demote_d16
 	color_transform
+	color_clear
 )
 file(GLOB gpu_tiler_shader_includes CONFIGURE_DEPENDS "${gpu_tiler_shader_dir}/gpu_tiler_*.inc")
 foreach(shader_name IN LISTS gpu_tiler_shader_names)
@@ -336,7 +337,7 @@ if (KYTY_CLANG_CL)
 	set_source_files_properties(${kyty_emulator_src} PROPERTIES COMPILE_FLAGS "-Wno-pragma-pack -Wno-deprecated-declarations -D_TIMESPEC_DEFINED")
 endif()
 
-set(kyty_emulator_link_libraries common Vulkan::Headers spirv-tools-opt spirv-tools SDL3::SDL3-static imgui xxhash FFmpeg::ffmpeg fmt::fmt nlohmann_json::nlohmann_json LibAtrac9 kyty_git_version Zydis::Zydis xbyak::xbyak opus OpenXR::openxr_loader)
+set(kyty_emulator_link_libraries common Vulkan::Headers spirv-tools-opt spirv-tools SDL3::SDL3-static imgui xxhash FFmpeg::ffmpeg fmt::fmt nlohmann_json::nlohmann_json LibAtrac9 kyty_git_version Zydis::Zydis xbyak::xbyak opus OpenXR::openxr_loader libzstd_static)
 
 # Linux system libraries.
 if(LINUX)
@@ -406,6 +407,8 @@ add_executable(scalar_provenance_tests EXCLUDE_FROM_ALL
 target_link_libraries(scalar_provenance_tests fmt::fmt xbyak::xbyak)
 target_include_directories(scalar_provenance_tests PRIVATE ${inc_headers})
 
+add_executable(build_flags_tests EXCLUDE_FROM_ALL "${KYTY_TESTS_DIR}/BuildFlagsTests.cpp")
+
 add_executable(page_manager_tests EXCLUDE_FROM_ALL
 	"${KYTY_TESTS_DIR}/PageManagerTests.cpp"
 	"${KYTY_SOURCE_DIR}/graphics/host_gpu/pageManager.cpp"
@@ -450,6 +453,11 @@ add_executable(lru_cache_tests EXCLUDE_FROM_ALL
 )
 target_include_directories(lru_cache_tests PRIVATE ${inc_headers})
 
+add_executable(slot_vector_tests EXCLUDE_FROM_ALL
+	"${KYTY_TESTS_DIR}/SlotVectorTests.cpp"
+)
+target_include_directories(slot_vector_tests PRIVATE ${inc_headers})
+
 add_executable(ime_dialog_tests EXCLUDE_FROM_ALL
 	"${KYTY_TESTS_DIR}/ImeDialogTests.cpp"
 	"${KYTY_SOURCE_DIR}/libs/imeCommon.cpp"
@@ -540,6 +548,19 @@ add_executable(png_enc_tests EXCLUDE_FROM_ALL "${KYTY_TESTS_DIR}/PngEncTests.cpp
 target_link_libraries(png_enc_tests common fmt::fmt)
 target_include_directories(png_enc_tests PRIVATE ${inc_headers})
 
+add_executable(jpeg_dec_tests EXCLUDE_FROM_ALL "${KYTY_TESTS_DIR}/JpegDecTests.cpp")
+target_link_libraries(jpeg_dec_tests common fmt::fmt)
+target_include_directories(jpeg_dec_tests PRIVATE ${inc_headers})
+
+add_executable(jpeg_enc_tests EXCLUDE_FROM_ALL "${KYTY_TESTS_DIR}/JpegEncTests.cpp")
+target_link_libraries(jpeg_enc_tests common fmt::fmt)
+target_include_directories(jpeg_enc_tests PRIVATE ${inc_headers})
+
+add_executable(occlusion_queries_tests EXCLUDE_FROM_ALL
+	"${KYTY_TESTS_DIR}/OcclusionQueriesTests.cpp")
+target_link_libraries(occlusion_queries_tests common fmt::fmt Vulkan::Headers)
+target_include_directories(occlusion_queries_tests PRIVATE ${inc_headers})
+
 add_executable(ces_tests EXCLUDE_FROM_ALL "${KYTY_TESTS_DIR}/CesTests.cpp")
 target_link_libraries(ces_tests common fmt::fmt SDL3::SDL3-static)
 target_include_directories(ces_tests PRIVATE ${inc_headers})
@@ -631,6 +652,7 @@ if(BUILD_TESTING)
 	add_test(NAME image_page_table COMMAND $<TARGET_FILE:image_page_table_tests>)
 	add_test(NAME memory_tracker COMMAND $<TARGET_FILE:memory_tracker_tests>)
 	add_test(NAME page_manager COMMAND $<TARGET_FILE:page_manager_tests>)
+	add_test(NAME build_flags COMMAND $<TARGET_FILE:build_flags_tests>)
 	add_test(NAME archive_file COMMAND $<TARGET_FILE:archive_file_tests>)
 	add_test(NAME trophy_system COMMAND $<TARGET_FILE:trophy_system_tests>)
 	add_test(NAME avplayer_file COMMAND $<TARGET_FILE:avplayer_file_tests>)
@@ -651,6 +673,9 @@ if(BUILD_TESTING)
 	add_test(NAME host_input COMMAND $<TARGET_FILE:host_input_tests>)
 	add_test(NAME save_data_memory COMMAND $<TARGET_FILE:save_data_memory_tests>)
 	add_test(NAME png_enc COMMAND $<TARGET_FILE:png_enc_tests>)
+	add_test(NAME jpeg_dec COMMAND $<TARGET_FILE:jpeg_dec_tests>)
+	add_test(NAME jpeg_enc COMMAND $<TARGET_FILE:jpeg_enc_tests>)
+	add_test(NAME occlusion_queries COMMAND $<TARGET_FILE:occlusion_queries_tests>)
 	add_test(NAME ces COMMAND $<TARGET_FILE:ces_tests>)
 	add_test(NAME save_data_dialog COMMAND $<TARGET_FILE:save_data_dialog_tests>)
 	add_test(NAME http_uri_parse COMMAND $<TARGET_FILE:http_uri_parse_tests>)
@@ -675,6 +700,14 @@ if(BUILD_TESTING)
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --stream-buffer-only)
 	add_test(NAME gpu_command_lane
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --gpu-command-lane-only)
+	add_test(NAME wait_until_safe_for_rendering
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --wait-until-safe-for-rendering-only)
+	add_test(NAME pm4_flip
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --flip-only)
+	add_test(NAME depth_bits
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --depth-bits-only)
+	add_test(NAME msaa_storage
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --msaa-storage-only)
 	add_test(NAME pm4_context_state
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --context-state-only)
 	add_test(NAME gpu_tiler
@@ -713,6 +746,7 @@ if(BUILD_TESTING)
 		image_page_table_tests
 		memory_tracker_tests
 		page_manager_tests
+		build_flags_tests
 		archive_file_tests
 		trophy_system_tests
 		avplayer_file_tests
@@ -735,6 +769,9 @@ if(BUILD_TESTING)
 		host_input_tests
 		save_data_memory_tests
 		png_enc_tests
+		jpeg_dec_tests
+		jpeg_enc_tests
+		occlusion_queries_tests
 		ces_tests
 		save_data_dialog_tests
 		http_uri_parse_tests
@@ -778,6 +815,19 @@ add_custom_target(kyty_touch_cursors
 add_dependencies(kyty_emulator kyty_touch_cursors)
 install(DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/assets/cursors" DESTINATION assets)
 
+set(KYTY_NOTO_SANS_FONTS
+	"${KYTY_NOTO_SANS_DIR}/NotoSans-Regular.ttf"
+	"${KYTY_NOTO_SANS_DIR}/NotoSans-Bold.ttf"
+	"${KYTY_NOTO_SANS_DIR}/NotoSans-Italic.ttf"
+	"${KYTY_NOTO_SANS_DIR}/NotoSans-BoldItalic.ttf")
+add_custom_target(kyty_noto_sans_fonts
+	COMMAND ${CMAKE_COMMAND} -E make_directory "$<TARGET_FILE_DIR:kyty_emulator>/assets/fonts"
+	COMMAND ${CMAKE_COMMAND} -E copy_if_different ${KYTY_NOTO_SANS_FONTS}
+		"$<TARGET_FILE_DIR:kyty_emulator>/assets/fonts"
+	VERBATIM)
+add_dependencies(kyty_emulator kyty_noto_sans_fonts)
+install(FILES ${KYTY_NOTO_SANS_FONTS} DESTINATION assets/fonts)
+
 target_link_libraries(kyty_emulator ${kyty_emulator_link_libraries})
 if (WIN32)
 	target_link_libraries(kyty_emulator opengl32)

README.md

@@ -46,6 +46,16 @@ or graphical glitches, so please include the version you tested when reporting a
 ## Screenshots
 
 <table align="center">
+  <tr>
+    <td align="center">
+      <strong>Cyberpunk 2077</strong><br>
+      <img src="docs/screenshots/ps5-07.png" width="300" alt="Cyberpunk 2077 running in KytyPS5">
+    </td>
+    <td align="center">
+      <strong>MLB The Show 23</strong><br>
+      <img src="docs/screenshots/ps5-08.png" width="300" alt="MLB The Show 23 running in KytyPS5">
+    </td>
+  </tr>
   <tr>
     <td align="center">
       <strong>Astro Bot</strong><br>

src/common/additionalContent.cpp

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

src/common/additionalContent.h

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

src/common/archive.h

@@ -5,6 +5,7 @@
 
 #include <cstdint>
 #include <filesystem>
+#include <functional>
 #include <memory>
 #include <vector>
 
@@ -45,6 +46,12 @@ private:
 // Accepts either a host archive filename or a virtual path inside the archive.
 [[nodiscard]] std::shared_ptr<ArchiveReader> OpenArchive(const std::filesystem::path& path);
 
+// Packs a host directory into a new ZArchive outside it. progress gets 0 < done <= total and
+// returns false to cancel. A failed or cancelled pack removes the partial archive.
+[[nodiscard]] bool PackArchive(const std::filesystem::path& directory,
+                               const std::filesystem::path& archive,
+                               const std::function<bool(uint64_t done, uint64_t total)>& progress);
+
 } // namespace Common
 
 #endif /* KYTY_COMMON_ARCHIVE_H_ */

src/common/archivePack.cpp

@@ -0,0 +1,84 @@
+// Kept out of archive.cpp so the statically linked emulator does not pull in the writer and the
+// zstd compressor.
+#include "common/archive.h"
+
+#include "common/stringUtils.h"
+
+#include <algorithm>
+#include <zarchive/zarchivewriter.h>
+
+namespace Common {
+
+bool PackArchive(const std::filesystem::path& directory, const std::filesystem::path& archive,
+                 const std::function<bool(uint64_t done, uint64_t total)>& progress) {
+	namespace fs = std::filesystem;
+	struct Output {
+		File file;
+		bool ok = true;
+	} output;
+	// Create the archive before walking, so a re-run started while a cancelled pack is still
+	// walking sees it exists instead of losing its archive to that pack's cleanup.
+	if (!output.file.Create(archive)) {
+		return false;
+	}
+	struct Entry {
+		fs::path path;
+		uint64_t size;
+		bool     is_file;
+	};
+	std::vector<Entry> entries;
+	uint64_t           total = 0;
+	std::error_code    error;
+	// Follow directory links, as the emulator does when it runs the folder.
+	fs::recursive_directory_iterator it(directory, fs::directory_options::follow_directory_symlink,
+	                                    error);
+	for (; !error && it != fs::end(it); it.increment(error)) {
+		const bool is_file = !it->is_directory(error);
+		const auto size    = is_file && !error ? it->file_size(error) : 0;
+		if (error) {
+			break;
+		}
+		entries.push_back({it->path(), size, is_file});
+		total += size;
+	}
+	output.ok = !error;
+	ZArchiveWriter writer(
+	    [](int32_t, void*) {},
+	    [](const void* data, size_t size, void* opaque) {
+		    auto&    out     = *static_cast<Output*>(opaque);
+		    uint32_t written = 0;
+		    out.file.Write(data, static_cast<uint32_t>(size), &written);
+		    out.ok = out.ok && written == size;
+	    },
+	    &output);
+	std::vector<uint8_t> buffer(1u << 20u);
+	uint64_t             done = 0;
+	for (size_t index = 0; output.ok && index < entries.size(); index++) {
+		const auto& entry = entries[index];
+		const auto  name  = PathToGenericString(entry.path.lexically_relative(directory));
+		File        input;
+		output.ok = entry.is_file ? writer.StartNewFile(name.c_str()) &&
+		                                input.Open(entry.path, File::Mode::Read)
+		                          : writer.MakeDir(name.c_str());
+		for (uint64_t left = entry.size; output.ok && left != 0;) {
+			uint32_t read = 0;
+			input.Read(buffer.data(), static_cast<uint32_t>(std::min<uint64_t>(left, buffer.size())),
+			           &read);
+			writer.AppendData(buffer.data(), read);
+			done += read;
+			left -= read;
+			output.ok = output.ok && read != 0 && progress(done, total);
+		}
+	}
+	if (output.ok) {
+		writer.Finalize();
+	}
+	output.ok = output.ok && output.file.Flush();
+	output.file.Close();
+	if (!output.ok) {
+		File::DeleteFile(archive);
+	}
+	return output.ok;
+}
+
+} // namespace Common

src/common/debugCounters.h

@@ -29,6 +29,9 @@ enum class Counter {
 	TextureUploadBytes,
 	TextureDownloadBytes,
 	TextureEvictions,
+	ColorMetadataReadbacks, // CPU blocked reading back GPU-written DCC/CMASK metadata.
+	ColorMetadataReadbackNs,
+	ColorMetadataChecks, // GPU clear-key checks of DCC/CMASK metadata.
 	BufferUploadBytes,
 	BufferDownloadBytes,
 	BufferEvictions,

src/common/emulatorConfig.cpp

@@ -133,6 +133,10 @@ bool ShaderValidationEnabled() {
 	return g_config->shader_validation_enabled;
 }
 
+bool ShaderCacheEnabled() {
+	return g_config->shader_cache_enabled;
+}
+
 ShaderOptimizationType GetShaderOptimizationType() {
 	return g_config->shader_optimization_type;
 }
@@ -197,6 +201,10 @@ bool TessellationEnabled() {
 	return g_config->tessellation_enabled;
 }
 
+bool OcclusionQueriesEnabled() {
+	return g_config->occlusion_queries_enabled;
+}
+
 bool TrophyEnabled() {
 	return g_config->trophy_enabled;
 }

src/common/emulatorConfig.h

@@ -72,6 +72,7 @@ struct ConfigOptions {
 	uint32_t               console_language            = DEFAULT_CONSOLE_LANGUAGE;
 	bool                   vulkan_validation_enabled   = false;
 	bool                   shader_validation_enabled   = false;
+	bool                   shader_cache_enabled        = false;
 	ShaderOptimizationType shader_optimization_type    = ShaderOptimizationType::None;
 	LogDirection           shader_log_direction        = LogDirection::Silent;
 	std::filesystem::path  shader_log_folder           = "_Shaders";
@@ -88,6 +89,7 @@ struct ConfigOptions {
 	bool                   readback_tiled_images       = false;
 	bool                   sync_raw_image_buffers      = false;
 	bool                   tessellation_enabled        = false;
+	bool                   occlusion_queries_enabled   = false;
 	bool                   trophy_enabled              = true;
 	uint32_t               trophy_duration_seconds     = 6;
 	bool                   perf_overlay_enabled            = false;
@@ -131,6 +133,7 @@ uint32_t GetConsoleLanguage();
 bool     VulkanValidationEnabled();
 
 bool                   ShaderValidationEnabled();
+bool                   ShaderCacheEnabled();
 ShaderOptimizationType GetShaderOptimizationType();
 LogDirection           GetShaderLogDirection();
 std::filesystem::path  GetShaderLogFolder();
@@ -154,6 +157,7 @@ bool ReadbackLinearImagesEnabled();
 bool ReadbackTiledImagesEnabled();
 bool SyncRawImageBuffersEnabled();
 bool TessellationEnabled();
+bool OcclusionQueriesEnabled();
 bool TrophyEnabled();
 uint32_t GetTrophyDurationSeconds();
 bool PlayGoHackEnabled();

src/common/slotVector.h

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

src/emulator.cpp

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

src/emulator.h

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

src/graphics/guest_gpu/command_processor/commandProcessor.h

@@ -27,6 +27,7 @@ struct IndirectDrawRegisters {
 	uint32_t vertex_offset;
 	uint32_t instance_offset;
 	uint32_t index_offset;
+	uint32_t draw_index;
 };
 
 class Pm4Execution {
@@ -50,13 +51,6 @@ private:
 
 class CommandProcessor {
 public:
-	struct FlipInfo {
-		int     handle    = 0;
-		int     index     = 0;
-		int     flip_mode = 0;
-		int64_t flip_arg  = 0;
-	};
-
 	CommandProcessor(RenderContext& renderer, int interrupt_event_id)
 	    : m_renderer(renderer), m_interrupt_event_id(interrupt_event_id) {}
 	~CommandProcessor() = default;
@@ -90,43 +84,21 @@ public:
 	void DrawIndirectMulti(uint32_t data_offset, uint32_t max_count_or_count,
 	                       const volatile uint32_t* count_addr, uint32_t stride_in_bytes,
 	                       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,
-	                        uint32_t interrupt_selector, uint32_t interrupt_context_id = 0);
-	void WriteAtEndOfPipe64(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, uint64_t value,
-	                        uint32_t interrupt_selector, uint32_t interrupt_context_id = 0);
-	void Flip();
-	void Flip(void* dst_gpu_addr, uint32_t value);
-	void FlipWithInterrupt(uint32_t eop_event_type, uint32_t cache_action, void* dst_gpu_addr,
-	                       uint32_t value);
+	void ReleaseMem(uint32_t data_sel, uint32_t interrupt_selector, void* address, uint64_t data,
+	                uint32_t interrupt_context_id);
+	void Flip(int handle, int index, int flip_mode, int64_t flip_arg);
 	void PrepareCpuFlip(uint64_t request_id);
-	void SynchronizeGpu();
-	void TriggerEopEventAtEndOfPipe(uint32_t interrupt_context_id);
 	void DispatchDirect(uint32_t thread_group_x, uint32_t thread_group_y, uint32_t thread_group_z,
 	                    uint32_t mode);
 	void DispatchIndirect(uint64_t args_addr, uint32_t mode);
 	void WaitFlipDone(uint32_t video_out_handle, uint32_t display_buffer_index);
 	void TriggerEvent(uint32_t event_type, uint32_t event_index, uint64_t event_address = 0);
 
-	void SetUserDataMarker(HW::UserSgprType type) { m_user_data_marker = type; }
-	[[nodiscard]] HW::UserSgprType GetUserDataMarker() const { return m_user_data_marker; }
-
-	void ResetDeCe();
-	void SetCeComplete(bool complete) { m_ce_complete = complete; }
-	void WaitCe();
-	void WaitDeDiff(uint32_t diff);
 	void WaitForRewind(bool valid);
-	void IncrementDe();
-	void IncrementCe();
-
-	void WriteConstRam(uint32_t offset, const uint32_t* src, uint32_t dw_num);
-	void DumpConstRam(uint32_t* dst, uint32_t offset, uint32_t dw_num);
 
 	template <typename T>
 	void WaitRegMem(uint32_t func, const T* addr, T ref, T mask, uint32_t poll, uint32_t wait_op);
+	void MemSemaphore(uint64_t* counter, bool signal, bool write);
 	void WriteData(uint32_t* dst, const uint32_t* src, uint32_t dw_num, uint32_t write_control);
 	void WriteReferenceClock(uint64_t dst_address, uint32_t num_bytes);
 	void DmaData(uint8_t engine, uint8_t dst_sel, uint8_t dst_cache_policy,
@@ -140,45 +112,34 @@ public:
 	Pm4ProcessResult Process(Pm4Execution& execution, std::span<const uint32_t> commands);
 	void             ProcessIndirectBuffer(std::span<const uint32_t> commands, bool chain);
 
-	void SetFlip(const FlipInfo& flip) { m_flip = flip; }
-
 	[[nodiscard]] uint64_t GetSubmitId() const { return m_submit_id; }
 	void                   SetSubmitId(uint64_t submit_id) { m_submit_id = submit_id; }
 	[[nodiscard]] bool     IsAsyncComputeQueue() const { return m_interrupt_event_id >= 0x20; }
 
 private:
-	template <typename T>
-	void WriteAtEndOfPipe(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, T value, uint32_t interrupt_selector,
-	                      uint32_t interrupt_context_id);
 	void ProcessPm4(Pm4Execution& execution);
 	void SuspendPm4();
+	uint32_t            NumInstances();
+	const void*         IndexAddress(uint32_t first_index) const;
 	CommandScheduler&   GetScheduler() const { return m_renderer.GetCommandScheduler(); }
 	CommandBuffer&      CurrentBuffer() { return GetScheduler().Current(); }
 
-	RenderContext&   m_renderer;
-	HW::Context      m_ctx;
-	HW::Context      m_saved_ctx;
-	bool             m_context_state_pushed = false;
-	HW::UserConfig   m_ucfg;
-	HW::Shader       m_sh_ctx;
-	HW::UserSgprType m_user_data_marker                 = HW::UserSgprType::Unknown;
-	uint32_t         m_index_type_and_size              = 0;
-	uint32_t         m_index_buffer_size                = 0;
-	uint64_t         m_index_base_addr                  = 0;
-	uint64_t         m_draw_indirect_args_base_addr     = 0;
-	uint64_t         m_dispatch_indirect_args_base_addr = 0;
-	// Persistent draw state: indirect draws update it for subsequent draws.
-	uint32_t m_num_instances = 1;
+	RenderContext& m_renderer;
+	HW::Context    m_ctx;
+	HW::Context    m_saved_ctx;
+	bool           m_context_state_pushed = false;
+	HW::UserConfig m_ucfg;
+	HW::Shader     m_sh_ctx;
+	uint32_t       m_index_type_and_size              = 0;
+	uint32_t       m_index_buffer_size                = 0;
+	uint64_t       m_index_base_addr                  = 0;
+	uint64_t       m_draw_indirect_args_base_addr     = 0;
+	uint64_t       m_dispatch_indirect_args_base_addr = 0;
+	// Persistent draw state. Indirect draws leave the address of their last instance count,
+	// read only when a later draw inherits it.
+	uint32_t m_num_instances      = 1;
+	uint64_t m_num_instances_addr = 0;
 
-	uint32_t m_de_count    = 0;
-	uint32_t m_ce_count    = 0;
-	bool     m_ce_complete = false;
-
-	uint32_t m_const_ram[0x3000] = {0};
-
-	FlipInfo  m_flip;
 	const int m_interrupt_event_id;
 	uint64_t  m_submit_id                   = 0;
 	uint64_t  m_synthetic_occlusion_counter = 0;

src/graphics/guest_gpu/command_processor/pm4Dispatch.cpp

@@ -199,10 +199,10 @@ constexpr auto MakeOpcodeDispatchTable() {
 	g_cp_op_func[Pm4::IT_INDIRECT_BUFFER]           = CpOpIndirectBuffer;
 	g_cp_op_func[Pm4::IT_COPY_DATA]                 = CpOpCopyData;
 	g_cp_op_func[Pm4::IT_EVENT_WRITE]               = CpOpEventWrite;
-	g_cp_op_func[Pm4::IT_EVENT_WRITE_EOP]           = CpOpEventWriteEop;
-	g_cp_op_func[Pm4::IT_EVENT_WRITE_EOS]           = CpOpEventWriteEos;
 	g_cp_op_func[Pm4::IT_DMA_DATA]                  = CpOpDmaData;
 	g_cp_op_func[Pm4::IT_ACQUIRE_MEM]               = CpOpAcquireMem;
+	g_cp_op_func[Pm4::IT_RELEASE_MEM]               = CpOpReleaseMem;
+	g_cp_op_func[Pm4::IT_MEM_SEMAPHORE]             = CpOpMemSemaphore;
 	g_cp_op_func[Pm4::IT_REWIND]                    = CpOpRewind;
 	g_cp_op_func[Pm4::IT_SET_CONTEXT_REG]           = CpOpSetContextReg;
 	g_cp_op_func[Pm4::IT_SET_SH_REG]                = CpOpSetShaderReg;
@@ -211,12 +211,6 @@ constexpr auto MakeOpcodeDispatchTable() {
 	g_cp_op_func[Pm4::IT_SET_CONTEXT_REG_INDIRECT]  = CpOpIndirectCxRegs;
 	g_cp_op_func[Pm4::IT_SET_SH_REG_INDIRECT]       = CpOpIndirectShRegs;
 	g_cp_op_func[Pm4::IT_SET_UCONFIG_REG_INDIRECT]  = CpOpIndirectUcRegs;
-	g_cp_op_func[Pm4::IT_WRITE_CONST_RAM]           = CpOpWriteConstRam;
-	g_cp_op_func[Pm4::IT_DUMP_CONST_RAM]            = CpOpDumpConstRam;
-	g_cp_op_func[Pm4::IT_INCREMENT_CE_COUNTER]      = CpOpIncrementCeCounter;
-	g_cp_op_func[Pm4::IT_INCREMENT_DE_COUNTER]      = CpOpIncrementDeCounter;
-	g_cp_op_func[Pm4::IT_WAIT_ON_CE_COUNTER]        = CpOpWaitOnCeCounter;
-	g_cp_op_func[Pm4::IT_WAIT_ON_DE_COUNTER_DIFF]   = CpOpWaitOnDeCounterDiff;
 	g_cp_op_func[Pm4::IT_GET_LOD_STATS]             = CpOpGetLodStats;
 	g_cp_op_func[Pm4::IT_WAIT_REG_MEM_64]           = CpOpWaitRegMem64;
 
@@ -231,7 +225,6 @@ constexpr auto MakeCustomOpcodeDispatchTable() {
 	g_cp_op_custom_func[Pm4::R_PUSH_MARKER]    = CpOpPushMarker;
 	g_cp_op_custom_func[Pm4::R_POP_MARKER]     = CpOpPopMarker;
 	g_cp_op_custom_func[Pm4::R_FLIP]           = CpOpFlip;
-	g_cp_op_custom_func[Pm4::R_RELEASE_MEM]    = CpOpReleaseMem;
 	g_cp_op_custom_func[Pm4::R_CONTEXT_STATE]  = CpOpContextState;
 
 	return g_cp_op_custom_func;

src/graphics/guest_gpu/command_processor/pm4Dispatch.h

@@ -133,22 +133,16 @@ uint32_t CpOpWriteData(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, u
 uint32_t CpOpIndirectBuffer(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpCopyData(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpEventWrite(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
-uint32_t CpOpEventWriteEop(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
-uint32_t CpOpEventWriteEos(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpDmaData(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpAcquireMem(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
+uint32_t CpOpReleaseMem(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
+uint32_t CpOpMemSemaphore(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpSetContextReg(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpSetShaderReg(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpSetUconfigReg(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpIndirectCxRegs(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpIndirectShRegs(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpIndirectUcRegs(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
-uint32_t CpOpWriteConstRam(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
-uint32_t CpOpDumpConstRam(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
-uint32_t CpOpIncrementCeCounter(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
-uint32_t CpOpIncrementDeCounter(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
-uint32_t CpOpWaitOnCeCounter(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
-uint32_t CpOpWaitOnDeCounterDiff(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpGetLodStats(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpDispatchReset(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpWaitRegMem32(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
@@ -157,7 +151,6 @@ uint32_t CpOpPushMarker(CommandProcessor&, uint32_t, const uint32_t*, uint32_t,
 uint32_t CpOpPopMarker(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpWaitRegMem64(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpFlip(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
-uint32_t CpOpReleaseMem(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 
 } // namespace Libs::Graphics
 

src/graphics/guest_gpu/command_processor/pm4Handlers.cpp

@@ -54,34 +54,12 @@ namespace Libs::Graphics {
 
 namespace {
 
-constexpr uint32_t GcrGl2MetadataInvalidate = 1u << 1u;
-constexpr uint32_t GcrGl0VectorInvalidate   = 1u << 2u;
-constexpr uint32_t GcrGl1Invalidate         = 1u << 3u;
-constexpr uint32_t GcrGl2Unshared           = 1u << 4u;
-constexpr uint32_t GcrGl2Invalidate         = 1u << 8u;
-constexpr uint32_t GcrGl2Writeback          = 1u << 9u;
-constexpr uint32_t GcrOrder012              = 1u << 10u;
-constexpr uint32_t GcrOrder210              = 2u << 10u;
-constexpr uint32_t GcrKnownMask             = GcrGl2MetadataInvalidate | GcrGl0VectorInvalidate |
-                                              GcrGl1Invalidate | GcrGl2Unshared | GcrGl2Invalidate |
-                                              GcrGl2Writeback | GcrOrder012 | GcrOrder210;
-constexpr uint32_t RegisterSelectorMask     = 0x70000000u;
+constexpr uint32_t RegisterSelectorMask = 0x70000000u;
 
 constexpr uint32_t NormalizeRegisterOffset(uint32_t raw_offset) {
 	return raw_offset & ~RegisterSelectorMask;
 }
 
-uint32_t ReleaseMemCacheActionFromGcr(uint32_t gcr_cntl) {
-	return ((gcr_cntl & GcrGl2Writeback) != 0 ? 0x38u : 0x00u);
-}
-
-void LogUnknownReleaseMemGcr(uint32_t gcr_cntl) {
-	const uint32_t unknown = (gcr_cntl & ~GcrKnownMask);
-	if (unknown != 0) {
-		LOGF("\t warning: release_mem uses unknown GCR bits: 0x%04" PRIx32 "\n", unknown);
-	}
-}
-
 } // namespace
 
 static HW::BlendControl DecodeBlendControl(uint32_t value) {
@@ -1068,9 +1046,8 @@ KYTY_HW_SH_PARSER(HwShSetCsUserSgpr) {
 	auto write_num = UserSgprWriteNum(slot, reg_num, 16u, "cs");
 
 	for (uint32_t i = 0; i < write_num; i++) {
-		cp.GetShCtx().SetCsUserSgpr(slot + i, buffer[i], cp.GetUserDataMarker());
+		cp.GetShCtx().SetCsUserSgpr(slot + i, buffer[i]);
 	}
-	cp.SetUserDataMarker(HW::UserSgprType::Unknown);
 
 	return reg_num;
 }
@@ -1085,9 +1062,8 @@ KYTY_HW_SH_PARSER(HwShSetPsUserSgpr) {
 	auto write_num = UserSgprWriteNum(slot, reg_num, 32u, "ps");
 
 	for (uint32_t i = 0; i < write_num; i++) {
-		cp.GetShCtx().SetPsUserSgpr(slot + i, buffer[i], cp.GetUserDataMarker());
+		cp.GetShCtx().SetPsUserSgpr(slot + i, buffer[i]);
 	}
-	cp.SetUserDataMarker(HW::UserSgprType::Unknown);
 
 	return reg_num;
 }
@@ -1102,9 +1078,8 @@ KYTY_HW_SH_PARSER(HwShSetGsUserSgpr) {
 	auto write_num = UserSgprWriteNum(slot, reg_num, 32u, "gs");
 
 	for (uint32_t i = 0; i < write_num; i++) {
-		cp.GetShCtx().SetGsUserSgpr(slot + i, buffer[i], cp.GetUserDataMarker());
+		cp.GetShCtx().SetGsUserSgpr(slot + i, buffer[i]);
 	}
-	cp.SetUserDataMarker(HW::UserSgprType::Unknown);
 
 	return reg_num;
 }
@@ -1119,9 +1094,8 @@ KYTY_HW_SH_PARSER(HwShSetHsUserSgpr) {
 	auto write_num = UserSgprWriteNum(slot, reg_num, 32u, "hs");
 
 	for (uint32_t i = 0; i < write_num; i++) {
-		cp.GetShCtx().SetHsUserSgpr(slot + i, buffer[i], cp.GetUserDataMarker());
+		cp.GetShCtx().SetHsUserSgpr(slot + i, buffer[i]);
 	}
-	cp.SetUserDataMarker(HW::UserSgprType::Unknown);
 
 	return reg_num;
 }
@@ -1608,7 +1582,7 @@ KYTY_CP_OP_PARSER(CpOpDrawIndirect) {
 	// The native indexed packet enables the high-half first-index destination in word 3.
 	const IndirectDrawRegisters registers {
 	    buffer[1] & 0xffffu, buffer[2] & 0xffffu,
-	    indexed && (buffer[2] & (1u << 28u)) != 0 ? buffer[1] >> 16u : Pm4::SH_NOP};
+	    indexed && (buffer[2] & (1u << 28u)) != 0 ? buffer[1] >> 16u : Pm4::SH_NOP, Pm4::SH_NOP};
 
 	cp.DrawIndirect(data_offset, registers, draw_initiator, indexed);
 
@@ -1628,10 +1602,12 @@ KYTY_CP_OP_PARSER(CpOpDrawIndirectMulti) {
 	const auto stride_in_bytes = buffer[7];
 	const auto draw_initiator  = buffer[8];
 	const bool indexed         = (cmd_id == 0xc0083800);
-	// Multi-draw moves the first-index enable to word 4; the destinations stay in 2/3.
+	// Multi-draw moves the first-index enable to word 4; the destinations stay in 2/3. Word 4
+	// bit 31 enables the subdraw index, written to the register in its low half.
 	const IndirectDrawRegisters registers {
 	    buffer[1] & 0xffffu, buffer[2] & 0xffffu,
-	    indexed && (buffer[3] & (1u << 28u)) != 0 ? buffer[1] >> 16u : Pm4::SH_NOP};
+	    indexed && (buffer[3] & (1u << 28u)) != 0 ? buffer[1] >> 16u : Pm4::SH_NOP,
+	    (buffer[3] >> 31u) != 0 ? buffer[3] & 0xffffu : Pm4::SH_NOP};
 
 	if (count_indirect == 0) {
 		count_addr = nullptr;
@@ -1700,24 +1676,6 @@ KYTY_CP_OP_PARSER(CpOpContextState) {
 	return packet_size_dw - 1u;
 }
 
-KYTY_CP_OP_PARSER(CpOpDumpConstRam) {
-	KYTY_PROFILER_FUNCTION();
-
-	EXIT_NOT_IMPLEMENTED(cmd_id != 0xC0038300);
-
-	auto  offset = buffer[0];
-	auto  dw_num = buffer[1];
-	auto* dst = reinterpret_cast<uint32_t*>(buffer[2] | (static_cast<uint64_t>(buffer[3]) << 32u));
-
-	EXIT_NOT_IMPLEMENTED(dw_num >= 0x3000);
-	EXIT_NOT_IMPLEMENTED(offset > 0xbffc);
-	EXIT_NOT_IMPLEMENTED((offset & 0x3u) != 0);
-
-	cp.DumpConstRam(dst, offset, dw_num);
-
-	return 4;
-}
-
 KYTY_CP_OP_PARSER(CpOpEventWrite) {
 	KYTY_PROFILER_FUNCTION();
 
@@ -1738,90 +1696,16 @@ KYTY_CP_OP_PARSER(CpOpEventWrite) {
 	return packet_size_dw - 1u;
 }
 
-KYTY_CP_OP_PARSER(CpOpEventWriteEop) {
-	KYTY_PROFILER_FUNCTION();
-
-	EXIT_NOT_IMPLEMENTED(cmd_id != 0xC0044700);
-
-	uint32_t cache_policy       = (buffer[0] >> 25u) & 0x3u;
-	uint32_t event_write_dest   = ((buffer[0] >> 23u) & 0x10u) | ((buffer[2] >> 16u) & 0x01u);
-	uint32_t eop_event_type     = (buffer[0]) & 0x3fu;
-	uint32_t cache_action       = (buffer[0] >> 12u) & 0x3fu;
-	uint32_t event_index        = (buffer[0] >> 8u) & 0x7u;
-	uint32_t event_write_source = ((buffer[2] >> 29u) & 0x7u);
-	uint32_t interrupt_selector = (buffer[2] >> 24u) & 0x7u;
-	auto*    dst_gpu_addr =
-	    reinterpret_cast<void*>(buffer[1] | (static_cast<uint64_t>(buffer[2] & 0xffffu) << 32u));
-	uint64_t value = (buffer[3] | (static_cast<uint64_t>(buffer[4]) << 32u));
-
-	cp.WriteAtEndOfPipe64(cache_policy, event_write_dest, eop_event_type, cache_action, event_index,
-	                      event_write_source, dst_gpu_addr, value, interrupt_selector);
-
-	return 5;
-}
-
-KYTY_CP_OP_PARSER(CpOpEventWriteEos) {
-	KYTY_PROFILER_FUNCTION();
-
-	EXIT_NOT_IMPLEMENTED(cmd_id != 0xC0034802);
-
-	uint32_t cache_policy       = (buffer[0] >> 25u) & 0x3u;
-	uint32_t event_write_dest   = 0;
-	uint32_t eop_event_type     = (buffer[0]) & 0x3fu;
-	uint32_t cache_action       = (buffer[0] >> 12u) & 0x3fu;
-	uint32_t event_index        = (buffer[0] >> 8u) & 0x7u;
-	uint32_t event_write_source = ((buffer[2] >> 29u) & 0x7u);
-	uint32_t interrupt_selector = (buffer[2] >> 24u) & 0x7u;
-
-	auto* dst_gpu_addr =
-	    reinterpret_cast<void*>(buffer[1] | (static_cast<uint64_t>(buffer[2] & 0xffffu) << 32u));
-	uint32_t value = buffer[3];
-
-	cp.WriteAtEndOfPipe32(cache_policy, event_write_dest, eop_event_type, cache_action, event_index,
-	                      event_write_source, dst_gpu_addr, value, interrupt_selector);
-
-	return 4;
-}
-
 KYTY_CP_OP_PARSER(CpOpFlip) {
 	KYTY_PROFILER_FUNCTION();
 
 	EXIT_NOT_IMPLEMENTED(cmd_id != 0xc004105c);
 
-	CommandProcessor::FlipInfo f;
-
-	f.handle    = static_cast<int>(buffer[0]);
-	f.index     = static_cast<int>(buffer[1]);
-	f.flip_mode = static_cast<int>(buffer[2]);
-	f.flip_arg  = static_cast<int64_t>(buffer[3] | (static_cast<uint64_t>(buffer[4]) << 32u));
-
-	cp.SetFlip(f);
-	cp.Flip();
+	cp.Flip(static_cast<int>(buffer[0]), static_cast<int>(buffer[1]), static_cast<int>(buffer[2]),
+	        static_cast<int64_t>(buffer[3] | (static_cast<uint64_t>(buffer[4]) << 32u)));
 	return 5;
 }
 
-KYTY_CP_OP_PARSER(CpOpIncrementCeCounter) {
-	KYTY_PROFILER_FUNCTION();
-
-	EXIT_NOT_IMPLEMENTED(cmd_id != 0xC0008400);
-	EXIT_NOT_IMPLEMENTED(buffer[0] != 1);
-
-	cp.IncrementCe();
-
-	return 1;
-}
-
-KYTY_CP_OP_PARSER(CpOpIncrementDeCounter) {
-	KYTY_PROFILER_FUNCTION();
-
-	EXIT_NOT_IMPLEMENTED(cmd_id != 0xC0008500);
-	EXIT_NOT_IMPLEMENTED(buffer[0] != 0);
-
-	cp.IncrementDe();
-
-	return 1;
-}
-
 KYTY_CP_OP_PARSER(CpOpIndexType) {
 	KYTY_PROFILER_FUNCTION();
 
@@ -2103,50 +1987,12 @@ KYTY_CP_OP_PARSER(CpOpIndirectUcRegs) {
 	return KYTY_PM4_LEN(cmd_id) - 1u;
 }
 
-KYTY_CP_OP_PARSER(CpOpMarker) {
-	KYTY_PROFILER_FUNCTION();
-
-	uint32_t len_dw = ((cmd_id >> 16u) & 0x3fffu);
-
-	switch (buffer[0]) {
-		case 0x68750004: cp.SetUserDataMarker(HW::UserSgprType::Vsharp); break;
-		case 0x6875000d: cp.SetUserDataMarker(HW::UserSgprType::Region); break;
-		case 0x68750777: {
-			cp.Flip();
-			break;
-		}
-		case 0x68750778: {
-			auto* addr =
-			    reinterpret_cast<void*>(buffer[1] | (static_cast<uint64_t>(buffer[2]) << 32u));
-			uint32_t value = buffer[3];
-			cp.Flip(addr, value);
-			break;
-		}
-		case 0x68750781: {
-			auto* addr =
-			    reinterpret_cast<void*>(buffer[1] | (static_cast<uint64_t>(buffer[2]) << 32u));
-			uint32_t value          = buffer[3];
-			uint32_t eop_event_type = buffer[4];
-			uint32_t cache_action   = buffer[5];
-			cp.FlipWithInterrupt(eop_event_type, cache_action, addr, value);
-			break;
-		}
-		// NOP payloads are caller-owned data; only complete known marker tags execute.
-		default: break;
-	}
-
-	return len_dw + 1;
-}
-
 KYTY_CP_OP_PARSER(CpOpNop) {
 	KYTY_PROFILER_FUNCTION();
 
 	auto r = KYTY_PM4_R(cmd_id);
 
 	if (r == Pm4::R_ZERO) {
-		if ((buffer[0] & 0xffff0000u) == 0x68750000u) {
-			return CpOpMarker(cp, cmd_id, buffer, dw, num_dw);
-		}
 		return KYTY_PM4_LEN(cmd_id) - 1;
 	}
 
@@ -2202,112 +2048,33 @@ KYTY_CP_OP_PARSER(CpOpPushMarker) {
 KYTY_CP_OP_PARSER(CpOpReleaseMem) {
 	KYTY_PROFILER_FUNCTION();
 
-	EXIT_NOT_IMPLEMENTED(cmd_id != 0xc0061060);
+	EXIT_NOT_IMPLEMENTED(cmd_id != 0xc0064900);
 
-	uint32_t cache_policy       = (buffer[0] >> 25u) & 0x3u;
-	uint32_t eop_event_type     = buffer[0] & 0x3fu;
-	uint32_t event_index        = (buffer[0] >> 8u) & 0x7u;
-	uint32_t gcr_cntl           = (buffer[0] >> 12u) & 0xfffu;
-	uint32_t release_dst        = (buffer[1] >> 16u) & 0x3u;
-	uint32_t event_write_dest   = 0;
-	uint32_t data_sel           = (buffer[1] >> 29u) & 0x7u;
-	uint32_t interrupt_selector = (buffer[1] >> 24u) & 0x7u;
-	auto*    dst_gpu_addr =
-	    reinterpret_cast<void*>(buffer[2] | (static_cast<uint64_t>(buffer[3]) << 32u));
-	uint64_t value                = buffer[4] | (static_cast<uint64_t>(buffer[5]) << 32u);
-	uint32_t interrupt_context_id = buffer[6] & 0x07ffffffu;
-
-	constexpr uint32_t ReleaseMemDstMemory = 0;
-	constexpr uint32_t ReleaseMemDstTcL2   = 1;
-	EXIT_NOT_IMPLEMENTED(release_dst != ReleaseMemDstMemory && release_dst != ReleaseMemDstTcL2);
-	EXIT_NOT_IMPLEMENTED(data_sel != 0 && data_sel != 1 && data_sel != 2 && data_sel != 3 &&
-	                     data_sel != 5);
-
-	LogUnknownReleaseMemGcr(gcr_cntl);
-
-	const bool gl2_writeback = ((gcr_cntl & GcrGl2Writeback) != 0);
-
-	auto trigger_interrupt = [&]() {
-		switch (interrupt_selector) {
-			case 0x00:
-			case 0x03: break;
-			case 0x01:
-			case 0x02:
-			case 0x04:
-				cp.TriggerEopEventAtEndOfPipe(interrupt_context_id);
-				cp.BufferFlush();
-				break;
-			default: EXIT("unknown release_mem interrupt selector\n");
-		}
-	};
-
-	// Cache actions only order GPU work, which the renderer's resource barriers already order.
-	if (data_sel == 0 || interrupt_selector == 4 ||
-	    (release_dst == ReleaseMemDstMemory && dst_gpu_addr == nullptr)) {
-		trigger_interrupt();
-
-		return 7;
-	}
-
-	auto cache_action = ReleaseMemCacheActionFromGcr(gcr_cntl);
-	cache_policy      = 0;
-
-	if (data_sel == 1) {
-		eop_event_type    = 0x2f;
-		event_index       = 6;
-		auto event_source = 2u;
-
-		cp.WriteAtEndOfPipe32(cache_policy, event_write_dest, eop_event_type, cache_action,
-		                      event_index, event_source, dst_gpu_addr, static_cast<uint32_t>(value),
-		                      interrupt_selector, interrupt_context_id);
-		if (interrupt_selector == 0x01 || interrupt_selector == 0x02) {
-			cp.BufferFlush();
-		}
-
-		return 7;
-	}
-
-	if (data_sel == 5) {
-		if (gl2_writeback) {
-			cache_action = 0;
-		}
-
-		eop_event_type    = 0x2f;
-		event_index       = 6;
-		auto event_source = 1u;
-
-		cp.WriteAtEndOfPipe32(cache_policy, event_write_dest, eop_event_type, cache_action,
-		                      event_index, event_source, dst_gpu_addr, static_cast<uint32_t>(value),
-		                      interrupt_selector, interrupt_context_id);
-		if (interrupt_selector == 0x01) {
-			cp.BufferFlush();
-		}
-
-		return 7;
-	}
-
-	const bool keep_native_event_index =
-	    (eop_event_type == 0x04 && cache_action == 0x00) ||
-	    (eop_event_type == 0x2f && event_index == 0x06 && cache_action == 0x38);
-
-	if (data_sel == 2) {
-		if (!keep_native_event_index) {
-			event_index = 0;
-		}
-	} else {
-		EXIT_IF(data_sel != 3);
-		if (!keep_native_event_index) {
-			event_index = 0;
-		}
-		data_sel = 4;
-	}
-
-	cp.WriteAtEndOfPipe64(cache_policy, event_write_dest, eop_event_type, cache_action, event_index,
-	                      data_sel, dst_gpu_addr, value, interrupt_selector, interrupt_context_id);
+	// The event type, cache actions and destination cache only order GPU work, which the renderer's
+	// resource barriers already order.
+	cp.ReleaseMem(
+	    (buffer[1] >> 29u) & 0x7u, (buffer[1] >> 24u) & 0x7u,
+	    reinterpret_cast<void*>((buffer[2] & ~0x3u) | (static_cast<uint64_t>(buffer[3]) << 32u)),
+	    buffer[4] | (static_cast<uint64_t>(buffer[5]) << 32u), buffer[6] & 0x07ffffffu);
 
 	return 7;
 }
 
+KYTY_CP_OP_PARSER(CpOpMemSemaphore) {
+	KYTY_PROFILER_FUNCTION();
+
+	const auto control = buffer[2];
+	EXIT_NOT_IMPLEMENTED(cmd_id != 0xc0023900 || (control & ~0xe0110000u) != 0 ||
+	                     (control >> 29u) < 6u);
+
+	auto* counter = reinterpret_cast<uint64_t*>(
+	    (buffer[0] & ~0x7u) | (static_cast<uint64_t>(buffer[1] & 0x3ffffu) << 32u));
+	// A signal updates the counter before the next packet, so its mailbox stall never waits.
+	cp.MemSemaphore(counter, (control >> 29u) == 6u, ((control >> 20u) & 0x1u) != 0);
+
+	return 3;
+}
+
 KYTY_CP_OP_PARSER(CpOpSetContextReg) {
 	KYTY_PROFILER_FUNCTION();
 
@@ -2524,42 +2291,6 @@ KYTY_CP_OP_PARSER(CpOpWaitRegMem64) {
 	return CpOpWaitRegMemSized<uint64_t>(cp, cmd_id, buffer);
 }
 
-KYTY_CP_OP_PARSER(CpOpWaitOnCeCounter) {
-	KYTY_PROFILER_FUNCTION();
-
-	EXIT_NOT_IMPLEMENTED(cmd_id != 0xC0008600);
-	EXIT_NOT_IMPLEMENTED(buffer[0] != 1);
-
-	cp.WaitCe();
-
-	return 1;
-}
-
-KYTY_CP_OP_PARSER(CpOpWaitOnDeCounterDiff) {
-	KYTY_PROFILER_FUNCTION();
-
-	EXIT_NOT_IMPLEMENTED(cmd_id != 0xC0008800);
-
-	cp.WaitDeDiff(buffer[0]);
-
-	return 1;
-}
-
-KYTY_CP_OP_PARSER(CpOpWriteConstRam) {
-	KYTY_PROFILER_FUNCTION();
-
-	auto dw_num = (cmd_id >> 16u) & 0x3fffu;
-	auto offset = buffer[0];
-
-	EXIT_NOT_IMPLEMENTED(dw_num >= 0x3000);
-	EXIT_NOT_IMPLEMENTED(offset > 0xbffc);
-	EXIT_NOT_IMPLEMENTED((offset & 0x3u) != 0);
-
-	cp.WriteConstRam(offset, buffer + 1, dw_num);
-
-	return 1 + dw_num;
-}
-
 KYTY_CP_OP_PARSER(CpOpWriteData) {
 	KYTY_PROFILER_FUNCTION();
 
@@ -3519,9 +3250,7 @@ void GraphicsInitJmpTablesShIndirect() {
 
 	for (uint32_t slot = 0; slot < 16; slot++) {
 		g_hw_sh_indirect_func[Pm4::COMPUTE_USER_DATA_0 + slot] = [](KYTY_HW_SH_INDIRECT_ARGS) {
-			auto sgpr = cmd_offset - Pm4::COMPUTE_USER_DATA_0;
-			cp.GetShCtx().SetCsUserSgpr(sgpr, value, cp.GetUserDataMarker());
-			cp.SetUserDataMarker(HW::UserSgprType::Unknown);
+			cp.GetShCtx().SetCsUserSgpr(cmd_offset - Pm4::COMPUTE_USER_DATA_0, value);
 		};
 	}
 	for (uint32_t slot = 0; slot < 4; slot++) {
@@ -3532,21 +3261,15 @@ void GraphicsInitJmpTablesShIndirect() {
 	for (uint32_t slot = 0; slot < 32; slot++) {
 		g_hw_sh_indirect_func[Pm4::SPI_SHADER_USER_DATA_PS_0 + slot] =
 		    [](KYTY_HW_SH_INDIRECT_ARGS) {
-			    auto sgpr = cmd_offset - Pm4::SPI_SHADER_USER_DATA_PS_0;
-			    cp.GetShCtx().SetPsUserSgpr(sgpr, value, cp.GetUserDataMarker());
-			    cp.SetUserDataMarker(HW::UserSgprType::Unknown);
+			    cp.GetShCtx().SetPsUserSgpr(cmd_offset - Pm4::SPI_SHADER_USER_DATA_PS_0, value);
 		    };
 		g_hw_sh_indirect_func[Pm4::SPI_SHADER_USER_DATA_GS_0 + slot] =
 		    [](KYTY_HW_SH_INDIRECT_ARGS) {
-			    auto sgpr = cmd_offset - Pm4::SPI_SHADER_USER_DATA_GS_0;
-			    cp.GetShCtx().SetGsUserSgpr(sgpr, value, cp.GetUserDataMarker());
-			    cp.SetUserDataMarker(HW::UserSgprType::Unknown);
+			    cp.GetShCtx().SetGsUserSgpr(cmd_offset - Pm4::SPI_SHADER_USER_DATA_GS_0, value);
 		    };
 		g_hw_sh_indirect_func[Pm4::SPI_SHADER_USER_DATA_HS_0 + slot] =
 		    [](KYTY_HW_SH_INDIRECT_ARGS) {
-			    auto sgpr = cmd_offset - Pm4::SPI_SHADER_USER_DATA_HS_0;
-			    cp.GetShCtx().SetHsUserSgpr(sgpr, value, cp.GetUserDataMarker());
-			    cp.SetUserDataMarker(HW::UserSgprType::Unknown);
+			    cp.GetShCtx().SetHsUserSgpr(cmd_offset - Pm4::SPI_SHADER_USER_DATA_HS_0, value);
 		    };
 	}
 	for (uint32_t slot = 0; slot < 4; slot++) {

src/graphics/guest_gpu/graphicsRun.cpp

@@ -30,6 +30,7 @@
 #include <mutex>
 #include <semaphore>
 #include <thread>
+#include <utility>
 #include <vector>
 
 namespace Libs::Graphics {
@@ -39,21 +40,6 @@ static thread_local Pm4Execution*     g_current_execution = nullptr;
 static thread_local bool              g_gpu_thread        = false;
 static thread_local GuestGpu*         g_gpu_state         = nullptr;
 
-struct DrawIndirectArgs {
-	uint32_t vertex_count_per_instance;
-	uint32_t instance_count;
-	uint32_t start_vertex_location;
-	uint32_t start_instance_location;
-};
-
-struct DrawIndexedIndirectArgs {
-	uint32_t index_count_per_instance;
-	uint32_t instance_count;
-	uint32_t start_index_location;
-	uint32_t base_vertex_location;
-	uint32_t start_instance_location;
-};
-
 static bool GraphicsRunDebugDumpEnabled() {
 	return Config::GraphicsDebugDumpEnabled() &&
 	       Config::GetPrintfDirection() != Config::LogDirection::Silent;
@@ -106,6 +92,12 @@ void GuestGpu::SendCommand(Common::UniqueFunction<void>&& command) {
 	m_work_available.Signal();
 }
 
+void GuestGpu::Wake() {
+	Common::LockGuard lock(m_queue_mutex);
+	m_wake_pending = true;
+	m_work_available.Signal();
+}
+
 void GuestGpu::ProcessCommands() {
 	EXIT_IF(!IsGpuThread());
 	while (m_pending_commands.load(std::memory_order_acquire) != 0) {
@@ -135,16 +127,14 @@ void GuestGpu::SendCommandSync(Common::UniqueFunction<void>&& command) {
 	done.acquire();
 }
 
-void GuestGpu::Submit(std::span<const uint32_t> draw_commands,
-                      std::span<const uint32_t> constant_commands) {
-	if (draw_commands.empty()) {
+void GuestGpu::Submit(std::span<const uint32_t> commands) {
+	if (commands.empty()) {
 		return;
 	}
 	Submission submission;
-	submission.type              = SubmissionType::Graphics;
-	submission.queue_id          = 0;
-	submission.commands          = draw_commands;
-	submission.constant_commands = constant_commands;
+	submission.type     = SubmissionType::Graphics;
+	submission.queue_id = 0;
+	submission.commands = commands;
 	Enqueue(std::move(submission));
 }
 
@@ -206,12 +196,10 @@ void CommandProcessor::Reset() {
 	m_index_buffer_size                = 0;
 	m_index_base_addr                  = 0;
 	m_num_instances                    = 1;
+	m_num_instances_addr               = 0;
 	m_predicate_skip                   = false;
-	m_user_data_marker                 = HW::UserSgprType::Unknown;
 	m_draw_indirect_args_base_addr     = 0;
 	m_dispatch_indirect_args_base_addr = 0;
-
-	std::memset(m_const_ram, 0, sizeof(m_const_ram));
 }
 
 void CommandProcessor::ApplyContextStateOperation(ContextStateOperation operation) {
@@ -244,6 +232,8 @@ void CommandProcessor::BufferInit() {
 
 void CommandProcessor::BufferFlush() {
 	GetScheduler().Flush();
+	// The occlusion counters reach a guest that polls them without waiting for a later fence.
+	m_renderer.GetOcclusionQueries().Publish();
 }
 
 void CommandProcessor::BufferWait() {
@@ -251,47 +241,12 @@ void CommandProcessor::BufferWait() {
 	GetScheduler().Finish();
 }
 
-void CommandProcessor::ResetDeCe() {
-	m_de_count    = 0;
-	m_ce_count    = 0;
-	m_ce_complete = false;
-}
-
-void CommandProcessor::WaitCe() {
-	if (m_ce_count <= m_de_count && !m_ce_complete) {
-		SuspendPm4();
-	}
-}
-
-void CommandProcessor::WaitDeDiff(uint32_t diff) {
-	EXIT_IF(m_de_count > m_ce_count);
-	if (m_ce_count - m_de_count >= diff) {
-		SuspendPm4();
-	}
-}
-
 void CommandProcessor::WaitForRewind(bool valid) {
 	if (!valid) {
 		SuspendPm4();
 	}
 }
 
-void CommandProcessor::IncrementDe() {
-	m_de_count++;
-}
-
-void CommandProcessor::IncrementCe() {
-	m_ce_count++;
-}
-
-void CommandProcessor::WriteConstRam(uint32_t offset, const uint32_t* src, uint32_t dw_num) {
-	memcpy(m_const_ram + offset / 4, src, static_cast<size_t>(dw_num) * 4);
-}
-
-void CommandProcessor::DumpConstRam(uint32_t* dst, uint32_t offset, uint32_t dw_num) {
-	memcpy(dst, m_const_ram + offset / 4, static_cast<size_t>(dw_num) * 4);
-}
-
 bool TestWaitRegMemValue(uint64_t value, uint64_t ref, uint64_t mask, uint32_t func) {
 	switch (func) {
 		case 0: return true;
@@ -326,6 +281,16 @@ template void CommandProcessor::WaitRegMem<uint32_t>(uint32_t, const uint32_t*,
 template void CommandProcessor::WaitRegMem<uint64_t>(uint32_t, const uint64_t*, uint64_t, uint64_t,
                                                      uint32_t, uint32_t);
 
+void CommandProcessor::MemSemaphore(uint64_t* counter, bool signal, bool write) {
+	if (signal) {
+		*counter = write ? 1u : *counter + 1u;
+	} else if (*counter == 0) {
+		SuspendPm4();
+	} else {
+		(*counter)--;
+	}
+}
+
 void CommandProcessor::WriteData(uint32_t* dst, const uint32_t* src, uint32_t dw_num,
                                  uint32_t write_control) {
 	const uint32_t dst_sel = ((write_control >> 30u) & 0x1u) | ((write_control >> 7u) & 0x1eu);
@@ -478,7 +443,10 @@ void GuestGpu::ThreadRun(void* data) {
 				}
 				if (selected_queue < 0) {
 					gpu->m_processing = false;
-					gpu->m_work_available.WaitFor(&gpu->m_queue_mutex, 100);
+					// A wake that came before this wait must not be lost.
+					if (!std::exchange(gpu->m_wake_pending, false)) {
+						gpu->m_work_available.WaitFor(&gpu->m_queue_mutex, 100);
+					}
 					for (auto& queue: gpu->m_queues) {
 						if (!queue.empty()) {
 							queue.front().blocked = false;
@@ -543,64 +511,34 @@ bool GuestGpu::Process(Submission& submission) {
 	if (first_slice) {
 		submission.started = true;
 		cp.SetSubmitId(++m_submit_id);
-		cp.ResetDeCe();
-		cp.SetFlip({});
 	}
 
 	cp.BufferInit();
 	bool complete = true;
 
 	switch (submission.type) {
-		case SubmissionType::Graphics: {
-			bool progressed = false;
-			submission.constant_complete |= submission.constant_commands.empty();
-			for (;;) {
-				bool round_progress = false;
-				if (!submission.constant_complete) {
-					submission.constant_complete =
-					    cp.Process(submission.constant_execution, submission.constant_commands) ==
-					    Pm4ProcessResult::Complete;
-					round_progress |= submission.constant_execution.MadeProgress();
-				}
-				cp.SetCeComplete(submission.constant_complete);
-				if (!submission.command_complete) {
-					submission.command_complete =
-					    cp.Process(submission.command_execution, submission.commands) ==
-					    Pm4ProcessResult::Complete;
-					round_progress |= submission.command_execution.MadeProgress();
-				}
-				progressed |= round_progress;
-				complete = submission.command_complete && submission.constant_complete;
-				if (complete || !round_progress) {
-					break;
-				}
-			}
-			if (progressed) {
-				cp.BufferFlush();
-			}
-			break;
-		}
-		case SubmissionType::Compute: {
-			const auto      num_dw = static_cast<uint32_t>(submission.commands.size());
-			const auto*     buffer = submission.commands.data();
-			static uint32_t compute_batch_log_count = 0;
-			if (first_slice && num_dw <= 128 && compute_batch_log_count++ < 32) {
-				LOGF("compute direct batch: data=0x%016" PRIx64 ", num_dw=%" PRIu32 "\n",
-				     reinterpret_cast<uint64_t>(buffer), num_dw);
-				for (uint32_t i = 0; i < std::min<uint32_t>(num_dw, 16); i++) {
-					LOGF("\t compute[%02" PRIu32 "] = 0x%08" PRIx32 "\n", i, buffer[i]);
-				}
-			}
+		case SubmissionType::Compute:
 			if (first_slice) {
+				const auto      num_dw = static_cast<uint32_t>(submission.commands.size());
+				const auto*     buffer = submission.commands.data();
+				static uint32_t compute_batch_log_count = 0;
+				if (num_dw <= 128 && compute_batch_log_count++ < 32) {
+					LOGF("compute direct batch: data=0x%016" PRIx64 ", num_dw=%" PRIu32 "\n",
+					     reinterpret_cast<uint64_t>(buffer), num_dw);
+					for (uint32_t i = 0; i < std::min<uint32_t>(num_dw, 16); i++) {
+						LOGF("\t compute[%02" PRIu32 "] = 0x%08" PRIx32 "\n", i, buffer[i]);
+					}
+				}
 				GraphicsDbgDumpDcb("cc", num_dw, buffer);
 			}
-			complete = cp.Process(submission.command_execution, submission.commands) ==
-			           Pm4ProcessResult::Complete;
-			if (submission.command_execution.MadeProgress()) {
+			[[fallthrough]];
+		case SubmissionType::Graphics:
+			complete =
+			    cp.Process(submission.execution, submission.commands) == Pm4ProcessResult::Complete;
+			if (submission.execution.MadeProgress()) {
 				cp.BufferFlush();
 			}
 			break;
-		}
 		case SubmissionType::FlipPreparation:
 			cp.PrepareCpuFlip(submission.flip_request_id);
 			break;
@@ -703,18 +641,6 @@ void CommandProcessor::ProcessPm4(Pm4Execution& execution) {
 				     opcode, KYTY_PM4_R(packet_header), packet_dw,
 				     reinterpret_cast<uint64_t>(packet), packet_header);
 			}
-			if (opcode == Pm4::IT_NOP && KYTY_PM4_R(packet_header) == Pm4::R_RELEASE_MEM &&
-			    packet_dw >= 7) {
-				static std::atomic<uint32_t> log_count {0};
-				if (log_count.fetch_add(1) < 128) {
-					const auto dst = packet[3] | (static_cast<uint64_t>(packet[4]) << 32u);
-					const auto val = packet[5] | (static_cast<uint64_t>(packet[6]) << 32u);
-					LOGF("\t predicated skip: R_RELEASE_MEM dst=0x%016" PRIx64
-					     ", value=0x%016" PRIx64 ", action=0x%08" PRIx32
-					     ", gcr/data/int=0x%08" PRIx32 "\n",
-					     dst, val, packet[1], packet[2]);
-				}
-			}
 			cursor.offset_dw += packet_dw;
 			execution.m_made_progress = true;
 			continue;
@@ -784,7 +710,16 @@ void CommandProcessor::SetNumInstances(uint32_t num_instances) {
 		num_instances = 1;
 	}
 
-	m_num_instances = num_instances;
+	m_num_instances      = num_instances;
+	m_num_instances_addr = 0;
+}
+
+uint32_t CommandProcessor::NumInstances() {
+	if (m_num_instances_addr != 0) {
+		m_num_instances      = *reinterpret_cast<const uint32_t*>(m_num_instances_addr);
+		m_num_instances_addr = 0;
+	}
+	return m_num_instances;
 }
 
 void CommandProcessor::SetPredication(uint32_t condition, uint32_t op, uint32_t wait_op,
@@ -798,6 +733,7 @@ void CommandProcessor::SetPredication(uint32_t condition, uint32_t op, uint32_t
 			return;
 		case 0x01: {
 			EXIT_NOT_IMPLEMENTED(address == nullptr);
+			m_renderer.GetOcclusionQueries().Publish();
 			// One begin/end pair per DB; bit 63 marks each counter ready.
 			constexpr uint64_t ready_bit = 1ull << 63u;
 			const auto* results = reinterpret_cast<const volatile uint64_t*>(address);
@@ -841,16 +777,12 @@ void CommandProcessor::SetPredication(uint32_t condition, uint32_t op, uint32_t
 void CommandProcessor::DrawIndex(DrawIndexArgs args) {
 	args.index_type_and_size = m_index_type_and_size;
 	if (args.instance_count == 0) {
-		args.instance_count = m_num_instances;
-	}
-	if (GraphicsRunDebugDumpEnabled() && (args.base_vertex != 0 || args.first_instance != 0)) {
-		LOGF("\t draw indexed offsets: base_vertex = %" PRId32 ", first_instance = %" PRIu32 "\n",
-		     args.base_vertex, args.first_instance);
+		args.instance_count = NumInstances();
 	}
 	m_renderer.GetRenderExecutor().DrawIndex(m_submit_id, CurrentBuffer(), args);
 }
 
-void CommandProcessor::DrawIndexOffset(uint32_t index_offset, uint32_t index_count) {
+const void* CommandProcessor::IndexAddress(uint32_t first_index) const {
 	uint64_t index_size = 0;
 	switch (m_index_type_and_size) {
 		case 0: index_size = 2; break;
@@ -858,16 +790,16 @@ void CommandProcessor::DrawIndexOffset(uint32_t index_offset, uint32_t index_cou
 		case 2: index_size = 1; break;
 		default: EXIT("unknown index_type_and_size: %u\n", m_index_type_and_size);
 	}
+	return reinterpret_cast<const void*>(m_index_base_addr + first_index * index_size);
+}
 
-	auto* index_addr = reinterpret_cast<const void*>(
-	    m_index_base_addr + static_cast<uint64_t>(index_offset) * index_size);
-
-	DrawIndex({.index_count = index_count, .index_addr = index_addr});
+void CommandProcessor::DrawIndexOffset(uint32_t index_offset, uint32_t index_count) {
+	DrawIndex({.index_count = index_count, .index_addr = IndexAddress(index_offset)});
 }
 
 static void PatchIndirectDrawOffsets(CommandProcessor& cp, IndirectDrawRegisters registers,
                                      uint32_t vertex_offset, uint32_t instance_offset,
-                                     uint32_t first_index) {
+                                     uint32_t first_index, uint32_t draw_index) {
 	const auto write = [&](uint32_t location, uint32_t value) {
 		if (location == Pm4::SH_NOP) {
 			return;
@@ -879,11 +811,34 @@ static void PatchIndirectDrawOffsets(CommandProcessor& cp, IndirectDrawRegisters
 	write(registers.vertex_offset, vertex_offset);
 	write(registers.instance_offset, instance_offset);
 	write(registers.index_offset, first_index);
+	write(registers.draw_index, draw_index);
+}
+
+// Writes the GS user-data slots the packet patches; false when the shaders cannot read the
+// offsets from GPU records: a first-index or subdraw-index destination, or another stage's
+// user data.
+static bool IndirectDrawOffsets(IndirectDrawRegisters registers, ShaderIndirectOffsets& offsets) {
+	if (registers.index_offset != Pm4::SH_NOP || registers.draw_index != Pm4::SH_NOP) {
+		return false;
+	}
+	for (const auto [location, slot]: {std::pair {registers.vertex_offset, &offsets.vertex},
+	                                   std::pair {registers.instance_offset, &offsets.instance}}) {
+		if (location == Pm4::SH_NOP) {
+			continue;
+		}
+		const uint32_t index = location - Pm4::SPI_SHADER_USER_DATA_GS_0;
+		if (index >= HW::UserSgprInfo::SGPRS_MAX) {
+			return false;
+		}
+		*slot = static_cast<uint8_t>(index);
+	}
+	return true;
 }
 
 void CommandProcessor::DrawIndirect(uint32_t data_offset, IndirectDrawRegisters registers,
                                     uint32_t draw_initiator, bool indexed) {
-	const auto args_size = indexed ? sizeof(DrawIndexedIndirectArgs) : sizeof(DrawIndirectArgs);
+	const auto args_size =
+	    indexed ? sizeof(vk::DrawIndexedIndirectCommand) : sizeof(vk::DrawIndirectCommand);
 	DrawIndirectMulti(data_offset, 1, nullptr, args_size, registers, draw_initiator, indexed);
 }
 
@@ -893,77 +848,64 @@ void CommandProcessor::DrawIndirectMulti(uint32_t data_offset, uint32_t max_coun
                                          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);
+	// The guest records match Vulkan's indirect commands byte for byte.
+	EXIT_NOT_IMPLEMENTED((stride_in_bytes & 3u) != 0 ||
+	                     stride_in_bytes < (indexed ? sizeof(vk::DrawIndexedIndirectCommand)
+	                                                : sizeof(vk::DrawIndirectCommand)));
 
-	uint32_t draw_count = max_count_or_count;
-	if (count_addr != nullptr) {
-		draw_count = *count_addr;
-		if (draw_count > max_count_or_count) {
-			draw_count = max_count_or_count;
-		}
-	}
-
+	const auto draw_count = count_addr == nullptr
+	                            ? max_count_or_count
+	                            : std::min(uint32_t {*count_addr}, max_count_or_count);
 	if (draw_count == 0) {
 		return;
 	}
 	Common::DebugCounters::Add(Common::DebugCounters::Counter::IndirectDraws, draw_count);
 
-	const auto args_size = indexed ? sizeof(DrawIndexedIndirectArgs) : sizeof(DrawIndirectArgs);
-	EXIT_NOT_IMPLEMENTED(stride_in_bytes < args_size);
-
-	uint64_t index_size = 0;
-	if (indexed) {
-		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&      executor  = m_renderer.GetRenderExecutor();
+	const auto args_addr = m_draw_indirect_args_base_addr + data_offset;
+	// NUM_INSTANCES keeps the last draw's instance count.
+	m_num_instances_addr = args_addr + static_cast<uint64_t>(draw_count - 1) * stride_in_bytes +
+	                       offsetof(vk::DrawIndirectCommand, instanceCount);
+	// The GPU reads the records if the shaders allow; the VS prolog loads the patched SGPRs from
+	// Vulkan's bases. With no SGPR patched, Vulkan applies the offsets, which matches while they
+	// are zero.
+	DrawIndirectArgs records {args_addr, draw_count, stride_in_bytes};
+	if (IndirectDrawOffsets(registers, records.offsets) && (!indexed || m_index_buffer_size != 0) &&
+	    (indexed ? executor.DrawIndex(m_submit_id, CurrentBuffer(),
+	                                  {.index_count         = m_index_buffer_size,
+	                                   .index_addr          = IndexAddress(0),
+	                                   .index_type_and_size = m_index_type_and_size,
+	                                   .indirect            = records})
+	             : executor.DrawAuto(m_submit_id, CurrentBuffer(), {.indirect = records}))) {
+		return;
 	}
 
+	const uint32_t max_indices = m_index_buffer_size != 0 ? m_index_buffer_size : UINT32_MAX;
 	for (uint32_t i = 0; i < draw_count; i++) {
-		const auto args_addr = m_draw_indirect_args_base_addr + data_offset +
-		                       static_cast<uint64_t>(i) * stride_in_bytes;
-
+		const auto addr = args_addr + static_cast<uint64_t>(i) * stride_in_bytes;
 		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,
-			               .first_vertex   = args->start_vertex_location,
-			               .first_instance = args->start_instance_location,
-			               .offset_source  = DrawOffsetSource::IndirectArgs});
+			const auto& args = *reinterpret_cast<const vk::DrawIndirectCommand*>(addr);
+			PatchIndirectDrawOffsets(*this, registers, args.firstVertex, args.firstInstance, 0, i);
+			executor.DrawAuto(m_submit_id, CurrentBuffer(),
+			                  {.vertex_count   = args.vertexCount,
+			                   .instance_count = args.instanceCount,
+			                   .first_vertex   = args.firstVertex,
+			                   .first_instance = args.firstInstance,
+			                   .offset_source  = DrawOffsetSource::IndirectArgs});
 			continue;
 		}
 
-		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);
-
-		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 DrawIndexIndirectMulti: clamped index_count from %" PRIu32 " to %" PRIu32
-				     " using INDEX_BUFFER_SIZE\n",
-				     args->index_count_per_instance, index_count);
-			}
-		}
-
-		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});
+		const auto& args = *reinterpret_cast<const vk::DrawIndexedIndirectCommand*>(addr);
+		PatchIndirectDrawOffsets(*this, registers, static_cast<uint32_t>(args.vertexOffset),
+		                         args.firstInstance, args.firstIndex, i);
+		executor.DrawIndex(m_submit_id, CurrentBuffer(),
+		                   {.index_count         = std::min(args.indexCount, max_indices),
+		                    .index_addr          = IndexAddress(args.firstIndex),
+		                    .instance_count      = args.instanceCount,
+		                    .index_type_and_size = m_index_type_and_size,
+		                    .base_vertex         = args.vertexOffset,
+		                    .first_instance      = args.firstInstance,
+		                    .offset_source       = DrawOffsetSource::IndirectArgs});
 	}
 }
 
@@ -1012,221 +954,86 @@ void CommandProcessor::DispatchIndirect(uint64_t args_addr, uint32_t mode) {
 
 void CommandProcessor::DrawIndexAuto(DrawAutoArgs args) {
 	if (args.instance_count == 0) {
-		args.instance_count = m_num_instances;
+		args.instance_count = NumInstances();
 	}
 	m_renderer.GetRenderExecutor().DrawAuto(m_submit_id, CurrentBuffer(), args);
 }
 
 void CommandProcessor::WaitFlipDone(uint32_t video_out_handle, uint32_t display_buffer_index) {
-	BufferFlush();
-
-	m_renderer.GetVideoOut().WaitFlipDone(static_cast<int>(video_out_handle),
-	                                      static_cast<int>(display_buffer_index));
+	// Stall only this command processor, not the host command lane or other queues.
+	if (!m_renderer.GetVideoOut().IsSafeForRendering(static_cast<int>(video_out_handle),
+	                                                 static_cast<int>(display_buffer_index))) {
+		SuspendPm4();
+	}
 }
 
-template <typename T>
-void CommandProcessor::WriteAtEndOfPipe(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, T value, uint32_t interrupt_selector,
-                                        uint32_t interrupt_context_id) {
-	static_assert(sizeof(T) == sizeof(uint32_t) || sizeof(T) == sizeof(uint64_t));
-
-	auto& command = CurrentBuffer();
-
+void CommandProcessor::ReleaseMem(uint32_t data_sel, uint32_t interrupt_selector, void* address,
+                                  uint64_t data, uint32_t interrupt_context_id) {
 	if (GraphicsRunDebugDumpEnabled()) {
-		const auto bits      = static_cast<unsigned>(sizeof(T) * 8u);
-		const auto log_width = static_cast<int>(sizeof(T) * 2u);
-
-		LOGF("CommandProcessor::WriteAtEndOfPipe%u()\n"
-		     "\t cache_policy        = 0x%08" PRIx32 "\n"
-		     "\t event_write_dest    = 0x%08" PRIx32 "\n"
-		     "\t eop_event_type      = 0x%08" PRIx32 "\n"
-		     "\t cache_action        = 0x%08" PRIx32 "\n"
-		     "\t event_index         = 0x%08" PRIx32 "\n"
-		     "\t event_write_source  = 0x%08" PRIx32 "\n"
-		     "\t interrupt_selector  = 0x%08" PRIx32 "\n"
-		     "\t interrupt_context   = 0x%08" PRIx32 "\n"
-		     "\t dst_gpu_addr        = 0x%016" PRIx64 "\n"
-		     "\t value               = 0x%0*" PRIx64 "\n",
-		     bits, cache_policy, event_write_dest, eop_event_type, cache_action, event_index,
-		     event_write_source, interrupt_selector, interrupt_context_id,
-		     reinterpret_cast<uint64_t>(dst_gpu_addr), log_width, static_cast<uint64_t>(value));
+		LOGF("CommandProcessor::ReleaseMem()\n"
+		     "\t data_sel           = %" PRIu32 "\n"
+		     "\t interrupt_selector = %" PRIu32 "\n"
+		     "\t address            = 0x%016" PRIx64 "\n"
+		     "\t data               = 0x%016" PRIx64 "\n"
+		     "\t interrupt_context  = 0x%08" PRIx32 "\n",
+		     data_sel, interrupt_selector, reinterpret_cast<uint64_t>(address), data,
+		     interrupt_context_id);
 	}
 
-	EXIT_NOT_IMPLEMENTED(cache_policy != 0x00000000);
-	EXIT_NOT_IMPLEMENTED(event_write_dest != 0x00000000);
-
-	bool with_interrupt = false;
-	switch (interrupt_selector) {
-		case 0x00:
-		case 0x03: with_interrupt = false; break;
-		case 0x01:
-			if (!IsAsyncComputeQueue()) {
-				TriggerEopEventAtEndOfPipe(interrupt_context_id);
-				return;
-			}
-			with_interrupt = true;
-			break;
-		case 0x02: with_interrupt = true; break;
-		default: EXIT("unknown interrupt_selector\n");
+	// QueuedInterrupt: kNone, kOnly, kAfterWriteConfirm, kDataAfterWriteConfirm, kSendIntCtx.
+	EXIT_NOT_IMPLEMENTED(interrupt_selector > 4);
+	const bool interrupt = (interrupt_selector != 0 && interrupt_selector != 3);
+	// kSendIntCtx has no label, and kOnly writes none on graphics queues.
+	if (address == nullptr || interrupt_selector == 4 ||
+	    (interrupt_selector == 1 && !IsAsyncComputeQueue())) {
+		data_sel = 0;
+	}
+	if (data_sel == 0 && !interrupt) {
+		return;
 	}
 
-	auto write32 = [&] {
-		auto* dst  = static_cast<uint32_t*>(dst_gpu_addr);
-		auto  data = static_cast<uint32_t>(value);
-		EXIT_IF(dst == nullptr);
-		(void)command.Handle();
-		std::memcpy(dst, &data, sizeof(data));
+	// Dumped occlusion counters are written before any end-of-pipe signal reaches the guest.
+	m_renderer.GetOcclusionQueries().Publish();
 
-		if (with_interrupt) {
-			TriggerEopEventAtEndOfPipe(interrupt_context_id);
+	// A label beside GPU-written bytes goes to the backing store, so they are not downloaded.
+	// Other labels go through the guest mapping: its fault keeps a clean cached copy current and
+	// downloads GPU-written bytes under the label before the store.
+	const auto write_label = [this, address](const void* value, uint64_t size) {
+		const auto vaddr   = reinterpret_cast<uint64_t>(address);
+		auto&      buffers = m_renderer.GetBufferCache();
+		if (!GuestRange {vaddr, size}.Valid() || !buffers.IsRegionGpuModified(vaddr, size) ||
+		    buffers.HasGpuDirtyBytes(vaddr, size) ||
+		    !LibKernel::Memory::TryWriteBacking(vaddr, value, size)) {
+			std::memcpy(address, value, size);
 		}
 	};
-
-	switch (event_write_source) {
-		case 0x01:
-			if constexpr (sizeof(T) == sizeof(uint32_t)) {
-				if (eop_event_type == 0x2f && cache_action == 0x00 && event_index == 0x06) {
-					auto* dst = static_cast<uint32_t*>(dst_gpu_addr);
-					SynchronizeGpu();
-					Sync::ReadGds(*m_renderer.GetBufferCache().GetGdsBuffer(), dst, value & 0xffffu,
-					              value >> 16u);
-					if (with_interrupt) {
-						m_renderer.TriggerInterrupt(m_interrupt_event_id, interrupt_context_id);
-					}
-					return;
-				}
-			} else if (eop_event_type == 0x04 && cache_action == 0x00 && event_index == 0x05) {
-				write32();
-				return;
-			}
+	// EndOfPipeData: kNone, kImmediate32b, kImmediate64b, kRefClk; 5 reads GDS.
+	switch (data_sel) {
+		case 0: break;
+		case 1: {
+			const auto value = static_cast<uint32_t>(data);
+			write_label(&value, sizeof(value));
 			break;
-		case 0x02:
-		case 0x04:
-			if constexpr (sizeof(T) == sizeof(uint32_t)) {
-				if (event_write_source == 0x02 && eop_event_type == 0x2f && event_index == 0x06) {
-					switch (cache_action) {
-						case 0x00:
-						case 0x38: write32(); return;
-						default: break;
-					}
-				}
-			} else {
-				if (event_write_source == 0x04) {
-					value = Sync::ReadReferenceClock();
-				}
-				auto write64 = [&] {
-					auto* dst = static_cast<uint64_t*>(dst_gpu_addr);
-					EXIT_IF(dst == nullptr);
-					(void)command.Handle();
-					std::memcpy(dst, &value, sizeof(value));
-
-					if (with_interrupt) {
-						TriggerEopEventAtEndOfPipe(interrupt_context_id);
-					}
-				};
-
-				switch (cache_action) {
-					case 0x00:
-						switch (eop_event_type) {
-							case 0x04:
-								if (event_index == 0x05) {
-									write64();
-									return;
-								}
-								break;
-							case 0x14:
-							case 0x28:
-							case 0x2f:
-								if (event_index == 0x00) {
-									write64();
-									return;
-								}
-								break;
-							case 0x2b:
-							case 0x2d:
-							case 0x30:
-								if (event_index == 0x00 && !with_interrupt) {
-									write64();
-									return;
-								}
-								break;
-							default: break;
-						}
-						break;
-					case 0x38:
-						switch (eop_event_type) {
-							case 0x04:
-							case 0x14:
-							case 0x28:
-								if (((eop_event_type == 0x04 || eop_event_type == 0x28) &&
-								     event_index == 0x05) ||
-								    (event_index == 0x00)) {
-									write64();
-									return;
-								}
-								break;
-							case 0x2b:
-							case 0x2d:
-								if (event_index == 0x00 && !with_interrupt) {
-									write64();
-									return;
-								}
-								break;
-							case 0x2f:
-								if (event_index == 0x06 && !with_interrupt) {
-									write64();
-									return;
-								}
-								break;
-							default: break;
-						}
-						break;
-					case 0x3b:
-						if (eop_event_type == 0x04 && event_index == 0x05 && with_interrupt) {
-							write64();
-							return;
-						}
-						break;
-					default: break;
-				}
-			}
+		}
+		case 3: data = Sync::ReadReferenceClock(); [[fallthrough]];
+		case 2: write_label(&data, sizeof(data)); break;
+		case 5: {
+			GetScheduler().Finish();
+			const auto gds = static_cast<uint32_t>(data);
+			Sync::ReadGds(*m_renderer.GetBufferCache().GetGdsBuffer(),
+			              static_cast<uint32_t*>(address), gds & 0xffffu, gds >> 16u);
 			break;
-		default: break;
+		}
+		default: EXIT("unknown release_mem data selector: %" PRIu32 "\n", data_sel);
 	}
 
-	EXIT("unknown event type\n");
-}
-
-void CommandProcessor::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,
-                                          uint32_t interrupt_selector,
-                                          uint32_t interrupt_context_id) {
-	WriteAtEndOfPipe(cache_policy, event_write_dest, eop_event_type, cache_action, event_index,
-	                 event_write_source, dst_gpu_addr, value, interrupt_selector,
-	                 interrupt_context_id);
-}
-
-void CommandProcessor::WriteAtEndOfPipe64(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, uint64_t value,
-                                          uint32_t interrupt_selector,
-                                          uint32_t interrupt_context_id) {
-	WriteAtEndOfPipe(cache_policy, event_write_dest, eop_event_type, cache_action, event_index,
-	                 event_write_source, dst_gpu_addr, value, interrupt_selector,
-	                 interrupt_context_id);
-}
-
-void CommandProcessor::TriggerEopEventAtEndOfPipe(uint32_t interrupt_context_id) {
-	(void)CurrentBuffer().Handle();
-	GetScheduler().DeferPriorityOperation(
-	    [&renderer = m_renderer, event_id = m_interrupt_event_id, interrupt_context_id] {
-		    renderer.TriggerInterrupt(event_id, interrupt_context_id);
-	    });
+	// GuestGpu::Process's end-of-slice flush submits the work this interrupt waits on.
+	if (interrupt) {
+		GetScheduler().DeferPriorityOperation(
+		    [&renderer = m_renderer, event_id = m_interrupt_event_id, interrupt_context_id] {
+			    renderer.TriggerInterrupt(event_id, interrupt_context_id);
+		    });
+	}
 }
 
 void CommandProcessor::TriggerEvent(uint32_t event_type, uint32_t event_index,
@@ -1279,15 +1086,19 @@ void CommandProcessor::TriggerEvent(uint32_t event_type, uint32_t event_index,
 				     "\n",
 				     event_index, event_address);
 			}
+			if (auto& queries = m_renderer.GetOcclusionQueries(); queries.Enabled()) {
+				queries.Dump(event_address);
+				break;
+			}
 			static std::once_flag warning_once;
 			std::call_once(warning_once, [] {
-				std::printf("Warning: game uses occlusion queries, which are currently treated as "
-				            "always visible; GPU usage may be higher and FPS may be lower.\n");
+				std::printf("Warning: game uses occlusion queries, which are treated as always "
+				            "visible without --occlusion-queries; GPU usage may be higher and FPS "
+				            "may be lower.\n");
 			});
 
-			// Until host occlusion queries are implemented, publish an always-visible result. The
-			// PS5 layout contains one interleaved begin/end pair per DB, and bit 63 marks a result
-			// ready.
+			// Without host occlusion queries, publish an always-visible result. The PS5 layout
+			// contains one interleaved begin/end pair per DB, and bit 63 marks a result ready.
 			constexpr uint64_t ready_bit    = 1ull << 63u;
 			constexpr uint64_t counter_mask = ready_bit - 1u;
 			auto*              results      = reinterpret_cast<volatile uint64_t*>(event_address);
@@ -1303,69 +1114,20 @@ void CommandProcessor::TriggerEvent(uint32_t event_type, uint32_t event_index,
 	}
 }
 
-void CommandProcessor::Flip() {
+void CommandProcessor::Flip(int handle, int index, int flip_mode, int64_t flip_arg) {
 	if (GraphicsRunDebugDumpEnabled()) {
 		LOGF("CommandProcessor::Flip()\n");
 	}
+	m_renderer.GetOcclusionQueries().Publish();
 
 	auto& command = CurrentBuffer();
-	auto request = Sync::PrepareVideoOutFlip(command, m_flip.handle, m_flip.index, m_flip.flip_mode,
-	                                         m_flip.flip_arg);
+	auto  request = Sync::PrepareVideoOutFlip(command, handle, index, flip_mode, flip_arg);
 	(void)command.Handle();
 	GetScheduler().DeferPriorityOperation(
 	    [&renderer = m_renderer, request] { renderer.GetVideoOut().CompleteFlip(request); });
 	GetScheduler().Flush();
 }
 
-void CommandProcessor::Flip(void* dst_gpu_addr, uint32_t value) {
-	auto& command = CurrentBuffer();
-
-	if (GraphicsRunDebugDumpEnabled()) {
-		LOGF("CommandProcessor::Flip()\n"
-		     "\t dst_gpu_addr = 0x%016" PRIx64 "\n"
-		     "\t value        = 0x%08" PRIx32 "\n",
-		     reinterpret_cast<uint64_t>(dst_gpu_addr), value);
-	}
-
-	EXIT_IF(dst_gpu_addr == nullptr);
-	std::memcpy(dst_gpu_addr, &value, sizeof(value));
-	auto request = Sync::PrepareVideoOutFlip(command, m_flip.handle, m_flip.index, m_flip.flip_mode,
-	                                         m_flip.flip_arg);
-	(void)command.Handle();
-	GetScheduler().DeferPriorityOperation(
-	    [&renderer = m_renderer, request] { renderer.GetVideoOut().CompleteFlip(request); });
-	GetScheduler().Flush();
-}
-
-void CommandProcessor::FlipWithInterrupt(uint32_t eop_event_type, uint32_t cache_action,
-                                         void* dst_gpu_addr, uint32_t value) {
-	auto& command = CurrentBuffer();
-
-	if (GraphicsRunDebugDumpEnabled()) {
-		LOGF("CommandProcessor::FlipWithInterrupt()\n"
-		     "\t eop_event_type      = 0x%08" PRIx32 "\n"
-		     "\t cache_action        = 0x%08" PRIx32 "\n"
-		     "\t dst_gpu_addr        = 0x%016" PRIx64 "\n"
-		     "\t value               = 0x%08" PRIx32 "\n",
-		     eop_event_type, cache_action, reinterpret_cast<uint64_t>(dst_gpu_addr), value);
-	}
-
-	if (eop_event_type != 0x00000004 || cache_action != 0x00000038) {
-		EXIT("unknown event type\n");
-	}
-	EXIT_IF(dst_gpu_addr == nullptr);
-	std::memcpy(dst_gpu_addr, &value, sizeof(value));
-	auto request = Sync::PrepareVideoOutFlip(command, m_flip.handle, m_flip.index, m_flip.flip_mode,
-	                                         m_flip.flip_arg);
-	(void)command.Handle();
-	GetScheduler().DeferPriorityOperation(
-	    [&renderer = m_renderer, event_id = m_interrupt_event_id, request] {
-		    renderer.GetVideoOut().CompleteFlip(request);
-		    renderer.TriggerInterrupt(event_id, 0);
-	    });
-	GetScheduler().Flush();
-}
-
 void CommandProcessor::PrepareCpuFlip(uint64_t request_id) {
 	auto& command = CurrentBuffer();
 	if (g_current_processor != nullptr) {
@@ -1383,10 +1145,6 @@ void CommandProcessor::PrepareCpuFlip(uint64_t request_id) {
 	GetScheduler().Flush();
 }
 
-void CommandProcessor::SynchronizeGpu() {
-	GetScheduler().Finish();
-}
-
 bool GuestGpu::IsGpuThread() noexcept {
 	return g_gpu_thread;
 }

src/graphics/guest_gpu/graphicsRun.h

@@ -31,10 +31,11 @@ public:
 	[[nodiscard]] bool IsStopping();
 	void               SendCommand(Common::UniqueFunction<void>&& command);
 	void               SendCommandSync(Common::UniqueFunction<void>&& command);
+	// Retries blocked submissions now instead of after the 100 ms stall timeout.
+	void               Wake();
 
 	// Submitted command memory is borrowed and must remain valid until GPU execution completes.
-	void Submit(std::span<const uint32_t> draw_commands,
-	            std::span<const uint32_t> constant_commands);
+	void Submit(std::span<const uint32_t> commands);
 	void SubmitCompute(uint32_t queue, std::span<const uint32_t> commands);
 	void SubmitFlipPreparation(uint64_t request_id);
 	// Insert an ordered graphics drain; only a previous suspend point can block the caller.
@@ -58,14 +59,10 @@ private:
 		SubmissionType            type     = SubmissionType::Graphics;
 		uint32_t                  queue_id = 0;
 		std::span<const uint32_t> commands;
-		std::span<const uint32_t> constant_commands;
-		Pm4Execution              command_execution;
-		Pm4Execution              constant_execution;
-		bool                      started           = false;
-		bool                      command_complete  = false;
-		bool                      constant_complete = false;
-		bool                      blocked           = false;
-		uint64_t                  flip_request_id   = 0;
+		Pm4Execution              execution;
+		bool                      started         = false;
+		bool                      blocked         = false;
+		uint64_t                  flip_request_id = 0;
 	};
 
 	void              Enqueue(Submission submission);
@@ -87,6 +84,7 @@ private:
 	bool                                           m_processing        = false;
 	bool                                           m_accepting         = true;
 	bool                                           m_stopping          = false;
+	bool                                           m_wake_pending      = false;
 	bool                                           m_shutdown_complete = false;
 	// Completion callbacks can outlive GuestGpu during renderer shutdown.
 	std::shared_ptr<std::binary_semaphore> m_suspend_point_ready =

src/graphics/guest_gpu/hardwareContext.h

@@ -597,14 +597,10 @@ struct ShaderRegisters {
 	}
 };
 
-enum class UserSgprType { Unknown, Region, Vsharp };
-
 struct UserSgprInfo {
 	static constexpr int SGPRS_MAX = 32;
 
-	uint32_t     value[SGPRS_MAX] = {0};
-	UserSgprType type[SGPRS_MAX]  = {};
-	uint32_t     count            = 0;
+	uint32_t value[SGPRS_MAX] = {0};
 };
 
 struct VertexShaderInfo {
@@ -1056,30 +1052,10 @@ public:
 	void SetCsShader(const CsStageRegisters& cs_regs) { m_cs.cs_regs = cs_regs; }
 	void SetCsWaveSize(uint8_t wave_size) { m_cs.cs_regs.wave_size = wave_size; }
 
-	void SetPsUserSgpr(uint32_t id, uint32_t value, UserSgprType type) {
-		m_ps.ps_user_sgpr.value[id] = value;
-		m_ps.ps_user_sgpr.type[id]  = type;
-		m_ps.ps_user_sgpr.count =
-		    ((id + 1) > m_ps.ps_user_sgpr.count ? (id + 1) : m_ps.ps_user_sgpr.count);
-	}
-	void SetCsUserSgpr(uint32_t id, uint32_t value, UserSgprType type) {
-		m_cs.cs_user_sgpr.value[id] = value;
-		m_cs.cs_user_sgpr.type[id]  = type;
-		m_cs.cs_user_sgpr.count =
-		    ((id + 1) > m_cs.cs_user_sgpr.count ? (id + 1) : m_cs.cs_user_sgpr.count);
-	}
-	void SetHsUserSgpr(uint32_t id, uint32_t value, UserSgprType type) {
-		m_vs.hs_user_sgpr.value[id] = value;
-		m_vs.hs_user_sgpr.type[id]  = type;
-		m_vs.hs_user_sgpr.count =
-		    ((id + 1) > m_vs.hs_user_sgpr.count ? (id + 1) : m_vs.hs_user_sgpr.count);
-	}
-	void SetGsUserSgpr(uint32_t id, uint32_t value, UserSgprType type) {
-		m_vs.gs_user_sgpr.value[id] = value;
-		m_vs.gs_user_sgpr.type[id]  = type;
-		m_vs.gs_user_sgpr.count =
-		    ((id + 1) > m_vs.gs_user_sgpr.count ? (id + 1) : m_vs.gs_user_sgpr.count);
-	}
+	void SetPsUserSgpr(uint32_t id, uint32_t value) { m_ps.ps_user_sgpr.value[id] = value; }
+	void SetCsUserSgpr(uint32_t id, uint32_t value) { m_cs.cs_user_sgpr.value[id] = value; }
+	void SetHsUserSgpr(uint32_t id, uint32_t value) { m_vs.hs_user_sgpr.value[id] = value; }
+	void SetGsUserSgpr(uint32_t id, uint32_t value) { m_vs.gs_user_sgpr.value[id] = value; }
 
 	[[nodiscard]] const PixelShaderInfo&   GetPs() const { return m_ps; }
 	[[nodiscard]] const VertexShaderInfo&  GetVs() const { return m_vs; }

src/graphics/guest_gpu/pm4.cpp

@@ -25,7 +25,6 @@ constexpr auto MakeRegisterNames() {
 	names[R_POP_MARKER]     = "R_POP_MARKER";
 	names[R_WRITE_DATA]     = "R_WRITE_DATA";
 	names[R_FLIP]           = "R_FLIP";
-	names[R_RELEASE_MEM]    = "R_RELEASE_MEM";
 	names[R_DMA_DATA]       = "R_DMA_DATA";
 	names[R_CONTEXT_STATE]  = "R_CONTEXT_STATE";
 	return names;
@@ -54,7 +53,6 @@ constexpr auto MakeOpcodeNames() {
 	names[IT_DRAW_INDIRECT_MULTI]       = "IT_DRAW_INDIRECT_MULTI";
 	names[IT_DRAW_INDEX_AUTO]           = "IT_DRAW_INDEX_AUTO";
 	names[IT_NUM_INSTANCES]             = "IT_NUM_INSTANCES";
-	names[IT_INDIRECT_BUFFER_CNST]      = "IT_INDIRECT_BUFFER_CNST";
 	names[IT_DRAW_INDEX_OFFSET_2]       = "IT_DRAW_INDEX_OFFSET_2";
 	names[IT_WRITE_DATA]                = "IT_WRITE_DATA";
 	names[IT_MEM_SEMAPHORE]             = "IT_MEM_SEMAPHORE";
@@ -66,8 +64,6 @@ constexpr auto MakeOpcodeNames() {
 	names[IT_PFP_SYNC_ME]               = "IT_PFP_SYNC_ME";
 	names[IT_SURFACE_SYNC]              = "IT_SURFACE_SYNC";
 	names[IT_EVENT_WRITE]               = "IT_EVENT_WRITE";
-	names[IT_EVENT_WRITE_EOP]           = "IT_EVENT_WRITE_EOP";
-	names[IT_EVENT_WRITE_EOS]           = "IT_EVENT_WRITE_EOS";
 	names[IT_RELEASE_MEM]               = "IT_RELEASE_MEM";
 	names[IT_DMA_DATA]                  = "IT_DMA_DATA";
 	names[IT_ACQUIRE_MEM]               = "IT_ACQUIRE_MEM";
@@ -80,12 +76,6 @@ constexpr auto MakeOpcodeNames() {
 	names[IT_SET_QUEUE_REG]             = "IT_SET_QUEUE_REG";
 	names[IT_SET_UCONFIG_REG]           = "IT_SET_UCONFIG_REG";
 	names[IT_SET_UCONFIG_REG_INDEX]     = "IT_SET_UCONFIG_REG_INDEX";
-	names[IT_WRITE_CONST_RAM]           = "IT_WRITE_CONST_RAM";
-	names[IT_DUMP_CONST_RAM]            = "IT_DUMP_CONST_RAM";
-	names[IT_INCREMENT_CE_COUNTER]      = "IT_INCREMENT_CE_COUNTER";
-	names[IT_INCREMENT_DE_COUNTER]      = "IT_INCREMENT_DE_COUNTER";
-	names[IT_WAIT_ON_CE_COUNTER]        = "IT_WAIT_ON_CE_COUNTER";
-	names[IT_WAIT_ON_DE_COUNTER_DIFF]   = "IT_WAIT_ON_DE_COUNTER_DIFF";
 	names[IT_DISPATCH_DRAW_PREAMBLE]    = "IT_DISPATCH_DRAW_PREAMBLE";
 	names[IT_DISPATCH_DRAW]             = "IT_DISPATCH_DRAW";
 	names[IT_GET_LOD_STATS]             = "IT_GET_LOD_STATS";

src/graphics/guest_gpu/pm4.h

@@ -40,7 +40,6 @@ constexpr uint32_t IT_INDEX_TYPE                = 0x2A;
 constexpr uint32_t IT_DRAW_INDIRECT_MULTI       = 0x2C;
 constexpr uint32_t IT_DRAW_INDEX_AUTO           = 0x2D;
 constexpr uint32_t IT_NUM_INSTANCES             = 0x2F;
-constexpr uint32_t IT_INDIRECT_BUFFER_CNST      = 0x33;
 constexpr uint32_t IT_DRAW_INDEX_OFFSET_2       = 0x35;
 constexpr uint32_t IT_WRITE_DATA                = 0x37;
 constexpr uint32_t IT_MEM_SEMAPHORE             = 0x39;
@@ -53,8 +52,6 @@ constexpr uint32_t IT_CP_DMA                    = 0x41;
 constexpr uint32_t IT_PFP_SYNC_ME               = 0x42;
 constexpr uint32_t IT_SURFACE_SYNC              = 0x43;
 constexpr uint32_t IT_EVENT_WRITE               = 0x46;
-constexpr uint32_t IT_EVENT_WRITE_EOP           = 0x47;
-constexpr uint32_t IT_EVENT_WRITE_EOS           = 0x48;
 constexpr uint32_t IT_RELEASE_MEM               = 0x49;
 constexpr uint32_t IT_DMA_DATA                  = 0x50;
 constexpr uint32_t IT_ACQUIRE_MEM               = 0x58;
@@ -67,12 +64,6 @@ constexpr uint32_t IT_SET_SH_REG                = 0x76;
 constexpr uint32_t IT_SET_QUEUE_REG             = 0x78;
 constexpr uint32_t IT_SET_UCONFIG_REG           = 0x79;
 constexpr uint32_t IT_SET_UCONFIG_REG_INDEX     = 0x7A;
-constexpr uint32_t IT_WRITE_CONST_RAM           = 0x81;
-constexpr uint32_t IT_DUMP_CONST_RAM            = 0x83;
-constexpr uint32_t IT_INCREMENT_CE_COUNTER      = 0x84;
-constexpr uint32_t IT_INCREMENT_DE_COUNTER      = 0x85;
-constexpr uint32_t IT_WAIT_ON_CE_COUNTER        = 0x86;
-constexpr uint32_t IT_WAIT_ON_DE_COUNTER_DIFF   = 0x88;
 constexpr uint32_t IT_DISPATCH_DRAW             = 0x8D;
 constexpr uint32_t IT_GET_LOD_STATS             = 0x8E;
 constexpr uint32_t IT_WAIT_REG_MEM_64           = 0x93;
@@ -88,7 +79,6 @@ constexpr uint32_t R_PUSH_MARKER    = 0x0B;
 constexpr uint32_t R_POP_MARKER     = 0x0C;
 constexpr uint32_t R_WRITE_DATA     = 0x15;
 constexpr uint32_t R_FLIP           = 0x17;
-constexpr uint32_t R_RELEASE_MEM    = 0x18;
 constexpr uint32_t R_DMA_DATA       = 0x19;
 constexpr uint32_t R_CONTEXT_STATE  = 0x1A;
 

src/graphics/host_gpu/graphicContext.h

@@ -35,6 +35,9 @@ struct GraphicContext {
 	bool                               attachment_feedback_loop_enabled      = false;
 	bool                               provoking_vertex_last_enabled         = false;
 	bool                               maintenance8_enabled                  = false;
+	bool                               conditional_rendering_enabled         = false;
+	bool                               occlusion_queries_enabled             = false;
+	bool                               occlusion_query_precise_enabled       = false;
 	bool                               supports_block_texel_view              = false;
 	bool                                      mesh_shader_enabled                   = false;
 	vk::PhysicalDeviceMeshShaderPropertiesEXT mesh_shader_properties                = {};
@@ -42,6 +45,8 @@ struct GraphicContext {
 	uint32_t                           min_subgroup_size                     = 0;
 	uint32_t                           max_subgroup_size                     = 0;
 	uint32_t                           max_push_descriptors                  = 0;
+	uint32_t                           device_clock_period_ps                = 0;
+	vk::DescriptorSetLayout            image_heap_layout                     = nullptr;
 	vk::ShaderStageFlags               required_subgroup_size_stages         = {};
 	Common::Mutex                      queue_mutex;
 	uint32_t                           queue_family = static_cast<uint32_t>(-1);

src/graphics/host_gpu/memoryTracker.cpp

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

src/graphics/host_gpu/memoryTracker.h

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

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

@@ -127,7 +127,7 @@ void BufferCache::ChangeRegister(BufferId id) {
 		}
 		WriteDataBuffer(m_bda_pagetable_buffer, table_offset,
 		                addresses.data(), addresses.size() * sizeof(vk::DeviceAddress));
-		m_memory_tracker.QueueBdaSync(buffer.CpuAddress(), buffer.Size());
+		QueueBdaSync(buffer.CpuAddress(), buffer.Size());
 	} else {
 		const auto found = m_buffers.find(buffer.CpuAddress());
 		EXIT_IF(found == m_buffers.end() || found->second != id);
@@ -174,6 +174,7 @@ bool BufferCache::DownloadBufferMemory(Buffer& buffer, uint64_t vaddr, uint64_t
 			    total_size += Common::AlignUp(end - start, 64);
 		    });
 		    m_gpu_modified_ranges.Subtract(address, bytes);
+		    ForgetColorMetadataChecks(address, bytes);
 	    });
 	if (copies.empty()) {
 		return false;
@@ -298,6 +299,7 @@ void BufferCache::InvalidateMemory(uint64_t vaddr, uint64_t size) {
 	}
 	m_memory_tracker.InvalidateRegion(vaddr, size,
 	                                  [this, vaddr, size] { ReadMemory(vaddr, size, true); });
+	QueueBdaSync(vaddr, size);
 }
 
 void BufferCache::ReadMemory(uint64_t vaddr, uint64_t size, bool is_write) {
@@ -317,6 +319,7 @@ void BufferCache::ReadMemory(uint64_t vaddr, uint64_t size, bool is_write) {
 		}
 		if (is_write) {
 			m_memory_tracker.MarkRegionAsCpuModified(vaddr, size);
+			QueueBdaSync(vaddr, size);
 		}
 	});
 }
@@ -519,6 +522,7 @@ std::pair<Buffer*, uint64_t> BufferCache::ObtainBuffer(uint64_t vaddr, uint64_t
 	(void)SynchronizeBuffer(buffer, vaddr, size, is_written, is_texel_buffer);
 	if (is_written) {
 		m_gpu_modified_ranges.Add(vaddr, size);
+		ForgetColorMetadataChecks(vaddr, size);
 	}
 	return {&buffer, buffer.Offset(vaddr)};
 }
@@ -623,6 +627,29 @@ bool BufferCache::IsRegionRegistered(uint64_t vaddr, uint64_t size) {
 	return address + m_slot_buffers[id].Size() > vaddr;
 }
 
+void BufferCache::MarkColorMetadataChecked(uint64_t vaddr, uint64_t size, uint32_t keys) {
+	for (uint32_t slot = 0; slot < m_checked_color_metadata.size(); slot++) {
+		if ((keys & (1u << slot)) != 0) {
+			m_checked_color_metadata[slot].Add(vaddr, size);
+		}
+	}
+}
+
+bool BufferCache::IsColorMetadataChecked(uint64_t vaddr, uint64_t size, uint32_t keys) const {
+	for (uint32_t slot = 0; slot < m_checked_color_metadata.size(); slot++) {
+		if ((keys & (1u << slot)) != 0 && !m_checked_color_metadata[slot].Contains(vaddr, size)) {
+			return false;
+		}
+	}
+	return true;
+}
+
+void BufferCache::ForgetColorMetadataChecks(uint64_t vaddr, uint64_t size) {
+	for (auto& checked: m_checked_color_metadata) {
+		checked.Subtract(vaddr, size);
+	}
+}
+
 bool BufferCache::IsRegionGpuModified(uint64_t vaddr, uint64_t size) {
 	return m_memory_tracker.IsRegionGpuModified(vaddr, size);
 }
@@ -694,14 +721,16 @@ void BufferCache::ProcessFaultBuffer() {
 	m_fault_manager.ProcessFaultBuffer();
 }
 
-void BufferCache::SynchronizeBda(const RangeSet& mapped) {
-	// Take before walking: writes queued during synchronization remain pending.
-	const auto pending = m_memory_tracker.TakeBdaSync();
-	pending.ForEach([&](uint64_t begin, uint64_t end) {
-		mapped.ForEachInRange(begin, end - begin, [this](uint64_t start, uint64_t finish) {
-			SynchronizeBuffersInRange(start, finish - start);
-		});
-	});
+void BufferCache::QueueBdaSync(uint64_t vaddr, uint64_t size) {
+	const auto begin = Common::AlignDown(vaddr, TRACKER_PAGE_SIZE);
+	const auto end   = Common::AlignUp(vaddr + size, TRACKER_PAGE_SIZE);
+	std::lock_guard lock(m_bda_sync_mutex);
+	m_bda_sync.Add(begin, end - begin);
+}
+
+RangeSet BufferCache::TakeBdaSync() {
+	std::lock_guard lock(m_bda_sync_mutex);
+	return std::exchange(m_bda_sync, {});
 }
 
 void BufferCache::SynchronizeBuffersInRange(uint64_t vaddr, uint64_t size) {

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

@@ -11,7 +11,9 @@
 #include "graphics/host_gpu/renderer/cache/multiLevelPageTable.h"
 #include "graphics/host_gpu/renderer/cache/streamBuffer.h"
 
+#include <array>
 #include <map>
+#include <mutex>
 #include <span>
 #include <utility>
 #include <vector>
@@ -79,20 +81,21 @@ public:
 	[[nodiscard]] bool HasGpuDirtyBytes(uint64_t vaddr, uint64_t size);
 	[[nodiscard]] bool IsRegionCpuModified(uint64_t vaddr, uint64_t size);
 	[[nodiscard]] bool IsRegionGpuModified(uint64_t vaddr, uint64_t size);
+	// Color metadata the GPU found free of the clear keys in `keys` (bits of code >> 5) stays so
+	// until a GPU write or download.
+	void               MarkColorMetadataChecked(uint64_t vaddr, uint64_t size, uint32_t keys);
+	[[nodiscard]] bool IsColorMetadataChecked(uint64_t vaddr, uint64_t size, uint32_t keys) const;
 	void               ProcessFaultBuffer();
-	void               SynchronizeBda(const RangeSet& mapped);
-	void               QueueBdaSync(uint64_t vaddr, uint64_t size) {
-		m_memory_tracker.QueueBdaSync(vaddr, size);
-	}
+	void               SynchronizeBuffersInRange(uint64_t vaddr, uint64_t size);
+	// Queue after the range's state change becomes visible, or a drain could consume it first.
+	void                   QueueBdaSync(uint64_t vaddr, uint64_t size);
+	[[nodiscard]] RangeSet TakeBdaSync();
 	void               RunGarbageCollector();
 
 private:
 	friend struct BufferCacheTestAccess;
 
-	bool IsBufferInvalid(BufferId id) const {
-		const auto* buffer = m_slot_buffers.try_get(id);
-		return buffer == nullptr || buffer->is_deleted;
-	}
+	bool IsBufferInvalid(BufferId id) const { return !id || m_slot_buffers[id].is_deleted; }
 
 	using BufferMap = std::map<uint64_t, BufferId>;
 	struct OverlapResult {
@@ -117,13 +120,13 @@ private:
 	void DeleteBuffer(BufferId id);
 	[[nodiscard]] bool SynchronizeBuffer(Buffer& buffer, uint64_t vaddr, uint64_t size,
 	                                     bool is_written, bool is_texel_buffer);
-	void SynchronizeBuffersInRange(uint64_t vaddr, uint64_t size);
 	[[nodiscard]] vk::Buffer UploadCopies(Buffer& buffer, std::span<vk::BufferCopy> copies,
 	                                      uint64_t total_size);
 	[[nodiscard]] bool SynchronizeBufferFromImage(Buffer& buffer, uint64_t vaddr, uint64_t size);
 	// Synchronous downloads publish before returning; asynchronous callers wait before reuse.
 	template <bool async>
 	[[nodiscard]] bool DownloadBufferMemory(Buffer& buffer, uint64_t vaddr, uint64_t size);
+	void               ForgetColorMetadataChecks(uint64_t vaddr, uint64_t size);
 
 	GraphicContext&                                   m_graphics;
 	CommandScheduler&                                 m_scheduler;
@@ -135,7 +138,11 @@ private:
 	BufferMap                                         m_buffers;
 	PageTable                                         m_page_table;
 	RangeSet                                          m_gpu_modified_ranges;
+	// One per TileManager::ColorClearSlots key slot.
+	std::array<RangeSet, 8>                           m_checked_color_metadata;
 	MemoryTracker                                     m_memory_tracker;
+	std::mutex                                        m_bda_sync_mutex;
+	RangeSet                                          m_bda_sync;
 	StreamBuffer                                      m_staging_buffer;
 	StreamBuffer                                      m_stream_buffer;
 	StreamBuffer                                      m_download_buffer;

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

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

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

@@ -229,6 +229,7 @@ void TextureCache::UnregisterImage(ImageId id) {
 	if (!image.registered) {
 		return;
 	}
+	image.HeapChanged();
 	UntrackImage(id);
 	ImagePageTable::PageRange pages {};
 	if (!ImagePageTable::TryGetPageRange(image.info.data.address, image.info.data.size, pages)) {
@@ -250,24 +251,24 @@ void TextureCache::UnregisterImage(ImageId id) {
 }
 
 void TextureCache::DeleteImage(ImageId id) {
-	auto* image = m_slot_images.try_get(id);
-	if (image == nullptr || !image->registered) {
+	auto& image = m_slot_images[id];
+	if (!image.registered) {
 		return;
 	}
-	if (image->depth_id) {
-		std::erase(m_slot_images[image->depth_id].stencil_ids, id);
+	if (image.depth_id) {
+		std::erase(m_slot_images[image.depth_id].stencil_ids, id);
 	} else {
 		// Detach first: deleting each association also unlinks it from its depth owner.
-		for (const auto association: std::exchange(image->stencil_ids, {})) {
+		for (const auto association: std::exchange(image.stencil_ids, {})) {
 			FreeImage(association);
 		}
 	}
-	if (image->IsGpuModified()) {
+	if (image.IsGpuModified()) {
 		EXIT("TextureCache: deleting a GPU-modified image without resolving its contents\n");
 	}
 	m_download_images.erase(id);
-	if (image->info.metadata.kind == ImageMetadataKind::Htile) {
-		m_surface_metas.erase(image->info.metadata.range.address);
+	if (image.info.metadata.kind == ImageMetadataKind::Htile) {
+		m_surface_metas.erase(image.info.metadata.range.address);
 	}
 	UnregisterImage(id);
 	if (m_scheduler.Active()) {
@@ -378,6 +379,7 @@ void TextureCache::UntrackImageHead(ImageId id) {
 	const auto address = Common::AlignDown(begin + TRACKER_PAGE_SIZE, TRACKER_PAGE_SIZE);
 	const auto size    = address - begin;
 	image.track_addr   = address;
+	image.HeapChanged();
 	if (image.track_addr == image.track_addr_end) {
 		MarkAsMaybeDirty(id, image);
 	}
@@ -395,6 +397,7 @@ void TextureCache::UntrackImageTail(ImageId id) {
 	const auto address   = Common::AlignDown(end, TRACKER_PAGE_SIZE);
 	const auto size      = end - address;
 	image.track_addr_end = address;
+	image.HeapChanged();
 	if (image.track_addr == image.track_addr_end) {
 		MarkAsMaybeDirty(id, image);
 	}
@@ -426,12 +429,6 @@ TextureCache::ImageIds TextureCache::FindImagesInRegion(uint64_t address, uint64
 		return {};
 	}
 
-	uint32_t query_epoch = ++m_image_query_epoch;
-	if (query_epoch == 0) {
-		m_slot_images.ForEach([](ImageId, const Image& image) { image.query_epoch = 0; });
-		query_epoch = ++m_image_query_epoch;
-	}
-
 	ImageIds result;
 	ForEachPage(address, size, [&](uint64_t page) {
 		const auto* owners = m_image_page_table.Find(page);
@@ -439,19 +436,17 @@ TextureCache::ImageIds TextureCache::FindImagesInRegion(uint64_t address, uint64
 			return;
 		}
 		owners->ForEach([&](ImageId id) {
-			auto* image = m_slot_images.try_get(id);
-			if (image == nullptr) {
+			const auto& image = m_slot_images[id];
+			if (image.picked || !image.Overlaps(address, size, page_overlap)) {
 				return;
 			}
-			if (image->query_epoch == query_epoch) {
-				return;
-			}
-			image->query_epoch = query_epoch;
-			if (image->Overlaps(address, size, page_overlap)) {
-				result.push_back(id);
-			}
+			image.picked = true;
+			result.push_back(id);
 		});
 	});
+	for (const auto id: result) {
+		m_slot_images[id].picked = false;
+	}
 	return result;
 }
 
@@ -493,16 +488,15 @@ void TextureCache::RefreshCopySource(ImageId id) {
 }
 
 bool TextureCache::CopyD16(Image& destination, Image& source) {
-	const bool source_depth      = source.info.IsDepth();
-	const bool destination_depth = destination.info.IsDepth();
-	if (source_depth == destination_depth) {
+	// Converts guest D16 between a four-byte host depth aspect and two-byte color or D16 depth.
+	const bool source_wide = DepthAspectTransferBytes(source.backing.format) == sizeof(uint32_t);
+	if (source_wide == (DepthAspectTransferBytes(destination.backing.format) == sizeof(uint32_t))) {
 		return false;
 	}
-	auto&      depth          = source_depth ? source : destination;
-	auto&      color          = source_depth ? destination : source;
-	const auto transfer_bytes = DepthAspectTransferBytes(depth.backing.format);
+	auto& depth  = source_wide ? source : destination;
+	auto& narrow = source_wide ? destination : source;
 	if (depth.info.bytes_per_block != sizeof(uint16_t) ||
-	    color.info.bytes_per_block != sizeof(uint16_t) || transfer_bytes != sizeof(uint32_t)) {
+	    narrow.info.bytes_per_block != sizeof(uint16_t)) {
 		return false;
 	}
 	EXIT_IF(source.backing.samples != 1 || destination.backing.samples != 1 ||
@@ -512,42 +506,42 @@ bool TextureCache::CopyD16(Image& destination, Image& source) {
 
 	const auto     layers = depth.info.resources.layers;
 	const uint64_t depth_slice =
-	    static_cast<uint64_t>(depth.info.pitch) * depth.info.extent.height * transfer_bytes;
-	const uint64_t color_slice =
-	    static_cast<uint64_t>(color.info.pitch) * color.info.extent.height * sizeof(uint16_t);
-	EXIT_IF(layers == 0 || depth_slice > UINT64_MAX / layers || color_slice > UINT64_MAX / layers);
-	const auto                       depth_size = depth_slice * layers;
-	const auto                       color_size = color_slice * layers;
-	const std::array                 depth_copies {
+	    static_cast<uint64_t>(depth.info.pitch) * depth.info.extent.height * sizeof(uint32_t);
+	const uint64_t narrow_slice =
+	    static_cast<uint64_t>(narrow.info.pitch) * narrow.info.extent.height * sizeof(uint16_t);
+	EXIT_IF(layers == 0 || depth_slice > UINT64_MAX / layers || narrow_slice > UINT64_MAX / layers);
+	const std::array depth_copies {
 	    vk::BufferImageCopy {0,
 	                         depth.info.pitch,
 	                         depth.info.extent.height,
 	                         {vk::ImageAspectFlagBits::eDepth, 0, 0, layers},
 	                         {},
 	                         depth.info.extent}};
-	const std::array color_copies {
+	const std::array narrow_copies {
 	    vk::BufferImageCopy {0,
-	                         color.info.pitch,
-	                         color.info.extent.height,
-	                         {vk::ImageAspectFlagBits::eColor, 0, 0, layers},
+	                         narrow.info.pitch,
+	                         narrow.info.extent.height,
+	                         {narrow.info.IsDepth() ? vk::ImageAspectFlagBits::eDepth
+	                                                : vk::ImageAspectFlagBits::eColor,
+	                          0, 0, layers},
 	                         {},
-	                         color.info.extent}};
+	                         narrow.info.extent}};
 
-	auto                         depth_buffer = m_tiler.GetScratchBuffer(depth_size);
-	auto                         color_buffer = m_tiler.GetScratchBuffer(color_size, depth_buffer.buffer);
+	auto depth_buffer  = m_tiler.GetScratchBuffer(depth_slice * layers);
+	auto narrow_buffer = m_tiler.GetScratchBuffer(narrow_slice * layers, depth_buffer.buffer);
 	const TileManager::D16Layout promote_layout {
 	    .width               = depth.info.extent.width,
 	    .height              = depth.info.extent.height,
 	    .layers              = layers,
-	    .source_row_stride   = static_cast<uint64_t>(color.info.pitch) * sizeof(uint16_t),
-	    .target_row_stride   = static_cast<uint64_t>(depth.info.pitch) * transfer_bytes,
-	    .source_slice_stride = color_slice,
+	    .source_row_stride   = static_cast<uint64_t>(narrow.info.pitch) * sizeof(uint16_t),
+	    .target_row_stride   = static_cast<uint64_t>(depth.info.pitch) * sizeof(uint32_t),
+	    .source_slice_stride = narrow_slice,
 	    .target_slice_stride = depth_slice,
 	};
 	const bool d32 = DepthAspectTransferFormat(depth.backing.format) == vk::Format::eD32Sfloat;
-	if (source_depth) {
+	if (source_wide) {
 		source.Download(depth_copies, depth_buffer.buffer, depth_buffer.offset, depth_buffer.size);
-		m_tiler.ConvertD16(depth_buffer, color_buffer, TileManager::D16Direction::Demote, d32,
+		m_tiler.ConvertD16(depth_buffer, narrow_buffer, TileManager::D16Direction::Demote, d32,
 		                   {.width               = promote_layout.width,
 		                    .height              = promote_layout.height,
 		                    .layers              = promote_layout.layers,
@@ -555,11 +549,12 @@ bool TextureCache::CopyD16(Image& destination, Image& source) {
 		                    .target_row_stride   = promote_layout.source_row_stride,
 		                    .source_slice_stride = promote_layout.target_slice_stride,
 		                    .target_slice_stride = promote_layout.source_slice_stride});
-		destination.Upload(color_copies, color_buffer.buffer, color_buffer.offset,
-		                   color_buffer.size);
+		destination.Upload(narrow_copies, narrow_buffer.buffer, narrow_buffer.offset,
+		                   narrow_buffer.size);
 	} else {
-		source.Download(color_copies, color_buffer.buffer, color_buffer.offset, color_buffer.size);
-		m_tiler.ConvertD16(color_buffer, depth_buffer, TileManager::D16Direction::Promote, d32,
+		source.Download(narrow_copies, narrow_buffer.buffer, narrow_buffer.offset,
+		                narrow_buffer.size);
+		m_tiler.ConvertD16(narrow_buffer, depth_buffer, TileManager::D16Direction::Promote, d32,
 		                   promote_layout);
 		destination.Upload(depth_copies, depth_buffer.buffer, depth_buffer.offset,
 		                   depth_buffer.size);
@@ -784,18 +779,26 @@ void Resolve(TextureCache& cache, ImageId cached_id, ImageId merged_id) {
 TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& requested,
                                                          BindingType binding, ImageId cached_id,
                                                          ImageId merged_id) {
-	auto owner = m_slot_images.try_get(cached_id);
-	if (owner == nullptr || !owner->registered) {
+	auto& cached = m_slot_images[cached_id];
+	if (!cached.registered) {
 		return {merged_id};
 	}
-	if (BcAlias::Matches(requested, owner->info) && !owner->IsCpuDirty()) {
+	// PS5: image heaps keep BC images and raw block views of their layers resident together, so a
+	// raw read of a layer subset with the same levels views the BC image instead of evicting it.
+	const bool raw_view = binding == BindingType::Texture && !requested.IsBlock() &&
+	                      requested.type == cached.info.type &&
+	                      requested.data.address == cached.info.data.address &&
+	                      requested.BlockExtent() == cached.info.BlockExtent() &&
+	                      requested.resources.levels == cached.info.resources.levels &&
+	                      requested.resources.layers < cached.info.resources.layers;
+	if (!raw_view && BcAlias::Matches(requested, cached.info) && !cached.IsCpuDirty()) {
 		BcAlias::Resolve(*this, cached_id, merged_id);
 		return {merged_id};
 	}
-	auto&      cached       = *owner;
 	const auto current_tick = m_scheduler.CurrentTick();
 	const bool safe_to_delete =
-	    current_tick - std::min(current_tick, cached.tick_accessed_last) > NumFramesBeforeRemoval;
+	    current_tick - std::min(current_tick, cached.tick_accessed_last) > NumFramesBeforeRemoval &&
+	    cached.heap_refs.load(std::memory_order_relaxed) == 0u;
 
 	const uint32_t requested_block = requested.bytes_per_block * requested.samples;
 	const uint32_t cached_block    = cached.info.bytes_per_block * cached.info.samples;
@@ -865,9 +868,8 @@ TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& reques
 		if (requested.pixel_format != cached.info.pixel_format ||
 		    requested.data.size <= cached.info.data.size) {
 			const auto result_id = merged_id ? merged_id : cached_id;
-			const auto result    = m_slot_images.try_get(result_id);
-			return {result != nullptr && ImageViewOps::FormatsCompatible(result->info.pixel_format,
-			                                                             requested.pixel_format)
+			return {ImageViewOps::FormatsCompatible(m_slot_images[result_id].info.pixel_format,
+			                                        requested.pixel_format)
 			            ? result_id
 			            : ImageId {}};
 		}
@@ -1067,12 +1069,78 @@ void TextureCache::MaterializeColorClear(ImageId id, const ImageDesc& desc,
 	if (first >= layers || count > layers - first) {
 		EXIT("TextureCache: color view exceeds its native metadata slices\n");
 	}
-	// Finish native metadata writes before reading backing bytes. This can submit the scheduler,
-	// so discovery runs before final draw uploads and never holds the texture lock across it.
-	if (m_buffer_cache.IsRegionGpuModified(range.address, range.size)) {
-		m_buffer_cache.ReadMemory(range.address, range.size, false);
+	const auto       slice_size = range.size / layers;
+	const GuestRange viewed {range.address + slice_size * first, slice_size * count};
+	if (m_buffer_cache.IsRegionGpuModified(viewed.address, viewed.size)) {
+		std::array<vk::ClearAttachment, TileManager::ColorClearSlots> clears {};
+		uint32_t keys = 0;
+		for (uint32_t key = 0; key < clears.size(); key++) {
+			clears[key].aspectMask = vk::ImageAspectFlagBits::eColor;
+			if (DecodeColorClear(desc, static_cast<uint8_t>(key << 5u),
+			                     clears[key].clearValue.color)) {
+				keys |= 1u << key;
+			}
+		}
+		// No selected key to clear, or none left since the last check.
+		if (m_buffer_cache.IsColorMetadataChecked(viewed.address, viewed.size, keys)) {
+			return;
+		}
+		{
+			std::scoped_lock lock {m_lock};
+			auto&            image = m_slot_images[id];
+			// Compressed images cannot upload, so they must already hold native contents.
+			const bool native_current = (image.usage.render_target || image.IsGpuModified()) &&
+			                            !image.IsBufferModified() && !image.IsCpuDirty();
+			// 2D-array attachment views select 2D layers or, of a volume, depth slices.
+			if (m_graphics.conditional_rendering_enabled &&
+			    image.backing.image_type != vk::ImageType::e1D &&
+			    count <= m_graphics.GetPhysicalDeviceProperties().limits.maxFramebufferLayers &&
+			    (image.backing.usage & vk::ImageUsageFlagBits::eColorAttachment) &&
+			    (m_graphics.GetFormatProperties(view.format).optimalTilingFeatures &
+			     vk::FormatFeatureFlagBits::eColorAttachment) &&
+			    (desc.info.metadata.compression == VideoOutCompression::Uncompressed ||
+			     native_current)) {
+				// The GPU selects each slice's clear, so the CPU never waits for the metadata.
+				Common::DebugCounters::Add(Common::DebugCounters::Counter::ColorMetadataChecks);
+				InitializeImage(id);
+				const bool expand = desc.type != BindingType::VideoOut;
+				const auto [buffer, offset] =
+				    m_buffer_cache.ObtainBuffer(viewed.address, viewed.size, expand);
+				const auto predicates = m_tiler.CheckColorClear(
+				    {buffer->Handle(), offset, viewed.size}, count, keys, expand);
+				if (expand) {
+					m_buffer_cache.MarkColorMetadataChecked(viewed.address, viewed.size, keys);
+				}
+				const auto command = m_scheduler.Current().Handle();
+				const auto area    = BeginColorRendering(image, view.format, view.base_level,
+				                                         image_first, count, nullptr, command);
+				vk::ClearRect rect {.rect = area, .layerCount = 1};
+				for (uint32_t slice = 0; slice < count; slice++) {
+					rect.baseArrayLayer = slice;
+					for (uint32_t key = 0; key < clears.size(); key++) {
+						if ((keys & (1u << key)) != 0) {
+							const vk::ConditionalRenderingBeginInfoEXT predicate {
+							    .buffer = predicates,
+							    .offset = (slice * clears.size() + key) * sizeof(uint32_t)};
+							command.beginConditionalRenderingEXT(&predicate);
+							command.clearAttachments(1, &clears[key], 1, &rect);
+							command.endConditionalRenderingEXT();
+						}
+					}
+				}
+				command.endRendering();
+				CommitGpuWrite(image);
+				return;
+			}
+		}
+		// Finish native metadata writes before reading backing bytes. This can submit the
+		// scheduler, so discovery runs before final draw uploads and never holds the texture
+		// lock across it.
+		Common::DebugCounters::ScopedTimer timer(
+		    Common::DebugCounters::Counter::ColorMetadataReadbacks,
+		    Common::DebugCounters::Counter::ColorMetadataReadbackNs);
+		m_buffer_cache.ReadMemory(viewed.address, viewed.size, false);
 	}
-	const auto slice_size = range.size / layers;
 	for (uint32_t slice = 0; slice < count; slice++) {
 		const auto address = range.address + slice_size * (first + slice);
 		uint8_t code = 0;
@@ -1148,9 +1216,8 @@ ImageId TextureCache::AssociateStencil(ImageId depth_id, GuestRange stencil) {
 
 	ImageId association {};
 	for (const auto id: FindImagesInRegion(stencil.address, stencil.size, false)) {
-		const auto owner = m_slot_images.try_get(id);
-		if (owner != nullptr && owner->info.data == stencil &&
-		    owner->info.extent == depth.info.extent) {
+		const auto& owner = m_slot_images[id];
+		if (owner.info.data == stencil && owner.info.extent == depth.info.extent) {
 			association = id;
 		}
 	}
@@ -1243,12 +1310,24 @@ ImageId TextureCache::FindImage(ImageDesc& desc, bool exact_format) {
 					view_layer = overlap.layer;
 				}
 			}
+			// PS5: image heaps hold a descriptor per dynamic-resolution extent of a render target,
+			// and only one extent at an address holds data, so heap lookups create no second one.
+			if (!result && desc.heap &&
+			    std::any_of(candidates.begin(), candidates.end(), [&](ImageId id) {
+				    const auto* image = m_slot_images.try_get(id);
+				    return image != nullptr && image->registered &&
+					       image->heap_refs.load(std::memory_order_relaxed) != 0u &&
+					       image->info.data.address == desc.info.data.address &&
+					       image->info.type == desc.info.type &&
+					       image->info.extent != desc.info.extent;
+			    })) {
+				return {};
+			}
 		} else if (block_alias) {
 			for (const auto candidate: candidates) {
-				const auto* owner = m_slot_images.try_get(candidate);
-				if (candidate != result && owner != nullptr && owner->registered &&
-				    !owner->IsCpuDirty() &&
-				    BcAlias::Matches(m_slot_images[result].info, owner->info)) {
+				const auto& owner = m_slot_images[candidate];
+				if (candidate != result && owner.registered && !owner.IsCpuDirty() &&
+				    BcAlias::Matches(m_slot_images[result].info, owner.info)) {
 					BcAlias::Resolve(*this, candidate, result);
 				}
 			}
@@ -1311,14 +1390,14 @@ ImageId TextureCache::FindImageFromRange(uint64_t address, uint64_t size, bool e
 	std::scoped_lock lock {m_lock};
 	ImageIds         matches;
 	for (const auto id: FindImagesInRegion(address, size, false)) {
-		auto owner = m_slot_images.try_get(id);
-		if (owner == nullptr || owner->info.data.address != address) {
+		const auto* owner = &m_slot_images[id];
+		if (owner->info.data.address != address) {
 			continue;
 		}
 		if (ensure_valid && owner->depth_id) {
-			owner = m_slot_images.try_get(owner->depth_id);
+			owner = &m_slot_images[owner->depth_id];
 		}
-		if (owner == nullptr || (ensure_valid && !owner->SafeToDownload())) {
+		if (ensure_valid && !owner->SafeToDownload()) {
 			continue;
 		}
 		matches.push_back(id);
@@ -1335,11 +1414,8 @@ ImageId TextureCache::FindImageFromRange(uint64_t address, uint64_t size, bool e
 			}
 		}
 	}
-	if (selected && ensure_valid) {
-		const auto owner = m_slot_images.try_get(selected);
-		if (owner != nullptr && owner->depth_id) {
-			selected = owner->depth_id;
-		}
+	if (selected && ensure_valid && m_slot_images[selected].depth_id) {
+		selected = m_slot_images[selected].depth_id;
 	}
 	return selected;
 }
@@ -1360,9 +1436,8 @@ vk::ImageView TextureCache::FindTexture(ImageId id, const ImageDesc& desc) {
 	    desc.info.data.End() <= image.info.stencil.End()) {
 		for (const auto stencil_id:
 		     FindImagesInRegion(image.info.stencil.address, image.info.stencil.size, false)) {
-			const auto* stencil = m_slot_images.try_get(stencil_id);
-			if (stencil != nullptr && stencil->depth_id == id &&
-			    stencil->info.data == image.info.stencil) {
+			const auto& stencil = m_slot_images[stencil_id];
+			if (stencil.depth_id == id && stencil.info.data == image.info.stencil) {
 				RefreshImage(stencil_id);
 				break;
 			}
@@ -1473,27 +1548,24 @@ bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address
 	vk::ImageAspectFlags  aspect {};
 	ImageSubresourceRange subresources;
 	for (const auto id: FindImagesInRegion(address, size, false)) {
-		auto owner = m_slot_images.try_get(id);
-		if (owner == nullptr) {
-			continue;
-		}
+		const auto&          owner = m_slot_images[id];
 		vk::ImageAspectFlags candidate {};
 		ImageId              candidate_id = id;
 		ImageSubresourceRange candidate_subresources {
-		    0, owner->info.resources.levels, 0, owner->info.TransferLayers()};
-		if (owner->depth_id && owner->info.data.address == address &&
-		    owner->info.data.size == size) {
+		    0, owner.info.resources.levels, 0, owner.info.TransferLayers()};
+		if (owner.depth_id && owner.info.data.address == address &&
+		    owner.info.data.size == size) {
 			candidate              = vk::ImageAspectFlagBits::eStencil;
-			candidate_id           = owner->depth_id;
-			candidate_subresources = owner->stencil_subresources;
-			owner        = m_slot_images.try_get(candidate_id);
-			if (owner == nullptr || owner->backing.image == nullptr || !owner->info.HasStencil()) {
+			candidate_id           = owner.depth_id;
+			candidate_subresources = owner.stencil_subresources;
+			const auto& depth      = m_slot_images[candidate_id];
+			if (depth.backing.image == nullptr || !depth.info.HasStencil()) {
 				continue;
 			}
-		} else if (!owner->depth_id && owner->info.data.address == address &&
-		           owner->info.data.size == size) {
-			candidate = owner->info.IsDepth() ? vk::ImageAspectFlagBits::eDepth
-			                                  : vk::ImageAspectFlagBits::eColor;
+		} else if (!owner.depth_id && owner.info.data.address == address &&
+		           owner.info.data.size == size) {
+			candidate = owner.info.IsDepth() ? vk::ImageAspectFlagBits::eDepth
+			                                 : vk::ImageAspectFlagBits::eColor;
 		}
 		if (!candidate) {
 			continue;
@@ -1574,29 +1646,8 @@ void TextureCache::ClearImage(CommandBuffer& command, ImageId id, vk::Format for
 	if (format != image.backing.format || (image.info.IsVolume() && !full_image)) {
 		EXIT_NOT_IMPLEMENTED(range.aspectMask != vk::ImageAspectFlagBits::eColor ||
 		                     range.levelCount != 1);
-		ImageViewInfo view {};
-		view.format = format;
-		view.type   = range.layerCount == 1 ? vk::ImageViewType::e2D : vk::ImageViewType::e2DArray;
-		view.base_level  = range.baseMipLevel;
-		view.base_layer  = range.baseArrayLayer;
-		view.layer_count = range.layerCount;
-		view.usage       = vk::ImageUsageFlagBits::eColorAttachment;
-		image.Transit(vk::ImageLayout::eColorAttachmentOptimal,
-		              vk::AccessFlagBits2::eColorAttachmentWrite, {}, command.Handle());
-		vk::RenderingAttachmentInfo attachment {};
-		attachment.imageView   = image.FindView(view);
-		attachment.imageLayout = vk::ImageLayout::eColorAttachmentOptimal;
-		attachment.loadOp      = vk::AttachmentLoadOp::eClear;
-		attachment.storeOp     = vk::AttachmentStoreOp::eStore;
-		attachment.clearValue  = clear;
-		vk::RenderingInfo rendering {};
-		rendering.renderArea.extent = {
-		    std::max(image.info.extent.width >> range.baseMipLevel, 1u),
-		    std::max(image.info.extent.height >> range.baseMipLevel, 1u)};
-		rendering.layerCount           = range.layerCount;
-		rendering.colorAttachmentCount = 1;
-		rendering.pColorAttachments    = &attachment;
-		command.Handle().beginRendering(&rendering);
+		BeginColorRendering(image, format, range.baseMipLevel, range.baseArrayLayer,
+		                    range.layerCount, &clear, command.Handle());
 		command.Handle().endRendering();
 		CommitGpuWrite(image);
 		return;
@@ -1619,6 +1670,40 @@ void TextureCache::ClearImage(CommandBuffer& command, ImageId id, vk::Format for
 	CommitGpuWrite(image);
 }
 
+vk::Rect2D TextureCache::BeginColorRendering(Image& image, vk::Format format, uint32_t level,
+                                             uint32_t layer, uint32_t layers,
+                                             const vk::ClearValue* clear,
+                                             vk::CommandBuffer     command) {
+	ImageViewInfo view {};
+	view.format      = format;
+	view.type        = layers == 1 ? vk::ImageViewType::e2D : vk::ImageViewType::e2DArray;
+	view.base_level  = level;
+	view.base_layer  = layer;
+	view.layer_count = layers;
+	view.usage       = vk::ImageUsageFlagBits::eColorAttachment;
+	image.Transit(vk::ImageLayout::eColorAttachmentOptimal,
+	              vk::AccessFlagBits2::eColorAttachmentRead |
+	                  vk::AccessFlagBits2::eColorAttachmentWrite,
+	              {}, command);
+	vk::RenderingAttachmentInfo attachment {};
+	attachment.imageView   = image.FindView(view);
+	attachment.imageLayout = vk::ImageLayout::eColorAttachmentOptimal;
+	attachment.loadOp      = vk::AttachmentLoadOp::eLoad;
+	attachment.storeOp     = vk::AttachmentStoreOp::eStore;
+	if (clear != nullptr) {
+		attachment.loadOp     = vk::AttachmentLoadOp::eClear;
+		attachment.clearValue = *clear;
+	}
+	vk::RenderingInfo rendering {};
+	rendering.renderArea.extent    = {std::max(image.info.extent.width >> level, 1u),
+	                                  std::max(image.info.extent.height >> level, 1u)};
+	rendering.layerCount           = layers;
+	rendering.colorAttachmentCount = 1;
+	rendering.pColorAttachments    = &attachment;
+	command.beginRendering(&rendering);
+	return rendering.renderArea;
+}
+
 void TextureCache::InvalidateMemory(uint64_t address, uint64_t size) {
 	if (!GuestRange {address, size}.Valid()) {
 		EXIT("TextureCache: invalid memory-invalidation range\n");
@@ -1752,6 +1837,7 @@ bool BufferCache::SynchronizeBufferFromImage(Buffer& buffer, uint64_t vaddr, uin
 	     copy_size != image.info.data.size)) {
 		return false;
 	}
+	ForgetColorMetadataChecks(image.info.data.address, copy_size);
 	m_texture_cache.DownloadImage(image, buffer, buf_offset, copy_size, levels);
 	return true;
 }
@@ -1820,23 +1906,20 @@ void TextureCache::InvalidateCpuAliases(uint64_t address, uint64_t size) {
 	const auto page_begin = Common::AlignDown(address, TRACKER_PAGE_SIZE);
 	const auto page_end   = Common::AlignUp(address + size, TRACKER_PAGE_SIZE);
 	for (const auto id: FindImagesInRegion(address, size, true)) {
-		auto owner = m_slot_images.try_get(id);
-		if (owner == nullptr) {
-			continue;
-		}
-		if (owner->Overlaps(address, size)) {
-			owner->MarkCpuDirty();
+		auto& owner = m_slot_images[id];
+		if (owner.Overlaps(address, size)) {
+			owner.MarkCpuDirty();
 			UntrackImage(id);
 			continue;
 		}
-		const auto image_begin = owner->info.data.address;
-		const auto image_end   = owner->info.data.End();
+		const auto image_begin = owner.info.data.address;
+		const auto image_end   = owner.info.data.End();
 		if (page_end < image_end) {
 			UntrackImageHead(id);
 		} else if (image_begin < page_begin) {
 			UntrackImageTail(id);
 		} else {
-			MarkAsMaybeDirty(id, *owner);
+			MarkAsMaybeDirty(id, owner);
 		}
 	}
 }
@@ -1935,6 +2018,11 @@ void TextureCache::RunGarbageCollector() {
 			if (owner == nullptr || !owner->registered || owner->depth_id) {
 				continue;
 			}
+			if (owner->heap_refs.load(std::memory_order_relaxed) != 0u) {
+				// Resident heaps sample it without touching it.
+				TouchImage(*owner);
+				continue;
+			}
 			if (owner->IsGpuModified()) {
 				const bool safe = owner->SafeToDownload();
 				if (safe && owner->info.IsTiled()) {
@@ -1971,8 +2059,7 @@ void TextureCache::ProcessDownloadImages() {
 	}
 	std::scoped_lock lock {m_lock};
 	for (const auto id: m_download_images) {
-		const auto owner = m_slot_images.try_get(id);
-		if (owner != nullptr && owner->registered && owner->IsGpuModified()) {
+		if (m_slot_images[id].IsGpuModified()) {
 			(void)DownloadImageMemory(id);
 		}
 	}

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

@@ -43,6 +43,7 @@ public:
 		ImageInfo     info;
 		ImageViewInfo view_info;
 		BindingType   type = BindingType::Texture;
+		bool          heap = false; // PS5: resolved for a resident image heap.
 	};
 
 	TextureCache(GraphicContext& graphics, CommandScheduler& scheduler, PageManager& page_manager,
@@ -150,6 +151,12 @@ private:
 	// Caller holds m_lock. Volume layer ranges select depth slices.
 	void ClearImage(CommandBuffer& command, ImageId id, vk::Format format,
 	                const vk::ImageSubresourceRange& range, const vk::ClearValue& clear);
+	// Caller holds m_lock and ends rendering. Begins rendering `layers` layers of one color mip
+	// through a `format` view, clearing them or, when `clear` is null, loading them; returns the
+	// render area.
+	vk::Rect2D BeginColorRendering(Image& image, vk::Format format, uint32_t level,
+	                               uint32_t layer, uint32_t layers, const vk::ClearValue* clear,
+	                               vk::CommandBuffer command);
 	void PrepareImageCopy(Image& image);
 	void RefreshCopySource(ImageId id);
 	[[nodiscard]] bool CopyD16(Image& destination, Image& source);
@@ -178,13 +185,13 @@ private:
 	uint64_t                                          m_pressure_gc_memory = 1536ull * 1024 * 1024;
 	uint64_t         m_critical_gc_memory     = 3ull * 1024 * 1024 * 1024;
 	uint64_t         m_gc_tick                = 0;
-	mutable uint32_t m_image_query_epoch      = 0;
 	bool             m_readback_linear_images = false;
 	bool             m_readback_tiled_images  = false;
 
 	friend struct TextureCacheTestAccess;
 	friend class BufferCache;
 	friend class RenderExecutor;
+	friend class ImageHeap;
 	friend bool BcAlias::Matches(const ImageInfo&, const ImageInfo&);
 	friend bool BcAlias::Publish(TextureCache&, Image&);
 	friend void BcAlias::Resolve(TextureCache&, ImageId, ImageId);

src/graphics/host_gpu/renderer/debug.cpp

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

src/graphics/host_gpu/renderer/dynamicState.cpp

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

src/graphics/host_gpu/renderer/dynamicState.h

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

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

@@ -30,16 +30,18 @@ namespace {
 	}
 }
 
-[[nodiscard]] vk::ImageCreateFlags ImageCreateFlags(const GraphicContext& graphics,
-                                                   const ImageInfo& info) {
+[[nodiscard]] vk::ImageCreateFlags
+ImageCreateFlags(const GraphicContext& graphics, const ImageInfo& info, vk::ImageUsageFlags usage) {
 	vk::ImageCreateFlags flags {};
 	if (DepthAspectTransferFormat(info.pixel_format) == vk::Format::eUndefined) {
 		flags |= vk::ImageCreateFlagBits::eMutableFormat;
-		flags |= vk::ImageCreateFlagBits::eExtendedUsage;
 		if (info.IsBlock() && graphics.supports_block_texel_view) {
 			flags |= vk::ImageCreateFlagBits::eBlockTexelViewCompatible;
 		}
 	}
+	if (usage & vk::ImageUsageFlagBits::eStorage) {
+		flags |= vk::ImageCreateFlagBits::eExtendedUsage;
+	}
 	if (info.IsVolume()) {
 		flags |= vk::ImageCreateFlagBits::e2DArrayCompatible;
 	}
@@ -59,7 +61,7 @@ namespace {
 			const auto storage = usage | vk::ImageUsageFlagBits::eStorage;
 			if (graphics.GetImageFormatProperties(info.pixel_format, HostImageType(info.type),
 			                                      vk::ImageTiling::eOptimal, storage,
-			                                      ImageCreateFlags(graphics, info),
+			                                      ImageCreateFlags(graphics, info, storage),
 			                                      nullptr) == vk::Result::eSuccess) {
 				usage = storage;
 			} else {
@@ -88,7 +90,8 @@ namespace {
 	if (HasFormatFeature(properties, vk::FormatFeatureFlagBits::eColorAttachment)) {
 		usage |= vk::ImageUsageFlagBits::eColorAttachment;
 	}
-	if (info.samples == 1) {
+	if (graphics.GetPhysicalDeviceProperties().limits.storageImageSampleCounts &
+	    static_cast<vk::SampleCountFlagBits>(info.samples)) {
 		usage |= vk::ImageUsageFlagBits::eStorage;
 	}
 	return usage;
@@ -209,6 +212,13 @@ Image::Barriers Image::GetBarriers(vk::ImageLayout                      destinat
 		barriers.push_back(barrier);
 	}
 
+	// Resident heaps sample their images in the read state; only leaving it invalidates them.
+	if (std::ranges::any_of(barriers, [&](const vk::ImageMemoryBarrier2& barrier) {
+		    return barrier.oldLayout == ReadLayout() &&
+		           barrier.srcAccessMask == vk::AccessFlagBits2::eShaderRead;
+	    })) {
+		HeapChanged();
+	}
 	state = {destination_stage, destination_access, destination_layout};
 	return barriers;
 }
@@ -415,15 +425,15 @@ void Image::Resolve(Image& source, const ImageSubresourceRange& source_range,
 	const auto destination_height =
 	    std::max(backing.extent.height >> destination_range.base_level, 1u);
 	const bool copy = source.backing.samples == 1;
-	EXIT_IF(layers == 0 || info.extent.width > source_width || info.extent.height > source_height ||
-	        info.extent.width > destination_width || info.extent.height > destination_height ||
+	EXIT_IF(layers == 0 ||
 	        (copy ? !ImageViewOps::FormatsCompatible(source.backing.format, backing.format)
 	              : source.backing.format != backing.format));
 	auto resolved_source_range             = source_range;
 	auto resolved_destination_range        = destination_range;
 	resolved_source_range.layer_count      = layers;
 	resolved_destination_range.layer_count = layers;
-	const vk::Extent3D resolve_extent {info.extent.width, info.extent.height, 1};
+	const vk::Extent3D resolve_extent {std::min(source_width, destination_width),
+	                                   std::min(source_height, destination_height), 1};
 
 	m_scheduler.EndRendering();
 	auto command = m_scheduler.Current().Handle();
@@ -702,7 +712,8 @@ Image::Image(GraphicContext& graphics, CommandScheduler& scheduler, const ImageI
 	}
 
 	vk::ImageCreateInfo create {};
-	create.flags         = ImageCreateFlags(graphics, info);
+	create.usage         = ImageUsageFlags(graphics, info);
+	create.flags         = ImageCreateFlags(graphics, info, create.usage);
 	create.imageType     = HostImageType(info.type);
 	create.extent        = info.extent;
 	create.mipLevels     = info.resources.levels;
@@ -710,7 +721,6 @@ Image::Image(GraphicContext& graphics, CommandScheduler& scheduler, const ImageI
 	create.format        = info.pixel_format;
 	create.tiling        = vk::ImageTiling::eOptimal;
 	create.initialLayout = vk::ImageLayout::eUndefined;
-	create.usage         = ImageUsageFlags(graphics, info);
 	create.samples       = vulkan_sample_count(info.samples);
 
 	vk::ImageFormatProperties properties {};

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

@@ -7,6 +7,7 @@
 #include "graphics/host_gpu/graphicContext.h"
 #include "graphics/host_gpu/renderer/image/imageInfo.h"
 
+#include <atomic>
 #include <compare>
 #include <limits>
 #include <optional>
@@ -71,6 +72,7 @@ public:
 	void MarkCpuDirty() {
 		m_cpu_dirty       = true;
 		m_maybe_cpu_dirty = false;
+		HeapChanged();
 	}
 
 	[[nodiscard]] bool IsCpuDirty() const { return m_cpu_dirty || m_maybe_cpu_dirty; }
@@ -80,6 +82,7 @@ public:
 		if (!IsCpuDirty()) {
 			m_maybe_cpu_hash  = HashGuestEdges();
 			m_maybe_cpu_dirty = true;
+			HeapChanged();
 		}
 	}
 	[[nodiscard]] bool ResolveMaybeCpuHash(uint64_t hash) {
@@ -104,7 +107,10 @@ public:
 	void               ClearGpuModified() noexcept { m_gpu_modified = false; }
 
 	[[nodiscard]] bool IsBufferModified() const noexcept { return m_buffer_modified; }
-	void               MarkBufferModified() noexcept { m_buffer_modified = true; }
+	void               MarkBufferModified() noexcept {
+		m_buffer_modified = true;
+		HeapChanged();
+	}
 	void               ClearBufferModified() noexcept { m_buffer_modified = false; }
 
 	[[nodiscard]] bool Overlaps(uint64_t address, uint64_t size,
@@ -120,14 +126,30 @@ public:
 		return backing.image == nullptr ? 0 : Common::AlignUp(info.data.size, 1024);
 	}
 	[[nodiscard]] uint64_t HashGuestEdges() const;
+	[[nodiscard]] vk::ImageLayout ReadLayout() const noexcept {
+		return info.IsDepth() ? vk::ImageLayout::eDepthStencilReadOnlyOptimal
+		                      : vk::ImageLayout::eShaderReadOnlyOptimal;
+	}
+	// Resident image heaps sample an image without rediscovering it until one changes its
+	// contents, tracking, identity or layout.
+	void HeapChanged() const noexcept {
+		if (heap_refs.load(std::memory_order_relaxed) != 0u) {
+			heap_epoch.fetch_add(1u, std::memory_order_relaxed);
+		}
+	}
 
+	static inline std::atomic<uint64_t> heap_epoch {0};
+	static inline std::atomic<uint64_t> serials {0};
+
+	// Unique per image, unlike ImageId: freed slots are reused.
+	const uint64_t   serial = serials.fetch_add(1u, std::memory_order_relaxed) + 1u;
 	ImageInfo        info;
 	VulkanImage      backing;
 	std::vector<CachedImageView> views;
 	ImageUsage       usage;
 	ImageBinding     binding;
 	bool             registered     = false;
-	mutable uint32_t query_epoch    = 0;
+	mutable bool     picked         = false; // Set only while FindImagesInRegion collects it.
 	uint64_t         track_addr     = 0;
 	uint64_t         track_addr_end = 0;
 	ImageId          depth_id {};
@@ -137,6 +159,8 @@ public:
 	ImageSubresourceRange stencil_subresources;
 	uint64_t         tick_accessed_last = 0;
 	size_t           lru_id             = 0;
+	// Resident image heap elements that select this image.
+	std::atomic<uint32_t> heap_refs {0};
 
 private:
 	friend struct ImageTestAccess;

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

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

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

@@ -101,6 +101,7 @@ IsSupportedStorageImageResource(const ShaderRecompiler::IR::ImageResource& resou
 	       (resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim1D ||
 	        resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim1DArray ||
 	        resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim2D ||
+	        resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim2DMsaa ||
 	        resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim3D ||
 	        resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim2DArray) &&
 	       ((resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::None &&

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

@@ -14,6 +14,7 @@
 #include "gpu_tiler_shaders/gpu_tiler_standard64_3d_spv.h"
 #include "gpu_tiler_shaders/gpu_tiler_standard64_spv.h"
 #include "gpu_tiler_shaders/gpu_tiler_color_transform_spv.h"
+#include "gpu_tiler_shaders/gpu_tiler_color_clear_spv.h"
 #include "graphics/host_gpu/graphicContext.h"
 #include "graphics/host_gpu/renderer/cache/streamBuffer.h"
 #include "graphics/host_gpu/renderer/commandScheduler.h"
@@ -71,20 +72,11 @@ TileManager::~TileManager() {
 			m_graphics.device.destroyPipeline(pipeline, nullptr);
 		}
 	}
-	if (m_d16_to_d24 != nullptr) {
-		m_graphics.device.destroyPipeline(m_d16_to_d24, nullptr);
-	}
-	if (m_d16_to_d32 != nullptr) {
-		m_graphics.device.destroyPipeline(m_d16_to_d32, nullptr);
-	}
-	if (m_d24_to_d16 != nullptr) {
-		m_graphics.device.destroyPipeline(m_d24_to_d16, nullptr);
-	}
-	if (m_d32_to_d16 != nullptr) {
-		m_graphics.device.destroyPipeline(m_d32_to_d16, nullptr);
-	}
-	if (m_color_transform != nullptr) {
-		m_graphics.device.destroyPipeline(m_color_transform, nullptr);
+	for (auto pipeline: {m_d16_to_d24, m_d16_to_d32, m_d24_to_d16, m_d32_to_d16, m_color_transform,
+	                     m_color_clear}) {
+		if (pipeline != nullptr) {
+			m_graphics.device.destroyPipeline(pipeline, nullptr);
+		}
 	}
 	if (m_pipeline_layout != nullptr) {
 		m_graphics.device.destroyPipelineLayout(m_pipeline_layout, nullptr);
@@ -220,42 +212,37 @@ vk::Pipeline TileManager::GetPipeline(uint32_t slot) {
 	if (m_pipelines[slot] != nullptr) {
 		return m_pipelines[slot];
 	}
-	struct Shader {
-		const uint32_t* code;
-		size_t          words;
-	};
-	static constexpr std::array<Shader, FamilyCount> shaders {{
-	    {GPU_TILER_STANDARD256_SPV, std::size(GPU_TILER_STANDARD256_SPV)},
-	    {GPU_TILER_STANDARD4_SPV, std::size(GPU_TILER_STANDARD4_SPV)},
-	    {GPU_TILER_STANDARD4_3D_SPV, std::size(GPU_TILER_STANDARD4_3D_SPV)},
-	    {GPU_TILER_STANDARD64_SPV, std::size(GPU_TILER_STANDARD64_SPV)},
-	    {GPU_TILER_STANDARD64_3D_SPV, std::size(GPU_TILER_STANDARD64_3D_SPV)},
-	    {GPU_TILER_PRT_SPV, std::size(GPU_TILER_PRT_SPV)},
-	    {GPU_TILER_PRT_3D_SPV, std::size(GPU_TILER_PRT_3D_SPV)},
-	    {GPU_TILER_RENDER_TARGET_SPV, std::size(GPU_TILER_RENDER_TARGET_SPV)},
-	    {GPU_TILER_DEPTH_SPV, std::size(GPU_TILER_DEPTH_SPV)},
-	}};
-	const uint32_t                                   element_index = slot % BytesPerElementCount;
+	static constexpr std::array<std::span<const uint32_t>, FamilyCount> shaders {
+	    GPU_TILER_STANDARD256_SPV, GPU_TILER_STANDARD4_SPV,     GPU_TILER_STANDARD4_3D_SPV,
+	    GPU_TILER_STANDARD64_SPV,  GPU_TILER_STANDARD64_3D_SPV, GPU_TILER_PRT_SPV,
+	    GPU_TILER_PRT_3D_SPV,      GPU_TILER_RENDER_TARGET_SPV, GPU_TILER_DEPTH_SPV};
+	const uint32_t                   element_index   = slot % BytesPerElementCount;
 	const uint32_t                   direction_index = slot / (FamilyCount * BytesPerElementCount);
 	const uint32_t                   family_index    = (slot / BytesPerElementCount) % FamilyCount;
 	const uint32_t                   values[] {1u << element_index, direction_index};
 	const vk::SpecializationMapEntry entries[] {{0, 0, 4}, {1, 4, 4}};
 	const vk::SpecializationInfo     specialization {2, entries, sizeof(values), values};
-	const auto module =
-	    CompileSPV({shaders[family_index].code, shaders[family_index].words}, m_graphics.device);
+	m_pipelines[slot] = CreatePipeline(shaders[family_index], &specialization);
+	return m_pipelines[slot];
+}
+
+vk::Pipeline TileManager::CreatePipeline(std::span<const uint32_t>     code,
+                                         const vk::SpecializationInfo* specialization) const {
+	const auto                        module = CompileSPV(code, m_graphics.device);
 	vk::PipelineShaderStageCreateInfo stage {};
 	stage.stage               = vk::ShaderStageFlagBits::eCompute;
 	stage.module              = module;
 	stage.pName               = "main";
-	stage.pSpecializationInfo = &specialization;
+	stage.pSpecializationInfo = specialization;
 	vk::ComputePipelineCreateInfo create {};
-	create.stage  = stage;
-	create.layout = m_pipeline_layout;
-	const auto result =
-	    m_graphics.device.createComputePipelines(nullptr, 1, &create, nullptr, &m_pipelines[slot]);
+	create.stage          = stage;
+	create.layout         = m_pipeline_layout;
+	vk::Pipeline pipeline = nullptr;
+	const auto   result =
+	    m_graphics.device.createComputePipelines(nullptr, 1, &create, nullptr, &pipeline);
 	m_graphics.device.destroyShaderModule(module, nullptr);
 	RequireVulkanSuccess(result, "create TileManager pipeline");
-	return m_pipelines[slot];
+	return pipeline;
 }
 
 void TileManager::Record(vk::Buffer source, uint64_t source_offset, uint64_t source_capacity,
@@ -434,39 +421,21 @@ uint32_t TileManager::ConversionRows(uint64_t offset, uint64_t row_stride, uint6
 
 void TileManager::ConvertD16(Result source, Result target, D16Direction direction, bool d32,
                              const D16Layout& layout) {
-	vk::Pipeline* pipeline_pointer = nullptr;
+	vk::Pipeline*             pipeline_pointer = nullptr;
+	std::span<const uint32_t> code;
 	if (direction == D16Direction::Promote) {
 		pipeline_pointer = d32 ? &m_d16_to_d32 : &m_d16_to_d24;
+		code             = GPU_TILER_PROMOTE_D16_SPV;
 	} else {
 		pipeline_pointer = d32 ? &m_d32_to_d16 : &m_d24_to_d16;
+		code             = GPU_TILER_DEMOTE_D16_SPV;
 	}
 	auto& pipeline = *pipeline_pointer;
 	if (pipeline == nullptr) {
 		const uint32_t                   value = d32 ? 1u : 0u;
 		const vk::SpecializationMapEntry entry {0, 0, sizeof(value)};
 		const vk::SpecializationInfo     specialization {1, &entry, sizeof(value), &value};
-		const uint32_t*                  code  = nullptr;
-		size_t                           words = 0;
-		if (direction == D16Direction::Promote) {
-			code  = GPU_TILER_PROMOTE_D16_SPV;
-			words = std::size(GPU_TILER_PROMOTE_D16_SPV);
-		} else {
-			code  = GPU_TILER_DEMOTE_D16_SPV;
-			words = std::size(GPU_TILER_DEMOTE_D16_SPV);
-		}
-		const auto module = CompileSPV({code, words}, m_graphics.device);
-		vk::PipelineShaderStageCreateInfo stage {};
-		stage.stage               = vk::ShaderStageFlagBits::eCompute;
-		stage.module              = module;
-		stage.pName               = "main";
-		stage.pSpecializationInfo = &specialization;
-		vk::ComputePipelineCreateInfo create {};
-		create.stage  = stage;
-		create.layout = m_pipeline_layout;
-		const auto result =
-		    m_graphics.device.createComputePipelines(nullptr, 1, &create, nullptr, &pipeline);
-		m_graphics.device.destroyShaderModule(module, nullptr);
-		RequireVulkanSuccess(result, "create D16 conversion pipeline");
+		pipeline = CreatePipeline(code, &specialization);
 	}
 
 	const uint64_t source_element =
@@ -515,7 +484,7 @@ void TileManager::ConvertD16(Result source, Result target, D16Direction directio
 	                            vk::AccessFlagBits::eHostWrite |
 	                            vk::AccessFlagBits::eTransferWrite |
 	                            vk::AccessFlagBits::eShaderWrite;
-	barriers[1].dstAccessMask = vk::AccessFlagBits::eShaderRead | vk::AccessFlagBits::eShaderWrite;
+	barriers[1].dstAccessMask = vk::AccessFlagBits::eShaderWrite;
 	command.pipelineBarrier(
 	    vk::PipelineStageFlagBits::eAllCommands | vk::PipelineStageFlagBits::eHost,
 	    vk::PipelineStageFlagBits::eComputeShader, {}, 0, nullptr, 2, barriers, 0, nullptr);
@@ -599,18 +568,7 @@ void TileManager::TransformColor(Result input, Result output, ColorTransform tra
 	        output.size < input.size);
 	const auto pixels = static_cast<uint32_t>(input.size / element_bytes);
 	if (m_color_transform == nullptr) {
-		const auto module = CompileSPV(GPU_TILER_COLOR_TRANSFORM_SPV, m_graphics.device);
-		vk::PipelineShaderStageCreateInfo stage {};
-		stage.stage  = vk::ShaderStageFlagBits::eCompute;
-		stage.module = module;
-		stage.pName  = "main";
-		vk::ComputePipelineCreateInfo create {};
-		create.stage  = stage;
-		create.layout = m_pipeline_layout;
-		const auto result =
-		    m_graphics.device.createComputePipelines(nullptr, 1, &create, nullptr, &m_color_transform);
-		m_graphics.device.destroyShaderModule(module, nullptr);
-		RequireVulkanSuccess(result, "create color transform pipeline");
+		m_color_transform = CreatePipeline(GPU_TILER_COLOR_TRANSFORM_SPV, nullptr);
 	}
 	const auto input_binding  = BindStorage(input, input.size);
 	const auto output_binding = BindStorage(output, input.size);
@@ -670,4 +628,60 @@ TileManager::Result TileManager::TransformColor(Result input, ColorTransform tra
 	return result;
 }
 
+vk::Buffer TileManager::CheckColorClear(Result metadata, uint32_t slices, uint32_t keys,
+                                        bool expand) {
+	struct Check {
+		uint32_t metadata_base;
+		uint32_t words;
+		uint32_t keys;
+		uint32_t expand;
+	};
+	static_assert(sizeof(Check) <= sizeof(Conversion));
+	const uint32_t max_slices =
+	    m_graphics.GetPhysicalDeviceProperties().limits.maxFramebufferLayers;
+	EXIT_IF(slices == 0 || slices > max_slices || metadata.size % (4ull * slices) != 0 ||
+	        metadata.size / 4 > UINT32_MAX);
+	if (m_color_clear == nullptr) {
+		m_color_clear      = CreatePipeline(GPU_TILER_COLOR_CLEAR_SPV, nullptr);
+		m_clear_predicates = std::make_unique<Buffer>(
+		    m_graphics, m_scheduler, MemoryUsage::DeviceLocal, 0,
+		    vk::BufferUsageFlagBits::eStorageBuffer |
+		        vk::BufferUsageFlagBits::eConditionalRenderingEXT,
+		    uint64_t {max_slices} * ColorClearSlots * sizeof(uint32_t));
+	}
+	const auto source = BindStorage(metadata, metadata.size);
+	const vk::DescriptorBufferInfo infos[] {source.info,
+	                                        {m_clear_predicates->Handle(), 0, VK_WHOLE_SIZE}};
+	std::array<vk::WriteDescriptorSet, 2> writes {};
+	for (uint32_t index = 0; index < writes.size(); index++) {
+		writes[index].dstBinding      = index;
+		writes[index].descriptorCount = 1;
+		writes[index].descriptorType  = vk::DescriptorType::eStorageBuffer;
+		writes[index].pBufferInfo     = &infos[index];
+	}
+	// Orders the metadata writers and the previous check's predicate reads before this check.
+	vk::MemoryBarrier barrier {
+	    .srcAccessMask = vk::AccessFlagBits::eMemoryWrite,
+	    .dstAccessMask = vk::AccessFlagBits::eShaderRead | vk::AccessFlagBits::eShaderWrite};
+	m_scheduler.EndRendering();
+	auto command = m_scheduler.Current().Handle();
+	command.pipelineBarrier(vk::PipelineStageFlagBits::eAllCommands,
+	                        vk::PipelineStageFlagBits::eComputeShader, {}, 1, &barrier, 0, nullptr,
+	                        0, nullptr);
+	command.bindPipeline(vk::PipelineBindPoint::eCompute, m_color_clear);
+	command.pushDescriptorSetKHR(vk::PipelineBindPoint::eCompute, m_pipeline_layout, 0,
+	                             static_cast<uint32_t>(writes.size()), writes.data());
+	const Check push {source.base, static_cast<uint32_t>(metadata.size / 4 / slices), keys,
+	                  expand};
+	command.pushConstants(m_pipeline_layout, vk::ShaderStageFlagBits::eCompute, 0, sizeof(push),
+	                      &push);
+	command.dispatch(slices, 1, 1);
+	barrier.srcAccessMask = vk::AccessFlagBits::eShaderWrite;
+	barrier.dstAccessMask = vk::AccessFlagBits::eMemoryRead | vk::AccessFlagBits::eMemoryWrite;
+	command.pipelineBarrier(vk::PipelineStageFlagBits::eComputeShader,
+	                        vk::PipelineStageFlagBits::eAllCommands, {}, 1, &barrier, 0, nullptr,
+	                        0, nullptr);
+	return m_clear_predicates->Handle();
+}
+
 } // namespace Libs::Graphics

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

@@ -54,6 +54,13 @@ public:
 	                                const D16Layout& layout);
 	[[nodiscard]] Result TransformColor(Result input, ColorTransform transform, bool to_host);
 	void TransformColor(Result input, Result output, ColorTransform transform, bool to_host);
+	// Color-clear keys are multiples of 0x20, so key >> 5 selects one of these slots.
+	static constexpr uint32_t ColorClearSlots = 8;
+	// Returns the buffer whose dword slice * ColorClearSlots + (key >> 5) is set when every byte
+	// of that metadata slice equals a key whose bit `keys` selects; a hit expands to 0xff when
+	// `expand` is set.
+	[[nodiscard]] vk::Buffer CheckColorClear(Result metadata, uint32_t slices, uint32_t keys,
+	                                         bool expand);
 
 private:
 	friend struct TileManagerTestAccess;
@@ -117,6 +124,8 @@ private:
 	                            vk::Buffer target, uint64_t target_offset, uint64_t target_capacity,
 	                            const Tiling& tiling, bool clear_target);
 	[[nodiscard]] vk::Pipeline GetPipeline(uint32_t slot);
+	[[nodiscard]] vk::Pipeline CreatePipeline(std::span<const uint32_t>     code,
+	                                          const vk::SpecializationInfo* specialization) const;
 
 	GraphicContext&                         m_graphics;
 	CommandScheduler&                       m_scheduler;
@@ -130,6 +139,8 @@ private:
 	vk::Pipeline                            m_d24_to_d16      = nullptr;
 	vk::Pipeline                            m_d32_to_d16      = nullptr;
 	vk::Pipeline                            m_color_transform = nullptr;
+	vk::Pipeline                            m_color_clear     = nullptr;
+	std::unique_ptr<Buffer>                 m_clear_predicates;
 };
 
 } // namespace Libs::Graphics

src/graphics/host_gpu/renderer/occlusionQueries.cpp

@@ -0,0 +1,82 @@
+#include "graphics/host_gpu/renderer/occlusionQueries.h"
+
+#include "graphics/host_gpu/graphicContext.h"
+#include "graphics/host_gpu/renderer/commandScheduler.h"
+
+#include <algorithm>
+
+namespace Libs::Graphics {
+
+// Queries no dump has read yet; past them a draw counts one sample.
+constexpr uint32_t QUERY_COUNT = 16384;
+
+OcclusionQueries::OcclusionQueries(GraphicContext& graphics, CommandScheduler& scheduler)
+    : m_scheduler(scheduler) {
+	if (!graphics.occlusion_queries_enabled) {
+		return;
+	}
+	m_device = graphics.device;
+	const vk::QueryPoolCreateInfo info {.queryType  = vk::QueryType::eOcclusion,
+	                                    .queryCount = QUERY_COUNT};
+	RequireVulkanSuccess(m_device.createQueryPool(&info, nullptr, &m_pool), "vkCreateQueryPool");
+	m_device.resetQueryPool(m_pool, 0, QUERY_COUNT);
+	m_results.resize(QUERY_COUNT);
+	if (graphics.occlusion_query_precise_enabled) {
+		m_control = vk::QueryControlFlagBits::ePrecise;
+	}
+}
+
+OcclusionQueries::~OcclusionQueries() {
+	if (m_pool) {
+		m_device.destroyQueryPool(m_pool, nullptr);
+	}
+}
+
+int32_t OcclusionQueries::BeginQuery(vk::CommandBuffer command) {
+	if (m_begun - m_read == QUERY_COUNT) {
+		m_uncounted++;
+		return -1;
+	}
+	const auto query = static_cast<uint32_t>(m_begun++ % QUERY_COUNT);
+	command.beginQuery(m_pool, query, m_control);
+	return static_cast<int32_t>(query);
+}
+
+void OcclusionQueries::Dump(uint64_t address) {
+	// Begins dump to the results, ends 8 bytes into them.
+	if ((address & 8u) == 0) {
+		m_open++;
+	} else if (m_open != 0) {
+		m_open--;
+	}
+	m_dumps.push_back({address, m_scheduler.CurrentTick(), m_begun, m_uncounted});
+}
+
+void OcclusionQueries::PublishDumps() {
+	if (const auto last = m_dumps.back().queries; last != m_read) {
+		m_scheduler.Wait(m_dumps.back().tick);
+		// The unread queries wrap around the pool at most once.
+		for (uint64_t read = m_read; read < last;) {
+			const auto first = static_cast<uint32_t>(read % QUERY_COUNT);
+			const auto count =
+			    static_cast<uint32_t>(std::min<uint64_t>(last - read, QUERY_COUNT - first));
+			RequireVulkanSuccess(
+			    m_device.getQueryPoolResults(m_pool, first, count, count * sizeof(uint64_t),
+				                             m_results.data() + first, sizeof(uint64_t),
+				                             vk::QueryResultFlagBits::e64),
+			    "vkGetQueryPoolResults");
+			m_device.resetQueryPool(m_pool, first, count);
+			read += count;
+		}
+	}
+	for (const auto& dump: m_dumps) {
+		for (; m_read < dump.queries; m_read++) {
+			m_samples += m_results[m_read % QUERY_COUNT];
+		}
+		WriteZPassCounters(reinterpret_cast<volatile uint64_t*>(dump.address),
+		                   m_samples + dump.uncounted, (dump.address & 8u) == 0);
+	}
+	m_dumps.clear();
+}
+
+} // namespace Libs::Graphics

src/graphics/host_gpu/renderer/occlusionQueries.h

@@ -0,0 +1,85 @@
+#ifndef EMULATOR_SRC_GRAPHICS_HOST_GPU_RENDERER_OCCLUSIONQUERIES_H_
+#define EMULATOR_SRC_GRAPHICS_HOST_GPU_RENDERER_OCCLUSIONQUERIES_H_
+
+#include "common/common.h"
+#include "graphics/host_gpu/vulkanCommon.h"
+
+#include <cstdint>
+#include <vector>
+
+namespace Libs::Graphics {
+
+class CommandScheduler;
+struct GraphicContext;
+
+// The guest's query results (16-byte aligned): a begin/end pair of running z-pass counters per DB,
+// 16 bytes apart; bit 63 marks a written counter. A query's samples are the sum of end - begin.
+constexpr uint32_t OCCLUSION_QUERY_DBS   = 16;
+constexpr uint64_t OCCLUSION_QUERY_READY = 1ull << 63u;
+
+// A dump of the running count, all of it in DB 0. A begin also clears its end counters.
+inline void WriteZPassCounters(volatile uint64_t* results, uint64_t samples, bool begin) {
+	for (uint32_t db = 0; db < OCCLUSION_QUERY_DBS; db++) {
+		results[db * 2] = OCCLUSION_QUERY_READY | (db == 0 ? samples : 0);
+		if (begin) {
+			results[db * 2 + 1] = 0;
+		}
+	}
+}
+
+// The guest's z-pass counters from host occlusion queries: a draw counts its samples while a guest
+// query is open, and the dumps' counters are written before the guest can see the GPU ran them.
+class OcclusionQueries {
+public:
+	OcclusionQueries(GraphicContext& graphics, CommandScheduler& scheduler);
+	~OcclusionQueries();
+	KYTY_CLASS_NO_COPY(OcclusionQueries);
+
+	[[nodiscard]] bool Enabled() const { return m_pool != nullptr; }
+
+	// In the draw's rendering: the query that counts its samples, or -1.
+	[[nodiscard]] int32_t Begin(vk::CommandBuffer command) {
+		return m_open == 0 ? -1 : BeginQuery(command);
+	}
+	void End(vk::CommandBuffer command, int32_t query) const {
+		if (query >= 0) {
+			command.endQuery(m_pool, static_cast<uint32_t>(query));
+		}
+	}
+
+	// A PIXEL_PIPE_STAT_DUMP to address, after the draws recorded so far.
+	void Dump(uint64_t address);
+	// Waits for the GPU to run the dumps and writes their counters.
+	void Publish() {
+		if (!m_dumps.empty()) {
+			PublishDumps();
+		}
+	}
+
+private:
+	struct PendingDump {
+		uint64_t address;
+		uint64_t tick;
+		uint64_t queries;
+		uint64_t uncounted;
+	};
+
+	int32_t BeginQuery(vk::CommandBuffer command);
+	void    PublishDumps();
+
+	CommandScheduler&        m_scheduler;
+	vk::Device               m_device;
+	vk::QueryPool            m_pool;
+	vk::QueryControlFlags    m_control;
+	std::vector<PendingDump> m_dumps;
+	std::vector<uint64_t>    m_results;
+	uint32_t                 m_open      = 0;
+	uint64_t                 m_begun     = 0;
+	uint64_t                 m_read      = 0;
+	uint64_t                 m_samples   = 0;
+	uint64_t                 m_uncounted = 0;
+};
+
+} // namespace Libs::Graphics
+
+#endif // EMULATOR_SRC_GRAPHICS_HOST_GPU_RENDERER_OCCLUSIONQUERIES_H_

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

@@ -254,6 +254,10 @@ static bool IsSupportedStorageTextureDescriptor(const ShaderRecompiler::IR::Imag
 	const bool is_3d = resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim3D &&
 	                   descriptor.Type() == Prospero::ImageType::kColor3D &&
 	                   descriptor.BaseArray5() == 0;
+	const bool is_2d_msaa =
+	    resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim2DMsaa &&
+	    descriptor.Type() == Prospero::ImageType::kColor2DMsaa && descriptor.Depth() == 0 &&
+	    descriptor.BaseArray5() == 0;
 	TileTextureBlockLayout tile_layout {};
 	bool                   supported_tile = false;
 	switch (tile) {
@@ -284,8 +288,8 @@ static bool IsSupportedStorageTextureDescriptor(const ShaderRecompiler::IR::Imag
 	const bool supported_swizzle =
 	    IsValidImageSwizzle(swizzle) &&
 	    (swizzle == DstSel(4, 5, 6, 7) || !resource.read || resource.atomic);
-	return (is_1d || is_1d_array || is_2d || is_2d_array || is_3d) && supported_tile &&
-	       descriptor.BaseLevel() <= descriptor.LastLevel() &&
+	return (is_1d || is_1d_array || is_2d || is_2d_array || is_3d || is_2d_msaa) &&
+	       supported_tile && descriptor.BaseLevel() <= descriptor.LastLevel() &&
 	       descriptor.MinLod() == 0 && supported_swizzle && descriptor.BCSwizzle() == 0 &&
 	       !descriptor.MsaaDepth();
 }
@@ -356,7 +360,7 @@ static ImageViewInfo TextureViewInfo(const ShaderRecompiler::IR::ImageResource&
 	}
 	view.usage = storage ? vk::ImageUsageFlagBits::eStorage : vk::ImageUsageFlagBits::eSampled;
 	view.mapping =
-	    storage || surface_format.conversion_format != Prospero::BufferFormat::kInvalid
+	    storage || resource.conversion_format != Prospero::BufferFormat::kInvalid
 	        ? vk::ComponentMapping {}
 	        : TextureGetComponentMapping(descriptor.DstSelXYZW(), surface_format.host_to_storage);
 	switch (resource.dimension) {
@@ -515,7 +519,9 @@ TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageR
 	}
 	const auto samples = multisampled ? 1u << last_level : 1u;
 	const auto depth          = static_cast<uint32_t>(descriptor.Depth()) + 1u;
-	const auto format         = descriptor.Format();
+	const auto format         = ShaderRecompiler::IR::ReadsDepthBits(resource)
+	                                ? Prospero::BufferFormat::k16UNorm
+	                                : descriptor.Format();
 	const auto surface_format = TextureGetSurfaceFormatInfo(format);
 	const bool shader_conversion =
 	    surface_format.conversion_format != Prospero::BufferFormat::kInvalid;
@@ -579,6 +585,10 @@ TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageR
 	const auto alignment = desc.info.DataAlignment();
 	EXIT_NOT_IMPLEMENTED(desc.info.data.size == 0 || alignment == 0 ||
 	                     (address & (static_cast<uint64_t>(alignment) - 1u)) != 0);
+	// Image heaps keep the descriptors of unmapped textures, which PS5 never samples.
+	if (resource.heap && !m_context.IsMapped(address, desc.info.data.size)) {
+		return {.layout = vk::ImageLayout::eGeneral};
+	}
 	if (storage) {
 		ValidateStorageTexture(resource, descriptor, desc.info.data.size);
 	}
@@ -594,8 +604,13 @@ TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageR
 	                                 view_levels, desc.info.resources.layers);
 	desc.view_info.base_level = view_base;
 	desc.type = storage ? TextureCache::BindingType::Storage : TextureCache::BindingType::Texture;
+	desc.heap = resource.heap;
 
-	auto       id                  = texture_cache.FindImage(desc, shader_conversion);
+	auto id = texture_cache.FindImage(desc, shader_conversion);
+	if (!id) {
+		// Another extent at the address holds it; the heap retries once an image is created.
+		return {.desc = std::move(desc), .layout = vk::ImageLayout::eGeneral};
+	}
 	auto*      image               = &texture_cache.GetImage(id);
 	const bool stencil_association = static_cast<bool>(image->depth_id);
 	if (stencil_association) {
@@ -662,9 +677,7 @@ void RenderExecutor::BindRenderTarget(ImageId id) {
 
 void RenderExecutor::ResetBindings() {
 	for (const auto id: m_bound_images) {
-		if (auto* image = m_context.GetTextureCache().m_slot_images.try_get(id); image != nullptr) {
-			image->binding = {};
-		}
+		m_context.GetTextureCache().m_slot_images[id].binding = {};
 	}
 	m_bound_images.clear();
 }
@@ -673,7 +686,7 @@ void RenderExecutor::PrepareBindings(const ShaderStageRuntime& runtime,
                                      PreparedBindings& prepared) {
 	KYTY_PROFILER_FUNCTION();
 	EXIT_IF(!runtime);
-	const auto& program  = *runtime.program;
+	const auto& program  = runtime.module->info;
 	const auto& snapshot = *runtime.resources;
 	prepared.runtime = &runtime;
 	prepared.gds = {nullptr, 0, VK_WHOLE_SIZE};
@@ -684,6 +697,10 @@ void RenderExecutor::PrepareBindings(const ShaderStageRuntime& runtime,
 	prepared.samplers.clear();
 	prepared.shader_data.clear();
 	for (uint32_t i = 0; i < program.info.images.size(); i++) {
+		if (program.info.images[i].heap) {
+			prepared.images[i] = {};
+			continue;
+		}
 		auto binding = ResolveTexture(program.info.images[i], snapshot.images[i]);
 		BindImage(binding.image_id, binding.desc.type == TextureCache::BindingType::Storage);
 		binding.mip_views.swap(prepared.images[i].mip_views);
@@ -711,7 +728,7 @@ void RenderExecutor::FindBuffers(std::span<PreparedBindings* const> stages) {
 	for (auto* stage: stages) {
 		auto& prepared = *stage;
 		EXIT_IF(prepared.runtime == nullptr || !*prepared.runtime);
-		const auto& program  = *prepared.runtime->program;
+		const auto& program  = prepared.runtime->module->info;
 		const auto& snapshot = *prepared.runtime->resources;
 		const auto& layout = program.bindings;
 		prepared.buffer_sources.resize(layout.descriptor_counts[static_cast<size_t>(BindingKind::Buffers)]);
@@ -768,7 +785,7 @@ void RenderExecutor::FindBuffers(std::span<PreparedBindings* const> stages) {
 void RenderExecutor::RebindBuffers(PreparedBindings& prepared) {
 	KYTY_PROFILER_FUNCTION();
 	EXIT_IF(prepared.runtime == nullptr || !*prepared.runtime);
-	const auto& program   = *prepared.runtime->program;
+	const auto& program   = prepared.runtime->module->info;
 	const auto& snapshot  = *prepared.runtime->resources;
 	const auto& layout    = program.bindings;
 	EXIT_IF(prepared.buffer_sources.size() != layout.descriptor_counts[static_cast<size_t>(BindingKind::Buffers)]);
@@ -789,6 +806,9 @@ void RenderExecutor::RebindBuffers(PreparedBindings& prepared) {
 		prepared.buffers[index] = NativeStorageBuffer(m_context, prepared.buffer_sources[index],
 		                                              resource, buffer_offset);
 		pack_memory_offset(index, buffer_offset);
+		// Null bindings allow 0 bytes, not the shared null buffer's 16, so no store reaches it.
+		prepared.shader_data[layout.MemorySizeDword() + index] =
+		    static_cast<uint32_t>(prepared.buffer_sources[index].size + buffer_offset);
 	}
 	if (layout.descriptor_counts[static_cast<size_t>(BindingKind::FlattenedSrt)] != 0) {
 		prepared.flattened_srt = NativeUpload(m_context, snapshot.flattened_srt);
@@ -801,7 +821,7 @@ void RenderExecutor::RebindBuffers(PreparedBindings& prepared) {
 void RenderExecutor::RebindImages(PreparedBindings& prepared) {
 	KYTY_PROFILER_FUNCTION();
 	EXIT_IF(prepared.runtime == nullptr || !*prepared.runtime);
-	const auto& program  = *prepared.runtime->program;
+	const auto& program  = prepared.runtime->module->info;
 	const auto& snapshot = *prepared.runtime->resources;
 	auto&       images   = prepared.images;
 	EXIT_IF(images.size() != program.info.images.size());
@@ -809,12 +829,9 @@ void RenderExecutor::RebindImages(PreparedBindings& prepared) {
 	// Acquire each view before a later overlapping descriptor can replace its image.
 	for (uint32_t i = 0; i < program.info.images.size(); i++) {
 		if (!images[i].image_id) continue;
-		const auto old_image = texture_cache.m_slot_images.try_get(images[i].image_id);
-		if (old_image == nullptr || !old_image->registered ||
-		    old_image->binding.needs_rebind) {
-			if (old_image != nullptr) {
-				old_image->binding = {};
-			}
+		auto& old_image = texture_cache.m_slot_images[images[i].image_id];
+		if (!old_image.registered || old_image.binding.needs_rebind) {
+			old_image.binding = {};
 			images[i] = ResolveTexture(program.info.images[i], snapshot.images[i]);
 			BindImage(images[i].image_id,
 			          images[i].desc.type == TextureCache::BindingType::Storage);
@@ -846,26 +863,28 @@ void RenderExecutor::RebindImages(PreparedBindings& prepared) {
 
 void RenderExecutor::PrepareGraphicsBindings(std::span<PreparedBindings* const> stages,
                                              std::span<RenderColorInfo> colors) {
-	bool uses_dma = false;
+	bool uses_dma        = false;
+	bool uses_image_heap = false;
 	FindBuffers(stages);
 	for (auto* stage: stages) {
-		uses_dma |= stage->runtime->program->info.uses_dma;
+		uses_dma |= stage->runtime->module->info.info.uses_dma;
+		uses_image_heap |= stage->runtime->module->info.info.uses_image_heap;
 	}
 	if (uses_dma) {
 		m_context.PrepareBda();
 	}
+	if (uses_image_heap) {
+		m_context.GetImageHeap().Rebind(*this);
+	}
 	for (auto* stage: stages) {
 		RebindImages(*stage);
 	}
 	auto& cache = m_context.GetTextureCache();
 	for (auto& target: colors) {
 		EXIT_IF(!target.image_id);
-		const auto old_image = cache.m_slot_images.try_get(target.image_id);
-		if (old_image == nullptr || !old_image->registered ||
-		    old_image->binding.needs_rebind) {
-			if (old_image != nullptr) {
-				old_image->binding = {};
-			}
+		auto& old_image = cache.m_slot_images[target.image_id];
+		if (!old_image.registered || old_image.binding.needs_rebind) {
+			old_image.binding = {};
 			target.desc.view_info.base_level = target.guest_mip_level;
 			target.desc.view_info.base_layer = target.guest_array_layer;
 			target.image_id = cache.FindImage(target.desc);
@@ -898,7 +917,7 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 	                                       : vk::ShaderStageFlags {};
 	for (const auto* prepared: prepared_bindings) {
 		EXIT_IF(prepared == nullptr || prepared->runtime == nullptr || !*prepared->runtime);
-		const auto& program = *prepared->runtime->program;
+		const auto& program = prepared->runtime->module->info;
 		write_count += std::popcount(program.bindings.descriptor_mask);
 		for (auto mask = program.bindings.descriptor_mask; mask != 0; mask &= mask - 1) {
 			const auto index = std::countr_zero(mask);
@@ -919,17 +938,16 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 		const auto& reads = reader->runtime->resources->specialization_reads;
 		if (reads.empty()) continue;
 		for (const auto* writer: prepared_bindings) {
-			if (writer->runtime->program->has_address_writes) {
+			if (writer->runtime->module->info.has_address_writes) {
 				EXIT("scalar resource reads cannot be proven disjoint from shader address writes\n");
 			}
 		}
 		for (const auto [address, size]: reads) {
 			for (const auto id: m_bound_images) {
-				const auto* image = m_context.GetTextureCache().m_slot_images.try_get(id);
-				if (image == nullptr ||
-				    (!image->binding.shader_write && !image->binding.is_target)) continue;
-				for (const auto written: {image->info.data, image->info.stencil,
-				                          image->info.metadata.range}) {
+				const auto& image = m_context.GetTextureCache().m_slot_images[id];
+				if (!image.binding.shader_write && !image.binding.is_target) continue;
+				for (const auto written: {image.info.data, image.info.stencil,
+				                          image.info.metadata.range}) {
 					if (written.size != 0 && ImageRangeOverlaps(address, size,
 					                                          written.address, written.size)) {
 						EXIT("scalar resource reads overlap an image or attachment write\n");
@@ -937,7 +955,7 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 				}
 			}
 			for (const auto* writer: prepared_bindings) {
-				const auto& program = *writer->runtime->program;
+				const auto& program = writer->runtime->module->info;
 				for (const auto& resource: program.info.buffers) {
 					if (!resource.written || resource.descriptor_index == UINT32_MAX) continue;
 					const auto& written = writer->buffer_sources[resource.descriptor_index];
@@ -953,6 +971,9 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 			}
 		}
 	}
+	if (pipeline.uses_image_heap) {
+		m_context.GetImageHeap().Commit(buffer, pipeline_bind_point, pipeline.pipeline_layout);
+	}
 	m_descriptor_images.resize(image_descriptor_count);
 	m_descriptor_writes.clear();
 	m_descriptor_writes.reserve(write_count);
@@ -960,7 +981,7 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 	std::array<vk::DescriptorBufferInfo, 2> address_buffers;
 
 	for (auto* prepared: prepared_bindings) {
-		const auto& program       = *prepared->runtime->program;
+		const auto& program       = prepared->runtime->module->info;
 		auto&       descriptors   = *prepared;
 		const auto  shader_stage  = NativeShaderStage(program.stage);
 		const auto  shader_stages = ShaderPipelineStages(shader_stage);
@@ -1010,9 +1031,7 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 				              vk::AccessFlagBits2::eShaderRead | vk::AccessFlagBits2::eShaderWrite,
 				              range, vk_buffer);
 			} else {
-				image.Transit(image.info.IsDepth() ? vk::ImageLayout::eDepthStencilReadOnlyOptimal
-				                                   : vk::ImageLayout::eShaderReadOnlyOptimal,
-				              vk::AccessFlagBits2::eShaderRead, range, vk_buffer);
+				image.Transit(image.ReadLayout(), vk::AccessFlagBits2::eShaderRead, range, vk_buffer);
 			}
 			binding.layout = image.backing.state.layout;
 		}
@@ -1083,6 +1102,7 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 		for (uint32_t i = 0; i < program.info.images.size(); ++i) {
 			const auto& resource = program.info.images[i];
 			const auto& texture = descriptors.images[i];
+			if (resource.heap) continue;
 			EXIT_IF(texture.image_id &&
 			        resource.mip_count != (texture.mip_views.empty() ? 1u : texture.mip_views.size()));
 			const auto kind = *ShaderRecompiler::IR::DescriptorBindingForImage(resource);

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

@@ -0,0 +1,638 @@
+#include "graphics/host_gpu/renderer/pipeline/imageHeap.h"
+
+#include "common/alignment.h"
+#include "common/assert.h"
+#include "common/profiler.h"
+#include "graphics/host_gpu/pageManager.h"
+#include "graphics/host_gpu/regionDefinitions.h"
+#include "graphics/host_gpu/renderer/cache/streamBuffer.h"
+#include "graphics/host_gpu/renderer/render.h"
+#include "graphics/host_gpu/renderer/renderContext.h"
+#include "graphics/shader/recompiler/ir/passes/ResourceMaterialization.h"
+#include "kernel/memory.h"
+
+#include <algorithm>
+#include <bit>
+#include <xxhash.h>
+
+namespace Libs::Graphics {
+namespace {
+
+namespace IR = ShaderRecompiler::IR;
+
+// Smaller tables are often written for one draw, so they enter on a second draw of a frame.
+constexpr uint64_t MinTableBytes = 64ull * 1024ull;
+constexpr uint32_t ElementMask   = (1u << IR::ImageHeapKindShift) - 1u;
+
+template <typename Pred, typename Fn>
+void ForEachRun(size_t begin, size_t end, Pred&& pred, Fn&& fn) {
+	while (begin < end) {
+		if (!pred(begin)) {
+			++begin;
+			continue;
+		}
+		auto run = begin + 1u;
+		while (run < end && pred(run)) ++run;
+		fn(begin, run);
+		begin = run;
+	}
+}
+
+// Grows [begin, end) to cover [first, last).
+void Widen(size_t& begin, size_t& end, size_t first, size_t last) {
+	const bool empty = begin == end;
+	begin            = empty ? first : std::min(begin, first);
+	end              = empty ? last : std::max(end, last);
+}
+
+// Probes at first + k * step whose first dword lies within a table of size bytes.
+uint64_t Probes(uint64_t size, uint64_t first, uint64_t step) {
+	return size >= first + 4u ? (size - first - 4u) / step + 1u : 0u;
+}
+
+// Tables grow a little on most frames, so their slot words get room to double.
+std::unique_ptr<Buffer> SlotBuffer(RenderContext& context, uint64_t size) {
+	return std::make_unique<Buffer>(
+	    context.GetGraphics(), context.GetCommandScheduler(), MemoryUsage::DeviceLocal, 0,
+	    vk::BufferUsageFlagBits::eTransferDst | vk::BufferUsageFlagBits::eStorageBuffer |
+	        vk::BufferUsageFlagBits::eShaderDeviceAddress,
+	    std::bit_ceil(size + 4u));
+}
+
+// Guest bytes of a table that its tracker pages [begin, end) cover.
+std::pair<uint64_t, uint64_t> PageRange(uint64_t base, uint64_t size, size_t begin, size_t end) {
+	const auto origin = Common::AlignDown(base, TRACKER_PAGE_SIZE);
+	return {std::max(base, origin + begin * TRACKER_PAGE_SIZE),
+	        std::min(base + size, origin + end * TRACKER_PAGE_SIZE)};
+}
+
+} // namespace
+
+size_t ImageHeap::DescriptorHash::operator()(const DescriptorValue& value) const noexcept {
+	return XXH3_64bits(value.dwords.data(), sizeof(value.dwords));
+}
+
+ImageHeap::ImageHeap(RenderContext& context, PageManager& page_manager)
+    : m_context(context), m_page_manager(page_manager) {
+	auto& graphics = context.GetGraphics();
+	vk::PhysicalDeviceVulkan12Properties limits {};
+	vk::PhysicalDeviceProperties2        properties {};
+	properties.pNext = &limits;
+	graphics.physical_device.getProperties2(&properties);
+	EXIT_NOT_IMPLEMENTED(std::min({limits.maxDescriptorSetUpdateAfterBindSampledImages,
+	                               limits.maxPerStageDescriptorUpdateAfterBindSampledImages,
+	                               limits.maxPerStageUpdateAfterBindResources}) <
+	                     IR::ImageHeapCapacity);
+
+	const vk::DescriptorBindingFlags binding_flags =
+	    vk::DescriptorBindingFlagBits::ePartiallyBound |
+	    vk::DescriptorBindingFlagBits::eUpdateAfterBind |
+	    vk::DescriptorBindingFlagBits::eUpdateUnusedWhilePending;
+	vk::DescriptorSetLayoutBindingFlagsCreateInfo flags_info {};
+	flags_info.bindingCount  = 1;
+	flags_info.pBindingFlags = &binding_flags;
+	const vk::DescriptorSetLayoutBinding binding {0, vk::DescriptorType::eSampledImage,
+	                                              IR::ImageHeapCapacity,
+	                                              vk::ShaderStageFlagBits::eAll, nullptr};
+	vk::DescriptorSetLayoutCreateInfo layout_info {};
+	layout_info.pNext        = &flags_info;
+	layout_info.flags        = vk::DescriptorSetLayoutCreateFlagBits::eUpdateAfterBindPool;
+	layout_info.bindingCount = 1;
+	layout_info.pBindings    = &binding;
+	EXIT_IF(graphics.device.createDescriptorSetLayout(&layout_info, nullptr, &m_layout) !=
+	        vk::Result::eSuccess);
+	const vk::DescriptorPoolSize pool_size {vk::DescriptorType::eSampledImage,
+	                                        IR::ImageHeapCapacity};
+	vk::DescriptorPoolCreateInfo pool_info {};
+	pool_info.flags         = vk::DescriptorPoolCreateFlagBits::eUpdateAfterBind;
+	pool_info.maxSets       = 1;
+	pool_info.poolSizeCount = 1;
+	pool_info.pPoolSizes    = &pool_size;
+	EXIT_IF(graphics.device.createDescriptorPool(&pool_info, nullptr, &m_pool) !=
+	        vk::Result::eSuccess);
+	vk::DescriptorSetAllocateInfo allocate {};
+	allocate.descriptorPool     = m_pool;
+	allocate.descriptorSetCount = 1;
+	allocate.pSetLayouts        = &m_layout;
+	EXIT_IF(graphics.device.allocateDescriptorSets(&allocate, &m_set) != vk::Result::eSuccess);
+	graphics.image_heap_layout = m_layout;
+	// Slot word 0 selects element 0, which stays a null image.
+	m_elements.push_back({.binding = {.layout = vk::ImageLayout::eShaderReadOnlyOptimal}});
+	m_writes.push_back(0u);
+	WriteDescriptors();
+}
+
+ImageHeap::~ImageHeap() {
+	for (const auto& memory: m_memories) Unwatch(*memory);
+	auto& graphics             = m_context.GetGraphics();
+	graphics.image_heap_layout = nullptr;
+	graphics.device.destroyDescriptorPool(m_pool, nullptr);
+	graphics.device.destroyDescriptorSetLayout(m_layout, nullptr);
+}
+
+bool ImageHeap::Acquire(const IR::ImageHeapRequest& request, IR::ImageHeapView& result) {
+	KYTY_PROFILER_FUNCTION();
+	const auto size = request.size & ~uint64_t {3};
+	if (request.count == 0u || request.step == 0u ||
+	    ((request.base | request.first | request.step) & 3u) != 0u) {
+		return false;
+	}
+	auto memory = std::ranges::find_if(
+	    m_memories, [&](const auto& candidate) { return candidate->base == request.base; });
+	if (memory == m_memories.end() && size < MinTableBytes &&
+	    m_sightings.try_emplace(request.base, m_draw).first->second == m_draw) {
+		return false;
+	}
+	// A table follows its descriptor's size, as the first program of a draw acquires it.
+	if (memory == m_memories.end() || ((*memory)->size != size && (*memory)->draw != m_draw)) {
+		if (!m_context.IsMapped(request.base, size)) return false;
+		if (memory == m_memories.end()) {
+			std::scoped_lock lock {m_mutex};
+			memory = m_memories.insert(m_memories.end(),
+			                           std::make_unique<Memory>(Memory {.base = request.base}));
+		}
+		Resize(**memory, size);
+	}
+	// A draw refreshes a table and a view takes its changes once, so all programs of the draw see
+	// one snapshot and one set of kinds.
+	if ((*memory)->draw != m_draw) {
+		if (!Refresh(**memory)) return false;
+		(*memory)->draw = m_draw;
+	}
+	auto&      views   = (*memory)->views;
+	const auto bytes   = (*memory)->size;
+	const auto profile = IR::ImageHeapResource(*request.root, 0u);
+	const auto probes  = Probes(size, request.first, request.step);
+	const auto limit   = request.count < probes ? request.count : UINT64_MAX;
+	auto       view    = std::ranges::find_if(views, [&](const auto& candidate) {
+		return candidate->step == request.step && candidate->first == request.first &&
+		       candidate->limit == limit && candidate->profile == profile;
+	});
+	if (view == views.end()) {
+		const auto epoch   = Image::heap_epoch.load(std::memory_order_relaxed);
+		const auto count   = std::min(limit, Probes(bytes, request.first, request.step));
+		auto       created = std::make_unique<View>(View {.memory        = **memory,
+		                                                  .step          = request.step,
+		                                                  .first         = request.first,
+		                                                  .count         = count,
+		                                                  .limit         = limit,
+		                                                  .profile       = profile,
+		                                                  .resolve_epoch = epoch,
+		                                                  .layout_epoch  = epoch});
+		created->slots.resize(bytes / 4u + 1u);
+		created->buffer = SlotBuffer(m_context, bytes);
+		UpdateProbes(*created, 0u, bytes);
+		// The first upload also clears the slots no probe selects.
+		created->dirty_begin = 0;
+		created->dirty_end   = created->slots.size();
+		view = views.insert(views.end(), std::move(created));
+	} else {
+		// Empty after the first acquire in a draw, since the table refreshes once per draw.
+		UpdateProbes(**view, (*view)->changed_begin, (*view)->changed_end);
+		(*view)->changed_begin = (*view)->changed_end = 0;
+	}
+	if (std::ranges::find(m_active, view->get()) == m_active.end()) m_active.push_back(view->get());
+	const auto& selected = **view;
+	result = {.slots      = selected.buffer->BufferDeviceAddress(),
+	          .slot_count = static_cast<uint32_t>(selected.slots.size() - 1u),
+	          .kinds      = selected.kinds};
+	return true;
+}
+
+bool ImageHeap::Refresh(Memory& memory) {
+	// GPU writes leave no fault; reading the table downloads them.
+	const bool gpu_dirty = m_context.GetBufferCache().HasGpuDirtyBytes(memory.base, memory.size);
+	const auto pages     = memory.dirty_pages.size();
+	m_runs.clear();
+	{
+		std::scoped_lock lock {m_mutex};
+		// Watch before reading, so a write during the read faults the page again.
+		ForEachRun(
+		    0, pages, [&](size_t page) { return memory.dirty_pages[page] != 0u; },
+		    [&](size_t begin, size_t end) {
+			    std::fill(memory.dirty_pages.begin() + begin, memory.dirty_pages.begin() + end,
+			              uint8_t {0});
+			    Watch<true>(memory, begin, end);
+			    m_runs.emplace_back(begin, end);
+		    });
+	}
+	if (gpu_dirty) m_runs.assign(1u, {0u, pages});
+	for (size_t run = 0; run < m_runs.size(); ++run) {
+		const auto [address, end] = PageRange(memory.base, memory.size, m_runs[run].first,
+		                                      m_runs[run].second);
+		m_scratch.resize((end - address) / 4u);
+		if (!LibKernel::Memory::TryReadBufferBacking(address, m_scratch.data(), end - address)) {
+			std::scoped_lock lock {m_mutex};
+			for (; run < m_runs.size(); ++run) {
+				MarkDirty(memory, m_runs[run].first, m_runs[run].second);
+			}
+			return false;
+		}
+		const auto words = std::span(memory.words).subspan((address - memory.base) / 4u,
+		                                                   m_scratch.size());
+		size_t     first = 0;
+		size_t     last  = words.size();
+		while (first < last && words[first] == m_scratch[first]) ++first;
+		while (first < last && words[last - 1u] == m_scratch[last - 1u]) --last;
+		if (first == last) continue;
+		std::copy(m_scratch.begin() + first, m_scratch.begin() + last, words.begin() + first);
+		const auto offset = address - memory.base;
+		for (auto& view: memory.views) {
+			Widen(view->changed_begin, view->changed_end, offset + first * 4u, offset + last * 4u);
+		}
+	}
+	return true;
+}
+
+void ImageHeap::MarkDirty(Memory& memory, size_t begin, size_t end) {
+	ForEachRun(
+	    begin, end, [&](size_t page) { return memory.dirty_pages[page] == 0u; },
+	    [&](size_t first, size_t last) {
+		    std::fill(memory.dirty_pages.begin() + first, memory.dirty_pages.begin() + last,
+		              uint8_t {1});
+		    Watch<false>(memory, first, last);
+	    });
+}
+
+template <bool track>
+void ImageHeap::Watch(const Memory& memory, size_t begin, size_t end) {
+	const auto [address, limit] = PageRange(memory.base, memory.size, begin, end);
+	m_page_manager.UpdatePageWatchers<track>(address, limit - address);
+}
+
+void ImageHeap::Unwatch(const Memory& memory) {
+	ForEachRun(
+	    0, memory.dirty_pages.size(), [&](size_t page) { return memory.dirty_pages[page] == 0u; },
+	    [&](size_t begin, size_t end) { Watch<false>(memory, begin, end); });
+}
+
+void ImageHeap::InvalidateMemory(uint64_t address, uint64_t size) {
+	std::scoped_lock lock {m_mutex};
+	for (auto& memory: m_memories) {
+		// Watches cover whole pages, so a write next to the table must unwatch its page too.
+		const auto origin = Common::AlignDown(memory->base, TRACKER_PAGE_SIZE);
+		const auto begin  = std::max(address, origin);
+		const auto end    = std::min(address + size,
+		                             origin + memory->dirty_pages.size() * TRACKER_PAGE_SIZE);
+		if (begin >= end) continue;
+		MarkDirty(*memory, (begin - origin) / TRACKER_PAGE_SIZE,
+		          (end - 1u - origin) / TRACKER_PAGE_SIZE + 1u);
+	}
+}
+
+void ImageHeap::UpdateProbes(View& view, uint64_t begin, uint64_t end) {
+	// A probe reads its descriptor from first + probe * step.
+	const uint64_t bytes = view.profile.r128 ? 16u : 32u;
+	auto probe = begin < view.first + bytes ? 0u : (begin - view.first - bytes) / view.step + 1u;
+	for (; probe < view.count && view.first + probe * view.step < end; ++probe) {
+		UpdateProbe(view, probe);
+	}
+}
+
+void ImageHeap::UpdateProbe(View& view, uint64_t probe) {
+	const auto&     words  = view.memory.words;
+	const auto      offset = (view.first + probe * view.step) / 4u;
+	DescriptorValue value {.dword_count = 8u};
+	std::copy_n(words.begin() + offset,
+	            std::min<size_t>(view.profile.r128 ? 4u : 8u, words.size() - offset),
+	            value.dwords.begin());
+	// A buried word stays null in every view until its descriptor changes.
+	auto& dead = view.memory.dead[offset];
+	if (dead != 0u && dead != DescriptorHash {}(value)) dead = 0u;
+	const auto kind = dead != 0u ? 0u : IR::ImageHeapKind(view.profile, value);
+	Assign(view, offset, kind == 0u ? 0u : Intern(view, kind, value));
+}
+
+void ImageHeap::Assign(View& view, size_t slot, uint32_t entry) {
+	const auto previous = std::exchange(view.slots[slot], entry);
+	Release(view, previous);
+	if (previous != entry) Widen(view.dirty_begin, view.dirty_end, slot, slot + 1u);
+}
+
+void ImageHeap::Bury() {
+	if (m_unmapped.empty()) return;
+	for (const auto& memory: m_memories) {
+		for (const auto& view: memory->views) {
+			bool buried = false;
+			for (const auto& [value, index]: view->elements) {
+				const auto& binding = m_elements[index].binding;
+				if (binding.image_id && std::ranges::any_of(m_unmapped, [&](GuestRange range) {
+					    return ImageRangeOverlaps(binding.desc.info.data, range);
+				    })) {
+					m_elements[index].stale = buried = true;
+				}
+			}
+			if (!buried) continue;
+			for (uint64_t probe = 0; probe < view->count; ++probe) {
+				const auto  slot    = (view->first + probe * view->step) / 4u;
+				const auto& element = m_elements[view->slots[slot] & ElementMask];
+				if (!element.stale) continue;
+				memory->dead[slot] = DescriptorHash {}(element.value);
+				Assign(*view, slot, 0u);
+			}
+		}
+	}
+	m_unmapped.clear();
+}
+
+uint32_t ImageHeap::Intern(View& view, uint32_t kind, const DescriptorValue& value) {
+	const auto arm = static_cast<uint32_t>(std::ranges::find(view.kind_keys, kind) -
+	                                       view.kind_keys.begin());
+	if (arm == view.kind_keys.size()) {
+		view.kind_keys.push_back(kind);
+		view.kinds.push_back(value);
+	}
+	auto [found, inserted] = view.elements.try_emplace(value, 0u);
+	if (inserted) {
+		found->second    = Allocate();
+		auto& element    = m_elements[found->second];
+		element.value    = value;
+		element.kind     = kind;
+		view.pending.push_back(found->second);
+	}
+	++m_elements[found->second].refs;
+	return arm << IR::ImageHeapKindShift | found->second;
+}
+
+void ImageHeap::Release(View& view, uint32_t entry) {
+	const auto index = entry & ElementMask;
+	if (index == 0u || --m_elements[index].refs != 0u) return;
+	auto& element = m_elements[index];
+	if (const auto found = view.elements.find(element.value);
+	    found != view.elements.end() && found->second == index) {
+		view.elements.erase(found);
+	}
+	Unreference(element);
+	view.released.push_back(index);
+}
+
+void ImageHeap::Unreference(Element& element) {
+	if (auto* image = ImageOf(element)) {
+		image->heap_refs.fetch_sub(1u, std::memory_order_relaxed);
+		image->tick_accessed_last = m_context.GetCommandScheduler().CurrentTick();
+	}
+}
+
+Image* ImageHeap::ImageOf(const Element& element) const {
+	auto* image = m_context.GetTextureCache().m_slot_images.try_get(element.binding.image_id);
+	return image != nullptr && image->serial == element.serial ? image : nullptr;
+}
+
+uint32_t ImageHeap::Allocate() {
+	auto& scheduler = m_context.GetCommandScheduler();
+	while (!m_retired.empty() && scheduler.IsFree(m_retired.front().first)) {
+		m_free.push_back(m_retired.front().second);
+		m_retired.pop_front();
+	}
+	if (m_free.empty()) {
+		if (m_elements.size() == IR::ImageHeapCapacity) {
+			EXIT("image heaps exceed %u resident images\n", IR::ImageHeapCapacity);
+		}
+		m_elements.emplace_back();
+		return static_cast<uint32_t>(m_elements.size() - 1u);
+	}
+	const auto index = m_free.back();
+	m_free.pop_back();
+	m_elements[index] = {};
+	return index;
+}
+
+void ImageHeap::Retire(View& view) {
+	if (view.released.empty()) return;
+	// A retired index is reused once its tick ends, so no held entry may keep it.
+	std::erase_if(view.held, [&](uint32_t index) { return m_elements[index].refs == 0u; });
+	const auto tick = m_context.GetCommandScheduler().CurrentTick();
+	for (const auto index: view.released) m_retired.emplace_back(tick, index);
+	view.released.clear();
+}
+
+void ImageHeap::Resize(Memory& memory, uint64_t size) {
+	{
+		std::scoped_lock lock {m_mutex};
+		Unwatch(memory);
+		memory.size = size;
+		// The next refresh rereads the whole table.
+		memory.dirty_pages.assign((Common::AlignUp(memory.base + size, TRACKER_PAGE_SIZE) -
+		                           Common::AlignDown(memory.base, TRACKER_PAGE_SIZE)) /
+		                              TRACKER_PAGE_SIZE,
+		                          uint8_t {1});
+	}
+	memory.words.resize(size / 4u);
+	memory.dead.resize(size / 4u);
+	auto& scheduler = m_context.GetCommandScheduler();
+	for (auto& view: memory.views) {
+		const auto count = std::min(view->limit, Probes(size, view->first, view->step));
+		for (auto probe = count; probe < view->count; ++probe) {
+			Assign(*view, (view->first + probe * view->step) / 4u, 0u);
+		}
+		view->count       = count;
+		const auto slots  = view->slots.size();
+		view->slots.resize(size / 4u + 1u);
+		view->dirty_end   = std::min(view->dirty_end, view->slots.size());
+		view->dirty_begin = std::min(view->dirty_begin, view->dirty_end);
+		if (view->slots.size() * sizeof(uint32_t) > view->buffer->Size()) {
+			if (scheduler.Active()) {
+				scheduler.DeferOperation([buffer = std::move(view->buffer)] {});
+			}
+			view->buffer = SlotBuffer(m_context, size);
+			Widen(view->dirty_begin, view->dirty_end, 0u, view->slots.size());
+		} else if (slots < view->slots.size()) {
+			// Device words past the old end may hold stale slots.
+			Widen(view->dirty_begin, view->dirty_end, slots, view->slots.size());
+		}
+	}
+}
+
+void ImageHeap::Rebind(RenderExecutor& executor) {
+	KYTY_PROFILER_FUNCTION();
+	auto&      textures = m_context.GetTextureCache();
+	const auto epoch    = Image::heap_epoch.load(std::memory_order_relaxed);
+	const auto serial   = Image::serials.load(std::memory_order_relaxed);
+	for (auto* view: m_active) {
+		bool       stale = false;
+		const auto mark  = [&](uint32_t index) { m_elements[index].stale = stale = true; };
+		if (view->resolve_epoch != epoch) {
+			view->resolve_epoch = epoch;
+			for (const auto& [value, index]: view->elements) {
+				const auto& binding = m_elements[index].binding;
+				if (!binding.image_id) continue;
+				const auto* image = ImageOf(m_elements[index]);
+				if (image == nullptr || !image->registered || image->depth_id ||
+				    image->binding.needs_rebind) {
+					mark(index);
+				} else {
+					(void)textures.FindTexture(binding.image_id, binding.desc);
+				}
+			}
+		}
+		// Discovery applies the color metadata clears written since the last draw.
+		std::erase_if(view->dcc, [&](uint32_t index) {
+			const auto& element = m_elements[index];
+			if (element.refs == 0u || element.stale) return true;
+			if (executor.ResolveTexture(IR::ImageHeapResource(view->profile, element.kind),
+			                            element.value).image_id == element.binding.image_id) {
+				return false;
+			}
+			mark(index);
+			return true;
+		});
+		// Held elements retry once an image is created, such as a render target of their extent.
+		if (view->held_serial != serial) {
+			view->held_serial = serial;
+			std::erase_if(view->held, [&](uint32_t index) {
+				const auto& element = m_elements[index];
+				if (element.refs == 0u || element.stale) return true;
+				const auto binding = executor.ResolveTexture(
+				    IR::ImageHeapResource(view->profile, element.kind), element.value);
+				if (binding.image_id) mark(index);
+				// An unmapped texture stays null.
+				return binding.image_id || binding.desc.info.data.Empty();
+			});
+		}
+		if (stale) {
+			// A changed image needs a new element; pending draws still read the old one. The new
+			// element keeps the old value, since a view takes table changes only when acquired.
+			std::erase_if(view->elements,
+			              [&](const auto& item) { return m_elements[item.second].stale; });
+			for (uint64_t probe = 0; probe < view->count; ++probe) {
+				const auto  slot    = (view->first + probe * view->step) / 4u;
+				const auto& element = m_elements[view->slots[slot] & ElementMask];
+				if (!element.stale) continue;
+				const auto value = element.value; // Intern may grow m_elements.
+				Assign(*view, slot, Intern(*view, element.kind, value));
+			}
+		}
+		for (const auto index: view->pending) {
+			if (m_elements[index].refs == 0u) continue;
+			auto binding = executor.ResolveTexture(
+			    IR::ImageHeapResource(view->profile, m_elements[index].kind),
+			    m_elements[index].value);
+			if (binding.image_id) {
+				auto& image = textures.GetImage(binding.image_id);
+				image.heap_refs.fetch_add(1u, std::memory_order_relaxed);
+				m_elements[index].serial = image.serial;
+				binding.layout     = image.ReadLayout();
+				binding.image_view = textures.FindTexture(binding.image_id, binding.desc);
+				if (binding.desc.info.metadata.kind == ImageMetadataKind::Dcc) {
+					view->dcc.push_back(index);
+				}
+				view->away.push_back(index);
+			} else if (!binding.desc.info.data.Empty()) {
+				view->held.push_back(index);
+			}
+			m_elements[index].binding = std::move(binding);
+			m_writes.push_back(index);
+		}
+		view->pending.clear();
+	}
+	WriteDescriptors();
+}
+
+void ImageHeap::WriteDescriptors() {
+	if (m_writes.empty()) return;
+	std::vector<vk::DescriptorImageInfo> infos;
+	std::vector<vk::WriteDescriptorSet>  writes(m_writes.size());
+	infos.reserve(m_writes.size());
+	for (size_t write = 0; write < m_writes.size(); ++write) {
+		const auto& binding = m_elements[m_writes[write]].binding;
+		infos.push_back({nullptr, binding.image_view, binding.layout});
+		writes[write].dstSet          = m_set;
+		writes[write].dstArrayElement = m_writes[write];
+		writes[write].descriptorCount = 1;
+		writes[write].descriptorType  = vk::DescriptorType::eSampledImage;
+		writes[write].pImageInfo      = &infos.back();
+	}
+	m_context.GetGraphics().device.updateDescriptorSets(static_cast<uint32_t>(writes.size()),
+	                                                    writes.data(), 0, nullptr);
+	m_writes.clear();
+}
+
+void ImageHeap::Commit(CommandBuffer& buffer, vk::PipelineBindPoint bind_point,
+                       vk::PipelineLayout layout) {
+	KYTY_PROFILER_FUNCTION();
+	const auto epoch = Image::heap_epoch.load(std::memory_order_relaxed);
+	for (auto* view: m_active) {
+		if (view->dirty_begin != view->dirty_end) {
+			const auto slots = std::span(view->slots).subspan(view->dirty_begin,
+			                                                  view->dirty_end - view->dirty_begin);
+			auto& stream = m_context.GetBufferCache().GetUtilityBuffer(MemoryUsage::Stream);
+			const auto offset = stream.Copy(slots.data(), slots.size_bytes(), sizeof(uint32_t));
+			view->buffer->CopyFrom(buffer, stream, offset, view->dirty_begin * sizeof(uint32_t),
+			                       slots.size_bytes());
+			view->dirty_begin = view->dirty_end = 0;
+		}
+		Retire(*view);
+		if (view->layout_epoch != epoch) {
+			view->layout_epoch = epoch;
+			view->away.clear();
+			for (const auto& [value, index]: view->elements) view->away.push_back(index);
+		}
+		// This draw's own bindings own the layouts of its attachments and storage images. An image
+		// this draw's discovery replaced lives until the tick ends, and its slots still select it.
+		std::erase_if(view->away, [&](uint32_t index) {
+			const auto& element = m_elements[index];
+			auto*       image   = ImageOf(element);
+			if (element.refs == 0u || image == nullptr) return true;
+			if (image->binding.is_target || image->binding.shader_write ||
+			    image->binding.force_general) {
+				return false;
+			}
+			const auto& view_info = element.binding.desc.view_info;
+			image->Transit(image->ReadLayout(), vk::AccessFlagBits2::eShaderRead,
+			               ImageSubresourceRange {view_info.base_level, view_info.level_count,
+			                                      view_info.base_layer, view_info.layer_count},
+			               buffer.Handle());
+			return true;
+		});
+	}
+	m_active.clear();
+	buffer.Handle().bindDescriptorSets(bind_point, layout, IR::ImageHeapSet, 1, &m_set, 0,
+	                                   nullptr);
+}
+
+void ImageHeap::UnmapMemory(uint64_t address, uint64_t size) {
+	// Descriptors of textures in the range stay null even once it is mapped again.
+	if (!m_unmapped.empty() && m_unmapped.back().End() == address) {
+		m_unmapped.back().size += size;
+	} else {
+		m_unmapped.push_back({address, size});
+	}
+	std::scoped_lock lock {m_mutex};
+	std::erase_if(m_memories, [&](const std::unique_ptr<Memory>& memory) {
+		if (address >= memory->base + memory->size || memory->base >= address + size) {
+			return false;
+		}
+		Drop(*memory);
+		return true;
+	});
+}
+
+void ImageHeap::RunGarbageCollector() {
+	// A smaller table rebuilt each frame moves to a new base. Larger ones keep their dead words.
+	std::scoped_lock lock {m_mutex};
+	std::erase_if(m_memories, [&](const std::unique_ptr<Memory>& memory) {
+		if (memory->size >= MinTableBytes || memory->draw > m_suspend_draw) return false;
+		Drop(*memory);
+		return true;
+	});
+	m_suspend_draw = m_draw;
+	m_sightings.clear();
+}
+
+void ImageHeap::Drop(Memory& memory) {
+	auto& scheduler = m_context.GetCommandScheduler();
+	Unwatch(memory);
+	for (auto& view: memory.views) {
+		std::erase(m_active, view.get());
+		for (const auto& [value, index]: view->elements) {
+			Unreference(m_elements[index]);
+			view->released.push_back(index);
+		}
+		Retire(*view);
+		if (scheduler.Active()) {
+			scheduler.DeferOperation([buffer = std::move(view->buffer)] {});
+		}
+	}
+}
+
+} // namespace Libs::Graphics

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

@@ -0,0 +1,150 @@
+#ifndef EMULATOR_SRC_GRAPHICS_HOST_GPU_RENDERER_IMAGEHEAP_H_
+#define EMULATOR_SRC_GRAPHICS_HOST_GPU_RENDERER_IMAGEHEAP_H_
+
+#include "common/common.h"
+#include "graphics/host_gpu/renderer/pipeline/descriptors.h"
+#include "graphics/shader/recompiler/ir/passes/SrtWalker.h"
+
+#include <deque>
+#include <memory>
+#include <mutex>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+namespace Libs::Graphics {
+
+class Buffer;
+class CommandBuffer;
+class PageManager;
+class RenderContext;
+class RenderExecutor;
+struct ImageHeapTestAccess;
+
+// Guest image tables that shaders index by GPU-computed offsets, kept resident in one bindless
+// sampled-image array. A table is keyed by its base and follows its descriptor's size. Each view
+// of a table owns device slot words, one per table dword, that select an array element. CPU
+// writes to a table fault in its snapshot, which each draw refreshes once; a view takes the
+// changes when a draw first acquires it. Changes to the images behind elements bump
+// Image::heap_epoch. A smaller table leaves at a suspend point if the frame did not acquire it.
+class ImageHeap final: public ShaderRecompiler::IR::ImageHeapSource {
+public:
+	ImageHeap(RenderContext& context, PageManager& page_manager);
+	~ImageHeap();
+	KYTY_CLASS_NO_COPY(ImageHeap);
+
+	// Starts a draw or dispatch, before its programs acquire views.
+	void BeginDraw() {
+		++m_draw;
+		Bury();
+	}
+	bool Acquire(const ShaderRecompiler::IR::ImageHeapRequest& request,
+	             ShaderRecompiler::IR::ImageHeapView&          result) override;
+	// Image discovery: resolves new elements and rediscovers changed images.
+	void Rebind(RenderExecutor& executor);
+	// Recording: uploads slots, moves resident images to their read layout and binds the array.
+	void Commit(CommandBuffer& buffer, vk::PipelineBindPoint bind_point, vk::PipelineLayout layout);
+	// CPU write fault, on any thread.
+	void InvalidateMemory(uint64_t address, uint64_t size);
+	void UnmapMemory(uint64_t address, uint64_t size);
+	// At a suspend point.
+	void RunGarbageCollector();
+
+private:
+	using DescriptorValue = ShaderRecompiler::IR::DescriptorValue;
+
+	struct DescriptorHash {
+		size_t operator()(const DescriptorValue& value) const noexcept;
+	};
+
+	struct Element {
+		DescriptorValue value;
+		TextureBinding  binding;
+		uint32_t        kind   = 0;
+		uint32_t        refs   = 0; // Slots of the owning view that select it.
+		bool            stale  = false;
+		uint64_t        serial = 0; // Image::serial of binding.image_id.
+	};
+
+	struct Memory;
+
+	struct View {
+		Memory&                             memory;
+		uint64_t                            step  = 0;
+		uint64_t                            first = 0;
+		uint64_t                            count = 0;
+		uint64_t                            limit = 0; // Probes a key domain allows, if fewer.
+		ShaderRecompiler::IR::ImageResource profile;
+		std::vector<uint32_t>               slots; // Plus a trailing null slot.
+		std::vector<uint32_t>               kind_keys;
+		std::vector<DescriptorValue>        kinds; // One representative per kind_keys entry.
+		std::unordered_map<DescriptorValue, uint32_t, DescriptorHash> elements;
+		std::vector<uint32_t>   pending;  // Not yet resolved.
+		std::vector<uint32_t>   released; // No slot selects them after the next upload.
+		std::vector<uint32_t>   away;     // Not yet moved to the read layout.
+		std::vector<uint32_t>   dcc;      // Color metadata clears apply on each discovery.
+		std::vector<uint32_t>   held;     // Another extent holds their address.
+		std::unique_ptr<Buffer> buffer;
+		size_t                  dirty_begin   = 0; // Slots to upload.
+		size_t                  dirty_end     = 0;
+		size_t                  changed_begin = 0; // Table bytes changed since it was probed.
+		size_t                  changed_end   = 0;
+		uint64_t                resolve_epoch = 0;
+		uint64_t                layout_epoch  = 0;
+		uint64_t                held_serial   = 0; // Image::serials when held was last retried.
+	};
+
+	struct Memory {
+		uint64_t                           base = 0;
+		uint64_t                           size = 0;
+		std::vector<uint32_t>              words;
+		std::vector<uint8_t>               dirty_pages; // Unwatched, so unread since a write.
+		std::vector<uint64_t>              dead;        // Per word, a buried descriptor's hash.
+		uint64_t                           draw = 0;    // Last draw that refreshed it.
+		std::vector<std::unique_ptr<View>> views;
+	};
+
+	[[nodiscard]] bool Refresh(Memory& memory);
+	void               MarkDirty(Memory& memory, size_t begin, size_t end);
+	template <bool track>
+	void     Watch(const Memory& memory, size_t begin, size_t end);
+	void     Unwatch(const Memory& memory);
+	void     Drop(Memory& memory);
+	void     UpdateProbes(View& view, uint64_t begin, uint64_t end);
+	void     UpdateProbe(View& view, uint64_t probe);
+	void     Assign(View& view, size_t slot, uint32_t entry);
+	void     Bury();
+	uint32_t Intern(View& view, uint32_t kind, const DescriptorValue& value);
+	void     Release(View& view, uint32_t entry);
+	void     Unreference(Element& element);
+	Image*   ImageOf(const Element& element) const;
+	uint32_t Allocate();
+	void     Retire(View& view);
+	void     Resize(Memory& memory, uint64_t size);
+	void     WriteDescriptors();
+
+	RenderContext&                       m_context;
+	PageManager&                         m_page_manager;
+	std::mutex                           m_mutex; // Fault-visible Memory state and m_memories.
+	std::vector<std::unique_ptr<Memory>> m_memories;
+	std::vector<View*>                   m_active; // Acquired since the last commit.
+	uint64_t                             m_draw         = 1;
+	uint64_t                             m_suspend_draw = 0; // m_draw at the last suspend point.
+	std::vector<Element>                 m_elements;
+	std::vector<uint32_t>                m_free;
+	std::unordered_map<uint64_t, uint64_t>    m_sightings; // Smaller tables: base, first draw.
+	std::deque<std::pair<uint64_t, uint32_t>> m_retired; // Tick, element.
+	std::vector<std::pair<size_t, size_t>>    m_runs;
+	std::vector<uint32_t>                     m_scratch;
+	std::vector<uint32_t>                     m_writes; // Elements whose descriptors changed.
+	std::vector<GuestRange>                   m_unmapped; // Since the last Bury.
+	vk::DescriptorSetLayout                   m_layout = nullptr;
+	vk::DescriptorPool                        m_pool   = nullptr;
+	vk::DescriptorSet                         m_set    = nullptr;
+
+	friend struct ImageHeapTestAccess;
+};
+
+} // namespace Libs::Graphics
+
+#endif // EMULATOR_SRC_GRAPHICS_HOST_GPU_RENDERER_IMAGEHEAP_H_

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

@@ -4,31 +4,48 @@
 #include "common/abi.h"
 #include "common/assert.h"
 #include "common/common.h"
+#include "common/file.h"
+#include "common/uniqueFunction.h"
+#include "graphics/host_gpu/renderer/colorRenderTarget.h"
+#include "graphics/host_gpu/renderer/depthRenderTarget.h"
 #include "graphics/host_gpu/renderer/renderTarget.h"
 #include "graphics/host_gpu/vulkanCommon.h"
+#include "graphics/shader/recompiler/ir/passes/ResourceMaterialization.h"
+#include "graphics/shader/rectListShader.h"
 #include "graphics/shader/shader.h"
 
 #include <chrono>
+#include <condition_variable>
 #include <cstddef>
 #include <filesystem>
 #include <memory>
+#include <mutex>
+#include <queue>
 #include <span>
+#include <thread>
 #include <type_traits>
 #include <unordered_map>
+#include <vector>
 
 namespace Libs::Graphics {
 
 struct GraphicContext;
-struct RenderColorInfo;
-struct RenderDepthInfo;
+struct ShaderParams;
 class CommandBuffer;
 
-namespace HW {
-class Context;
-class Shader;
-class UserConfig;
-struct ComputeShaderInfo;
-} // namespace HW
+namespace ShaderRecompiler {
+struct CompileOptions;
+struct ShaderSource;
+struct TranslateResult;
+} // namespace ShaderRecompiler
+
+namespace ShaderRecompiler::IR {
+class ImageHeapSource;
+} // namespace ShaderRecompiler::IR
+
+namespace ShaderRecompiler::Spirv {
+struct Profile;
+} // namespace ShaderRecompiler::Spirv
 
 #pragma pack(push, 1)
 
@@ -36,14 +53,13 @@ struct PipelineStaticParameters {
 	bool                       negative_one_to_one      = false;
 	bool                       depth_clip_enable        = true;
 	vk::PrimitiveTopology      topology                 = vk::PrimitiveTopology::ePointList;
+	// 0 without tessellation, 3 for a rectangle list, else the input control points.
+	uint32_t                   patch_control_points     = 0;
 	bool                       primitive_restart_enable = false;
 	uint32_t                   samples                  = 1;
 	bool                       sample_shading_enable    = false;
 	bool                       alpha_to_coverage_enable = false;
 	uint32_t                   color_mask[RENDER_COLOR_ATTACHMENTS_MAX]           = {};
-	bool                       cull_front                                         = false;
-	bool                       cull_back                                          = false;
-	bool                       face                                               = false;
 	bool                       provoking_vtx_last                                 = false;
 	vk::PolygonMode            polygon_mode                                       = vk::PolygonMode::eFill;
 	uint8_t                    color_srcblend[RENDER_COLOR_ATTACHMENTS_MAX]       = {};
@@ -54,57 +70,110 @@ struct PipelineStaticParameters {
 	uint8_t                    alpha_destblend[RENDER_COLOR_ATTACHMENTS_MAX]      = {};
 	bool                       separate_alpha_blend[RENDER_COLOR_ATTACHMENTS_MAX] = {};
 	bool                       blend_enable[RENDER_COLOR_ATTACHMENTS_MAX]         = {};
-
-	bool operator==(const PipelineStaticParameters& other) const noexcept;
 };
 
-#pragma pack(pop)
-
-static_assert(std::is_trivially_copyable_v<PipelineStaticParameters>);
-static_assert(std::is_standard_layout_v<PipelineStaticParameters>);
-static_assert(alignof(PipelineStaticParameters) == 1);
-static_assert(sizeof(PipelineStaticParameters) == 117);
-
 struct PipelineRenderingState {
 	std::array<vk::Format, RENDER_COLOR_ATTACHMENTS_MAX> color_formats {};
 	vk::Format                                           depth_format   = vk::Format::eUndefined;
 	vk::Format                                           stencil_format = vk::Format::eUndefined;
 	uint32_t                                             color_count    = 0;
-
-	bool operator==(const PipelineRenderingState&) const = default;
 };
 
 struct PipelineVertexInputState {
 	struct Binding {
-		uint32_t stride                           = 0;
-		bool     instance                         = false;
-		bool     operator==(const Binding&) const = default;
+		uint32_t stride   = 0;
+		uint8_t  instance = 0;
 	};
 	struct Attribute {
-		uint32_t offset                             = 0;
-		uint8_t  binding                            = 0;
-		bool     operator==(const Attribute&) const = default;
+		uint32_t offset  = 0;
+		uint8_t  binding = 0;
 	};
 
 	std::array<Binding, ShaderVertexInputInfo::RES_MAX>   bindings {};
 	std::array<Attribute, ShaderVertexInputInfo::RES_MAX> attributes {};
 	uint8_t                                               binding_count   = 0;
 	uint8_t                                               attribute_count = 0;
-
-	bool operator==(const PipelineVertexInputState&) const = default;
 };
 
-struct ShaderProgram {
-	uint64_t         id     = 0;
-	vk::ShaderModule module = nullptr;
+// Everything a graphics pipeline is made of besides the stage modules its ids name. Aligned so
+// that rendering.color_formats can be passed to Vulkan.
+struct alignas(4) GraphicsPipelineKey {
+	// VS, LS or mesh; HS; TES; PS. 0 marks an absent stage.
+	std::array<uint64_t, 4>  stage_ids {};
+	PipelineRenderingState   rendering;
+	PipelineVertexInputState vertex_input;
+	PipelineStaticParameters static_params;
 
-	explicit operator bool() const { return id != 0 && module != nullptr; }
+	bool operator==(const GraphicsPipelineKey& other) const;
+};
+
+#pragma pack(pop)
+
+// Keys are compared and hashed as bytes.
+static_assert(std::has_unique_object_representations_v<GraphicsPipelineKey>);
+
+// A compiled stage with what pipeline creation needs from it besides the key.
+struct ShaderModule {
+	uint64_t                                     id     = 0;
+	vk::ShaderModule                             module = nullptr;
+	ShaderRecompiler::IR::CompiledShaderInfo     info;
+	ShaderRecompiler::IR::ResourceSpecialization specialization;
+	std::vector<uint32_t>                        function_code;
+	// VS or LS: the host format of each vertex input attribute.
+	std::vector<vk::Format>                      vertex_formats;
+	// Rectangle-list PS: the parameters it reads.
+	std::vector<RectListParameter>               rect_parameters;
+};
+
+// A draw's topology, shader stages (from PipelineCache::GetGraphicsPrograms) and targets.
+struct DrawRenderState {
+	RenderDepthInfo                      depth_info;
+	RenderColorInfo                      color_info[RENDER_COLOR_ATTACHMENTS_MAX] = {};
+	uint32_t                             color_count                              = 0;
+	// Color targets the pixel program writes.
+	uint32_t                             mrt_mask                                 = 0;
+	vk::PrimitiveTopology                topology = vk::PrimitiveTopology::ePointList;
+	bool                                 primitive_restart_enable                 = false;
+	bool                                 ps_active                                = false;
+	std::array<ShaderVertexInputInfo, 3> vertex_info;
+	ShaderPixelInputInfo                 ps_input_info;
+	// Input: the user-data slots a GPU-read indirect draw patches.
+	ShaderIndirectOffsets                indirect_offsets;
+
+	[[nodiscard]] std::span<const ShaderVertexInputInfo> VertexStages() const {
+		return {vertex_info.data(), vertex_info[1].stage ? 3u : 1u};
+	}
+};
+
+// Writes the shader store on its own thread, running jobs in the order they were posted.
+class ShaderStoreWriter {
+public:
+	// Keeps the first `kept` bytes of the store at `path`, or starts it over with `header`. The
+	// thread runs only if that succeeded.
+	void Open(const std::filesystem::path& path, std::span<const uint8_t> header, uint64_t kept);
+	explicit operator bool() const { return m_thread.joinable(); }
+	// Compresses `record` and appends it on the store thread, in call order.
+	void Append(std::vector<uint8_t>&& record);
+
+private:
+	friend struct PipelineCacheTestAccess;
+
+	void Post(Common::UniqueFunction<void>&& job);
+
+	Common::File                             m_file;
+	std::mutex                               m_mutex;
+	std::condition_variable_any              m_ready;
+	std::queue<Common::UniqueFunction<void>> m_jobs;
+	// Declared last, so it drains the queue and joins before the rest is destroyed.
+	std::jthread m_thread;
 };
 
 // The owning renderer serializes access, including saves while the GPU is running.
 class PipelineCache {
 public:
-	explicit PipelineCache(GraphicContext& graphics);
+	// Without a shader store path nothing is kept across runs.
+	PipelineCache(GraphicContext& graphics, ShaderRecompiler::IR::ImageHeapSource* image_heap,
+	              const std::filesystem::path& shader_store = {});
 	~PipelineCache();
 	KYTY_CLASS_NO_COPY(PipelineCache);
 	void Save();
@@ -114,89 +183,109 @@ public:
 		vk::Pipeline            pipeline              = nullptr;
 		vk::DescriptorSetLayout descriptor_set_layout = nullptr;
 		bool                    uses_push_descriptors = false;
+		bool                    uses_image_heap       = false;
 	};
 
-	struct GraphicsPrograms {
-		std::array<ShaderProgram, 3> vertex;
-		ShaderProgram pixel;
+	// Reads the shaders and registers bound in `command` into the state's shader stages.
+	void GetGraphicsPrograms(const CommandBuffer& command, DrawRenderState& state);
+	// Null skips the dispatch.
+	const ShaderModule* GetComputeProgram(const HW::ComputeShaderInfo& regs,
+	                                      ShaderComputeInputInfo&      input_info);
 
-		[[nodiscard]] uint32_t VertexStageCount() const { return vertex[1] ? 3u : 1u; }
-	};
-
-	GraphicsPrograms
-	GetGraphicsPrograms(const HW::VertexShaderInfo& vertex_regs,
-	                    const HW::PixelShaderInfo& pixel_regs, const HW::ShaderRegisters& sh,
-	                    const HW::Context& context, const HW::UserConfig& user_config,
-	                    std::span<const Prospero::ColorComponentMapping, 8> target_export_mapping,
-	                    bool pixel_active, std::array<ShaderVertexInputInfo, 3>& vertex_info,
-	                    ShaderPixelInputInfo& pixel_info);
-	ShaderProgram GetComputeProgram(const HW::ComputeShaderInfo& regs,
-	                                const HW::ShaderRegisters&   sh,
-	                                ShaderComputeInputInfo&      input_info);
-
-	Pipeline& GetGraphicsPipeline(std::span<const RenderColorInfo>       colors,
-	                              const RenderDepthInfo&                 depth,
-	                              std::span<const ShaderVertexInputInfo> vertex_info,
-	                              CommandBuffer& command, const ShaderPixelInputInfo* ps_input_info,
-	                              vk::PrimitiveTopology topology, bool primitive_restart_enable,
-	                              const GraphicsPrograms& programs);
-	Pipeline& GetComputePipeline(const ShaderComputeInputInfo& input_info,
-	                             const ShaderProgram&          compute_program);
+	Pipeline& GetGraphicsPipeline(const CommandBuffer& command, const DrawRenderState& state);
+	Pipeline& GetComputePipeline(const ShaderComputeInputInfo& input_info);
 
 private:
-	struct ProgramCache;
+	friend struct PipelineCacheTestAccess;
 
-	struct GraphicsPipelineKey {
-		PipelineRenderingState   rendering;
-		std::array<uint64_t, 3>  vertex_shader_ids {};
-		uint64_t                 ps_shader_id = 0;
-		PipelineVertexInputState vertex_input;
-		PipelineStaticParameters static_params;
+	struct Program {
+		std::unique_ptr<ShaderRecompiler::ShaderSource> call_source;
+		ShaderRecompiler::IR::ResourcePlan           resource_plan;
+		ShaderRecompiler::IR::ResourceSnapshot       resources;
+		ShaderRecompiler::IR::ResourceSpecialization specialization;
+		std::vector<ShaderModule>                    modules;
+		bool                                         skip_dispatch = false;
 
-		bool operator==(const GraphicsPipelineKey& other) const {
-			return rendering == other.rendering && vertex_shader_ids == other.vertex_shader_ids &&
-			       ps_shader_id == other.ps_shader_id && vertex_input == other.vertex_input &&
-			       static_params == other.static_params;
-		}
+		[[nodiscard]] std::span<const uint32_t> FunctionCode() const;
+		bool RefreshCallSource(const ShaderRecompiler::IR::SrtRuntime& runtime);
+		template <bool HasCalls>
+		ShaderModule* FindModule(const ShaderRecompiler::IR::SrtRuntime& runtime,
+		                         uint32_t                                push_data_cursor);
 	};
 
-	struct PipelineKeyHash {
-		static void Mix(std::size_t& hash, std::size_t value) {
-			hash ^= value + static_cast<std::size_t>(0x9e3779b97f4a7c15ull) + (hash << 6u) +
-			        (hash >> 2u);
-		}
+	struct ProgramKey {
+		ShaderType            stage           = ShaderType::Unknown;
+		uint64_t              hash            = 0;
+		uint32_t              user_data_count = 0;
+		std::vector<uint32_t> static_state;
+
+		bool operator==(const ProgramKey&) const = default;
+	};
+
+	// Variants of one source share a bucket; ProgramKey equality compares the state.
+	struct ProgramKeyHash {
+		std::size_t operator()(const ProgramKey& key) const { return key.hash; }
 	};
 
 	struct GraphicsPipelineKeyHash {
 		std::size_t operator()(const GraphicsPipelineKey& key) const;
 	};
 
-	GraphicContext&               m_graphics;
-	std::unique_ptr<ProgramCache> m_program_cache;
+	using ModuleMap = std::unordered_map<uint64_t, const ShaderModule*>;
+
+	GraphicContext&                        m_graphics;
+	ShaderRecompiler::IR::ImageHeapSource* m_image_heap;
+	std::unordered_map<ProgramKey, Program, ProgramKeyHash> m_program_cache;
+	// Modules compiled per stage.
+	std::array<uint32_t, static_cast<size_t>(ShaderType::TessellationEvaluation) + 1> m_compiled {};
+	ProgramKey                             m_lookup_key;
 	vk::PipelineCache             m_driver_cache = nullptr;
 	std::filesystem::path         m_driver_cache_path;
 	std::chrono::steady_clock::time_point m_next_driver_cache_save;
 	bool                          m_driver_cache_dirty = false;
-	std::unordered_map<GraphicsPipelineKey, std::unique_ptr<Pipeline>, GraphicsPipelineKeyHash>
-	                                                        m_graphics_pipelines;
-	std::unordered_map<uint64_t, std::unique_ptr<Pipeline>> m_compute_pipelines;
+	std::unordered_map<GraphicsPipelineKey, Pipeline, GraphicsPipelineKeyHash> m_graphics_pipelines;
+	std::unordered_map<uint64_t, Pipeline>                                     m_compute_pipelines;
+	std::unordered_map<uint64_t, ShaderTessellationInputInfo>                  m_tess_layouts;
+	// Records everything first created; closed without a shader store.
+	ShaderStoreWriter m_store;
 
+	template <typename InputInfo>
+	const ShaderModule* GetProgram(const ShaderParams& params, InputInfo& input_info,
+	                               uint32_t& push_data_cursor);
+	ShaderModule& CompileModule(Program& program, const ProgramKey& key,
+	                            ShaderRecompiler::CompileOptions  options,
+	                            ShaderRecompiler::TranslateResult translated,
+	                            uint32_t                          push_data_cursor);
 	void InitializeDriverCache();
 	void CheckpointDriverCache();
+
+	// pipelineSerialization.cpp
+	void WarmUp(const std::filesystem::path& path);
+	bool LoadRecord(std::span<const uint8_t>& records, std::vector<GraphicsPipelineKey>& pipelines);
+	void StoreModule(const ProgramKey& key, const Program& program, const ShaderModule& module,
+	                 const std::vector<uint32_t>& spirv);
+	void StorePipeline(const GraphicsPipelineKey& key);
+	void StoreLayout(uint64_t key, const ShaderTessellationInputInfo& layout);
+	// Every module by id, with id 0 naming an absent stage.
+	[[nodiscard]] ModuleMap MapModules() const;
+	// False when a stage id of the key is not in `modules`.
+	bool CreateGraphicsPipeline(Pipeline& pipeline, const GraphicsPipelineKey& key,
+	                            const ModuleMap& modules);
 };
 
-void LogPipelineTrace(const char* phase, uint64_t vertex_program_id, uint64_t pixel_program_id);
+// The shader store of the running title; empty unless --shader-cache is on in a Release build
+// without debug dumps, which change the SPIR-V.
+std::filesystem::path ShaderStorePath();
+// Host capabilities that change the emitted SPIR-V.
+ShaderRecompiler::Spirv::Profile SpirvProfile(const GraphicContext& graphics);
+
+// Stages are indexed like GraphicsPipelineKey::stage_ids.
 void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& pipeline,
-                            const PipelineRenderingState&          rendering,
-                            const PipelineVertexInputState&        vertex_input,
-                            std::span<const ShaderVertexInputInfo> vertex_info,
-                            const ShaderPixelInputInfo*            ps_input_info,
-                            const PipelineCache::GraphicsPrograms& programs,
-                            const PipelineStaticParameters&        static_params,
-                            vk::PipelineCache                      driver_cache);
+                            const GraphicsPipelineKey&                 key,
+                            const std::array<const ShaderModule*, 4>& stages,
+                            vk::PipelineCache                          driver_cache);
 void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& pipeline,
-                            const ShaderComputeInputInfo& input_info,
-                            vk::ShaderModule compute_module, vk::PipelineCache driver_cache);
+                            const ShaderModule& module, vk::PipelineCache driver_cache);
 
 } // namespace Libs::Graphics
 

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

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

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

@@ -1,6 +1,7 @@
 #include "graphics/host_gpu/renderer/pipeline/shaderResourceBarrier.h"
 
 #include "common/assert.h"
+#include "graphics/host_gpu/renderer/pipeline/pipelineCache.h"
 #include "graphics/shader/shader.h"
 #include "graphics/shader/shaderBindings.h"
 
@@ -52,6 +53,7 @@ vk::MemoryBarrier MakeShaderWriteDependency() {
 	barrier.dstAccessMask = vk::AccessFlagBits::eShaderRead | vk::AccessFlagBits::eShaderWrite |
 	                        vk::AccessFlagBits::eVertexAttributeRead |
 	                        vk::AccessFlagBits::eIndexRead | vk::AccessFlagBits::eUniformRead |
+	                        vk::AccessFlagBits::eIndirectCommandRead |
 	                        vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite |
 	                        vk::AccessFlagBits::eColorAttachmentRead |
 	                        vk::AccessFlagBits::eColorAttachmentWrite;
@@ -89,7 +91,7 @@ vk::BufferMemoryBarrier MakeGdsDependency(vk::Buffer buffer) {
 
 bool HasShaderBufferWrites(const ShaderStageRuntime& runtime) {
 	EXIT_IF(!runtime);
-	const auto& program   = *runtime.program;
+	const auto& program   = runtime.module->info;
 	const auto& resources = *runtime.resources;
 	EXIT_IF(resources.buffers.size() != program.info.buffers.size());
 	// Address writes reach memory outside the bound buffer descriptors.

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

@@ -13,7 +13,6 @@
 #include "graphics/host_gpu/renderer/renderContext.h"
 #include "graphics/host_gpu/renderer/renderTarget.h"
 #include "graphics/host_gpu/vulkanCommon.h"
-#include "graphics/shader/recompiler/BufferFormat.h"
 #include "graphics/shader/recompiler/ir/ShaderIR.h"
 #include "graphics/shader/rectListShader.h"
 #include "graphics/shader/shader.h"
@@ -21,124 +20,11 @@
 #include <algorithm>
 #include <array>
 #include <bit>
-#include <limits>
 #include <span>
 #include <vector>
 
 namespace Libs::Graphics {
 
-// IDK: maybe we can remove it?
-constexpr uint8_t kTemporaryVertexAttribFormat113 =
-    static_cast<uint8_t>(Prospero::VertexAttribFormat::k16_16SInt);
-constexpr uint32_t kTemporaryPs5BufferFormat121 = 121u;
-
-static bool NarrowInputFormat(vk::Format& format, uint32_t& size, uint32_t used_components) {
-	if (used_components == 0 || used_components >= size) {
-		return false;
-	}
-
-	switch (format) {
-		case vk::Format::eR32G32B32A32Sfloat:
-			switch (used_components) {
-				case 1: format = vk::Format::eR32Sfloat; break;
-				case 2: format = vk::Format::eR32G32Sfloat; break;
-				case 3: format = vk::Format::eR32G32B32Sfloat; break;
-				default: return false;
-			}
-			size = used_components;
-			return true;
-		case vk::Format::eR32G32B32Sfloat:
-			switch (used_components) {
-				case 1: format = vk::Format::eR32Sfloat; break;
-				case 2: format = vk::Format::eR32G32Sfloat; break;
-				default: return false;
-			}
-			size = used_components;
-			return true;
-		case vk::Format::eR16G16B16A16Sfloat:
-			switch (used_components) {
-				case 1: format = vk::Format::eR16Sfloat; break;
-				case 2: format = vk::Format::eR16G16Sfloat; break;
-				default: return false;
-			}
-			size = used_components;
-			return true;
-		case vk::Format::eR8G8B8A8Unorm:
-			switch (used_components) {
-				case 1: format = vk::Format::eR8Unorm; break;
-				case 2: format = vk::Format::eR8G8Unorm; break;
-				default: return false;
-			}
-			size = used_components;
-			return true;
-		case vk::Format::eR8G8B8A8Snorm:
-			if (used_components != 2) {
-				return false;
-			}
-			format = vk::Format::eR8G8Snorm;
-			size   = 2;
-			return true;
-		case vk::Format::eR8G8B8A8Uint:
-			switch (used_components) {
-				case 1: format = vk::Format::eR8Uint; break;
-				case 2: format = vk::Format::eR8G8Uint; break;
-				default: return false;
-			}
-			size = used_components;
-			return true;
-		default: break;
-	}
-
-	return false;
-}
-
-static void GetInputFormat(const ShaderBufferResource& res, vk::Format& format, uint32_t& size,
-                           uint32_t used_components) {
-	const auto fmt        = res.Format();
-	const auto raw_format = res.RawFormat();
-	if (raw_format == kTemporaryVertexAttribFormat113) {
-		static bool logged_113 = false;
-		if (!logged_113) {
-			LOGF("InputFormat: temporary: accepting invalid PS5 buffer format 113 as "
-			     "vk::Format::eR32G32B32A32Sfloat\n");
-			logged_113 = true;
-		}
-		format = vk::Format::eR32G32B32A32Sfloat;
-		size   = 4;
-		if (NarrowInputFormat(format, size, used_components)) {
-			LOGF("InputFormat: narrowing fmt=%u to %s for used_components=%u\n", raw_format,
-			     vk::to_string(format).c_str(), used_components);
-		}
-		return;
-	}
-	if (raw_format == kTemporaryPs5BufferFormat121) {
-		static bool logged_121 = false;
-		if (!logged_121) {
-			LOGF("InputFormat: accepting PS5 buffer format 121 as vk::Format::eR16G16Sfloat\n");
-			logged_121 = true;
-		}
-		format = vk::Format::eR16G16Sfloat;
-		size   = 2;
-		return;
-	}
-
-	format = VulkanFormat(fmt);
-	size   = ShaderRecompiler::Format::GetFormatInfo(fmt).component_count;
-	if (format == vk::Format::eUndefined || size == 0) {
-		EXIT("unknown vertex format: fmt = %u\n", raw_format);
-	}
-
-	if (NarrowInputFormat(format, size, used_components)) {
-		static std::atomic<uint64_t> log_count = 0;
-		auto                         log_id    = log_count.fetch_add(1);
-		if (log_id < 32) {
-			LOGF("VertexInput: narrowed vertex format to %" PRIu32
-			     " component(s) for shader fetch\n",
-			     used_components);
-		}
-	}
-}
-
 static vk::BlendFactor GetBlendFactor(uint32_t factor) {
 	switch (static_cast<Prospero::BlendFactor>(factor)) {
 		case Prospero::BlendFactor::kZero: return vk::BlendFactor::eZero;
@@ -242,33 +128,27 @@ static void CreateDescriptorLayout(GraphicContext& graphics, PipelineCache::Pipe
 
 // NOLINTNEXTLINE(readability-function-cognitive-complexity)
 void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& pipeline,
-                            const PipelineRenderingState&          rendering,
-                            const PipelineVertexInputState&        vertex_input,
-                            std::span<const ShaderVertexInputInfo> vertex_info,
-                            const ShaderPixelInputInfo*            ps_input_info,
-                            const PipelineCache::GraphicsPrograms& programs,
-                            const PipelineStaticParameters&        static_params,
-                            vk::PipelineCache                      driver_cache) {
-	const auto& vs_input_info  = vertex_info.front();
-	const auto& vertex_program = programs.vertex[0];
-	const auto& pixel_program  = programs.pixel;
-	const bool  tessellation   = vertex_info.size() == 3;
-	const bool ps_active = ps_input_info != nullptr;
-	EXIT_IF(!vertex_program || (ps_active && !pixel_program));
-	const bool with_depth = rendering.depth_format != vk::Format::eUndefined ||
-	                        rendering.stencil_format != vk::Format::eUndefined;
-	EXIT_IF(!vs_input_info.stage);
-	const bool mesh = vs_input_info.stage.program->stage == ShaderType::Mesh;
+                            const GraphicsPipelineKey&                 key,
+                            const std::array<const ShaderModule*, 4>& stages,
+                            vk::PipelineCache                          driver_cache) {
+	const auto& rendering     = key.rendering;
+	const auto& vertex_input  = key.vertex_input;
+	const auto& static_params = key.static_params;
+	const auto& vertex        = *stages[0];
+	const auto* pixel         = stages[3];
+	const bool  with_depth    = rendering.depth_format != vk::Format::eUndefined ||
+	                            rendering.stencil_format != vk::Format::eUndefined;
+	const bool  mesh          = vertex.info.stage == ShaderType::Mesh;
 	EXIT_NOT_IMPLEMENTED(mesh && !graphics.mesh_shader_enabled);
-	const bool rect_list =
-	    !mesh && !tessellation && static_params.topology == vk::PrimitiveTopology::ePatchList;
+	const bool rect_list = static_params.patch_control_points != 0 && stages[1] == nullptr;
 
 	vk::ShaderModule tess_control_shader_module = nullptr;
 	vk::ShaderModule tess_eval_shader_module    = nullptr;
 
 	if (rect_list) {
-		const auto shaders =
-		    BuildRectListShaders(vs_input_info, ps_active ? ps_input_info : nullptr);
+		std::span<const RectListParameter> parameters;
+		if (pixel != nullptr) parameters = pixel->rect_parameters;
+		const auto shaders = BuildRectListShaders(parameters, vertex.info.param_export_mask);
 		tess_control_shader_module = CompileSPV(shaders.control, graphics.device);
 		if (graphics_debug_dump_enabled()) {
 			LOGF("PipelineTrace: vkCreateShaderModule RectList TCS done module=%p\n",
@@ -287,11 +167,12 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 
 	vk::PipelineShaderStageCreateInfo shader_stages[4] {};
 	uint32_t                          shader_stage_count = 0;
-	for (uint32_t i = 0; i < vertex_info.size(); i++) {
-		shader_stages[shader_stage_count++] = {.stage =
-		                                           NativeShaderStage(vertex_info[i].logical_stage),
-		                                       .module = programs.vertex[i].module,
-		                                       .pName  = "main"};
+	for (const auto* stage: stages) {
+		if (stage != nullptr) {
+			shader_stages[shader_stage_count++] = {.stage  = NativeShaderStage(stage->info.stage),
+			                                       .module = stage->module,
+			                                       .pName  = "main"};
+		}
 	}
 	if (rect_list) {
 		shader_stages[shader_stage_count++] = {.stage =
@@ -303,11 +184,6 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 		                                       .module = tess_eval_shader_module,
 		                                       .pName  = "main"};
 	}
-	if (ps_active) {
-		shader_stages[shader_stage_count++] = {.stage  = vk::ShaderStageFlagBits::eFragment,
-		                                       .module = pixel_program.module,
-		                                       .pName  = "main"};
-	}
 
 	vk::VertexInputAttributeDescription input_attr[ShaderVertexInputInfo::RES_MAX] {};
 	vk::VertexInputBindingDescription   input_desc[ShaderVertexInputInfo::RES_MAX] {};
@@ -323,46 +199,7 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 		input_attr[index].binding  = vertex_input.attributes[index].binding;
 		input_attr[index].location = index;
 		input_attr[index].offset   = vertex_input.attributes[index].offset;
-
-		uint32_t   attr_size     = 4;
-		const auto registers_num = vs_input_info.resources_dst[index].registers_num;
-		const auto compiled_components =
-		    vs_input_info.stage.program->info.vertex_fetch_components[index];
-		const auto used_components =
-		    compiled_components > 0 ? static_cast<int>(compiled_components) : registers_num;
-		GetInputFormat(vs_input_info.resources[index], input_attr[index].format, attr_size,
-		               static_cast<uint32_t>(used_components));
-
-		if (graphics_debug_dump_enabled()) {
-			static std::atomic_uint log_count = 0;
-			const auto              log_id    = log_count.fetch_add(1, std::memory_order_relaxed);
-			if (log_id < 128) {
-				LOGF("VertexInputState[%u]: attr=%u binding=%u offset=%u stride=%u fmt=%d "
-				     "src_fmt=%u dst=v%u regs=%u"
-				     " fetched_components=%u attr_size=%u swizzle=%u,%u,%u,%u\n",
-				     log_id, index, input_attr[index].binding, input_attr[index].offset,
-				     input_desc[input_attr[index].binding].stride,
-				     static_cast<int>(input_attr[index].format),
-				     static_cast<uint32_t>(vs_input_info.resources[index].Format()),
-				     static_cast<uint32_t>(vs_input_info.resources_dst[index].register_start),
-				     static_cast<uint32_t>(registers_num), static_cast<uint32_t>(used_components),
-				     attr_size, static_cast<uint32_t>(vs_input_info.resources[index].DstSelX()),
-				     static_cast<uint32_t>(vs_input_info.resources[index].DstSelY()),
-				     static_cast<uint32_t>(vs_input_info.resources[index].DstSelZ()),
-				     static_cast<uint32_t>(vs_input_info.resources[index].DstSelW()));
-			}
-		}
-
-		if (vs_input_info.resources[index].OutOfBounds() != 0) {
-			static bool logged = false;
-			if (!logged) {
-				LOGF("VertexInput: temporary: accepting PS5 out-of-bounds behavior %" PRIu8 "\n",
-				     vs_input_info.resources[index].OutOfBounds());
-				logged = true;
-			}
-		}
-
-		EXIT_IF(registers_num < 1 || registers_num > 4);
+		input_attr[index].format   = vertex.vertex_formats[index];
 	}
 
 	vk::PipelineVertexInputStateCreateInfo vertex_input_info {};
@@ -382,17 +219,6 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	vk::PipelineViewportStateCreateInfo viewport_state {};
 	viewport_state.pNext = &depth_clip_control;
 
-	vk::CullModeFlags cull_mode = vk::CullModeFlagBits::eNone;
-	if (static_params.cull_back) {
-		cull_mode |= vk::CullModeFlagBits::eBack;
-	}
-	if (static_params.cull_front) {
-		cull_mode |= vk::CullModeFlagBits::eFront;
-	}
-
-	vk::FrontFace front_face =
-	    (static_params.face ? vk::FrontFace::eClockwise : vk::FrontFace::eCounterClockwise);
-
 	vk::PipelineRasterizationDepthClipStateCreateInfoEXT clip_ext {};
 	clip_ext.depthClipEnable = static_params.depth_clip_enable ? VK_TRUE : VK_FALSE;
 
@@ -413,8 +239,6 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 		provoking_vertex.pNext = rasterizer.pNext;
 		rasterizer.pNext = &provoking_vertex;
 	}
-	rasterizer.cullMode  = cull_mode;
-	rasterizer.frontFace = front_face;
 	rasterizer.polygonMode = static_params.polygon_mode;
 	rasterizer.lineWidth = 1.0f;
 
@@ -470,23 +294,22 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 
 	std::vector<vk::DescriptorSetLayoutBinding> descriptor_bindings;
 	vk::ShaderStageFlags graphics_stages = vk::ShaderStageFlagBits::eFragment;
-	for (const auto& stage: vertex_info) {
-		const auto native_stage = NativeShaderStage(stage.logical_stage);
-		AddLayoutBindings(descriptor_bindings, *stage.stage.program, native_stage);
-		graphics_stages |= native_stage;
-	}
-	if (ps_active) {
-		EXIT_IF(!ps_input_info->stage);
-		AddLayoutBindings(descriptor_bindings, *ps_input_info->stage.program,
-		                  vk::ShaderStageFlagBits::eFragment);
+	for (const auto* stage: stages) {
+		if (stage != nullptr) {
+			const auto native_stage = NativeShaderStage(stage->info.stage);
+			AddLayoutBindings(descriptor_bindings, stage->info, native_stage);
+			pipeline.uses_image_heap |= stage->info.info.uses_image_heap;
+			graphics_stages |= native_stage;
+		}
 	}
 	CreateDescriptorLayout(graphics, pipeline, descriptor_bindings);
 	const vk::PushConstantRange push_constants {graphics_stages, 0,
 	                                            ShaderRecompiler::IR::NativePushConstantSize};
+	const std::array set_layouts {pipeline.descriptor_set_layout, graphics.image_heap_layout};
 
 	vk::PipelineLayoutCreateInfo pipeline_layout_info {};
-	pipeline_layout_info.setLayoutCount         = 1;
-	pipeline_layout_info.pSetLayouts            = &pipeline.descriptor_set_layout;
+	pipeline_layout_info.setLayoutCount         = pipeline.uses_image_heap ? 2u : 1u;
+	pipeline_layout_info.pSetLayouts            = set_layouts.data();
 	pipeline_layout_info.pushConstantRangeCount = 1;
 	pipeline_layout_info.pPushConstantRanges    = &push_constants;
 
@@ -495,7 +318,7 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	if (graphics_debug_dump_enabled()) {
 		LOGF("PipelineTrace: vkCreatePipelineLayout begin VS=%" PRIu64 " PS=%" PRIu64
 		     " set_layouts=1 push_constants=%" PRIu32 "\n",
-		     vertex_program.id, ps_active ? pixel_program.id : 0, 1u);
+		     key.stage_ids[0], key.stage_ids[3], 1u);
 	}
 	auto result = graphics.device.createPipelineLayout(&pipeline_layout_info, nullptr,
 	                                                   &pipeline.pipeline_layout);
@@ -505,14 +328,14 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	}
 	EXIT_NOT_IMPLEMENTED(result != vk::Result::eSuccess);
 
-	EXIT_NOT_IMPLEMENTED(pipeline.pipeline_layout == nullptr);
-
 	vk::PipelineDepthStencilStateCreateInfo depth_stencil_info {};
 
 	std::vector<vk::DynamicState> dynamic_states {
 	    vk::DynamicState::eViewportWithCount,
 	    vk::DynamicState::eScissorWithCount,
 	    vk::DynamicState::eLineWidth,
+	    vk::DynamicState::eCullMode,
+	    vk::DynamicState::eFrontFace,
 	    vk::DynamicState::eDepthTestEnable,
 	    vk::DynamicState::eDepthWriteEnable,
 	    vk::DynamicState::eDepthCompareOp,
@@ -552,9 +375,9 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	pipeline_info.pVertexInputState        = mesh ? nullptr : &vertex_input_info;
 	pipeline_info.pInputAssemblyState      = mesh ? nullptr : &input_assembly;
 	vk::PipelineTessellationStateCreateInfo tessellation_state {};
-	tessellation_state.patchControlPoints =
-	    tessellation ? vs_input_info.tess.input_control_points : 3u;
-	pipeline_info.pTessellationState = (rect_list || tessellation) ? &tessellation_state : nullptr;
+	tessellation_state.patchControlPoints = static_params.patch_control_points;
+	pipeline_info.pTessellationState =
+	    static_params.patch_control_points != 0 ? &tessellation_state : nullptr;
 	pipeline_info.pViewportState          = &viewport_state;
 	pipeline_info.pRasterizationState     = &rasterizer;
 	pipeline_info.pMultisampleState       = &multisampling;
@@ -570,7 +393,7 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 		LOGF("PipelineTrace: vkCreateGraphicsPipelines begin VS=%" PRIu64 " PS=%" PRIu64
 		     " topology=%" PRIu32 " color_mask=0x%08" PRIx32
 		     " depth=%s blend=%s dyn_states=%" PRIu32 "\n",
-		     vertex_program.id, ps_active ? pixel_program.id : 0,
+		     key.stage_ids[0], key.stage_ids[3],
 		     static_cast<uint32_t>(static_params.topology), static_params.color_mask[0],
 		     (with_depth ? "true" : "false"), (static_params.blend_enable[0] ? "true" : "false"),
 		     dynamic_state.dynamicStateCount);
@@ -583,29 +406,18 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	}
 	EXIT_NOT_IMPLEMENTED(result != vk::Result::eSuccess);
 
-	EXIT_NOT_IMPLEMENTED(pipeline.pipeline == nullptr);
-
-	if (tess_control_shader_module != nullptr) {
-		graphics.device.destroyShaderModule(tess_control_shader_module, nullptr);
-	}
-	if (tess_eval_shader_module != nullptr) {
-		graphics.device.destroyShaderModule(tess_eval_shader_module, nullptr);
-	}
+	graphics.device.destroyShaderModule(tess_control_shader_module, nullptr);
+	graphics.device.destroyShaderModule(tess_eval_shader_module, nullptr);
 }
 
-// NOLINTNEXTLINE(readability-function-cognitive-complexity)
 void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& pipeline,
-                            const ShaderComputeInputInfo& input_info,
-                            vk::ShaderModule compute_module, vk::PipelineCache driver_cache) {
-	EXIT_IF(compute_module == nullptr);
-
+                            const ShaderModule& module, vk::PipelineCache driver_cache) {
 	vk::PipelineShaderStageCreateInfo                     comp_shader_stage_info {};
 	vk::PipelineShaderStageRequiredSubgroupSizeCreateInfo comp_subgroup_size {};
 	comp_shader_stage_info.stage  = vk::ShaderStageFlagBits::eCompute;
-	comp_shader_stage_info.module = compute_module;
+	comp_shader_stage_info.module = module.module;
 	comp_shader_stage_info.pName  = "main";
-	EXIT_IF(!input_info.stage);
-	const auto wave_size = input_info.stage.program->wave_size;
+	const auto wave_size = module.info.wave_size;
 	if (graphics.compute_subgroup_size_control_enabled &&
 	    wave_size >= graphics.min_subgroup_size && wave_size <= graphics.max_subgroup_size) {
 		comp_subgroup_size.requiredSubgroupSize = wave_size;
@@ -613,32 +425,27 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	}
 
 	std::vector<vk::DescriptorSetLayoutBinding> descriptor_bindings;
-	AddLayoutBindings(descriptor_bindings, *input_info.stage.program,
-	                  vk::ShaderStageFlagBits::eCompute);
+	AddLayoutBindings(descriptor_bindings, module.info, vk::ShaderStageFlagBits::eCompute);
+	pipeline.uses_image_heap = module.info.info.uses_image_heap;
 	CreateDescriptorLayout(graphics, pipeline, descriptor_bindings);
 	const vk::PushConstantRange push_constants {vk::ShaderStageFlagBits::eCompute, 0,
 	                                            ShaderRecompiler::IR::NativePushConstantSize};
+	const std::array set_layouts {pipeline.descriptor_set_layout, graphics.image_heap_layout};
 
 	vk::PipelineLayoutCreateInfo pipeline_layout_info {};
-	pipeline_layout_info.setLayoutCount         = 1;
-	pipeline_layout_info.pSetLayouts            = &pipeline.descriptor_set_layout;
+	pipeline_layout_info.setLayoutCount         = pipeline.uses_image_heap ? 2u : 1u;
+	pipeline_layout_info.pSetLayouts            = set_layouts.data();
 	pipeline_layout_info.pushConstantRangeCount = 1;
 	pipeline_layout_info.pPushConstantRanges    = &push_constants;
 
 	EXIT_IF(pipeline.pipeline_layout != nullptr);
 
-	LOGF("PipelineTrace: vkCreatePipelineLayout CS begin set_layouts=1 push_constants=%u\n",
-	     1u);
 	auto result = graphics.device.createPipelineLayout(&pipeline_layout_info, nullptr,
 	                                                  &pipeline.pipeline_layout);
-	LOGF("PipelineTrace: vkCreatePipelineLayout CS done result=%s layout=%p\n",
-	     vk::to_string(result).c_str(), static_cast<void*>(pipeline.pipeline_layout));
 	EXIT_NOT_IMPLEMENTED(result != vk::Result::eSuccess);
 
-	EXIT_NOT_IMPLEMENTED(pipeline.pipeline_layout == nullptr);
-
 	vk::ComputePipelineCreateInfo info {};
-	if (input_info.stage.program->info.uses_gds) {
+	if (module.info.info.uses_gds) {
 		info.flags = vk::PipelineCreateFlagBits::eDispatchBase;
 	}
 	info.stage             = comp_shader_stage_info;
@@ -647,15 +454,9 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 
 	EXIT_IF(pipeline.pipeline != nullptr);
 
-	LOGF("PipelineTrace: vkCreateComputePipelines begin layout=%p\n",
-	     static_cast<void*>(pipeline.pipeline_layout));
 	result = graphics.device.createComputePipelines(driver_cache, 1, &info, nullptr,
 	                                                &pipeline.pipeline);
-	LOGF("PipelineTrace: vkCreateComputePipelines done result=%s pipeline=%p\n",
-	     vk::to_string(result).c_str(), static_cast<void*>(pipeline.pipeline));
 	EXIT_NOT_IMPLEMENTED(result != vk::Result::eSuccess);
-
-	EXIT_NOT_IMPLEMENTED(pipeline.pipeline == nullptr);
 }
 
 } // namespace Libs::Graphics

src/graphics/host_gpu/renderer/render.h

@@ -23,14 +23,9 @@ class Shader;
 } // namespace HW
 
 struct GraphicContext;
-struct ShaderBufferResource;
-struct ShaderComputeInputInfo;
-struct RenderDepthInfo;
-struct RenderColorInfo;
 struct DrawCallInfo;
 struct DrawEmitInfo;
 struct DrawIndexBufferSource;
-struct DrawRenderState;
 class RenderContext;
 class CommandScheduler;
 struct RenderExecutorTestAccess;
@@ -48,6 +43,14 @@ enum class DrawOffsetSource : uint8_t {
 	IndirectArgs,
 };
 
+// Nonzero address: the GPU reads count Vulkan indirect draw records, stride bytes apart.
+struct DrawIndirectArgs {
+	uint64_t              address = 0;
+	uint32_t              count   = 0;
+	uint32_t              stride  = 0;
+	ShaderIndirectOffsets offsets;
+};
+
 struct DrawIndexArgs {
 	uint32_t         index_count                = 0;
 	const void*      index_addr                 = nullptr;
@@ -57,6 +60,8 @@ struct DrawIndexArgs {
 	uint32_t         first_instance             = 0;
 	DrawOffsetSource offset_source              = DrawOffsetSource::DrawState;
 	uint32_t         render_target_slice_offset = 0;
+	// With indirect records, index_count is INDEX_BUFFER_SIZE.
+	DrawIndirectArgs indirect;
 };
 
 struct DrawAutoArgs {
@@ -66,6 +71,7 @@ struct DrawAutoArgs {
 	uint32_t         first_instance             = 0;
 	DrawOffsetSource offset_source              = DrawOffsetSource::DrawState;
 	uint32_t         render_target_slice_offset = 0;
+	DrawIndirectArgs indirect;
 };
 
 struct SubmitInfo {
@@ -161,18 +167,19 @@ public:
 	void CommitBindings(CommandBuffer& buffer, vk::PipelineBindPoint pipeline_bind_point,
 	                    const PipelineCache::Pipeline&     pipeline,
 	                    std::span<PreparedBindings* const> bindings);
+	[[nodiscard]] TextureBinding ResolveTexture(const ShaderRecompiler::IR::ImageResource&   resource,
+	                                            const ShaderRecompiler::IR::DescriptorValue& value);
 
 private:
-	void DrawIndex(uint64_t submit_id, CommandBuffer& buffer, const DrawIndexArgs& args);
-	void DrawAuto(uint64_t submit_id, CommandBuffer& buffer, const DrawAutoArgs& args);
+	// False, with nothing drawn, when the draw's indirect records need CPU counts.
+	bool DrawIndex(uint64_t submit_id, CommandBuffer& buffer, const DrawIndexArgs& args);
+	bool DrawAuto(uint64_t submit_id, CommandBuffer& buffer, const DrawAutoArgs& args);
 
 	struct GraphicsBindings {
 		std::array<PreparedBindings, 3> vertex;
 		std::optional<PreparedBindings> pixel;
 	};
 
-	[[nodiscard]] TextureBinding ResolveTexture(const ShaderRecompiler::IR::ImageResource& resource,
-	                                            const ShaderRecompiler::IR::DescriptorValue& value);
 	void PrepareGraphicsBindings(std::span<PreparedBindings* const> stages,
 	                             std::span<RenderColorInfo> colors);
 	void ResolveRenderColorTarget(CommandBuffer& buffer, RenderColorInfo& target,
@@ -187,9 +194,8 @@ private:
 	                                          uint32_t            render_target_slice_offset,
 	                                          DrawRenderState& state);
 	void ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buffer, const DrawCallInfo& draw,
-	                         DrawRenderState& state, vk::PrimitiveTopology topology,
-	                         const DrawEmitInfo& emit, const DrawIndexBufferSource& index_source,
-	                         bool primitive_restart_enable);
+	                         DrawRenderState& state, const DrawEmitInfo& emit,
+	                         const DrawIndexBufferSource& index_source);
 	[[nodiscard]] RenderState AcquireRenderTargets(CommandBuffer& buffer, RenderColorInfo* colors,
 	                                               uint32_t color_count, RenderDepthInfo& depth,
 	                                               vk::ImageAspectFlags& feedback_aspects,

src/graphics/host_gpu/renderer/renderCompute.cpp

@@ -53,7 +53,7 @@ static bool FillSourcesDisjoint(std::span<const ShaderRecompiler::IR::Descriptor
 
 bool RenderExecutor::TryConsumeComputeMetaClear(const ShaderComputeInputInfo& input,
                                                 const CommandBuffer&          buffer) {
-	const auto& program   = *input.stage.program;
+	const auto& program   = input.stage.module->info;
 	const auto& resources = *input.stage.resources;
 	if (resources.buffers.size() != program.info.buffers.size()) {
 		EXIT("compute runtime buffer count does not match shader metadata\n");
@@ -135,7 +135,7 @@ bool ResolveComputeBufferFill(const ShaderComputeInputInfo& input, uint32_t grou
 bool RenderExecutor::TryConsumeComputeFill(const ShaderComputeInputInfo& input,
                                           CommandBuffer& command, uint32_t group_x,
                                           uint32_t group_y, uint32_t group_z, uint32_t mode) {
-	const auto& program   = *input.stage.program;
+	const auto& program   = input.stage.module->info;
 	const auto& resources = *input.stage.resources;
 	const auto& fill      = resources.uniform_fill;
 	auto&       cache     = command.GetContext().GetTextureCache();
@@ -208,7 +208,7 @@ bool RenderExecutor::TryConsumeComputeFill(const ShaderComputeInputInfo& input,
 	if (logged_clears.fetch_add(1, std::memory_order_relaxed) < 32) {
 		LOGF("GraphicsRenderDispatchDirect: compute %s shader=0x%016" PRIx64
 		     " addr=0x%016" PRIx64 " size=0x%016" PRIx64 " value=0x%08" PRIx32 "\n",
-		     image_clear ? "image clear" : "buffer fill", input.stage.program->shader_hash,
+		     image_clear ? "image clear" : "buffer fill", input.stage.module->info.shader_hash,
 		     descriptor.Base48(), size, packed_clear);
 	}
 	return true;
@@ -216,7 +216,7 @@ bool RenderExecutor::TryConsumeComputeFill(const ShaderComputeInputInfo& input,
 
 static void BindSharedMemory(RenderContext& context, ShaderComputeInputInfo& input,
                              PreparedBindings& bindings, uint64_t indirect_args = 0) {
-	if (input.stage.program->bindings.descriptor_counts[
+	if (input.stage.module->info.bindings.descriptor_counts[
 	        static_cast<size_t>(ShaderRecompiler::IR::DescriptorBindingKind::SharedMemory)] == 0) {
 		return;
 	}
@@ -248,7 +248,6 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
                                     uint32_t thread_group_z, uint32_t mode, bool async_compute) {
 	EXIT_IF(buffer.IsInvalid());
 	m_context.GetCommandScheduler().PopPendingOperations();
-	auto& ctx    = buffer.GetRegisters();
 	auto& sh_ctx = buffer.GetShaders();
 
 	buffer.SetDebugInfo(static_cast<uint32_t>(CommandBufferDebugOp::DispatchDirect), submit_id,
@@ -293,7 +292,6 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 	}
 
 	const auto& cs_regs = sh_ctx.GetCs();
-	const auto& sh_regs = ctx.GetShaderRegisters();
 
 	ShaderComputeInputInfo input_info {};
 	input_info.async_compute = async_compute;
@@ -311,16 +309,15 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 			input_info.workgroup_counts[axis] = threads / size + (threads % size != 0u);
 		}
 	}
-	const auto compute_program =
-	    m_context.GetPipelineCache().GetComputeProgram(cs_regs, sh_regs, input_info);
-	if (!compute_program) return;
+	m_context.GetImageHeap().BeginDraw();
+	if (!m_context.GetPipelineCache().GetComputeProgram(cs_regs, input_info)) return;
 	if (use_thread_dimensions) {
 		input_info.dispatch_threads_num[0]    = thread_group_x;
 		input_info.dispatch_threads_num[1]    = thread_group_y;
 		input_info.dispatch_threads_num[2]    = thread_group_z;
 	}
 
-	const auto& program   = *input_info.stage.program;
+	const auto& program   = input_info.stage.module->info;
 	const auto& resources = *input_info.stage.resources;
 	if (resources.specialization_reads.empty() &&
 	    (TryConsumeComputeMetaClear(input_info, buffer) ||
@@ -350,8 +347,7 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 	}
 
 	buffer.EndRendering();
-	auto& pipeline =
-	    m_context.GetPipelineCache().GetComputePipeline(input_info, compute_program);
+	auto& pipeline = m_context.GetPipelineCache().GetComputePipeline(input_info);
 	auto& bindings = m_compute_bindings;
 	PrepareBindings(input_info.stage, bindings);
 	if (program.bindings.dispatch_thread_dword != ShaderRecompiler::IR::PushData::NoStart) {
@@ -363,6 +359,9 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 	if (program.info.uses_dma) {
 		m_context.PrepareBda();
 	}
+	if (program.info.uses_image_heap) {
+		m_context.GetImageHeap().Rebind(*this);
+	}
 	RebindImages(bindings);
 	BindSharedMemory(m_context, input_info, bindings);
 	RebindBuffers(bindings);
@@ -497,20 +496,22 @@ void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
 	ShaderComputeInputInfo input_info {};
 	input_info.async_compute = async_compute;
 	input_info.indirect_args_address = args_addr;
-	const auto compute_program = m_context.GetPipelineCache().GetComputeProgram(
-	    cs_regs, buffer.GetRegisters().GetShaderRegisters(), input_info);
-	if (!compute_program) return;
+	m_context.GetImageHeap().BeginDraw();
+	if (!m_context.GetPipelineCache().GetComputeProgram(cs_regs, input_info)) return;
 	buffer.EndRendering();
-	auto& pipeline = m_context.GetPipelineCache().GetComputePipeline(input_info, compute_program);
+	auto& pipeline = m_context.GetPipelineCache().GetComputePipeline(input_info);
 	auto& bindings = m_compute_bindings;
 	PrepareBindings(input_info.stage, bindings);
 	PreparedBindings* descriptor_stage = &bindings;
 	FindBuffers(std::span {&descriptor_stage, 1u});
-	const auto& program = *input_info.stage.program;
+	const auto& program = input_info.stage.module->info;
 	if (program.info.uses_dma) {
 		m_context.PrepareBda();
 	}
 	BindSharedMemory(m_context, input_info, bindings, args_addr);
+	if (program.info.uses_image_heap) {
+		m_context.GetImageHeap().Rebind(*this);
+	}
 	RebindImages(bindings);
 	// Acquiring arguments can merge cache buffers; finalize shader bindings afterward.
 	const auto [args_buffer, args_offset] = m_context.GetBufferCache().ObtainBuffer(

src/graphics/host_gpu/renderer/renderContext.cpp

@@ -10,12 +10,13 @@
 
 namespace Libs::Graphics {
 
-RenderContext::RenderContext(GraphicContext& graphics)
+RenderContext::RenderContext(GraphicContext& graphics, const std::filesystem::path& shader_store)
     : m_graphics(graphics), m_render_executor(*this), m_command_scheduler(*this, graphics),
       m_descriptor_heap(graphics, m_command_scheduler.GetMasterSemaphore()),
-      m_pipeline_cache(graphics), m_sampler_cache(graphics),
+      m_pipeline_cache(graphics, &m_image_heap, shader_store), m_sampler_cache(graphics),
       m_buffer_cache(graphics, m_command_scheduler, m_page_manager, m_texture_cache),
-      m_texture_cache(graphics, m_command_scheduler, m_page_manager, m_buffer_cache) {
+      m_texture_cache(graphics, m_command_scheduler, m_page_manager, m_buffer_cache),
+      m_image_heap(*this, m_page_manager), m_occlusion_queries(graphics, m_command_scheduler) {
 	EXIT_NOT_IMPLEMENTED(!Common::Thread::IsMainThread());
 }
 
@@ -64,6 +65,7 @@ bool RenderContext::HandleFault(PageFaultAccess access, uint64_t fault_vaddr) no
 	if (access == PageFaultAccess::Write) {
 		m_buffer_cache.InvalidateMemory(fault_vaddr, fault_size);
 		m_texture_cache.InvalidateMemory(fault_vaddr, fault_size);
+		m_image_heap.InvalidateMemory(fault_vaddr, fault_size);
 	} else {
 		m_buffer_cache.ReadMemory(fault_vaddr, fault_size);
 	}
@@ -76,6 +78,7 @@ bool RenderContext::InvalidateMemory(uint64_t vaddr, uint64_t size) {
 	}
 	m_buffer_cache.InvalidateMemory(vaddr, size);
 	m_texture_cache.InvalidateMemory(vaddr, size);
+	m_image_heap.InvalidateMemory(vaddr, size);
 	return true;
 }
 
@@ -112,6 +115,7 @@ void RenderContext::UnmapMemory(uint64_t vaddr, uint64_t size) {
 		}
 		m_buffer_cache.InvalidateMemory(vaddr, size);
 		m_texture_cache.UnmapMemory(vaddr, size);
+		m_image_heap.UnmapMemory(vaddr, size);
 		std::lock_guard lock(m_mapped_ranges_mutex);
 		m_mapped_ranges.Subtract(vaddr, size);
 	};
@@ -130,7 +134,11 @@ void RenderContext::PrepareBda() {
 		m_bda_logged = true;
 	}
 	std::shared_lock lock(m_mapped_ranges_mutex);
-	m_buffer_cache.SynchronizeBda(m_mapped_ranges);
+	m_buffer_cache.TakeBdaSync().ForEach([this](uint64_t begin, uint64_t end) {
+		m_mapped_ranges.ForEachInRange(begin, end - begin, [this](uint64_t start, uint64_t finish) {
+			m_buffer_cache.SynchronizeBuffersInRange(start, finish - start);
+		});
+	});
 	m_fault_process_pending = true;
 }
 
@@ -139,6 +147,7 @@ void RenderContext::RunGarbageCollector() {
 		m_fault_process_pending = false;
 		m_buffer_cache.ProcessFaultBuffer();
 	}
+	m_image_heap.RunGarbageCollector();
 	m_texture_cache.ProcessDownloadImages();
 	m_texture_cache.RunGarbageCollector();
 	m_buffer_cache.RunGarbageCollector();

src/graphics/host_gpu/renderer/renderContext.h

@@ -11,7 +11,9 @@
 #include "graphics/host_gpu/renderer/cache/samplerCache.h"
 #include "graphics/host_gpu/renderer/cache/textureCache.h"
 #include "graphics/host_gpu/renderer/commandScheduler.h"
+#include "graphics/host_gpu/renderer/occlusionQueries.h"
 #include "graphics/host_gpu/renderer/pipeline/descriptorHeap.h"
+#include "graphics/host_gpu/renderer/pipeline/imageHeap.h"
 #include "graphics/host_gpu/renderer/pipeline/pipelineCache.h"
 #include "kernel/eventQueue.h"
 
@@ -29,7 +31,8 @@ class GuestGpu;
 
 class RenderContext {
 public:
-	explicit RenderContext(GraphicContext& graphics);
+	explicit RenderContext(GraphicContext&              graphics,
+	                       const std::filesystem::path& shader_store = ShaderStorePath());
 	~RenderContext();
 	KYTY_CLASS_NO_COPY(RenderContext);
 
@@ -43,10 +46,12 @@ public:
 	CommandScheduler&   GetCommandScheduler() { return m_command_scheduler; }
 	PipelineCache&      GetPipelineCache() { return m_pipeline_cache; }
 	DescriptorHeap&     GetDescriptorHeap() { return m_descriptor_heap; }
+	ImageHeap&          GetImageHeap() { return m_image_heap; }
 	SamplerCache&       GetSamplerCache() { return m_sampler_cache; }
 	BufferCache&        GetBufferCache() { return m_buffer_cache; }
 	TextureCache&       GetTextureCache() { return m_texture_cache; }
 	RenderExecutor&     GetRenderExecutor() { return m_render_executor; }
+	OcclusionQueries&   GetOcclusionQueries() { return m_occlusion_queries; }
 
 	[[nodiscard]] bool HandleFault(PageFaultAccess access, uint64_t fault_vaddr) noexcept;
 	[[nodiscard]] bool InvalidateMemory(uint64_t vaddr, uint64_t size);
@@ -76,6 +81,8 @@ private:
 	PageManager               m_page_manager;
 	BufferCache               m_buffer_cache;
 	TextureCache              m_texture_cache;
+	ImageHeap                 m_image_heap;
+	OcclusionQueries          m_occlusion_queries;
 	mutable std::shared_mutex m_mapped_ranges_mutex;
 	RangeSet                  m_mapped_ranges;
 	std::unique_ptr<GuestGpu> m_gpu;

src/graphics/host_gpu/renderer/renderDraw.cpp

@@ -17,7 +17,6 @@
 #include "graphics/host_gpu/renderer/colorRenderTarget.h"
 #include "graphics/host_gpu/renderer/debug.h"
 #include "graphics/host_gpu/renderer/depthRenderTarget.h"
-#include "graphics/host_gpu/renderer/image/textureCommon.h"
 #include "graphics/host_gpu/renderer/pipeline/pipelineCache.h"
 #include "graphics/host_gpu/renderer/pipeline/shaderResourceBarrier.h"
 #include "graphics/host_gpu/renderer/render.h"
@@ -58,7 +57,7 @@ std::pair<int32_t, uint32_t> ResolveDrawOffsets(uint32_t index_offset,
 	}
 
 	EXIT_IF(!vs_input_info.stage);
-	const auto& program   = *vs_input_info.stage.program;
+	const auto& program   = vs_input_info.stage.module->info;
 	const auto& resources = *vs_input_info.stage.resources;
 	if (index_offset == 0 &&
 	    program.info.vertex_offset_sgpr >= static_cast<int32_t>(program.user_data_base)) {
@@ -198,7 +197,7 @@ static void LogDrawTargetState(const char* draw_name, const RenderColorInfo& col
 	const auto& dc             = ctx.GetDepthControl();
 	const auto& vp             = ctx.GetScreenViewport();
 	const auto& vp0            = vp.viewports[0];
-	const auto& ps_resources   = ps_input_info.stage.program->info;
+	const auto& ps_resources   = ps_input_info.stage.module->info.info;
 	const auto  sampled_images = std::count_if(
 	    ps_resources.images.begin(), ps_resources.images.end(), [](const auto& image) {
 		    return image.resource_class == ShaderRecompiler::IR::ImageResourceClass::Sampled;
@@ -321,7 +320,7 @@ static void UpdateGraphicsDynamicState(const CommandBuffer& buffer, DynamicState
 	const auto& ctx = buffer.GetRegisters();
 	const auto&        vp  = ctx.GetScreenViewport();
 	const vk::Extent2D framebuffer_extent {rendering.width, rendering.height};
-	const auto& outputs = vs_input_info.stage.program->info.outputs;
+	const auto& outputs = vs_input_info.stage.module->info.info.outputs;
 	const bool  indexed_viewports =
 	    std::any_of(outputs.begin(), outputs.end(), [](const auto& output) {
 		    return output.kind == ShaderRecompiler::IR::StageOutputKind::ViewportIndex;
@@ -404,6 +403,20 @@ static void UpdateGraphicsDynamicState(const CommandBuffer& buffer, DynamicState
 		dynamic_state.SetDepthBias(constant_factor, poly_offset.clamp, slope_factor);
 	}
 
+	// Rect lists are drawn without culling.
+	vk::CullModeFlags cull_mode {};
+	if (!Prospero::IsRectList(buffer.GetUserConfig().GetPrimType())) {
+		if (mode.cull_front) {
+			cull_mode |= vk::CullModeFlagBits::eFront;
+		}
+		if (mode.cull_back) {
+			cull_mode |= vk::CullModeFlagBits::eBack;
+		}
+	}
+	dynamic_state.SetCullMode(cull_mode);
+	dynamic_state.SetFrontFace(mode.face ? vk::FrontFace::eClockwise
+	                                     : vk::FrontFace::eCounterClockwise);
+
 	dynamic_state.SetStencilTestEnabled(depth.stencil_test_enable);
 	if (depth.stencil_test_enable) {
 		const auto& front = depth.stencil_front;
@@ -433,22 +446,6 @@ static bool DrawHasValidVertexShader(const HW::Shader& sh_ctx) {
 	return vs.es_regs.data_addr != 0;
 }
 
-static bool PixelShaderHasDepthOrCoverageSideEffects(const HW::ShaderRegisters& sh_regs) {
-	const auto& db = sh_regs.db_shader_control;
-	return db.shader_kill_enable || db.shader_z_export_enable || db.shader_mask_export_enable ||
-	       db.shader_dual_export_enable || db.shader_execute_on_noop || sh_regs.AlphaToMask();
-}
-
-struct DrawRenderState {
-	RenderDepthInfo       depth_info;
-	RenderColorInfo       color_info[RENDER_COLOR_ATTACHMENTS_MAX] = {};
-	uint32_t              color_count                              = 0;
-	bool                  ps_active                                = true;
-	std::array<ShaderVertexInputInfo, 3> vertex_info;
-	ShaderPixelInputInfo  ps_input_info;
-	PipelineCache::GraphicsPrograms programs;
-};
-
 struct DrawCallInfo {
 	CommandBufferDebugOp debug_op       = CommandBufferDebugOp::DrawIndex;
 	uint32_t             index_count    = 0;
@@ -474,9 +471,8 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
 	for (uint32_t i = 0; i < color_count; i++) {
 		auto& target = colors[i];
 		EXIT_IF(!target.image_id);
-		const auto owner = cache.m_slot_images.try_get(target.image_id);
-		if (owner == nullptr || (!owner->registered && !owner->info.data.Empty()) ||
-		    owner->binding.needs_rebind) {
+		const auto& owner = cache.m_slot_images[target.image_id];
+		if ((!owner.registered && !owner.info.data.Empty()) || owner.binding.needs_rebind) {
 			EXIT("color target changed after render-state discovery\n");
 		}
 		const auto image_view = cache.FindRenderTarget(target.image_id, target.desc);
@@ -502,8 +498,8 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
 		attachment.image_layout = layout;
 	}
 	if (depth.image_id) {
-		const auto owner = cache.m_slot_images.try_get(depth.image_id);
-		if (owner == nullptr || !owner->registered || owner->binding.needs_rebind) {
+		const auto& owner = cache.m_slot_images[depth.image_id];
+		if (!owner.registered || owner.binding.needs_rebind) {
 			EXIT("depth target changed after render-state discovery\n");
 		}
 		const auto  image_view = cache.FindDepthTarget(depth.image_id, depth.desc);
@@ -589,23 +585,6 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
 	return state;
 }
 
-static uint32_t DrawColorOutputMask(const HW::Context& ctx) {
-	const auto& sh_regs     = ctx.GetShaderRegisters();
-	const auto  write_mask  = ctx.GetRenderTargetMask() & sh_regs.m_cbShaderMask;
-	uint32_t    output_mask = ctx.HasObjectPrimitiveIdOutput() ? 1u << 7u : 0;
-	for (uint32_t index = 0; index < RENDER_COLOR_ATTACHMENTS_MAX; index++) {
-		const auto slot = ShaderPixelExportTarget(sh_regs.m_cbShaderMask, index);
-		if (slot >= RENDER_COLOR_ATTACHMENTS_MAX) {
-			break;
-		}
-		if (sh_regs.target_output_mode[index] != 0 &&
-		    render_target_mask_slot(write_mask, slot) != 0) {
-			output_mask |= 1u << slot;
-		}
-	}
-	return output_mask;
-}
-
 enum class CbColorMode : uint8_t {
 	Disable            = 0,
 	Normal             = 1,
@@ -631,10 +610,11 @@ bool RenderExecutor::ConsumeMetadataColorOperation(CommandBuffer& buffer,
 	       mode == static_cast<uint8_t>(CbColorMode::DccDecompress);
 }
 
+// vertex_offset is the first vertex of non-indexed draws.
 struct DrawEmitInfo {
-	int32_t  vertex_offset = 0;
-	uint32_t first_vertex  = 0;
-	uint32_t first_instance = 0;
+	int32_t          vertex_offset  = 0;
+	uint32_t         first_instance = 0;
+	DrawIndirectArgs indirect;
 };
 
 struct DrawIndexBufferSource {
@@ -869,61 +849,24 @@ static bool ResolvePrimitiveRestart(const CommandBuffer& buffer,
 	return false;
 }
 
-static void RefreshShaders(CommandBuffer& buffer, const DrawCallInfo& draw,
-                           uint32_t color_output_mask, DrawRenderState& state) {
-	auto& ctx    = buffer.GetRegisters();
-	auto& sh_ctx = buffer.GetShaders();
-
-	const auto& vertex_shader_info = sh_ctx.GetVs();
-	const auto& pixel_shader_info  = sh_ctx.GetPs();
-	const auto& shader_regs        = ctx.GetShaderRegisters();
-
-	state.programs      = {};
-	state.ps_input_info = {};
-	std::array<Prospero::ColorComponentMapping, RENDER_COLOR_ATTACHMENTS_MAX>
-	    target_export_mapping {};
-	for (uint32_t slot = 0; slot < RENDER_COLOR_ATTACHMENTS_MAX; slot++) {
-		const auto& rt = ctx.GetRenderTarget(slot);
-		if ((color_output_mask & (1u << slot)) != 0 && rt.base.addr != 0) {
-			target_export_mapping[slot] =
-			    TextureGetRenderTargetFormat(rt.info.format, rt.info.channel_type,
-			                                 rt.info.channel_order)
-			        .export_mapping;
-		}
-	}
-	auto& pipeline_cache = buffer.GetContext().GetPipelineCache();
-	if (draw.IsIndexed()) {
-		LogDrawPhase(draw.Name(), "GetGraphicsPrograms");
-	}
-	state.programs = pipeline_cache.GetGraphicsPrograms(
-	    vertex_shader_info, pixel_shader_info, shader_regs, ctx, buffer.GetUserConfig(),
-	    target_export_mapping, state.ps_active, state.vertex_info, state.ps_input_info);
-}
-
 bool RenderExecutor::PrepareDrawRenderState(CommandBuffer& buffer, const DrawCallInfo& draw,
                                             uint32_t            render_target_slice_offset,
 	                                        DrawRenderState& state) {
-	const auto& shader_regs       = buffer.GetRegisters().GetShaderRegisters();
-	const auto  color_output_mask = DrawColorOutputMask(buffer.GetRegisters());
-	state.ps_active = buffer.GetShaders().GetPs().ps_regs.data_addr != 0 &&
-	                  (color_output_mask != 0 ||
-	                   PixelShaderHasDepthOrCoverageSideEffects(shader_regs));
-	RefreshShaders(buffer, draw, color_output_mask, state);
-	uint32_t mrt_mask = 0;
-	if (state.ps_active) {
-		for (const auto& output: state.ps_input_info.stage.program->info.outputs) {
-			if (output.kind == ShaderRecompiler::IR::StageOutputKind::Mrt ||
-			    output.kind == ShaderRecompiler::IR::StageOutputKind::ObjectPrimitiveId) {
-				mrt_mask |= 1u << output.location;
-			}
-		}
+	m_context.GetImageHeap().BeginDraw();
+	if (draw.IsIndexed()) {
+		LogDrawPhase(draw.Name(), "GetGraphicsPrograms");
+	}
+	m_context.GetPipelineCache().GetGraphicsPrograms(buffer, state);
+	// Vertex programs beyond the host mesh limits or with unsupported call targets are not
+	// compiled.
+	if (state.vertex_info[0].stage.module == nullptr) {
+		return false;
 	}
-	mrt_mask &= color_output_mask;
 	if (draw.IsIndexed()) {
 		LogDrawPhase(draw.Name(), "ResolveRenderColorTarget");
 	}
 	for (uint32_t slot = 0; slot < RENDER_COLOR_ATTACHMENTS_MAX; slot++) {
-		if ((mrt_mask & (1u << slot)) != 0) {
+		if ((state.mrt_mask & (1u << slot)) != 0) {
 			ResolveRenderColorTarget(buffer, state.color_info[state.color_count],
 			                         render_target_slice_offset, slot);
 			if (state.color_info[state.color_count].image_id) {
@@ -1043,8 +986,8 @@ static void EmitDrawPrimitives(const HW::UserConfig& ucfg, vk::CommandBuffer vk_
 				vk_buffer.drawIndexed(draw.index_count, draw.instance_count, 0, emit.vertex_offset,
 				                      emit.first_instance);
 			} else {
-				vk_buffer.draw(draw.index_count, draw.instance_count, emit.first_vertex,
-				               emit.first_instance);
+				vk_buffer.draw(draw.index_count, draw.instance_count,
+				               static_cast<uint32_t>(emit.vertex_offset), emit.first_instance);
 			}
 			break;
 		case Prospero::PrimitiveType::kQuadListLegacy:
@@ -1054,7 +997,8 @@ static void EmitDrawPrimitives(const HW::UserConfig& ucfg, vk::CommandBuffer vk_
 					vk_buffer.drawIndexed(4, draw.instance_count, i, emit.vertex_offset,
 					                      emit.first_instance);
 				} else {
-					vk_buffer.draw(4, draw.instance_count, i + emit.first_vertex,
+					vk_buffer.draw(4, draw.instance_count,
+					               i + static_cast<uint32_t>(emit.vertex_offset),
 					               emit.first_instance);
 				}
 			}
@@ -1075,7 +1019,7 @@ bool RenderExecutor::CopyDepthToColor(CommandBuffer& buffer, const DrawCallInfo&
 	EXIT_NOT_IMPLEMENTED(!m_context.GetGraphics().maintenance8_enabled ||
 	                     !Prospero::IsRectList(ucfg.GetPrimType()) ||
 	                     draw.IsIndexed() || draw.index_count != 3 || draw.instance_count != 1 ||
-	                     emit.first_vertex != 0 || emit.first_instance != 0 ||
+	                     emit.vertex_offset != 0 || emit.first_instance != 0 ||
 	                     state.color_count != 1 || color.target_slot != 0 || !depth.image_id ||
 	                     color.desc.info.samples != 1 ||
 	                     depth.desc.info.extent.width != color.Extent().width ||
@@ -1121,16 +1065,15 @@ bool RenderExecutor::CopyDepthToColor(CommandBuffer& buffer, const DrawCallInfo&
 
 void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buffer,
                                          const DrawCallInfo& draw, DrawRenderState& state,
-                                         vk::PrimitiveTopology topology, const DrawEmitInfo& emit,
-                                         const DrawIndexBufferSource& index_source,
-	                                     bool primitive_restart_enable) {
-	auto& ucfg = buffer.GetUserConfig();
+                                         const DrawEmitInfo&          emit,
+                                         const DrawIndexBufferSource& index_source) {
+	auto&      ucfg                     = buffer.GetUserConfig();
+	const bool primitive_restart_enable = state.primitive_restart_enable;
 	if (CopyDepthToColor(buffer, draw, emit, state)) {
 		return;
 	}
-	const auto vertex_stages =
-	    std::span {state.vertex_info.data(), state.programs.VertexStageCount()};
-	const bool mesh_active = state.vertex_info[0].stage.program->stage == ShaderType::Mesh;
+	const auto vertex_stages = state.VertexStages();
+	const bool mesh_active = state.vertex_info[0].stage.module->info.stage == ShaderType::Mesh;
 	uint32_t   mesh_groups = 0;
 	if (mesh_active) {
 		const auto& mesh = state.vertex_info[0].mesh;
@@ -1189,13 +1132,18 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 		vertex_bindings = AcquireVertexBuffers(buffer, state.vertex_info[0]);
 		index_binding   = PrepareIndexBuffer(buffer, index_source);
 	}
+	std::pair<Buffer*, uint64_t> args_buffer {};
+	if (emit.indirect.address != 0) {
+		const auto size = static_cast<uint64_t>(emit.indirect.count - 1) * emit.indirect.stride +
+		                  (draw.IsIndexed() ? sizeof(vk::DrawIndexedIndirectCommand)
+		                                    : sizeof(vk::DrawIndirectCommand));
+		args_buffer = m_context.GetBufferCache().ObtainBuffer(emit.indirect.address, size, false);
+		EXIT_IF((args_buffer.second & 3u) != 0);
+	}
 	if (draw.IsIndexed()) {
 		LogDrawPhase(draw.Name(), "CreatePipeline");
 	}
-	auto& pipeline = m_context.GetPipelineCache().GetGraphicsPipeline(
-	    std::span {state.color_info, state.color_count}, state.depth_info, vertex_stages, buffer,
-	    state.ps_active ? &state.ps_input_info : nullptr, topology, primitive_restart_enable,
-	    state.programs);
+	auto& pipeline = m_context.GetPipelineCache().GetGraphicsPipeline(buffer, state);
 	vk::ImageAspectFlags feedback_aspects;
 	const auto rendering =
 	    AcquireRenderTargets(buffer, state.color_info, state.color_count, state.depth_info,
@@ -1225,12 +1173,11 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 	if (draw_dwords != 0) {
 		std::array<uint32_t, PushData::MeshDrawDwordCount> draw_data {};
 		if (mesh_active) {
-			draw_data = {
-			    draw.index_count,
-			    draw.IsIndexed() ? static_cast<uint32_t>(emit.vertex_offset) : emit.first_vertex,
-			    emit.first_instance, index_source.guest_element_size,
-			    static_cast<uint32_t>(index_source.address),
-			    static_cast<uint32_t>(index_source.address >> 32u)};
+			draw_data = {draw.index_count,
+			             static_cast<uint32_t>(emit.vertex_offset),
+			             index_source.guest_element_size,
+			             static_cast<uint32_t>(index_source.address),
+			             static_cast<uint32_t>(index_source.address >> 32u)};
 		} else {
 			draw_data[0] = buffer.GetUserConfig().GetObjectId();
 		}
@@ -1273,11 +1220,20 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 	if (!draw.IsIndexed()) {
 		SetDrawDebugPhase(buffer, submit_id, draw, 0x500u);
 	}
+	auto&         occlusion_queries = m_context.GetOcclusionQueries();
+	const int32_t occlusion_query   = occlusion_queries.Begin(vk_buffer);
 	if (mesh_active) {
 		vk_buffer.drawMeshTasksEXT(mesh_groups, draw.instance_count, 1);
-	} else {
+	} else if (args_buffer.first == nullptr) {
 		EmitDrawPrimitives(ucfg, vk_buffer, draw, emit);
+	} else if (draw.IsIndexed()) {
+		vk_buffer.drawIndexedIndirect(args_buffer.first->Handle(), args_buffer.second,
+		                              emit.indirect.count, emit.indirect.stride);
+	} else {
+		vk_buffer.drawIndirect(args_buffer.first->Handle(), args_buffer.second,
+		                       emit.indirect.count, emit.indirect.stride);
 	}
+	occlusion_queries.End(vk_buffer, occlusion_query);
 	Common::DebugCounters::Add(Common::DebugCounters::Counter::Draws);
 	Common::DebugCounters::Add(Common::DebugCounters::Counter::DrawInstances, draw.instance_count);
 
@@ -1294,7 +1250,32 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 	}
 }
 
-void RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
+// Vulkan offsets for a prepared draw, or nullopt when its indirect records need the CPU: mesh
+// and quad-list draws for counts, tessellation draws for patched SGPRs. Indirect records carry
+// their own offsets.
+static std::optional<DrawEmitInfo> ResolveDrawEmit(const HW::UserConfig&        ucfg,
+                                                   const ShaderVertexInputInfo& vs,
+                                                   DrawOffsetSource             source,
+                                                   int32_t                      first_vertex,
+                                                   uint32_t                     first_instance,
+                                                   const DrawIndirectArgs&      indirect) {
+	if (indirect.address != 0) {
+		if (vs.stage.module->info.stage == ShaderType::Mesh ||
+		    ucfg.GetPrimType() == Prospero::PrimitiveType::kQuadListLegacy ||
+		    vs.indirect_offsets != indirect.offsets) {
+			return std::nullopt;
+		}
+		return DrawEmitInfo {.indirect = indirect};
+	}
+	if (source == DrawOffsetSource::IndirectArgs) {
+		// Native fetches consume the patched SGPRs; rewritten fetches need Vulkan offsets.
+		return vs.fetch_embedded ? DrawEmitInfo {first_vertex, first_instance} : DrawEmitInfo {};
+	}
+	const auto [vertex_offset, instance_offset] = ResolveDrawOffsets(ucfg.GetIndexOffset(), vs);
+	return DrawEmitInfo {vertex_offset + first_vertex, instance_offset};
+}
+
+bool RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
                                const DrawIndexArgs& args) {
 	KYTY_PROFILER_FUNCTION();
 
@@ -1309,19 +1290,19 @@ void RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
 	                    reinterpret_cast<uint64_t>(args.index_addr));
 
 	Common::LockGuard lock(m_context.GetMutex());
-	if (args.index_count == 0 || args.instance_count == 0) {
-		return;
+	if (args.index_count == 0 || (args.instance_count == 0 && args.indirect.address == 0)) {
+		return true;
 	}
 
 	if (ConsumeMetadataColorOperation(buffer, args.render_target_slice_offset) ||
 	    DepthStencilCopy(buffer) ||
 	    ResolveColorTargets(buffer, args.render_target_slice_offset)) {
 		ResetBindings();
-		return;
+		return true;
 	}
 
 	if (!DrawHasValidVertexShader(sh_ctx)) {
-		return;
+		return true;
 	}
 
 	if (graphics_debug_dump_enabled()) {
@@ -1345,9 +1326,9 @@ void RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
 
 	hw_check(buffer);
 
-	vk::PrimitiveTopology topology = vk::PrimitiveTopology::ePointList;
-	if (!GetDrawTopology(ucfg, topology)) {
-		return;
+	DrawRenderState state {.indirect_offsets = args.indirect.offsets};
+	if (!GetDrawTopology(ucfg, state.topology)) {
+		return true;
 	}
 
 	DrawIndexBufferSource index_source {};
@@ -1367,7 +1348,7 @@ void RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
 		default: EXIT("unknown index_type_and_size: %u\n", args.index_type_and_size);
 	}
 	index_source.size = static_cast<uint64_t>(args.index_count) * index_source.guest_element_size;
-	const bool primitive_restart = ResolvePrimitiveRestart(buffer, index_source);
+	state.primitive_restart_enable = ResolvePrimitiveRestart(buffer, index_source);
 
 	std::vector<uint16_t> expanded_indices;
 	if (index_source.guest_element_size == 1) {
@@ -1375,7 +1356,8 @@ void RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
 		const auto* src = static_cast<const uint8_t*>(args.index_addr);
 		expanded_indices.resize(args.index_count);
 		for (uint32_t i = 0; i < args.index_count; i++) {
-			expanded_indices[i] = primitive_restart && src[i] == 0xffu ? 0xffffu : src[i];
+			expanded_indices[i] =
+			    state.primitive_restart_enable && src[i] == 0xffu ? 0xffffu : src[i];
 		}
 		index_source.host_data = expanded_indices.data();
 		index_source.size      = expanded_indices.size() * sizeof(uint16_t);
@@ -1383,33 +1365,24 @@ void RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
 
 	const DrawCallInfo draw {CommandBufferDebugOp::DrawIndex, args.index_count,
 	                        args.instance_count, args.first_instance};
-	DrawRenderState state {};
 	if (!PrepareDrawRenderState(buffer, draw, args.render_target_slice_offset, state)) {
 		ResetBindings();
-		return;
+		return true;
 	}
 
 	LogDrawStateIfNeeded(buffer, draw, state, args.index_type_and_size,
 	                     args.index_addr);
-
-	const bool indirect = args.offset_source == DrawOffsetSource::IndirectArgs;
-	const auto [vertex_offset, instance_offset] =
-	    indirect ? std::pair<int32_t, uint32_t> {0, args.first_instance}
-	             : ResolveDrawOffsets(ucfg.GetIndexOffset(), state.vertex_info[0]);
-
-	DrawEmitInfo emit {};
-	// 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);
+	const auto emit = ResolveDrawEmit(ucfg, state.vertex_info[0], args.offset_source,
+	                                  args.base_vertex, args.first_instance, args.indirect);
+	if (emit) {
+		ExecutePreparedDraw(submit_id, buffer, draw, state, *emit, index_source);
+	}
 	ResetBindings();
+	return emit.has_value();
 }
 
 // NOLINTNEXTLINE(readability-function-cognitive-complexity)
-void RenderExecutor::DrawAuto(uint64_t submit_id, CommandBuffer& buffer, const DrawAutoArgs& args) {
+bool RenderExecutor::DrawAuto(uint64_t submit_id, CommandBuffer& buffer, const DrawAutoArgs& args) {
 	KYTY_PROFILER_FUNCTION();
 
 	EXIT_IF(buffer.IsInvalid());
@@ -1423,19 +1396,19 @@ void RenderExecutor::DrawAuto(uint64_t submit_id, CommandBuffer& buffer, const D
 	                    args.first_instance);
 
 	Common::LockGuard lock(m_context.GetMutex());
-	if (args.vertex_count == 0 || args.instance_count == 0) {
-		return;
+	if (args.indirect.address == 0 && (args.vertex_count == 0 || args.instance_count == 0)) {
+		return true;
 	}
 
 	if (ConsumeMetadataColorOperation(buffer, args.render_target_slice_offset) ||
 	    DepthStencilCopy(buffer) ||
 	    ResolveColorTargets(buffer, args.render_target_slice_offset)) {
 		ResetBindings();
-		return;
+		return true;
 	}
 
 	if (!DrawHasValidVertexShader(sh_ctx)) {
-		return;
+		return true;
 	}
 
 	if (graphics_debug_dump_enabled()) {
@@ -1458,20 +1431,19 @@ void RenderExecutor::DrawAuto(uint64_t submit_id, CommandBuffer& buffer, const D
 	const DrawCallInfo draw {CommandBufferDebugOp::DrawIndexAuto,
 	                         args.vertex_count, args.instance_count, args.first_instance};
 
-	vk::PrimitiveTopology topology = vk::PrimitiveTopology::ePointList;
-	if (!GetDrawTopology(ucfg, topology)) {
+	DrawRenderState state {.indirect_offsets = args.indirect.offsets};
+	if (!GetDrawTopology(ucfg, state.topology)) {
 		ResetBindings();
-		return;
+		return true;
 	}
-	DrawRenderState state {};
 	if (!PrepareDrawRenderState(buffer, draw, args.render_target_slice_offset, state)) {
 		ResetBindings();
-		return;
+		return true;
 	}
 
 	const bool rect_list = Prospero::IsRectList(ucfg.GetPrimType());
 	if (rect_list && state.vertex_info[0].buffers_num == 0 &&
-	    state.vertex_info[0].stage.program->param_export_mask == 0 &&
+	    state.vertex_info[0].stage.module->info.param_export_mask == 0 &&
 	    state.ps_input_info.input_num != 0) {
 		if (graphics_debug_dump_enabled()) {
 			LOGF("DrawIndexAuto: skipping rect-list draw with no VS param exports and PS inputs: "
@@ -1480,24 +1452,18 @@ void RenderExecutor::DrawAuto(uint64_t submit_id, CommandBuffer& buffer, const D
 			     sh_ctx.GetVs().es_regs.data_addr, sh_ctx.GetVs().gs_regs.data_addr);
 		}
 		ResetBindings();
-		return;
+		return true;
 	}
 
 	LogDrawStateIfNeeded(buffer, draw, state, 0, nullptr);
-
-	const bool indirect = args.offset_source == DrawOffsetSource::IndirectArgs;
-	const auto [vertex_offset, instance_offset] =
-	    indirect ? std::pair<int32_t, uint32_t> {0, args.first_instance}
-	             : ResolveDrawOffsets(ucfg.GetIndexOffset(), state.vertex_info[0]);
-	DrawEmitInfo emit {};
-	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);
+	const auto emit = ResolveDrawEmit(ucfg, state.vertex_info[0], args.offset_source,
+	                                  static_cast<int32_t>(args.first_vertex), args.first_instance,
+	                                  args.indirect);
+	if (emit) {
+		ExecutePreparedDraw(submit_id, buffer, draw, state, *emit, {});
+	}
 	ResetBindings();
+	return emit.has_value();
 }
 
 bool RenderExecutor::ResolveColorTargets(CommandBuffer& buffer, uint32_t render_target_slice_offset) {
@@ -1529,8 +1495,9 @@ bool RenderExecutor::ResolveColorTargets(CommandBuffer& buffer, uint32_t render_
 	cache.MarkGpuWritten(dst.image_id);
 	auto& source      = cache.GetImage(src.image_id);
 	auto& destination = cache.GetImage(dst.image_id);
-	destination.Resolve(source, {src.guest_mip_level, 1, src.guest_array_layer, 1},
-	                    {dst.guest_mip_level, 1, dst.guest_array_layer, 1});
+	destination.Resolve(source,
+	                    {src.desc.view_info.base_level, 1, src.desc.view_info.base_layer, 1},
+	                    {dst.desc.view_info.base_level, 1, dst.desc.view_info.base_layer, 1});
 	return true;
 }
 

src/graphics/host_gpu/shaders/gpu_tiler_color_clear.comp

@@ -0,0 +1,50 @@
+#version 450
+
+// One workgroup per metadata slice; see TileManager::CheckColorClear.
+layout(local_size_x = 256) in;
+
+layout(set = 0, binding = 0, std430) buffer Metadata { uint data[]; } metadata;
+layout(set = 0, binding = 1, std430) writeonly buffer Predicates { uint data[]; } predicates;
+
+layout(push_constant, std430) uniform Check {
+	uint metadata_base;
+	uint words;
+	uint keys;
+	uint expand;
+} params;
+
+const uint Slots = 8u; // TileManager::ColorClearSlots
+
+shared uint key;
+shared uint mismatch;
+
+void main() {
+	const uint slice = gl_WorkGroupID.x;
+	const uint base  = (params.metadata_base >> 2u) + slice * params.words;
+	// Each invocation reads, and may rewrite, only its own words.
+	if (gl_LocalInvocationIndex == 0u) {
+		key      = metadata.data[base] & 0xffu;
+		mismatch = 0u;
+	}
+	barrier();
+	const uint pattern = key * 0x01010101u;
+	bool       same    = true;
+	for (uint i = gl_LocalInvocationIndex; same && i < params.words; i += gl_WorkGroupSize.x) {
+		same = metadata.data[base + i] == pattern;
+	}
+	if (!same) {
+		atomicOr(mismatch, 1u);
+	}
+	barrier();
+	const uint slot = key >> 5u;
+	const bool hit  = mismatch == 0u && (key & 0x1fu) == 0u && (params.keys & (1u << slot)) != 0u;
+	if (gl_LocalInvocationIndex < Slots) {
+		predicates.data[slice * Slots + gl_LocalInvocationIndex] =
+		    uint(hit && slot == gl_LocalInvocationIndex);
+	}
+	if (hit && params.expand != 0u) {
+		for (uint i = gl_LocalInvocationIndex; i < params.words; i += gl_WorkGroupSize.x) {
+			metadata.data[base + i] = 0xffffffffu;
+		}
+	}
+}

src/graphics/host_gpu/shaders/gpu_tiler_demote_d16.comp

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

src/graphics/presentation/performanceOverlay.cpp

@@ -310,6 +310,10 @@ void GpuActivityRows(const Snapshot& shown) {
 	Row("CPU waiting on GPU",
 	    Fmt("%.0f ms/s  (%.0f waits/s)", Rate(shown, Counter::GpuWaitNs) / 1e6,
 	        Rate(shown, Counter::GpuWaits)));
+	Row("DCC/CMASK readbacks",
+	    Fmt("%.0f ms/s  (%.0f/s)", Rate(shown, Counter::ColorMetadataReadbackNs) / 1e6,
+	        Rate(shown, Counter::ColorMetadataReadbacks)));
+	Row("DCC/CMASK GPU checks", Fmt("%.0f/s", Rate(shown, Counter::ColorMetadataChecks)));
 }
 
 void CompilationRows(const Snapshot& shown) {
@@ -529,16 +533,16 @@ void DrawMemoryAndIo(const Snapshot& shown) {
 		    Bytes(guest.pooled_allocated) + " (" + Bytes(guest.pool_committed) + " committed)");
 		Row("Flexible", Bytes(guest.flexible_used) + " / " + Bytes(guest.flexible_total));
 		Row("Extended", Bytes(guest.automatic_allocated));
-		// Past the pool size the guest sees no page-table entries available.
-		const auto entries = std::max(guest.cpu_page_entries, guest.gpu_page_entries);
+		// Past its total a side sees no page-table entries available.
+		const auto percent = std::max(guest.cpu_page_entries * 100 / guest.cpu_page_entries_total,
+		                              guest.gpu_page_entries * 100 / guest.gpu_page_entries_total);
 		Row("Page table (2 MB)",
-		    Fmt("CPU %llu  GPU %llu  / %llu",
+		    Fmt("CPU %llu / %llu  GPU %llu / %llu",
 		        static_cast<unsigned long long>(guest.cpu_page_entries),
+		        static_cast<unsigned long long>(guest.cpu_page_entries_total),
 		        static_cast<unsigned long long>(guest.gpu_page_entries),
-		        static_cast<unsigned long long>(guest.page_entries_total)),
-		    entries > guest.page_entries_total            ? BAD
-		    : entries * 10 > guest.page_entries_total * 9 ? MEDIUM
-		                                                  : GOOD);
+		        static_cast<unsigned long long>(guest.gpu_page_entries_total)),
+		    LevelColor(static_cast<double>(percent), 90, 100));
 		ImGui::EndTable();
 	}
 	if (BeginSection("Emulator process")) {

src/graphics/presentation/videoOut.cpp

@@ -221,7 +221,7 @@ public:
 	void WaitForSubmitSlot(VideoOutConfig& cfg);
 	bool Flip(uint32_t micros);
 	void GetFlipStatus(VideoOutConfig& cfg, VideoOutFlipStatus& out);
-	void Wait(VideoOutConfig& cfg, int index);
+	bool IsSafeForRendering(VideoOutConfig& cfg, int index);
 
 private:
 	enum class RequestState { Reserved, Recording, Ready, Presenting };
@@ -240,6 +240,7 @@ private:
 		Graphics::Presenter::Frame* frame;
 		bool                        premultiplied_alpha;
 	};
+	void SignalDoneLocked();
 
 	Graphics::Presenter& m_presenter;
 	Common::Mutex        m_mutex;
@@ -250,6 +251,7 @@ private:
 	std::list<Request>   m_cpu_requests;
 	std::list<Request>   m_cancelled_requests;
 	bool                 m_processing      = false;
+	bool                 m_gpu_waiting     = false;
 	uint64_t             m_next_request_id = 1;
 };
 
@@ -953,7 +955,7 @@ bool FlipQueue::Reserve(VideoOutConfig& cfg, int index, int64_t flip_arg, FlipRe
 				}
 			}
 		}
-		m_done_cond_var.SignalAll();
+		SignalDoneLocked();
 	}
 	if (cfg.flip_status.flipPendingNum >= VIDEO_OUT_FLIP_QUEUE_CAPACITY) {
 		Common::DebugCounters::Add(Common::DebugCounters::Counter::FlipsRejected);
@@ -1033,7 +1035,7 @@ void FlipQueue::Cancel(VideoOutConfig& cfg) {
 			it = queue->erase(it);
 		}
 	}
-	m_done_cond_var.SignalAll();
+	SignalDoneLocked();
 	m_submit_slot_cond_var.SignalAll();
 	m_submit_cond_var.SignalAll();
 	m_mutex.Unlock();
@@ -1205,19 +1207,26 @@ void FlipQueue::WaitForSubmitSlot(VideoOutConfig& cfg) {
 	}
 }
 
-void FlipQueue::Wait(VideoOutConfig& cfg, int index) {
-	Common::LockGuard lock(m_mutex);
-
-	auto has_request = [this, &cfg, index] {
-		auto matches = [&cfg, index](const auto& r) { return r.cfg == &cfg && r.index == index; };
-		return std::any_of(m_requests.begin(), m_requests.end(), matches) ||
-		       std::any_of(m_cpu_requests.begin(), m_cpu_requests.end(), matches);
-	};
-	while (has_request()) {
-		m_done_cond_var.Wait(&m_mutex);
+void FlipQueue::SignalDoneLocked() {
+	m_done_cond_var.SignalAll();
+	// A stalled GPU cannot drain and shut down without re-checking under m_mutex, so it is alive here.
+	if (m_gpu_waiting) {
+		m_gpu_waiting = false;
+		m_presenter.Renderer().GetGpu().Wake();
 	}
 }
 
+// A display buffer is safe for rendering once it is neither displayed nor waiting to be.
+bool FlipQueue::IsSafeForRendering(VideoOutConfig& cfg, int index) {
+	Common::LockGuard lock(m_mutex);
+	auto matches = [&cfg, index](const auto& r) { return r.cfg == &cfg && r.index == index; };
+	const bool safe = cfg.flip_status.currentBuffer != index &&
+	                  std::none_of(m_requests.begin(), m_requests.end(), matches) &&
+	                  std::none_of(m_cpu_requests.begin(), m_cpu_requests.end(), matches);
+	m_gpu_waiting = !safe;
+	return safe;
+}
+
 bool FlipQueue::Flip(uint32_t micros) {
 	KYTY_PROFILER_BLOCK("FlipQueue::Flip");
 
@@ -1311,7 +1320,7 @@ bool FlipQueue::Flip(uint32_t micros) {
 		it = m_requests.erase(it);
 	}
 	m_processing = false;
-	m_done_cond_var.SignalAll();
+	SignalDoneLocked();
 	m_submit_slot_cond_var.SignalAll();
 	m_mutex.Unlock();
 	for (size_t i = count; i != 0; i--) {
@@ -1694,12 +1703,12 @@ void VideoOutDriver::WaitForSubmitSlot(int handle) {
 	m_impl->GetFlipQueue().WaitForSubmitSlot(*cfg);
 }
 
-void VideoOutDriver::WaitFlipDone(int handle, int index) {
+bool VideoOutDriver::IsSafeForRendering(int handle, int index) {
 	auto* ctx = m_impl->Get(handle);
 	EXIT_IF(ctx == nullptr);
 
 	EXIT_NOT_IMPLEMENTED(!IsValidBufferIndex(index));
-	m_impl->GetFlipQueue().Wait(*ctx, index);
+	return m_impl->GetFlipQueue().IsSafeForRendering(*ctx, index);
 }
 
 KYTY_SYSV_ABI int VideoOutGetFlipStatus(int handle, VideoOutFlipStatus* status) {

src/graphics/presentation/videoOut.h

@@ -38,7 +38,7 @@ public:
 	void CompleteFlip(uint64_t request_id);
 	void SubmitFlipPreparation(uint64_t request_id);
 	void WaitForSubmitSlot(int handle);
-	void WaitFlipDone(int handle, int index);
+	[[nodiscard]] bool IsSafeForRendering(int handle, int index);
 
 	[[nodiscard]] Impl& State() noexcept;
 

src/graphics/presentation/window/vulkanWindow.cpp

@@ -24,6 +24,7 @@
 #include <SDL3/SDL.h>
 #include <SDL3/SDL_vulkan.h>
 #include <algorithm>
+#include <cmath>
 #include <cstdio>
 #include <cstring>
 #include <fmt/format.h>
@@ -50,6 +51,7 @@ struct VulkanExtensions {
 vk::PhysicalDeviceVulkan11Features WindowContext::RequiredVulkan11Features() noexcept {
 	vk::PhysicalDeviceVulkan11Features features {};
 	features.storageBuffer16BitAccess = VK_TRUE;
+	features.shaderDrawParameters     = VK_TRUE;
 	return features;
 }
 
@@ -65,6 +67,9 @@ vk::PhysicalDeviceVulkan12Features WindowContext::RequiredVulkan12Features() noe
 	features.shaderSampledImageArrayNonUniformIndexing = VK_TRUE;
 	features.shaderStorageImageArrayNonUniformIndexing = VK_TRUE;
 	features.shaderStorageBufferArrayNonUniformIndexing = VK_TRUE;
+	features.descriptorBindingSampledImageUpdateAfterBind = VK_TRUE;
+	features.descriptorBindingUpdateUnusedWhilePending    = VK_TRUE;
+	features.descriptorBindingPartiallyBound              = VK_TRUE;
 	return features;
 }
 
@@ -274,6 +279,8 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 		              required_features12.samplerMirrorClampToEdge);
 		check_feature(features11.storageBuffer16BitAccess, "storageBuffer16BitAccess",
 		              WindowContext::RequiredVulkan11Features().storageBuffer16BitAccess);
+		check_feature(features11.shaderDrawParameters, "shaderDrawParameters",
+		              WindowContext::RequiredVulkan11Features().shaderDrawParameters);
 		check_feature(features12.storageBuffer8BitAccess, "storageBuffer8BitAccess",
 		              required_features12.storageBuffer8BitAccess);
 		check_feature(features12.timelineSemaphore, "timelineSemaphore",
@@ -295,6 +302,14 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 		check_feature(features12.shaderStorageBufferArrayNonUniformIndexing,
 		              "shaderStorageBufferArrayNonUniformIndexing",
 		              required_features12.shaderStorageBufferArrayNonUniformIndexing);
+		check_feature(features12.descriptorBindingSampledImageUpdateAfterBind,
+		              "descriptorBindingSampledImageUpdateAfterBind",
+		              required_features12.descriptorBindingSampledImageUpdateAfterBind);
+		check_feature(features12.descriptorBindingUpdateUnusedWhilePending,
+		              "descriptorBindingUpdateUnusedWhilePending",
+		              required_features12.descriptorBindingUpdateUnusedWhilePending);
+		check_feature(features12.descriptorBindingPartiallyBound, "descriptorBindingPartiallyBound",
+		              required_features12.descriptorBindingPartiallyBound);
 		check_feature(features13.robustImageAccess, "robustImageAccess");
 		check_feature(robustness2.nullDescriptor, "nullDescriptor");
 		check_feature(features13.dynamicRendering, "dynamicRendering",
@@ -325,9 +340,13 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 
 		check_feature(device_features2.features.shaderImageGatherExtended,
 		              "shaderImageGatherExtended");
+		check_feature(device_features2.features.shaderResourceMinLod, "shaderResourceMinLod");
 
 		check_feature(device_features2.features.independentBlend, "independentBlend");
 		check_feature(device_features2.features.tessellationShader, "tessellationShader");
+		check_feature(device_features2.features.drawIndirectFirstInstance,
+		              "drawIndirectFirstInstance");
+		check_feature(device_features2.features.multiDrawIndirect, "multiDrawIndirect");
 
 		if (!skip_device) {
 			auto available_extensions = EnumerateVulkan<vk::ExtensionProperties>(
@@ -492,6 +511,13 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 		maintenance8.pNext = supported_features2.pNext;
 		supported_features2.pNext = &maintenance8;
 	}
+	const bool shader_clock_extension =
+	    HasExtension(device_extensions, VK_KHR_SHADER_CLOCK_EXTENSION_NAME);
+	vk::PhysicalDeviceShaderClockFeaturesKHR shader_clock {};
+	if (shader_clock_extension) {
+		shader_clock.pNext = supported_features2.pNext;
+		supported_features2.pNext = &shader_clock;
+	}
 	const bool image_atomic_int64_extension =
 	    HasExtension(device_extensions, VK_EXT_SHADER_IMAGE_ATOMIC_INT64_EXTENSION_NAME);
 	vk::PhysicalDeviceShaderImageAtomicInt64FeaturesEXT image_atomic_int64 {};
@@ -499,6 +525,13 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 		image_atomic_int64.pNext = supported_features2.pNext;
 		supported_features2.pNext = &image_atomic_int64;
 	}
+	const bool conditional_rendering_extension =
+	    HasExtension(device_extensions, VK_EXT_CONDITIONAL_RENDERING_EXTENSION_NAME);
+	vk::PhysicalDeviceConditionalRenderingFeaturesEXT conditional_rendering {};
+	if (conditional_rendering_extension) {
+		conditional_rendering.pNext = supported_features2.pNext;
+		supported_features2.pNext   = &conditional_rendering;
+	}
 	const bool barycentric_extension =
 	    HasExtension(device_extensions, VK_KHR_FRAGMENT_SHADER_BARYCENTRIC_EXTENSION_NAME);
 	vk::PhysicalDeviceFragmentShaderBarycentricFeaturesKHR supported_barycentric {};
@@ -572,6 +605,13 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 	    feedback_extensions && feedback_layout.attachmentFeedbackLoopLayout &&
 	    feedback_dynamic.attachmentFeedbackLoopDynamicState;
 	graphics.maintenance8_enabled = maintenance8_extension && maintenance8.maintenance8;
+	// The device clock ticks at timestampPeriod nanoseconds.
+	if (shader_clock_extension && shader_clock.shaderDeviceClock) {
+		graphics.device_clock_period_ps = static_cast<uint32_t>(
+		    std::lround(graphics.physical_device_properties.limits.timestampPeriod * 1000.0f));
+	}
+	graphics.conditional_rendering_enabled =
+	    conditional_rendering_extension && conditional_rendering.conditionalRendering;
 	LOGF("Vulkan maintenance8 support: %s\n",
 	     graphics.maintenance8_enabled ? "true" : "false");
 	LOGF("Vulkan depth feedback support: %s\n",
@@ -597,10 +637,15 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 	device_features.depthClamp  = VK_TRUE;
 #endif
 	device_features.shaderStorageImageWriteWithoutFormat = VK_TRUE;
+	device_features.shaderStorageImageMultisample        =
+	    supported_features2.features.shaderStorageImageMultisample;
 	device_features.shaderImageGatherExtended            = VK_TRUE;
+	device_features.shaderResourceMinLod                 = VK_TRUE;
 	device_features.independentBlend                     = VK_TRUE;
 	device_features.dualSrcBlend                         = VK_TRUE;
 	device_features.tessellationShader                   = VK_TRUE;
+	device_features.drawIndirectFirstInstance            = VK_TRUE;
+	device_features.multiDrawIndirect                    = VK_TRUE;
 	device_features.sampleRateShading                    = VK_TRUE;
 	device_features.depthBiasClamp                       = VK_TRUE;
 	device_features.shaderClipDistance                   = VK_TRUE;
@@ -609,7 +654,22 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 	device_features.multiViewport                        = VK_TRUE;
 	device_features.fillModeNonSolid                      = VK_TRUE;
 	device_features.vertexPipelineStoresAndAtomics       = VK_TRUE;
+	// Storage image sample counts beyond 1x require the feature.
+	if (!device_features.shaderStorageImageMultisample) {
+		graphics.physical_device_properties.limits.storageImageSampleCounts =
+		    vk::SampleCountFlagBits::e1;
+	}
 	graphics.sample_rate_shading_enabled                 = true;
+	// Occlusion queries reset their queries from the host and count samples exactly where they can.
+	graphics.occlusion_queries_enabled =
+	    Config::OcclusionQueriesEnabled() && supported_features12.hostQueryReset;
+	graphics.occlusion_query_precise_enabled =
+	    graphics.occlusion_queries_enabled && supported_features2.features.occlusionQueryPrecise;
+	features12.hostQueryReset             = graphics.occlusion_queries_enabled;
+	device_features.occlusionQueryPrecise = graphics.occlusion_query_precise_enabled;
+	LOGF("Vulkan occlusion queries: %s, precise: %s\n",
+	     graphics.occlusion_queries_enabled ? "true" : "false",
+	     graphics.occlusion_query_precise_enabled ? "true" : "false");
 	device_features.shaderInt64 = VK_TRUE;
 	device_features.shaderFloat64 =
 	    supported_features2.features.shaderFloat64 &&
@@ -670,6 +730,23 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 		maintenance8.pNext = const_cast<void*>(create_info.pNext);
 		create_info.pNext = &maintenance8;
 	}
+	if (graphics.device_clock_period_ps != 0u) {
+		shader_clock.pNext = const_cast<void*>(create_info.pNext);
+		create_info.pNext = &shader_clock;
+	}
+	if (graphics.conditional_rendering_enabled) {
+		conditional_rendering.pNext = const_cast<void*>(create_info.pNext);
+		create_info.pNext = &conditional_rendering;
+	}
+	// The renderer only uses the Vulkan 1.3 core dynamic states, but RADV (Mesa 26.3) creates its
+	// PS epilog cache only when this EXT feature is enabled and crashes in vkCmdDraw when a
+	// depth- or stencil-writing fragment shader needs an epilog.
+	vk::PhysicalDeviceExtendedDynamicStateFeaturesEXT extended_dynamic_state {};
+	extended_dynamic_state.extendedDynamicState = VK_TRUE;
+	if (HasExtension(device_extensions, VK_EXT_EXTENDED_DYNAMIC_STATE_EXTENSION_NAME)) {
+		extended_dynamic_state.pNext = const_cast<void*>(create_info.pNext);
+		create_info.pNext = &extended_dynamic_state;
+	}
 	create_info.pQueueCreateInfos       = &queue_create_info;
 	create_info.queueCreateInfoCount    = 1;
 	create_info.enabledExtensionCount   = static_cast<uint32_t>(device_extensions.size());
@@ -1041,12 +1118,15 @@ void WindowContext::CreateVulkan() {
 			graphic_ctx.memory_budget_ext_enabled = true;
 		}
 		for (const auto* extension: {VK_KHR_MAINTENANCE_8_EXTENSION_NAME,
+		                             VK_KHR_SHADER_CLOCK_EXTENSION_NAME,
 		                             VK_EXT_SHADER_IMAGE_ATOMIC_INT64_EXTENSION_NAME,
 		                             VK_EXT_PROVOKING_VERTEX_EXTENSION_NAME,
 		                             VK_EXT_MESH_SHADER_EXTENSION_NAME,
 		                             VK_KHR_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_EXTENSION_NAME,
 		                             VK_EXT_DEPTH_RANGE_UNRESTRICTED_EXTENSION_NAME,
-		                             VK_KHR_FRAGMENT_SHADER_BARYCENTRIC_EXTENSION_NAME}) {
+		                             VK_KHR_FRAGMENT_SHADER_BARYCENTRIC_EXTENSION_NAME,
+		                             VK_EXT_CONDITIONAL_RENDERING_EXTENSION_NAME,
+		                             VK_EXT_EXTENDED_DYNAMIC_STATE_EXTENSION_NAME}) {
 			if (HasExtension(available_extensions, extension)) {
 				device_extensions.push_back(extension);
 			}

src/graphics/presentation/window/window.cpp

@@ -32,8 +32,6 @@
 #include <vector>
 #include <vulkan/vk_platform.h>
 
-#define STB_IMAGE_IMPLEMENTATION
-#define STBI_NO_SIMD
 #include "stb_image.h"
 
 // IWYU pragma: no_include <intrin.h>

src/graphics/shader/recompiler/ShaderRecompiler.cpp

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

src/graphics/shader/recompiler/ShaderRecompiler.h

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

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

@@ -169,6 +169,9 @@ uint32_t Builder::AddString(const char* text) {
 }
 
 void Builder::AddName(uint32_t target, const char* name) {
+	if (!m_emit_names) {
+		return;
+	}
 	std::vector<uint32_t> operands = {target};
 	AppendString(operands, name);
 	AppendInstruction(m_debug, spv::OpName, operands);

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

@@ -67,6 +67,7 @@ public:
 	}
 
 	uint32_t AddString(const char* text);
+	void EnableNames() { m_emit_names = true; }
 	void AddName(uint32_t target, const char* name);
 	template <typename... Args>
 	void AddAnnotation(spv::Op opcode, const Args&... operands) {
@@ -159,6 +160,7 @@ private:
 	std::map<std::string, uint32_t>           m_import_ids;
 	std::map<std::vector<uint32_t>, uint32_t> m_declaration_ids;
 	size_t                                    m_unpatched_phi_incomings = 0;
+	bool                                      m_emit_names              = false;
 };
 
 } // namespace Libs::Graphics::ShaderRecompiler::Spirv

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

@@ -161,6 +161,7 @@ void ValidateNativeProgram(const IR::Program& program, bool lds_storage) {
 		if (!kind.has_value()) {
 			Fail(program, "native shader plan has an invalid image class");
 		}
+		if (image.heap) continue;
 		if (image.mip_count == 0u ||
 		    (image.mip_mode != IR::ImageMipMode::Dynamic && image.mip_count != 1u)) {
 			Fail(program, "native shader plan has an invalid image mip descriptor count");
@@ -226,6 +227,7 @@ std::vector<uint32_t> EmitProgram(IR::Program& program, ShaderStageInputInfo inp
 		std::snprintf(name, sizeof(name), "rdna2-%016" PRIx64 "-block-byte-pc-plus-one",
 		              program.shader_hash);
 		state.debug_source = state.builder.AddString(name);
+		state.builder.EnableNames();
 	}
 	const auto* workgroup = ShaderWorkgroupInput(program.stage, input_info);
 	state.lane_count =

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

@@ -13,6 +13,8 @@ namespace Libs::Graphics::ShaderRecompiler::Spirv {
 struct Profile {
 	bool maintenance8         = false; // Dynamic texel offsets on image samples.
 	bool native_float_min_max = false; // NMin/NMax give the guest's NaN and signed-zero results.
+	// OpReadClockKHR Device tick in picoseconds; 0 without shaderDeviceClock.
+	uint32_t device_clock_period_ps = 0;
 };
 
 inline constexpr Profile DefaultProfile {};

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

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

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

@@ -30,6 +30,13 @@ bool PixelParameterIsCustom(const EmitterState& state, uint32_t attr) {
 	       ShaderPixelParameterIsCustom(*state.input_info.pixel, attr);
 }
 
+uint32_t MrtOutputType(EmitterState& state, uint32_t index) {
+	const auto& pixel = *state.input_info.pixel;
+	if ((pixel.target_sint_mask & (1u << index)) != 0u) return TypeI32Vector(state, 4);
+	if ((pixel.target_uint_mask & (1u << index)) != 0u) return TypeU32Vector(state, 4);
+	return TypeF32Vector(state, 4);
+}
+
 uint32_t OutputVariableForExport(const EmitterState& state, const IR::ExportInfo& exp) {
 	if (exp.kind == IR::ExportTargetKind::Position && exp.index == 0) {
 		return state.per_vertex_variable;
@@ -120,7 +127,14 @@ uint32_t ImageType(EmitterState& state, const IR::ImageResource& image) {
 	} else {
 		EXIT("invalid image resource class");
 	}
+	if (image.dimension == ImageDimension::Dim1D || image.dimension == ImageDimension::Dim1DArray) {
+		state.builder.RequireCapability(sampled == 1 ? spv::CapabilitySampled1D
+		                                             : spv::CapabilityImage1D);
+	}
 	const auto& info = ImageDimensionInfoFor(image.dimension);
+	if (sampled == 2 && info.multisampled != 0u) {
+		state.builder.RequireCapability(spv::CapabilityStorageImageMultisample);
+	}
 	const auto  scalar_type =
 	    image.atomic64 ? TypeU64(state) : ImageScalarType(state, image.numeric_class);
 	return state.builder.Type(spv::OpTypeImage, scalar_type,
@@ -144,14 +158,20 @@ void DecorateNonUniformDescriptor(EmitterState& state, uint32_t id, spv::Capabil
 	state.builder.RequireCapability(capability);
 	state.builder.AddAnnotation(spv::OpDecorate, id, spv::DecorationNonUniform);
 }
+
+ImageDefinition& ImageDescriptorDefinition(EmitterState& state, const IR::ImageResource& image,
+                                           IR::DescriptorBindingKind kind) {
+	return (image.heap ? state.heap_images : state.images)[IR::ImageBindingIndex(kind)];
+}
 } // namespace
 
 uint32_t LoadImageDescriptor(EmitterState& state, uint32_t resource, uint32_t mip,
                              uint32_t array_index) {
 	const auto pointer = ImageDescriptorPointer(state, resource, mip, array_index);
 	const auto image = state.builder.AllocateId();
-	const auto kind = *IR::DescriptorBindingForImage(state.program.info.images[resource]);
-	state.builder.AddFunction(spv::OpLoad, state.images[IR::ImageBindingIndex(kind)].type,
+	const auto& resource_info = state.program.info.images[resource];
+	const auto kind = *IR::DescriptorBindingForImage(resource_info);
+	state.builder.AddFunction(spv::OpLoad, ImageDescriptorDefinition(state, resource_info, kind).type,
 	                          image, pointer);
 	if (array_index != 0u) {
 		DecorateNonUniformDescriptor(state, image,
@@ -184,13 +204,13 @@ uint32_t MakeSampledImage(EmitterState& state, uint32_t resource, uint32_t sampl
 	EXIT_IF(image_resource.resource_class != IR::ImageResourceClass::Sampled);
 	const auto  image          = LoadImageDescriptor(state, resource, mip, array_index);
 	const auto  sampled_image = state.builder.AllocateId();
-	const auto kind = *IR::DescriptorBindingForImage(image_resource);
-	auto& sampled_type = state.images[IR::ImageBindingIndex(kind)].sampled_type;
-	if (sampled_type == 0) {
-		sampled_type = state.builder.Type(
-		    spv::OpTypeSampledImage, state.images[IR::ImageBindingIndex(kind)].type);
+	auto& definition = ImageDescriptorDefinition(state, image_resource,
+	                                             *IR::DescriptorBindingForImage(image_resource));
+	if (definition.sampled_type == 0) {
+		definition.sampled_type = state.builder.Type(spv::OpTypeSampledImage, definition.type);
 	}
-	state.builder.AddFunction(spv::OpSampledImage, sampled_type, sampled_image, image, sampler_id);
+	state.builder.AddFunction(spv::OpSampledImage, definition.sampled_type, sampled_image, image,
+	                          sampler_id);
 	if (array_index != 0u || sampler_dynamic) {
 		DecorateNonUniformDescriptor(state, sampled_image, spv::CapabilitySampledImageArrayNonUniformIndexing);
 	}
@@ -202,26 +222,31 @@ uint32_t ImageDescriptorPointer(EmitterState& state, uint32_t resource, uint32_t
 	const auto& image = state.program.info.images.at(resource);
 	EXIT_IF(mip >= image.mip_count);
 	const auto kind = IR::DescriptorBindingForImage(image);
-	EXIT_IF(!kind.has_value());
+	EXIT_IF(!kind.has_value() || (image.heap && array_index == 0u));
 	if (array_index == 0u) {
 		array_index = ConstantU32(state, image.descriptor_index + mip);
 	}
-	const auto& definition = state.images[IR::ImageBindingIndex(*kind)];
+	const auto& definition = ImageDescriptorDefinition(state, image, *kind);
 	return DescriptorElementPointer(state, definition.pointer_type, definition.variable,
 	                                array_index, *kind, resource,
 	                                "image descriptor array was not emitted");
 }
 
 void EmitStorageImageWrite(EmitterState& state, uint32_t resource, uint32_t mip_lod, uint32_t coord,
-                           uint32_t texel, uint32_t array_index) {
+                           uint32_t texel, uint32_t array_index, uint32_t sample) {
 	const auto& image = state.program.info.images.at(resource);
 	EXIT_IF(image.resource_class != IR::ImageResourceClass::Storage);
 	if (!image.atomic) {
 		state.builder.RequireCapability(spv::CapabilityStorageImageWriteWithoutFormat);
 	}
 	const auto EmitWrite = [&](uint32_t mip) {
-		state.builder.AddFunction(spv::OpImageWrite, LoadImageDescriptor(state, resource, mip, array_index),
-		                          coord, texel);
+		const auto descriptor = LoadImageDescriptor(state, resource, mip, array_index);
+		if (sample == 0u) {
+			state.builder.AddFunction(spv::OpImageWrite, descriptor, coord, texel);
+		} else {
+			state.builder.AddFunction(spv::OpImageWrite, descriptor, coord, texel,
+			                          spv::ImageOperandsSampleMask, sample);
+		}
 	};
 	if (image.mip_mode != IR::ImageMipMode::Dynamic) {
 		EmitWrite(0);

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

@@ -4,6 +4,8 @@
 
 #include <algorithm>
 #include <atomic>
+#include <bit>
+#include <optional>
 
 namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter {
 namespace {
@@ -61,7 +63,8 @@ uint32_t EmitBuiltinU32(EmitterState& state, IR::StageInputKind kind, uint32_t c
 	}
 	if (kind == IR::StageInputKind::VertexIndex || kind == IR::StageInputKind::InstanceIndex ||
 	    kind == IR::StageInputKind::InvocationId || kind == IR::StageInputKind::PrimitiveId ||
-	    kind == IR::StageInputKind::Layer || kind == IR::StageInputKind::SampleId) {
+	    kind == IR::StageInputKind::Layer || kind == IR::StageInputKind::SampleId ||
+	    kind == IR::StageInputKind::BaseVertex || kind == IR::StageInputKind::BaseInstance) {
 		const auto value = state.builder.AllocateId();
 		const auto bits  = state.builder.AllocateId();
 		state.builder.AddFunction(spv::OpLoad, TypeI32(state), value, variable);
@@ -112,6 +115,18 @@ uint32_t EmitBuiltinU32(EmitterState& state, IR::StageInputKind kind, uint32_t c
 	return EmitInputComponentU32(state, kind, component);
 }
 
+// Whether EXEC enables a source lane within the caller's 32-lane half.
+uint32_t EmitSourceLaneActive(ValueEmitContext& ctx, uint32_t exec, uint32_t lane) {
+	auto&      state  = ctx.state;
+	const auto scope  = ConstantU32(state, spv::ScopeSubgroup);
+	const auto ballot = EmitNative<spv::OpGroupNonUniformBallot, IR::Type::U32x4>(state, scope, exec);
+	if (ctx.half != 0) {
+		lane = EmitAndConstant(state, lane, 31);
+	}
+	return EmitNative<spv::OpGroupNonUniformBallotBitExtract, IR::Type::U1>(state, scope, ballot,
+	                                                                        lane);
+}
+
 uint32_t EmitDppWriteCondition(ValueEmitContext& ctx, const IR::DppMoveFlags& flags,
                                uint32_t exec) {
 	auto&    state    = ctx.state;
@@ -131,15 +146,98 @@ uint32_t EmitDppWriteCondition(ValueEmitContext& ctx, const IR::DppMoveFlags& fl
 	enable(flags.bank_mask, 2u);
 	enable(flags.row_mask, 4u);
 	// Only row shifts (0x101-0x10f, 0x111-0x11f) can vacate a lane; other controls always read one.
-	const bool row_shift = !flags.dpp8 && flags.control > 0x100u && flags.control < 0x120u &&
+	const bool row_shift = flags.control > 0x100u && flags.control < 0x120u &&
 	                       flags.control != 0x110u;
-	if (!flags.bound_control && row_shift) {
+	if (flags.bound_control || (flags.fetch_inactive && !row_shift)) {
+		return writable;
+	}
+	// Without BOUND_CTRL, a vacated source or, without FI, an EXEC-disabled one disables the write.
+	const auto target = EmitDppTargetLane(state, flags);
+	if (row_shift) {
+		writable = Binary(state, spv::OpLogicalAnd, TypeBool(state), writable, target.valid);
+	}
+	if (!flags.fetch_inactive) {
+		// Vacated lanes target outside their row; wrapping keeps the ballot index in range.
+		auto source = target.lane;
+		if (row_shift) {
+			lane   = lane != 0 ? lane : EmitSubgroupLocalInvocationId(state);
+			source = EmitNative<spv::OpBitFieldInsert, IR::Type::U32>(
+			    state, lane, source, ConstantU32(state, 0), ConstantU32(state, 4));
+		}
 		writable = Binary(state, spv::OpLogicalAnd, TypeBool(state), writable,
-		                  EmitDppTargetLane(state, flags).valid);
+		                  EmitSourceLaneActive(ctx, exec, source));
 	}
 	return writable;
 }
 
+uint32_t LoadVertexParameter(EmitterState& state, const InputBinding& input, uint32_t vertex,
+                             uint32_t chan) {
+	const auto pointer = state.builder.AllocateId();
+	const auto value   = state.builder.AllocateId();
+	state.builder.AddFunction(spv::OpAccessChain,
+	                          TypePointer(state, spv::StorageClassInput, TypeF32(state)), pointer,
+	                          input.variable_id, ConstantU32(state, vertex),
+	                          ConstantU32(state, chan & 3u));
+	state.builder.AddFunction(spv::OpLoad, TypeF32(state), value, pointer);
+	return value;
+}
+
+// Interpolates a pixel input channel to f32. A half (0 low, 1 high) selects one f16 of a packed
+// FP16_INTERP_MODE dword before blending; smooth inputs read by halves are always per-vertex.
+uint32_t EmitPixelAttribute(EmitterState& state, const InputBinding& input, uint32_t attr,
+                            uint32_t chan, std::optional<uint32_t> half) {
+	const auto select = [&](uint32_t value) {
+		if (!half) {
+			return value;
+		}
+		const auto bits     = state.builder.AllocateId();
+		const auto unpacked = state.builder.AllocateId();
+		const auto word     = state.builder.AllocateId();
+		state.builder.AddFunction(spv::OpBitcast, TypeU32(state), bits, value);
+		state.builder.AddFunction(spv::OpExtInst, TypeF32Vector(state, 2), unpacked,
+		                          GlslStd450(state), GLSLstd450UnpackHalf2x16, bits);
+		state.builder.AddFunction(spv::OpCompositeExtract, TypeF32(state), word, unpacked, *half);
+		return word;
+	};
+	if (!input.per_vertex) {
+		const auto vector    = state.builder.AllocateId();
+		const auto component = state.builder.AllocateId();
+		state.builder.AddFunction(spv::OpLoad, TypeF32Vector(state, 4), vector, input.variable_id);
+		state.builder.AddFunction(spv::OpCompositeExtract, TypeF32(state), component, vector,
+		                          chan & 3u);
+		return select(component);
+	}
+	if (PixelParameterIsFlat(state, attr)) {
+		const auto vertex =
+		    state.input_info.pixel->parameter_mode == ShaderPixelParameterMode::LastVertex ? 2u : 0u;
+		return select(LoadVertexParameter(state, input, vertex, chan));
+	}
+	const auto barycentric_kind = state.input_info.pixel->ps_no_perspective
+	                                  ? IR::StageInputKind::BaryCoordNoPerspective
+	                                  : IR::StageInputKind::BaryCoordSmooth;
+	const auto barycentric      = InputVariableForKind(state, barycentric_kind);
+	uint32_t   sum              = 0;
+	for (uint32_t vertex = 0; vertex < 3u; vertex++) {
+		const auto pointer = state.builder.AllocateId();
+		const auto weight  = state.builder.AllocateId();
+		const auto product = state.builder.AllocateId();
+		state.builder.AddFunction(spv::OpAccessChain,
+		                          TypePointer(state, spv::StorageClassInput, TypeF32(state)),
+		                          pointer, barycentric, ConstantU32(state, vertex));
+		state.builder.AddFunction(spv::OpLoad, TypeF32(state), weight, pointer);
+		state.builder.AddFunction(spv::OpFMul, TypeF32(state), product,
+		                          select(LoadVertexParameter(state, input, vertex, chan)), weight);
+		if (vertex == 0u) {
+			sum = product;
+		} else {
+			const auto next = state.builder.AllocateId();
+			state.builder.AddFunction(spv::OpFAdd, TypeF32(state), next, sum, product);
+			sum = next;
+		}
+	}
+	return sum;
+}
+
 uint32_t EmitAttribute(EmitterState& state, uint32_t attr, uint32_t chan) {
 	const auto* input = InputBindingForParameter(state, attr);
 	if (input == nullptr || input->variable_id == 0) {
@@ -148,57 +246,9 @@ uint32_t EmitAttribute(EmitterState& state, uint32_t attr, uint32_t chan) {
 	if (state.program.stage == ShaderType::Vertex || state.program.stage == ShaderType::Local) {
 		return EmitVertexParameterComponentU32(state, *input, chan & 3u);
 	}
-	const auto load_per_vertex = [&](uint32_t vertex) {
-		const auto pointer = state.builder.AllocateId();
-		const auto value   = state.builder.AllocateId();
-		state.builder.AddFunction(
-		    spv::OpAccessChain, TypePointer(state, spv::StorageClassInput, TypeF32(state)), pointer,
-		    input->variable_id, ConstantU32(state, vertex), ConstantU32(state, chan & 3u));
-		state.builder.AddFunction(spv::OpLoad, TypeF32(state), value, pointer);
-		return value;
-	};
-	if (input->per_vertex) {
-		if (PixelParameterIsFlat(state, attr)) {
-			const auto value = load_per_vertex(
-			    state.input_info.pixel->parameter_mode == ShaderPixelParameterMode::LastVertex ? 2u : 0u);
-			const auto bits  = state.builder.AllocateId();
-			state.builder.AddFunction(spv::OpBitcast, TypeU32(state), bits, value);
-			return bits;
-		}
-		const auto barycentric_kind = state.input_info.pixel->ps_no_perspective
-		                                  ? IR::StageInputKind::BaryCoordNoPerspective
-		                                  : IR::StageInputKind::BaryCoordSmooth;
-		const auto barycentric      = InputVariableForKind(state, barycentric_kind);
-		uint32_t   sum              = 0;
-		for (uint32_t vertex = 0; vertex < 3u; vertex++) {
-			const auto pointer = state.builder.AllocateId();
-			const auto weight  = state.builder.AllocateId();
-			const auto product = state.builder.AllocateId();
-			state.builder.AddFunction(spv::OpAccessChain,
-			                          TypePointer(state, spv::StorageClassInput, TypeF32(state)),
-			                          pointer, barycentric, ConstantU32(state, vertex));
-			state.builder.AddFunction(spv::OpLoad, TypeF32(state), weight, pointer);
-			state.builder.AddFunction(spv::OpFMul, TypeF32(state), product, load_per_vertex(vertex),
-			                          weight);
-			if (vertex == 0u) {
-				sum = product;
-			} else {
-				const auto next = state.builder.AllocateId();
-				state.builder.AddFunction(spv::OpFAdd, TypeF32(state), next, sum, product);
-				sum = next;
-			}
-		}
-		const auto bits = state.builder.AllocateId();
-		state.builder.AddFunction(spv::OpBitcast, TypeU32(state), bits, sum);
-		return bits;
-	}
-	const auto vector    = state.builder.AllocateId();
-	const auto component = state.builder.AllocateId();
-	const auto bits      = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpLoad, TypeF32Vector(state, 4), vector, input->variable_id);
-	state.builder.AddFunction(spv::OpCompositeExtract, TypeF32(state), component, vector,
-	                          chan & 3u);
-	state.builder.AddFunction(spv::OpBitcast, TypeU32(state), bits, component);
+	const auto value = EmitPixelAttribute(state, *input, attr, chan, std::nullopt);
+	const auto bits  = state.builder.AllocateId();
+	state.builder.AddFunction(spv::OpBitcast, TypeU32(state), bits, value);
 	return bits;
 }
 
@@ -209,25 +259,16 @@ uint32_t EmitInterpolationParameter(ValueEmitContext& ctx, uint32_t attr, uint32
 	if (!input->per_vertex) {
 		return EmitAttribute(ctx.state, attr, chan);
 	}
-	const auto load_vertex = [&](uint32_t vertex) {
-		const auto pointer = state.builder.AllocateId();
-		const auto value   = state.builder.AllocateId();
-		state.builder.AddFunction(
-		    spv::OpAccessChain, TypePointer(state, spv::StorageClassInput, TypeF32(state)), pointer,
-		    input->variable_id, ConstantU32(state, vertex), ConstantU32(state, chan & 3u));
-		state.builder.AddFunction(spv::OpLoad, TypeF32(state), value, pointer);
-		return value;
-	};
-
 	uint32_t selected_vertex = (mode + 1u) % 3u;
 	if (mode == 2u && PixelParameterIsFlat(state, attr)) {
 		selected_vertex =
 		    state.input_info.pixel->parameter_mode == ShaderPixelParameterMode::LastVertex ? 2u : 0u;
 	}
-	uint32_t value = load_vertex(selected_vertex);
+	uint32_t value = LoadVertexParameter(state, *input, selected_vertex, chan);
 	if (!PixelParameterIsCustom(state, attr) && mode < 2u) {
 		const auto delta = state.builder.AllocateId();
-		state.builder.AddFunction(spv::OpFSub, TypeF32(state), delta, value, load_vertex(0));
+		state.builder.AddFunction(spv::OpFSub, TypeF32(state), delta, value,
+		                          LoadVertexParameter(state, *input, 0, chan));
 		value = delta;
 	}
 	const auto bits = state.builder.AllocateId();
@@ -251,9 +292,10 @@ uint32_t ExportRawComponent(ValueEmitContext& ctx, uint32_t vector, uint32_t com
 }
 
 uint32_t ExportVector(ValueEmitContext& ctx, uint32_t data, const IR::ExportInfo& exp,
-                      bool uint_output) {
-	auto& state = ctx.state;
-	if (exp.compr && !uint_output) {
+                      uint32_t vector_type) {
+	auto&      state          = ctx.state;
+	const bool integer_output = vector_type != TypeF32Vector(state, 4);
+	if (exp.compr && !integer_output) {
 		auto unpack = GLSLstd450UnpackHalf2x16;
 		switch (MrtOutputMode(state, exp)) {
 			case 5: unpack = GLSLstd450UnpackUnorm2x16; break;
@@ -290,9 +332,11 @@ uint32_t ExportVector(ValueEmitContext& ctx, uint32_t data, const IR::ExportInfo
 		    ConstantU32(state, 0),
 		    ConstantU32(state, 0),
 		    ConstantU32(state, 0),
-		    ConstantU32(state, uint_output ? 1u : 0x3f800000u),
+		    ConstantU32(state, integer_output ? 1u : 0x3f800000u),
 		};
 		if (exp.compr) {
+			const auto extract = vector_type == TypeU32Vector(state, 4) ? spv::OpBitFieldUExtract
+			                                                            : spv::OpBitFieldSExtract;
 			for (uint32_t pair = 0; pair < 2u; pair++) {
 				if ((exp.en & (3u << (pair * 2u))) == 0u) {
 					continue;
@@ -304,8 +348,8 @@ uint32_t ExportVector(ValueEmitContext& ctx, uint32_t data, const IR::ExportInfo
 						continue;
 					}
 					raw[component] = state.builder.AllocateId();
-					state.builder.AddFunction(spv::OpBitFieldUExtract, TypeU32(state), raw[component],
-					                          packed, ConstantU32(state, lane * 16u),
+					state.builder.AddFunction(extract, TypeU32(state), raw[component], packed,
+					                          ConstantU32(state, lane * 16u),
 					                          ConstantU32(state, 16));
 				}
 			}
@@ -320,12 +364,10 @@ uint32_t ExportVector(ValueEmitContext& ctx, uint32_t data, const IR::ExportInfo
 		state.builder.AddFunction(spv::OpCompositeConstruct, TypeU32Vector(state, 4), data,
 		                          raw[0], raw[1], raw[2], raw[3]);
 	}
-	if (uint_output) {
+	if (vector_type == TypeU32Vector(state, 4)) {
 		return data;
 	}
-	const auto vector = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpBitcast, TypeF32Vector(state, 4), vector, data);
-	return vector;
+	return Unary(state, spv::OpBitcast, vector_type, data);
 }
 
 uint32_t DistancePointer(EmitterState& state, IR::StageOutputKind kind, uint32_t index) {
@@ -523,11 +565,11 @@ void EmitSetAttribute(ValueEmitContext& ctx, const IR::Inst& inst) {
 		if (state.program.stage != ShaderType::Mesh && variable == 0) {
 			return;
 		}
-		const bool uint_output = MrtOutputMode(state, exp) == 7u;
-		const auto vector_type = uint_output ? TypeU32Vector(state, 4) : TypeF32Vector(state, 4);
-		auto       value       = ExportVector(ctx, data, exp, uint_output);
+		const auto vector_type = MrtOutputMode(state, exp) != 0u ? MrtOutputType(state, exp.index)
+		                                                         : TypeF32Vector(state, 4);
+		auto       value       = ExportVector(ctx, data, exp, vector_type);
 		if (state.program.stage == ShaderType::Pixel && exp.kind == IR::ExportTargetKind::Mrt &&
-		    exp.index == 0 && !uint_output &&
+		    exp.index == 0 && vector_type == TypeF32Vector(state, 4) &&
 		    state.input_info.pixel->alpha_blend_source != ShaderAlphaBlendSource::None) {
 			// Broadcast logical alpha before swizzling the primary output.
 			const auto blend_output =
@@ -671,9 +713,16 @@ void EmitBarrier(EmitterState& state) {
 		return;
 	}
 	const auto memory_scope = tessellation ? spv::ScopeInvocation : spv::ScopeWorkgroup;
-	const auto memory = state.lds_storage_class == spv::StorageClassStorageBuffer
-	                        ? spv::MemorySemanticsUniformMemoryMask
-	                        : spv::MemorySemanticsWorkgroupMemoryMask;
+	auto memory = state.lds_storage_class == spv::StorageClassStorageBuffer
+	                  ? spv::MemorySemanticsUniformMemoryMask
+	                  : spv::MemorySemanticsWorkgroupMemoryMask;
+	// Shaders read other waves' buffer and image stores with plain loads right after S_BARRIER.
+	const auto written = [](const auto& resource) { return resource.written; };
+	if (std::ranges::any_of(state.program.info.buffers, written) ||
+	    std::ranges::any_of(state.program.info.images, written)) {
+		memory = memory | spv::MemorySemanticsUniformMemoryMask |
+		         spv::MemorySemanticsImageMemoryMask;
+	}
 	const auto semantics = tessellation ? spv::MemorySemanticsMaskNone
 	                                    : spv::MemorySemanticsAcquireReleaseMask | memory;
 	state.builder.AddFunction(spv::OpControlBarrier, ConstantU32(state, spv::ScopeWorkgroup),
@@ -771,6 +820,13 @@ uint32_t EmitBallot(ValueEmitContext& ctx, IR::Value predicate) {
 	return ctx.Ballot(predicate);
 }
 
+uint32_t EmitLiveMask(EmitterState& state) {
+	const auto helper =
+	    EmitNative<spv::OpLoad, IR::Type::U1>(state, state.helper_invocation_variable);
+	return EmitNative<spv::OpGroupNonUniformBallot, IR::Type::U32x4>(
+	    state, ConstantU32(state, spv::ScopeSubgroup), EmitLogicalNotBool(state, helper));
+}
+
 uint32_t EmitReadFirstLane(ValueEmitContext& ctx, const IR::Inst& inst) {
 	const auto ballot = ctx.Ballot(inst.Arg(1));
 	const auto lane   = ctx.FirstLane(ballot);
@@ -781,18 +837,28 @@ uint32_t EmitReadLane(ValueEmitContext& ctx, const IR::Inst& inst) {
 	return ctx.Shuffle(inst, 0, ctx.Arg(inst, 1));
 }
 
-uint32_t EmitSubgroupUMinU32(ValueEmitContext& ctx, const IR::Inst& inst) {
+template <spv::Op Reduce, auto Combine>
+static uint32_t EmitSubgroupReduceU32(ValueEmitContext& ctx, const IR::Inst& inst) {
 	auto& state = ctx.state;
 	auto value = ctx.Arg(inst, 0);
 	if (ctx.other_half != nullptr) {
-		value = EmitUMin32(state, ctx.HalfArg(inst, 0, 0), ctx.HalfArg(inst, 0, 1));
+		value = Combine(state, ctx.HalfArg(inst, 0, 0), ctx.HalfArg(inst, 0, 1));
 	}
 	const auto result = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpGroupNonUniformUMin, TypeU32(state), result,
-	                          ConstantU32(state, spv::ScopeSubgroup), spv::GroupOperationReduce, value);
+	state.builder.AddFunction(Reduce, TypeU32(state), result,
+	                          ConstantU32(state, spv::ScopeSubgroup), spv::GroupOperationReduce,
+	                          value);
 	return result;
 }
 
+uint32_t EmitSubgroupUMinU32(ValueEmitContext& ctx, const IR::Inst& inst) {
+	return EmitSubgroupReduceU32<spv::OpGroupNonUniformUMin, EmitUMin32>(ctx, inst);
+}
+
+uint32_t EmitSubgroupUMaxU32(ValueEmitContext& ctx, const IR::Inst& inst) {
+	return EmitSubgroupReduceU32<spv::OpGroupNonUniformUMax, EmitUMax32>(ctx, inst);
+}
+
 uint32_t EmitWriteLane(ValueEmitContext& ctx, const IR::Inst& inst) {
 	auto&      state = ctx.state;
 	const auto hit   = state.builder.AllocateId();
@@ -803,27 +869,23 @@ uint32_t EmitWriteLane(ValueEmitContext& ctx, const IR::Inst& inst) {
 }
 
 uint32_t EmitPermlane16U32(ValueEmitContext& ctx, const IR::Inst& inst) {
-	auto&      state     = ctx.state;
-	const auto flags     = inst.Flags<IR::PermlaneFlags>();
-	const auto subid     = EmitSubgroupLocalInvocationId(state);
-	const auto row       = state.builder.AllocateId();
-	const auto row_value = state.builder.AllocateId();
-	const auto lane      = state.builder.AllocateId();
-	const auto lane8     = state.builder.AllocateId();
-	const auto shift     = state.builder.AllocateId();
-	const auto upper     = state.builder.AllocateId();
-	const auto selected  = state.builder.AllocateId();
-	const auto shifted   = state.builder.AllocateId();
-	const auto index     = state.builder.AllocateId();
-	const auto target    = state.builder.AllocateId();
+	auto&      state    = ctx.state;
+	const auto flags    = inst.Flags<IR::PermlaneFlags>();
+	const auto subid    = EmitSubgroupLocalInvocationId(state);
+	const auto row      = state.builder.AllocateId();
+	const auto lane     = state.builder.AllocateId();
+	const auto lane8    = state.builder.AllocateId();
+	const auto shift    = state.builder.AllocateId();
+	const auto upper    = state.builder.AllocateId();
+	const auto selected = state.builder.AllocateId();
+	const auto shifted  = state.builder.AllocateId();
+	const auto index    = state.builder.AllocateId();
+	const auto target   = state.builder.AllocateId();
 	state.builder.AddFunction(spv::OpBitwiseAnd, TypeU32(state), row, subid,
 	                          ConstantU32(state, 0xfffffff0u));
-	if (flags.x16) {
-		state.builder.AddFunction(spv::OpBitwiseXor, TypeU32(state), row_value, row,
-		                          ConstantU32(state, 16));
-	} else {
-		state.builder.AddFunction(spv::OpCopyObject, TypeU32(state), row_value, row);
-	}
+	const auto row_value =
+	    flags.x16 ? Binary(state, spv::OpBitwiseXor, TypeU32(state), row, ConstantU32(state, 16))
+		          : row;
 	state.builder.AddFunction(spv::OpBitwiseAnd, TypeU32(state), lane, subid,
 	                          ConstantU32(state, 15));
 	state.builder.AddFunction(spv::OpBitwiseAnd, TypeU32(state), lane8, lane,
@@ -839,20 +901,23 @@ uint32_t EmitPermlane16U32(ValueEmitContext& ctx, const IR::Inst& inst) {
 	                          ConstantU32(state, 15));
 	state.builder.AddFunction(spv::OpBitwiseOr, TypeU32(state), target, row_value, index);
 	const auto shuffled = ctx.Shuffle(inst, 0, target);
-	uint32_t   result   = shuffled;
-	if (!flags.fetch_inactive) {
-		const auto source_exec = ctx.Shuffle(inst, 3, target);
-		result                 = state.builder.AllocateId();
-		state.builder.AddFunction(spv::OpSelect, TypeU32(state), result, source_exec, shuffled,
-		                          ConstantU32(state, 0));
+	if (flags.fetch_inactive) {
+		return shuffled;
 	}
-	return result;
+	return Select(state, TypeU32(state), EmitSourceLaneActive(ctx, ctx.Arg(inst, 3), target),
+	              shuffled, ctx.Arg(inst, 4));
 }
 
 uint32_t EmitGetAttribute(ValueEmitContext& ctx, const IR::Inst& inst) {
 	return EmitAttribute(ctx.state, inst.Arg(0).U32(), inst.Arg(1).U32());
 }
 
+uint32_t EmitGetAttributeF16(ValueEmitContext& ctx, const IR::Inst& inst) {
+	const auto attr = inst.Arg(0).U32();
+	return EmitPixelAttribute(ctx.state, *InputBindingForParameter(ctx.state, attr), attr,
+	                          inst.Arg(1).U32(), inst.Arg(2).U32());
+}
+
 uint32_t EmitGetInterpolationParameter(ValueEmitContext& ctx, const IR::Inst& inst) {
 	return EmitInterpolationParameter(ctx, inst.Arg(0).U32(), inst.Arg(1).U32(), inst.Arg(2).U32());
 }
@@ -863,6 +928,22 @@ uint32_t EmitGetShaderBase(ValueEmitContext& ctx) {
 	return ctx.Def(IR::Value(uint64_t {0}));
 }
 
+uint32_t EmitReadRealtimeClock(ValueEmitContext& ctx, const IR::Inst& inst) {
+	auto&      state  = ctx.state;
+	const auto period = state.profile.device_clock_period_ps;
+	if (period == 0u) {
+		ctx.Fail(inst, "requires shaderDeviceClock");
+	}
+	state.builder.RequireExtension("SPV_KHR_shader_clock");
+	state.builder.RequireCapability(spv::CapabilityShaderClockKHR);
+	const auto clock =
+	    Unary(state, spv::OpReadClockKHR, TypeU64(state), ConstantU32(state, spv::ScopeDevice));
+	// S_MEMREALTIME counts at 100 MHz. Shift the host clock by the most bits that keep its tick at
+	// most 15000 ps (2/3 of 100 MHz): 1 GHz shifts by 3 to 125 MHz; 100 MHz and slower stay.
+	const uint32_t shift = std::bit_width(7500u / period);
+	return EmitShiftRightLogical64(ctx, clock, IR::Value(shift));
+}
+
 void EmitUnreachable(ValueEmitContext& ctx, const IR::Inst& inst) {
 	ctx.Fail(inst, "must be lowered before SPIR-V emission");
 }

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

@@ -38,7 +38,8 @@ bool HasFlag(const IR::MemoryInfo& mem, uint32_t flag) {
 
 uint32_t AddressU32(ValueEmitContext& ctx, const IR::MemoryInfo& mem, const IR::Inst& address,
                     uint32_t component) {
-	const auto layout = Decoder::ImageAddressComponentLayout(mem.image_sample_flags, component);
+	const auto layout = Decoder::ImageAddressComponentLayout(mem.image_sample_flags,
+	                                                         mem.image_dimension, component);
 	const auto packed = layout.bit_offset / 32u;
 	if (packed >= address.NumArgs()) return ConstantU32(ctx.state, 0);
 	auto value = ctx.Def(address.Arg(packed));
@@ -55,10 +56,11 @@ uint32_t AddressU32(ValueEmitContext& ctx, const IR::MemoryInfo& mem, const IR::
 
 uint32_t AddressF32(ValueEmitContext& ctx, const IR::MemoryInfo& mem, const IR::Inst& address,
                     uint32_t component) {
-	const auto value = AddressU32(ctx, mem, address, component);
-	return Decoder::ImageAddressComponentLayout(mem.image_sample_flags, component).bit_width == 16u
-	           ? EmitF16BitsToF32(ctx.state, value)
-	           : Unary(ctx.state, spv::OpBitcast, TypeF32(ctx.state), value);
+	const auto value  = AddressU32(ctx, mem, address, component);
+	const auto layout = Decoder::ImageAddressComponentLayout(mem.image_sample_flags,
+	                                                         mem.image_dimension, component);
+	return layout.bit_width == 16u ? EmitF16BitsToF32(ctx.state, value)
+	                               : Unary(ctx.state, spv::OpBitcast, TypeF32(ctx.state), value);
 }
 
 ImageSampleLayout Layout(const IR::MemoryInfo& mem) {
@@ -77,6 +79,7 @@ ImageSampleLayout Layout(const IR::MemoryInfo& mem) {
 	layout.coord = cursor;
 	cursor += info.coordinate_components;
 	if (HasFlag(mem, Decoder::ImageSampleFlagLod)) layout.lod = cursor++;
+	if (HasFlag(mem, Decoder::ImageSampleFlagLodClamp)) layout.clamp = cursor++;
 	return layout;
 }
 
@@ -389,6 +392,7 @@ Format::BufferFormatInfo ImageConversionFormat(const EmitterState&   state,
 	const auto format = image.conversion_format;
 	if (format == Prospero::BufferFormat::kInvalid) return {};
 	const auto info = Format::GetFormatInfo(format);
+	if (IR::ReadsDepthBits(image)) return info;
 	const bool raw_signed = IR::RawSignedStorageFormat(image, format) != Prospero::BufferFormat::kInvalid;
 	EXIT_IF((Prospero::RemapTextureFormat(format) == format && !raw_signed) || info.component_count == 0u ||
 	        info.component_count > 4u);
@@ -424,6 +428,7 @@ uint32_t UnpackImageTexel(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uint
 
 	const auto numeric_class = ctx.state.program.info.images[mem.resource].numeric_class;
 	const auto scalar_type   = ImageScalarType(ctx.state, numeric_class);
+	const bool integer       = info.packed_bitfield || info.type == Format::ComponentType::Uint;
 	uint32_t   packed        = 0;
 	if (info.packed_bitfield) {
 		packed = ctx.state.builder.AllocateId();
@@ -441,8 +446,16 @@ uint32_t UnpackImageTexel(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uint
 			ctx.state.builder.AddFunction(spv::OpCompositeExtract, scalar_type,
 			                              components[component], texel, component);
 			// UNorm backing preserves the guest's filtering; scaling precedes swizzle constants.
+			const auto unorm_max = static_cast<float>((1u << info.component_bits[component]) - 1u);
 			components[component] = Binary(ctx.state, spv::OpFMul, scalar_type, components[component],
-			                               ConstantF32Value(ctx.state, 255.0f));
+			                               ConstantF32Value(ctx.state, unorm_max));
+			if (integer) {
+				// Depth bits: round back to the stored UNorm code.
+				components[component] =
+				    Unary(ctx.state, spv::OpConvertFToU, TypeU32(ctx.state),
+				          Binary(ctx.state, spv::OpFAdd, scalar_type, components[component],
+				                 ConstantF32Value(ctx.state, 0.5f)));
+			}
 		}
 	}
 	for (uint32_t component = info.component_count; component < 4u; component++) {
@@ -454,20 +467,19 @@ uint32_t UnpackImageTexel(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uint
 	for (uint32_t component = 0; component < 4u; component++) {
 		const auto selector = (swizzle >> (component * 3u)) & 7u;
 		if (selector == 1u) {
-			selected[component] = info.packed_bitfield
+			selected[component] = integer
 			    ? ConstantU32(ctx.state, Format::FormattedConstantBits(info, Format::FormattedSourceKind::One))
 			    : ConstantF32Value(ctx.state, 1.0f);
 		} else if (selector >= 4u) {
 			selected[component] = components[selector - 4u];
 		} else {
-			selected[component] = info.packed_bitfield ? ConstantU32(ctx.state, 0u)
-			                                          : ZeroF32(ctx.state);
+			selected[component] = integer ? ConstantU32(ctx.state, 0u) : ZeroF32(ctx.state);
 		}
 	}
 	const auto result = ctx.state.builder.AllocateId();
 	ctx.state.builder.AddFunction(spv::OpCompositeConstruct,
-	                              ImageVectorType(ctx.state, numeric_class, 4), result, selected[0],
-	                              selected[1], selected[2], selected[3]);
+	                              integer ? TypeU32Vector(ctx.state, 4) : TypeF32Vector(ctx.state, 4),
+	                              result, selected[0], selected[1], selected[2], selected[3]);
 	return info.type == Format::ComponentType::Unorm
 	           ? Unary(ctx.state, spv::OpBitcast, TypeF32Vector(ctx.state, 4), result) : result;
 }
@@ -691,6 +703,21 @@ auto EmitImageAccess(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t resul
 	    image.indirect_resources.size() < 2u) {
 		ctx.Fail(inst, "has no indirect image runtime mapping");
 	}
+	if (image.heap) {
+		const auto entry = EmitImageHeapEntry(state, key, image.indirect_mapping_offset);
+		const auto kinds = static_cast<uint32_t>(image.indirect_resources.size()) - 1u;
+		// One kind leaves every slot word a bare element.
+		if (kinds == 1u) return emit(image.indirect_resources[1], entry);
+		const auto element = Binary(state, spv::OpBitwiseAnd, TypeU32(state), entry,
+		                            ConstantU32(state, (1u << IR::ImageHeapKindShift) - 1u));
+		const auto EmitKind = [&](uint32_t kind) {
+			return emit(image.indirect_resources[kind + 1u], element);
+		};
+		return EmitIndexSwitch(state,
+		                       Binary(state, spv::OpShiftRightLogical, TypeU32(state), entry,
+		                              ConstantU32(state, IR::ImageHeapKindShift)),
+		                       kinds, result_type, EmitKind);
+	}
 	const auto selected = EmitIndirectResourceIndex(
 	    state, key, image.indirect_mapping_offset, image.indirect_search_iterations, 0u);
 	const auto SameType = [&](uint32_t lhs, uint32_t rhs) {
@@ -786,6 +813,25 @@ uint32_t EmitImageResult(ValueEmitContext& ctx, const IR::Inst& inst, const Emit
 	return mixed ? value : result_bits(mem.resource, value);
 }
 
+// Ordinals outside the table select candidates[0]. Past ordinal 0, a consecutive table is an
+// offset of ordinal - 1, which wraps at 0 so that one unsigned compare checks both ends.
+uint32_t SamplerArrayIndex(EmitterState& state, uint32_t ordinal,
+                           const std::vector<uint32_t>& candidates) {
+	const auto count = static_cast<uint32_t>(candidates.size());
+	const auto gap = std::adjacent_find(candidates.begin() + 1, candidates.end(),
+	    [](uint32_t lhs, uint32_t rhs) { return rhs != lhs + 1u; });
+	if (count == 1u || gap != candidates.end()) {
+		return EmitIndexSwitch(state, ordinal, count, TypeU32(state),
+		                       [&](uint32_t item) { return ConstantU32(state, candidates[item]); });
+	}
+	const auto offset = Binary(state, spv::OpISub, TypeU32(state), ordinal, ConstantU32(state, 1u));
+	const auto in_table = Binary(state, spv::OpULessThan, TypeBool(state), offset,
+	                             ConstantU32(state, count - 1u));
+	const auto index = Binary(state, spv::OpIAdd, TypeU32(state), offset,
+	                          ConstantU32(state, candidates[1]));
+	return Select(state, TypeU32(state), in_table, index, ConstantU32(state, candidates[0]));
+}
+
 struct ImageSampler {
 	uint32_t id = 0;
 	bool dynamic = false;
@@ -821,9 +867,7 @@ std::array<ImageSampler, 2> LoadImageSamplers(ValueEmitContext& ctx, const IR::I
 				selected = EmitIndirectResourceIndex(state, selected,
 				    samplers[index].indirect_mapping_offset,
 				    samplers[index].indirect_search_iterations, 0u);
-			array_index = EmitIndexSwitch(
-			    state, selected, static_cast<uint32_t>(candidates.size()),
-			    TypeU32(state), [&](uint32_t ordinal) { return ConstantU32(state, candidates[ordinal]); });
+			array_index = SamplerArrayIndex(state, selected, candidates);
 		}
 		sampler = {LoadSamplerDescriptor(state, index, array_index), array_index != 0u};
 	};
@@ -870,14 +914,19 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 	}
 	if (op == IR::ValueOpcode::ImageQueryLod) {
 		state.builder.RequireCapability(spv::CapabilityImageQuery);
-		const auto dimension = image.dimension;
-		const auto sampled = MakeSampledImage(state, mem.resource,
-		                                     LoadSamplerDescriptor(state, mem.sampler));
-		const auto lod       = state.builder.AllocateId();
-		state.builder.AddFunction(
-		    spv::OpImageQueryLod, TypeF32Vector(state, 2), lod, sampled,
-		    CoordF32(ctx, mem, *address, 0, ImageDimensionInfoFor(dimension).spatial_components,
-		             image.cube));
+		const auto sampler = LoadSamplerDescriptor(state, mem.sampler);
+		const auto lod = EmitImageAccess(ctx, inst, TypeF32Vector(state, 2),
+		    [&](uint32_t resource, uint32_t array_index) {
+			    const auto& candidate = state.program.info.images[resource];
+			    const auto query = state.builder.AllocateId();
+			    state.builder.AddFunction(
+			        spv::OpImageQueryLod, TypeF32Vector(state, 2), query,
+			        MakeSampledImage(state, resource, sampler, 0u, array_index),
+			        CoordF32(ctx, mem, *address, 0,
+			                 ImageDimensionInfoFor(candidate.dimension).spatial_components,
+			                 candidate.cube));
+			    return query;
+		    });
 		uint32_t values[4] = {ConstantU32(state, 0), ConstantU32(state, 0), ConstantU32(state, 0),
 		                      ConstantU32(state, 0)};
 		for (uint32_t index = 0; index < 2u; index++) {
@@ -932,9 +981,12 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 				const auto mip_lod = candidate.mip_mode == IR::ImageMipMode::Dynamic
 				    ? LodU32(ctx, mem, *address, mem.image_dimension) : 0u;
 				const auto coord = CoordU32(ctx, mem, *address, candidate.dimension);
+				const auto& dimension_info = ImageDimensionInfoFor(candidate.dimension);
+				const auto sample = dimension_info.multisampled != 0u
+				    ? AddressU32(ctx, mem, *address, dimension_info.coordinate_components) : 0u;
 				const auto texel = StoreTexel(ctx, selected, ctx.Arg(inst, 2),
 				    candidate.numeric_class == Prospero::TextureNumericClass::Uint);
-				EmitStorageImageWrite(state, resource, mip_lod, coord, texel, array_index);
+				EmitStorageImageWrite(state, resource, mip_lod, coord, texel, array_index, sample);
 			});
 		});
 		return;
@@ -1037,7 +1089,12 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 				dref_value = AddressF32(ctx, mem, *address, layout.dref);
 			}
 		}
-		std::array<uint32_t, 4> operands {};
+		// _cl clamps the computed LOD from below.
+		uint32_t min_lod = 0;
+		if (layout.clamp != NoImageComponent) {
+			min_lod = AddressF32(ctx, mem, *address, layout.clamp);
+		}
+		std::array<uint32_t, 5> operands {};
 		size_t operand_count = 0;
 		if (HasFlag(mem, Decoder::ImageSampleFlagDerivative)) {
 			operands[0] = spv::ImageOperandsGradMask;
@@ -1049,6 +1106,11 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 			if (HasFlag(mem, Decoder::ImageSampleFlagLod) && layout.lod != NoImageComponent) {
 				lod = AddressF32(ctx, mem, *address, layout.lod);
 			}
+			// SPIR-V takes MinLod only with implicit LOD or Grad.
+			if (min_lod != 0) {
+				lod     = EmitGlsl<GLSLstd450FMax, IR::Type::F32>(state, lod, min_lod);
+				min_lod = 0;
+			}
 			operands[0] = spv::ImageOperandsLodMask;
 			operands[1] = lod;
 			operand_count = 2;
@@ -1072,6 +1134,13 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 				sample_operands[sample_operand_count++] = PackedOffset(
 				    ctx, offset_value, candidate.dimension, sample_operands[0]);
 			}
+			// MinLod follows the offset in operand order.
+			if (min_lod != 0) {
+				state.builder.RequireCapability(spv::CapabilityMinLod);
+				sample_operand_count = std::max(size_t{1}, sample_operand_count);
+				sample_operands[0] |= spv::ImageOperandsMinLodMask;
+				sample_operands[sample_operand_count++] = min_lod;
+			}
 			const auto coord =
 			    CoordF32(ctx, mem, *address, layout.coord,
 			             ImageDimensionInfoFor(candidate.dimension).coordinate_components,

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

@@ -56,6 +56,7 @@ EMIT_NATIVE(BitFieldSExtract, OpBitFieldSExtract, U32, uint32_t, uint32_t, uint3
 EMIT_NATIVE(SelectU1, OpSelect, U1, uint32_t, uint32_t, uint32_t)
 EMIT_NATIVE(SelectU32, OpSelect, U32, uint32_t, uint32_t, uint32_t)
 EMIT_NATIVE(SelectF32, OpSelect, F32, uint32_t, uint32_t, uint32_t)
+EMIT_NATIVE(SelectF64, OpSelect, F64, uint32_t, uint32_t, uint32_t)
 EMIT_NATIVE(IAdd32, OpIAdd, U32, uint32_t, uint32_t)
 EMIT_NATIVE(ISub32, OpISub, U32, uint32_t, uint32_t)
 EMIT_NATIVE(IMul32, OpIMul, U32, uint32_t, uint32_t)
@@ -229,16 +230,20 @@ uint32_t              EmitDppUpdateU32(ValueEmitContext& ctx, const IR::Inst& in
 uint32_t              EmitWqmU64(EmitterState& state, uint32_t value);
 uint32_t              EmitLaneId(EmitterState& state);
 uint32_t              EmitBallot(ValueEmitContext& ctx, IR::Value predicate);
+uint32_t              EmitLiveMask(EmitterState& state);
 uint32_t              EmitConditionRef(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitReadFirstLane(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitReadLane(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitSubgroupUMinU32(ValueEmitContext& ctx, const IR::Inst& inst);
+uint32_t              EmitSubgroupUMaxU32(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitWriteLane(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitPermlane16U32(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitGetAttribute(ValueEmitContext& ctx, const IR::Inst& inst);
+uint32_t              EmitGetAttributeF16(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitGetInterpolationParameter(ValueEmitContext& ctx, const IR::Inst& inst);
 void                  EmitSetAttribute(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitGetShaderBase(ValueEmitContext& ctx);
+uint32_t              EmitReadRealtimeClock(ValueEmitContext& ctx, const IR::Inst& inst);
 inline constexpr auto EmitGetSrtResource     = EmitVoid;
 inline constexpr auto EmitGetBufferResource  = EmitGetSrtResource;
 uint32_t              EmitGetBufferResourceIndex(ValueEmitContext& ctx, const IR::Inst& inst);
@@ -286,6 +291,7 @@ inline constexpr auto EmitBufferAtomicIAdd32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicISub32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicSMin32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicUMin32    = EmitAtomic32;
+inline constexpr auto EmitBufferAtomicUMin64    = EmitBufferAtomic64;
 inline constexpr auto EmitBufferAtomicSMax32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicUMax32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicAnd32     = EmitAtomic32;

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

@@ -112,6 +112,7 @@ struct EmitterState {
 	uint32_t                                         lane_half               = 0;
 	std::array<BufferDefinition, 4>                  storage_buffers {};
 	std::vector<uint32_t>                           memory_byte_offsets;
+	std::vector<uint32_t>                           memory_byte_sizes;
 	uint32_t                                         bda_pagetable_variable  = 0;
 	uint32_t                                         fault_buffer_variable   = 0;
 	uint32_t                                         bda_pointer_function    = 0;
@@ -130,6 +131,7 @@ struct EmitterState {
 	uint32_t                                         lds_u64_variable        = 0;
 	std::array<uint32_t, 2>                          scratch_variable {};
 	std::array<ImageDefinition, IR::ImageBindingCount> images {};
+	std::array<ImageDefinition, IR::ImageBindingCount> heap_images {};
 	uint32_t                   sampler_variable                      = 0;
 	uint32_t                   sampler_type                          = 0;
 	uint32_t                   sampler_pointer_type                  = 0;
@@ -144,6 +146,7 @@ struct EmitterState {
 	const IR::Block*           current_block                         = nullptr;
 	uint32_t                   pixel_valid_mask_variable             = 0;
 	uint32_t                   subgroup_local_invocation_id_variable = 0;
+	uint32_t                   helper_invocation_variable            = 0;
 	uint32_t                   per_vertex_variable                   = 0;
 	uint32_t                   point_size_variable                   = 0;
 	uint32_t                   clip_distance_variable                = 0;
@@ -281,6 +284,7 @@ struct ImageSampleLayout {
 	uint32_t bias   = NoImageComponent;
 	uint32_t coord  = 0;
 	uint32_t lod    = NoImageComponent;
+	uint32_t clamp  = NoImageComponent;
 	uint32_t grad_x = NoImageComponent;
 	uint32_t grad_y = NoImageComponent;
 };
@@ -304,6 +308,7 @@ uint32_t VertexParameterScalarType(EmitterState& state, VertexInputScalarKind ki
 
 
 uint32_t OutputVariableForExport(const EmitterState& state, const IR::ExportInfo& exp);
+uint32_t MrtOutputType(EmitterState& state, uint32_t index);
 
 uint32_t ConstantU32(EmitterState& state, uint32_t value);
 
@@ -339,8 +344,9 @@ uint32_t MakeSampledImage(EmitterState& state, uint32_t resource, uint32_t sampl
                           uint32_t mip = 0, uint32_t array_index = 0,
                           bool sampler_dynamic = false);
 
+// sample is 0 (no id) for single-sampled images.
 void EmitStorageImageWrite(EmitterState& state, uint32_t resource, uint32_t mip_lod, uint32_t coord,
-                           uint32_t texel, uint32_t array_index);
+                           uint32_t texel, uint32_t array_index, uint32_t sample);
 
 spv::ExecutionModel ExecutionModelForStage(ShaderType stage);
 
@@ -412,7 +418,7 @@ Prospero::BufferFormat StorageBufferFormat(const EmitterState& state, const IR::
 
 uint32_t SharedIndirectBufferResource(const EmitterState& state, uint32_t resource);
 
-void EmitMemoryOffsets(EmitterState& state);
+void EmitMemoryBindings(EmitterState& state);
 
 uint32_t LdsDwordCount(const EmitterState& state);
 void EnsureLdsStorage(EmitterState& state);
@@ -430,6 +436,8 @@ struct MemoryResourceAccess {
 
 uint32_t EmitIndirectResourceIndex(EmitterState& state, uint32_t key, uint32_t mapping_offset,
                                    uint32_t search_iterations, uint32_t default_resource);
+// Slot word of an image heap byte offset; offsets past the heap select the null element.
+uint32_t EmitImageHeapEntry(EmitterState& state, uint32_t key, uint32_t mapping_offset);
 MemoryResourceAccess PrepareMemoryResourceAccess(EmitterState& state, const IR::MemoryInfo& mem,
                                                  uint32_t array_index = 0u);
 
@@ -527,6 +535,7 @@ inline constexpr auto EmitFNegateValue = EmitNative<spv::OpFNegate, IR::Type::F3
 inline constexpr auto EmitFAbsValue = EmitGlsl<GLSLstd450FAbs, IR::Type::F32, uint32_t>;
 
 uint32_t EmitF16BitsToF32(EmitterState& state, uint32_t bits);
+uint32_t EmitF32BitsToF16RtzBits(EmitterState& state, uint32_t bits);
 
 void EmitProgram(EmitterState& state);
 

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

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

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

@@ -38,19 +38,39 @@ uint32_t EmitBinaryU32(EmitterState& state, spv::Op opcode, uint32_t lhs, uint32
 	return ret;
 }
 
+static uint32_t LoadFlattenedSrt(EmitterState& state, uint32_t index) {
+	const auto pointer = state.builder.AllocateId();
+	state.builder.AddFunction(spv::OpAccessChain, TypeStorageBufferElementPointer(state), pointer,
+	                          state.flattened_srt_variable, ConstantU32(state, 0), index);
+	const auto value = state.builder.AllocateId();
+	state.builder.AddFunction(spv::OpLoad, TypeU32(state), value, pointer);
+	return value;
+}
+
+uint32_t EmitImageHeapEntry(EmitterState& state, uint32_t key, uint32_t mapping_offset) {
+	const auto Mapping = [&](uint32_t word) {
+		return LoadFlattenedSrt(state, ConstantU32(state, mapping_offset + word));
+	};
+	// Keys past the table select its trailing null slot.
+	const auto index = EmitUMin32(state, Binary(state, spv::OpShiftRightLogical, TypeU32(state), key,
+	                                            ConstantU32(state, 2u)),
+	                              Mapping(0u));
+	const auto offset = Unary(state, spv::OpUConvert, TypeU64(state),
+	                          Binary(state, spv::OpShiftLeftLogical, TypeU32(state), index,
+	                                 ConstantU32(state, 2u)));
+	const auto pointer = Unary(state, spv::OpConvertUToPtr, TypePhysicalU32Pointer(state),
+	                           Binary(state, spv::OpIAdd, TypeU64(state),
+	                                  PackU64(state, Mapping(1u), Mapping(2u)), offset));
+	const auto entry = state.builder.AllocateId();
+	state.builder.AddFunction(spv::OpLoad, TypeU32(state), entry, pointer,
+	                          spv::MemoryAccessAlignedMask, 4u);
+	return entry;
+}
+
 uint32_t EmitIndirectResourceIndex(EmitterState& state, uint32_t key, uint32_t mapping_offset,
                                    uint32_t search_iterations, uint32_t default_resource) {
-	const auto LoadMapping = [&](uint32_t index) {
-		const auto pointer = state.builder.AllocateId();
-		state.builder.AddFunction(spv::OpAccessChain, TypeStorageBufferElementPointer(state),
-		                          pointer, state.flattened_srt_variable, ConstantU32(state, 0),
-		                          index);
-		const auto value = state.builder.AllocateId();
-		state.builder.AddFunction(spv::OpLoad, TypeU32(state), value, pointer);
-		return value;
-	};
 	const auto mapping = ConstantU32(state, mapping_offset);
-	const auto count = LoadMapping(mapping);
+	const auto count = LoadFlattenedSrt(state, mapping);
 	if (search_iterations == 0u) {
 		const auto index = Binary(state, spv::OpShiftRightLogical, TypeU32(state), key,
 		                          ConstantU32(state, 2u));
@@ -58,7 +78,7 @@ uint32_t EmitIndirectResourceIndex(EmitterState& state, uint32_t key, uint32_t m
 		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),
+		return Select(state, TypeU32(state), valid, LoadFlattenedSrt(state, entry),
 		              ConstantU32(state, default_resource));
 	}
 	const auto entry_at = [&](uint32_t index) {
@@ -76,7 +96,7 @@ uint32_t EmitIndirectResourceIndex(EmitterState& state, uint32_t key, uint32_t m
 		                        ConstantU32(state, 1u));
 		const auto probe = Select(state, TypeU32(state), active, mid, ConstantU32(state, 0u));
 		const auto less = Binary(state, spv::OpULessThan, TypeBool(state),
-		                         LoadMapping(entry_at(probe)), key);
+		                         LoadFlattenedSrt(state, entry_at(probe)), key);
 		low = Select(state, TypeU32(state),
 		             Binary(state, spv::OpLogicalAnd, TypeBool(state), active, less),
 		             Binary(state, spv::OpIAdd, TypeU32(state), mid, ConstantU32(state, 1u)), low);
@@ -87,9 +107,10 @@ uint32_t EmitIndirectResourceIndex(EmitterState& state, uint32_t key, uint32_t m
 	const auto in_range = Binary(state, spv::OpULessThan, TypeBool(state), low, count);
 	const auto entry = entry_at(Select(state, TypeU32(state), in_range, low, ConstantU32(state, 0u)));
 	const auto match = Binary(state, spv::OpLogicalAnd, TypeBool(state), in_range,
-	                          Binary(state, spv::OpIEqual, TypeBool(state), LoadMapping(entry), key));
-	const auto resource = LoadMapping(Binary(state, spv::OpIAdd, TypeU32(state), entry,
-	                                         ConstantU32(state, 1u)));
+	                          Binary(state, spv::OpIEqual, TypeBool(state),
+	                                 LoadFlattenedSrt(state, entry), key));
+	const auto resource = LoadFlattenedSrt(
+	    state, Binary(state, spv::OpIAdd, TypeU32(state), entry, ConstantU32(state, 1u)));
 	return Select(state, TypeU32(state), match, resource, ConstantU32(state, default_resource));
 }
 
@@ -133,9 +154,10 @@ uint32_t SharedIndirectBufferResource(const EmitterState& state, uint32_t resour
 	return representative;
 }
 
-void EmitMemoryOffsets(EmitterState& state) {
+void EmitMemoryBindings(EmitterState& state) {
 	const auto count = state.program.bindings.descriptor_counts[static_cast<size_t>(IR::DescriptorBindingKind::Buffers)];
 	state.memory_byte_offsets.resize(count);
+	state.memory_byte_sizes.resize(count);
 	std::vector<bool> dynamic_offsets(count);
 	for (uint32_t resource = 0; resource < state.program.info.buffers.size(); ++resource) {
 		if (SharedIndirectBufferResource(state, resource) == UINT32_MAX) continue;
@@ -153,6 +175,8 @@ void EmitMemoryOffsets(EmitterState& state) {
 			if (dynamic_offsets[i + lane]) continue;
 			state.memory_byte_offsets[i + lane] = EmitNative<spv::OpBitFieldUExtract, IR::Type::U32>(
 			    state, word, ConstantU32(state, lane * 8u), ConstantU32(state, 8u));
+			state.memory_byte_sizes[i + lane] =
+			    EmitShaderDataDwordLoad(state, state.program.bindings.MemorySizeDword() + i + lane);
 		}
 	}
 }
@@ -236,22 +260,28 @@ MemoryResourceAccess PrepareStorageBufferResourceAccess(EmitterState& state,
 		state.builder.RequireCapability(spv::CapabilityShaderNonUniform);
 		state.builder.RequireCapability(spv::CapabilityStorageBufferArrayNonUniformIndexing);
 		state.builder.AddAnnotation(spv::OpDecorate, access.object_pointer, spv::DecorationNonUniform);
-		const auto offsets_start = state.program.bindings.memory_offset_dword +
-		    (state.program.bindings.UsesPushData() ? state.program.bindings.push_data_start_dword : 0u);
+		const auto& bindings   = state.program.bindings;
+		const auto  data_start = bindings.UsesPushData() ? bindings.push_data_start_dword : 0u;
 		const auto word = EmitShaderDataDwordLoadAt(state, EmitAddU32(
-		    state, ConstantU32(state, offsets_start),
+		    state, ConstantU32(state, bindings.memory_offset_dword + data_start),
 		    Binary(state, spv::OpShiftRightLogical, TypeU32(state), array_index, ConstantU32(state, 2))));
 		const auto shift = Binary(state, spv::OpShiftLeftLogical, TypeU32(state),
 		    Binary(state, spv::OpBitwiseAnd, TypeU32(state), array_index, ConstantU32(state, 3)),
 		    ConstantU32(state, 3));
 		access.byte_offset = EmitNative<spv::OpBitFieldUExtract, IR::Type::U32>(
 		    state, word, shift, ConstantU32(state, 8));
+		access.length = EmitShaderDataDwordLoadAt(state, EmitAddU32(
+		    state, ConstantU32(state, bindings.MemorySizeDword() + data_start), array_index));
 	} else {
 		access.byte_offset = state.memory_byte_offsets[static_index];
+		access.length      = state.memory_byte_sizes[static_index];
+	}
+	// storage_buffers[i] views the binding as 1 << i byte elements.
+	const auto shift = static_cast<uint32_t>(&buffer - state.storage_buffers.data());
+	if (shift != 0u) {
+		access.length = Binary(state, spv::OpShiftRightLogical, TypeU32(state), access.length,
+		                       ConstantU32(state, shift));
 	}
-	access.length      = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpArrayLength, TypeU32(state), access.length,
-	                          access.object_pointer, 0);
 	return access;
 }
 
@@ -451,6 +481,17 @@ uint32_t EncodeFormattedStoreComponent(EmitterState& state, const Format::Buffer
                                        UnormRounding rounding) {
 	EXIT_IF(rounding != UnormRounding::RoundEven && info.type != Format::ComponentType::Unorm);
 	const auto bits = info.component_bits[component];
+	if (info.type == Format::ComponentType::Float && bits < 16u) {
+		// Truncate to f16, whose bias the unsigned 11/10-bit floats share, and drop mantissa
+		// bits; negatives and NaN, whose f16 bits lie above +inf, store zero.
+		const auto half = data_bits == 16u ? EmitAndConstant(state, data, 0xffffu)
+		                                   : EmitF32BitsToF16RtzBits(state, data);
+		return Select(state, TypeU32(state),
+		              EmitCompareU32Constant(state, spv::OpUGreaterThan, half, 0x7c00u),
+		              ConstantU32(state, 0),
+		              EmitBinaryU32(state, spv::OpShiftRightLogical, half,
+		                            ConstantU32(state, 15u - bits)));
+	}
 	if (data_bits == 16u) {
 		if (info.type == Format::ComponentType::Uint) {
 			data = EmitBitFieldUExtract(state, data, ConstantU32(state, 0), ConstantU32(state, 16));
@@ -544,6 +585,7 @@ spv::Op SpirvAtomicOpcode(IR::ValueOpcode opcode) {
 		case IR::ValueOpcode::ImageAtomicUMin32:
 		case IR::ValueOpcode::ImageAtomicUMin64:
 		case IR::ValueOpcode::BufferAtomicUMin32:
+		case IR::ValueOpcode::BufferAtomicUMin64:
 		case IR::ValueOpcode::SharedAtomicUMin32: return spv::OpAtomicUMin;
 		case IR::ValueOpcode::ImageAtomicSMax32:
 		case IR::ValueOpcode::BufferAtomicSMax32:

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

@@ -187,7 +187,7 @@ void EmitMeshEntryPoint(EmitterState& state) {
 				vertex[component] = state.builder.AllocateId();
 				state.builder.AddFunction(
 				    spv::OpBitFieldUExtract, TypeU32(state), vertex[component], packed,
-				    ConstantU32(state, component * 10u), ConstantU32(state, 10));
+				    ConstantU32(state, component * 10u), ConstantU32(state, 9));
 			}
 			const auto triangle = state.builder.AllocateId();
 			state.builder.AddFunction(spv::OpCompositeConstruct, TypeU32Vector(state, 3), triangle,

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

@@ -206,6 +206,17 @@ uint32_t F32ArrayType(EmitterState& state, uint32_t count) {
 	return state.builder.Type(spv::OpTypeArray, TypeF32(state), ConstantU32(state, count));
 }
 
+void DefineImageArray(EmitterState& state, ImageDefinition& definition, const IR::ImageResource& image,
+                      uint32_t count, uint32_t set, uint32_t binding) {
+	definition.type     = ImageType(state, image);
+	const auto array    = state.builder.Type(spv::OpTypeArray, definition.type, ConstantU32(state, count));
+	definition.variable = state.builder.DefineGlobalVariable(
+	    TypePointer(state, spv::StorageClassUniformConstant, array), spv::StorageClassUniformConstant);
+	state.builder.AddAnnotation(spv::OpDecorate, definition.variable, spv::DecorationDescriptorSet, set);
+	state.builder.AddAnnotation(spv::OpDecorate, definition.variable, spv::DecorationBinding, binding);
+	definition.pointer_type = TypePointer(state, spv::StorageClassUniformConstant, definition.type);
+}
+
 void DefineDescriptors(EmitterState& state) {
 	using Kind = IR::DescriptorBindingKind;
 	auto& info = state.program.info;
@@ -222,6 +233,14 @@ void DefineDescriptors(EmitterState& state) {
 		    (image.mip_mode != IR::ImageMipMode::Dynamic && image.mip_count != 1u)) {
 			EXIT("shader image has an invalid descriptor type or mip count");
 		}
+		if (image.heap) {
+			// Heap kinds alias one array; the root itself is never accessed.
+			auto& definition = state.heap_images[IR::ImageBindingIndex(*kind)];
+			if (image.indirect_resources.empty() && definition.variable == 0) {
+				DefineImageArray(state, definition, image, IR::ImageHeapCapacity, IR::ImageHeapSet, 0u);
+			}
+			continue;
+		}
 		auto& count = counts[static_cast<size_t>(*kind)];
 		image.descriptor_index = count;
 		if (count == 0) image_types[IR::ImageBindingIndex(*kind)] = &image;
@@ -331,21 +350,11 @@ void DefineDescriptors(EmitterState& state) {
 			default: {
 				EXIT_IF(IR::ImageBindingResourceClass(kind) ==
 				        IR::ImageResourceClass::None);
-				const auto& image = *image_types[IR::ImageBindingIndex(kind)];
-				const auto  name  = "image_" + std::to_string(static_cast<uint32_t>(kind));
 				auto& definition = state.images[IR::ImageBindingIndex(kind)];
-				definition.type = ImageType(state, image);
-				definition.variable = Define(ArrayType(definition.type), name.c_str(),
-				                             spv::StorageClassUniformConstant);
-				definition.pointer_type = TypePointer(
-				    state, spv::StorageClassUniformConstant, definition.type);
-				if (image.dimension == ImageDimension::Dim1D ||
-				    image.dimension == ImageDimension::Dim1DArray) {
-					state.builder.RequireCapability(image.resource_class ==
-					                                        IR::ImageResourceClass::Sampled
-					                                    ? spv::CapabilitySampled1D
-					                                    : spv::CapabilityImage1D);
-				}
+				DefineImageArray(state, definition, *image_types[IR::ImageBindingIndex(kind)],
+				                 counts[index], 0u, IR::NativeBinding(state.program.stage, kind));
+				state.builder.AddName(definition.variable,
+				                      ("image_" + std::to_string(static_cast<uint32_t>(kind))).c_str());
 				break;
 			}
 		}
@@ -467,6 +476,8 @@ uint32_t BuiltInForInput(IR::StageInputKind kind) {
 		case IR::StageInputKind::LocalInvocationId: return spv::BuiltInLocalInvocationId;
 		case IR::StageInputKind::LocalInvocationIndex: return spv::BuiltInLocalInvocationIndex;
 		case IR::StageInputKind::GlobalInvocationId: return spv::BuiltInGlobalInvocationId;
+		case IR::StageInputKind::BaseVertex: return spv::BuiltInBaseVertex;
+		case IR::StageInputKind::BaseInstance: return spv::BuiltInBaseInstance;
 		default: return UINT32_MAX;
 	}
 }
@@ -517,7 +528,9 @@ void DefineInputs(EmitterState& state) {
 			case IR::StageInputKind::PrimitiveId:
 			case IR::StageInputKind::InstanceIndex:
 			case IR::StageInputKind::Layer:
-			case IR::StageInputKind::SampleId: type = TypeI32(state); break;
+			case IR::StageInputKind::SampleId:
+			case IR::StageInputKind::BaseVertex:
+			case IR::StageInputKind::BaseInstance: type = TypeI32(state); break;
 			case IR::StageInputKind::WorkgroupId:
 			case IR::StageInputKind::NumWorkgroups:
 			case IR::StageInputKind::LocalInvocationId:
@@ -586,6 +599,13 @@ void DefineInputs(EmitterState& state) {
 			state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationFlat);
 		}
 	}
+	if (state.program.info.helper_invocation) {
+		const auto variable = DefineInterfaceVariable(state, TypeBool(state), spv::StorageClassInput,
+		                                              "gl_HelperInvocation");
+		state.helper_invocation_variable = variable;
+		state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationBuiltIn,
+		                            spv::BuiltInHelperInvocation);
+	}
 }
 
 void DefineOutputs(EmitterState& state) {
@@ -682,13 +702,15 @@ void DefineOutputs(EmitterState& state) {
 				break;
 			case IR::StageOutputKind::Parameter:
 			case IR::StageOutputKind::Mrt: {
-				const bool uint_output =
-				    binding.kind == IR::StageOutputKind::Mrt &&
-				    state.program.stage == ShaderType::Pixel &&
-				    binding.index < std::size(state.input_info.pixel->target_output_mode) &&
-				    (state.input_info.pixel->target_output_mode[binding.index] == 7u ||
-				     (state.input_info.pixel->target_packed_unorm_mask & (1u << binding.index)) != 0u);
-				const auto type = uint_output ? TypeU32Vector(state, 4) : TypeF32Vector(state, 4);
+				auto type = TypeF32Vector(state, 4);
+				if (binding.kind == IR::StageOutputKind::Mrt &&
+				    state.program.stage == ShaderType::Pixel && binding.index < 8u) {
+					// Packed UNorm stores its encoded bits.
+					const auto bit = 1u << binding.index;
+					type = (state.input_info.pixel->target_packed_unorm_mask & bit) != 0u
+					           ? TypeU32Vector(state, 4)
+					           : MrtOutputType(state, binding.index);
+				}
 				binding.variable_id = DefineInterfaceVariable(state, type, spv::StorageClassOutput,
 				                                              binding.debug_name.c_str());
 				const bool dual_source = binding.kind == IR::StageOutputKind::Mrt &&
@@ -752,7 +774,7 @@ void DefineModule(EmitterState& state) {
 	if (state.program.info.buffer_u16) {
 		state.builder.RequireCapability(spv::CapabilityStorageBuffer16BitAccess);
 	}
-	if (state.program.info.uses_dma) {
+	if (state.program.info.uses_dma || state.program.info.uses_image_heap) {
 		state.builder.RequireCapability(spv::CapabilityInt64);
 		state.builder.RequireCapability(spv::CapabilityPhysicalStorageBufferAddresses);
 		state.builder.RequireExtension("SPV_KHR_physical_storage_buffer");
@@ -784,6 +806,10 @@ void DefineModule(EmitterState& state) {
 	if (InputVariableForKind(state, IR::StageInputKind::SampleId) != 0) {
 		state.builder.RequireCapability(spv::CapabilitySampleRateShading);
 	}
+	if (InputVariableForKind(state, IR::StageInputKind::BaseVertex) != 0 ||
+	    InputVariableForKind(state, IR::StageInputKind::BaseInstance) != 0) {
+		state.builder.RequireCapability(spv::CapabilityDrawParameters);
+	}
 	if (state.program.info.image_gather_extended) {
 		state.builder.RequireCapability(spv::CapabilityImageGatherExtended);
 	}
@@ -820,7 +846,7 @@ void DefineModule(EmitterState& state) {
 		state.builder.RequireExtension("SPV_KHR_fragment_shader_barycentric");
 	}
 	state.builder.RequireExtension("SPV_KHR_float_controls");
-	state.builder.AddMemoryModel(state.program.info.uses_dma
+	state.builder.AddMemoryModel(state.program.info.uses_dma || state.program.info.uses_image_heap
 	                                 ? spv::AddressingModelPhysicalStorageBuffer64
 	                                 : spv::AddressingModelLogical,
 	                             spv::MemoryModelGLSL450);

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

@@ -235,9 +235,9 @@ void EmitDirectInstruction(ValueEmitContext& ctx, const IR::Inst& inst) {
 	const auto op = inst.GetOpcode();
 	if (ctx.half != 0 &&
 	    (op == Op::Ballot || op == Op::ReadFirstLane || op == Op::SubgroupUMinU32 ||
-	     op == Op::ReadConstBuffer ||
+	     op == Op::SubgroupUMaxU32 || op == Op::ReadConstBuffer || op == Op::ReadRealtimeClock ||
 	     (op == Op::LoadAddressU32 && ctx.Memory(inst).kind == IR::ResourceKind::ScalarAddress))) {
-		// Whole-wave results and SMEM loads (SGPR operands, no EXEC) are identical in both halves.
+		// Whole-wave results, clocks and SMEM loads (SGPR operands, no EXEC) match in both halves.
 		ctx.Define(inst, ctx.other_half->Result(inst));
 		return;
 	}
@@ -746,7 +746,7 @@ void EmitProgram(EmitterState& state) {
 		state.lds_base_dwords = EmitBinaryU32(state, spv::OpIMul, index,
 		                                     ConstantU32(state, LdsDwordCount(state)));
 	}
-	EmitMemoryOffsets(state);
+	EmitMemoryBindings(state);
 	if (program.blocks.empty()) {
 		EmitReturn(ctx);
 	} else if (state.program.dispatcher_fallback) {

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

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

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

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

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

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

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

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

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

@@ -6,13 +6,14 @@ namespace Libs::Graphics::ShaderRecompiler::Decoder {
 namespace {
 
 struct MemoryOpcodeInfo {
-	uint32_t encoding    = 0;
-	Opcode   decoded     = Opcode::UNSUPPORTED;
-	uint32_t data_dwords = 1;
-	uint32_t data_bits   = 32;
-	bool     data_signed = false;
-	bool     typed       = false;
-	bool     formatted   = false;
+	uint32_t encoding        = 0;
+	Opcode   decoded         = Opcode::UNSUPPORTED;
+	uint32_t data_dwords     = 1;
+	uint32_t data_bits       = 32;
+	bool     data_signed     = false;
+	bool     typed           = false;
+	bool     formatted       = false;
+	uint32_t data_components = 0; // D16 formats pack two components per dword
 };
 
 constexpr MemoryOpcodeInfo SMEM_OPCODE_LIST[] = {
@@ -21,7 +22,7 @@ constexpr MemoryOpcodeInfo SMEM_OPCODE_LIST[] = {
     {0x04u, Opcode::S_LOAD_DWORDX16, 16, 32},      {0x08u, Opcode::S_BUFFER_LOAD_DWORD, 1, 32},
     {0x09u, Opcode::S_BUFFER_LOAD_DWORDX2, 2, 32}, {0x0au, Opcode::S_BUFFER_LOAD_DWORDX4, 4, 32},
     {0x0bu, Opcode::S_BUFFER_LOAD_DWORDX8, 8, 32}, {0x0cu, Opcode::S_BUFFER_LOAD_DWORDX16, 16, 32},
-    {0x25u, Opcode::S_MEMREALTIME, 2, 32},
+    {0x20u, Opcode::S_DCACHE_INV, 0, 0},           {0x25u, Opcode::S_MEMREALTIME, 2, 32},
 };
 
 constexpr MemoryOpcodeInfo MUBUF_OPCODE_LIST[] = {
@@ -68,14 +69,19 @@ constexpr MemoryOpcodeInfo MUBUF_OPCODE_LIST[] = {
     {0x40u, Opcode::BUFFER_ATOMIC_FMAX, 1, 32},
     {0x50u, Opcode::BUFFER_ATOMIC_SWAP_X2, 2, 32},
     {0x52u, Opcode::BUFFER_ATOMIC_ADD_X2, 2, 32},
+    {0x56u, Opcode::BUFFER_ATOMIC_UMIN_X2, 2, 32},
     {0x59u, Opcode::BUFFER_ATOMIC_AND_X2, 2, 32},
     {0x5au, Opcode::BUFFER_ATOMIC_OR_X2, 2, 32},
     {0x71u, Opcode::BUFFER_GL0_INV, 0, 0},
     {0x72u, Opcode::BUFFER_GL1_INV, 0, 0},
     {0x80u, Opcode::BUFFER_LOAD_FORMAT_D16_X, 1, 16, false, false, true},
-    {0x83u, Opcode::BUFFER_LOAD_FORMAT_D16_XYZW, 2, 16, false, false, true},
+    {0x81u, Opcode::BUFFER_LOAD_FORMAT_D16_XY, 1, 16, false, false, true, 2},
+    {0x82u, Opcode::BUFFER_LOAD_FORMAT_D16_XYZ, 2, 16, false, false, true, 3},
+    {0x83u, Opcode::BUFFER_LOAD_FORMAT_D16_XYZW, 2, 16, false, false, true, 4},
     {0x84u, Opcode::BUFFER_STORE_FORMAT_D16_X, 1, 16, false, false, true},
-    {0x87u, Opcode::BUFFER_STORE_FORMAT_D16_XYZW, 2, 16, false, false, true},
+    {0x85u, Opcode::BUFFER_STORE_FORMAT_D16_XY, 1, 16, false, false, true, 2},
+    {0x86u, Opcode::BUFFER_STORE_FORMAT_D16_XYZ, 2, 16, false, false, true, 3},
+    {0x87u, Opcode::BUFFER_STORE_FORMAT_D16_XYZW, 2, 16, false, false, true, 4},
 };
 
 constexpr MemoryOpcodeInfo MTBUF_OPCODE_LIST[] = {
@@ -90,14 +96,14 @@ constexpr MemoryOpcodeInfo MTBUF_OPCODE_LIST[] = {
 };
 
 constexpr MemoryOpcodeInfo FLAT_OPCODE_LIST[] = {
-    {0x08u, Opcode::FLAT_LOAD_UBYTE, 1, 8},     {0x09u, Opcode::FLAT_LOAD_SBYTE, 1, 8, true},
-    {0x0au, Opcode::FLAT_LOAD_USHORT, 1, 16},   {0x0bu, Opcode::FLAT_LOAD_SSHORT, 1, 16, true},
-    {0x0cu, Opcode::FLAT_LOAD_DWORD, 1, 32},    {0x0du, Opcode::FLAT_LOAD_DWORDX2, 2, 32},
-    {0x0eu, Opcode::FLAT_LOAD_DWORDX4, 4, 32},  {0x0fu, Opcode::FLAT_LOAD_DWORDX3, 3, 32},
-    {0x18u, Opcode::FLAT_STORE_BYTE, 1, 8},     {0x1au, Opcode::FLAT_STORE_SHORT, 1, 16},
-    {0x1cu, Opcode::FLAT_STORE_DWORD, 1, 32},   {0x1du, Opcode::FLAT_STORE_DWORDX2, 2, 32},
-    {0x1eu, Opcode::FLAT_STORE_DWORDX4, 4, 32}, {0x1fu, Opcode::FLAT_STORE_DWORDX3, 3, 32},
-    {0x24u, Opcode::FLAT_LOAD_SHORT_D16, 1, 16},
+    {0x08u, Opcode::FLAT_LOAD_UBYTE, 1, 8},      {0x09u, Opcode::FLAT_LOAD_SBYTE, 1, 8, true},
+    {0x0au, Opcode::FLAT_LOAD_USHORT, 1, 16},    {0x0bu, Opcode::FLAT_LOAD_SSHORT, 1, 16, true},
+    {0x0cu, Opcode::FLAT_LOAD_DWORD, 1, 32},     {0x0du, Opcode::FLAT_LOAD_DWORDX2, 2, 32},
+    {0x0eu, Opcode::FLAT_LOAD_DWORDX4, 4, 32},   {0x0fu, Opcode::FLAT_LOAD_DWORDX3, 3, 32},
+    {0x18u, Opcode::FLAT_STORE_BYTE, 1, 8},      {0x1au, Opcode::FLAT_STORE_SHORT, 1, 16},
+    {0x1cu, Opcode::FLAT_STORE_DWORD, 1, 32},    {0x1du, Opcode::FLAT_STORE_DWORDX2, 2, 32},
+    {0x1eu, Opcode::FLAT_STORE_DWORDX4, 4, 32},  {0x1fu, Opcode::FLAT_STORE_DWORDX3, 3, 32},
+    {0x24u, Opcode::FLAT_LOAD_SHORT_D16, 1, 16}, {0x25u, Opcode::FLAT_LOAD_SHORT_D16_HI, 1, 16},
 };
 
 constexpr MemoryOpcodeInfo DS_OPCODE_LIST[] = {
@@ -158,12 +164,13 @@ void ApplyMemoryInfo(Instruction& inst, const MemoryOpcodeInfo* info) {
 		inst.opcode = Opcode::UNSUPPORTED;
 		return;
 	}
-	inst.opcode      = info->decoded;
-	inst.data_dwords = info->data_dwords;
-	inst.data_bits   = info->data_bits;
-	inst.data_signed = info->data_signed;
-	inst.typed       = info->typed;
-	inst.formatted   = info->formatted;
+	inst.opcode          = info->decoded;
+	inst.data_dwords     = info->data_dwords;
+	inst.data_bits       = info->data_bits;
+	inst.data_signed     = info->data_signed;
+	inst.typed           = info->typed;
+	inst.formatted       = info->formatted;
+	inst.data_components = info->data_components;
 }
 
 } // namespace
@@ -279,6 +286,9 @@ void DecodeSmem(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	if (inst.opcode == Opcode::UNSUPPORTED) {
 		SetUnsupported(inst, Family::SMEM, opcode, "SMEM opcode is not implemented");
 	}
+	if (inst.opcode == Opcode::S_DCACHE_INV) {
+		return;
+	}
 
 	DecodeScalarDestination(sdst, pc, inst.dst);
 	if (inst.opcode == Opcode::S_MEMREALTIME) {
@@ -323,12 +333,7 @@ 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_XYZW ||
-	    inst.opcode == Opcode::BUFFER_STORE_FORMAT_D16_XYZW) {
-		inst.data_components = 4u;
-	}
-	if (inst.opcode == Opcode::BUFFER_LOAD_FORMAT_D16_X ||
-	    inst.opcode == Opcode::BUFFER_LOAD_UBYTE_D16 ||
+	if (inst.opcode == Opcode::BUFFER_LOAD_UBYTE_D16 ||
 	    inst.opcode == Opcode::BUFFER_LOAD_SHORT_D16 ||
 	    inst.opcode == Opcode::BUFFER_LOAD_SHORT_D16_HI) {
 		inst.dst.sdwa_sel = inst.opcode == Opcode::BUFFER_LOAD_SHORT_D16_HI ? 5u : 4u;
@@ -418,9 +423,10 @@ void DecodeFlat(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	}
 
 	DecodeVectorGpr(IsFlatStoreOpcode(inst.opcode) ? data : vdst, inst.dst);
-	if (inst.opcode == Opcode::FLAT_LOAD_SHORT_D16) {
-		// D16 loads reuse partial destinations to preserve the untouched high half.
-		inst.dst.sdwa_sel = 4u;
+	if (inst.opcode == Opcode::FLAT_LOAD_SHORT_D16 ||
+	    inst.opcode == Opcode::FLAT_LOAD_SHORT_D16_HI) {
+		// D16 loads reuse partial destinations to preserve the untouched half.
+		inst.dst.sdwa_sel = inst.opcode == Opcode::FLAT_LOAD_SHORT_D16_HI ? 5u : 4u;
 	}
 	DecodeVectorGpr(addr, inst.src0);
 	inst.src_count = 1;

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

@@ -57,6 +57,8 @@ constexpr OpcodeMap SOP1_OPCODE_LIST[] = {
     {0x14u, Opcode::S_FF1_I32_B64},
     {0x15u, Opcode::S_FLBIT_I32_B32},
     {0x16u, Opcode::S_FLBIT_I32_B64},
+    {0x17u, Opcode::S_FLBIT_I32},
+    {0x18u, Opcode::S_FLBIT_I32_I64},
     {0x19u, Opcode::S_SEXT_I32_I8},
     {0x1au, Opcode::S_SEXT_I32_I16},
     {0x1bu, Opcode::S_BITSET0_B32},

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

@@ -123,10 +123,9 @@ std::string ImageSampleFlagsToString(uint32_t flags) {
 }
 
 std::string FormatMimg(const Instruction& inst) {
-	const char* sample_name =
-	    inst.opcode == Opcode::IMAGE_SAMPLE ? MimgSampleOpcodeName(inst.opcode_id) : nullptr;
+	const char*            sampled_name = MimgSampledOpcodeName(inst.opcode_id);
 	const std::string_view name =
-	    sample_name != nullptr ? sample_name : magic_enum::enum_name(inst.opcode);
+	    sampled_name != nullptr ? sampled_name : magic_enum::enum_name(inst.opcode);
 	std::string text =
 	    fmt::format("0x{:08x}: {} {}, {}, {}, {} ; dmask=0x{:x} image_dim={}", inst.pc, name,
 	                OperandToString(inst.dst).c_str(), OperandToString(inst.src0).c_str(),
@@ -138,22 +137,10 @@ std::string FormatMimg(const Instruction& inst) {
 	if (inst.image_r128) {
 		text += " r128=1";
 	}
-	switch (inst.opcode) {
-		case Opcode::IMAGE_SAMPLE:
-		case Opcode::IMAGE_GATHER4_L:
-		case Opcode::IMAGE_GATHER4_LZ:
-		case Opcode::IMAGE_GATHER4_C:
-		case Opcode::IMAGE_GATHER4_C_L:
-		case Opcode::IMAGE_GATHER4_C_LZ:
-		case Opcode::IMAGE_GATHER4_LZ_O:
-		case Opcode::IMAGE_GATHER4_C_O:
-		case Opcode::IMAGE_GATHER4_C_LZ_O:
-		case Opcode::IMAGE_GATHER4H:
-			text += fmt::format(" sample_flags={} addr_components={}",
-			                    ImageSampleFlagsToString(inst.image_sample_flags).c_str(),
-			                    inst.image_address_components);
-			break;
-		default: break;
+	if (sampled_name != nullptr) {
+		text += fmt::format(" sample_flags={} addr_components={}",
+		                    ImageSampleFlagsToString(inst.image_sample_flags).c_str(),
+		                    inst.image_address_components);
 	}
 	if (inst.image_nsa_dwords != 0) {
 		text += fmt::format(" nsa_dwords={} nsa_addr=", inst.image_nsa_dwords);
@@ -258,7 +245,9 @@ void DecodeScalarSource(uint32_t code, uint32_t pc, Operand& operand) {
 		case 126u: operand.kind = OperandKind::ExecLo; return;
 		case 127u: operand.kind = OperandKind::ExecHi; return;
 		case 235u: operand.kind = OperandKind::SharedBase; return;
+		case 236u: operand.kind = OperandKind::SharedLimit; return;
 		case 237u: operand.kind = OperandKind::PrivateBase; return;
+		case 238u: operand.kind = OperandKind::PrivateLimit; return;
 		case 239u: operand.kind = OperandKind::PopsExitingWaveId; return;
 		case 248u:
 			operand.kind      = OperandKind::FloatInlineConstant;
@@ -459,7 +448,9 @@ std::string OperandToString(const Operand& operand) {
 		case OperandKind::M0: text = "m0"; break;
 		case OperandKind::PopsExitingWaveId: text = "pops_exiting_wave_id"; break;
 		case OperandKind::SharedBase: text = "shared_base"; break;
+		case OperandKind::SharedLimit: text = "shared_limit"; break;
 		case OperandKind::PrivateBase: text = "private_base"; break;
+		case OperandKind::PrivateLimit: text = "private_limit"; break;
 		case OperandKind::Null: text = "null"; break;
 		default: text = "unknown"; break;
 	}
@@ -566,6 +557,7 @@ std::string InstructionToString(const Instruction& inst) {
 			                                               OperandToString(inst.src0).c_str()));
 		case Opcode::S_BARRIER:
 			return WithUnsupportedReason(inst, fmt::format("0x{:08x}: s_barrier", inst.pc));
+		case Opcode::S_DCACHE_INV:
 		case Opcode::BUFFER_GL0_INV:
 		case Opcode::BUFFER_GL1_INV:
 			return fmt::format("0x{:08x}: {}", inst.pc, magic_enum::enum_name(inst.opcode));
@@ -612,15 +604,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::IMAGE_LOAD_MIP:
 		case Opcode::IMAGE_GET_RESINFO:
 		case Opcode::IMAGE_GET_LOD:
-		case Opcode::IMAGE_GATHER4_L:
-		case Opcode::IMAGE_GATHER4_LZ:
-		case Opcode::IMAGE_GATHER4_C:
-		case Opcode::IMAGE_GATHER4_C_L:
-		case Opcode::IMAGE_GATHER4_C_LZ:
-		case Opcode::IMAGE_GATHER4_LZ_O:
-		case Opcode::IMAGE_GATHER4_C_O:
-		case Opcode::IMAGE_GATHER4_C_LZ_O:
-		case Opcode::IMAGE_GATHER4H: return WithUnsupportedReason(inst, FormatMimg(inst));
+		case Opcode::IMAGE_GATHER4: return WithUnsupportedReason(inst, FormatMimg(inst));
 		case Opcode::S_LOAD_DWORD:
 		case Opcode::S_LOAD_DWORDX2:
 		case Opcode::S_LOAD_DWORDX4:
@@ -636,6 +620,8 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::BUFFER_LOAD_FORMAT_XYZ:
 		case Opcode::BUFFER_LOAD_FORMAT_XYZW:
 		case Opcode::BUFFER_LOAD_FORMAT_D16_X:
+		case Opcode::BUFFER_LOAD_FORMAT_D16_XY:
+		case Opcode::BUFFER_LOAD_FORMAT_D16_XYZ:
 		case Opcode::BUFFER_LOAD_FORMAT_D16_XYZW:
 		case Opcode::BUFFER_STORE_FORMAT_X:
 		case Opcode::BUFFER_STORE_FORMAT_XY:
@@ -643,6 +629,8 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::BUFFER_STORE_FORMAT_XYZW:
 		case Opcode::BUFFER_STORE_FORMAT_D16_X:
 		case Opcode::BUFFER_STORE_FORMAT_D16_HI_X:
+		case Opcode::BUFFER_STORE_FORMAT_D16_XY:
+		case Opcode::BUFFER_STORE_FORMAT_D16_XYZ:
 		case Opcode::BUFFER_STORE_FORMAT_D16_XYZW:
 		case Opcode::BUFFER_LOAD_UBYTE:
 		case Opcode::BUFFER_LOAD_UBYTE_D16:
@@ -677,6 +665,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::BUFFER_ATOMIC_SUB:
 		case Opcode::BUFFER_ATOMIC_SMIN:
 		case Opcode::BUFFER_ATOMIC_UMIN:
+		case Opcode::BUFFER_ATOMIC_UMIN_X2:
 		case Opcode::BUFFER_ATOMIC_SMAX:
 		case Opcode::BUFFER_ATOMIC_UMAX:
 		case Opcode::BUFFER_ATOMIC_AND:
@@ -693,6 +682,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::FLAT_LOAD_USHORT:
 		case Opcode::FLAT_LOAD_SSHORT:
 		case Opcode::FLAT_LOAD_SHORT_D16:
+		case Opcode::FLAT_LOAD_SHORT_D16_HI:
 		case Opcode::FLAT_LOAD_DWORD:
 		case Opcode::FLAT_LOAD_DWORDX2:
 		case Opcode::FLAT_LOAD_DWORDX3:

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

@@ -60,6 +60,8 @@ enum class Opcode {
 	S_FF1_I32_B64,
 	S_FLBIT_I32_B32,
 	S_FLBIT_I32_B64,
+	S_FLBIT_I32,
+	S_FLBIT_I32_I64,
 	S_BITREPLICATE_B64_B32,
 	S_QUADMASK_B64,
 	S_GETPC_B64,
@@ -542,11 +544,14 @@ enum class Opcode {
 	S_BUFFER_LOAD_DWORDX8,
 	S_BUFFER_LOAD_DWORDX16,
 	S_MEMREALTIME,
+	S_DCACHE_INV,
 	BUFFER_LOAD_FORMAT_X,
 	BUFFER_LOAD_FORMAT_XY,
 	BUFFER_LOAD_FORMAT_XYZ,
 	BUFFER_LOAD_FORMAT_XYZW,
 	BUFFER_LOAD_FORMAT_D16_X,
+	BUFFER_LOAD_FORMAT_D16_XY,
+	BUFFER_LOAD_FORMAT_D16_XYZ,
 	BUFFER_LOAD_FORMAT_D16_XYZW,
 	BUFFER_STORE_FORMAT_X,
 	BUFFER_STORE_FORMAT_XY,
@@ -554,6 +559,8 @@ enum class Opcode {
 	BUFFER_STORE_FORMAT_XYZW,
 	BUFFER_STORE_FORMAT_D16_X,
 	BUFFER_STORE_FORMAT_D16_HI_X,
+	BUFFER_STORE_FORMAT_D16_XY,
+	BUFFER_STORE_FORMAT_D16_XYZ,
 	BUFFER_STORE_FORMAT_D16_XYZW,
 	BUFFER_LOAD_UBYTE,
 	BUFFER_LOAD_UBYTE_D16,
@@ -590,6 +597,7 @@ enum class Opcode {
 	BUFFER_ATOMIC_SUB,
 	BUFFER_ATOMIC_SMIN,
 	BUFFER_ATOMIC_UMIN,
+	BUFFER_ATOMIC_UMIN_X2,
 	BUFFER_ATOMIC_SMAX,
 	BUFFER_ATOMIC_UMAX,
 	BUFFER_ATOMIC_AND,
@@ -606,6 +614,7 @@ enum class Opcode {
 	FLAT_LOAD_USHORT,
 	FLAT_LOAD_SSHORT,
 	FLAT_LOAD_SHORT_D16,
+	FLAT_LOAD_SHORT_D16_HI,
 	FLAT_LOAD_DWORD,
 	FLAT_LOAD_DWORDX2,
 	FLAT_LOAD_DWORDX3,
@@ -701,18 +710,12 @@ enum class Opcode {
 	IMAGE_ATOMIC_FMIN,
 	IMAGE_ATOMIC_FMAX,
 	IMAGE_SAMPLE,
-	IMAGE_GATHER4_L,
-	IMAGE_GATHER4_LZ,
-	IMAGE_GATHER4_C,
-	IMAGE_GATHER4_C_L,
-	IMAGE_GATHER4_C_LZ,
-	IMAGE_GATHER4_LZ_O,
-	IMAGE_GATHER4_C_O,
-	IMAGE_GATHER4_C_LZ_O,
-	IMAGE_GATHER4H,
+	IMAGE_GATHER4,
 	V_INTERP_P1_F32,
 	V_INTERP_P2_F32,
 	V_INTERP_MOV_F32,
+	V_INTERP_P1LL_F16,
+	V_INTERP_P2_F16,
 
 	S_NOP,
 	S_WAITCNT,
@@ -758,7 +761,9 @@ enum class OperandKind {
 	M0,
 	PopsExitingWaveId,
 	SharedBase,
+	SharedLimit,
 	PrivateBase,
+	PrivateLimit,
 	Null,
 	Vgpr,
 };

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

@@ -391,6 +391,8 @@ constexpr OpcodeMap VOP3_OPCODE_LIST[] = {
     {0x377u, Opcode::V_PERMLANE16_B32},
     {0x378u, Opcode::V_PERMLANEX16_B32},
     {0x34bu, Opcode::V_FMA_F16},
+    {0x342u, Opcode::V_INTERP_P1LL_F16},
+    {0x35au, Opcode::V_INTERP_P2_F16},
     {0x36du, Opcode::V_ADD3_U32},
     {0x14fu, Opcode::V_ALIGNBYTE_B32},
 };
@@ -1553,6 +1555,9 @@ bool HasUnsupportedNativeVop3Modifiers(Opcode opcode, bool permlane, bool mad_mi
 	}
 	switch (opcode) {
 		case Opcode::V_LDEXP_F32: return (abs & ~1u) != 0u || op_sel != 0u || (neg & ~1u) != 0u;
+		case Opcode::V_INTERP_P1LL_F16:
+		case Opcode::V_INTERP_P2_F16:
+			return abs != 0u || (op_sel & 0x7u) != 0u || clamp != 0u || omod != 0u || neg != 0u;
 		case Opcode::V_CNDMASK_B32:
 			return (abs & ~0x3u) != 0u || op_sel != 0u || clamp != 0u || omod != 0u ||
 			       (neg & ~0x3u) != 0u;
@@ -1664,6 +1669,17 @@ bool IsVopcCompareExec(Opcode opcode) {
 	}
 }
 
+// VINTRP and VOP3 interpolation share one layout: src0 = I/J VGPR, src1 = attr, src2 = chan.
+void SetInterpolationAttribute(Instruction& inst, uint32_t attr, uint32_t chan) {
+	inst.src1.kind       = OperandKind::IntegerInlineConstant;
+	inst.src1.value      = attr;
+	inst.src1.signed_val = static_cast<int32_t>(attr);
+	inst.src2.kind       = OperandKind::IntegerInlineConstant;
+	inst.src2.value      = chan;
+	inst.src2.signed_val = static_cast<int32_t>(chan);
+	inst.src_count       = 3;
+}
+
 } // namespace
 
 void DecodeVop2(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index,
@@ -1830,6 +1846,15 @@ void DecodeVop3(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 		inst.src_count = 0;
 		return;
 	}
+	if (inst.opcode == Opcode::V_INTERP_P1LL_F16 || inst.opcode == Opcode::V_INTERP_P2_F16) {
+		// SRC0 packs {attr_word, attr_chan, attr}; attr_word selects the high f16 of the dword.
+		DecodeVectorGpr(vdst, inst.dst);
+		DecodeScalarSource(src1, pc, inst.src0);
+		SetInterpolationAttribute(inst, src0 & 0x3fu, (src0 >> 6u) & 0x3u);
+		inst.src1.op_sel  = (src0 & 0x100u) != 0u;
+		inst.dst.sdwa_sel = (op_sel & 0x8u) != 0u ? 5u : 4u;
+		return;
+	}
 	if (compare_exec) {
 		inst.dst.kind = OperandKind::ExecLo;
 	} else if (scalar_dst) {
@@ -2004,13 +2029,7 @@ void DecodeVintrp(uint32_t pc, std::span<const uint32_t> code, uint32_t word_ind
 	} else {
 		DecodeVectorGpr(vsrc, inst.src0);
 	}
-	inst.src1.kind       = OperandKind::IntegerInlineConstant;
-	inst.src1.value      = attr;
-	inst.src1.signed_val = static_cast<int32_t>(attr);
-	inst.src2.kind       = OperandKind::IntegerInlineConstant;
-	inst.src2.value      = chan;
-	inst.src2.signed_val = static_cast<int32_t>(chan);
-	inst.src_count       = 3;
+	SetInterpolationAttribute(inst, attr, chan);
 }
 
 } // namespace Libs::Graphics::ShaderRecompiler::Decoder

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

@@ -73,14 +73,19 @@ void Translator::TranslateEmbeddedFetch(const Decoder::Instruction& inst, uint32
 	}
 }
 
-void Translator::V_INTERP_P1_F32() {}
-
 void Translator::V_INTERP_P2_F32(const Decoder::Instruction& inst) {
 	const auto value = ir.Emit(IR::ValueOpcode::GetAttribute,
 	                           {IR::Value(inst.src1.value), IR::Value(inst.src2.value)});
 	WriteOperand(inst.dst, value);
 }
 
+void Translator::V_INTERP_P2_F16(const Decoder::Instruction& inst) {
+	const auto value = ir.Emit(IR::ValueOpcode::GetAttributeF16,
+	                           {IR::Value(inst.src1.value), IR::Value(inst.src2.value),
+	                            IR::Value(uint32_t {inst.src1.op_sel})});
+	WriteF16(inst.dst, IR::F32(value));
+}
+
 void Translator::V_INTERP_MOV_F32(const Decoder::Instruction& inst) {
 	if (inst.src0.value >= 3u) {
 		EXIT("v_interp_mov_f32 mode %u is reserved at pc 0x%08x", inst.src0.value, inst.pc);
@@ -108,8 +113,11 @@ void Translator::EXP(const Decoder::Instruction& inst) {
 
 void Translator::EmitInterpolation(const Decoder::Instruction& inst) {
 	switch (inst.opcode) {
-		case Decoder::Opcode::V_INTERP_P1_F32: V_INTERP_P1_F32(); return;
+		// P2 interpolates the whole attribute, so the P1 halves have nothing to compute.
+		case Decoder::Opcode::V_INTERP_P1_F32:
+		case Decoder::Opcode::V_INTERP_P1LL_F16: return;
 		case Decoder::Opcode::V_INTERP_P2_F32: V_INTERP_P2_F32(inst); return;
+		case Decoder::Opcode::V_INTERP_P2_F16: V_INTERP_P2_F16(inst); return;
 		case Decoder::Opcode::V_INTERP_MOV_F32: V_INTERP_MOV_F32(inst); return;
 		default: return FailMissingTranslation(inst);
 	}

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

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

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

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

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

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

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

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

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

@@ -25,8 +25,7 @@ void Translator::V_MAD_I16(const Decoder::Instruction& inst) {
 	const auto rhs    = ReadU16AsU32(inst.src1, true);
 	auto       result = ir.IAdd(ir.IMul(lhs, rhs), ReadU16AsU32(inst.src2, true));
 	if (inst.dst.clamp) {
-		result = IR::U32(ir.Emit(IR::ValueOpcode::SMax32, {result, IR::Value(0xffff8000u)}));
-		result = IR::U32(ir.Emit(IR::ValueOpcode::SMin32, {result, IR::Value(0x7fffu)}));
+		result = Saturate16(result, true);
 	}
 	Write16Bits(DestinationOperand(inst), result);
 }
@@ -108,19 +107,7 @@ void Translator::S_U64_MASK(const Decoder::Instruction& inst, IR::ValueOpcode lo
 			default: return false;
 		}
 	};
-	if (is_exec_or_vcc(inst.dst) || is_exec_or_vcc(inst.src0) ||
-	    (inst.src_count > 1u && is_exec_or_vcc(inst.src1))) {
-		const auto mask = WriteMask(inst.dst, invocation_result, true);
-		ir.SetScc(ir.INotEqual(ir.BitwiseOr(mask[0], mask[1]), IR::U32(IR::Value(0u))));
-		return;
-	}
-
-	// Descriptor tracking reasons about each scalar word independently.
-	const auto lhs        = ReadU32Pair(inst.src0);
-	auto       mask_valid = ReadMaskValid(inst.src0);
-	if (inst.src_count > 1u) {
-		mask_valid = ir.LogicalAnd(mask_valid, ReadMaskValid(inst.src1));
-	}
+	const auto lhs = ReadU32Pair(inst.src0);
 	std::array<IR::U32, 2> result;
 	if (unary) {
 		result = {ir.BitwiseNot(lhs[0]), ir.BitwiseNot(lhs[1])};
@@ -133,11 +120,21 @@ void Translator::S_U64_MASK(const Decoder::Instruction& inst, IR::ValueOpcode lo
 			result[component] = negate_result ? ir.BitwiseNot(IR::U32(value)) : IR::U32(value);
 		}
 	}
-	WriteU32Pair(inst.dst, result);
-	if (inst.dst.kind == Decoder::OperandKind::Sgpr) {
-		const auto dst = static_cast<IR::ScalarReg>(inst.dst.reg);
-		ir.SetThreadBitScalarReg(dst, invocation_result);
-		ir.SetScalarMaskTag(dst, mask_valid);
+	if (is_exec_or_vcc(inst.dst) || is_exec_or_vcc(inst.src0) ||
+	    (inst.src_count > 1u && is_exec_or_vcc(inst.src1))) {
+		WriteMask(inst.dst, invocation_result, result, true);
+	} else {
+		// Descriptor tracking reasons about each scalar word independently.
+		auto mask_valid = ReadMaskValid(inst.src0);
+		if (inst.src_count > 1u) {
+			mask_valid = ir.LogicalAnd(mask_valid, ReadMaskValid(inst.src1));
+		}
+		WriteU32Pair(inst.dst, result);
+		if (inst.dst.kind == Decoder::OperandKind::Sgpr) {
+			const auto dst = static_cast<IR::ScalarReg>(inst.dst.reg);
+			ir.SetThreadBitScalarReg(dst, invocation_result);
+			ir.SetScalarMaskTag(dst, mask_valid);
+		}
 	}
 	ir.SetScc(ir.INotEqual(ir.BitwiseOr(result[0], result[1]), IR::U32(IR::Value(0u))));
 }
@@ -221,8 +218,17 @@ void Translator::V_XOR3_B32(const Decoder::Instruction& inst) {
 	WriteOperand(DestinationOperand(inst), result);
 }
 
-void Translator::S_FF_I32_B64(const Decoder::Instruction& inst, bool find_zero, bool from_msb) {
+void Translator::S_FF_I32_B64(const Decoder::Instruction& inst, bool find_zero, bool from_msb,
+                              bool sign) {
 	auto source = ReadU32Pair(inst.src0);
+	if (sign) {
+		// A signed source sign-extends a literal; then bits equal to the sign become zero.
+		if (inst.src0.kind == Decoder::OperandKind::LiteralConstant) {
+			source[1] = ir.ShiftRightArithmetic(source[0], IR::U32(IR::Value(31u)));
+		}
+		const auto mask = ir.ShiftRightArithmetic(source[1], IR::U32(IR::Value(31u)));
+		source          = {ir.BitwiseXor(source[0], mask), ir.BitwiseXor(source[1], mask)};
+	}
 	for (auto& word: source) {
 		word = find_zero ? ir.BitwiseNot(word) : word;
 	}
@@ -240,12 +246,19 @@ void Translator::S_FF_I32_B64(const Decoder::Instruction& inst, bool find_zero,
 
 void Translator::V_FFBH_32(const Decoder::Instruction& inst, bool sign) {
 	const auto source = ReadU32(inst.src0);
-	const auto value  = sign ? ir.BitwiseXor(
-	                              source, ir.ShiftRightArithmetic(source, IR::U32(IR::Value(31u))))
-	                         : source;
-	const auto msb      = IR::U32(ir.Emit(IR::ValueOpcode::FindUMsb32, {value}));
+	if (sign) {
+		// Bits equal to the sign become zero; FindILsb of the reversed bits is ~0 when none.
+		const auto bits =
+		    ir.BitwiseXor(source, ir.ShiftRightArithmetic(source, IR::U32(IR::Value(31u))));
+		const auto reversed = ir.Emit(IR::ValueOpcode::BitReverse32, {bits});
+		WriteOperand(DestinationOperand(inst),
+		             IR::U32(ir.Emit(IR::ValueOpcode::FindILsb32, {reversed})));
+		return;
+	}
+	// ResourceTracking matches this 31 - FindUMsb32 form.
+	const auto msb      = IR::U32(ir.Emit(IR::ValueOpcode::FindUMsb32, {source}));
 	const auto position = ir.ISub(IR::U32(IR::Value(31u)), msb);
-	const auto result   = ir.Select(ir.INotEqual(value, IR::U32(IR::Value(0u))), position,
+	const auto result   = ir.Select(ir.INotEqual(source, IR::U32(IR::Value(0u))), position,
 	                                IR::U32(IR::Value(0xffffffffu)));
 	WriteOperand(DestinationOperand(inst), result);
 }
@@ -277,7 +290,8 @@ void Translator::V_MAD_U64_U32(const Decoder::Instruction& inst) {
 	if (inst.dst2.kind != Decoder::OperandKind::Null &&
 	    inst.dst2.kind != Decoder::OperandKind::Unknown) {
 		// The 32x32 product fits in 64 bits, so only the addition can carry.
-		WriteMask(inst.dst2, IR::U1(ir.Emit(IR::ValueOpcode::ULessThan64, {result, product})));
+		const auto carry = IR::U1(ir.Emit(IR::ValueOpcode::ULessThan64, {result, product}));
+		WriteVectorMask(inst.dst2, carry);
 	}
 }
 
@@ -457,7 +471,7 @@ void Translator::S_BITCMP_B32(const Decoder::Instruction& inst, bool expected) {
 	const auto offset = ir.BitwiseAnd(ReadU32(inst.src1), IR::U32(IR::Value(31u)));
 	const auto bit =
 	    IR::U32(ir.Emit(IR::ValueOpcode::BitFieldUExtract, {value, offset, IR::Value(1u)}));
-	WriteCompareResult(inst.dst, ir.IEqual(bit, IR::U32(IR::Value(expected ? 1u : 0u))));
+	ir.SetScc(ir.IEqual(bit, IR::U32(IR::Value(expected ? 1u : 0u))));
 }
 
 void Translator::S_BITCMP_B64(const Decoder::Instruction& inst, bool expected) {
@@ -468,7 +482,7 @@ void Translator::S_BITCMP_B64(const Decoder::Instruction& inst, bool expected) {
 	    ir.Select(ir.ULessThan(offset, IR::U32(IR::Value(32u))), value[0], value[1]);
 	const auto bit =
 	    IR::U32(ir.Emit(IR::ValueOpcode::BitFieldUExtract, {word, word_bit, IR::Value(1u)}));
-	WriteCompareResult(inst.dst, ir.IEqual(bit, IR::U32(IR::Value(expected ? 1u : 0u))));
+	ir.SetScc(ir.IEqual(bit, IR::U32(IR::Value(expected ? 1u : 0u))));
 }
 
 void Translator::V_ALIGN_B32(const Decoder::Instruction& inst, bool byte_offset) {

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

@@ -1,11 +1,9 @@
 #include "common/assert.h"
-#include "common/logging/log.h"
 #include "graphics/shader/recompiler/frontend/translate/Translator.h"
 #include "graphics/shader/recompiler/frontend/decode/ImageOps.h"
 
 #include <algorithm>
 #include <array>
-#include <atomic>
 
 namespace Libs::Graphics::ShaderRecompiler::Frontend {
 
@@ -324,8 +322,8 @@ IR::Value Translator::MakeImageAddress(const Decoder::Instruction& inst,
                                        const Decoder::Operand&     base) {
 	std::array<IR::Value, 13> components {};
 	components.fill(IR::Value(0u));
-	const auto count =
-	    Decoder::ImageAddressDwordCount(inst.image_sample_flags, inst.image_address_components);
+	const auto count = Decoder::ImageAddressDwordCount(
+	    inst.image_sample_flags, inst.image_dimension, inst.image_address_components);
 	EXIT_IF(count > components.size());
 	const auto nsa_components =
 	    std::min(inst.image_nsa_dwords * 4u, Decoder::MaxImageNsaAddressComponents);
@@ -455,12 +453,17 @@ void Translator::BUFFER_LOAD(const Decoder::Instruction& inst) {
 	            AddMemoryInfo(memory, inst.pc));
 	if (bits != 32u) {
 		WriteOperand(inst.dst, WidenSubdword(loaded, bits, sign));
-	} else if (memory.data_dwords == 1u) {
-		WriteOperand(inst.dst, loaded);
-	} else {
-		for (uint32_t component = 0; component < memory.data_dwords; component++) {
-			WriteOperand(OffsetOperand(inst.dst, component),
-			             ir.CompositeExtract(loaded, component));
+		return;
+	}
+	for (uint32_t component = 0; component < memory.data_dwords; component++) {
+		const auto dst = OffsetOperand(inst.dst, component);
+		const auto value =
+		    memory.data_dwords == 1u ? IR::U32(loaded) : ir.CompositeExtract(loaded, component);
+		// D16 formats with an odd component count keep the high half of their final VGPR.
+		if (memory.data_bits == 16u && component * 2u + 1u == memory.component_count) {
+			Write16Bits(dst, value);
+		} else {
+			WriteOperand(dst, value);
 		}
 	}
 }
@@ -860,9 +863,10 @@ void Translator::DS_WRITE2(const Decoder::Instruction& inst) {
 
 void Translator::DS_APPEND_CONSUME(const Decoder::Instruction& inst, IR::ValueOpcode opcode) {
 	const auto memory = MemoryInfoFromDecoded(inst);
-	WriteOperand(inst.dst, ir.Emit(opcode,
-	                               {ReadU32(MemorySourceAt(inst, 0)), ir.GetExec()},
-	                               AddMemoryInfo(memory, inst.pc)));
+	const auto exec   = ExecMask();
+	WriteOperand(inst.dst,
+	             ir.Emit(opcode, {ReadU32(MemorySourceAt(inst, 0)), exec[0], exec[1], ir.GetExec()},
+	                     AddMemoryInfo(memory, inst.pc)));
 }
 
 void Translator::DS_ADDTID(const Decoder::Instruction& inst, bool write) {
@@ -898,10 +902,7 @@ void Translator::DS_ORDERED_COUNT(const Decoder::Instruction& inst) {
 		const auto packer = ir.BitwiseAnd(m0, IR::U32(IR::Value(0xfu)));
 		address = ir.IAdd(address, ir.ShiftLeftLogical(packer, IR::U32(IR::Value(2u))));
 	}
-	const auto exec   = ir.GetExec();
-	const auto ballot = ir.Emit(IR::ValueOpcode::Ballot, {exec});
-	const auto low = ir.CompositeExtract(ballot, 0u);
-	const auto high = ir.CompositeExtract(ballot, 1u);
+	const auto [low, high] = ExecMask();
 	const auto active = ir.INotEqual(ir.BitwiseOr(low, high), IR::U32(IR::Value(0u)));
 	const auto first = ir.Select(
 	    ir.INotEqual(low, IR::U32(IR::Value(0u))),
@@ -940,6 +941,8 @@ void Translator::DS_PERMUTE(const Decoder::Instruction& inst, bool backward) {
 
 void Translator::EmitMemory(const Decoder::Instruction& inst) {
 	switch (inst.opcode) {
+		// Translated scalar loads have no cache of their own to invalidate.
+		case Decoder::Opcode::S_DCACHE_INV: return;
 		case Decoder::Opcode::BUFFER_GL0_INV:
 		case Decoder::Opcode::BUFFER_GL1_INV:
 			ir.Emit(IR::ValueOpcode::DataCacheInvalidate);
@@ -964,17 +967,9 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::S_BUFFER_LOAD_DWORDX4:
 		case Decoder::Opcode::S_BUFFER_LOAD_DWORDX8:
 		case Decoder::Opcode::S_BUFFER_LOAD_DWORDX16: return S_LOAD(inst, false);
-		case Decoder::Opcode::S_MEMREALTIME: {
-			static std::atomic_flag warned = ATOMIC_FLAG_INIT;
-			if (!warned.test_and_set(std::memory_order_relaxed)) {
-				Log::WriteToConsoleAndLog(
-				    "Warning: S_MEMREALTIME uses placeholder UINT64_MAX; real-time clock not implemented.\n");
-			}
-			for (uint32_t component = 0; component < 2; component++) {
-				WriteOperand(ScalarDestinationOperand(inst.dst, component), IR::Value(UINT32_MAX));
-			}
-			return;
-		}
+		case Decoder::Opcode::S_MEMREALTIME:
+			return WriteU32Pair(inst.dst,
+			                    ExtractU64(IR::U64(ir.Emit(IR::ValueOpcode::ReadRealtimeClock))));
 
 		case Decoder::Opcode::BUFFER_LOAD_UBYTE:
 		case Decoder::Opcode::BUFFER_LOAD_UBYTE_D16:
@@ -992,6 +987,8 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 		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::BUFFER_LOAD_FORMAT_D16_XY:
+		case Decoder::Opcode::BUFFER_LOAD_FORMAT_D16_XYZ:
 		case Decoder::Opcode::BUFFER_LOAD_FORMAT_D16_XYZW:
 		case Decoder::Opcode::TBUFFER_LOAD_FORMAT_X:
 		case Decoder::Opcode::TBUFFER_LOAD_FORMAT_XY:
@@ -1012,6 +1009,8 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::BUFFER_STORE_FORMAT_XYZW:
 		case Decoder::Opcode::BUFFER_STORE_FORMAT_D16_X:
 		case Decoder::Opcode::BUFFER_STORE_FORMAT_D16_HI_X:
+		case Decoder::Opcode::BUFFER_STORE_FORMAT_D16_XY:
+		case Decoder::Opcode::BUFFER_STORE_FORMAT_D16_XYZ:
 		case Decoder::Opcode::BUFFER_STORE_FORMAT_D16_XYZW:
 		case Decoder::Opcode::TBUFFER_STORE_FORMAT_X:
 		case Decoder::Opcode::TBUFFER_STORE_FORMAT_XY:
@@ -1032,6 +1031,8 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 			return BUFFER_ATOMIC(inst, IR::ValueOpcode::BufferAtomicSMin32);
 		case Decoder::Opcode::BUFFER_ATOMIC_UMIN:
 			return BUFFER_ATOMIC(inst, IR::ValueOpcode::BufferAtomicUMin32);
+		case Decoder::Opcode::BUFFER_ATOMIC_UMIN_X2:
+			return BUFFER_ATOMIC(inst, IR::ValueOpcode::BufferAtomicUMin64);
 		case Decoder::Opcode::BUFFER_ATOMIC_SMAX:
 			return BUFFER_ATOMIC(inst, IR::ValueOpcode::BufferAtomicSMax32);
 		case Decoder::Opcode::BUFFER_ATOMIC_UMAX:
@@ -1149,6 +1150,7 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::FLAT_LOAD_USHORT:
 		case Decoder::Opcode::FLAT_LOAD_SSHORT:
 		case Decoder::Opcode::FLAT_LOAD_SHORT_D16:
+		case Decoder::Opcode::FLAT_LOAD_SHORT_D16_HI:
 		case Decoder::Opcode::FLAT_LOAD_DWORD:
 		case Decoder::Opcode::FLAT_LOAD_DWORDX2:
 		case Decoder::Opcode::FLAT_LOAD_DWORDX3:
@@ -1168,15 +1170,7 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::IMAGE_STORE:
 		case Decoder::Opcode::IMAGE_STORE_MIP: return IMAGE_STORE(inst);
 		case Decoder::Opcode::IMAGE_SAMPLE: return IMAGE_SAMPLE(inst);
-		case Decoder::Opcode::IMAGE_GATHER4_L:
-		case Decoder::Opcode::IMAGE_GATHER4_LZ:
-		case Decoder::Opcode::IMAGE_GATHER4_C:
-		case Decoder::Opcode::IMAGE_GATHER4_C_L:
-		case Decoder::Opcode::IMAGE_GATHER4_C_LZ:
-		case Decoder::Opcode::IMAGE_GATHER4_LZ_O:
-		case Decoder::Opcode::IMAGE_GATHER4_C_O:
-		case Decoder::Opcode::IMAGE_GATHER4_C_LZ_O:
-		case Decoder::Opcode::IMAGE_GATHER4H: return IMAGE_GATHER(inst);
+		case Decoder::Opcode::IMAGE_GATHER4: return IMAGE_GATHER(inst);
 
 		case Decoder::Opcode::DS_MIN_F32:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicFMin32, false);

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

@@ -73,43 +73,41 @@ void Translator::EmitScalar(const Decoder::Instruction& inst) {
 
 		case O::S_CMP_EQ_U32:
 		case O::S_CMP_EQ_I32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::IEqual32, IR::Type::U32, true, false);
+			EmitIntegerCompare(inst, IR::ValueOpcode::IEqual32, IR::Type::U32, true);
 			return;
 		case O::S_CMP_LG_U32:
 		case O::S_CMP_LG_I32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::INotEqual32, IR::Type::U32, true, false);
+			EmitIntegerCompare(inst, IR::ValueOpcode::INotEqual32, IR::Type::U32, true);
 			return;
 		case O::S_CMP_GT_U32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::UGreaterThan32, IR::Type::U32, true, false);
+			EmitIntegerCompare(inst, IR::ValueOpcode::UGreaterThan32, IR::Type::U32, true);
 			return;
 		case O::S_CMP_GE_U32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::UGreaterThanEqual32, IR::Type::U32, true,
-			                   false);
+			EmitIntegerCompare(inst, IR::ValueOpcode::UGreaterThanEqual32, IR::Type::U32, true);
 			return;
 		case O::S_CMP_LT_U32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThan32, IR::Type::U32, true, false);
+			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThan32, IR::Type::U32, true);
 			return;
 		case O::S_CMP_LE_U32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThanEqual32, IR::Type::U32, true, false);
+			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThanEqual32, IR::Type::U32, true);
 			return;
 		case O::S_CMP_GT_I32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::SGreaterThan32, IR::Type::U32, true, false);
+			EmitIntegerCompare(inst, IR::ValueOpcode::SGreaterThan32, IR::Type::U32, true);
 			return;
 		case O::S_CMP_GE_I32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::SGreaterThanEqual32, IR::Type::U32, true,
-			                   false);
+			EmitIntegerCompare(inst, IR::ValueOpcode::SGreaterThanEqual32, IR::Type::U32, true);
 			return;
 		case O::S_CMP_LT_I32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThan32, IR::Type::U32, true, false);
+			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThan32, IR::Type::U32, true);
 			return;
 		case O::S_CMP_LE_I32:
-			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThanEqual32, IR::Type::U32, true, false);
+			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThanEqual32, IR::Type::U32, true);
 			return;
 		case O::S_CMP_EQ_U64:
-			EmitIntegerCompare(inst, IR::ValueOpcode::IEqual64, IR::Type::U64, true, false);
+			EmitIntegerCompare(inst, IR::ValueOpcode::IEqual64, IR::Type::U64, true);
 			return;
 		case O::S_CMP_LG_U64:
-			EmitIntegerCompare(inst, IR::ValueOpcode::INotEqual64, IR::Type::U64, true, false);
+			EmitIntegerCompare(inst, IR::ValueOpcode::INotEqual64, IR::Type::U64, true);
 			return;
 
 		case O::S_AND_B64:
@@ -238,10 +236,12 @@ void Translator::EmitScalar(const Decoder::Instruction& inst) {
 			return ComposedIntegerBinary(inst, IR::ValueOpcode::BitwiseOr32, false, true, true);
 		case O::S_XNOR_B32:
 			return ComposedIntegerBinary(inst, IR::ValueOpcode::BitwiseXor32, false, true, true);
-		case O::S_FF0_I32_B64: return S_FF_I32_B64(inst, true, false);
-		case O::S_FF1_I32_B64: return S_FF_I32_B64(inst, false, false);
+		case O::S_FF0_I32_B64: return S_FF_I32_B64(inst, true, false, false);
+		case O::S_FF1_I32_B64: return S_FF_I32_B64(inst, false, false, false);
 		case O::S_FLBIT_I32_B32: return V_FFBH_32(inst, false);
-		case O::S_FLBIT_I32_B64: return S_FF_I32_B64(inst, false, true);
+		case O::S_FLBIT_I32_B64: return S_FF_I32_B64(inst, false, true, false);
+		case O::S_FLBIT_I32: return V_FFBH_32(inst, true);
+		case O::S_FLBIT_I32_I64: return S_FF_I32_B64(inst, false, true, true);
 
 		case O::S_BITSET0_B32: return S_BITSET_B32(inst, false);
 		case O::S_BITSET1_B32: return S_BITSET_B32(inst, true);

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

@@ -129,6 +129,20 @@ std::array<IR::U32, 2> Translator::BallotMask(IR::U1 value) {
 	        program.wave_size == 64u ? ir.CompositeExtract(mask, 1) : IR::U32(IR::Value(0u))};
 }
 
+std::array<IR::U32, 2> Translator::ExecMask() {
+	return {ir.GetExecLo(), program.wave_size == 64u ? ir.GetExecHi() : IR::U32(IR::Value(0u))};
+}
+
+// Pixel EXEC words exclude the helper invocations that the per-lane EXEC keeps enabled.
+std::array<IR::U32, 2> Translator::VectorMask(IR::U1 value) {
+	auto mask = BallotMask(value);
+	if (program.stage == ShaderType::Pixel) {
+		const auto exec = ExecMask();
+		mask = {ir.BitwiseAnd(mask[0], exec[0]), ir.BitwiseAnd(mask[1], exec[1])};
+	}
+	return mask;
+}
+
 IR::U32 Translator::ReadRawU32(const Decoder::Operand& operand) {
 	switch (operand.kind) {
 		case Decoder::OperandKind::LiteralConstant:
@@ -138,6 +152,8 @@ IR::U32 Translator::ReadRawU32(const Decoder::Operand& operand) {
 		case Decoder::OperandKind::SharedBase:
 		case Decoder::OperandKind::PrivateBase:
 		case Decoder::OperandKind::PopsExitingWaveId: return IR::U32(IR::Value(0u));
+		case Decoder::OperandKind::SharedLimit:
+		case Decoder::OperandKind::PrivateLimit: return IR::U32(IR::Value(0xffffffffu));
 		case Decoder::OperandKind::Sgpr:
 			return ir.GetScalarReg(static_cast<IR::ScalarReg>(operand.reg));
 		case Decoder::OperandKind::Vgpr:
@@ -319,7 +335,10 @@ void Translator::WriteRawU32(const Decoder::Operand& operand, IR::U32 value) {
 		case Decoder::OperandKind::Vgpr: {
 			const auto reg = static_cast<IR::VectorReg>(operand.reg);
 			const auto old = ir.GetVectorReg(reg);
-			if (operand.dpp) {
+			// DPP skips EXEC lanes only through row/bank masks, or for invalid sources without
+			// BOUND_CTRL (dpp8 reads zero); otherwise it is a plain masked write.
+			if (operand.dpp && (operand.dpp_row_mask != 0xfu || operand.dpp_bank_mask != 0xfu ||
+			                    (!operand.dpp_bound_ctrl && !operand.dpp8))) {
 				value = IR::U32(ir.Emit(IR::ValueOpcode::DppUpdateU32, {value, old, ir.GetExec()},
 				                        DppFlags(operand)));
 			} else {
@@ -385,21 +404,9 @@ void Translator::WriteOperand(const Decoder::Operand& operand, IR::Value value)
 		switch (operand.kind) {
 			case Decoder::OperandKind::Scc: ir.SetScc(IR::U1(value)); return;
 			case Decoder::OperandKind::ExecLo:
-			case Decoder::OperandKind::ExecHi: {
-				const auto mask = BallotMask(IR::U1(value));
-				ir.SetExec(IR::U1(value));
-				ir.SetExecLo(mask[0]);
-				ir.SetExecHi(mask[1]);
-				return;
-			}
+			case Decoder::OperandKind::ExecHi:
 			case Decoder::OperandKind::VccLo:
-			case Decoder::OperandKind::VccHi: {
-				const auto mask = BallotMask(IR::U1(value));
-				ir.SetVcc(IR::U1(value));
-				ir.SetVccLo(mask[0]);
-				ir.SetVccHi(mask[1]);
-				return;
-			}
+			case Decoder::OperandKind::VccHi: EXIT("lane masks are written through WriteMask");
 			default:
 				WriteRawU32(operand, ir.Select(IR::U1(value), IR::U32(IR::Value(1u)),
 				                               IR::U32(IR::Value(0u))));
@@ -468,17 +475,17 @@ IR::U32 Translator::ReadU32(const Decoder::Operand& operand) {
 }
 
 std::array<IR::U32, 2> Translator::ReadU32Pair(const Decoder::Operand& operand) {
-	if (operand.kind == Decoder::OperandKind::PrivateBase ||
-	    operand.kind == Decoder::OperandKind::SharedBase) {
-		return {IR::U32(IR::Value(0u)), IR::U32(IR::Value(
-		    operand.kind == Decoder::OperandKind::PrivateBase ? Decoder::PrivateApertureHigh
-		                                                      : Decoder::SharedApertureHigh))};
-	}
-	if (operand.kind == Decoder::OperandKind::ExecLo) {
-		return {ir.GetExecLo(), ir.GetExecHi()};
-	}
-	if (operand.kind == Decoder::OperandKind::VccLo) {
-		return {ir.GetVccLo(), ir.GetVccHi()};
+	// Each aperture limit is its base plus 0xffffffff.
+	switch (operand.kind) {
+		case Decoder::OperandKind::SharedBase:
+		case Decoder::OperandKind::SharedLimit:
+			return {ReadRawU32(operand), IR::U32(IR::Value(Decoder::SharedApertureHigh))};
+		case Decoder::OperandKind::PrivateBase:
+		case Decoder::OperandKind::PrivateLimit:
+			return {ReadRawU32(operand), IR::U32(IR::Value(Decoder::PrivateApertureHigh))};
+		case Decoder::OperandKind::ExecLo: return {ir.GetExecLo(), ir.GetExecHi()};
+		case Decoder::OperandKind::VccLo: return {ir.GetVccLo(), ir.GetVccHi()};
+		default: break;
 	}
 	if (operand.kind == Decoder::OperandKind::FloatInlineConstant) {
 		const auto bits =
@@ -636,7 +643,7 @@ IR::U1 Translator::ReadMask(const Decoder::Operand& operand) {
 		}
 		case Decoder::OperandKind::ExecLo:
 		case Decoder::OperandKind::ExecHi: return ir.GetExec();
-		case Decoder::OperandKind::VccLo:
+		case Decoder::OperandKind::VccLo: return ir.GetVcc();
 		case Decoder::OperandKind::VccHi:
 			return program.wave_size == 32u
 			           ? ThreadBit({ReadRawU32(operand), IR::U32(IR::Value(0u))})
@@ -674,9 +681,8 @@ IR::U1 Translator::ReadMaskValid(const Decoder::Operand& operand) {
 	}
 }
 
-std::array<IR::U32, 2> Translator::WriteMask(const Decoder::Operand& operand, IR::U1 value,
-                                             bool write_64) {
-	const auto mask = BallotMask(value);
+void Translator::WriteMask(const Decoder::Operand& operand, IR::U1 value,
+                           const std::array<IR::U32, 2>& mask, bool write_64) {
 	switch (operand.kind) {
 		case Decoder::OperandKind::Sgpr: {
 			const auto reg  = static_cast<IR::ScalarReg>(operand.reg);
@@ -697,43 +703,43 @@ std::array<IR::U32, 2> Translator::WriteMask(const Decoder::Operand& operand, IR
 				ir.SetThreadBitScalarReg(high, IR::U1(IR::Value(false)));
 				ir.SetScalarMaskTag(high, IR::U1(IR::Value(false)));
 			}
-			return mask;
+			return;
 		}
 		case Decoder::OperandKind::ExecLo:
 		case Decoder::OperandKind::ExecHi: {
 			ir.SetExec(value);
 			ir.SetExecLo(mask[0]);
 			ir.SetExecHi(mask[1]);
-			return mask;
+			return;
 		}
 		case Decoder::OperandKind::VccLo:
 		case Decoder::OperandKind::VccHi: {
-			if (!write_64 && program.wave_size == 32u) {
+			if (operand.kind == Decoder::OperandKind::VccHi && !write_64 &&
+			    program.wave_size == 32u) {
 				WriteRawU32(operand, mask[0]);
-				return mask;
+				return;
 			}
 			ir.SetVcc(value);
 			ir.SetVccLo(mask[0]);
-			ir.SetVccHi(mask[1]);
-			return mask;
+			if (write_64 || program.wave_size == 64u) {
+				ir.SetVccHi(mask[1]);
+			}
+			return;
 		}
-		case Decoder::OperandKind::Scc: ir.SetScc(value); return mask;
+		case Decoder::OperandKind::Scc: ir.SetScc(value); return;
 		default:
 			WriteRawU32(operand, mask[0]);
 			WriteRawU32(OffsetOperand(operand, 1), mask[1]);
-			return mask;
+			return;
 	}
 }
 
-void Translator::WriteCompareResult(const Decoder::Operand& operand, IR::U1 value) {
-	if (operand.kind == Decoder::OperandKind::Scc) {
-		WriteOperand(operand, value);
-		return;
-	}
+void Translator::WriteVectorMask(const Decoder::Operand& operand, IR::U1 value) {
 	// VALU instructions only write lanes enabled by the incoming EXEC mask. This is
 	// especially important for V_CMPX: replacing EXEC with the ungated comparison would
 	// reactivate lanes disabled by an enclosing divergent region.
-	WriteMask(operand, ir.LogicalAnd(ir.GetExec(), value));
+	const auto active = ir.LogicalAnd(ir.GetExec(), value);
+	WriteMask(operand, active, VectorMask(active));
 }
 
 void Translator::AddBranchCondition(const CFG::Graph& graph, const CFG::BasicBlock& source,
@@ -1150,7 +1156,10 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 			}
 		}
 		entry_ir.SetExec(initial_exec);
-		const auto initial_mask = entry_ir.Emit(IR::ValueOpcode::Ballot, {initial_exec});
+		// Pixel EXEC words start with the live (non-helper) lanes; helpers keep the per-lane EXEC.
+		const auto initial_mask = options.stage == ShaderType::Pixel
+		                              ? entry_ir.Emit(IR::ValueOpcode::LiveMask)
+		                              : entry_ir.Emit(IR::ValueOpcode::Ballot, {initial_exec});
 		const auto initial_lo = entry_ir.CompositeExtract(initial_mask, 0);
 		const auto initial_hi = options.wave_size == 64u ? entry_ir.CompositeExtract(initial_mask, 1)
 		                                               : IR::U32(IR::Value(0u));
@@ -1235,12 +1244,12 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 			    entry_ir.BitwiseOr(wave_info, entry_ir.BitwiseOr(entry_ir.ShiftLeftLogical(
 			                                                         primitive_count, u32(8)),
 			                                                     vertex_count)));
+			// Instance IDs exclude the start instance, which shaders add from user data.
 			if (mesh.fast_launch) {
 				// Fast launch broadcasts the subgroup's base vertex and instance to every lane.
 				entry_ir.SetVectorReg(static_cast<IR::VectorReg>(5), entry_ir.IAdd(draw(1), group));
-				entry_ir.SetVectorReg(
-				    static_cast<IR::VectorReg>(6),
-				    entry_ir.IAdd(draw(2), builtin(IR::StageInputKind::WorkgroupId, 1)));
+				entry_ir.SetVectorReg(static_cast<IR::VectorReg>(6),
+				                      builtin(IR::StageInputKind::WorkgroupId, 1));
 			} else {
 				// GS adjacency addresses local ES records in LDS. Fans retain the draw's
 				// center in every subgroup; strip winding follows the global primitive.
@@ -1268,14 +1277,14 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 				                                         entry_ir.ShiftLeftLogical(second, u32(18))));
 				entry_ir.SetVectorReg(static_cast<IR::VectorReg>(1),
 				                      entry_ir.ShiftLeftLogical(third, u32(2)));
-				const auto index_bytes  = draw(3);
+				const auto index_bytes  = draw(2);
 				const auto indexed      = entry_ir.INotEqual(index_bytes, u32(0));
-				const auto index_low    = draw(4);
+				const auto index_low    = draw(3);
 				const auto byte_offset  = entry_ir.IAdd(entry_ir.BitwiseAnd(index_low, u32(3)),
 				                                           entry_ir.IMul(input_vertex, index_bytes));
 				const auto index_resource = entry_ir.Emit(
 				    IR::ValueOpcode::GetAddressResource,
-				    {entry_ir.BitwiseAnd(index_low, u32(~3u)), draw(5)});
+				    {entry_ir.BitwiseAnd(index_low, u32(~3u)), draw(4)});
 				const auto memory_index = static_cast<uint32_t>(result.memory_info.size());
 				result.memory_info.push_back({.kind = IR::ResourceKind::Global});
 				const auto packed_index = entry_ir.Emit(
@@ -1289,9 +1298,8 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 				     entry_ir.IMul(index_bytes, u32(8))}));
 				entry_ir.SetVectorReg(static_cast<IR::VectorReg>(5),
 				                      entry_ir.IAdd(draw(1), entry_ir.Select(indexed, index, input_vertex)));
-				entry_ir.SetVectorReg(
-				    static_cast<IR::VectorReg>(8),
-				    entry_ir.IAdd(draw(2), builtin(IR::StageInputKind::WorkgroupId, 1)));
+				entry_ir.SetVectorReg(static_cast<IR::VectorReg>(8),
+				                      builtin(IR::StageInputKind::WorkgroupId, 1));
 			}
 		} else if (options.stage == ShaderType::Local) {
 			entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(3), IR::U32(IR::Value(64u)));
@@ -1381,10 +1389,27 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 			                      IR::U32(IR::Value(options.wave_size << 12u)));
 			entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(3),
 			                      entry_ir.BitwiseOr(vertex_count, IR::U32(IR::Value(1u << 28u))));
+			// Vulkan adds an indirect record's bases to the indices; the hardware passes them
+			// only in the user SGPRs the packet patches.
+			const auto& offsets = options.input_info.vertex->indirect_offsets;
+			const auto  index   = [&](uint8_t slot, IR::StageInputKind kind,
+			                          IR::StageInputKind base_kind) {
+				if (offsets == ShaderIndirectOffsets {}) {
+					return builtin(kind);
+				}
+				const auto base = builtin(base_kind);
+				if (slot < options.user_data_count) {
+					entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(options.user_data_base + slot),
+					                      base);
+				}
+				return entry_ir.ISub(builtin(kind), base);
+			};
 			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(5),
-			                      builtin(IR::StageInputKind::VertexIndex));
+			                      index(offsets.vertex, IR::StageInputKind::VertexIndex,
+			                            IR::StageInputKind::BaseVertex));
 			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(8),
-			                      builtin(IR::StageInputKind::InstanceIndex));
+			                      index(offsets.instance, IR::StageInputKind::InstanceIndex,
+			                            IR::StageInputKind::BaseInstance));
 		}
 	}
 	const bool flush_f32_inputs = options.stage == ShaderType::Compute &&

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

@@ -27,6 +27,8 @@ private:
 	Decoder::Operand ScalarDestinationOperand(const Decoder::Operand& operand, uint32_t offset);
 	Decoder::Operand PlainOperand(const Decoder::Operand& operand);
 	std::array<IR::U32, 2> BallotMask(IR::U1 value);
+	std::array<IR::U32, 2> ExecMask();
+	std::array<IR::U32, 2> VectorMask(IR::U1 value);
 	IR::U32                ReadRawU32(const Decoder::Operand& operand);
 	IR::U32                ReadScalarCode(uint32_t code);
 	IR::U32                ApplyBitSourceModifiers(const Decoder::Operand& operand, IR::U32 value);
@@ -52,9 +54,9 @@ private:
 	IR::U32 ConditionBit(const Decoder::Operand& operand);
 	IR::U1  ReadMask(const Decoder::Operand& operand);
 	IR::U1  ReadMaskValid(const Decoder::Operand& operand);
-	std::array<IR::U32, 2> WriteMask(const Decoder::Operand& operand, IR::U1 value,
-	                                 bool write_64 = false);
-	void    WriteCompareResult(const Decoder::Operand& operand, IR::U1 value);
+	void    WriteMask(const Decoder::Operand& operand, IR::U1 value,
+	                  const std::array<IR::U32, 2>& mask, bool write_64 = false);
+	void    WriteVectorMask(const Decoder::Operand& operand, IR::U1 value);
 
 	IR::MemoryFlags AddMemoryInfo(const IR::MemoryInfo& memory, uint32_t pc);
 	IR::ExportFlags AddExportInfo(const Decoder::Instruction& inst);
@@ -114,17 +116,16 @@ private:
 
 	IR::F32 SelectF32(IR::U1 condition, IR::F32 true_value, IR::F32 false_value);
 	IR::U32 PackU16Lanes(IR::U32 low, IR::U32 high);
-	void EmitCompareResult(const Decoder::Instruction& inst, IR::U1 value, bool scalar, bool cmpx);
-	void EmitCompareConstant(const Decoder::Instruction& inst, bool value, bool scalar, bool cmpx);
+	IR::U32 Saturate16(IR::U32 value, bool signed_value);
+	void EmitCompareResult(const Decoder::Instruction& inst, IR::U1 value, bool scalar);
+	void EmitCompareConstant(const Decoder::Instruction& inst, bool value, bool scalar);
 	void EmitIntegerCompare(const Decoder::Instruction& inst, IR::ValueOpcode opcode, IR::Type type,
-	                        bool scalar, bool cmpx, bool swap_operands = false);
+	                        bool scalar, bool swap_operands = false);
 	void EmitInteger16Compare(const Decoder::Instruction& inst, IR::ValueOpcode opcode,
-	                          bool signed_value, bool cmpx);
-	void EmitFloatCompare(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool half,
-	                      bool cmpx);
-	void EmitFloatOrderedCompare(const Decoder::Instruction& inst, bool ordered, bool cmpx,
-	                             bool is64 = false);
-	void EmitFloatClassCompare(const Decoder::Instruction& inst, bool cmpx, bool half = false);
+	                          bool signed_value);
+	void EmitFloatCompare(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool half);
+	void EmitFloatOrderedCompare(const Decoder::Instruction& inst, bool ordered, bool is64 = false);
+	void EmitFloatClassCompare(const Decoder::Instruction& inst, bool half = false);
 	void V_CVT_F32_UBYTE(const Decoder::Instruction& inst, uint32_t byte_index);
 	void V_CVT_F32_U32(const Decoder::Instruction& inst);
 	void V_CVT_F32_I32(const Decoder::Instruction& inst);
@@ -141,6 +142,7 @@ private:
 	void V_CVT_PKRTZ_F16_F32(const Decoder::Instruction& inst);
 	void V_CVT_PKNORM_F32(const Decoder::Instruction& inst, IR::ValueOpcode opcode);
 	void V_CVT_PK_U8_F32(const Decoder::Instruction& inst);
+	void V_CVT_PK_16_32(const Decoder::Instruction& inst, bool signed_value);
 	void V_PACK_B32_F16(const Decoder::Instruction& inst);
 	void PackedFloat16(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool accumulator,
 	                   bool quiet_snan);
@@ -151,6 +153,7 @@ private:
 	void Float16Ternary(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool accumulator,
 	                    bool mix);
 	void FloatUnary(const Decoder::Instruction& inst, IR::ValueOpcode opcode);
+	void V_FRACT(const Decoder::Instruction& inst, IR::Type type);
 	void FloatBinary(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool reverse);
 	void FloatTernary(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool accumulator,
 	                  bool mix);
@@ -182,7 +185,7 @@ private:
 	void V_AND_OR_B32(const Decoder::Instruction& inst);
 	void V_OR3_B32(const Decoder::Instruction& inst);
 	void V_XOR3_B32(const Decoder::Instruction& inst);
-	void S_FF_I32_B64(const Decoder::Instruction& inst, bool find_zero, bool from_msb);
+	void S_FF_I32_B64(const Decoder::Instruction& inst, bool find_zero, bool from_msb, bool sign);
 	void V_FFBH_32(const Decoder::Instruction& inst, bool sign);
 	void Integer24(const Decoder::Instruction& inst, bool sign, bool addend);
 	void V_MAD_U64_U32(const Decoder::Instruction& inst);
@@ -240,8 +243,8 @@ private:
 	void V_READLANE_B32(const Decoder::Instruction& inst);
 	void V_WRITELANE_B32(const Decoder::Instruction& inst);
 	void V_PERMLANE16_B32(const Decoder::Instruction& inst, bool x16);
-	void V_INTERP_P1_F32();
 	void V_INTERP_P2_F32(const Decoder::Instruction& inst);
+	void V_INTERP_P2_F16(const Decoder::Instruction& inst);
 	void V_INTERP_MOV_F32(const Decoder::Instruction& inst);
 	void EXP(const Decoder::Instruction& inst);