DualSense haptics and controller speaker sound bluethooth (#969)

* pad: send DualSense speaker and haptic audio over Bluetooth

* ci: fix Opus clang-cl build and macOS memory tracker test

* docs: remove Bluetooth notes from project README

* pad: keep Bluetooth audio work off emulation timing path

* pad: smooth DualSense speaker audio on USB and Bluetooth

* tests: find available macOS guest range for page manager

* pad: keep haptic waveform when Bluetooth queue overflows

* audio: reject zero output sample frequency

* Simplify DualSense audio routing and fix Bluetooth batching

---------

Co-authored-by: Codex <codex@openai.com>
Co-authored-by: nmzik <Nmzik@mail.ru>

.github/workflows/build.yml

@@ -112,12 +112,16 @@ jobs:
       - name: Build
         shell: cmd
         run: |
-          cmake --build _Build/windows --target launcher audio_out2_port_tests ime_dialog_tests virtual_memory_allocation_tests --parallel
+          cmake --build _Build/windows --target launcher ^
+            audio_out2_port_tests pad_haptics_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests ^
+            shader_recompiler_compute_tests ^
+            resource_tracking_tests scalar_provenance_tests ^
+            resource_materialization_tests shader_vertex_metadata_tests archive_file_tests avplayer_file_tests --parallel
 
       - name: Test
         shell: cmd
         run: |
-          ctest --test-dir _Build/windows --output-on-failure -R "^(audio_out2_port|ime_dialog|virtual_memory_allocation)$"
+          ctest --test-dir _Build/windows --output-on-failure -R "^(audio_out2_port|pad_haptics|ime_dialog|lru_cache|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|archive_file|avplayer_file|virtual_memory_allocation)$"
 
       - name: Install
         shell: cmd
@@ -198,14 +202,17 @@ jobs:
         shell: bash
         run: |
           cmake --build _Build/macos \
-            --target launcher audio_out2_port_tests ime_dialog_tests virtual_memory_allocation_tests \
+            --target launcher audio_out2_port_tests pad_haptics_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests \
+              memory_tracker_tests page_manager_tests shader_recompiler_compute_tests \
+              resource_tracking_tests scalar_provenance_tests \
+              resource_materialization_tests shader_vertex_metadata_tests archive_file_tests avplayer_file_tests \
             --parallel
 
       - name: Test
         shell: bash
         run: |
           ctest --test-dir _Build/macos --output-on-failure \
-            -R '^(audio_out2_port|ime_dialog|virtual_memory_allocation)$'
+            -R '^(audio_out2_port|pad_haptics|ime_dialog|lru_cache|memory_tracker|page_manager|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|archive_file|avplayer_file|virtual_memory_allocation)$'
 
       - name: Install
         shell: bash
@@ -379,14 +386,16 @@ jobs:
         run: |
           cmake --build _Build/linux \
             --target launcher page_manager_tests memory_tracker_tests \
-              audio_out2_port_tests ime_dialog_tests virtual_memory_allocation_tests \
+              audio_out2_port_tests pad_haptics_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests \
+              resource_tracking_tests scalar_provenance_tests \
+              resource_materialization_tests shader_vertex_metadata_tests archive_file_tests avplayer_file_tests \
             --parallel
 
       - name: Test
         shell: bash
         run: |
           ctest --test-dir _Build/linux --output-on-failure \
-            -R '^(audio_out2_port|ime_dialog|page_manager|memory_tracker|virtual_memory_allocation)$'
+            -R '^(audio_out2_port|pad_haptics|ime_dialog|lru_cache|memory_tracker|page_manager|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|archive_file|avplayer_file|virtual_memory_allocation)$'
 
       - name: Install
         shell: bash

3rdparty/CMakeLists.txt

@@ -2,10 +2,31 @@ set(BUILD_SHARED_LIBS OFF CACHE BOOL "" FORCE)
 
 add_subdirectory(cpuinfo)
 
-# The Windows guest red-zone patcher needs a decoder/encoder and a small
-# trampoline assembler. Pin the same revisions used by shadPS4's implementation.
 include(FetchContent)
 
+# The Bluetooth DualSense transports speaker audio as fixed-size Opus packets.
+set(kyty_build_testing_saved "${BUILD_TESTING}")
+set(BUILD_TESTING OFF)
+set(OPUS_BUILD_PROGRAMS OFF CACHE BOOL "" FORCE)
+set(OPUS_INSTALL_PKG_CONFIG_MODULE OFF CACHE BOOL "" FORCE)
+set(OPUS_INSTALL_CMAKE_CONFIG_MODULE OFF CACHE BOOL "" FORCE)
+if(MSVC AND CMAKE_C_COMPILER_ID STREQUAL "Clang")
+	# Opus assumes MSVC accepts SSE4.1 intrinsics without per-file flags, but
+	# clang-cl rejects them in its baseline x64 files. Its AVX2 dispatch also
+	# depends on the SSE4.1 declarations, so retain the x64 SSE2 baseline only.
+	set(OPUS_X86_MAY_HAVE_SSE4_1 OFF CACHE BOOL "" FORCE)
+	set(OPUS_X86_MAY_HAVE_AVX2 OFF CACHE BOOL "" FORCE)
+endif()
+FetchContent_Declare(opus
+	GIT_REPOSITORY https://github.com/xiph/opus.git
+	GIT_TAG ddbe48383984d56acd9e1ab6a090c54ca6b735a6
+	GIT_SHALLOW FALSE)
+FetchContent_MakeAvailable(opus)
+set(BUILD_TESTING "${kyty_build_testing_saved}")
+install(FILES "${opus_SOURCE_DIR}/COPYING" DESTINATION "licenses/opus")
+
+# The Windows guest red-zone 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)
 FetchContent_Declare(
 	xbyak
@@ -28,7 +49,54 @@ FetchContent_Declare(
 
 FetchContent_MakeAvailable(xbyak zydis)
 
+# ZArchive provides random-access reads for compressed game dumps. Build only
+# the reader/writer library and its zstd dependency; the command-line tool is
+# not part of KytyPS5's runtime distribution.
+set(ZSTD_BUILD_PROGRAMS OFF CACHE BOOL "" FORCE)
+set(ZSTD_BUILD_SHARED OFF CACHE BOOL "" FORCE)
+set(ZSTD_BUILD_STATIC ON CACHE BOOL "" FORCE)
+set(ZSTD_BUILD_TESTS OFF CACHE BOOL "" FORCE)
+set(ZSTD_LEGACY_SUPPORT OFF CACHE BOOL "" FORCE)
+set(ZSTD_MULTITHREAD_SUPPORT OFF CACHE BOOL "" FORCE)
+FetchContent_Declare(
+	zstd
+	GIT_REPOSITORY https://github.com/facebook/zstd.git
+	GIT_TAG f8745da6ff1ad1e7bab384bd1f9d742439278e99
+	GIT_SHALLOW FALSE
+	SOURCE_SUBDIR build/cmake
+)
+FetchContent_Declare(
+	zarchive_source
+	GIT_REPOSITORY https://github.com/Exzap/ZArchive.git
+	GIT_TAG d2c717730092c7bf8cbb033b12fd4001b7c4d932
+	GIT_SHALLOW FALSE
+	PATCH_COMMAND "${GIT_EXECUTABLE}" apply --ignore-whitespace
+		"${CMAKE_CURRENT_SOURCE_DIR}/patches/zarchive-reader.patch"
+	SOURCE_SUBDIR kyty-library-only
+)
+FetchContent_MakeAvailable(zstd zarchive_source)
+
+add_library(zarchive STATIC
+	"${zarchive_source_SOURCE_DIR}/src/zarchivereader.cpp"
+	"${zarchive_source_SOURCE_DIR}/src/zarchivewriter.cpp"
+	"${zarchive_source_SOURCE_DIR}/src/sha_256.c"
+)
+add_library(ZArchive::zarchive ALIAS zarchive)
+target_include_directories(zarchive SYSTEM PUBLIC "${zarchive_source_SOURCE_DIR}/include")
+target_include_directories(zarchive PRIVATE "${zstd_SOURCE_DIR}/lib")
+target_link_libraries(zarchive PRIVATE libzstd_static)
+target_compile_features(zarchive PRIVATE cxx_std_20)
+if(MSVC)
+	target_compile_options(zarchive PRIVATE /W0)
+else()
+	target_compile_options(zarchive PRIVATE -w)
+endif()
+
 set(SDL_SHARED OFF CACHE BOOL "" FORCE)
+if(APPLE)
+	# SDL's built-in iconv fallback does not support CP932.
+	set(SDL_SYSTEM_ICONV ON CACHE BOOL "Use system iconv for CP932 conversion" FORCE)
+endif()
 add_subdirectory(SDL3)
 
 add_library(Vulkan-Headers INTERFACE)
@@ -120,6 +188,7 @@ endif()
 
 set(TRACY_STATIC ON CACHE BOOL "" FORCE)
 set(TRACY_ENABLE ON CACHE BOOL "" FORCE)
+set(TRACY_NO_CRASH_HANDLER ON CACHE BOOL "" FORCE)
 set(TRACY_ON_DEMAND ON CACHE BOOL "" FORCE)
 set(TRACY_DELAYED_INIT ON CACHE BOOL "" FORCE)
 set(TRACY_MANUAL_LIFETIME ON CACHE BOOL "" FORCE)

3rdparty/patches/zarchive-reader.patch

@@ -0,0 +1,100 @@
+--- a/include/zarchive/zarchivecommon.h
++++ b/include/zarchive/zarchivecommon.h
+@@ -256,7 +256,7 @@
+ 
+ 			bool IsWithinValidRange(uint64_t fileSize) const
+ 			{
+-				return (offset + size) <= fileSize;
++				return offset <= fileSize && size <= fileSize - offset;
+ 			}
+ 		};
+ 
+--- a/src/zarchivereader.cpp
++++ b/src/zarchivereader.cpp
+@@ -14,6 +14,7 @@
+ 
+ static bool _ifstream_readBytes(std::ifstream& file, uint64_t offset, void* buffer, uint32_t size)
+ {
++	file.clear();
+ 	file.seekg(offset, std::ios_base::beg);
+ 	file.read((char*)buffer, size);
+ 	return file.gcount() == size;
+@@ -80,9 +81,27 @@
+ 	// verify file tree
+ 	if (fileTree[0].IsFile())
+ 		return nullptr; // first entry must be root directory
+-	auto rootName = GetName(nameTable, fileTree[0].GetNameOffset());
+-	if (!rootName.empty())
++	if (fileTree[0].GetNameOffset() != 0x7FFFFFFF)
+ 		return nullptr; // root node must not have a name
++	const uint64_t dataCapacity = (uint64_t)offsetRecords.size() * _ZARCHIVE::ENTRIES_PER_OFFSETRECORD * _ZARCHIVE::COMPRESSED_BLOCK_SIZE;
++	for (size_t i = 0; i < fileTree.size(); ++i)
++	{
++		auto& entry = fileTree[i];
++		if (i != 0)
++		{
++			auto name = GetName(nameTable, entry.GetNameOffset());
++			if (name.empty() || name == "." || name == ".." || name.find_first_of("/\\") != std::string_view::npos || name.find('\0') != std::string_view::npos)
++				return nullptr;
++		}
++		if (entry.IsFile())
++		{
++			if (entry.GetFileOffset() > dataCapacity || entry.GetFileSize() > dataCapacity - entry.GetFileOffset())
++				return nullptr;
++		}
++		else if (entry.directoryRecord.nodeStartIndex > fileTree.size() ||
++			entry.directoryRecord.count > fileTree.size() - entry.directoryRecord.nodeStartIndex)
++			return nullptr;
++	}
+ 	// read meta data
+ 	// todo
+ 
+@@ -238,7 +257,7 @@
+ 		uint32_t stepSize = std::min<uint32_t>(remainingBytes, _ZARCHIVE::COMPRESSED_BLOCK_SIZE - blockOffset);
+ 		CacheBlock* block = GetCachedBlock(blockIdx);
+ 		if (!block)
+-			return 0;
++			return bytesToRead - remainingBytes;
+ 		std::memcpy(bufferU8, block->data + blockOffset, stepSize);
+ 		rawReadOffset += stepSize;
+ 		remainingBytes -= stepSize;
+@@ -325,14 +344,18 @@
+ 	// determine offset and size of compressed block
+ 	auto& record = m_offsetRecords[recordIndex];
+ 	uint64_t offset = record.baseOffset;
++	if (offset > m_compressedDataSize)
++		return false;
+ 	for (uint32_t i = 0; i < recordSubIndex; i++)
+ 	{
+-		offset += (uint64_t)record.size[i];
+-		offset++;
++		const uint64_t size = (uint64_t)record.size[i] + 1;
++		if (size > m_compressedDataSize - offset)
++			return false;
++		offset += size;
+ 	}
+ 	uint32_t compressedSize = (uint32_t)record.size[recordSubIndex] + 1;
+ 	// load file data
+-	if ((offset + compressedSize) > m_compressedDataSize)
++	if (compressedSize > m_compressedDataSize - offset)
+ 		return false;
+ 	offset += m_compressedDataOffset;
+ 	if (compressedSize == _ZARCHIVE::COMPRESSED_BLOCK_SIZE)
+@@ -350,7 +373,7 @@
+ // returns empty view on failure
+ std::string_view ZArchiveReader::GetName(const std::vector<uint8_t>& nameTable, uint32_t nameOffset)
+ {
+-	if (nameOffset == 0x7FFFFFFF || nameOffset > nameTable.size())
++	if (nameOffset == 0x7FFFFFFF || nameOffset >= nameTable.size())
+ 		return "";
+ 	// parse header
+ 	uint16_t nameLength = nameTable[nameOffset] & 0x7F;
+@@ -359,7 +382,7 @@
+ 		// extended 2-byte length
+ 		if (nameOffset + 1 >= nameTable.size())
+ 			return "";
+-		nameLength |= ((uint16_t)nameTable[nameOffset] << 7);
++		nameLength |= ((uint16_t)nameTable[nameOffset + 1] << 7);
+ 		nameOffset += 2;
+ 	}
+ 	else

CMakeLists.txt

@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.16...4.3)
+cmake_minimum_required(VERSION 3.18...4.3)
 
 if(CMAKE_VERSION VERSION_GREATER_EQUAL 4.0)
 	set(CMAKE_POLICY_VERSION_MINIMUM 3.5 CACHE STRING "Minimum policy version for bundled dependencies")
@@ -301,6 +301,7 @@ set(gpu_blit_generated_dir "${PROJECT_BINARY_DIR}/gpu_blit_shaders")
 set(gpu_blit_shader_sources
 	"${gpu_tiler_shader_dir}/gpu_blit_fs_triangle.vert"
 	"${gpu_tiler_shader_dir}/gpu_blit_color_to_ms_depth.frag"
+	"${gpu_tiler_shader_dir}/gpu_video_out_overlay.frag"
 )
 foreach(shader_source IN LISTS gpu_blit_shader_sources)
 	get_filename_component(shader_name "${shader_source}" NAME_WE)
@@ -331,7 +332,7 @@ if (KYTY_CLANG_CL)
 	set_source_files_properties(${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)
+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)
 
 # Linux system libraries.
 if(LINUX)
@@ -406,11 +407,26 @@ add_executable(page_manager_tests EXCLUDE_FROM_ALL
 target_link_libraries(page_manager_tests common)
 target_include_directories(page_manager_tests PRIVATE ${inc_headers})
 
+add_executable(archive_file_tests EXCLUDE_FROM_ALL
+	"${KYTY_TESTS_DIR}/ArchiveFileTests.cpp"
+)
+target_link_libraries(archive_file_tests common ZArchive::zarchive)
+target_include_directories(archive_file_tests PRIVATE ${inc_headers})
+
+add_executable(avplayer_file_tests EXCLUDE_FROM_ALL "${KYTY_TESTS_DIR}/AvPlayerFileTests.cpp")
+target_link_libraries(avplayer_file_tests common FFmpeg::ffmpeg ZArchive::zarchive)
+target_include_directories(avplayer_file_tests PRIVATE ${inc_headers})
+
 add_executable(bit_array_tests EXCLUDE_FROM_ALL
 	"${KYTY_TESTS_DIR}/BitArrayTests.cpp"
 )
 target_include_directories(bit_array_tests PRIVATE ${inc_headers})
 
+add_executable(lru_cache_tests EXCLUDE_FROM_ALL
+	"${KYTY_TESTS_DIR}/LruCacheTests.cpp"
+)
+target_include_directories(lru_cache_tests PRIVATE ${inc_headers})
+
 add_executable(ime_dialog_tests EXCLUDE_FROM_ALL
 	"${KYTY_TESTS_DIR}/ImeDialogTests.cpp"
 	"${KYTY_SOURCE_DIR}/libs/imeCommon.cpp"
@@ -465,6 +481,10 @@ add_executable(audio_out2_port_tests EXCLUDE_FROM_ALL
 target_link_libraries(audio_out2_port_tests common fmt::fmt)
 target_include_directories(audio_out2_port_tests PRIVATE ${inc_headers})
 
+add_executable(audio_out_timing_tests EXCLUDE_FROM_ALL "${KYTY_TESTS_DIR}/AudioOutTimingTests.cpp")
+target_link_libraries(audio_out_timing_tests common fmt::fmt SDL3::SDL3-static)
+target_include_directories(audio_out_timing_tests PRIVATE ${inc_headers})
+
 add_executable(ngs2_sampler_tests EXCLUDE_FROM_ALL
 	"${KYTY_TESTS_DIR}/Ngs2SamplerTests.cpp"
 	"${KYTY_SOURCE_DIR}/loader/timer.cpp"
@@ -472,12 +492,36 @@ add_executable(ngs2_sampler_tests EXCLUDE_FROM_ALL
 target_link_libraries(ngs2_sampler_tests common fmt::fmt LibAtrac9)
 target_include_directories(ngs2_sampler_tests PRIVATE ${inc_headers})
 
+add_executable(pad_haptics_tests EXCLUDE_FROM_ALL
+	"${KYTY_TESTS_DIR}/PadHapticsTests.cpp"
+)
+target_link_libraries(pad_haptics_tests common fmt::fmt SDL3::SDL3-static opus)
+target_include_directories(pad_haptics_tests PRIVATE ${inc_headers})
+
 add_executable(save_data_memory_tests EXCLUDE_FROM_ALL
 	"${KYTY_TESTS_DIR}/SaveDataMemoryTests.cpp"
 )
 target_link_libraries(save_data_memory_tests common fmt::fmt)
 target_include_directories(save_data_memory_tests PRIVATE ${inc_headers})
 
+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 ${inc_headers})
+
+add_executable(ces_tests EXCLUDE_FROM_ALL "${KYTY_TESTS_DIR}/CesTests.cpp")
+target_link_libraries(ces_tests common fmt::fmt SDL3::SDL3-static)
+target_include_directories(ces_tests PRIVATE ${inc_headers})
+
+add_executable(save_data_dialog_tests EXCLUDE_FROM_ALL
+	"${KYTY_TESTS_DIR}/SaveDataDialogTests.cpp"
+	"${KYTY_SOURCE_DIR}/libs/dialog.cpp"
+)
+target_link_libraries(save_data_dialog_tests common fmt::fmt)
+target_include_directories(save_data_dialog_tests PRIVATE ${inc_headers})
+
+add_kyty_full_emulator_test(http_uri_parse_tests
+	"${KYTY_TESTS_DIR}/HttpUriParseTests.cpp")
+
 add_executable(event_queue_lifetime_tests EXCLUDE_FROM_ALL
 	"${KYTY_TESTS_DIR}/EventQueueLifetimeTests.cpp"
 	"${KYTY_SOURCE_DIR}/kernel/eventQueue.cpp"
@@ -488,6 +532,7 @@ target_include_directories(event_queue_lifetime_tests PRIVATE ${inc_headers})
 
 add_kyty_full_emulator_test(kernel_file_system_tests
 	"${KYTY_TESTS_DIR}/KernelFileSystemTests.cpp")
+target_link_libraries(kernel_file_system_tests ZArchive::zarchive)
 
 add_executable(sync_on_address_tests EXCLUDE_FROM_ALL
 	"${KYTY_TESTS_DIR}/SyncOnAddressTests.cpp"
@@ -554,7 +599,10 @@ 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 archive_file COMMAND $<TARGET_FILE:archive_file_tests>)
+	add_test(NAME avplayer_file COMMAND $<TARGET_FILE:avplayer_file_tests>)
 	add_test(NAME bit_array COMMAND $<TARGET_FILE:bit_array_tests>)
+	add_test(NAME lru_cache COMMAND $<TARGET_FILE:lru_cache_tests>)
 	add_test(NAME shader_vertex_metadata COMMAND $<TARGET_FILE:shader_vertex_metadata_tests>)
 	add_test(NAME resource_materialization COMMAND $<TARGET_FILE:resource_materialization_tests>)
 	add_test(NAME resource_tracking COMMAND $<TARGET_FILE:resource_tracking_tests>)
@@ -562,8 +610,14 @@ if(BUILD_TESTING)
 	add_test(NAME kernel_file_system COMMAND $<TARGET_FILE:kernel_file_system_tests>)
 	add_test(NAME sync_on_address COMMAND $<TARGET_FILE:sync_on_address_tests>)
 	add_test(NAME audio_out2_port COMMAND $<TARGET_FILE:audio_out2_port_tests>)
+	add_test(NAME audio_out_timing COMMAND $<TARGET_FILE:audio_out_timing_tests>)
 	add_test(NAME ngs2_sampler COMMAND $<TARGET_FILE:ngs2_sampler_tests>)
+	add_test(NAME pad_haptics COMMAND $<TARGET_FILE:pad_haptics_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 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>)
 	add_test(NAME shader_recompiler_compute COMMAND $<TARGET_FILE:shader_recompiler_compute_tests>)
 	add_test(NAME shader_recompiler_alignbyte
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --alignbyte-only)
@@ -600,6 +654,8 @@ if(BUILD_TESTING)
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --compute-meta-clear-only)
 	add_test(NAME buffer_cache_dirty_gc
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --buffer-cache-gc-only)
+	add_test(NAME color_render_target_1d
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --rt1d-only)
 
 	if(WIN32)
 		# These tests still depend on the Windows multisample-depth path.
@@ -618,7 +674,10 @@ if(BUILD_TESTING)
 		image_page_table_tests
 		memory_tracker_tests
 		page_manager_tests
+		archive_file_tests
+		avplayer_file_tests
 		bit_array_tests
+		lru_cache_tests
 		shader_vertex_metadata_tests
 		resource_materialization_tests
 		resource_tracking_tests
@@ -628,7 +687,14 @@ if(BUILD_TESTING)
 		shader_recompiler_compute_tests
 		virtual_memory_allocation_tests
 		audio_out2_port_tests
+		audio_out_timing_tests
 		ngs2_sampler_tests
+		pad_haptics_tests
+		save_data_memory_tests
+		png_enc_tests
+		ces_tests
+		save_data_dialog_tests
+		http_uri_parse_tests
 	)
 endif()
 
@@ -713,6 +779,10 @@ endif()
 install(TARGETS kyty_emulator DESTINATION .)
 if(KYTY_BUILD_LAUNCHER)
 	add_subdirectory("${KYTY_SOURCE_DIR}/launcher" "${CMAKE_CURRENT_BINARY_DIR}/launcher")
+
+	if(LINUX)
+		target_link_options(launcher PRIVATE $<$<CONFIG:Release>:-Wl,--strip-all>)
+	endif()
 endif()
 if(KYTY_CLANG_CL)
 	install(FILES "${KYTY_THIRD_PARTY_DIR}/winpthread/bin/libwinpthread-1.dll" DESTINATION .)

README.md

@@ -7,6 +7,10 @@
 [![Status](https://img.shields.io/badge/status-active%20development-orange.svg)](#current-status)
 [![License](https://img.shields.io/badge/license-GPL--2.0-blue.svg)](LICENSE)
 
+**[Weekly updates](https://github.com/KytyPS5/KytyPS5/discussions/862)** — game progress, recent fixes and ongoing development.
+
+**[Development on Discord](https://discord.gg/UNrkMqGaBg)** — KytyPS5 development.
+
 KytyPS5 is a free and open-source PlayStation 5 emulator written in C++ for Windows and Linux,
 with experimental macOS support. It is based on a heavily modified version of
 [Kyty](https://github.com/InoriRus/Kyty). The project is in active development, and behavior
@@ -68,8 +72,8 @@ or graphical glitches, so please include the version you tested when reporting a
       <img src="docs/screenshots/ps5-02.png" width="300" alt="Demon's Souls running in KytyPS5">
     </td>
     <td align="center">
-      <strong>Hellboy</strong><br>
-      <img src="docs/screenshots/ps5-06.png" width="300" alt="Hellboy running in KytyPS5">
+      <strong>UFC 5</strong><br>
+      <img src="docs/screenshots/ps5-06.png" width="300" alt="UFC 5 running in KytyPS5">
     </td>
   </tr>
 </table>
@@ -232,9 +236,10 @@ cmake --build _Build/linux --target launcher --parallel
 cmake --install _Build/linux --prefix _Build/linux/install
 ```
 
-The configure step downloads the FFmpeg prebuilts and the `xbyak`/`zydis` sources, so it needs
-network access; a fully sandboxed `nix build` would require vendoring those inputs. A Vulkan 1.3
-driver must be available at runtime (on NixOS, `hardware.graphics.enable = true`).
+The configure step downloads the FFmpeg prebuilts and the `xbyak`, `zydis`, `zstd`, and ZArchive
+sources, so it needs network access; a fully sandboxed `nix build` would require vendoring those
+inputs. A Vulkan 1.3 driver must be available at runtime (on NixOS,
+`hardware.graphics.enable = true`).
 
 ### Building on macOS
 
@@ -328,17 +333,21 @@ open _Build/macos/install/KytyPS5.app  # or double-click in Finder
 ```
 
 On first launch, add one or more game folders in the global settings. The launcher searches those
-folders recursively for game directories containing `eboot.bin`. Select a detected game and run it
-from the game list.
+folders recursively for game directories containing `eboot.bin` and ZArchive (`.zar`) game dumps
+whose archive root contains `eboot.bin`. Select a detected game and run it from the game list.
+ZArchive dumps are mounted read-only and streamed directly; they do not need to be extracted first.
 
-The emulator can also be started directly with a legally obtained game directory or ELF file:
+The emulator can also be started directly with a legally obtained game directory, ELF file, or
+ZArchive dump:
 
 ```powershell
 .\_Build\windows\install\kyty_emulator.exe --game "D:\Games\ExampleGame"
+.\_Build\windows\install\kyty_emulator.exe --game "D:\Games\ExampleGame.zar"
 ```
 
 ```bash
 ./_Build/linux/install/kyty_emulator --game "/games/ExampleGame"
+./_Build/linux/install/kyty_emulator --game "/games/ExampleGame.zar"
 ```
 
 On macOS, the adjacent flat or app-bundled `libMoltenVK.dylib` is found automatically; no

src/common/CMakeLists.txt

@@ -26,6 +26,7 @@ target_link_libraries(common
 		cpuinfo
 		kyty_git_version
 		spdlog::spdlog
+		ZArchive::zarchive
 )
 
 get_property(inc_headers TARGET common PROPERTY INCLUDE_DIRECTORIES)

src/common/archive.cpp

@@ -0,0 +1,397 @@
+#include "common/archive.h"
+
+#include "common/stringUtils.h"
+
+#include <algorithm>
+#include <condition_variable>
+#include <cstring>
+#include <mutex>
+#include <optional>
+#include <string_view>
+#include <thread>
+#include <unordered_map>
+#include <utility>
+#include <zarchive/zarchivereader.h>
+
+// POSIX hosts and MinGW's pthread-backed libstdc++ have cancellable condition waits.
+// MSVC/ClangCL and MinGW's native Win32 threading backend do not use pthread waits.
+#if !defined(_WIN32) || defined(_GLIBCXX_GCC_GTHR_POSIX_H)
+#include <pthread.h>
+#define KYTY_ARCHIVE_PTHREAD_CANCELLATION
+#endif
+
+namespace Common {
+
+// Backends expose immutable metadata and offset-based reads. Cursor management, path routing,
+// reader lifetime and the host-stack worker are shared by every archive format.
+class ArchiveReader {
+public:
+	struct Entry {
+		uint64_t id;
+		uint64_t size;
+		bool     is_file;
+	};
+
+	virtual ~ArchiveReader()                                                       = default;
+	virtual std::optional<Entry>        Find(std::string_view member)              = 0;
+	virtual std::vector<File::DirEntry> List(std::string_view member)              = 0;
+	virtual uint64_t Read(uint64_t id, uint64_t offset, uint32_t size, void* data) = 0;
+};
+
+namespace {
+
+class ZArchiveReaderBackend final: public ArchiveReader {
+public:
+	explicit ZArchiveReaderBackend(std::unique_ptr<ZArchiveReader> reader)
+	    : m_reader(std::move(reader)) {}
+
+	std::optional<Entry> Find(std::string_view member) override {
+		const auto node = m_reader->LookUp(member);
+		if (node == ZARCHIVE_INVALID_NODE) {
+			return {};
+		}
+		return Entry {node, m_reader->GetFileSize(node), m_reader->IsFile(node)};
+	}
+
+	std::vector<File::DirEntry> List(std::string_view member) override {
+		const auto node = m_reader->LookUp(member);
+		if (!m_reader->IsDirectory(node)) {
+			return {};
+		}
+		const auto count = m_reader->GetDirEntryCount(node);
+		if (count > (1u << 20u)) {
+			return {};
+		}
+		std::vector<File::DirEntry> result;
+		result.reserve(count);
+		for (uint32_t index = 0; index < count; ++index) {
+			ZArchiveReader::DirEntry entry {};
+			if (m_reader->GetDirEntry(node, index, entry)) {
+				result.push_back({std::string(entry.name), entry.isFile});
+			}
+		}
+		return result;
+	}
+
+	uint64_t Read(uint64_t id, uint64_t offset, uint32_t size, void* data) override {
+		// ZArchiveReader already serializes access to its stream and decompression cache.
+		return m_reader->ReadFromFile(static_cast<ZArchiveNodeHandle>(id), offset, size, data);
+	}
+
+private:
+	std::unique_ptr<ZArchiveReader> m_reader;
+};
+
+std::shared_ptr<ArchiveReader> OpenZArchive(const std::filesystem::path& path) {
+	auto reader = std::unique_ptr<ZArchiveReader>(ZArchiveReader::OpenFromFile(path));
+	return reader ? std::make_shared<ZArchiveReaderBackend>(std::move(reader)) : nullptr;
+}
+
+struct ArchiveFormat {
+	std::string_view extension;
+	std::shared_ptr<ArchiveReader> (*open)(const std::filesystem::path&);
+};
+
+constexpr ArchiveFormat Formats[] = {{".zar", OpenZArchive}};
+using NativeView                  = std::basic_string_view<std::filesystem::path::value_type>;
+
+const ArchiveFormat* FindFormat(NativeView path) {
+	for (const auto& format: Formats) {
+		if (path.size() < format.extension.size()) {
+			continue;
+		}
+		const auto extension = path.substr(path.size() - format.extension.size());
+		if (std::equal(
+		        extension.begin(), extension.end(), format.extension.begin(),
+		        [](auto a, char b) { return (a >= 'A' && a <= 'Z' ? a + ('a' - 'A') : a) == b; })) {
+			return &format;
+		}
+	}
+	return nullptr;
+}
+
+size_t FindMarker(NativeView path) {
+	// Native files use this check on every operation. The usual path has no marker and allocates
+	// nothing.
+	for (auto marker = path.find('!'); marker != NativeView::npos;
+	     marker      = path.find('!', marker + 1)) {
+		if ((marker + 1 == path.size() || path[marker + 1] == '/' || path[marker + 1] == '\\') &&
+		    FindFormat(path.substr(0, marker)) != nullptr) {
+			return marker;
+		}
+	}
+	return NativeView::npos;
+}
+
+struct ParsedPath {
+	std::filesystem::path archive;
+	std::string           member;
+};
+
+std::optional<ParsedPath> ParsePath(const std::filesystem::path& path) {
+	const auto& native = path.native();
+	const auto  marker = FindMarker(native);
+	if (marker == NativeView::npos) {
+		return {};
+	}
+	ParsedPath parsed {.archive = std::filesystem::path(native.substr(0, marker))};
+	auto       member = PathToGenericString(std::filesystem::path(native.substr(marker + 1)));
+	std::replace(member.begin(), member.end(), '\\', '/');
+	std::string_view remaining = member;
+	while (!remaining.empty()) {
+		const auto separator = remaining.find('/');
+		const auto component = remaining.substr(0, separator);
+		remaining.remove_prefix(separator == std::string_view::npos ? remaining.size() : separator + 1);
+		if (component.empty() || component == ".") {
+			continue;
+		}
+		if (component == "..") {
+			if (parsed.member.empty()) {
+				return {};
+			}
+			const auto separator = parsed.member.rfind('/');
+			parsed.member.resize(separator == std::string::npos ? 0 : separator);
+		} else {
+			if (!parsed.member.empty()) {
+				parsed.member += '/';
+			}
+			parsed.member += component;
+		}
+	}
+	return parsed;
+}
+
+std::mutex                                                              g_readers_mutex;
+std::unordered_map<std::filesystem::path, std::weak_ptr<ArchiveReader>> g_readers;
+
+// Guest stacks can be only 64 KiB. Decompress on an ordinary host stack, then copy into guest
+// memory on the caller so GPU-tracked writes use its fault handler. A single worker bounds host
+// threads and avoids duplicating the archive's own block cache. Requests live on waiting callers.
+class ArchiveIoThread {
+public:
+	static ArchiveIoThread& Instance() {
+		// Guest exit can run static destructors while other guest threads are still reading.
+		// Keep the worker alive until process termination, as with the guest threads themselves.
+		static auto* instance = new ArchiveIoThread;
+		return *instance;
+	}
+
+	uint64_t Read(ArchiveReader& reader, uint64_t id, uint64_t offset, uint32_t size, void* data) {
+		Request          request {reader, id, offset, size, data};
+		std::unique_lock lock(m_mutex);
+		if (m_last != nullptr) {
+			m_last->next = &request;
+		} else {
+			m_first = &request;
+		}
+		m_last = &request;
+		m_wake.notify_one();
+		request.ready.wait(lock, [&] { return request.done; });
+		return request.result;
+	}
+
+	KYTY_CLASS_NO_COPY(ArchiveIoThread);
+
+private:
+	struct Request {
+		ArchiveReader&          reader;
+		uint64_t                id;
+		uint64_t                offset;
+		uint32_t                size;
+		void*                   data;
+		Request*                next   = nullptr;
+		uint64_t                result = 0;
+		bool                    done   = false;
+		std::condition_variable ready {};
+	};
+
+	ArchiveIoThread() {
+		std::thread([this] { Loop(); }).detach();
+	}
+
+	void Loop() {
+		for (;;) {
+			Request* request;
+			{
+				std::unique_lock lock(m_mutex);
+				m_wake.wait(lock, [this] { return m_first != nullptr; });
+				request = m_first;
+				m_first = request->next;
+				if (m_first == nullptr) {
+					m_last = nullptr;
+				}
+			}
+			const auto result =
+			    request->reader.Read(request->id, request->offset, request->size, request->data);
+			{
+				std::lock_guard lock(m_mutex);
+				request->result = result;
+				request->done   = true;
+				// Notify before releasing the mutex: the waiting caller owns the condition
+				// variable.
+				request->ready.notify_one();
+			}
+		}
+	}
+
+	std::mutex              m_mutex;
+	std::condition_variable m_wake;
+	Request*                m_first = nullptr;
+	Request*                m_last  = nullptr;
+};
+
+} // namespace
+
+struct ArchiveFile::Private {
+	std::shared_ptr<ArchiveReader> reader;
+	ArchiveReader::Entry           entry;
+	uint64_t                       position = 0;
+};
+
+ArchiveFile::ArchiveFile(std::unique_ptr<Private> p): m_p(std::move(p)) {}
+ArchiveFile::~ArchiveFile() = default;
+
+uint64_t ArchiveFile::Size() const {
+	return m_p->entry.size;
+}
+uint64_t ArchiveFile::Tell() const {
+	return m_p->position;
+}
+bool ArchiveFile::Seek(uint64_t offset) {
+	m_p->position = offset;
+	return true;
+}
+
+void ArchiveFile::Read(void* data, uint32_t size, uint32_t* bytes_read) {
+#if defined(KYTY_ARCHIVE_PTHREAD_CANCELLATION)
+	// A pending cancellation must not destroy the waiting request or its TLS buffer while the
+	// worker still uses them. Restore the caller's state only after the copy is complete.
+	int cancellation_state = PTHREAD_CANCEL_ENABLE;
+	pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &cancellation_state);
+#endif
+	constexpr uint32_t                chunk_size = 1u << 20u;
+	thread_local std::vector<uint8_t> chunk;
+	uint32_t                          read = 0;
+	if (data != nullptr && size != 0 && m_p->position < Size()) {
+		const auto requested =
+		    static_cast<uint32_t>(std::min<uint64_t>(size, Size() - m_p->position));
+		if (chunk.size() < std::min(chunk_size, requested)) {
+			chunk.resize(std::min(chunk_size, requested));
+		}
+		while (read < requested) {
+			const auto step = std::min(chunk_size, requested - read);
+			const auto got  = ArchiveIoThread::Instance().Read(*m_p->reader, m_p->entry.id,
+			                                                   m_p->position, step, chunk.data());
+			if (got == 0 || got > step) {
+				break;
+			}
+			std::memcpy(static_cast<uint8_t*>(data) + read, chunk.data(), got);
+			read += static_cast<uint32_t>(got);
+			m_p->position += got;
+			if (got < step) {
+				break;
+			}
+		}
+	}
+	if (bytes_read != nullptr) {
+		*bytes_read = read;
+	}
+#if defined(KYTY_ARCHIVE_PTHREAD_CANCELLATION)
+	pthread_setcancelstate(cancellation_state, nullptr);
+#endif
+}
+
+std::filesystem::path MakeArchivePath(const std::filesystem::path& archive,
+                                      const std::filesystem::path& member) {
+	auto root = archive;
+	root += "!";
+	return member.empty() ? root : root / member.relative_path();
+}
+
+bool IsSupportedArchive(const std::filesystem::path& path) {
+	return FindFormat(path.native()) != nullptr;
+}
+
+bool IsArchivePath(const std::filesystem::path& path) {
+	return FindMarker(path.native()) != NativeView::npos;
+}
+
+std::filesystem::path GetArchiveHostPath(const std::filesystem::path& path) {
+	const auto marker = FindMarker(path.native());
+	return marker == NativeView::npos ? std::filesystem::path {}
+	                                  : std::filesystem::path(path.native().substr(0, marker));
+}
+
+std::shared_ptr<ArchiveReader> OpenArchive(const std::filesystem::path& path) {
+	const auto  archive = IsArchivePath(path) ? GetArchiveHostPath(path) : path;
+	const auto* format  = FindFormat(archive.native());
+	if (format == nullptr) {
+		return {};
+	}
+	std::error_code error;
+	auto            key = std::filesystem::absolute(archive, error);
+	if (error) {
+		return {};
+	}
+	{
+		std::lock_guard lock(g_readers_mutex);
+		if (auto it = g_readers.find(key); it != g_readers.end()) {
+			if (auto reader = it->second.lock()) {
+				return reader;
+			}
+		}
+	}
+	// Do not hold the cache mutex while opening an unrelated archive from disk.
+	auto reader = format->open(archive);
+	if (reader == nullptr) {
+		return {};
+	}
+	std::lock_guard lock(g_readers_mutex);
+	std::erase_if(g_readers, [](const auto& entry) { return entry.second.expired(); });
+	auto& cached = g_readers[key];
+	if (auto existing = cached.lock()) {
+		return existing;
+	}
+	cached = reader;
+	return reader;
+}
+
+std::optional<File::Info> GetArchiveInfo(const std::filesystem::path& path) {
+	const auto parsed = ParsePath(path);
+	if (parsed) {
+		if (auto reader = OpenArchive(parsed->archive)) {
+			if (auto entry = reader->Find(parsed->member)) {
+				return File::Info {entry->is_file, entry->size};
+			}
+		}
+	}
+	return {};
+}
+
+std::vector<File::DirEntry> GetArchiveDirEntries(const std::filesystem::path& path) {
+	const auto parsed = ParsePath(path);
+	if (parsed) {
+		if (auto reader = OpenArchive(parsed->archive)) {
+			return reader->List(parsed->member);
+		}
+	}
+	return {};
+}
+
+std::unique_ptr<ArchiveFile> OpenArchiveFile(const std::filesystem::path& path) {
+	auto parsed = ParsePath(path);
+	if (!parsed) {
+		return {};
+	}
+	auto       reader = OpenArchive(parsed->archive);
+	const auto entry  = reader ? reader->Find(parsed->member) : std::nullopt;
+	if (!entry || !entry->is_file) {
+		return {};
+	}
+	auto p     = std::make_unique<ArchiveFile::Private>();
+	p->reader  = std::move(reader);
+	p->entry   = *entry;
+	return std::unique_ptr<ArchiveFile>(new ArchiveFile(std::move(p)));
+}
+
+} // namespace Common

src/common/archive.h

@@ -0,0 +1,50 @@
+#ifndef KYTY_COMMON_ARCHIVE_H_
+#define KYTY_COMMON_ARCHIVE_H_
+
+#include "common/file.h"
+
+#include <cstdint>
+#include <filesystem>
+#include <memory>
+#include <vector>
+
+namespace Common {
+
+class ArchiveReader;
+
+class ArchiveFile {
+public:
+	~ArchiveFile();
+
+	[[nodiscard]] uint64_t Size() const;
+	[[nodiscard]] uint64_t Tell() const;
+	bool                   Seek(uint64_t offset);
+	void                   Read(void* data, uint32_t size, uint32_t* bytes_read);
+
+	KYTY_CLASS_NO_COPY(ArchiveFile);
+
+private:
+	struct Private;
+	explicit ArchiveFile(std::unique_ptr<Private> p);
+	std::unique_ptr<Private> m_p;
+
+	friend std::unique_ptr<ArchiveFile> OpenArchiveFile(const std::filesystem::path& path);
+};
+
+// Virtual paths use "archive.ext!/member" and retain native host-path encoding.
+[[nodiscard]] std::filesystem::path       MakeArchivePath(const std::filesystem::path& archive,
+                                                          const std::filesystem::path& member = {});
+[[nodiscard]] bool                        IsSupportedArchive(const std::filesystem::path& path);
+[[nodiscard]] bool                        IsArchivePath(const std::filesystem::path& path);
+[[nodiscard]] std::optional<File::Info>   GetArchiveInfo(const std::filesystem::path& path);
+[[nodiscard]] std::vector<File::DirEntry> GetArchiveDirEntries(const std::filesystem::path& path);
+[[nodiscard]] std::filesystem::path       GetArchiveHostPath(const std::filesystem::path& path);
+[[nodiscard]] std::unique_ptr<ArchiveFile> OpenArchiveFile(const std::filesystem::path& path);
+
+// Retain this handle while a mount or metadata scan is active to reuse the index and block cache.
+// Accepts either a host archive filename or a virtual path inside the archive.
+[[nodiscard]] std::shared_ptr<ArchiveReader> OpenArchive(const std::filesystem::path& path);
+
+} // namespace Common
+
+#endif /* KYTY_COMMON_ARCHIVE_H_ */

src/common/emulatorConfig.cpp

@@ -47,6 +47,10 @@ const std::string& GetAudioInputDevice() {
 	return g_config->audio_input_device;
 }
 
+const std::optional<ControllerColor>& GetControllerColor() {
+	return g_config->controller_color;
+}
+
 PresentMode GetPresentMode() {
 	return g_config->present_mode;
 }

src/common/emulatorConfig.h

@@ -4,7 +4,9 @@
 #include "common/common.h"
 
 #include <cstddef>
+#include <array>
 #include <filesystem>
+#include <optional>
 #include <string>
 #include <vector>
 
@@ -26,6 +28,7 @@ enum class LogDirection { Silent, Console, File };
 enum class PresentMode { Fifo, Mailbox, Immediate };
 
 using Keymap = std::vector<std::string>;
+using ControllerColor = std::array<uint8_t, 3>;
 
 constexpr uint32_t DEFAULT_CONSOLE_LANGUAGE = 1;
 constexpr uint32_t MAX_CONSOLE_LANGUAGE     = 29;
@@ -44,6 +47,7 @@ struct ConfigOptions {
 	std::string            user_name                   = "Kyty";
 	int32_t                user_id                     = DEFAULT_USER_ID;
 	std::string            audio_input_device;
+	std::optional<ControllerColor> controller_color;
 	PresentMode            present_mode                = PresentMode::Mailbox;
 	int32_t                gpu_index                   = -1;
 	bool                   fullscreen_enabled          = false;
@@ -81,6 +85,7 @@ uint32_t GetScreenHeight();
 const std::string& GetUserName();
 int32_t  GetUserId();
 const std::string& GetAudioInputDevice();
+const std::optional<ControllerColor>& GetControllerColor();
 PresentMode GetPresentMode();
 int32_t GetGpuIndex();
 bool     FullscreenEnabled();

src/common/file.cpp

@@ -1,5 +1,6 @@
 #include "common/file.h"
 
+#include "common/archive.h"
 #include "common/assert.h"
 #include "common/common.h"
 #include "common/platform/sysFileIO.h"
@@ -61,7 +62,8 @@ SysFileTimeStruct DateTimeToFileTimeUtc(const DateTime& date_time) {
 } // namespace
 
 struct File::FilePrivate {
-	sys_file_t* f;
+	sys_file_t*                  f;
+	std::unique_ptr<ArchiveFile> archive;
 };
 
 File::File(): m_p(std::make_unique<FilePrivate>()) {
@@ -85,11 +87,15 @@ File::File(const std::filesystem::path& name, Mode mode): m_p(std::make_unique<F
 }
 
 bool File::IsInvalid() const {
-	return m_p->f == nullptr;
+	return m_p->f == nullptr && m_p->archive == nullptr;
 }
 
 bool File::Create(const std::filesystem::path& name) {
-	EXIT_IF(m_p->f != nullptr);
+	EXIT_IF(!IsInvalid());
+
+	if (IsArchivePath(name)) {
+		return false;
+	}
 
 	m_file_name = name;
 
@@ -104,9 +110,16 @@ bool File::Create(const std::filesystem::path& name) {
 }
 
 bool File::Open(const std::filesystem::path& name, Mode mode) {
-	EXIT_IF(m_p->f != nullptr);
+	EXIT_IF(!IsInvalid());
 
 	m_file_name = name;
+	if (IsArchivePath(name)) {
+		if (mode != Mode::Read) {
+			return false;
+		}
+		m_p->archive = OpenArchiveFile(name);
+		return m_p->archive != nullptr;
+	}
 
 	switch (mode) {
 		case Mode::Read: m_p->f = SysFileOpenR(name); break;
@@ -124,7 +137,7 @@ bool File::Open(const std::filesystem::path& name, Mode mode) {
 }
 
 bool File::OpenInMem(void* buf, uint32_t buf_size) {
-	EXIT_IF(m_p->f != nullptr);
+	EXIT_IF(!IsInvalid());
 
 	m_p->f = SysFileOpen(static_cast<uint8_t*>(buf), buf_size);
 
@@ -137,7 +150,7 @@ bool File::OpenInMem(void* buf, uint32_t buf_size) {
 }
 
 bool File::CreateInMem() {
-	EXIT_IF(m_p->f != nullptr);
+	EXIT_IF(!IsInvalid());
 
 	m_p->f = SysFileCreate();
 
@@ -150,6 +163,7 @@ bool File::CreateInMem() {
 }
 
 void File::Close() {
+	m_p->archive.reset();
 	if (m_p->f != nullptr) {
 		SysFileClose(m_p->f);
 		m_p->f = nullptr;
@@ -157,7 +171,11 @@ void File::Close() {
 }
 
 uint64_t File::Size() const {
-	EXIT_IF(m_p->f == nullptr);
+	EXIT_IF(IsInvalid());
+
+	if (m_p->archive != nullptr) {
+		return m_p->archive->Size();
+	}
 
 	if (m_p->f == nullptr) {
 		return 0;
@@ -167,41 +185,76 @@ uint64_t File::Size() const {
 }
 
 uint64_t File::Remaining() const {
-	EXIT_IF(m_p->f == nullptr);
+	EXIT_IF(IsInvalid());
 
-	return Size() - Tell();
+	const auto size     = Size();
+	const auto position = Tell();
+	return position < size ? size - position : 0;
 }
 
 uint64_t File::Size(const std::filesystem::path& name) {
+	if (IsArchivePath(name)) {
+		const auto info = GetArchiveInfo(name);
+		return info ? info->size : 0;
+	}
 	return SysFileSize(name);
 }
 
+std::optional<File::Info> File::GetInfo(const std::filesystem::path& name) {
+	if (IsArchivePath(name)) {
+		return GetArchiveInfo(name);
+	}
+	Info info {};
+	return SysFileGetInfo(name, &info.is_file, &info.size) ? std::optional(info) : std::nullopt;
+}
+
 bool File::Seek(uint64_t offset) {
-	EXIT_IF(m_p->f == nullptr);
+	EXIT_IF(IsInvalid());
+
+	if (m_p->archive != nullptr) {
+		return m_p->archive->Seek(offset);
+	}
 
 	return SysFileSeek(*m_p->f, offset);
 }
 
 bool File::Truncate(uint64_t size) {
-	EXIT_IF(m_p->f == nullptr);
+	EXIT_IF(IsInvalid());
+
+	if (m_p->archive != nullptr) {
+		return false;
+	}
 
 	return SysFileTruncate(*m_p->f, size);
 }
 
 bool File::Unlink() {
-	EXIT_IF(m_p->f == nullptr);
+	EXIT_IF(IsInvalid());
+
+	if (m_p->archive != nullptr) {
+		return false;
+	}
 
 	return SysFileUnlink(*m_p->f, m_file_name);
 }
 
 uint64_t File::Tell() const {
-	EXIT_IF(m_p->f == nullptr);
+	EXIT_IF(IsInvalid());
+
+	if (m_p->archive != nullptr) {
+		return m_p->archive->Tell();
+	}
 
 	return SysFileTell(*m_p->f);
 }
 
 void File::Read(void* data, uint32_t size, uint32_t* bytes_read) {
-	EXIT_IF(m_p->f == nullptr);
+	EXIT_IF(IsInvalid());
+
+	if (m_p->archive != nullptr) {
+		m_p->archive->Read(data, size, bytes_read);
+		return;
+	}
 
 	if (m_p->f != nullptr) {
 		SysFileRead(data, size, *m_p->f, bytes_read);
@@ -209,7 +262,14 @@ void File::Read(void* data, uint32_t size, uint32_t* bytes_read) {
 }
 
 void File::Write(const void* data, uint32_t size, uint32_t* bytes_written) {
-	EXIT_IF(m_p->f == nullptr);
+	EXIT_IF(IsInvalid());
+
+	if (m_p->archive != nullptr) {
+		if (bytes_written != nullptr) {
+			*bytes_written = 0;
+		}
+		return;
+	}
 
 	SysFileWrite(data, size, *m_p->f, bytes_written);
 }
@@ -257,24 +317,41 @@ static std::filesystem::path WithoutTrailingSeparator(
 
 
 bool File::IsDirectoryExisting(const std::filesystem::path& path) {
+	if (IsArchivePath(path)) {
+		const auto info = GetArchiveInfo(path);
+		return info && !info->is_file;
+	}
 	return SysFileIsDirectoryExisting(WithoutTrailingSeparator(path));
 }
 
 bool File::IsFileExisting(const std::filesystem::path& name) {
+	if (IsArchivePath(name)) {
+		const auto info = GetArchiveInfo(name);
+		return info && info->is_file;
+	}
 	return SysFileIsFileExisting(name);
 }
 
 bool File::CreateDirectory(const std::filesystem::path& path) {
+	if (IsArchivePath(path)) {
+		return false;
+	}
     return SysFileCreateDirectory(
         WithoutTrailingSeparator(path));
 }
 
 bool File::DeleteDirectory(const std::filesystem::path& path) {
+	if (IsArchivePath(path)) {
+		return false;
+	}
     return SysFileDeleteDirectory(
         WithoutTrailingSeparator(path));
 }
 
 bool File::CreateDirectories(const std::filesystem::path& path) {
+	if (IsArchivePath(path)) {
+		return false;
+	}
     const auto normalized_path =
         WithoutTrailingSeparator(path);
 
@@ -297,6 +374,9 @@ bool File::CreateDirectories(const std::filesystem::path& path) {
 }
 
 bool File::DeleteDirectories(const std::filesystem::path& path) {
+	if (IsArchivePath(path)) {
+		return false;
+	}
     const auto normalized_path =
         WithoutTrailingSeparator(path);
 
@@ -322,11 +402,18 @@ bool File::DeleteDirectories(const std::filesystem::path& path) {
 bool File::DeleteFile(
     const std::filesystem::path& name) // @suppress("Member declaration not found")
 {
+	if (IsArchivePath(name)) {
+		return false;
+	}
 	return SysFileDeleteFile(name);
 }
 
 bool File::Flush() {
-	EXIT_IF(m_p->f == nullptr);
+	EXIT_IF(IsInvalid());
+
+	if (m_p->archive != nullptr) {
+		return true;
+	}
 
 	return SysFileFlush(*m_p->f);
 }
@@ -348,6 +435,9 @@ std::vector<std::byte> File::ReadWholeBuffer() {
 }
 
 DateTime File::GetLastAccessTimeUTC(const std::filesystem::path& name) {
+	if (IsArchivePath(name)) {
+		return GetLastAccessTimeUTC(GetArchiveHostPath(name));
+	}
 	SysTimeStruct t {};
 	SysFileToSystemTimeUtc(SysFileGetLastAccessTimeUtc(name), t);
 
@@ -359,6 +449,9 @@ DateTime File::GetLastAccessTimeUTC(const std::filesystem::path& name) {
 }
 
 DateTime File::GetLastWriteTimeUTC(const std::filesystem::path& name) {
+	if (IsArchivePath(name)) {
+		return GetLastWriteTimeUTC(GetArchiveHostPath(name));
+	}
 	SysTimeStruct t {};
 	SysFileToSystemTimeUtc(SysFileGetLastWriteTimeUtc(name), t);
 
@@ -373,6 +466,10 @@ void File::GetLastAccessAndWriteTimeUTC(const std::filesystem::path& name, DateT
                                         DateTime* write) {
 	EXIT_IF(access == nullptr);
 	EXIT_IF(write == nullptr);
+	if (IsArchivePath(name)) {
+		GetLastAccessAndWriteTimeUTC(GetArchiveHostPath(name), access, write);
+		return;
+	}
 
 	SysTimeStruct     at {};
 	SysTimeStruct     wt {};
@@ -399,7 +496,12 @@ void File::GetLastAccessAndWriteTimeUTC(DateTime* access, DateTime* write) {
 	EXIT_IF(access == nullptr);
 	EXIT_IF(write == nullptr);
 
-	EXIT_IF(m_p->f == nullptr);
+	EXIT_IF(IsInvalid());
+
+	if (m_p->archive != nullptr) {
+		GetLastAccessAndWriteTimeUTC(m_file_name, access, write);
+		return;
+	}
 
 	if (m_p->f == nullptr) {
 		*access = DateTime();
@@ -429,23 +531,35 @@ void File::GetLastAccessAndWriteTimeUTC(DateTime* access, DateTime* write) {
 }
 
 bool File::SetLastAccessTimeUTC(const std::filesystem::path& name, const DateTime& dt) {
+	if (IsArchivePath(name)) {
+		return false;
+	}
 	auto f = DateTimeToFileTimeUtc(dt);
 	return SysFileSetLastAccessTimeUtc(name, f);
 }
 
 bool File::SetLastWriteTimeUTC(const std::filesystem::path& name, const DateTime& dt) {
+	if (IsArchivePath(name)) {
+		return false;
+	}
 	auto f = DateTimeToFileTimeUtc(dt);
 	return SysFileSetLastWriteTimeUtc(name, f);
 }
 
 bool File::SetLastAccessAndWriteTimeUTC(const std::filesystem::path& name, const DateTime& access,
                                         const DateTime& write) {
+	if (IsArchivePath(name)) {
+		return false;
+	}
 	auto af = DateTimeToFileTimeUtc(access);
 	auto wf = DateTimeToFileTimeUtc(write);
 	return SysFileSetLastAccessAndWriteTimeUtc(name, af, wf);
 }
 
 std::vector<File::DirEntry> File::GetDirEntries(const std::filesystem::path& path) {
+	if (IsArchivePath(path)) {
+		return GetArchiveDirEntries(path);
+	}
 	std::vector<sys_dir_entry_t> files;
 
 	SysFileGetDents(path, files);
@@ -463,12 +577,18 @@ std::vector<File::DirEntry> File::GetDirEntries(const std::filesystem::path& pat
 bool File::CopyFile(const std::filesystem::path& src,
                     const std::filesystem::path& dst) // @suppress("Member declaration not found")
 {
+	if (IsArchivePath(src) || IsArchivePath(dst)) {
+		return false;
+	}
 	return SysFileCopyFile(src, dst);
 }
 
 bool File::RenameFile(const std::filesystem::path& src,
                       const std::filesystem::path& dst) // @suppress("Member declaration not found")
 {
+	if (IsArchivePath(src) || IsArchivePath(dst)) {
+		return false;
+	}
 	if (IsFileExisting(dst)) {
 		DeleteFile(dst); // @suppress("Invalid arguments")
 	}
@@ -477,6 +597,9 @@ bool File::RenameFile(const std::filesystem::path& src,
 }
 
 void File::RemoveReadonly(const std::filesystem::path& name) {
+	if (IsArchivePath(name)) {
+		return;
+	}
 	SysFileRemoveReadonly(name);
 }
 

src/common/file.h

@@ -7,6 +7,7 @@
 #include <cstddef>
 #include <filesystem>
 #include <memory>
+#include <optional>
 #include <string>
 #include <vector>
 
@@ -34,6 +35,10 @@ public:
 		std::string name;
 		bool        is_file;
 	};
+	struct Info {
+		bool     is_file;
+		uint64_t size;
+	};
 
 	File();
 	explicit File(const std::filesystem::path& name);
@@ -72,6 +77,7 @@ public:
 	std::vector<std::byte> ReadWholeBuffer();
 
 	static uint64_t Size(const std::filesystem::path& name);
+	static std::optional<Info> GetInfo(const std::filesystem::path& name);
 
 	static bool IsDirectoryExisting(const std::filesystem::path& path);
 	static bool IsFileExisting(const std::filesystem::path& name);

src/common/platform/sysFileIO.h

@@ -45,6 +45,7 @@ bool        SysFileTruncate(sys_file_t& f, uint64_t size); // NOLINT(google-runt
 bool        SysFileUnlink(sys_file_t&                  f,
                           const std::filesystem::path& name); // NOLINT(google-runtime-references)
 uint64_t    SysFileSize(const std::filesystem::path& file_name);
+bool        SysFileGetInfo(const std::filesystem::path& name, bool* is_file, uint64_t* size);
 bool        SysFileIsError(sys_file_t& f); // NOLINT(google-runtime-references)
 bool        SysFileIsDirectoryExisting(const std::filesystem::path& path);
 bool        SysFileIsFileExisting(const std::filesystem::path& name);

src/common/platform/sysLinuxFileIO.cpp

@@ -54,7 +54,7 @@ struct sys_file_t {
 #endif
 
 static std::filesystem::path get_internal_name(const std::filesystem::path& name) {
-	return name.is_absolute() ? name : (std::filesystem::path(".") / name);
+	return name.empty() || name.is_absolute() ? name : (std::filesystem::path(".") / name);
 }
 
 // Pass access-pattern hints to the host.
@@ -297,10 +297,21 @@ uint64_t SysFileSize(sys_file_t& f) {
 	return 0;
 }
 
+bool SysFileGetInfo(const std::filesystem::path& name, bool* is_file, uint64_t* size) {
+	const auto path = get_internal_name(name);
+	struct stat info {};
+	if (stat(path.c_str(), &info) != 0) {
+		return false;
+	}
+	*is_file = !S_ISDIR(info.st_mode);
+	*size = *is_file ? static_cast<uint64_t>(info.st_size) : 0;
+	return true;
+}
+
 uint64_t SysFileSize(const std::filesystem::path& file_name) {
-	sys_file_t* f    = SysFileOpenR(file_name);
-	uint64_t    size = SysFileSize(*f);
-	SysFileClose(f);
+	bool is_file;
+	uint64_t size = 0;
+	SysFileGetInfo(file_name, &is_file, &size);
 	return size;
 }
 
@@ -353,29 +364,15 @@ bool SysFileIsError(sys_file_t& f) {
 }
 
 bool SysFileIsDirectoryExisting(const std::filesystem::path& path) {
-	auto real_name     = get_internal_name(path);
-	auto real_name_str = real_name.string();
-
-	struct stat s {};
-
-	if (0 != stat(real_name_str.c_str(), &s)) {
-		return false;
-	}
-
-	return S_ISDIR(s.st_mode); // NOLINT
+	bool is_file;
+	uint64_t size;
+	return SysFileGetInfo(path, &is_file, &size) && !is_file;
 }
 
 bool SysFileIsFileExisting(const std::filesystem::path& name) {
-	auto real_name     = get_internal_name(name);
-	auto real_name_str = real_name.string();
-
-	struct stat s {};
-
-	if (0 != stat(real_name_str.c_str(), &s)) {
-		return false;
-	}
-
-	return !S_ISDIR(s.st_mode); // NOLINT
+	bool is_file;
+	uint64_t size;
+	return SysFileGetInfo(name, &is_file, &size) && is_file;
 }
 
 bool SysFileCreateDirectory(const std::filesystem::path& path) {

src/common/platform/sysWindowsFileIO.cpp

@@ -266,19 +266,27 @@ uint64_t SysFileSize(sys_file_t& f) {
 	return 0;
 }
 
-uint64_t SysFileSize(const std::filesystem::path& file_name) {
-	LARGE_INTEGER             s;
-	WIN32_FILE_ATTRIBUTE_DATA a;
-
-	auto wide = file_name.wstring();
-	if (GetFileAttributesExW(wide.c_str(), GetFileExInfoStandard, &a) == 0) {
-		return 0;
+bool SysFileGetInfo(const std::filesystem::path& name, bool* is_file, uint64_t* size) {
+	const bool directory_path = !name.empty() && !name.has_filename() && name != name.root_path();
+	WIN32_FILE_ATTRIBUTE_DATA info {};
+	if (GetFileAttributesExW(directory_path ? name.parent_path().c_str() : name.c_str(),
+	                         GetFileExInfoStandard, &info) == 0) {
+		return false;
 	}
+	const bool file = (info.dwFileAttributes & FILE_ATTRIBUTE_DIRECTORY) == 0;
+	if (directory_path && file) {
+		return false;
+	}
+	*is_file = file;
+	*size = *is_file ? ((static_cast<uint64_t>(info.nFileSizeHigh) << 32u) | info.nFileSizeLow) : 0;
+	return true;
+}
 
-	s.HighPart = static_cast<LONG>(a.nFileSizeHigh);
-	s.LowPart  = a.nFileSizeLow;
-
-	return s.QuadPart;
+uint64_t SysFileSize(const std::filesystem::path& file_name) {
+	bool is_file;
+	uint64_t size = 0;
+	SysFileGetInfo(file_name, &is_file, &size);
+	return size;
 }
 
 bool SysFileTruncate(sys_file_t& f, uint64_t size) {
@@ -346,17 +354,15 @@ bool SysFileIsError(sys_file_t& f) {
 }
 
 bool SysFileIsDirectoryExisting(const std::filesystem::path& path) {
-	auto  wide = path.wstring();
-	DWORD a    = GetFileAttributesW(wide.c_str());
-	return a != INVALID_FILE_ATTRIBUTES &&
-	       ((a & static_cast<DWORD>(FILE_ATTRIBUTE_DIRECTORY)) != 0u);
+	bool is_file;
+	uint64_t size;
+	return SysFileGetInfo(path, &is_file, &size) && !is_file;
 }
 
 bool SysFileIsFileExisting(const std::filesystem::path& name) {
-	auto  wide = name.wstring();
-	DWORD a    = GetFileAttributesW(wide.c_str());
-	return a != INVALID_FILE_ATTRIBUTES &&
-	       ((a & static_cast<DWORD>(FILE_ATTRIBUTE_DIRECTORY)) == 0u);
+	bool is_file;
+	uint64_t size;
+	return SysFileGetInfo(name, &is_file, &size) && is_file;
 }
 
 bool SysFileCreateDirectory(const std::filesystem::path& path) {

src/common/profiler.cpp

@@ -2,52 +2,13 @@
 
 #include "common/emulatorConfig.h"
 
-#include <algorithm>
 #include <common/TracyProtocol.hpp>
 #include <common/TracyVersion.hpp>
 #include <cstdio>
 #include <tracy/Tracy.hpp>
-#include <vector>
-
-namespace {
-
-thread_local std::vector<Profiler::ScopedBlock*> g_block_stack;
-
-void RemoveBlock(Profiler::ScopedBlock* block) {
-	auto block_it = std::find(g_block_stack.rbegin(), g_block_stack.rend(), block);
-	if (block_it != g_block_stack.rend()) {
-		g_block_stack.erase(std::next(block_it).base());
-	}
-}
-
-} // namespace
 
 namespace Profiler {
 
-ScopedBlock::ScopedBlock(const tracy::SourceLocationData* source_location) {
-	if (tracy::ProfilerAvailable()) {
-		m_zone.emplace(source_location, TRACY_CALLSTACK, true);
-		g_block_stack.push_back(this);
-	}
-}
-
-ScopedBlock::~ScopedBlock() {
-	End();
-}
-
-void ScopedBlock::End() {
-	if (m_zone.has_value()) {
-		m_zone.reset();
-		RemoveBlock(this);
-	}
-}
-
-void EndBlock() {
-	if (!g_block_stack.empty()) {
-		g_block_stack.back()->End();
-	}
-}
-
 void SetThreadName(const char* name) {
 	if (tracy::ProfilerAvailable() && name != nullptr) {
 		tracy::SetThreadName(name);

src/common/profiler.h

@@ -4,7 +4,6 @@
 #include "common/common.h"
 
 #include <cstdint>
-#include <optional>
 #include <tracy/Tracy.hpp> // IWYU pragma: export
 
 namespace profiler::colors {
@@ -23,22 +22,6 @@ inline constexpr uint32_t DeepOrangeA200 = 0xff6e40;
 
 namespace Profiler {
 
-class ScopedBlock {
-public:
-	explicit ScopedBlock(const tracy::SourceLocationData* source_location);
-	ScopedBlock(const ScopedBlock&)            = delete;
-	ScopedBlock& operator=(const ScopedBlock&) = delete;
-	ScopedBlock(ScopedBlock&&)                 = delete;
-	ScopedBlock& operator=(ScopedBlock&&)      = delete;
-	~ScopedBlock();
-
-	void End();
-
-private:
-	std::optional<tracy::ScopedZone> m_zone;
-};
-
-void EndBlock();
 void SetThreadName(const char* name);
 
 void Initialize();
@@ -66,19 +49,11 @@ struct Lifecycle {
 	    kyty_profiler_source_location_, line) {name, TracyFunction, TracyFile,                     \
 	                                           static_cast<uint32_t>(line),                        \
 	                                           KYTY_PROFILER_COLOR_OR_DEFAULT(0, __VA_ARGS__)};    \
-	Profiler::ScopedBlock KYTY_PROFILER_CONCAT(kyty_profiler_block_, line)(                        \
-	    &KYTY_PROFILER_CONCAT(kyty_profiler_source_location_, line))
+	tracy::ScopedZone KYTY_PROFILER_CONCAT(kyty_profiler_block_, line)(                           \
+	    &KYTY_PROFILER_CONCAT(kyty_profiler_source_location_, line),                              \
+	    TRACY_CALLSTACK, tracy::ProfilerAvailable())
 
-#define KYTY_PROFILER_FUNCTION(...) KYTY_PROFILER_FUNCTION_IMPL(__LINE__ __VA_OPT__(, ) __VA_ARGS__)
-#define KYTY_PROFILER_FUNCTION_IMPL(line, ...)                                                     \
-	static constexpr tracy::SourceLocationData KYTY_PROFILER_CONCAT(                               \
-	    kyty_profiler_source_location_, line) {nullptr, TracyFunction, TracyFile,                  \
-	                                           static_cast<uint32_t>(line),                        \
-	                                           KYTY_PROFILER_COLOR_OR_DEFAULT(0, __VA_ARGS__)};    \
-	Profiler::ScopedBlock KYTY_PROFILER_CONCAT(kyty_profiler_block_, line)(                        \
-	    &KYTY_PROFILER_CONCAT(kyty_profiler_source_location_, line))
-
-#define KYTY_PROFILER_END_BLOCK Profiler::EndBlock()
+#define KYTY_PROFILER_FUNCTION(...) KYTY_PROFILER_BLOCK(nullptr __VA_OPT__(, ) __VA_ARGS__)
 
 #define KYTY_PROFILER_THREAD(name) Profiler::SetThreadName(name)
 

src/common/threads.cpp

@@ -2,7 +2,6 @@
 
 #include "common/assert.h"
 
-#include <algorithm>
 #include <atomic>
 #include <cerrno>
 #include <chrono>             // IWYU pragma: keep
@@ -26,8 +25,7 @@
 #ifdef KYTY_WIN_CS
 #include <windows.h> // IWYU pragma: keep
 // IWYU pragma: no_include <winbase.h>
-constexpr DWORD    KYTY_CS_SPIN_COUNT          = 4000;
-constexpr uint64_t KYTY_SLEEP_SPIN_LIMIT_100NS = 500; // 50 us
+constexpr DWORD KYTY_CS_SPIN_COUNT = 4000;
 
 #ifndef CREATE_WAITABLE_TIMER_HIGH_RESOLUTION
 #define CREATE_WAITABLE_TIMER_HIGH_RESOLUTION 0x00000002
@@ -38,30 +36,6 @@ static void SleepHighResolution100ns(uint64_t units_100ns) {
 		return;
 	}
 
-	// Keep spinning only where a kernel transition is
-	// likely to cost more than the requested delay; ordinary millisecond sleeps use the
-	// per-thread high-resolution waitable timer below.
-	if (units_100ns <= KYTY_SLEEP_SPIN_LIMIT_100NS) {
-		LARGE_INTEGER frequency {};
-		LARGE_INTEGER start {};
-		if (QueryPerformanceFrequency(&frequency) != 0 && QueryPerformanceCounter(&start) != 0 &&
-		    frequency.QuadPart > 0) {
-			const auto wait_ticks =
-			    static_cast<LONGLONG>((static_cast<long double>(units_100ns) *
-			                           static_cast<long double>(frequency.QuadPart)) /
-			                          10000000.0L);
-			const auto    deadline = start.QuadPart + std::max<LONGLONG>(wait_ticks, 1);
-			LARGE_INTEGER now {};
-			do {
-				if (QueryPerformanceCounter(&now) == 0) {
-					break;
-				}
-				YieldProcessor();
-			} while (now.QuadPart < deadline);
-			return;
-		}
-	}
-
 	thread_local HANDLE timer = CreateWaitableTimerExW(
 	    nullptr, nullptr, CREATE_WAITABLE_TIMER_HIGH_RESOLUTION, TIMER_ALL_ACCESS);
 	if (timer == nullptr) {
@@ -131,14 +105,13 @@ static SleepConditionVariableCS_func_t ResolveSleepConditionVariableCS() {
 #endif
 
 #ifdef KYTY_POSIX_HIGH_RES_SLEEP
-// Spin for very short waits; use an absolute deadline for longer waits.
+// An absolute deadline preserves the requested sleep across signal interruptions.
 static void SleepHighResolutionNanos(uint64_t nanos) {
 	if (nanos == 0) {
 		return;
 	}
 
 	constexpr uint64_t NANOS_PER_SEC = 1000000000;
-	constexpr uint64_t SPIN_LIMIT_NS = 50000; // below this a context switch dominates
 
 	timespec deadline {};
 	if (clock_gettime(CLOCK_MONOTONIC, &deadline) != 0) {
@@ -150,17 +123,6 @@ static void SleepHighResolutionNanos(uint64_t nanos) {
 	deadline.tv_sec += static_cast<time_t>(target_nsec / NANOS_PER_SEC);
 	deadline.tv_nsec = static_cast<long>(target_nsec % NANOS_PER_SEC);
 
-	if (nanos <= SPIN_LIMIT_NS) {
-		timespec now {};
-		do {
-			if (clock_gettime(CLOCK_MONOTONIC, &now) != 0) {
-				return;
-			}
-		} while (now.tv_sec < deadline.tv_sec ||
-		         (now.tv_sec == deadline.tv_sec && now.tv_nsec < deadline.tv_nsec));
-		return;
-	}
-
 	while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME, &deadline, nullptr) == EINTR) {
 	}
 }

src/graphics/guest_gpu/command_processor/commandProcessor.h

@@ -62,7 +62,6 @@ public:
 
 	void            BufferInit();
 	void            BufferFlush();
-	void            BufferFlushAndWait();
 	void            BufferWait();
 	HW::Context&    GetCtx() { return m_ctx; }
 	HW::UserConfig& GetUcfg() { return m_ucfg; }

src/graphics/guest_gpu/command_processor/pm4Dispatch.cpp

@@ -140,6 +140,8 @@ constexpr auto MakeShaderDispatchTable() {
 	g_hw_sh_func[Pm4::COMPUTE_PGM_RSRC3]               = HwShSetCsRegisters;
 	g_hw_sh_func[Pm4::COMPUTE_PACE_ID]                 = HwShSetCsRegisters;
 	g_hw_sh_func[Pm4::SPI_GRAPHICS_SHADER_CONTROL_PS]  = HwShSetRegisters;
+	g_hw_sh_func[Pm4::SPI_SHADER_USER_DATA_ADDR_LO_PS] = HwShSetRegisters;
+	g_hw_sh_func[Pm4::SPI_SHADER_USER_DATA_ADDR_HI_PS] = HwShSetRegisters;
 	g_hw_sh_func[Pm4::SPI_GRAPHICS_SHADER_CONTROL_GS]  = HwShSetRegisters;
 	g_hw_sh_func[Pm4::SPI_SHADER_USER_DATA_ADDR_LO_GS] = HwShSetRegisters;
 	g_hw_sh_func[Pm4::SPI_SHADER_USER_DATA_ADDR_HI_GS] = HwShSetRegisters;

src/graphics/guest_gpu/command_processor/pm4Handlers.cpp

@@ -2266,7 +2266,6 @@ KYTY_CP_OP_PARSER(CpOpReleaseMem) {
 	const bool gl2_writeback = ((gcr_cntl & GcrGl2Writeback) != 0);
 
 	auto trigger_interrupt = [&]() {
-		bool queued = false;
 		switch (interrupt_selector) {
 			case 0x00:
 			case 0x03: break;
@@ -2274,13 +2273,10 @@ KYTY_CP_OP_PARSER(CpOpReleaseMem) {
 			case 0x02:
 			case 0x04:
 				cp.TriggerEopEventAtEndOfPipe(interrupt_context_id);
-				queued = true;
+				cp.BufferFlush();
 				break;
 			default: EXIT("unknown release_mem interrupt selector\n");
 		}
-		if (queued) {
-			cp.BufferFlush();
-		}
 	};
 
 	if (data_sel == 0 || interrupt_selector == 4) {
@@ -2318,7 +2314,9 @@ KYTY_CP_OP_PARSER(CpOpReleaseMem) {
 		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);
-		cp.BufferFlush();
+		if (interrupt_selector == 0x01 || interrupt_selector == 0x02) {
+			cp.BufferFlush();
+		}
 
 		return 7;
 	}
@@ -3624,6 +3622,18 @@ void GraphicsInitJmpTablesShIndirect() {
 	g_hw_sh_indirect_func[Pm4::SPI_GRAPHICS_SHADER_CONTROL_PS] = [](KYTY_HW_SH_INDIRECT_ARGS) {
 		HwShIgnoreShaderRegister(cmd_offset, value);
 	};
+	for (uint32_t offset = Pm4::SPI_SHADER_USER_DATA_ADDR_LO_PS;
+	     offset <= Pm4::SPI_SHADER_USER_DATA_ADDR_HI_PS; offset++) {
+		g_hw_sh_indirect_func[offset] = [](KYTY_HW_SH_INDIRECT_ARGS) {
+			cp.GetShCtx().SetPsUserDataAddress(cmd_offset - Pm4::SPI_SHADER_USER_DATA_ADDR_LO_PS,
+			                                   value);
+		};
+	}
+	for (uint32_t offset = 4; offset <= 5; offset++) {
+		g_hw_sh_indirect_func[offset] = [](KYTY_HW_SH_INDIRECT_ARGS) {
+			cp.GetShCtx().SetPsAuxiliaryTableAddress(cmd_offset - 4, value);
+		};
+	}
 	g_hw_sh_indirect_func[Pm4::SPI_SHADER_PACE_ID_GS] = [](KYTY_HW_SH_INDIRECT_ARGS) {
 		HwShIgnoreShaderRegister(cmd_offset, value);
 	};

src/graphics/guest_gpu/graphicsRun.cpp

@@ -35,7 +35,6 @@ namespace Libs::Graphics {
 
 static thread_local CommandProcessor* g_current_processor = nullptr;
 static thread_local Pm4Execution*     g_current_execution = nullptr;
-static thread_local bool              g_gpu_mutex_owned   = false;
 static thread_local bool              g_gpu_thread        = false;
 static thread_local GuestGpu*         g_gpu_state         = nullptr;
 
@@ -54,27 +53,6 @@ struct DrawIndexedIndirectArgs {
 	uint32_t start_instance_location;
 };
 
-class GpuMutexLock final {
-public:
-	explicit GpuMutexLock(Common::Mutex& mutex): m_mutex(mutex) {
-		if (g_gpu_mutex_owned) {
-			EXIT("recursive GPU mutex acquisition\n");
-		}
-		g_gpu_mutex_owned = true;
-		m_mutex.Lock();
-	}
-	~GpuMutexLock() {
-		if (!g_gpu_mutex_owned) {
-			EXIT("invalid GPU mutex release\n");
-		}
-		m_mutex.Unlock();
-		g_gpu_mutex_owned = false;
-	}
-
-private:
-	Common::Mutex& m_mutex;
-};
-
 static bool GraphicsRunDebugDumpEnabled() {
 	return Config::GraphicsDebugDumpEnabled() &&
 	       Config::GetPrintfDirection() != Config::LogDirection::Silent;
@@ -84,6 +62,7 @@ GuestGpu::GuestGpu(RenderContext& renderer): m_renderer(renderer) {
 	EXIT_NOT_IMPLEMENTED(!Common::Thread::IsMainThread());
 	GraphicsInitJmpTables();
 	m_gfx_cp = std::make_unique<CommandProcessor>(renderer, 0);
+	m_gfx_cp->Reset();
 	m_thread = std::jthread(ThreadRun, this);
 }
 
@@ -160,20 +139,16 @@ void GuestGpu::Submit(std::span<const uint32_t> draw_commands,
 	if (draw_commands.empty()) {
 		return;
 	}
-	GpuMutexLock lock(m_submission_mutex);
-	Submission   submission;
+	Submission submission;
 	submission.type              = SubmissionType::Graphics;
 	submission.queue_id          = 0;
 	submission.commands          = draw_commands;
 	submission.constant_commands = constant_commands;
-	submission.reset_processor   = m_graphics_done;
-	m_graphics_done              = false;
 	Enqueue(std::move(submission));
 }
 
 void GuestGpu::SubmitCompute(uint32_t queue, std::span<const uint32_t> commands) {
 	EXIT_IF(commands.empty());
-	GpuMutexLock lock(m_submission_mutex);
 
 	EXIT_NOT_IMPLEMENTED(queue < ComputeQueueBase || queue >= ComputeQueueBase + ComputeQueueCount);
 
@@ -186,22 +161,21 @@ void GuestGpu::SubmitCompute(uint32_t queue, std::span<const uint32_t> commands)
 }
 
 void GuestGpu::SubmitFlipPreparation(uint64_t request_id) {
-	GpuMutexLock lock(m_submission_mutex);
-	Submission   submission;
+	Submission submission;
 	submission.type            = SubmissionType::FlipPreparation;
 	submission.queue_id        = 0;
-	submission.reset_processor = m_graphics_done;
 	submission.flip_request_id = request_id;
-	m_graphics_done            = false;
 	Enqueue(std::move(submission));
 }
 
-void GuestGpu::Done() {
-	GpuMutexLock lock(m_submission_mutex);
-	if (!IsGpuThread()) {
-		WaitForIdle();
-	}
-	m_graphics_done = true;
+void GuestGpu::SuspendPoint() {
+	EXIT_IF(IsGpuThread() || CommandScheduler::InDeferredOperation());
+	// Do not hold a queue lock while waiting: asynchronous work may be needed to
+	// finish the preceding graphics frame. The first point returns immediately.
+	m_suspend_point_ready->acquire();
+	Submission submission;
+	submission.type = SubmissionType::SuspendPoint;
+	Enqueue(std::move(submission));
 	m_done_num++;
 }
 
@@ -229,6 +203,9 @@ void CommandProcessor::Reset() {
 	m_context_state_pushed             = false;
 	m_index_type_and_size              = 0;
 	m_index_buffer_size                = 0;
+	m_index_base_addr                  = 0;
+	m_num_instances                    = 1;
+	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;
@@ -268,10 +245,6 @@ void CommandProcessor::BufferFlush() {
 	GetScheduler().Flush();
 }
 
-void CommandProcessor::BufferFlushAndWait() {
-	GetScheduler().FlushAndWait();
-}
-
 void CommandProcessor::BufferWait() {
 	BufferInit();
 	GetScheduler().Finish();
@@ -459,6 +432,7 @@ void GuestGpu::Enqueue(Submission submission) {
 }
 
 void GuestGpu::WaitForIdle() {
+	EXIT_IF(IsGpuThread() || CommandScheduler::InDeferredOperation());
 	Common::LockGuard lock(m_queue_mutex);
 	while (m_processing || !m_commands.empty() || m_submission_count != 0) {
 		m_idle.Wait(&m_queue_mutex);
@@ -564,11 +538,7 @@ void GuestGpu::ThreadRun(void* data) {
 
 bool GuestGpu::Process(Submission& submission) {
 	const bool first_slice = !submission.started;
-	auto& cp = GetProcessor(submission.queue_id);
-
-	if (first_slice && submission.reset_processor) {
-		cp.Reset();
-	}
+	auto&      cp          = GetProcessor(submission.queue_id);
 
 	if (first_slice) {
 		submission.started = true;
@@ -645,6 +615,13 @@ bool GuestGpu::Process(Submission& submission) {
 			m_renderer.RunGarbageCollector();
 			cp.PrepareCpuFlip(submission.flip_request_id);
 			break;
+		case SubmissionType::SuspendPoint:
+			cp.EmitGlobalBarrier();
+			m_renderer.GetCommandScheduler().DeferPriorityOperation(
+			    [ready = m_suspend_point_ready] { ready->release(); });
+			cp.BufferFlush();
+			cp.Reset();
+			break;
 	}
 
 	return complete;
@@ -849,9 +826,7 @@ void CommandProcessor::SetPredication(uint32_t condition, uint32_t op, uint32_t
 			}
 		} break;
 		case 0x03:
-			if (wait_op != 0) {
-				BufferFlushAndWait();
-			}
+			// The wait selector applies only to Z-pass query readiness.
 			EXIT_NOT_IMPLEMENTED(address == nullptr);
 			value = *reinterpret_cast<const volatile uint64_t*>(address);
 			break;
@@ -1034,9 +1009,6 @@ void CommandProcessor::DispatchDirect(uint32_t thread_group_x, uint32_t thread_g
 	m_sh_ctx.SetCsWaveSize(Pm4::ComputeWaveSize(mode));
 
 	uint32_t frame_num = 0;
-	// uint32_t local_x   = 1;
-	// uint32_t local_y   = 1;
-	// uint32_t local_z   = 1;
 
 	{
 		frame_num = m_renderer.GetGpu().GetFrameNum();
@@ -1058,35 +1030,9 @@ void CommandProcessor::DispatchDirect(uint32_t thread_group_x, uint32_t thread_g
 			}
 		}
 
-		const auto& cs = m_sh_ctx.GetCs().cs_regs;
-		// local_x        = std::max(cs.num_thread_x, 1u);
-		// local_y        = std::max(cs.num_thread_y, 1u);
-		// local_z        = std::max(cs.num_thread_z, 1u);
 		m_renderer.GetRenderExecutor().DispatchDirect(m_submit_id, CurrentBuffer(), thread_group_x,
 		                                              thread_group_y, thread_group_z, mode);
 	}
-
-	/*constexpr uint32_t DispatchInitiatorUseThreadDimensions = 1u << 5u;
-	auto               group_count = [](uint32_t threads, uint32_t group_size) {
-	    return (threads == 0
-	                ? 0u
-	                : (threads + std::max(group_size, 1u) - 1u) / std::max(group_size, 1u));
-	};
-
-	auto groups_x = thread_group_x;
-	auto groups_y = thread_group_y;
-	auto groups_z = thread_group_z;
-	if ((mode & DispatchInitiatorUseThreadDimensions) != 0) {
-	    groups_x = group_count(thread_group_x, local_x);
-	    groups_y = group_count(thread_group_y, local_y);
-	    groups_z = group_count(thread_group_z, local_z);
-	}
-
-	const uint64_t invocations =
-	    static_cast<uint64_t>(groups_x) * groups_y * groups_z * local_x * local_y * local_z;
-	if (invocations != 0) {
-	    BufferFlushAndWait();
-	}*/
 }
 
 void CommandProcessor::DispatchIndirect(uint64_t args_addr, uint32_t mode) {
@@ -1502,8 +1448,9 @@ void CommandProcessor::PrepareCpuFlip(uint64_t request_id) {
 	ProcessorScope processor_scope(*this);
 
 	m_renderer.GetVideoOut().PrepareFlip(request_id, command);
+	GetScheduler().DeferPriorityOperation(
+	    [this, request_id] { m_renderer.GetVideoOut().CompleteFlip(request_id); });
 	GetScheduler().Flush();
-	m_renderer.GetVideoOut().CompleteFlip(request_id);
 }
 
 void CommandProcessor::SynchronizeGpu() {

src/graphics/guest_gpu/graphicsRun.h

@@ -13,6 +13,7 @@
 #include <deque>
 #include <memory>
 #include <mutex>
+#include <semaphore>
 #include <span>
 #include <thread>
 
@@ -32,11 +33,14 @@ public:
 	void               SendCommandSync(Common::UniqueFunction<void>&& command);
 
 	// 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              SubmitCompute(uint32_t queue, std::span<const uint32_t> commands);
-	void              SubmitFlipPreparation(uint64_t request_id);
-	void              Done();
+	void Submit(std::span<const uint32_t> draw_commands,
+	            std::span<const uint32_t> constant_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.
+	void SuspendPoint();
+	// Wait for guest command processing, including all compute queues (not native GPU completion).
+	void              WaitForIdle();
 	[[nodiscard]] int GetFrameNum() const;
 
 	[[nodiscard]] static bool IsGpuThread() noexcept;
@@ -48,7 +52,7 @@ private:
 	static constexpr uint32_t ComputeQueueBase     = 0x20;
 	static constexpr uint32_t QueueCount           = 1 + ComputeQueueCount;
 
-	enum class SubmissionType { Graphics, Compute, FlipPreparation };
+	enum class SubmissionType { Graphics, Compute, FlipPreparation, SuspendPoint };
 
 	struct Submission {
 		SubmissionType            type     = SubmissionType::Graphics;
@@ -57,7 +61,6 @@ private:
 		std::span<const uint32_t> constant_commands;
 		Pm4Execution              command_execution;
 		Pm4Execution              constant_execution;
-		bool                      reset_processor   = false;
 		bool                      started           = false;
 		bool                      command_complete  = false;
 		bool                      constant_complete = false;
@@ -66,14 +69,12 @@ private:
 	};
 
 	void              Enqueue(Submission submission);
-	void              WaitForIdle();
 	void              ProcessCommands();
 	bool              Process(Submission& submission);
 	static void       ThreadRun(void* data);
 	CommandProcessor& GetProcessor(uint32_t queue_id);
 
 	RenderContext&                                 m_renderer;
-	Common::Mutex                                  m_submission_mutex;
 	Common::Mutex                                  m_queue_mutex;
 	std::mutex                                     m_shutdown_mutex;
 	Common::CondVar                                m_work_available;
@@ -84,10 +85,12 @@ private:
 	uint32_t                                       m_next_queue        = 0;
 	uint32_t                                       m_submission_count  = 0;
 	bool                                           m_processing        = false;
-	bool                                           m_graphics_done     = true;
 	bool                                           m_accepting         = true;
 	bool                                           m_stopping          = false;
 	bool                                           m_shutdown_complete = false;
+	// Completion callbacks can outlive GuestGpu during renderer shutdown.
+	std::shared_ptr<std::binary_semaphore> m_suspend_point_ready =
+	    std::make_shared<std::binary_semaphore>(1);
 
 	std::unique_ptr<CommandProcessor>                                m_gfx_cp;
 	std::array<std::unique_ptr<CommandProcessor>, ComputeQueueCount> m_compute_cp;

src/graphics/guest_gpu/hardwareContext.h

@@ -458,6 +458,9 @@ struct PsShaderResource2 {
 
 struct PsStageRegisters {
 	uint64_t          data_addr = 0;
+	uint64_t          user_data_addr = 0;
+	// Hades II uses PS offsets 4:5 for a byte-addressed descriptor table.
+	uint64_t          auxiliary_table_addr = 0;
 	PsShaderResource1 rsrc1;
 	PsShaderResource2 rsrc2;
 };
@@ -1017,6 +1020,12 @@ public:
 	void SetGsShaderResource2(const GsShaderResource2& rsrc2) { m_vs.gs_regs.rsrc2 = rsrc2; }
 
 	void SetPsShaderBase(uint64_t addr) { m_ps.ps_regs.data_addr = addr; }
+	void SetPsUserDataAddress(uint32_t word, uint32_t value) {
+		SetUserDataAddressWord(m_ps.ps_regs.user_data_addr, word, value);
+	}
+	void SetPsAuxiliaryTableAddress(uint32_t word, uint32_t value) {
+		SetUserDataAddressWord(m_ps.ps_regs.auxiliary_table_addr, word, value);
+	}
 	void SetPsShaderResource1(const PsShaderResource1& rsrc1) { m_ps.ps_regs.rsrc1 = rsrc1; }
 	void SetPsShaderResource2(const PsShaderResource2& rsrc2) { m_ps.ps_regs.rsrc2 = rsrc2; }
 

src/graphics/guest_gpu/pm4.h

@@ -697,6 +697,8 @@ constexpr uint32_t CX_NUM = 0x3FF + 1;
 
 /* Shader registers */
 
+constexpr uint32_t SPI_SHADER_USER_DATA_ADDR_LO_PS = 0x2;
+constexpr uint32_t SPI_SHADER_USER_DATA_ADDR_HI_PS = 0x3;
 constexpr uint32_t SPI_SHADER_PACE_ID_PS            = 0x6;
 constexpr uint32_t SPI_GRAPHICS_SHADER_CONTROL_PS    = 0x7;
 constexpr uint32_t SPI_SHADER_PGM_LO_PS              = 0x8;

src/graphics/guest_gpu/tile.cpp

@@ -1213,20 +1213,12 @@ bool TileGetRenderTargetSize(uint32_t width, uint32_t height, uint32_t pitch,
 	return true;
 }
 
-bool TileGetDccSize(uint32_t width, uint32_t height, uint32_t slices,
-                    uint32_t bytes_per_element, uint32_t levels, Prospero::TileMode tile,
-                    TileSizeAlign& total_size, uint32_t num_fragments_log2) {
-	total_size = {};
-	if (width == 0 || height == 0 || slices == 0 || levels != 1 || num_fragments_log2 != 0 ||
-	    !std::has_single_bit(bytes_per_element) || bytes_per_element > 16 ||
-	    (tile != Prospero::TileMode::kRenderTarget && tile != Prospero::TileMode::kDepth)) {
+static bool TileGetColorMetadataSize(uint32_t width, uint32_t height, uint32_t slices,
+                                      uint32_t block_width, uint32_t block_height,
+                                      TileSizeAlign& total_size) {
+	if (width == 0 || height == 0 || slices == 0) {
 		return false;
 	}
-	// Gen5 color metadata uses 4 KiB blocks, with one byte per 256 bytes of color data.
-	// Metadata stays thin for volume surfaces, so every depth slice has its own block raster.
-	const uint32_t coverage_bits = 20u - std::countr_zero(bytes_per_element);
-	const uint32_t block_width   = 1u << ((coverage_bits + 1u) / 2u);
-	const uint32_t block_height  = 1u << (coverage_bits / 2u);
 	const uint64_t blocks_x = (static_cast<uint64_t>(width) + block_width - 1u) / block_width;
 	const uint64_t blocks_y = (static_cast<uint64_t>(height) + block_height - 1u) / block_height;
 	const uint64_t blocks   = blocks_x * blocks_y;
@@ -1238,36 +1230,27 @@ bool TileGetDccSize(uint32_t width, uint32_t height, uint32_t slices,
 	return true;
 }
 
-bool TileGetRenderTargetMipLayout(uint32_t width, uint32_t height, uint32_t pitch,
-                                  uint32_t bytes_per_element, uint32_t levels,
-                                  TileSizeAlign& total_size, TileSizeOffset* level_sizes,
-                                  TilePaddedSize* padded_size) {
+bool TileGetDccSize(uint32_t width, uint32_t height, uint32_t slices,
+                    uint32_t bytes_per_element, uint32_t levels, Prospero::TileMode tile,
+                    TileSizeAlign& total_size, uint32_t num_fragments_log2) {
 	total_size = {};
-	if (width == 0 || height == 0 || levels == 0 || levels > 16 ||
-	    pitch != TileGetRenderTargetPitch(width, bytes_per_element)) {
+	if (levels != 1 || num_fragments_log2 != 0 ||
+	    !std::has_single_bit(bytes_per_element) || bytes_per_element > 16 ||
+	    (tile != Prospero::TileMode::kRenderTarget && tile != Prospero::TileMode::kDepth)) {
 		return false;
 	}
-	uint32_t max_levels    = 1;
-	uint32_t max_dimension = std::max(width, height);
-	while (max_dimension > 1) {
-		max_dimension >>= 1u;
-		max_levels++;
-	}
-	if (levels > max_levels) {
-		return false;
-	}
-	auto format = Prospero::BufferFormat::kInvalid;
-	switch (bytes_per_element) {
-		case 1: format = Prospero::BufferFormat::k8UNorm; break;
-		case 2: format = Prospero::BufferFormat::k16UNorm; break;
-		case 4: format = Prospero::BufferFormat::k32Float; break;
-		case 8: format = Prospero::BufferFormat::k16_16_16_16Float; break;
-		case 16: format = Prospero::BufferFormat::k32_32_32_32Float; break;
-		default: return false;
-	}
-	TileGetTextureSize(format, width, height, levels, Prospero::TileMode::kRenderTarget,
-	                   &total_size, level_sizes, padded_size);
-	return total_size.size != 0 && total_size.align == 65536;
+	// Gen5 DCC uses 4 KiB blocks, with one byte per 256 bytes of color data.
+	// Metadata stays thin for volume surfaces, so every depth slice has its own block raster.
+	const uint32_t coverage_bits = 20u - std::countr_zero(bytes_per_element);
+	return TileGetColorMetadataSize(width, height, slices, 1u << ((coverage_bits + 1u) / 2u),
+	                                1u << (coverage_bits / 2u), total_size);
+}
+
+bool TileGetCmaskSize(uint32_t width, uint32_t height, uint32_t slices, uint32_t levels,
+                      TileSizeAlign& total_size) {
+	total_size = {};
+	// Gen5 CMASK stores four bits per 8x8 tile in 4 KiB blocks covering 1024x512 pixels.
+	return levels == 1 && TileGetColorMetadataSize(width, height, slices, 1024, 512, total_size);
 }
 
 void TileGetTextureSize(Prospero::BufferFormat format, uint32_t width, uint32_t height,

src/graphics/guest_gpu/tile.h

@@ -126,10 +126,8 @@ bool     TileGetRenderTargetSize(uint32_t width, uint32_t height, uint32_t pitch
 bool     TileGetDccSize(uint32_t width, uint32_t height, uint32_t slices,
                         uint32_t bytes_per_element, uint32_t levels, Prospero::TileMode tile,
                         TileSizeAlign& total_size, uint32_t num_fragments_log2 = 0);
-bool     TileGetRenderTargetMipLayout(uint32_t width, uint32_t height, uint32_t pitch,
-                                      uint32_t bytes_per_element, uint32_t levels,
-                                      TileSizeAlign& total_size, TileSizeOffset* level_sizes,
-                                      TilePaddedSize* padded_size);
+bool     TileGetCmaskSize(uint32_t width, uint32_t height, uint32_t slices, uint32_t levels,
+                          TileSizeAlign& total_size);
 void     TileGetTextureSize(Prospero::BufferFormat format, uint32_t width, uint32_t height,
                             uint32_t levels, Prospero::TileMode tile, TileSizeAlign* total_size,
                             TileSizeOffset* level_sizes, TilePaddedSize* padded_size);

src/graphics/host_gpu/graphicContext.h

@@ -29,6 +29,7 @@ struct GraphicContext {
 	bool                               memory_budget_ext_enabled             = false;
 	bool                               compute_subgroup_size_control_enabled = false;
 	bool                               sample_rate_shading_enabled           = false;
+	// bool fp64_denorm_preserve = false; // Temporarily disabled.
 	bool                               attachment_feedback_loop_enabled      = false;
 	bool                               provoking_vertex_last_enabled         = false;
 	bool                               supports_block_texel_view              = false;

src/graphics/host_gpu/regionDefinitions.h

@@ -3,6 +3,7 @@
 
 #include "common/bitArray.h"
 #include "common/common.h"
+#include "kernel/memory.h"
 
 #include <compare>
 
@@ -10,7 +11,8 @@ namespace Libs::Graphics {
 
 constexpr uint64_t TRACKER_PAGE_SIZE    = 4ull * 1024ull;
 constexpr uint64_t TRACKER_REGION_SIZE  = 4ull * 1024ull * 1024ull;
-constexpr uint64_t TRACKER_ADDRESS_SIZE = 1ull << 40u;
+constexpr uint64_t TRACKER_ADDRESS_SIZE = 1ull << 44u;
+constexpr uint64_t LOWER_ADDRESS_SIZE   = 1ull << 40u;
 constexpr size_t   TRACKER_REGION_PAGES = TRACKER_REGION_SIZE / TRACKER_PAGE_SIZE;
 
 struct GuestRange {
@@ -19,8 +21,11 @@ struct GuestRange {
 
 	[[nodiscard]] constexpr bool Empty() const noexcept { return address == 0 && size == 0; }
 	[[nodiscard]] constexpr bool Valid() const noexcept {
-		return address != 0 && size != 0 && address < TRACKER_ADDRESS_SIZE &&
-		       size <= TRACKER_ADDRESS_SIZE - address;
+		if (address == 0 || size == 0) return false;
+		if (address < LOWER_ADDRESS_SIZE) return size <= LOWER_ADDRESS_SIZE - address;
+		constexpr auto base = LibKernel::Memory::kExtendedMemoryBase;
+		constexpr auto length = LibKernel::Memory::kExtendedMemorySize;
+		return address >= base && address - base < length && size <= length - (address - base);
 	}
 	[[nodiscard]] constexpr bool     ValidOrEmpty() const noexcept { return Empty() || Valid(); }
 	[[nodiscard]] constexpr uint64_t End() const noexcept { return address + size; }

src/graphics/host_gpu/regionManager.h

@@ -30,21 +30,14 @@ class TrackingSpinLock final {
 public:
 	void lock() noexcept {
 		const auto thread = CurrentThread();
-		if (m_owner.load(std::memory_order_relaxed) == thread) {
-			EXIT("recursive region tracking lock\n");
-		}
 		while (m_lock.test_and_set(std::memory_order_acquire)) {
-			if (m_owner.load(std::memory_order_relaxed) == thread) {
-				EXIT("recursive region tracking lock while contended\n");
-			}
+			EXIT_NOT_IMPLEMENTED(m_owner.load(std::memory_order_relaxed) == thread);
 			std::atomic_signal_fence(std::memory_order_seq_cst);
 		}
 		m_owner.store(thread, std::memory_order_relaxed);
 	}
 	void unlock() noexcept {
-		if (m_owner.load(std::memory_order_relaxed) != CurrentThread()) {
-			EXIT("region tracking lock released by non-owner\n");
-		}
+		EXIT_NOT_IMPLEMENTED(m_owner.load(std::memory_order_relaxed) != CurrentThread());
 		m_owner.store(0, std::memory_order_relaxed);
 		m_lock.clear(std::memory_order_release);
 	}
@@ -88,8 +81,7 @@ public:
 	template <DirtySource source>
 	[[nodiscard]] bool IsModified(uint64_t offset, uint64_t size) const {
 		const auto [start, end] = GetPageRange(m_cpu_addr + offset, size);
-		const auto& bits        = GetBits<source>();
-		return RegionBits(bits, start, end).Any();
+		return GetBits<source>().FirstRangeFrom(start).first < end;
 	}
 
 	template <DirtySource source, bool enable>

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

@@ -55,7 +55,8 @@ template <bool insert>
 void BufferCache::ChangeRegister(BufferId id) {
 	auto& buffer = m_slot_buffers[id];
 	PageTable::PageRange pages {};
-	EXIT_IF(!PageTable::TryGetPageRange(buffer.CpuAddress(), buffer.Size(), pages));
+	EXIT_IF(!(GuestRange {buffer.CpuAddress(), buffer.Size()}.Valid()) ||
+	        !PageTable::TryGetPageRange(buffer.CpuAddress(), buffer.Size(), pages));
 	for (size_t page = pages.first; page < pages.last_exclusive; ++page) {
 		if constexpr (insert) {
 			m_page_table[page] = id;
@@ -64,6 +65,7 @@ void BufferCache::ChangeRegister(BufferId id) {
 		}
 	}
 	const auto size_pages = pages.last_exclusive - pages.first;
+	const auto table_offset = PageIndex(buffer.CpuAddress()) * sizeof(vk::DeviceAddress);
 	if constexpr (insert) {
 		const auto [it, inserted] = m_buffers.emplace(buffer.CpuAddress(), id);
 		(void)it;
@@ -75,7 +77,7 @@ void BufferCache::ChangeRegister(BufferId id) {
 		for (uint64_t i = 0; i < size_pages; ++i) {
 			addresses.push_back(buffer.BufferDeviceAddress() + (i << CACHING_PAGEBITS));
 		}
-		WriteDataBuffer(m_bda_pagetable_buffer, pages.first * sizeof(vk::DeviceAddress),
+		WriteDataBuffer(m_bda_pagetable_buffer, table_offset,
 		                addresses.data(), addresses.size() * sizeof(vk::DeviceAddress));
 	} else {
 		const auto found = m_buffers.find(buffer.CpuAddress());
@@ -84,7 +86,7 @@ void BufferCache::ChangeRegister(BufferId id) {
 		EXIT_IF(buffer.Size() > m_total_used_memory);
 		m_total_used_memory -= buffer.Size();
 		m_lru_cache.Free(buffer.lru_id);
-		m_bda_pagetable_buffer.Fill(pages.first * sizeof(vk::DeviceAddress),
+		m_bda_pagetable_buffer.Fill(table_offset,
 		                            size_pages * sizeof(vk::DeviceAddress), 0);
 		buffer.is_deleted = true;
 	}
@@ -108,6 +110,7 @@ void BufferCache::DeleteBuffer(BufferId id) {
 	}
 }
 
+template <bool async>
 bool BufferCache::DownloadBufferMemory(Buffer& buffer, uint64_t vaddr, uint64_t size) {
 	std::vector<vk::BufferCopy> copies;
 	uint64_t                    total_size     = 0;
@@ -127,11 +130,16 @@ bool BufferCache::DownloadBufferMemory(Buffer& buffer, uint64_t vaddr, uint64_t
 		return false;
 	}
 
-	const auto [mapped, offset] = m_download_buffer.Map(total_size, 64);
+	auto [mapped, offset] = m_download_buffer.Map(total_size, 64);
+	std::unique_ptr<Buffer> temporary;
 	if (mapped == nullptr) {
-		EXIT("BufferCache: download exceeds 64 MiB staging buffer capacity\n");
+		temporary = std::make_unique<Buffer>(m_graphics, m_scheduler, MemoryUsage::Download, 0,
+		                                     vk::BufferUsageFlagBits::eTransferDst, total_size);
+		mapped = temporary->Mapped().data();
+	} else {
+		m_download_buffer.Commit();
 	}
-	m_download_buffer.Commit();
+	const auto& download = temporary ? *temporary : m_download_buffer;
 	for (auto& copy: copies) {
 		copy.dstOffset += offset;
 	}
@@ -150,27 +158,35 @@ bool BufferCache::DownloadBufferMemory(Buffer& buffer, uint64_t vaddr, uint64_t
 	native.pipelineBarrier(vk::PipelineStageFlagBits::eAllCommands,
 	                       vk::PipelineStageFlagBits::eTransfer, {}, 0, nullptr, 1, &before, 0,
 	                       nullptr);
-	native.copyBuffer(buffer.Handle(), m_download_buffer.Handle(),
+	native.copyBuffer(buffer.Handle(), download.Handle(),
 	                  static_cast<uint32_t>(copies.size()), copies.data());
 
 	auto after          = before;
 	after.srcAccessMask = vk::AccessFlagBits::eTransferWrite;
 	after.dstAccessMask = vk::AccessFlagBits::eHostRead;
-	after.buffer        = m_download_buffer.Handle();
+	after.buffer        = download.Handle();
 	after.offset        = offset;
 	after.size          = total_size;
 	native.pipelineBarrier(vk::PipelineStageFlagBits::eTransfer,
 	                       vk::PipelineStageFlagBits::eAllCommands |
 	                           vk::PipelineStageFlagBits::eHost,
 	                       {}, 0, nullptr, 1, &after, 0, nullptr);
-	m_scheduler.DeferPriorityOperation([this, mapped, offset, total_size, buffer_address,
-	                                    copies = std::move(copies)] {
-		m_download_buffer.Invalidate(offset, total_size);
+	auto publish = [this, mapped, offset, total_size, buffer_address,
+	                copies = std::move(copies), owner = std::move(temporary)] {
+		(owner ? *owner : m_download_buffer).Invalidate(offset, total_size);
 		for (const auto& copy: copies) {
 			Libs::LibKernel::Memory::WriteBacking(buffer_address + copy.srcOffset,
 			                                      mapped + (copy.dstOffset - offset), copy.size);
 		}
-	});
+	};
+	if constexpr (async) {
+		m_scheduler.DeferPriorityOperation(std::move(publish));
+	} else {
+		const auto tick = m_scheduler.CurrentTick();
+		m_scheduler.Wait(tick);
+		m_scheduler.WaitPriorityOperations(tick);
+		publish();
+	}
 	return true;
 }
 
@@ -249,10 +265,7 @@ void BufferCache::ReadMemory(uint64_t vaddr, uint64_t size, bool is_write) {
 		const auto window_begin = std::max(Common::AlignDown(vaddr, WindowSize), buffer_begin);
 		const auto window_end = std::min(std::max(window_begin + WindowSize, vaddr + size), buffer_end);
 
-		if (DownloadBufferMemory(buffer, window_begin, window_end - window_begin)) {
-			const auto tick = m_scheduler.CurrentTick();
-			m_scheduler.Wait(tick);
-			m_scheduler.WaitPriorityOperations(tick);
+		if (DownloadBufferMemory<false>(buffer, window_begin, window_end - window_begin)) {
 			m_memory_tracker.UnmarkRegionAsGpuModified(window_begin, window_end - window_begin);
 		}
 		if (is_write) {
@@ -312,10 +325,14 @@ BufferCache::OverlapResult BufferCache::ResolveOverlaps(uint64_t vaddr, uint64_t
 			// Reserve space in the incoming stream's direction of growth.
 			// The old buffer extending left of the request predicts growth to the right, and vice versa.
 			if (expands_left) {
-				end += std::min(StreamLeapSize, PageTable::kAddressSpaceSize - end);
+				end += std::min(StreamLeapSize, (vaddr < LOWER_ADDRESS_SIZE ? LOWER_ADDRESS_SIZE
+				                                       : LibKernel::Memory::kExtendedMemoryBase +
+				                                             LibKernel::Memory::kExtendedMemorySize) - end);
 			}
 			if (expands_right) {
-				const auto minimum = CACHING_PAGESIZE * 2;
+				const auto minimum = vaddr < LOWER_ADDRESS_SIZE
+				                         ? CACHING_PAGESIZE * 2
+				                         : LibKernel::Memory::kExtendedMemoryBase;
 				if (begin > minimum) {
 					begin -= std::min(StreamLeapSize, begin - minimum);
 				}
@@ -448,7 +465,8 @@ std::pair<Buffer*, uint64_t> BufferCache::ObtainBuffer(uint64_t vaddr, uint64_t
 		const auto alignment = std::max<uint64_t>(
 		    m_graphics.physical_device_properties.limits.minUniformBufferOffsetAlignment, 1);
 		auto [mapped, offset] = m_stream_buffer.Map(size, alignment, false);
-		if (mapped != nullptr && Libs::LibKernel::Memory::TryReadBacking(vaddr, mapped, size)) {
+		if (mapped != nullptr) {
+			std::memcpy(mapped, reinterpret_cast<const void*>(vaddr), size);
 			m_stream_buffer.Commit();
 			return {&m_stream_buffer, offset};
 		}
@@ -484,8 +502,7 @@ std::pair<Buffer*, uint64_t> BufferCache::ObtainBufferForImage(uint64_t vaddr, u
 	}
 
 	auto [staging, stage_offset] = m_staging_buffer.Map(size, 16);
-	if (staging == nullptr || (!Libs::LibKernel::Memory::TryReadBacking(vaddr, staging, size) &&
-	                           !Libs::LibKernel::Memory::TryReadPrtBacking(vaddr, staging, size))) {
+	if (staging == nullptr || !Libs::LibKernel::Memory::TryReadSparseBacking(vaddr, staging, size)) {
 		EXIT("BufferCache: failed to read mapped guest image backing\n");
 	}
 	m_staging_buffer.Commit();
@@ -603,7 +620,7 @@ void BufferCache::RunGarbageCollector() {
 			return false;
 		}
 		if (dirty) {
-			EXIT_IF(!DownloadBufferMemory(buffer, buffer.CpuAddress(), buffer.Size()));
+			EXIT_NOT_IMPLEMENTED(!DownloadBufferMemory<true>(buffer, buffer.CpuAddress(), buffer.Size()));
 			dirty_buffers.push_back(id);
 		} else {
 			m_memory_tracker.UntrackMemory(buffer.CpuAddress(), buffer.Size());

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

@@ -29,10 +29,22 @@ class BufferCache {
 public:
 	static constexpr uint32_t CACHING_PAGEBITS  = 14;
 	static constexpr uint64_t CACHING_PAGESIZE  = uint64_t {1} << CACHING_PAGEBITS;
-	static constexpr uint64_t CACHING_NUMPAGES  = uint64_t {1} << (40 - CACHING_PAGEBITS);
+	static constexpr uint64_t CACHING_NUMPAGES  = (LOWER_ADDRESS_SIZE + LibKernel::Memory::kExtendedMemorySize) >> CACHING_PAGEBITS;
 	static constexpr uint64_t BDA_PAGETABLE_SIZE =
 	    CACHING_NUMPAGES * sizeof(vk::DeviceAddress);
 
+	static constexpr uint64_t PageIndex(uint64_t address) {
+		return (address < LOWER_ADDRESS_SIZE
+		            ? address
+		            : address - LibKernel::Memory::kExtendedMemoryBase + LOWER_ADDRESS_SIZE) >>
+		       CACHING_PAGEBITS;
+	}
+	static constexpr uint64_t GuestAddress(uint64_t offset) {
+		return offset < LOWER_ADDRESS_SIZE
+		           ? offset
+		           : offset - LOWER_ADDRESS_SIZE + LibKernel::Memory::kExtendedMemoryBase;
+	}
+
 	BufferCache(GraphicContext& graphics, CommandScheduler& scheduler, PageManager& page_manager,
 	            TextureCache& texture_cache);
 	~BufferCache();
@@ -88,7 +100,7 @@ private:
 		bool                has_stream_leap;
 	};
 
-	using PageTable = MultiLevelPageTable<BufferId, CACHING_PAGEBITS, 40, 16>;
+	using PageTable = MultiLevelPageTable<BufferId, CACHING_PAGEBITS, 44, 20>;
 	static_assert(CACHING_PAGESIZE == (uint64_t {1} << PageTable::kPageBits));
 	void WriteDataBuffer(Buffer& buffer, uint64_t address, const void* source, uint64_t size);
 	void TouchBuffer(const Buffer& buffer);
@@ -105,7 +117,8 @@ private:
 	[[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);
-	// Queues backing publication; callers wait before clearing dirty pages or reusing their data.
+	// Synchronous downloads publish before returning; asynchronous callers wait before reuse.
+	template <bool async>
 	[[nodiscard]] bool DownloadBufferMemory(Buffer& buffer, uint64_t vaddr, uint64_t size);
 
 	GraphicContext&                                   m_graphics;

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

@@ -134,8 +134,9 @@ void FaultManager::ProcessFaultBuffer() {
 		const auto* faults = std::bit_cast<const uint64_t*>(mapped);
 		const auto  count  = static_cast<uint32_t>(faults[0]);
 		for (uint32_t index = 1; index <= count; ++index) {
-			fault_ranges.Add(faults[index], BufferCache::CACHING_PAGESIZE);
-			LOGF("Accessed non-GPU cached memory at 0x%016" PRIx64 "\n", faults[index]);
+			const auto address = BufferCache::GuestAddress(faults[index]);
+			fault_ranges.Add(address, BufferCache::CACHING_PAGESIZE);
+			LOGF("Accessed non-GPU cached memory at 0x%016" PRIx64 "\n", address);
 		}
 		fault_ranges.ForEach([this](uint64_t start, uint64_t end) {
 			EXIT_IF(end - start > std::numeric_limits<uint32_t>::max());

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

@@ -14,10 +14,10 @@ SamplerCache::~SamplerCache() {
 	}
 }
 
-vk::Sampler SamplerCache::GetSampler(const ShaderSamplerResource& r) {
+vk::Sampler SamplerCache::GetSampler(const ShaderSamplerResource& r, bool integer_border) {
 	Common::LockGuard lock(m_mutex);
 
-	const SamplerKey key {r.fields[0], r.fields[1], r.fields[2], r.fields[3]};
+	const SamplerKey key {r.fields[0], r.fields[1], r.fields[2], r.fields[3], integer_border};
 	if (auto iter = m_samplers.find(key); iter != m_samplers.end()) {
 		return iter->second;
 	}
@@ -55,11 +55,9 @@ vk::Sampler SamplerCache::GetSampler(const ShaderSamplerResource& r) {
 			case Prospero::SamplerAnisoRatio::kTwo: aniso_ratio = 2.0f; break;
 			case Prospero::SamplerAnisoRatio::kFour: aniso_ratio = 4.0f; break;
 			case Prospero::SamplerAnisoRatio::kEight: aniso_ratio = 8.0f; break;
-			case Prospero::SamplerAnisoRatio::kSixteen: aniso_ratio = 16.0f; break;
-			default:
-				EXIT("unknown ratio: %d dwords=%08x,%08x,%08x,%08x\n",
-				     static_cast<int>(r.MaxAnisoRatio()), r.fields[0], r.fields[1], r.fields[2],
-				     r.fields[3]);
+			case Prospero::SamplerAnisoRatio::kSixteen:
+			// Reserved encodings 5-7 fall back to the maximum ratio.
+			default: aniso_ratio = 16.0f; break;
 		}
 	}
 
@@ -95,23 +93,23 @@ vk::Sampler SamplerCache::GetSampler(const ShaderSamplerResource& r) {
 		return vk::SamplerAddressMode::eClampToBorder;
 	};
 
-	vk::BorderColor border = vk::BorderColor::eIntTransparentBlack;
+	vk::BorderColor border = integer_border ? vk::BorderColor::eIntTransparentBlack
+	                                       : vk::BorderColor::eFloatTransparentBlack;
 	switch (static_cast<Prospero::SamplerBorderColor>(r.BorderColorType())) {
-		case Prospero::SamplerBorderColor::kTransBlack:
-			border = vk::BorderColor::eIntTransparentBlack;
-			break;
+		case Prospero::SamplerBorderColor::kTransBlack: break;
 		case Prospero::SamplerBorderColor::kOpaqueBlack:
-			border = vk::BorderColor::eIntOpaqueBlack;
+			border = integer_border ? vk::BorderColor::eIntOpaqueBlack
+			                        : vk::BorderColor::eFloatOpaqueBlack;
 			break;
 		case Prospero::SamplerBorderColor::kOpaqueWhite:
-			border = vk::BorderColor::eIntOpaqueWhite;
+			border = integer_border ? vk::BorderColor::eIntOpaqueWhite
+			                        : vk::BorderColor::eFloatOpaqueWhite;
 			break;
 		case Prospero::SamplerBorderColor::kFromTable:
 			LOGF(
 			    "temporary: approximating table border color as transparent black, index = %" PRIu16
 			    "\n",
 			    r.BorderColorPtr());
-			border = vk::BorderColor::eIntTransparentBlack;
 			break;
 		default: EXIT("unknown border color: %d", static_cast<int>(r.BorderColorType()));
 	}
@@ -125,8 +123,11 @@ vk::Sampler SamplerCache::GetSampler(const ShaderSamplerResource& r) {
 	sampler_info.addressModeU = to_vk_address_mode(r.ClampX());
 	sampler_info.addressModeV = to_vk_address_mode(r.ClampY());
 	sampler_info.addressModeW = to_vk_address_mode(r.ClampZ());
-	sampler_info.mipLodBias =
-	    static_cast<float>(static_cast<int16_t>((r.LodBias() ^ 0x2000u) - 0x2000u)) / 256.0f;
+    const auto max_lod_bias = m_graphics.GetPhysicalDeviceProperties().limits.maxSamplerLodBias;
+    sampler_info.mipLodBias = std::clamp(
+        static_cast<float>(static_cast<int16_t>((r.LodBias() ^ 0x2000u) - 0x2000u)) / 256.0f,
+        -max_lod_bias, max_lod_bias
+    );
 	sampler_info.anisotropyEnable        = (aniso ? VK_TRUE : VK_FALSE);
 	sampler_info.maxAnisotropy           = aniso_ratio;
 	sampler_info.compareEnable           = (r.DepthCompareFunc() != 0 ? VK_TRUE : VK_FALSE);

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

@@ -24,10 +24,10 @@ public:
 	~SamplerCache();
 	KYTY_CLASS_NO_COPY(SamplerCache);
 
-	vk::Sampler GetSampler(const ShaderSamplerResource& r);
+	vk::Sampler GetSampler(const ShaderSamplerResource& r, bool integer_border);
 
 private:
-	using SamplerKey = std::array<uint32_t, 4>;
+	using SamplerKey = std::array<uint32_t, 5>;
 
 	struct SamplerKeyHash {
 		std::size_t operator()(const SamplerKey& key) const {

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

@@ -33,24 +33,27 @@ namespace {
 
 constexpr uint64_t NumFramesBeforeRemoval = 32;
 
-[[nodiscard]] bool DecodeDccClear(const TextureCache::ImageDesc& desc, uint8_t code,
+[[nodiscard]] bool DecodeColorClear(const TextureCache::ImageDesc& desc, uint8_t code,
                                   vk::ClearColorValue& clear) {
+	const auto& metadata = desc.info.metadata;
+	const auto  format   = desc.view_info.format;
+	const bool  cmask    = metadata.kind == ImageMetadataKind::Cmask;
+	if (cmask ? code == 0 : code == 0x20) {
+		// Register clears belong to the color buffer; the texture pipe cannot decode them.
+		return desc.type == TextureCache::BindingType::RenderTarget &&
+		       metadata.clear_register_valid &&
+		       DecodePackedColorClear(format, metadata.clear_word, clear);
+	}
+	if (cmask) {
+		return false;
+	}
 	switch (code) {
 		case 0x00:
-		case 0x20:
 		case 0x40:
 		case 0x80:
 		case 0xc0: break;
 		default: return false;
 	}
-	const auto& metadata = desc.info.metadata;
-	const auto  format   = desc.view_info.format;
-	if (code == 0x20) {
-		// Clear-to-register is a color-buffer operation; the texture pipe cannot decode it.
-		return desc.type == TextureCache::BindingType::RenderTarget &&
-		       metadata.dcc_clear_register_valid &&
-		       DecodePackedColorClear(format, metadata.dcc_clear_word, clear);
-	}
 	clear = {};
 	if (code == 0x00) {
 		return true;
@@ -193,7 +196,7 @@ bool TextureCache::SameBacking(const ImageInfo& cached, const ImageInfo& request
 		return false;
 	}
 	if (!ImageViewOps::FormatsCompatible(cached.pixel_format, requested.pixel_format) ||
-	    (cached.type != requested.type && requested.extent != vk::Extent3D {1, 1, 1})) {
+	    cached.type != requested.type) {
 		return false;
 	}
 	if (exact_format && cached.pixel_format != requested.pixel_format) {
@@ -219,14 +222,6 @@ TextureCache::BindingType TextureCache::UploadBinding(const Image& image) {
 	return image.usage.storage ? BindingType::Storage : BindingType::Texture;
 }
 
-bool TextureCache::SafeToDownload(const Image& image) {
-	if (!image.SafeToDownload()) {
-		return false;
-	}
-	const auto range = image.info.data;
-	return !m_buffer_cache.HasGpuDirtyBytes(range.address, range.size);
-}
-
 ImageId TextureCache::InsertImage(const ImageInfo& info) {
 	const auto id = m_slot_images.insert(m_graphics, m_scheduler, info);
 	if (!info.data.Empty()) {
@@ -588,7 +583,7 @@ bool TextureCache::CopyD16(Image& destination, Image& source) {
 	}
 
 	auto                         depth_buffer = m_tiler.GetScratchBuffer(depth_size);
-	auto                         color_buffer = m_tiler.GetScratchBuffer(color_size);
+	auto                         color_buffer = m_tiler.GetScratchBuffer(color_size, depth_buffer.buffer);
 	const TileManager::D16Layout promote_layout {
 	    .width               = depth.info.extent.width,
 	    .height              = depth.info.extent.height,
@@ -639,9 +634,12 @@ void TextureCache::CopyImage(ImageId destination_id, ImageId source_id) {
 	const bool source_depth = source.info.IsDepth();
 	const bool dest_depth   = destination.info.IsDepth();
 	const bool direct_copy =
-	    source.backing.format == destination.backing.format ||
-	    (!source_depth && !dest_depth &&
-	     vk::blockSize(source.backing.format) == vk::blockSize(destination.backing.format));
+	    (source.backing.image_type == destination.backing.image_type ||
+	     (source.backing.image_type != vk::ImageType::e1D &&
+	      destination.backing.image_type != vk::ImageType::e1D)) &&
+	    (source.backing.format == destination.backing.format ||
+	     (!source_depth && !dest_depth &&
+	      vk::blockSize(source.backing.format) == vk::blockSize(destination.backing.format)));
 	if (direct_copy) {
 		destination.CopyImage(source);
 	} else if (!CopyD16(destination, source)) {
@@ -675,7 +673,8 @@ void TextureCache::CopyImageMip(ImageId destination_id, ImageId source_id, uint3
 ImageId TextureCache::ResolveDepthOverlap(const ImageInfo& requested, BindingType binding,
                                           ImageId cached_id) {
 	auto& cached = m_slot_images[cached_id];
-	if (!cached.info.IsDepth() && !requested.IsDepth()) {
+	if ((!cached.info.IsDepth() && !requested.IsDepth()) ||
+	    cached.info.tile_mode != requested.tile_mode) {
 		return {};
 	}
 	const bool stencil_match = requested.HasStencil() == cached.info.HasStencil();
@@ -689,8 +688,8 @@ ImageId TextureCache::ResolveDepthOverlap(const ImageInfo& requested, BindingTyp
 	    requested.samples == 1 && requested.data == cached.info.data &&
 	    requested.extent == cached.info.extent && requested.resources == cached.info.resources &&
 	    requested.type == cached.info.type && requested.pitch == cached.info.pitch &&
-	    requested.tile_mode == cached.info.tile_mode && !requested.HasStencil() &&
-	    !cached.info.HasStencil() && !requested.HasMetadata() && !cached.info.HasMetadata();
+	    !requested.HasStencil() && !cached.info.HasStencil() && !requested.HasMetadata() &&
+	    !cached.info.HasMetadata();
 	// PPSA04264
 	const bool retain_cached_layout =
 	    requested.samples == 1 && cached.info.samples == 1 && cached.backing.samples == 1 &&
@@ -701,7 +700,7 @@ ImageId TextureCache::ResolveDepthOverlap(const ImageInfo& requested, BindingTyp
 	    requested.resources.layers != 0 && cached.info.resources.layers != 0 &&
 	    requested.resources.layers < cached.info.resources.layers &&
 	    requested.type == cached.info.type && requested.pitch == cached.info.pitch &&
-	    requested.tile_mode == cached.info.tile_mode && requested.mip_layout[0].offset == 0 &&
+	    requested.mip_layout[0].offset == 0 &&
 	    cached.info.mip_layout[0].offset == 0 &&
 	    requested.mip_layout[0].size == requested.data.size &&
 	    cached.info.mip_layout[0].size == cached.info.data.size &&
@@ -810,12 +809,15 @@ TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& reques
 			return {merged_id};
 		}
 		// PPSA08394
+		// A view cannot change the native image type or grow its extent.
 		if (requested.data.size == cached.info.data.size &&
-		    requested.resources == cached.info.resources && requested.type == cached.info.type &&
-		    requested.extent.width > cached.info.extent.width &&
-		    requested.extent.height >= cached.info.extent.height &&
-		    requested.extent.depth >= cached.info.extent.depth &&
-		    ImageViewOps::FormatsCompatible(cached.info.pixel_format, requested.pixel_format)) {
+		    requested.resources == cached.info.resources &&
+		    ImageViewOps::FormatsCompatible(cached.info.pixel_format, requested.pixel_format) &&
+		    (requested.type != cached.info.type
+		         ? requested.extent == cached.info.extent
+		         : requested.extent.width > cached.info.extent.width &&
+		               requested.extent.height >= cached.info.extent.height &&
+		               requested.extent.depth >= cached.info.extent.depth)) {
 			return {ExpandImage(requested, cached_id)};
 		}
 		// PS5 mip tails can expose more levels without increasing the guest allocation.
@@ -1069,7 +1071,7 @@ void TextureCache::UploadImage(Image& image, Buffer& source, uint64_t source_off
 		                     texels_per_slice > UINT64_MAX / transfer_bytes);
 		const uint64_t transfer_slice = texels_per_slice * transfer_bytes;
 		EXIT_NOT_IMPLEMENTED(transfer_slice > UINT64_MAX / layers);
-		auto promoted = m_tiler.GetScratchBuffer(transfer_slice * layers);
+		auto promoted = m_tiler.GetScratchBuffer(transfer_slice * layers, linear.buffer);
 		m_tiler.ConvertD16(
 		    linear, promoted, TileManager::D16Direction::Promote,
 		    info.pixel_format == vk::Format::eD32SfloatS8Uint,
@@ -1118,9 +1120,10 @@ void TextureCache::InitializeImage(ImageId id) {
 	}
 }
 
-void TextureCache::MaterializeDccClear(ImageId id, const ImageDesc& desc,
+void TextureCache::MaterializeColorClear(ImageId id, const ImageDesc& desc,
                                        uint32_t metadata_base_layer) {
-	if (desc.info.metadata.kind != ImageMetadataKind::Dcc) {
+	if (desc.info.metadata.kind != ImageMetadataKind::Dcc &&
+	    desc.info.metadata.kind != ImageMetadataKind::Cmask) {
 		return;
 	}
 	const auto range = desc.info.metadata.range;
@@ -1128,18 +1131,18 @@ void TextureCache::MaterializeDccClear(ImageId id, const ImageDesc& desc,
 		std::scoped_lock lock {m_lock};
 		auto& image         = m_slot_images[id];
 		image.info.metadata = desc.info.metadata;
-		// A native DCC allocation must not retain a reused HTile/CMask/FMask interpretation.
+		// Native color metadata must not retain a reused HTile/CMask/FMask clear flag.
 		m_surface_metas.erase(range.address);
 		if (range.size == 0 || desc.info.resources.levels != 1 || image.info.resources.levels != 1) {
 			return;
 		}
 	}
 	const auto layers = desc.info.TransferLayers();
-	// These one-mip surfaces use complete 4 KiB DCC metadata blocks.
+	// These one-mip surfaces use complete 4 KiB color metadata blocks.
 	constexpr uint64_t MetadataBlockSize = 0x1000;
 	if (!range.Valid() || range.address % MetadataBlockSize != 0 || layers == 0 ||
 	    range.size % layers != 0 || (range.size / layers) % MetadataBlockSize != 0) {
-		EXIT("TextureCache: DCC slices must contain aligned 4 KiB blocks\n");
+		EXIT("TextureCache: color metadata slices must contain aligned 4 KiB blocks\n");
 	}
 	const auto& view           = desc.view_info;
 	const bool  volume_texture = desc.info.IsVolume() && view.type == vk::ImageViewType::e3D;
@@ -1147,7 +1150,7 @@ void TextureCache::MaterializeDccClear(ImageId id, const ImageDesc& desc,
 	const auto  image_first    = volume_texture ? 0u : view.base_layer;
 	const auto  count          = volume_texture ? desc.info.extent.depth : view.layer_count;
 	if (first >= layers || count > layers - first) {
-		EXIT("TextureCache: DCC view exceeds its native metadata slices\n");
+		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.
@@ -1159,15 +1162,15 @@ void TextureCache::MaterializeDccClear(ImageId id, const ImageDesc& desc,
 		const auto address = range.address + slice_size * (first + slice);
 		uint8_t code = 0;
 		if (!LibKernel::Memory::TryReadBacking(address, &code, sizeof(code))) {
-			EXIT("TextureCache: failed to read DCC metadata backing\n");
+			EXIT("TextureCache: failed to read color metadata backing\n");
 		}
 		vk::ClearValue clear {};
-		if (!DecodeDccClear(desc, code, clear.color)) {
+		if (!DecodeColorClear(desc, code, clear.color)) {
 			continue;
 		}
 		std::vector<uint8_t> bytes(slice_size);
 		if (!LibKernel::Memory::TryReadBacking(address, bytes.data(), bytes.size())) {
-			EXIT("TextureCache: failed to read DCC metadata slice\n");
+			EXIT("TextureCache: failed to read color metadata slice\n");
 		}
 		if (!std::all_of(bytes.begin(), bytes.end(), [code](uint8_t byte) { return byte == code; })) {
 			continue;
@@ -1312,7 +1315,7 @@ ImageId TextureCache::FindImage(ImageDesc& desc, bool exact_format) {
 		image.tick_accessed_last = m_scheduler.CurrentTick();
 		TouchImage(image);
 	}
-	MaterializeDccClear(result, desc, metadata_base_layer);
+	MaterializeColorClear(result, desc, metadata_base_layer);
 	if (desc.type == BindingType::VideoOut &&
 	    desc.info.metadata.compression != VideoOutCompression::Uncompressed) {
 		std::scoped_lock lock {m_lock};
@@ -1349,7 +1352,7 @@ ImageId TextureCache::FindImageFromRange(uint64_t address, uint64_t size, bool e
 		if (ensure_valid && owner->depth_id) {
 			owner = m_slot_images.try_get(owner->depth_id);
 		}
-		if (owner == nullptr || (ensure_valid && !SafeToDownload(*owner))) {
+		if (owner == nullptr || (ensure_valid && !owner->SafeToDownload())) {
 			continue;
 		}
 		matches.push_back(id);
@@ -1530,7 +1533,7 @@ bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address
 	auto&          image = m_slot_images[selected];
 	vk::ClearValue clear {};
 	if (aspect == vk::ImageAspectFlagBits::eColor) {
-		if (!DecodePackedColorClear(image.info.pixel_format, packed_clear, clear.color)) {
+		if (!DecodeColorDwordFill(image.info.pixel_format, packed_clear, clear.color)) {
 			return false;
 		}
 	} else {
@@ -1676,7 +1679,7 @@ void TextureCache::DownloadDepth(Image& image, Buffer& destination, uint64_t des
 	auto host_linear = m_tiler.GetScratchBuffer(transfer_size);
 	image.Download(copies, host_linear.buffer, 0, host_linear.size);
 	const bool tiled        = info.IsTiled();
-	auto       guest_linear = tiled ? m_tiler.GetScratchBuffer(info.data.size)
+	auto       guest_linear = tiled ? m_tiler.GetScratchBuffer(info.data.size, host_linear.buffer)
 	                                : TileManager::Result {destination.Handle(), destination_offset,
 	                                                       destination.Size() - destination_offset};
 	m_tiler.ConvertD16(host_linear, guest_linear, TileManager::D16Direction::Demote,
@@ -1740,7 +1743,7 @@ bool BufferCache::SynchronizeBufferFromImage(Buffer& buffer, uint64_t vaddr, uin
 	std::scoped_lock lock {m_texture_cache.m_lock};
 	auto& image = m_texture_cache.m_slot_images[selected];
 	// The GPU thread owns image retirement; CPU invalidation can dirty this image after lookup.
-	if (!m_texture_cache.SafeToDownload(image)) {
+	if (!image.SafeToDownload()) {
 		return false;
 	}
 	if (!buffer.IsInBounds(image.info.data.address, 1)) {
@@ -1803,7 +1806,7 @@ bool TextureCache::DownloadImageMemory(ImageId id) {
 		return false;
 	}
 	auto transfer = BuildDownload(image);
-	if (!transfer.valid || !SafeToDownload(image)) {
+	if (!transfer.valid || !image.SafeToDownload()) {
 		return false;
 	}
 	const auto range    = image.info.data;
@@ -1864,7 +1867,10 @@ bool TextureCache::IsRegionGpuModified(uint64_t address, uint64_t size) {
 	std::scoped_lock lock {m_lock};
 	for (const auto id: FindImagesInRegion(address, size, false)) {
 		const auto& image = m_slot_images[id];
-		if (!image.depth_id && image.IsGpuModified()) {
+		// PPSA17168: S_LOAD_DWORD reads shader data at an address overlapping an old
+		// render target whose memory the CPU has reused. The cached image still retains
+		// its earlier GPU-modified flag.
+		if (!image.depth_id && image.IsGpuModified() && !image.IsDefinitelyCpuDirty()) {
 			return true;
 		}
 	}
@@ -1990,7 +1996,7 @@ void TextureCache::RunGarbageCollector() {
 				continue;
 			}
 			if (owner->IsGpuModified()) {
-				const bool safe = SafeToDownload(*owner);
+				const bool safe = owner->SafeToDownload();
 				if (safe && owner->info.IsTiled()) {
 					continue;
 				}

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

@@ -91,9 +91,9 @@ private:
 	};
 
 	using ImageIds       = InlinePageOwnerList<ImageId, 16>;
-	using ImagePageTable = MultiLevelPageTable<ImageIds, 20, 40, 10>;
+	using ImagePageTable = MultiLevelPageTable<ImageIds, 20, 44, 14>;
 
-	// Callers have validated the nonempty 40-bit range with TryGetPageRange.
+	// Callers have validated the nonempty 44-bit range with TryGetPageRange.
 	template <typename Func>
 	static void ForEachPage(uint64_t address, size_t size, Func&& func) {
 		using FuncReturn = typename std::invoke_result<Func, uint64_t>::type;
@@ -128,7 +128,6 @@ private:
 	[[nodiscard]] static bool SameBacking(const ImageInfo& cached, const ImageInfo& requested,
 	                                      bool exact_format);
 	[[nodiscard]] static BindingType UploadBinding(const Image& image);
-	[[nodiscard]] bool               SafeToDownload(const Image& image);
 
 	// Caller holds m_lock; it also serializes the per-image query epoch.
 	[[nodiscard]] ImageIds      FindImagesInRegion(uint64_t address, uint64_t size,
@@ -139,7 +138,7 @@ private:
 	                                                ImageId cached);
 	[[nodiscard]] ImageId       ExpandImage(const ImageInfo& info, ImageId source);
 	void                        RefreshImage(ImageId id);
-	void                        MaterializeDccClear(ImageId id, const ImageDesc& desc,
+	void                        MaterializeColorClear(ImageId id, const ImageDesc& desc,
 	                                                uint32_t metadata_base_layer);
 	void                        InitializeImage(ImageId id);
 	[[nodiscard]] TextureTransfer

src/graphics/host_gpu/renderer/colorRenderTarget.cpp

@@ -17,11 +17,10 @@
 #include <algorithm>
 #include <array>
 #include <atomic>
+#include <bit>
 
 namespace Libs::Graphics {
 
-static std::atomic<uint32_t> g_render_color_log_count = 0;
-
 static bool DccAlphaOnMsb(const HW::ColorInfo& info) {
 	switch (info.format) {
 		case Prospero::ChannelLayout::k10_10_10_2:
@@ -138,8 +137,7 @@ void RenderExecutor::ResolveRenderColorTarget(CommandBuffer& buffer, RenderColor
 	// Color-control state selects the color-buffer operation and logical blend operation.
 	// The normal copy operation is a regular color write, not an attachment clear.
 	// Nonlinear clear values are still stored as normalized components.
-	// Fast color clears are metadata driven and must be handled explicitly when
-	// that metadata path is implemented; render-pass load must preserve contents.
+	// Metadata clears are materialized during image discovery; render-pass loads preserve contents.
 	uint32_t   width  = 0;
 	uint32_t   height = 0;
 	uint32_t   pitch  = 0;
@@ -160,9 +158,6 @@ void RenderExecutor::ResolveRenderColorTarget(CommandBuffer& buffer, RenderColor
 			break;
 		default: EXIT("unknown tile mode: %u\n", static_cast<uint32_t>(rt.attrib3.tile_mode));
 	}
-	if (!tile && levels > 1) {
-		EXIT("linear mipmapped render targets are unsupported\n");
-	}
 	if (samples > 1 && (!tile || levels != 1)) {
 		EXIT("multisampled render targets require a single-mip tiled surface\n");
 	}
@@ -203,8 +198,7 @@ void RenderExecutor::ResolveRenderColorTarget(CommandBuffer& buffer, RenderColor
 		     " layer=%u/%u\n",
 		     rt.attrib3.dimension, rt.attrib3.depth, view.base_layer, view.image_layers);
 	}
-	// PPSA28068: the linear EULA target and its sampled view must share the padded pitch.
-	if (!tile || volume || texture_tile) {
+	if (samples == 1) {
 		pitch = TileGetTexturePitch(transfer_format, width, rt.attrib3.tile_mode);
 	} else {
 		pitch = TileGetRenderTargetPitch(width, bytes_per_element, rt.attrib.num_fragments);
@@ -232,37 +226,26 @@ void RenderExecutor::ResolveRenderColorTarget(CommandBuffer& buffer, RenderColor
 		}
 		size         = volume_layout.block_slice_size;
 		backing_size = volume_layout.total_size;
-	} else if (tile) {
+	} else {
 		TileSizeAlign layout {};
 		bool          valid_layout = false;
-		if (texture_tile) {
+		if (samples == 1) {
 			TileGetTextureSize(transfer_format, width, height, levels, rt.attrib3.tile_mode,
 			                   &layout, mip_sizes, mip_padded);
-			valid_layout = layout.size != 0 && layout.align == texture_tile_layout.block.block_size;
+			valid_layout = layout.size != 0 && layout.align != 0 &&
+			               (rt.attrib3.tile_mode != Prospero::TileMode::kRenderTarget ||
+			                levels <= std::bit_width(std::max(width, height)));
 		} else {
-			valid_layout =
-			    levels == 1 ? TileGetRenderTargetSize(width, height, pitch, bytes_per_element,
-			                                          layout, rt.attrib.num_fragments)
-			                : TileGetRenderTargetMipLayout(width, height, pitch, bytes_per_element,
-			                                               levels, layout, mip_sizes, mip_padded);
+			valid_layout = TileGetRenderTargetSize(width, height, pitch, bytes_per_element,
+			                                       layout, rt.attrib.num_fragments);
+			mip_sizes[0]  = {layout.size, 0, 0, 0, 0, 0};
+			mip_padded[0] = {pitch, height};
 		}
 		if (!valid_layout) {
 			EXIT("unsupported render-target layout: %ux%u pitch=%u bytes=%u levels=%u\n", width,
 			     height, pitch, bytes_per_element, levels);
 		}
 		size = layout.size;
-		EXIT_IF(size > UINT32_MAX);
-		if (levels == 1) {
-			mip_sizes[0]  = {static_cast<uint32_t>(size), 0, 0, 0, 0, 0};
-			mip_padded[0] = {pitch, height};
-		}
-	} else {
-		size = static_cast<uint64_t>(pitch) * height * bytes_per_element * samples;
-		if (size > UINT32_MAX) {
-			EXIT("linear render-target slice exceeds the supported layout size\n");
-		}
-		mip_sizes[0]  = {static_cast<uint32_t>(size), 0, 0, 0, 0, 0};
-		mip_padded[0] = {pitch, height};
 	}
 	if (size == 0 || (!volume && size > UINT64_MAX / view.image_layers)) {
 		EXIT("render-target memory footprint is invalid\n");
@@ -277,21 +260,7 @@ void RenderExecutor::ResolveRenderColorTarget(CommandBuffer& buffer, RenderColor
 		EXIT("render-target backing range is invalid\n");
 	}
 
-	const vk::Extent2D view_extent = {std::max(width >> rt.view.current_mip_level, 1u),
-	                                  std::max(height >> rt.view.current_mip_level, 1u)};
-
-	auto decision_log_id = g_render_color_log_count.fetch_add(1);
-	if (decision_log_id < 128) {
-		LOGF("RenderColorTarget: slot=%" PRIu32 " addr=0x%010" PRIx64 " size=0x%016" PRIx64
-		     " extent=%ux%ux%u view_mip=%u view_extent=%ux%u levels=%u pitch=%u"
-		     " fmt=0x%08" PRIx32 " nfmt=0x%08" PRIx32 " order=0x%08" PRIx32 " samples=%u tile=%s\n",
-		     rt_slot, rt.base.addr, backing_size, width, height, depth, rt.view.current_mip_level,
-		     view_extent.width, view_extent.height, levels, pitch,
-		     static_cast<uint32_t>(rt.info.format), static_cast<uint32_t>(rt.info.channel_type),
-		     static_cast<uint32_t>(rt.info.channel_order), samples, tile ? "tiled" : "linear");
-	}
-
-	TextureCache::ImageDesc desc {};
+	auto& desc = r.desc;
 	desc.type              = TextureCache::BindingType::RenderTarget;
 	desc.info.data         = {rt.base.addr, backing_size};
 	desc.info.pixel_format = target_format.format;
@@ -304,15 +273,25 @@ void RenderExecutor::ResolveRenderColorTarget(CommandBuffer& buffer, RenderColor
 	desc.info.samples         = samples;
 	desc.info.tile_mode       = rt.attrib3.tile_mode;
 	const bool has_dcc        = rt.info.dcc_compression_enable && rt.dcc_addr.addr != 0;
-	if (has_dcc) {
+	const bool has_cmask      = !rt.info.dcc_compression_enable && rt.info.cmask_fast_clear_enable &&
+	                            rt.cmask.addr != 0 && samples == 1 &&
+	                            !rt.info.fmask_compression_enable &&
+	                            !rt.attrib3.write_vrs_rate_hint_to_cmask;
+	if (has_dcc || has_cmask) {
 		TileSizeAlign metadata_size {};
-		(void)TileGetDccSize(width, height, volume ? depth : view.image_layers, bytes_per_element,
-		                     levels, rt.attrib3.tile_mode, metadata_size, rt.attrib.num_fragments);
-		desc.info.metadata.kind                     = ImageMetadataKind::Dcc;
-		desc.info.metadata.range                    = {rt.dcc_addr.addr, metadata_size.size};
-		desc.info.metadata.dcc_clear_word           = rt.clear_word0.word0;
-		desc.info.metadata.dcc_clear_register_valid = true;
-		desc.info.metadata.dcc_alpha_msb            = DccAlphaOnMsb(rt.info);
+		const auto layers = volume ? depth : view.image_layers;
+		if (has_dcc) {
+			(void)TileGetDccSize(width, height, layers, bytes_per_element, levels,
+			                     rt.attrib3.tile_mode, metadata_size, rt.attrib.num_fragments);
+			desc.info.metadata.dcc_alpha_msb = DccAlphaOnMsb(rt.info);
+		} else {
+			(void)TileGetCmaskSize(width, height, layers, levels, metadata_size);
+		}
+		// DCC owns the clear when both planes are enabled; single-sample CMASK stays expanded.
+		desc.info.metadata.kind = has_dcc ? ImageMetadataKind::Dcc : ImageMetadataKind::Cmask;
+		desc.info.metadata.range = {has_dcc ? rt.dcc_addr.addr : rt.cmask.addr, metadata_size.size};
+		desc.info.metadata.clear_word           = rt.clear_word0.word0;
+		desc.info.metadata.clear_register_valid = true;
 	}
 	for (uint32_t level = 0; level < levels; level++) {
 		if (volume) {
@@ -353,7 +332,6 @@ void RenderExecutor::ResolveRenderColorTarget(CommandBuffer& buffer, RenderColor
 	desc.view_info.layer_count = view.layer_count;
 	desc.view_info.usage       = vk::ImageUsageFlagBits::eColorAttachment;
 	auto& texture_cache        = m_context.GetTextureCache();
-	r.desc                     = std::move(desc);
 	r.guest_mip_level          = rt.view.current_mip_level;
 	r.guest_array_layer        = view.base_layer;
 	r.image_id                 = texture_cache.FindImage(r.desc, exact_format);

src/graphics/host_gpu/renderer/debug.cpp

@@ -219,17 +219,6 @@ static void RtCheck(const HW::RenderTarget& rt) {
 			}
 		}
 
-		if (rt.info.cmask_fast_clear_enable || rt.info.dcc_compression_enable) {
-			static bool logged = false;
-			if (!logged) {
-				LOGF("RenderTarget: temporary: ignoring PS5 color metadata fast_clear=%s dcc=%s "
-				     "cmask=0x%016" PRIx64 " dcc_addr=0x%016" PRIx64 "\n",
-				     rt.info.cmask_fast_clear_enable ? "true" : "false",
-				     rt.info.dcc_compression_enable ? "true" : "false", rt.cmask.addr,
-				     rt.dcc_addr.addr);
-				logged = true;
-			}
-		}
 		if (rt.info.blend_bypass) {
 			static bool logged = false;
 			if (!logged) {

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

@@ -1,6 +1,7 @@
 #include "graphics/host_gpu/renderer/image/image.h"
 
 #include "common/assert.h"
+#include "common/logging/log.h"
 #include "common/profiler.h"
 #include "graphics/host_gpu/renderer/cache/streamBuffer.h"
 #include "graphics/host_gpu/renderer/commandScheduler.h"
@@ -10,7 +11,9 @@
 
 #include <algorithm>
 #include <array>
+#include <atomic>
 #include <cstdint>
+#include <fmt/format.h>
 #include <xxhash.h>
 
 namespace Libs::Graphics {
@@ -48,12 +51,29 @@ namespace {
 }
 
 [[nodiscard]] vk::ImageUsageFlags ImageUsageFlags(GraphicContext& graphics, const ImageInfo& info) {
+	auto usage = vk::ImageUsageFlagBits::eTransferSrc | vk::ImageUsageFlagBits::eTransferDst;
 	if (info.IsBlock()) {
-		return vk::ImageUsageFlagBits::eTransferSrc | vk::ImageUsageFlagBits::eTransferDst |
-		       vk::ImageUsageFlagBits::eSampled;
+		usage |= vk::ImageUsageFlagBits::eSampled;
+		if (graphics.supports_block_texel_view) {
+			const auto storage = usage | vk::ImageUsageFlagBits::eStorage;
+			if (graphics.GetImageFormatProperties(info.pixel_format, HostImageType(info.type),
+			                                      vk::ImageTiling::eOptimal, storage,
+			                                      ImageCreateFlags(graphics, info),
+			                                      nullptr) == vk::Result::eSuccess) {
+				usage = storage;
+			} else {
+				static std::atomic_flag warned = ATOMIC_FLAG_INIT;
+				if (!warned.test_and_set(std::memory_order_relaxed)) {
+					Log::WriteToConsoleAndLog(fmt::format(
+					    "Warning: format {} does not support storage access; block-compressed "
+					    "textures written by the guest will not render.\n",
+					    vk::to_string(info.pixel_format)));
+				}
+			}
+		}
+		return usage;
 	}
 	const auto properties = graphics.GetFormatProperties(info.pixel_format);
-	auto       usage = vk::ImageUsageFlagBits::eTransferSrc | vk::ImageUsageFlagBits::eTransferDst;
 	if (HasFormatFeature(properties, vk::FormatFeatureFlagBits::eSampledImage)) {
 		usage |= vk::ImageUsageFlagBits::eSampled;
 	}
@@ -627,12 +647,11 @@ void Validate(const ImageInfo& info) {
 			}
 			break;
 		case ImageMetadataKind::Dcc:
-			if (info.metadata.range.address == 0 ||
-			    info.metadata.range.address >= TRACKER_ADDRESS_SIZE ||
-			    (info.metadata.range.size != 0 &&
-			     info.metadata.range.size > TRACKER_ADDRESS_SIZE - info.metadata.range.address) ||
+		case ImageMetadataKind::Cmask:
+			if (!GuestRange {info.metadata.range.address,
+			                 std::max<uint64_t>(info.metadata.range.size, 1)}.Valid() ||
 			    info.metadata.compression == VideoOutCompression::Unsupported) {
-				EXIT("invalid DCC metadata\n");
+				EXIT("invalid color metadata\n");
 			}
 			break;
 	}

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

@@ -17,17 +17,17 @@ namespace Libs::Graphics {
 
 enum class VideoOutCompression : uint8_t { Uncompressed, Dcc256_256_0, Dcc256_64_64, Unsupported };
 
-enum class ImageMetadataKind : uint8_t { None, Htile, Dcc };
+enum class ImageMetadataKind : uint8_t { None, Htile, Dcc, Cmask };
 
 struct ImageMetadataInfo {
 	GuestRange          range;
-	ImageMetadataKind   kind               = ImageMetadataKind::None;
-	uint32_t            control            = 0;
-	uint32_t            dcc_clear_word           = 0;
-	VideoOutCompression compression        = VideoOutCompression::Uncompressed;
-	bool                stencil_compressed = false;
-	bool                dcc_clear_register_valid = false;
-	bool                dcc_alpha_msb            = true;
+	ImageMetadataKind   kind                 = ImageMetadataKind::None;
+	uint32_t            control              = 0;
+	uint32_t            clear_word           = 0;
+	VideoOutCompression compression          = VideoOutCompression::Uncompressed;
+	bool                stencil_compressed   = false;
+	bool                clear_register_valid = false;
+	bool                dcc_alpha_msb        = true;
 };
 
 struct ImageSubresources {
@@ -428,6 +428,20 @@ inline constexpr std::array<VideoOutFormatPolicy, 7> VIDEO_OUT_FORMAT_POLICIES {
 	return true;
 }
 
+// Unlike a register clear, a DWORD fill repeats the same word across the entire pixel.
+// 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::eR16G16B16A16Sfloat) {
+		if (packed != 0) {
+			return false;
+		}
+		clear = vk::ClearColorValue {};
+		return true;
+	}
+	return DecodePackedColorClear(format, packed, clear);
+}
+
 [[nodiscard]] inline bool DecodePackedStencilClear(uint32_t packed, uint8_t& clear) {
 	const auto value = static_cast<uint8_t>(packed);
 	if (packed != static_cast<uint32_t>(value) * 0x01010101u) {

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

@@ -31,20 +31,14 @@ namespace ImageViewOps {
 	       ImageViewOps::IsFormatDepthCompatible(view_format);
 }
 
-[[nodiscard]] inline bool IsValidImageSwizzle(uint32_t swizzle) noexcept {
+[[nodiscard]] inline bool IsValidImageSwizzle(uint32_t swizzle,
+                                             uint32_t component_count = 4) noexcept {
 	if ((swizzle & ~0xfffu) != 0) {
 		return false;
 	}
 	for (uint32_t channel = 0; channel < 4; channel++) {
-		switch (GetDstSel(swizzle, channel)) {
-			case 0:
-			case 1:
-			case 4:
-			case 5:
-			case 6:
-			case 7: break;
-			default: return false;
-		}
+		const auto selector = GetDstSel(swizzle, channel);
+		if (selector > 1 && (selector < 4 || selector >= 4 + component_count)) return false;
 	}
 	return true;
 }
@@ -81,15 +75,8 @@ SelectSampledColorView(vk::Format image_format, vk::Format view_format, uint32_t
 [[nodiscard]] inline bool IsSupportedSampledDepthView(vk::Format image_format,
                                                       vk::Format view_format,
                                                       uint32_t   swizzle) noexcept {
-	if (!IsSupportedSampledDepthFormat(image_format, view_format)) {
-		return false;
-	}
-	switch (swizzle) {
-		case DstSel(4, 4, 4, 4):
-		case DstSel(4, 0, 0, 0):
-		case DstSel(4, 0, 0, 1): return true;
-		default: return false;
-	}
+	return IsSupportedSampledDepthFormat(image_format, view_format) &&
+	       IsValidImageSwizzle(swizzle, 1);
 }
 
 [[nodiscard]] inline bool

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

@@ -88,30 +88,6 @@ TileManager::~TileManager() {
 	}
 }
 
-TileManager::Scratch TileManager::AllocateScratch(uint64_t size) {
-	EXIT_IF(size == 0);
-	vk::BufferCreateInfo create {};
-	create.size  = size;
-	create.usage = vk::BufferUsageFlagBits::eStorageBuffer | vk::BufferUsageFlagBits::eTransferSrc |
-	               vk::BufferUsageFlagBits::eTransferDst;
-
-	VmaAllocationCreateInfo allocate {};
-	allocate.usage       = VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE;
-	VkBuffer      buffer = VK_NULL_HANDLE;
-	VmaAllocation memory = nullptr;
-	const auto    raw    = static_cast<VkBufferCreateInfo>(create);
-	RequireVulkanSuccess(static_cast<vk::Result>(vmaCreateBuffer(
-	                         m_graphics.allocator, &raw, &allocate, &buffer, &memory, nullptr)),
-	                     "allocate TileManager scratch buffer");
-	return {buffer, memory, size};
-}
-
-void TileManager::DeferDestroy(Scratch scratch) {
-	auto allocator = m_graphics.allocator;
-	m_scheduler.DeferOperation(
-	    [allocator, scratch] { vmaDestroyBuffer(allocator, scratch.buffer, scratch.allocation); });
-}
-
 void TileManager::Prepare(bool tile, uint64_t tiled_capacity, uint64_t linear_capacity,
                           std::span<const GpuTileInfo> infos, uint64_t source_base,
                           uint64_t target_base, std::vector<Dispatch>& dispatches) {
@@ -369,8 +345,7 @@ TileManager::Result TileManager::Detile(vk::Buffer tiled, uint64_t tiled_offset,
 	const uint64_t        source_base = tiled_offset & (descriptor_alignment - 1);
 	std::vector<Dispatch> dispatches;
 	Prepare(false, tiled_capacity, linear_capacity, infos, source_base, 0, dispatches);
-	auto scratch = AllocateScratch(Common::AlignUp(linear_capacity, 4));
-	DeferDestroy(scratch);
+	auto scratch = GetScratchBuffer(linear_capacity, tiled);
 	Record(tiled, tiled_offset, tiled_capacity, scratch.buffer, 0, scratch.size, dispatches,
 	       true);
 	return {scratch.buffer, 0, linear_capacity};
@@ -400,11 +375,9 @@ void TileManager::TileImage(Image& image, std::span<const vk::BufferImageCopy> r
 	    std::max<uint64_t>(limits.minStorageBufferOffsetAlignment, 4);
 	const uint64_t        target_base = tiled_offset & (descriptor_alignment - 1);
 	std::vector<Dispatch> dispatches;
-	// Reserve all stream parameters before creating a scheduler-lived scratch dependency:
-	// StreamBuffer::Map is allowed to submit the current tick when it wraps.
+	// Reserve stream parameters before recording the image download and conversion.
 	Prepare(true, tiled_capacity, linear_capacity, infos, 0, target_base, dispatches);
-	auto linear = AllocateScratch(Common::AlignUp(linear_capacity, 4));
-	DeferDestroy(linear);
+	auto linear = GetScratchBuffer(linear_capacity, tiled);
 	image.Download(regions, linear.buffer, 0, linear.size);
 	Result source {linear.buffer, 0, linear.size};
 	if (transform == ColorTransform::SwapBgra16) {
@@ -414,10 +387,19 @@ void TileManager::TileImage(Image& image, std::span<const vk::BufferImageCopy> r
 	       dispatches, false);
 }
 
-TileManager::Result TileManager::GetScratchBuffer(uint64_t size) {
-	auto scratch = AllocateScratch(Common::AlignUp(size, 4));
-	DeferDestroy(scratch);
-	return {scratch.buffer, 0, scratch.size};
+TileManager::Result TileManager::GetScratchBuffer(uint64_t size, vk::Buffer input) {
+	EXIT_IF(size == 0);
+	size = Common::AlignUp(size, 4);
+	auto& buffer = m_scratch[m_scratch[0] && m_scratch[0]->Handle() == input ? 1 : 0];
+	if (!buffer || buffer->Size() < size) {
+		if (buffer) {
+			m_scheduler.DeferOperation([old = std::move(buffer)]() mutable { old.reset(); });
+		}
+		buffer = std::make_unique<Buffer>(m_graphics, m_scheduler, MemoryUsage::DeviceLocal, 0,
+		    vk::BufferUsageFlagBits::eStorageBuffer | vk::BufferUsageFlagBits::eTransferSrc |
+		        vk::BufferUsageFlagBits::eTransferDst, size);
+	}
+	return {buffer->Handle(), 0, size};
 }
 
 TileManager::StorageBinding TileManager::BindStorage(Result buffer, uint64_t size) const {
@@ -668,9 +650,7 @@ void TileManager::SwapBgra16(Result input, Result output, uint32_t pixels) {
 
 TileManager::Result TileManager::SwapBgra16(Result input) {
 	EXIT_NOT_IMPLEMENTED(input.size == 0 || input.size % 8u != 0 || input.size / 8u > UINT32_MAX);
-	auto output = AllocateScratch(input.size);
-	DeferDestroy(output);
-	Result result {output.buffer, 0, output.size};
+	auto result = GetScratchBuffer(input.size, input.buffer);
 	SwapBgra16(input, result, static_cast<uint32_t>(input.size / 8u));
 	return result;
 }

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

@@ -6,13 +6,14 @@
 #include "graphics/host_gpu/vulkanCommon.h"
 
 #include <array>
+#include <memory>
 #include <span>
 #include <vector>
-#include <vk_mem_alloc.h>
 
 namespace Libs::Graphics {
 
 class CommandScheduler;
+class Buffer;
 class Image;
 class StreamBuffer;
 struct GraphicContext;
@@ -62,7 +63,7 @@ public:
 	~TileManager();
 	KYTY_CLASS_NO_COPY(TileManager);
 
-	// The returned device-local buffer remains alive through the current scheduler tick.
+	// Consume scratch results before the next acquisition, or pass their buffer as input.
 	[[nodiscard]] Result Detile(vk::Buffer tiled, uint64_t tiled_offset, uint64_t tiled_capacity,
 	                            uint64_t linear_capacity, std::span<const GpuTileInfo> infos);
 	void Tile(vk::Buffer linear, uint64_t linear_offset, uint64_t linear_capacity, vk::Buffer tiled,
@@ -71,7 +72,7 @@ public:
 	               uint64_t tiled_offset, uint64_t tiled_capacity, uint64_t linear_capacity,
 	               std::span<const GpuTileInfo> infos,
 	               ColorTransform               transform = ColorTransform::None);
-	[[nodiscard]] Result GetScratchBuffer(uint64_t size);
+	[[nodiscard]] Result GetScratchBuffer(uint64_t size, vk::Buffer input = nullptr);
 	void                 ConvertD16(Result source, Result target, D16Direction direction, bool d32,
 	                                const D16Layout& layout);
 	[[nodiscard]] Result SwapBgra16(Result input);
@@ -105,23 +106,16 @@ private:
 		uint32_t pipeline_slot = 0;
 		uint64_t params_offset = 0;
 	};
-	struct Scratch {
-		vk::Buffer    buffer     = nullptr;
-		VmaAllocation allocation = nullptr;
-		uint64_t      size       = 0;
-	};
 	struct StorageBinding {
 		vk::DescriptorBufferInfo info;
 		uint32_t                 base = 0;
 	};
 
-	[[nodiscard]] Scratch         AllocateScratch(uint64_t size);
 	[[nodiscard]] StorageBinding  BindStorage(Result buffer, uint64_t size) const;
 	[[nodiscard]] static uint32_t ConversionRows(uint64_t offset, uint64_t row_stride,
 	                                             uint64_t active, uint32_t remaining,
 	                                             uint64_t alignment, uint64_t max_range,
 	                                             uint32_t max_groups) noexcept;
-	void                          DeferDestroy(Scratch scratch);
 	void Prepare(bool tile, uint64_t tiled_capacity, uint64_t linear_capacity,
 	             std::span<const GpuTileInfo> infos, uint64_t source_base, uint64_t target_base,
 	             std::vector<Dispatch>& dispatches);
@@ -134,6 +128,7 @@ private:
 	GraphicContext&                         m_graphics;
 	CommandScheduler&                       m_scheduler;
 	StreamBuffer&                           m_stream_buffer;
+	std::array<std::unique_ptr<Buffer>, 2>   m_scratch;
 	vk::DescriptorSetLayout                 m_descriptor_layout = nullptr;
 	vk::PipelineLayout                      m_pipeline_layout   = nullptr;
 	std::array<vk::Pipeline, PipelineCount> m_pipelines {};

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

@@ -0,0 +1,58 @@
+#include "graphics/host_gpu/renderer/pipeline/blendMapping.h"
+
+#include "graphics/guest_gpu/hardwareContext.h"
+
+#include <initializer_list>
+
+namespace Libs::Graphics {
+
+bool BlendFactorIsDualSource(uint8_t factor) {
+	return factor >= static_cast<uint8_t>(Prospero::BlendFactor::kSrc1Color) &&
+	       factor <= static_cast<uint8_t>(Prospero::BlendFactor::kOneMinusSrc1Alpha);
+}
+
+namespace {
+
+bool BlendFactorIsConstantColor(uint8_t factor) {
+	return factor == static_cast<uint8_t>(Prospero::BlendFactor::kConstantColor) ||
+	       factor == static_cast<uint8_t>(Prospero::BlendFactor::kOneMinusConstantColor);
+}
+
+} // namespace
+
+BlendMappingSupport ClassifyBlendMapping(const HW::BlendControl&                blend,
+                                         const Prospero::ColorComponentMapping& mapping) {
+	// Color constants are not swizzled with the exports; scalar constant alpha is unaffected.
+	if (!mapping.IsIdentity() && (BlendFactorIsConstantColor(blend.color_srcblend) ||
+	                              BlendFactorIsConstantColor(blend.color_destblend))) {
+		return BlendMappingSupport::Unsupported;
+	}
+	if (mapping.Map(3) == 3u) {
+		return BlendMappingSupport::Direct;
+	}
+	// Moving alpha requires the same equation for all channels.
+	if (blend.separate_alpha_blend && (blend.alpha_srcblend != blend.color_srcblend ||
+	                                   blend.alpha_destblend != blend.color_destblend ||
+	                                   blend.alpha_comb_fcn != blend.color_comb_fcn)) {
+		return BlendMappingSupport::Unsupported;
+	}
+	auto support = BlendMappingSupport::Direct;
+	for (const auto factor: {blend.color_srcblend, blend.color_destblend}) {
+		if (BlendFactorIsDualSource(factor)) {
+			return BlendMappingSupport::Unsupported;
+		}
+		switch (static_cast<Prospero::BlendFactor>(factor)) {
+			case Prospero::BlendFactor::kSrcAlpha:
+			case Prospero::BlendFactor::kOneMinusSrcAlpha:
+				support = BlendMappingSupport::SourceAlpha;
+				break;
+			case Prospero::BlendFactor::kDstAlpha:
+			case Prospero::BlendFactor::kOneMinusDstAlpha:
+			case Prospero::BlendFactor::kSrcAlphaSaturate: return BlendMappingSupport::Unsupported;
+			default: break;
+		}
+	}
+	return support;
+}
+
+} // namespace Libs::Graphics

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

@@ -0,0 +1,27 @@
+#ifndef EMULATOR_SRC_GRAPHICS_HOST_GPU_RENDERER_PIPELINE_BLENDMAPPING_H_
+#define EMULATOR_SRC_GRAPHICS_HOST_GPU_RENDERER_PIPELINE_BLENDMAPPING_H_
+
+#include <cstdint>
+
+namespace Libs::Graphics {
+
+namespace HW {
+struct BlendControl;
+}
+namespace Prospero {
+struct ColorComponentMapping;
+}
+
+enum class BlendMappingSupport {
+	Direct,
+	SourceAlpha, // Requires logical alpha in the second blend source.
+	Unsupported,
+};
+
+bool                BlendFactorIsDualSource(uint8_t factor);
+BlendMappingSupport ClassifyBlendMapping(const HW::BlendControl&                blend,
+                                         const Prospero::ColorComponentMapping& mapping);
+
+} // namespace Libs::Graphics
+
+#endif // EMULATOR_SRC_GRAPHICS_HOST_GPU_RENDERER_PIPELINE_BLENDMAPPING_H_

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

@@ -91,19 +91,6 @@ vk::DescriptorImageInfo MakeImageInfo(const TextureBinding& texture, uint32_t el
 	return {nullptr, view, texture.layout};
 }
 
-static const char* ShaderStageResourceName(ShaderType stage) {
-	switch (stage) {
-		case ShaderType::Vertex: return "Vertex";
-		case ShaderType::Mesh: return "Mesh";
-		case ShaderType::Local: return "Local";
-		case ShaderType::TessellationControl: return "Hull";
-		case ShaderType::TessellationEvaluation: return "Domain";
-		case ShaderType::Pixel: return "Pixel";
-		case ShaderType::Compute: return "Compute";
-		default: return "Unknown";
-	}
-}
-
 static Prospero::ImageType TextureType(const ShaderTextureResource& descriptor) {
 	const auto type = descriptor.Type();
 	return type == Prospero::ImageType::kCube ? Prospero::ImageType::kColor2DArray : type;
@@ -126,8 +113,7 @@ static bool IsMultisampledTexture(Prospero::ImageType type) {
 
 static vk::DescriptorBufferInfo
 NativeStorageBuffer(RenderContext& context, const PreparedBindings::BufferSource& source,
-                    const ShaderRecompiler::IR::BufferResource& resource, ShaderType stage,
-                    uint32_t slot, uint32_t& buffer_offset) {
+                    const ShaderRecompiler::IR::BufferResource& resource, uint32_t& buffer_offset) {
 	buffer_offset = 0;
 
 	const auto& [address, size, id] = source;
@@ -152,16 +138,6 @@ NativeStorageBuffer(RenderContext& context, const PreparedBindings::BufferSource
 	if (resource.written) {
 		context.GetTextureCache().InvalidateMemoryFromGPU(address, size);
 	}
-	const char* access = "Read";
-	if (resource.written && resource.read) {
-		access = "ReadWrite";
-	} else if (resource.written) {
-		access = "Write";
-	}
-	SetVulkanObjectNameF(
-	    graphics.device, result.buffer,
-	    "Kyty.{}.StorageBuffer[slot={} guest=0x{:016x} size=0x{:x} access={} formatted={}]",
-	    ShaderStageResourceName(stage), slot, address, size, access, resource.formatted);
 	return result;
 }
 
@@ -196,8 +172,8 @@ bool IsSupportedDepthTextureEncoding(const ShaderTextureResource& descriptor, bo
 	constexpr uint32_t htile_control = 0x00280000u;
 	const uint32_t expected_control  = htile_control | (descriptor.MsaaDepth() ? (1u << 10u) : 0u);
 	const auto     metadata_addr     = descriptor.MetaAddr() << 8u;
-	return metadata_control == expected_control && metadata_addr != 0 &&
-	       metadata_addr < TRACKER_ADDRESS_SIZE && (metadata_addr & 0x7fffu) == 0 &&
+	return metadata_control == expected_control && GuestRange {metadata_addr, 1}.Valid() &&
+	       (metadata_addr & 0x7fffu) == 0 &&
 	       descriptor.TileMode() == Prospero::TileMode::kDepth;
 }
 
@@ -335,8 +311,10 @@ void ValidateStorageTexture(const ShaderRecompiler::IR::ImageResource& resource,
 	const bool raw_sint_storage = format == Prospero::BufferFormat::k32SInt && uint_resource &&
 	                              resource.written && !resource.read && !resource.atomic;
 	const auto numeric_class = Prospero::SampledTextureNumericClass(format);
+	const bool raw_float_atomic = format == Prospero::BufferFormat::k32Float && uint_resource &&
+	                              resource.atomic;
 	const bool format_ok =
-	    raw_sint_storage ||
+	    raw_sint_storage || raw_float_atomic ||
 	    (numeric_class != Prospero::TextureNumericClass::Unsupported &&
 	     numeric_class != Prospero::TextureNumericClass::Sint &&
 	     uint_resource == (numeric_class == Prospero::TextureNumericClass::Uint) &&
@@ -512,19 +490,50 @@ static ImageViewInfo TextureViewInfo(const ShaderRecompiler::IR::ImageResource&
 	return view;
 }
 
-static bool TextureViewPreservesMipLayout(const TileSurfaceDescription& description,
-                                          uint32_t                      view_levels) {
+static bool ResolveTextureMipView(const TileSurfaceDescription& description, bool metadata,
+                                   uint32_t view_levels, uint32_t& levels, uint32_t& base_level) {
 	TileSurfaceLayout physical {};
 	TileSurfaceLayout view {};
 	auto              view_description = description;
-	view_description.levels            = view_levels;
-	return TileGetTiledTextureLayout(description, physical) &&
-	       TileGetTiledTextureLayout(view_description, view) &&
-	       physical.first_tail_level == view.first_tail_level &&
-	       physical.block_slice_size == view.block_slice_size &&
-	       physical.total_size == view.total_size &&
-	       std::equal(std::begin(physical.mips), std::begin(physical.mips) + description.levels,
-	                  std::begin(view.mips));
+	view_description.levels            = levels;
+	if (!TileGetTiledTextureLayout(description, physical) ||
+	    !TileGetTiledTextureLayout(view_description, view)) {
+		return false;
+	}
+	if (physical.first_tail_level == view.first_tail_level &&
+	    physical.block_slice_size == view.block_slice_size &&
+	    physical.total_size == view.total_size &&
+	    std::equal(std::begin(physical.mips), std::begin(physical.mips) + description.levels,
+	               std::begin(view.mips))) {
+		return true;
+	}
+	if (metadata || ((description.layers > 1 || description.depth > 1) &&
+	                 physical.block_slice_size != view.block_slice_size)) {
+		return false;
+	}
+	// T# addresses the last mip. A view can select the same stored subresources
+	// with different mip indices; inaccessible mips need no host representation.
+	for (uint32_t base = 0; base + view_levels <= description.levels; ++base) {
+		bool matches = true;
+		for (uint32_t i = 0; i < view_levels; ++i) {
+			const auto source = base_level + i;
+			const auto target = base + i;
+			if (physical.mips[target] != view.mips[source] ||
+			    (target >= physical.first_tail_level) != (source >= view.first_tail_level) ||
+			    std::max(description.width >> target, 1u) != std::max(description.width >> source, 1u) ||
+			    std::max(description.height >> target, 1u) != std::max(description.height >> source, 1u) ||
+			    std::max(description.depth >> target, 1u) != std::max(description.depth >> source, 1u)) {
+				matches = false;
+				break;
+			}
+		}
+		if (matches) {
+			levels     = description.levels;
+			base_level = base;
+			return true;
+		}
+	}
+	return false;
 }
 
 TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageResource&   resource,
@@ -558,7 +567,7 @@ TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageR
 	const auto view_levels = multisampled || single_storage_mip
 	                             ? 1u
 	                             : static_cast<uint32_t>(last_level - base_level) + 1u;
-	const auto levels =
+	auto levels =
 	    multisampled ? 1u : std::max(physical_levels, base_level + view_levels);
 	const auto tile       = descriptor.TileMode();
 	const bool depth_tile = tile == Prospero::TileMode::kDepth;
@@ -600,12 +609,13 @@ TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageR
 	                             type == Prospero::ImageType::kColor2DArray ||
 	                             type == Prospero::ImageType::kColor2DMsaaArray;
 	const auto    image_layers = layered ? depth : 1u;
+	auto          view_base    = static_cast<uint32_t>(base_level);
 	if (levels > physical_levels) {
 		const TileSurfaceDescription physical {
 		    format, tile, volume ? TileSurfaceDimension::Dim3D : TileSurfaceDimension::Dim2D,
 		    width, height, volume ? depth : 1u, physical_levels, image_layers};
-		// Texture mip views take precedence over the resource count, but must keep its storage layout.
-		if (!TextureViewPreservesMipLayout(physical, levels)) {
+		if (!ResolveTextureMipView(physical, !resource.r128 && descriptor.MetaCompress(),
+		                           view_levels, levels, view_base)) {
 			EXIT("unsupported texture mip view changes physical layout: base=%u last=%u max=%u "
 			     "extent=%ux%ux%u tile=%u\n",
 			     base_level, last_level, max_mip, width, height, depth,
@@ -640,7 +650,8 @@ TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageR
 			pixel_format = depth_format->depth_attachment_format;
 		}
 	}
-	const auto storage_view_format = storage && format == Prospero::BufferFormat::k32SInt
+	const auto storage_view_format = storage && (resource.atomic ||
+	                                            format == Prospero::BufferFormat::k32SInt)
 	                                     ? vk::Format::eR32Uint
 	                                     : SrgbStorageViewFormat(pixel_format);
 	const auto view_format         = storage && storage_view_format != vk::Format::eUndefined
@@ -676,6 +687,7 @@ TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageR
 	}
 	desc.view_info = TextureViewInfo(resource, descriptor, view_format, surface_format, storage,
 	                                 view_levels, desc.info.resources.layers);
+	desc.view_info.base_level = view_base;
 	desc.type = storage ? TextureCache::BindingType::Storage : TextureCache::BindingType::Texture;
 
 	auto       id                  = texture_cache.FindImage(desc, shader_conversion);
@@ -702,14 +714,15 @@ static vk::Sampler NativeSampler(RenderContext&                       context,
                                  const ShaderRecompiler::IR::CompiledShaderInfo& program,
                                  uint32_t index,
                                  const ShaderRecompiler::IR::DescriptorValue& value) {
-	auto descriptor = DecodeNativeDescriptor<ShaderSamplerResource>(value);
-	if (!program.info.samplers[index].depth_compare) {
+	auto        descriptor = DecodeNativeDescriptor<ShaderSamplerResource>(value);
+	const auto& sampler = program.info.samplers[index];
+	if (!sampler.depth_compare) {
 		descriptor.fields[0] &= ~(0x7u << 12u);
 	}
-	if (program.info.samplers[index].force_point_filtering) {
+	if (sampler.force_point_filtering) {
 		descriptor.SetPointFiltering();
 	}
-	return context.GetSamplerCache().GetSampler(descriptor);
+	return context.GetSamplerCache().GetSampler(descriptor, sampler.integer_border);
 }
 
 static vk::DescriptorBufferInfo NativeUpload(RenderContext&            context,
@@ -795,9 +808,14 @@ void RenderExecutor::FindBuffers(PreparedBindings& prepared) {
 	auto&       cache    = m_context.GetBufferCache();
 
 	prepared.buffer_sources.clear();
-	prepared.buffer_sources.reserve(program.info.buffers.size());
-	for (uint32_t i = 0; i < program.info.buffers.size(); i++) {
-		auto descriptor = DecodeNativeDescriptor<ShaderBufferResource>(snapshot.buffers[i]);
+	const auto& layout = program.bindings;
+	if (layout.memory_offset_count == 0) {
+		return;
+	}
+	const auto& resources = layout.descriptors.front().resources;
+	prepared.buffer_sources.reserve(resources.size());
+	for (const auto resource: resources) {
+		auto descriptor = DecodeNativeDescriptor<ShaderBufferResource>(snapshot.buffers[resource]);
 		const auto address = descriptor.Base48();
 		const auto requested_size = descriptor.GetSize();
 		if (address == 0 || requested_size == 0) {
@@ -815,10 +833,10 @@ void RenderExecutor::RebindBuffers(PreparedBindings& prepared) {
 	const auto& program   = *prepared.runtime->program;
 	const auto& snapshot  = *prepared.runtime->resources;
 	const auto& layout    = program.bindings;
-	EXIT_IF(prepared.buffer_sources.size() != program.info.buffers.size());
+	EXIT_IF(prepared.buffer_sources.size() != layout.memory_offset_count);
 
 	prepared.buffers.clear();
-	prepared.buffers.reserve(program.info.buffers.size());
+	prepared.buffers.reserve(layout.memory_offset_count);
 	EXIT_IF(prepared.shader_data.size() != layout.ShaderDataDwords());
 	std::fill(prepared.shader_data.begin() + layout.memory_offset_dword,
 	          prepared.shader_data.end(), 0);
@@ -827,10 +845,11 @@ void RenderExecutor::RebindBuffers(PreparedBindings& prepared) {
 		const auto shift = (index % 4u) * 8u;
 		prepared.shader_data[dword] |= offset << shift;
 	};
-	for (uint32_t i = 0; i < program.info.buffers.size(); i++) {
+	for (uint32_t i = 0; i < layout.memory_offset_count; i++) {
+		const auto resource = layout.descriptors.front().resources[i];
 		uint32_t buffer_offset = 0;
 		prepared.buffers.push_back(NativeStorageBuffer(m_context, prepared.buffer_sources[i],
-		                                               program.info.buffers[i], program.stage, i,
+		                                               program.info.buffers[resource],
 		                                               buffer_offset));
 		pack_memory_offset(i, buffer_offset);
 	}
@@ -955,6 +974,41 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 		        (pipeline_bind_point == vk::PipelineBindPoint::eCompute &&
 		         shader_stage != vk::ShaderStageFlagBits::eCompute));
 	}
+	for (const auto* reader: prepared_bindings) {
+		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) {
+				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}) {
+					if (written.size != 0 && ImageRangeOverlaps(address, size,
+					                                          written.address, written.size)) {
+						EXIT("scalar resource reads overlap an image or attachment write\n");
+					}
+				}
+			}
+			for (const auto* writer: prepared_bindings) {
+				const auto& program = *writer->runtime->program;
+				for (uint32_t i = 0; i < writer->buffer_sources.size(); ++i) {
+					const auto resource = program.bindings.descriptors.front().resources[i];
+					if (!program.info.buffers[resource].written) continue;
+					const auto& written = writer->buffer_sources[i];
+					if (written.size != 0 && ImageRangeOverlaps(address, size,
+					                                          written.address, written.size)) {
+						EXIT("scalar resource reads overlap a shader buffer write\n");
+					}
+				}
+			}
+		}
+	}
 	m_descriptor_buffers.clear();
 	m_descriptor_images.clear();
 	m_descriptor_writes.clear();
@@ -1042,8 +1096,8 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 			} else {
 				switch (binding.kind) {
 					case BindingKind::Buffers:
-						for (const auto resource: binding.resources) {
-							const auto& view = descriptors.buffers.at(resource);
+						EXIT_IF(descriptors.buffers.size() != binding.resources.size());
+						for (const auto& view: descriptors.buffers) {
 							EXIT_IF(view.buffer == nullptr);
 							m_descriptor_buffers.push_back(view);
 						}

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

@@ -5,11 +5,13 @@
 #include "common/file.h"
 #include "common/logging/log.h"
 #include "common/profiler.h"
+#include "common/threads.h"
 #include "graphics/guest_gpu/hardwareContext.h"
 #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/imageView.h"
+#include "graphics/host_gpu/renderer/pipeline/blendMapping.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
 #include "graphics/shader/recompiler/ShaderRecompiler.h"
@@ -29,7 +31,6 @@
 #include <span>
 #include <spirv-tools/libspirv.hpp>
 #include <string_view>
-#include <tuple>
 #include <utility>
 #include <vector>
 #include <xxhash.h>
@@ -116,33 +117,22 @@ void DumpShaderSpirv(const char* stage_name, uint64_t shader_hash,
 }
 
 void DumpShaderOriginal(const char* stage_name, uint64_t shader_hash,
-                        std::span<const uint32_t> code, const std::string& decoded_dump) {
+                        std::span<const uint32_t> code) {
 	if (!Config::GraphicsDebugDumpEnabled()) {
 		return;
 	}
 	EXIT_IF(code.empty());
 	static std::atomic_int id = 0;
-	const auto base = Config::GetShaderLogFolder() / "original" /
-	                  fmt::format("{:04d}_new_shader_{}_{:016x}", id++, stage_name, shader_hash);
-	Common::File::CreateDirectories(base.parent_path());
-	for (const auto& [suffix, data, size]: {
-	         std::tuple {".bin", static_cast<const void*>(code.data()), code.size_bytes()},
-	         std::tuple {".rdna2", static_cast<const void*>(decoded_dump.data()),
-	                     decoded_dump.size()},
-	     }) {
-		if (size == 0) {
-			continue;
-		}
-		auto path = base;
-		path += suffix;
-		Common::File file(path);
-		if (file.IsInvalid()) {
-			const auto path_text = Common::PathToString(path);
-			LOGF_COLOR(Log::Color::BrightRed, "Can't create file: %s\n", path_text.c_str());
-		} else {
-			file.Write(data, size);
-		}
+	const auto path = Config::GetShaderLogFolder() / "original" /
+	                  fmt::format("{:04d}_new_shader_{}_{:016x}.bin", id++, stage_name, shader_hash);
+	Common::File::CreateDirectories(path.parent_path());
+	Common::File file(path);
+	if (file.IsInvalid()) {
+		const auto path_text = Common::PathToString(path);
+		LOGF_COLOR(Log::Color::BrightRed, "Can't create file: %s\n", path_text.c_str());
+		return;
 	}
+	file.Write(code.data(), code.size_bytes());
 }
 
 bool ValidateShaderSpirv(const char* label, uint64_t shader_hash,
@@ -177,6 +167,32 @@ bool ValidateShaderSpirv(const char* label, uint64_t shader_hash,
 
 } // namespace
 
+std::size_t PipelineCache::GraphicsPipelineKeyHash::operator()(const GraphicsPipelineKey& key) const {
+	std::size_t hash = 0;
+	PipelineKeyHash::Mix(hash, key.rendering.color_count);
+	for (uint32_t i = 0; i < key.rendering.color_count; i++) {
+		PipelineKeyHash::Mix(hash, static_cast<uint32_t>(key.rendering.color_formats[i]));
+	}
+	PipelineKeyHash::Mix(hash, static_cast<uint32_t>(key.rendering.depth_format));
+	PipelineKeyHash::Mix(hash, static_cast<uint32_t>(key.rendering.stencil_format));
+	for (const auto id: key.vertex_shader_ids) {
+		PipelineKeyHash::Mix(hash, id);
+	}
+	PipelineKeyHash::Mix(hash, key.ps_shader_id);
+	PipelineKeyHash::Mix(hash, key.vertex_input.binding_count);
+	for (uint32_t i = 0; i < key.vertex_input.binding_count; i++) {
+		PipelineKeyHash::Mix(hash, key.vertex_input.bindings[i].stride);
+		PipelineKeyHash::Mix(hash, key.vertex_input.bindings[i].instance);
+	}
+	PipelineKeyHash::Mix(hash, key.vertex_input.attribute_count);
+	for (uint32_t i = 0; i < key.vertex_input.attribute_count; i++) {
+		PipelineKeyHash::Mix(hash, key.vertex_input.attributes[i].offset);
+		PipelineKeyHash::Mix(hash, key.vertex_input.attributes[i].binding);
+	}
+	PipelineKeyHash::Mix(hash, XXH3_64bits(&key.static_params, sizeof(key.static_params)));
+	return hash;
+}
+
 struct PipelineCache::ProgramCache {
 	struct ProgramKey {
 		ShaderType            stage           = ShaderType::Unknown;
@@ -204,7 +220,6 @@ struct PipelineCache::ProgramCache {
 		ShaderRecompiler::IR::ResourceSnapshot       resources;
 		ShaderRecompiler::IR::ResourceSpecialization specialization;
 		std::vector<Permutation>                    permutations;
-		bool                                        skip_dispatch = false;
 	};
 
 	struct ProgramKeyHash {
@@ -218,32 +233,20 @@ struct PipelineCache::ProgramCache {
 			PipelineKeyHash::Mix(hash, key.code_size);
 			PipelineKeyHash::Mix(hash, key.static_state.size());
 			// Bucket same-shape static variants by source. ProgramKey equality performs the one
-			// exact state comparison needed on a stable hit without hashing up to 429 words first.
+			// exact state comparison needed on a stable hit without hashing the full state first.
 			return hash;
 		}
 	};
 
-	static constexpr std::size_t MaxStaticKeyWords = 13 + ShaderVertexInputInfo::RES_MAX * 13;
+	static constexpr std::size_t MaxStaticKeyWords = 32 + ShaderVertexInputInfo::RES_MAX * 6;
 
-	Permutation CompilePermutation(const ShaderParams&                          params,
+	Permutation CompilePermutation(const char*                                  stage_name,
 	                               const ShaderRecompiler::CompileOptions&      options,
 	                               ShaderRecompiler::TranslateResult            translated,
 	                               ShaderRecompiler::IR::ResourceSpecialization specialization,
 	                               uint32_t push_data_start_dword) {
-		const char* stage_name = nullptr;
-		switch (options.stage) {
-			case ShaderType::Vertex: stage_name = "vs"; break;
-			case ShaderType::Mesh: stage_name = "ms"; break;
-			case ShaderType::Local: stage_name = "ls"; break;
-			case ShaderType::TessellationControl: stage_name = "hs"; break;
-			case ShaderType::TessellationEvaluation: stage_name = "ds"; break;
-			case ShaderType::Pixel: stage_name = "ps"; break;
-			case ShaderType::Compute: stage_name = "cs"; break;
-			default: EXIT("invalid pipeline shader stage\n");
-		}
 		auto result = ShaderRecompiler::CompileProgram(std::move(translated), options,
 		                                               specialization, push_data_start_dword);
-		DumpShaderOriginal(stage_name, options.shader_hash, params.code, result.decoded_dump);
 		if (!ValidateShaderSpirv(options.dump_label, options.shader_hash, result.spirv)) {
 			DumpShaderSpirv(stage_name, options.shader_hash, result.spirv);
 			EXIT("%s failed hash=0x%016" PRIx64 ": SPIR-V validation failed\n", options.dump_label,
@@ -284,9 +287,6 @@ struct PipelineCache::ProgramCache {
 		lookup_key.code_size       = static_cast<uint32_t>(params.code.size());
 		BuildStageStaticKey(input_info, lookup_key.static_state);
 		auto                                         entry = programs.find(lookup_key);
-		if (entry != programs.end() && entry->second.skip_dispatch) {
-			return {};
-		}
 		const ShaderRecompiler::IR::SrtRuntime       runtime {
 		    .user_data                  = user_data,
 		    .shader_base                = params.Base(),
@@ -321,14 +321,15 @@ struct PipelineCache::ProgramCache {
 			stage_input.compute = &input_info;
 		}
 		const char* label = nullptr;
+		const char* stage_name = nullptr;
 		switch (stage) {
-			case ShaderType::Vertex: label = "ShaderRecompiler VS"; break;
-			case ShaderType::Mesh: label = "ShaderRecompiler MS"; break;
-			case ShaderType::Local: label = "ShaderRecompiler LS"; break;
-			case ShaderType::TessellationControl: label = "ShaderRecompiler HS"; break;
-			case ShaderType::TessellationEvaluation: label = "ShaderRecompiler DS"; break;
-			case ShaderType::Pixel: label = "ShaderRecompiler PS"; break;
-			case ShaderType::Compute: label = "ShaderRecompiler CS"; break;
+			case ShaderType::Vertex: label = "ShaderRecompiler VS"; stage_name = "vs"; break;
+			case ShaderType::Mesh: label = "ShaderRecompiler MS"; stage_name = "ms"; break;
+			case ShaderType::Local: label = "ShaderRecompiler LS"; stage_name = "ls"; break;
+			case ShaderType::TessellationControl: label = "ShaderRecompiler HS"; stage_name = "hs"; break;
+			case ShaderType::TessellationEvaluation: label = "ShaderRecompiler DS"; stage_name = "ds"; break;
+			case ShaderType::Pixel: label = "ShaderRecompiler PS"; stage_name = "ps"; break;
+			case ShaderType::Compute: label = "ShaderRecompiler CS"; stage_name = "cs"; break;
 			default: EXIT("invalid pipeline shader stage\n");
 		}
 		ShaderRecompiler::CompileOptions options;
@@ -351,12 +352,8 @@ struct PipelineCache::ProgramCache {
 		} else {
 			options.wave_size = input_info.wave_size;
 		}
+		DumpShaderOriginal(stage_name, options.shader_hash, params.code);
 		auto translated = ShaderRecompiler::TranslateProgram(params.code, options);
-		if (translated.skip_dispatch) {
-			entry = programs.try_emplace(lookup_key, ShaderRecompiler::IR::ResourcePlan {}).first;
-			entry->second.skip_dispatch = true;
-			return {};
-		}
 		if (entry == programs.end()) {
 			entry = programs.try_emplace(lookup_key,
 			    ShaderRecompiler::IR::ExtractResourcePlan(translated.program)).first;
@@ -365,7 +362,7 @@ struct PipelineCache::ProgramCache {
 			    entry->second.specialization));
 		}
 		entry->second.permutations.push_back(CompilePermutation(
-		    params, options, std::move(translated), entry->second.specialization, push_data_cursor));
+		    stage_name, options, std::move(translated), entry->second.specialization, push_data_cursor));
 		const auto& permutation = entry->second.permutations.back();
 		input_info.stage = {.program = &permutation.program, .resources = &entry->second.resources};
 		permutation.program.bindings.AdvancePushData(push_data_cursor);
@@ -514,7 +511,6 @@ void PipelineCache::InitializeDriverCache() {
 }
 
 void PipelineCache::Save() {
-	Common::LockGuard lock(m_mutex);
 	if (m_driver_cache == nullptr) {
 		return;
 	}
@@ -605,21 +601,30 @@ PipelineCache::GraphicsPrograms PipelineCache::GetGraphicsPrograms(
 	}
 	ShaderParams pixel_params;
 	if (pixel_active) {
-		pixel_params = PrepareProgram(pixel_regs, sh, target_export_mapping, pixel_info);
-		const auto& blend          = context.GetBlendControl(0);
-		const auto  is_dual_source = [](uint8_t factor) {
-			return factor >= static_cast<uint8_t>(Prospero::BlendFactor::kSrc1Color) &&
-			       factor <= static_cast<uint8_t>(Prospero::BlendFactor::kOneMinusSrc1Alpha);
-		};
+		pixel_params      = PrepareProgram(pixel_regs, sh, target_export_mapping, pixel_info);
+		const auto& blend = context.GetBlendControl(0);
 		pixel_info.dual_source_blending =
 		    blend.enable && !context.GetRenderTarget(0).info.blend_bypass &&
-		    (is_dual_source(blend.color_srcblend) || is_dual_source(blend.color_destblend) ||
-		     (blend.separate_alpha_blend &&
-		      (is_dual_source(blend.alpha_srcblend) || is_dual_source(blend.alpha_destblend))));
+		    (BlendFactorIsDualSource(blend.color_srcblend) ||
+		     BlendFactorIsDualSource(blend.color_destblend) ||
+		     (blend.separate_alpha_blend && (BlendFactorIsDualSource(blend.alpha_srcblend) ||
+		                                     BlendFactorIsDualSource(blend.alpha_destblend))));
 		if (pixel_info.dual_source_blending) {
-			// MRT1 supplies a second blend source for the same render target as MRT0.
+			// MRT1 supplies the second blend source for target 0.
 			pixel_info.target_output_mode[1]    = pixel_info.target_output_mode[0];
 			pixel_info.target_export_mapping[1] = pixel_info.target_export_mapping[0];
+		} else if (blend.enable && !context.GetRenderTarget(0).info.blend_bypass &&
+		           pixel_info.target_output_mode[0] != 0 && pixel_info.target_output_mode[0] != 7 &&
+		           std::all_of(std::begin(pixel_info.target_output_mode) + 1,
+		                       std::end(pixel_info.target_output_mode),
+		                       [](uint8_t mode) { return mode == 0; }) &&
+		           ClassifyBlendMapping(blend, pixel_info.target_export_mapping[0]) ==
+		               BlendMappingSupport::SourceAlpha) {
+			// Preserve logical alpha when the export mapping moves it.
+			pixel_info.alpha_blend_source_remap = true;
+			pixel_info.dual_source_blending     = true;
+			pixel_info.target_output_mode[1]    = pixel_info.target_output_mode[0];
+			pixel_info.target_export_mapping[1] = {};
 		}
 	}
 	if (context.GetClipControl().clip_disable) {
@@ -636,7 +641,6 @@ PipelineCache::GraphicsPrograms PipelineCache::GetGraphicsPrograms(
 		    static_cast<float>(std::min(limits.maxViewportDimensions[1], 16384u)) * 0.5f;
 		clip.enabled = true;
 	}
-	Common::LockGuard lock(m_mutex);
 	uint32_t          push_data_cursor =
 	    mesh_active ? ShaderRecompiler::IR::PushData::MeshDrawDwordCount : 0;
 	GraphicsPrograms  result;
@@ -654,7 +658,6 @@ ShaderProgram PipelineCache::GetComputeProgram(const HW::ComputeShaderInfo& regs
                                                ShaderComputeInputInfo&      input_info) {
 	input_info.host_subgroup_size = m_graphics.SupportsComputeWave64() ? 64u : 32u;
 	const auto        params      = PrepareProgram(regs, sh, input_info);
-	Common::LockGuard lock(m_mutex);
 	uint32_t          push_data_cursor = 0;
 	return m_program_cache->Get(params, input_info, push_data_cursor);
 }
@@ -679,7 +682,6 @@ PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(
 	EXIT_IF(ps_active && !pixel_program);
 	const auto color_count = static_cast<uint32_t>(colors.size());
 
-	Common::LockGuard lock(m_mutex);
 	auto&             ctx = command.GetRegisters();
 
 	const HW::ModeControl& mc = ctx.GetModeControl();
@@ -712,14 +714,35 @@ PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(
 		}
 		const auto& rt                        = ctx.GetRenderTarget(colors[i].target_slot);
 		const auto& bc                        = ctx.GetBlendControl(colors[i].target_slot);
-		static_params.color_srcblend[slot]       = bc.color_srcblend;
-		static_params.color_comb_fcn[slot]       = bc.color_comb_fcn;
-		static_params.color_destblend[slot]      = bc.color_destblend;
-		static_params.alpha_srcblend[slot]       = bc.alpha_srcblend;
-		static_params.alpha_comb_fcn[slot]       = bc.alpha_comb_fcn;
-		static_params.alpha_destblend[slot]      = bc.alpha_destblend;
-		static_params.separate_alpha_blend[slot] = bc.separate_alpha_blend;
-		static_params.blend_enable[slot]         = bc.enable && !rt.info.blend_bypass;
+		const bool alpha_remap =
+		    slot == 0 && ps_input_info != nullptr && ps_input_info->alpha_blend_source_remap;
+		static_params.blend_enable[slot] = bc.enable && !rt.info.blend_bypass;
+		if (static_params.blend_enable[slot] && !alpha_remap &&
+		    ClassifyBlendMapping(bc, colors[i].export_mapping) != BlendMappingSupport::Direct) {
+			static_params.blend_enable[slot] = false;
+			static std::atomic_bool warned = false;
+			if (!warned.exchange(true, std::memory_order_relaxed)) {
+				Log::WriteToConsoleAndLog(fmt::format(
+				    "Warning: blending disabled for unsupported color mapping "
+				    "(slot={} mapping=0x{:02x} color={}/{} alpha={}/{} separate={}).\n",
+				    slot, colors[i].export_mapping.packed, bc.color_srcblend, bc.color_destblend,
+				    bc.alpha_srcblend, bc.alpha_destblend, bc.separate_alpha_blend ? 1 : 0));
+			}
+		}
+		if (alpha_remap) {
+			static_params.blend_alpha_source_remap = true;
+		}
+		if (static_params.blend_enable[slot]) {
+			static_params.color_srcblend[slot]       = bc.color_srcblend;
+			static_params.color_comb_fcn[slot]       = bc.color_comb_fcn;
+			static_params.color_destblend[slot]      = bc.color_destblend;
+			static_params.separate_alpha_blend[slot] = bc.separate_alpha_blend;
+			if (bc.separate_alpha_blend) {
+				static_params.alpha_srcblend[slot]  = bc.alpha_srcblend;
+				static_params.alpha_comb_fcn[slot]  = bc.alpha_comb_fcn;
+				static_params.alpha_destblend[slot] = bc.alpha_destblend;
+			}
+		}
 	}
 	const bool with_depth =
 	    depth.desc.view_info.format != vk::Format::eUndefined && static_cast<bool>(depth.image_id);
@@ -783,24 +806,20 @@ PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(
 		        vs_input_info.resources_num > ShaderVertexInputInfo::RES_MAX);
 		key.vertex_input.binding_count   = static_cast<uint8_t>(vs_input_info.buffers_num);
 		key.vertex_input.attribute_count = static_cast<uint8_t>(vs_input_info.resources_num);
-		uint32_t attributes_num          = 0;
 		for (int binding = 0; binding < vs_input_info.buffers_num; binding++) {
 			const auto& buffer = vs_input_info.buffers[binding];
-			EXIT_IF(buffer.attr_num < 0 || buffer.attr_num > ShaderVertexInputBuffer::ATTR_MAX);
-			attributes_num += static_cast<uint32_t>(buffer.attr_num);
-			EXIT_IF(attributes_num > static_cast<uint32_t>(vs_input_info.resources_num));
 			key.vertex_input.bindings[binding] = {.stride   = buffer.stride,
 			                                      .instance = buffer.fetch_index != 0};
-			for (int attribute = 0; attribute < buffer.attr_num; attribute++) {
-				const auto index = buffer.attr_indices[attribute];
-				EXIT_IF(index < 0 || index >= vs_input_info.resources_num);
-				key.vertex_input.attributes[index] = {
-				    .offset  = buffer.attr_offsets[attribute],
-				    .binding = static_cast<uint8_t>(binding),
-				};
-			}
 		}
-		EXIT_IF(attributes_num != static_cast<uint32_t>(vs_input_info.resources_num));
+		for (int attribute = 0; attribute < vs_input_info.resources_num; attribute++) {
+			const auto binding = vs_input_info.resources_dst[attribute].buffer_index;
+			EXIT_IF(binding < 0 || binding >= vs_input_info.buffers_num);
+			key.vertex_input.attributes[attribute] = {
+			    .offset = static_cast<uint32_t>(vs_input_info.resources[attribute].Base48() -
+			                                    vs_input_info.buffers[binding].addr),
+			    .binding = static_cast<uint8_t>(binding),
+			};
+		}
 	}
 
 	if (auto iter = m_graphics_pipelines.find(key); iter != m_graphics_pipelines.end()) {
@@ -839,8 +858,6 @@ PipelineCache::GetComputePipeline(const ShaderComputeInputInfo& input_info,
 
 	EXIT_IF(!compute_program);
 
-	Common::LockGuard lock(m_mutex);
-
 	if (auto iter = m_compute_pipelines.find(compute_program.id);
 	    iter != m_compute_pipelines.end()) {
 		return *iter->second;

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

@@ -4,7 +4,6 @@
 #include "common/abi.h"
 #include "common/assert.h"
 #include "common/common.h"
-#include "common/threads.h"
 #include "graphics/host_gpu/renderer/renderTarget.h"
 #include "graphics/host_gpu/vulkanCommon.h"
 #include "graphics/shader/shader.h"
@@ -56,6 +55,7 @@ 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                       blend_alpha_source_remap                           = false;
 
 	bool operator==(const PipelineStaticParameters& other) const noexcept;
 };
@@ -65,7 +65,7 @@ struct PipelineStaticParameters {
 static_assert(std::is_trivially_copyable_v<PipelineStaticParameters>);
 static_assert(std::is_standard_layout_v<PipelineStaticParameters>);
 static_assert(alignof(PipelineStaticParameters) == 1);
-static_assert(sizeof(PipelineStaticParameters) == 125);
+static_assert(sizeof(PipelineStaticParameters) == 126);
 
 struct PipelineRenderingState {
 	std::array<vk::Format, RENDER_COLOR_ATTACHMENTS_MAX> color_formats {};
@@ -103,6 +103,7 @@ struct ShaderProgram {
 	explicit operator bool() const { return id != 0 && module != nullptr; }
 };
 
+// The owning renderer serializes access, including saves while the GPU is running.
 class PipelineCache {
 public:
 	explicit PipelineCache(GraphicContext& graphics);
@@ -166,45 +167,10 @@ private:
 			hash ^= value + static_cast<std::size_t>(0x9e3779b97f4a7c15ull) + (hash << 6u) +
 			        (hash >> 2u);
 		}
-
-		static void MixStaticParams(std::size_t& hash, const PipelineStaticParameters& params) {
-			const auto* bytes = reinterpret_cast<const uint8_t*>(&params);
-			for (std::size_t i = 0; i < sizeof(params); i++) {
-				Mix(hash, bytes[i]);
-			}
-		}
-
-		static void MixRendering(std::size_t& hash, const PipelineRenderingState& rendering) {
-			Mix(hash, rendering.color_count);
-			for (uint32_t i = 0; i < rendering.color_count; i++) {
-				Mix(hash, static_cast<uint32_t>(rendering.color_formats[i]));
-			}
-			Mix(hash, static_cast<uint32_t>(rendering.depth_format));
-			Mix(hash, static_cast<uint32_t>(rendering.stencil_format));
-		}
 	};
 
 	struct GraphicsPipelineKeyHash {
-		std::size_t operator()(const GraphicsPipelineKey& key) const {
-			std::size_t hash = 0;
-			PipelineKeyHash::MixRendering(hash, key.rendering);
-			for (const auto id: key.vertex_shader_ids) {
-				PipelineKeyHash::Mix(hash, id);
-			}
-			PipelineKeyHash::Mix(hash, key.ps_shader_id);
-			PipelineKeyHash::Mix(hash, key.vertex_input.binding_count);
-			for (uint32_t i = 0; i < key.vertex_input.binding_count; i++) {
-				PipelineKeyHash::Mix(hash, key.vertex_input.bindings[i].stride);
-				PipelineKeyHash::Mix(hash, key.vertex_input.bindings[i].instance);
-			}
-			PipelineKeyHash::Mix(hash, key.vertex_input.attribute_count);
-			for (uint32_t i = 0; i < key.vertex_input.attribute_count; i++) {
-				PipelineKeyHash::Mix(hash, key.vertex_input.attributes[i].offset);
-				PipelineKeyHash::Mix(hash, key.vertex_input.attributes[i].binding);
-			}
-			PipelineKeyHash::MixStaticParams(hash, key.static_params);
-			return hash;
-		}
+		std::size_t operator()(const GraphicsPipelineKey& key) const;
 	};
 
 	GraphicContext&               m_graphics;
@@ -214,7 +180,6 @@ private:
 	std::unordered_map<GraphicsPipelineKey, std::unique_ptr<Pipeline>, GraphicsPipelineKeyHash>
 	                                                        m_graphics_pipelines;
 	std::unordered_map<uint64_t, std::unique_ptr<Pipeline>> m_compute_pipelines;
-	Common::Mutex m_mutex;
 
 	void InitializeDriverCache();
 };

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

@@ -137,14 +137,17 @@ static void GetInputFormat(const ShaderBufferResource& res, vk::Format& format,
 	}
 }
 
-static vk::BlendFactor GetBlendFactor(uint32_t factor) {
+static vk::BlendFactor GetBlendFactor(uint32_t factor, bool remap_source_alpha) {
 	switch (static_cast<Prospero::BlendFactor>(factor)) {
 		case Prospero::BlendFactor::kZero: return vk::BlendFactor::eZero;
 		case Prospero::BlendFactor::kOne: return vk::BlendFactor::eOne;
 		case Prospero::BlendFactor::kSrcColor: return vk::BlendFactor::eSrcColor;
 		case Prospero::BlendFactor::kOneMinusSrcColor: return vk::BlendFactor::eOneMinusSrcColor;
-		case Prospero::BlendFactor::kSrcAlpha: return vk::BlendFactor::eSrcAlpha;
-		case Prospero::BlendFactor::kOneMinusSrcAlpha: return vk::BlendFactor::eOneMinusSrcAlpha;
+		case Prospero::BlendFactor::kSrcAlpha:
+			return remap_source_alpha ? vk::BlendFactor::eSrc1Color : vk::BlendFactor::eSrcAlpha;
+		case Prospero::BlendFactor::kOneMinusSrcAlpha:
+			return remap_source_alpha ? vk::BlendFactor::eOneMinusSrc1Color
+			                          : vk::BlendFactor::eOneMinusSrcAlpha;
 		case Prospero::BlendFactor::kDstAlpha: return vk::BlendFactor::eDstAlpha;
 		case Prospero::BlendFactor::kOneMinusDstAlpha: return vk::BlendFactor::eOneMinusDstAlpha;
 		case Prospero::BlendFactor::kDstColor: return vk::BlendFactor::eDstColor;
@@ -327,9 +330,6 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 			}
 		}
 
-		EXIT_NOT_IMPLEMENTED(vs_input_info.resources[index].AddTid());
-		EXIT_NOT_IMPLEMENTED(vs_input_info.resources[index].SwizzleEnabled());
-
 		EXIT_IF(registers_num < 1 || registers_num > 4);
 	}
 
@@ -397,17 +397,20 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 		color_blend_attachment[i].colorWriteMask =
 		    vk::ColorComponentFlags {static_params.color_mask[i]};
 		color_blend_attachment[i].blendEnable = static_params.blend_enable[i] ? VK_TRUE : VK_FALSE;
+		// Only target 0 can use the second blend source carrying logical alpha.
+		const bool remap_source_alpha         = i == 0 && static_params.blend_alpha_source_remap;
 		color_blend_attachment[i].srcColorBlendFactor =
-		    GetBlendFactor(static_params.color_srcblend[i]);
+		    GetBlendFactor(static_params.color_srcblend[i], remap_source_alpha);
 		color_blend_attachment[i].dstColorBlendFactor =
-		    GetBlendFactor(static_params.color_destblend[i]);
+		    GetBlendFactor(static_params.color_destblend[i], remap_source_alpha);
 		color_blend_attachment[i].colorBlendOp = GetBlendOp(static_params.color_comb_fcn[i]);
 		color_blend_attachment[i].srcAlphaBlendFactor =
-		    (static_params.separate_alpha_blend[i] ? GetBlendFactor(static_params.alpha_srcblend[i])
-		                                           : color_blend_attachment[i].srcColorBlendFactor);
+		    (static_params.separate_alpha_blend[i]
+		         ? GetBlendFactor(static_params.alpha_srcblend[i], remap_source_alpha)
+		         : color_blend_attachment[i].srcColorBlendFactor);
 		color_blend_attachment[i].dstAlphaBlendFactor =
 		    (static_params.separate_alpha_blend[i]
-		         ? GetBlendFactor(static_params.alpha_destblend[i])
+		         ? GetBlendFactor(static_params.alpha_destblend[i], remap_source_alpha)
 		         : color_blend_attachment[i].dstColorBlendFactor);
 		color_blend_attachment[i].alphaBlendOp =
 		    (static_params.separate_alpha_blend[i] ? GetBlendOp(static_params.alpha_comb_fcn[i])

src/graphics/host_gpu/renderer/renderCompute.cpp

@@ -256,10 +256,6 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 	input_info.dispatch_thread_dimensions = use_thread_dimensions;
 	const auto compute_program =
 	    m_context.GetPipelineCache().GetComputeProgram(cs_regs, sh_regs, input_info);
-	if (!compute_program) {
-		// Temporary until RT is implemented.
-		return;
-	}
 	if (use_thread_dimensions) {
 		input_info.dispatch_threads_num[0]    = thread_group_x;
 		input_info.dispatch_threads_num[1]    = thread_group_y;
@@ -268,15 +264,14 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 
 	const auto& program   = *input_info.stage.program;
 	const auto& resources = *input_info.stage.resources;
-	if (TryConsumeComputeMetaClear(input_info, buffer)) {
-		ResetBindings();
-		return;
-	}
-	if (TryConsumeComputeImageClear(input_info, buffer, thread_group_x, thread_group_y,
-	                                thread_group_z, mode)) {
+	if (resources.specialization_reads.empty() &&
+	    (TryConsumeComputeMetaClear(input_info, buffer) ||
+	     TryConsumeComputeImageClear(input_info, buffer, thread_group_x, thread_group_y,
+	                                 thread_group_z, mode))) {
 		ResetBindings();
 		return;
 	}
+
 	const bool large_workgroup =
 	    (input_info.threads_num[0] * input_info.threads_num[1] * input_info.threads_num[2] >= 512);
 	const bool                   has_sampler = !program.info.samplers.empty();
@@ -417,10 +412,6 @@ void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
 	ShaderComputeInputInfo input_info {};
 	const auto compute_program = m_context.GetPipelineCache().GetComputeProgram(
 	    cs_regs, buffer.GetRegisters().GetShaderRegisters(), input_info);
-	if (!compute_program) {
-		// Temporary until RT is implemented.
-		return;
-	}
 	buffer.EndRendering();
 	auto& pipeline = m_context.GetPipelineCache().GetComputePipeline(input_info, compute_program);
 	auto& bindings = m_compute_bindings;

src/graphics/host_gpu/renderer/renderContext.cpp

@@ -119,6 +119,10 @@ void RenderContext::UnmapMemory(uint64_t vaddr, uint64_t size) {
 }
 
 void RenderContext::PrepareBda() {
+	if (!m_bda_logged) {
+		Log::WriteToConsoleAndLog("GPU: using buffer device address (BDA) shader memory access.\n");
+		m_bda_logged = true;
+	}
 	std::shared_lock lock(m_mapped_ranges_mutex);
 	m_mapped_ranges.ForEach([this](uint64_t start, uint64_t end) {
 		m_buffer_cache.SynchronizeBuffersInRange(start, end - start);

src/graphics/host_gpu/renderer/renderContext.h

@@ -81,6 +81,7 @@ private:
 	std::unique_ptr<GuestGpu> m_gpu;
 	VideoOut::VideoOutDriver* m_video_out = nullptr;
 	bool                      m_fault_process_pending = false;
+	bool                      m_bda_logged = false;
 
 	Common::Mutex                        m_interrupt_mutex;
 	std::vector<InterruptEqRegistration> m_interrupt_eqs;

src/graphics/host_gpu/renderer/renderDraw.cpp

@@ -249,8 +249,8 @@ static void LogDrawInputState(const CommandBuffer& buffer, const RenderColorInfo
 	for (int bi = 0; bi < vs_input_info.buffers_num; bi++) {
 		const auto& b = vs_input_info.buffers[bi];
 		LOGF("DrawInputState[%u]: vb[%d] addr=0x%010" PRIx64
-		     " stride=%u records=%u fetch_index=%u attr_num=%d\n",
-		     log_id, bi, b.addr, b.stride, b.num_records, b.fetch_index, b.attr_num);
+		     " stride=%u records=%u fetch_index=%u\n",
+		     log_id, bi, b.addr, b.stride, b.num_records, b.fetch_index);
 
 		const auto* bytes = reinterpret_cast<const uint8_t*>(b.addr);
 		if (bytes != nullptr && b.stride != 0) {
@@ -272,11 +272,13 @@ static void LogDrawInputState(const CommandBuffer& buffer, const RenderColorInfo
 				     raw[5], raw[6], raw[7], raw[8], flt[0], flt[1], flt[2], flt[3], flt[4], flt[5],
 				     flt[6], flt[7], flt[8]);
 
-				for (int ai = 0; ai < b.attr_num; ai++) {
-					const auto  res_index = b.attr_indices[ai];
-					const auto& r         = vs_input_info.resources[res_index];
-					const auto& rd        = vs_input_info.resources_dst[res_index];
-					const auto  offset    = b.attr_offsets[ai];
+				for (int ai = 0; ai < vs_input_info.resources_num; ai++) {
+					const auto& r  = vs_input_info.resources[ai];
+					const auto& rd = vs_input_info.resources_dst[ai];
+					if (rd.buffer_index != bi) {
+						continue;
+					}
+					const auto offset = static_cast<uint32_t>(r.Base48() - b.addr);
 					if (offset + 4u <= b.stride &&
 					    r.Format() == Prospero::BufferFormat::k8_8_8_8UNorm) {
 						uint32_t packed = 0;
@@ -296,14 +298,16 @@ static void LogDrawInputState(const CommandBuffer& buffer, const RenderColorInfo
 			}
 		}
 
-		for (int ai = 0; ai < b.attr_num; ai++) {
-			const auto  res_index = b.attr_indices[ai];
-			const auto& r         = vs_input_info.resources[res_index];
-			const auto& rd        = vs_input_info.resources_dst[res_index];
-			LOGF("DrawInputState[%u]: attr[%d] res=%d offset=%u dst=v%d regs=%d fetch_index=%u "
+		for (int ai = 0; ai < vs_input_info.resources_num; ai++) {
+			const auto& r  = vs_input_info.resources[ai];
+			const auto& rd = vs_input_info.resources_dst[ai];
+			if (rd.buffer_index != bi) {
+				continue;
+			}
+			LOGF("DrawInputState[%u]: attr[%d] offset=%u dst=v%d regs=%d fetch_index=%u "
 			     "sharp=%08" PRIx32 " %08" PRIx32 " %08" PRIx32 " %08" PRIx32 "\n",
-			     log_id, ai, res_index, b.attr_offsets[ai], rd.register_start, rd.registers_num,
-			     rd.fetch_index, r.fields[0], r.fields[1], r.fields[2], r.fields[3]);
+			     log_id, ai, static_cast<uint32_t>(r.Base48() - b.addr), rd.register_start,
+			     rd.registers_num, rd.fetch_index, r.fields[0], r.fields[1], r.fields[2], r.fields[3]);
 		}
 	}
 }
@@ -575,12 +579,17 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
 	return state;
 }
 
-static bool DrawHasActivePixelShader(const CommandBuffer& buffer) {
-	const auto& ctx              = buffer.GetRegisters();
-	const auto& sh_regs          = ctx.GetShaderRegisters();
-	const bool  has_color_output = (ctx.GetRenderTargetMask() & sh_regs.m_cbShaderMask) != 0;
-	return buffer.GetShaders().GetPs().ps_regs.data_addr != 0 &&
-	       (has_color_output || PixelShaderHasDepthOrCoverageSideEffects(sh_regs));
+static uint32_t DrawColorOutputMask(const HW::Context& ctx) {
+	const auto& sh_regs     = ctx.GetShaderRegisters();
+	const auto  write_mask  = ctx.GetRenderTargetMask() & sh_regs.m_cbShaderMask;
+	uint32_t    output_mask = 0;
+	for (uint32_t slot = 0; slot < RENDER_COLOR_ATTACHMENTS_MAX; slot++) {
+		if (sh_regs.target_output_mode[slot] != 0 &&
+		    render_target_mask_slot(write_mask, slot) != 0) {
+			output_mask |= 1u << slot;
+		}
+	}
+	return output_mask;
 }
 
 enum class CbColorMode : uint8_t {
@@ -625,19 +634,22 @@ struct PreparedIndexBuffer {
 	vk::IndexType  type   = vk::IndexType::eUint16;
 };
 
-static uint64_t VertexBufferDescriptorSize(const ShaderVertexInputBuffer& buffer,
-                                           const ShaderVertexInputInfo& info) {
+static uint64_t VertexBufferDescriptorSize(int binding, const ShaderVertexInputInfo& info) {
+	const auto& buffer = info.buffers[binding];
 	if (buffer.stride != 0 || buffer.num_records == 0) {
 		return static_cast<uint64_t>(buffer.stride) * buffer.num_records;
 	}
 
 	uint64_t size = 0;
-	for (int i = 0; i < buffer.attr_num; i++) {
-		const auto& resource = info.resources[buffer.attr_indices[i]];
+	for (int i = 0; i < info.resources_num; i++) {
+		if (info.resources_dst[i].buffer_index != binding) {
+			continue;
+		}
+		const auto& resource = info.resources[i];
 		// RDNA2 OOB_SELECT=2 only checks NumRecords != 0. A constant attribute still
 		// fetches its entire format; NumRecords is not a byte count in this mode.
 		const uint64_t extent = resource.OutOfBounds() == 2
-		                            ? static_cast<uint64_t>(buffer.attr_offsets[i]) +
+		                            ? resource.Base48() - buffer.addr +
 		                                  ShaderRecompiler::Format::GetFormatInfo(resource.Format()).byte_size
 		                            : buffer.num_records;
 		size = std::max(size, extent);
@@ -674,7 +686,7 @@ static PreparedVertexBuffers AcquireVertexBuffers(CommandBuffer&               b
 	uint32_t                                                      range_count = 0;
 	for (int i = 0; i < vs_input_info.buffers_num; i++) {
 		const auto& vertex = vs_input_info.buffers[i];
-		const auto  size   = VertexBufferDescriptorSize(vertex, vs_input_info);
+		const auto  size   = VertexBufferDescriptorSize(i, vs_input_info);
 		sizes[i]           = size;
 		if (size == 0) {
 			continue;
@@ -713,9 +725,6 @@ static PreparedVertexBuffers AcquireVertexBuffers(CommandBuffer&               b
 		    Libs::LibKernel::Memory::ClampRangeSize(range.base_address, range.RequestedSize());
 		range.acquired_end = range.base_address + size;
 		range.binding      = cache.ObtainBuffer(range.base_address, size, false);
-		SetVulkanObjectNameF(
-		    buffer.GetContext().GetGraphics().device, range.binding.first->Handle(),
-		    "Kyty.VertexBufferRange[guest=0x{:016x} size=0x{:x}]", range.base_address, size);
 	}
 
 	// Rebuild slot bindings, offsetting non-empty slots into their acquired merged range.
@@ -747,10 +756,6 @@ static PreparedVertexBuffers AcquireVertexBuffers(CommandBuffer&               b
 		prepared.buffers[i] = range->binding.first->Handle();
 		prepared.offsets[i] = range->binding.second + vertex.addr - range->base_address;
 		prepared.sizes[i]   = std::min(size, range->acquired_end - vertex.addr);
-		SetVulkanObjectNameF(
-		    buffer.GetContext().GetGraphics().device, prepared.buffers[i],
-		    "Kyty.VertexBuffer[slot={} guest=0x{:016x} size=0x{:x} stride={} records={}]", i,
-		    vertex.addr, size, vertex.stride, vertex.num_records);
 	}
 
 	return prepared;
@@ -847,7 +852,7 @@ static bool ResolvePrimitiveRestart(const CommandBuffer& buffer,
 }
 
 static void RefreshShaders(CommandBuffer& buffer, const DrawCallInfo& draw,
-                           DrawRenderState& state) {
+                           uint32_t color_output_mask, DrawRenderState& state) {
 	auto& ctx    = buffer.GetRegisters();
 	auto& sh_ctx = buffer.GetShaders();
 
@@ -861,7 +866,7 @@ static void RefreshShaders(CommandBuffer& buffer, const DrawCallInfo& draw,
 	    target_export_mapping {};
 	for (uint32_t slot = 0; slot < RENDER_COLOR_ATTACHMENTS_MAX; slot++) {
 		const auto& rt = ctx.GetRenderTarget(slot);
-		if (rt.base.addr != 0 && render_target_mask_slot(ctx.GetRenderTargetMask(), slot) != 0) {
+		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)
@@ -880,8 +885,12 @@ static void RefreshShaders(CommandBuffer& buffer, const DrawCallInfo& draw,
 bool RenderExecutor::PrepareDrawRenderState(CommandBuffer& buffer, const DrawCallInfo& draw,
                                             uint32_t            render_target_slice_offset,
 	                                        DrawRenderState& state) {
-	state.ps_active = DrawHasActivePixelShader(buffer);
-	RefreshShaders(buffer, draw, state);
+	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) {
@@ -890,6 +899,7 @@ bool RenderExecutor::PrepareDrawRenderState(CommandBuffer& buffer, const DrawCal
 			}
 		}
 	}
+	mrt_mask &= color_output_mask;
 	if (draw.IsIndexed()) {
 		LogDrawPhase(draw.Name(), "ResolveRenderColorTarget");
 	}
@@ -927,17 +937,11 @@ static PreparedIndexBuffer PrepareIndexBuffer(CommandBuffer&               buffe
 		auto& stream = buffer.GetContext().GetBufferCache().GetUtilityBuffer(MemoryUsage::Stream);
 		prepared.offset = stream.Copy(source.host_data, source.size, 16);
 		prepared.buffer = stream.Handle();
-		SetVulkanObjectNameF(buffer.GetContext().GetGraphics().device, prepared.buffer,
-		                     "Kyty.IndexBuffer[guest=transient size=0x{:x} type={}]", source.size,
-		                     static_cast<uint32_t>(source.type));
 	} else {
 		auto [buffer_ptr, offset] =
 		    buffer.GetContext().GetBufferCache().ObtainBuffer(source.address, source.size, false);
 		prepared.buffer = buffer_ptr->Handle();
 		prepared.offset = offset;
-		SetVulkanObjectNameF(buffer.GetContext().GetGraphics().device, prepared.buffer,
-		                     "Kyty.IndexBuffer[guest=0x{:016x} size=0x{:x} type={}]",
-		                     source.address, source.size, static_cast<uint32_t>(source.type));
 	}
 	return prepared;
 }
@@ -1028,6 +1032,7 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 	uint32_t   mesh_groups = 0;
 	if (mesh_active) {
 		const auto& mesh = state.vertex_info[0].mesh;
+		EXIT_NOT_IMPLEMENTED(mesh.fast_launch && (draw.IsIndexed() || primitive_restart_enable));
 		static std::atomic_bool restart_warned = false;
 		if (primitive_restart_enable && !restart_warned.exchange(true, std::memory_order_relaxed)) {
 			std::printf("Warning: primitive restart is not implemented for mesh shaders; "

src/graphics/host_gpu/renderer/sync.cpp

@@ -172,7 +172,7 @@ uint64_t PrepareVideoOutFlip(CommandBuffer& buffer, int handle, int index, int f
 			     "\n",
 			     result, handle, index, flip_mode, flip_arg);
 		}
-		video_out.WaitForSubmitSlot();
+		video_out.WaitForSubmitSlot(handle);
 	}
 }
 

src/graphics/host_gpu/shaders/gpu_video_out_overlay.frag

@@ -0,0 +1,15 @@
+#version 450
+
+layout(location = 0) in vec2 uv;
+layout(location = 0) out vec4 color;
+layout(binding = 0) uniform sampler2D source;
+layout(push_constant) uniform AlphaControl {
+	uint premultiplied;
+};
+
+void main() {
+	color = texture(source, uv);
+	if (premultiplied == 0u) {
+		color.rgb *= color.a;
+	}
+}

src/graphics/presentation/presenter.h

@@ -4,6 +4,7 @@
 #include "common/common.h"
 
 #include <memory>
+#include <span>
 
 namespace Libs::Graphics {
 
@@ -15,6 +16,11 @@ struct WindowContext;
 class Presenter final {
 public:
 	struct Frame;
+	struct Layer {
+		Frame* frame;
+		int    bus;
+		bool   premultiplied_alpha;
+	};
 
 	explicit Presenter(WindowContext& window);
 	~Presenter();
@@ -23,11 +29,13 @@ public:
 	[[nodiscard]] Frame&         PrepareFrame(CommandBuffer& command, const ImageInfo& info);
 	[[nodiscard]] Frame&         PrepareBlankFrame(uint32_t width, uint32_t height, bool opaque,
 	                                               CommandBuffer* producer = nullptr);
-	[[nodiscard]] Frame*         PrepareLastFrame();
+	[[nodiscard]] bool           PresentLastFrame();
 	[[nodiscard]] bool           IsGuestPaused() const noexcept;
 	[[nodiscard]] bool           NeedsSystemOverlayRefresh() const noexcept;
 	[[nodiscard]] RenderContext& Renderer() const noexcept;
-	void                         Present(Frame& frame, bool reuse = false);
+	void                         Present(Frame& frame);
+	void                         Present(std::span<const Layer> layers);
+	void                         ClearLayer(int bus);
 	void                         Discard(Frame& frame);
 
 private:

src/graphics/presentation/videoOut.cpp

@@ -51,6 +51,8 @@ constexpr int      VIDEO_OUT_BUS_TYPE_OVERLAY                           = 1;
 constexpr int      VIDEO_OUT_BUS_TYPE_SUB                               = 2;
 constexpr int      VIDEO_OUT_FLIP_MODE_VSYNC                            = 1;
 constexpr int      VIDEO_OUT_FLIP_MODE_VSYNC_MULTI                      = 4;
+constexpr int      VIDEO_OUT_FLIP_MODE_SLAVE                            = 8;
+constexpr int      VIDEO_OUT_FLIP_MODE_MASTER                           = 9;
 constexpr int      VIDEO_OUT_BUFFER_INDEX_BLACK                         = -2;
 constexpr int      VIDEO_OUT_BUFFER_INDEX_BLANK                         = -1;
 constexpr int      VIDEO_OUT_BUFFER_NUM_MAX                             = 16;
@@ -63,6 +65,7 @@ constexpr uint64_t VIDEO_OUT_REFRESH_RATE_119_88HZ                      = 13;
 constexpr int      VIDEO_OUT_BUFFER_ATTRIBUTE_CATEGORY_UNCOMPRESSED     = 0;
 constexpr int      VIDEO_OUT_BUFFER_ATTRIBUTE_CATEGORY_COMPRESSED       = 1;
 constexpr uint64_t VIDEO_OUT_BUFFER_ATTRIBUTE_OPTION_STRICT_COLORIMETRY = 8;
+constexpr uint64_t VIDEO_OUT_BUFFER_ATTRIBUTE_OPTION_ALPHA_PREMULTIPLIED = 32;
 
 enum class VideoOutEventKind : uintptr_t {
 	Flip           = VIDEO_OUT_EVENT_FLIP,
@@ -183,6 +186,9 @@ struct VideoOutConfig {
 	uint64_t                            generation  = 0;
 	bool                                opened      = false;
 	bool                                closing     = false;
+	int                                 bus         = 0;
+	VideoOutConfig*                     master      = nullptr;
+	uint32_t                            slave_count = 0;
 	int                                 flip_rate   = 0;
 	uint64_t                            output_mode = VIDEO_OUT_OUTPUT_MODE_DEFAULT;
 	float                               gamma       = 1.0f;
@@ -206,7 +212,7 @@ public:
 	void Cancel(VideoOutConfig& cfg);
 	void Prepare(uint64_t request_id, Graphics::CommandBuffer& buffer);
 	void Complete(uint64_t request_id);
-	void WaitForSubmitSlot();
+	void WaitForSubmitSlot(VideoOutConfig& cfg);
 	bool Flip(uint32_t micros);
 	void GetFlipStatus(VideoOutConfig& cfg, VideoOutFlipStatus& out);
 	void Wait(VideoOutConfig& cfg, int index);
@@ -216,7 +222,9 @@ private:
 
 	struct Request {
 		uint64_t                    id;
+		uint64_t                    group;
 		VideoOutConfig*             cfg;
+		VideoOutConfig*             master;
 		uint64_t                    generation;
 		int                         index;
 		int64_t                     flip_arg;
@@ -224,6 +232,7 @@ private:
 		FlipRequestSource           source;
 		RequestState                state;
 		Graphics::Presenter::Frame* frame;
+		bool                        premultiplied_alpha;
 	};
 
 	Graphics::Presenter& m_presenter;
@@ -258,6 +267,7 @@ public:
 	VideoOutConfig* Get(int handle);
 	VideoOutConfig* Get(int handle, uint64_t& generation);
 	bool            IsOpened(int handle);
+	int             SetFlipMaster(int slave_handle, int master_handle);
 
 	void                     Init(uint32_t width, uint32_t height);
 	FlipQueue&               GetFlipQueue() { return m_flip_queue; }
@@ -498,7 +508,8 @@ static bool IsSpecialBufferIndex(int index) {
 }
 
 static bool IsValidFlipMode(int mode) {
-	return mode >= VIDEO_OUT_FLIP_MODE_VSYNC && mode <= VIDEO_OUT_FLIP_MODE_VSYNC_MULTI;
+	return (mode >= VIDEO_OUT_FLIP_MODE_VSYNC && mode <= VIDEO_OUT_FLIP_MODE_VSYNC_MULTI) ||
+	       mode == VIDEO_OUT_FLIP_MODE_SLAVE || mode == VIDEO_OUT_FLIP_MODE_MASTER;
 }
 
 static int ReserveFlipRequest(VideoOutDriver::Impl& driver, int handle, int index, int flip_mode,
@@ -508,13 +519,17 @@ static int ReserveFlipRequest(VideoOutDriver::Impl& driver, int handle, int inde
 		return VIDEO_OUT_ERROR_INVALID_HANDLE;
 	}
 	if (!IsValidFlipMode(flip_mode)) {
-		return VIDEO_OUT_ERROR_INVALID_VALUE;
+		return VIDEO_OUT_ERROR_INVALID_FLIP_MODE;
 	}
 	if (!IsValidBufferIndex(index)) {
 		return VIDEO_OUT_ERROR_INVALID_INDEX;
 	}
 
 	Common::LockGuard lock(video_out->mutex);
+	if ((video_out->master != nullptr) != (flip_mode == VIDEO_OUT_FLIP_MODE_SLAVE) ||
+	    (video_out->slave_count != 0) != (flip_mode == VIDEO_OUT_FLIP_MODE_MASTER)) {
+		return VIDEO_OUT_ERROR_INVALID_FLIP_MODE;
+	}
 	if (video_out->closing ||
 	    (!IsSpecialBufferIndex(index) && !video_out->buffers[index].Occupied())) {
 		return VIDEO_OUT_ERROR_INVALID_INDEX;
@@ -532,8 +547,8 @@ Graphics::ImageInfo BufferAttributeGroup::ImageInfo(const VideoOutBuffer& buffer
 	if (attribute.reserved0 != 0 || attribute.aspect_ratio != 0 || attribute.width == 0 ||
 	    attribute.height == 0 || attribute.width > 16384 || attribute.height > 16384 ||
 	    attribute.pitch_in_pixel != 0 ||
-	    (attribute.option != 0 &&
-	     attribute.option != VIDEO_OUT_BUFFER_ATTRIBUTE_OPTION_STRICT_COLORIMETRY) ||
+	    (attribute.option & ~(VIDEO_OUT_BUFFER_ATTRIBUTE_OPTION_STRICT_COLORIMETRY |
+	                          VIDEO_OUT_BUFFER_ATTRIBUTE_OPTION_ALPHA_PREMULTIPLIED)) != 0 ||
 	    attribute.tiling_mode != 0 || attribute.pad0 != 0 || attribute.reserved1[0] != 0 ||
 	    attribute.reserved1[1] != 0 || attribute.reserved1[2] != 0 || buffer.data_address == 0 ||
 	    compression == Graphics::VideoOutCompression::Unsupported) {
@@ -616,6 +631,11 @@ VideoOutDriver::Impl::~Impl() {
 		m_present_thread.request_stop();
 		m_present_thread.join();
 	}
+	for (int handle = 1; handle < VIDEO_OUT_NUM_MAX; handle++) {
+		if (m_video_out_ctx[handle].master != nullptr) {
+			(void)Close(handle);
+		}
+	}
 	for (int handle = 1; handle < VIDEO_OUT_NUM_MAX; handle++) {
 		(void)Close(handle);
 	}
@@ -660,6 +680,7 @@ int VideoOutDriver::Impl::Open(int bus_type, int index) {
 
 	config.closing = false;
 	config.opened  = true;
+	config.bus     = bus_type;
 	if (++config.generation == 0) {
 		EXIT("video-out port generation wrapped\n");
 	}
@@ -686,11 +707,13 @@ bool VideoOutDriver::Impl::Close(int handle) {
 	VideoOutEventQueues pre_vblank_events;
 	VideoOutEventQueues vblank_events;
 	VideoOutEventQueues output_mode_events;
+	VideoOutConfig* master = nullptr;
 	{
 		Common::LockGuard config_lock(config.mutex);
 		if (config.closing) {
 			return false;
 		}
+		EXIT_IF(config.slave_count != 0);
 		config.opened  = false;
 		config.closing = true;
 		if (++config.generation == 0) {
@@ -703,7 +726,9 @@ bool VideoOutDriver::Impl::Close(int handle) {
 			vblank_events      = std::move(config.events->vblank);
 			output_mode_events = std::move(config.events->output_mode);
 		}
-		config.flip_rate = 0;
+		config.flip_rate   = 0;
+		master             = config.master;
+		config.master      = nullptr;
 
 		for (const auto& buffer: config.buffers) {
 			if (buffer.Occupied() &&
@@ -720,8 +745,12 @@ bool VideoOutDriver::Impl::Close(int handle) {
 		}
 		config.vblank_cond.SignalAll();
 	}
-
+	if (master != nullptr) {
+		Common::LockGuard master_lock(master->mutex);
+		master->slave_count--;
+	}
 	m_flip_queue.Cancel(config);
+	m_presenter.ClearLayer(config.bus);
 	DeleteVideoOutEvents(flip_events, VideoOutEventKind::Flip);
 	DeleteVideoOutEvents(pre_vblank_events, VideoOutEventKind::PreVblankStart);
 	DeleteVideoOutEvents(vblank_events, VideoOutEventKind::Vblank);
@@ -759,6 +788,31 @@ bool VideoOutDriver::Impl::IsOpened(int handle) {
 	return handle > 0 && handle < VIDEO_OUT_NUM_MAX && m_video_out_ctx[handle].opened;
 }
 
+int VideoOutDriver::Impl::SetFlipMaster(int slave_handle, int master_handle) {
+	Common::LockGuard lock(m_mutex);
+	if (slave_handle <= 0 || slave_handle >= VIDEO_OUT_NUM_MAX ||
+	    master_handle <= 0 || master_handle >= VIDEO_OUT_NUM_MAX ||
+	    !m_video_out_ctx[slave_handle].opened || !m_video_out_ctx[master_handle].opened) {
+		return VIDEO_OUT_ERROR_INVALID_HANDLE;
+	}
+	auto& slave  = m_video_out_ctx[slave_handle];
+	auto& master = m_video_out_ctx[master_handle];
+	if (&slave == &master) {
+		return VIDEO_OUT_ERROR_UNSUPPORTED_OPERATION;
+	}
+	Common::LockGuard first_lock(m_video_out_ctx[std::min(slave_handle, master_handle)].mutex);
+	Common::LockGuard second_lock(m_video_out_ctx[std::max(slave_handle, master_handle)].mutex);
+	if (slave.master != nullptr || slave.slave_count != 0 || master.master != nullptr) {
+		return VIDEO_OUT_ERROR_UNSUPPORTED_OPERATION;
+	}
+	if (slave.flip_status.flipPendingNum != 0 || master.flip_status.flipPendingNum != 0) {
+		return VIDEO_OUT_ERROR_RESOURCE_BUSY;
+	}
+	slave.master = &master;
+	master.slave_count++;
+	return OK;
+}
+
 void VideoOutDriver::Impl::VblankBegin() {
 	Common::LockGuard lock(m_mutex);
 
@@ -821,9 +875,7 @@ void VideoOutDriver::Impl::PresentThread(std::stop_token token) {
 		const auto period  = std::max(frequency / refresh, uint64_t {1});
 
 		if (m_presenter.IsGuestPaused()) {
-			if (auto* frame = m_presenter.PrepareLastFrame(); frame != nullptr) {
-				m_presenter.Present(*frame, true);
-			}
+			(void)m_presenter.PresentLastFrame();
 			const auto frame_end = Common::Timer::QueryPerformanceCounter();
 			total_wait +=
 			    static_cast<int64_t>(period) - static_cast<int64_t>(frame_end - frame_begin);
@@ -833,8 +885,7 @@ void VideoOutDriver::Impl::PresentThread(std::stop_token token) {
 		VblankBegin();
 		bool presented = m_flip_queue.Flip(0);
 		if (!presented && m_presenter.NeedsSystemOverlayRefresh()) {
-			if (auto* frame = m_presenter.PrepareLastFrame(); frame != nullptr) {
-				m_presenter.Present(*frame, true);
+			if (m_presenter.PresentLastFrame()) {
 				presented = true;
 			} else {
 				uint32_t width  = 0;
@@ -877,25 +928,57 @@ bool FlipQueue::Reserve(VideoOutConfig& cfg, int index, int64_t flip_arg, FlipRe
                         uint64_t& request_id) {
 	Common::LockGuard lock(m_mutex);
 
-	if (m_requests.size() + m_cpu_requests.size() >= VIDEO_OUT_FLIP_QUEUE_CAPACITY) {
+	// A newer slave request replaces the one not yet captured by a master flip.
+	if (cfg.master != nullptr) {
+		for (auto* queue: {&m_requests, &m_cpu_requests}) {
+			for (auto it = queue->begin(); it != queue->end();) {
+				if (it->cfg != &cfg || it->group != 0) {
+					++it;
+					continue;
+				}
+				cfg.flip_status.flipPendingNum--;
+				cfg.flip_status.gcQueueNum -= it->source == FlipRequestSource::GpuEop;
+				if (it->state == RequestState::Ready) {
+					m_presenter.Discard(*it->frame);
+					it = queue->erase(it);
+				} else {
+					auto cancelled = it++;
+					m_cancelled_requests.splice(m_cancelled_requests.end(), *queue, cancelled);
+				}
+			}
+		}
+		m_done_cond_var.SignalAll();
+	}
+	if (cfg.flip_status.flipPendingNum >= VIDEO_OUT_FLIP_QUEUE_CAPACITY) {
 		return false;
 	}
 	auto& pending = source == FlipRequestSource::GpuEop ? m_requests : m_cpu_requests;
 
 	Request r {};
 	r.id         = m_next_request_id++;
+	r.group      = cfg.master == nullptr ? r.id : 0;
 	r.cfg        = &cfg;
+	r.master     = cfg.master;
 	r.generation = cfg.generation;
 	r.index      = index;
 	r.flip_arg   = flip_arg;
 	r.submit_ptc = LibKernel::KernelGetProcessTimeCounter();
 	r.source     = source;
 	r.state      = RequestState::Reserved;
+	if (cfg.slave_count != 0) {
+		for (auto* queue: {&m_requests, &m_cpu_requests}) {
+			for (auto& slave: *queue) {
+				if (slave.master == &cfg && slave.group == 0) {
+					slave.group = r.id;
+				}
+			}
+		}
+	}
 
 	pending.push_back(r);
 	request_id = r.id;
 
-	cfg.flip_status.flipPendingNum = static_cast<int>(m_requests.size() + m_cpu_requests.size());
+	cfg.flip_status.flipPendingNum++;
 	cfg.flip_status.submitProcessTimeCounter = r.submit_ptc;
 	if (source == FlipRequestSource::GpuEop) {
 		cfg.flip_status.gcQueueNum++;
@@ -917,7 +1000,9 @@ FlipQueue::~FlipQueue() {
 void FlipQueue::Cancel(VideoOutConfig& cfg) {
 	std::vector<Graphics::Presenter::Frame*> frames;
 	m_mutex.Lock();
-	while (m_processing && !m_requests.empty() && m_requests.front().cfg == &cfg) {
+	while (m_processing && std::any_of(m_requests.begin(), m_requests.end(), [&cfg](const auto& r) {
+		       return r.cfg == &cfg && r.state == RequestState::Presenting;
+	       })) {
 		m_done_cond_var.Wait(&m_mutex);
 	}
 	for (auto* queue: {&m_requests, &m_cpu_requests}) {
@@ -991,6 +1076,7 @@ void FlipQueue::Prepare(uint64_t request_id, Graphics::CommandBuffer& buffer) {
 	uint32_t            width   = 0;
 	uint32_t            height  = 0;
 	bool                current = false;
+	bool                premultiplied_alpha = false;
 	{
 		Common::LockGuard lock(cfg->mutex);
 		current = cfg->opened && !cfg->closing && cfg->generation == generation;
@@ -1012,6 +1098,9 @@ void FlipQueue::Prepare(uint64_t request_id, Graphics::CommandBuffer& buffer) {
 					     request_id, index, surface.group_index);
 				}
 				source_info = cfg->groups[surface.group_index].ImageInfo(surface);
+				premultiplied_alpha =
+				    (cfg->groups[surface.group_index].attribute.option &
+				     VIDEO_OUT_BUFFER_ATTRIBUTE_OPTION_ALPHA_PREMULTIPLIED) != 0;
 			}
 		}
 	}
@@ -1051,6 +1140,7 @@ void FlipQueue::Prepare(uint64_t request_id, Graphics::CommandBuffer& buffer) {
 		EXIT("video-out request changed while recording, id=%" PRIu64 "\n", request_id);
 	}
 	prepared->frame = frame;
+	prepared->premultiplied_alpha = premultiplied_alpha;
 }
 
 void FlipQueue::Complete(uint64_t request_id) {
@@ -1082,9 +1172,12 @@ void FlipQueue::Complete(uint64_t request_id) {
 	}
 }
 
-void FlipQueue::WaitForSubmitSlot() {
+void FlipQueue::WaitForSubmitSlot(VideoOutConfig& cfg) {
 	Common::LockGuard lock(m_mutex);
-	while (m_requests.size() + m_cpu_requests.size() >= VIDEO_OUT_FLIP_QUEUE_CAPACITY) {
+	auto belongs_to_port = [&cfg](const auto& request) { return request.cfg == &cfg; };
+	while (std::count_if(m_requests.begin(), m_requests.end(), belongs_to_port) +
+	           std::count_if(m_cpu_requests.begin(), m_cpu_requests.end(), belongs_to_port) >=
+	       VIDEO_OUT_FLIP_QUEUE_CAPACITY) {
 		if (m_requests.empty()) {
 			EXIT("video-out queue is saturated by CPU flips queued behind the current EOP\n");
 		}
@@ -1111,75 +1204,90 @@ bool FlipQueue::Flip(uint32_t micros) {
 	m_mutex.Lock();
 	if (m_requests.empty()) {
 		m_submit_cond_var.WaitFor(&m_mutex, micros);
-
-		if (m_requests.empty()) {
-			m_mutex.Unlock();
-			return false;
-		}
 	}
-	if (m_processing) {
-		EXIT("video-out flip queue processing is already active\n");
-	}
-	if (m_requests.front().state != RequestState::Ready) {
+	EXIT_IF(m_processing);
+	const auto owner = std::find_if(m_requests.begin(), m_requests.end(),
+	                               [](const auto& r) { return r.group == r.id; });
+	if (owner == m_requests.end()) {
 		m_mutex.Unlock();
 		return false;
 	}
-	m_processing = true;
-	auto r       = m_requests.front();
-	m_mutex.Unlock();
-
-	r.cfg->mutex.Lock();
-	if (!IsFlipDueLocked(*r.cfg, r.generation)) {
-		r.cfg->mutex.Unlock();
-		Common::LockGuard queue_lock(m_mutex);
-		m_processing = false;
-		m_done_cond_var.SignalAll();
+	const uint64_t group = owner->id;
+	if (std::any_of(m_cpu_requests.begin(), m_cpu_requests.end(),
+	                [group](const auto& r) { return r.group == group; }) ||
+	    std::any_of(m_requests.begin(), m_requests.end(), [group](const auto& r) {
+		    return r.group == group && r.state != RequestState::Ready;
+	    })) {
+		m_mutex.Unlock();
 		return false;
 	}
-
-	m_mutex.Lock();
-	if (m_requests.empty() || m_requests.front().id != r.id ||
-	    m_requests.front().state != RequestState::Ready || !m_processing) {
-		EXIT("video-out request changed before presentation, id=%" PRIu64 "\n", r.id);
+	std::array<Request, VideoOutDriver::Impl::VIDEO_OUT_BUSES.size()> requests {};
+	size_t count = 0;
+	for (auto& request: m_requests) {
+		if (request.group == group) {
+			EXIT_IF(count == requests.size());
+			request.state = RequestState::Presenting;
+			requests[count++] = request;
+		}
 	}
-	m_requests.front().state = RequestState::Presenting;
+	m_processing = true;
 	m_mutex.Unlock();
 
-	m_presenter.Present(*r.frame);
+	// Lock each port in bus order, then present and complete the whole group at one Vblank.
+	std::sort(requests.begin(), requests.begin() + count,
+	          [](const auto& a, const auto& b) { return a.cfg->bus < b.cfg->bus; });
+	bool due = true;
+	std::array<Graphics::Presenter::Layer, VideoOutDriver::Impl::VIDEO_OUT_BUSES.size()> layers {};
+	for (size_t i = 0; i < count; i++) {
+		auto& r = requests[i];
+		r.cfg->mutex.Lock();
+		due &= r.cfg->opened && !r.cfg->closing && r.cfg->generation == r.generation;
+		if (r.id == group) {
+			due &= IsFlipDueLocked(*r.cfg, r.generation);
+		}
+		layers[i] = {r.frame, r.cfg->bus, r.premultiplied_alpha};
+	}
+	if (due) {
+		m_presenter.Present(std::span(layers.data(), count));
+	}
 
 	m_mutex.Lock();
-	if (m_requests.empty() || m_requests.front().id != r.id ||
-	    m_requests.front().state != RequestState::Presenting) {
-		EXIT("video-out flip queue changed while processing its front request\n");
+	for (auto it = m_requests.begin(); it != m_requests.end();) {
+		if (it->group != group) {
+			++it;
+			continue;
+		}
+		if (!due) {
+			it++->state = RequestState::Ready;
+			continue;
+		}
+		auto& r = *it;
+		r.cfg->flip_status.count++;
+		r.cfg->flip_status.processTime              = LibKernel::KernelGetProcessTime();
+		r.cfg->flip_status.processTimeCounter       = LibKernel::KernelGetProcessTimeCounter();
+		r.cfg->flip_status.submitProcessTimeCounter = r.submit_ptc;
+		r.cfg->flip_status.flipArg                  = r.flip_arg;
+		r.cfg->flip_status.currentBuffer            = r.index;
+		r.cfg->flip_status.flipPendingNum--;
+		r.cfg->flip_status.gcQueueNum -= r.source == FlipRequestSource::GpuEop;
+		TriggerVideoOutEvents(*r.cfg, VideoOutEventKind::Flip, reinterpret_cast<void*>(r.flip_arg));
+		it = m_requests.erase(it);
 	}
-	m_requests.pop_front();
-
-	r.cfg->flip_status.count++;
-	r.cfg->flip_status.processTime              = LibKernel::KernelGetProcessTime();
-	r.cfg->flip_status.processTimeCounter       = LibKernel::KernelGetProcessTimeCounter();
-	r.cfg->flip_status.submitProcessTimeCounter = r.submit_ptc;
-	r.cfg->flip_status.flipArg                  = r.flip_arg;
-	r.cfg->flip_status.currentBuffer            = r.index;
-	r.cfg->flip_status.flipPendingNum = static_cast<int>(m_requests.size() + m_cpu_requests.size());
-	if (r.source == FlipRequestSource::GpuEop && r.cfg->flip_status.gcQueueNum > 0) {
-		r.cfg->flip_status.gcQueueNum--;
-	}
-	TriggerVideoOutEvents(*r.cfg, VideoOutEventKind::Flip, reinterpret_cast<void*>(r.flip_arg));
-
 	m_processing = false;
 	m_done_cond_var.SignalAll();
-	m_submit_slot_cond_var.Signal();
+	m_submit_slot_cond_var.SignalAll();
 	m_mutex.Unlock();
-	r.cfg->mutex.Unlock();
-
-	Graphics::RenderDocOnGuestFlip(m_presenter.Renderer());
-
-	if (Config::GraphicsDebugDumpEnabled() &&
-	    Config::GetPrintfDirection() != Config::LogDirection::Silent) {
-		LOGF("Flip done: %d\n", r.index);
+	for (size_t i = count; i != 0; i--) {
+		requests[i - 1].cfg->mutex.Unlock();
 	}
-
-	return true;
+	if (due) {
+		Graphics::RenderDocOnGuestFlip(m_presenter.Renderer());
+		if (Config::GraphicsDebugDumpEnabled() &&
+		    Config::GetPrintfDirection() != Config::LogDirection::Silent) {
+			LOGF("Flip done: %d\n", requests[0].index);
+		}
+	}
+	return due;
 }
 
 void FlipQueue::GetFlipStatus(VideoOutConfig& cfg, VideoOutFlipStatus& out) {
@@ -1188,6 +1296,18 @@ void FlipQueue::GetFlipStatus(VideoOutConfig& cfg, VideoOutFlipStatus& out) {
 	out = cfg.flip_status;
 }
 
+KYTY_SYSV_ABI int VideoOutSetFlipMaster(int slave_handle, int master_handle) {
+	PRINT_NAME();
+	return DriverState().SetFlipMaster(slave_handle, master_handle);
+}
+
+KYTY_SYSV_ABI void VideoOutAddBufferAttributeOption(VideoOutBufferAttribute2* attribute,
+                                                   uint64_t option) {
+	PRINT_NAME();
+	EXIT_IF(attribute == nullptr);
+	attribute->option |= option;
+}
+
 KYTY_SYSV_ABI int VideoOutOpen(int user_id, int bus_type, int index, const void* param) {
 	PRINT_NAME();
 
@@ -1467,7 +1587,7 @@ KYTY_SYSV_ABI int VideoOutSubmitFlip(int handle, int index, int flip_mode, int64
 	uint64_t  request_id = 0;
 	const int result     = ReserveFlipRequest(DriverState(), handle, index, flip_mode, flip_arg,
 	                                          FlipRequestSource::Cpu, request_id);
-	if (result == VIDEO_OUT_ERROR_INVALID_VALUE) {
+	if (result == VIDEO_OUT_ERROR_INVALID_FLIP_MODE) {
 		LOGF("\t unsupported flip_mode = %d\n", flip_mode);
 	}
 	if (result != OK) {
@@ -1504,8 +1624,10 @@ void VideoOutDriver::CompleteFlip(uint64_t request_id) {
 	m_impl->GetFlipQueue().Complete(request_id);
 }
 
-void VideoOutDriver::WaitForSubmitSlot() {
-	m_impl->GetFlipQueue().WaitForSubmitSlot();
+void VideoOutDriver::WaitForSubmitSlot(int handle) {
+	auto* cfg = m_impl->Get(handle);
+	EXIT_IF(cfg == nullptr);
+	m_impl->GetFlipQueue().WaitForSubmitSlot(*cfg);
 }
 
 void VideoOutDriver::WaitFlipDone(int handle, int index) {
@@ -1692,6 +1814,38 @@ KYTY_SYSV_ABI int VideoOutGetOutputStatus(int handle, VideoOutOutputStatus* stat
 	return OK;
 }
 
+KYTY_SYSV_ABI int VideoOutAllowOutputResolutionWqhdDetection(int handle) {
+	if (!DriverState().IsOpened(handle)) {
+		return VIDEO_OUT_ERROR_INVALID_HANDLE;
+	}
+	return OK;
+}
+
+KYTY_SYSV_ABI int VideoOutVrrPegToFixedRate(int handle, uint64_t arg1, uint64_t arg2) {
+	PRINT_NAME();
+
+	static std::atomic_bool logged {false};
+	if (!logged.exchange(true, std::memory_order_relaxed)) {
+		LOGF("\t handle = %d\n"
+		     "\t arg1   = 0x%016" PRIx64 "\n"
+		     "\t arg2   = 0x%016" PRIx64 "\n",
+		     handle, arg1, arg2);
+	}
+
+	return OK;
+}
+
+KYTY_SYSV_ABI int VideoOutVrrUnpegFromFixedRate(int handle) {
+	PRINT_NAME();
+
+	static std::atomic_bool logged {false};
+	if (!logged.exchange(true, std::memory_order_relaxed)) {
+		LOGF("\t handle = %d\n", handle);
+	}
+
+	return OK;
+}
+
 static int ValidateOutputConfig(int handle, uint64_t mode, const VideoOutOutputOptions* options,
                                 void* reserved_ptr, uint64_t reserved) {
 	if (!DriverState().IsOpened(handle)) {

src/graphics/presentation/videoOut.h

@@ -37,7 +37,7 @@ public:
 	void PrepareFlip(uint64_t request_id, Graphics::CommandBuffer& buffer);
 	void CompleteFlip(uint64_t request_id);
 	void SubmitFlipPreparation(uint64_t request_id);
-	void WaitForSubmitSlot();
+	void WaitForSubmitSlot(int handle);
 	void WaitFlipDone(int handle, int index);
 
 	[[nodiscard]] Impl& State() noexcept;
@@ -52,6 +52,9 @@ void                          VideoOutShutdown();
 
 KYTY_SYSV_ABI int  VideoOutOpen(int user_id, int bus_type, int index, const void* param);
 KYTY_SYSV_ABI int  VideoOutClose(int handle);
+KYTY_SYSV_ABI int  VideoOutSetFlipMaster(int slave_handle, int master_handle);
+KYTY_SYSV_ABI void VideoOutAddBufferAttributeOption(VideoOutBufferAttribute2* attribute,
+                                                   uint64_t option);
 KYTY_SYSV_ABI void VideoOutSetBufferAttribute2(VideoOutBufferAttribute2* attribute,
                                                uint64_t pixel_format, uint32_t tiling_mode,
                                                uint32_t width, uint32_t height, uint64_t option,
@@ -87,6 +90,9 @@ KYTY_SYSV_ABI int VideoOutGetEventData(const LibKernel::EventQueue::KernelEvent*
 KYTY_SYSV_ABI int VideoOutGetEventCount(const LibKernel::EventQueue::KernelEvent* ev);
 KYTY_SYSV_ABI int VideoOutWaitVblank(int handle);
 KYTY_SYSV_ABI int VideoOutGetOutputStatus(int handle, VideoOutOutputStatus* status);
+KYTY_SYSV_ABI int VideoOutAllowOutputResolutionWqhdDetection(int handle);
+KYTY_SYSV_ABI int VideoOutVrrPegToFixedRate(int handle, uint64_t arg1, uint64_t arg2);
+KYTY_SYSV_ABI int VideoOutVrrUnpegFromFixedRate(int handle);
 KYTY_SYSV_ABI int VideoOutInitializeOutputOptions(VideoOutOutputOptions* options);
 KYTY_SYSV_ABI int VideoOutIsOutputSupported(int handle, uint64_t mode,
                                             const VideoOutOutputOptions* options,

src/graphics/presentation/window/swapchain.cpp

@@ -4,6 +4,8 @@
 #include "common/logging/log.h"
 #include "common/profiler.h"
 #include "common/threads.h"
+#include "gpu_blit_shaders/gpu_blit_fs_triangle_spv.h"
+#include "gpu_blit_shaders/gpu_video_out_overlay_spv.h"
 #include "graphics/host_gpu/graphicContext.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
@@ -14,7 +16,7 @@
 #include "graphics/presentation/window/windowInternal.h"
 
 #include <algorithm>
-#include <deque>
+#include <array>
 #include <limits>
 #include <memory>
 #include <vector>
@@ -25,10 +27,10 @@
 namespace Libs::Graphics {
 
 struct Presenter::Frame {
-	VulkanImage image;
-	uint64_t    present_tick = 0;
-	bool        busy         = false;
-	bool        reusing_last = false;
+	VulkanImage   image;
+	vk::ImageView view         = nullptr;
+	uint64_t      present_tick = 0;
+	bool          busy         = false;
 
 	void Configure(GraphicContext& graphics, vk::Extent2D extent, vk::Format format);
 	void Transit(vk::CommandBuffer command, vk::ImageLayout layout, vk::AccessFlags2 access);
@@ -43,6 +45,7 @@ public:
 	~FramePool() {
 		m_scheduler.Wait(m_scheduler.CurrentTick() - 1);
 		for (auto& frame: m_frames) {
+			m_window.graphic_ctx.device.destroyImageView(frame->view, nullptr);
 			if (frame->image.image != nullptr) {
 				m_window.graphic_ctx.DeleteImage(frame->image);
 			}
@@ -60,6 +63,7 @@ public:
 		}
 		m_format = format;
 		m_frames.reserve(count);
+		m_free.reserve(count);
 		for (uint32_t i = 0; i < count; i++) {
 			auto frame = std::make_unique<Presenter::Frame>();
 			m_free.push_back(frame.get());
@@ -83,60 +87,45 @@ public:
 		return m_format;
 	}
 
-	Presenter::Frame* Acquire() {
+	Presenter::Frame* Acquire(vk::Extent2D extent, vk::Format format) {
 		m_mutex.Lock();
 		if (m_frames.empty()) {
 			EXIT("prepared-frame pool was used before swapchain initialization\n");
 		}
-		while (m_free.empty()) {
-			m_available.Wait(&m_mutex);
+		// A synchronized flip may need more frames than the swapchain has images.
+		// Queue capacity bounds growth; waiting here can prevent its master from recording.
+		if (m_free.empty()) {
+			auto frame = std::make_unique<Presenter::Frame>();
+			m_free.push_back(frame.get());
+			m_frames.push_back(std::move(frame));
 		}
-		auto* frame = m_free.front();
-		m_free.pop_front();
+		const auto compatible = std::ranges::find_if(m_free, [&](const auto* frame) {
+			return frame->image.extent.width == extent.width &&
+			       frame->image.extent.height == extent.height && frame->image.format == format;
+		});
+		auto*      frame      = compatible == m_free.end() ? m_free.back() : *compatible;
+		if (compatible != m_free.end()) {
+			*compatible = m_free.back();
+		}
+		m_free.pop_back();
 		if (frame->busy) {
 			EXIT("prepared-frame pool returned an invalid frame\n");
 		}
-		if (m_last_frame == frame) {
-			m_last_frame = nullptr;
-		}
-		frame->busy         = true;
-		frame->reusing_last = false;
+		frame->busy = true;
 		m_mutex.Unlock();
 
-		WaitForFrame(*frame);
+		m_scheduler.Wait(frame->present_tick);
 		return frame;
 	}
 
-	Presenter::Frame* AcquireLast() {
-		m_mutex.Lock();
-		auto* frame = m_last_frame;
-		if (frame == nullptr) {
-			m_mutex.Unlock();
-			return nullptr;
-		}
-		auto free = std::find(m_free.begin(), m_free.end(), frame);
-		if (free == m_free.end() || frame->busy) {
-			m_mutex.Unlock();
-			EXIT("last submitted frame is not available for reuse\n");
-		}
-		m_free.erase(free);
-		m_last_frame        = nullptr;
-		frame->busy         = true;
-		frame->reusing_last = true;
-		m_mutex.Unlock();
-
-		WaitForFrame(*frame);
-		return frame;
-	}
-
-	void ValidateForPresent(Presenter::Frame* frame, bool reuse) {
+	void ValidateForPresent(Presenter::Frame* frame) {
 		Common::LockGuard lock(m_mutex);
-		if (frame == nullptr || !frame->busy || frame->reusing_last != reuse) {
+		if (frame == nullptr || !frame->busy) {
 			EXIT("prepared frame has invalid presentation ownership\n");
 		}
 	}
 
-	void Release(Presenter::Frame* frame, bool make_last = false) {
+	void Release(Presenter::Frame* frame) {
 		if (frame == nullptr) {
 			EXIT("cannot release a null prepared frame\n");
 		}
@@ -144,26 +133,17 @@ public:
 		if (!frame->busy) {
 			EXIT("prepared frame was released twice\n");
 		}
-		frame->busy         = false;
-		frame->reusing_last = false;
-		if (make_last) {
-			m_last_frame = frame;
-		}
+		frame->busy = false;
 		m_free.push_back(frame);
-		m_available.Signal();
 	}
 
 private:
-	void WaitForFrame(Presenter::Frame& frame) { m_scheduler.Wait(frame.present_tick); }
-
 	WindowContext&                                 m_window;
 	CommandScheduler&                              m_scheduler;
 	Common::Mutex                                  m_mutex;
-	Common::CondVar                                m_available;
 	std::vector<std::unique_ptr<Presenter::Frame>> m_frames;
-	std::deque<Presenter::Frame*>                  m_free;
-	Presenter::Frame*                              m_last_frame = nullptr;
-	vk::Format                                     m_format     = vk::Format::eUndefined;
+	std::vector<Presenter::Frame*>                 m_free;
+	vk::Format                                     m_format = vk::Format::eUndefined;
 };
 
 void Presenter::Frame::Configure(GraphicContext& graphics, vk::Extent2D extent, vk::Format format) {
@@ -181,13 +161,16 @@ void Presenter::Frame::Configure(GraphicContext& graphics, vk::Extent2D extent,
 	const auto features = graphics.GetFormatProperties(format).optimalTilingFeatures;
 	const auto required =
 	    vk::FormatFeatureFlagBits::eBlitSrc | vk::FormatFeatureFlagBits::eSampledImageFilterLinear |
-	    vk::FormatFeatureFlagBits::eTransferSrc | vk::FormatFeatureFlagBits::eTransferDst;
+	    vk::FormatFeatureFlagBits::eSampledImage | vk::FormatFeatureFlagBits::eTransferSrc |
+	    vk::FormatFeatureFlagBits::eTransferDst;
 	if ((features & required) != required) {
 		EXIT("prepared presentation format lacks optimal blit support: format=%d features=0x%x\n",
 		     static_cast<int>(format), static_cast<vk::FormatFeatureFlags::MaskType>(features));
 	}
 
 	if (dst.image != nullptr) {
+		graphics.device.destroyImageView(view, nullptr);
+		view = nullptr;
 		graphics.DeleteImage(dst);
 	}
 
@@ -200,7 +183,8 @@ void Presenter::Frame::Configure(GraphicContext& graphics, vk::Extent2D extent,
 	create.format        = format;
 	create.tiling        = vk::ImageTiling::eOptimal;
 	create.initialLayout = vk::ImageLayout::eUndefined;
-	create.usage = vk::ImageUsageFlagBits::eTransferSrc | vk::ImageUsageFlagBits::eTransferDst;
+	create.usage = vk::ImageUsageFlagBits::eTransferSrc | vk::ImageUsageFlagBits::eTransferDst |
+	               vk::ImageUsageFlagBits::eSampled;
 	create.sharingMode = vk::SharingMode::eExclusive;
 	create.samples     = vk::SampleCountFlagBits::e1;
 	if (!graphics.CreateImage(create, dst)) {
@@ -258,7 +242,6 @@ void Presenter::Frame::CopyFrom(CommandBuffer& command_buffer, Image& source) {
 	EXIT_IF(copy.srcSubresource.layerCount != copy.dstSubresource.layerCount);
 	command.copyImage(source.backing.image, vk::ImageLayout::eTransferSrcOptimal, image.image,
 	                  vk::ImageLayout::eTransferDstOptimal, copy);
-	Transit(command, vk::ImageLayout::eTransferSrcOptimal, vk::AccessFlagBits2::eTransferRead);
 }
 
 void Presenter::Frame::Clear(CommandBuffer& command_buffer, const vk::ClearColorValue& color) {
@@ -267,7 +250,6 @@ void Presenter::Frame::Clear(CommandBuffer& command_buffer, const vk::ClearColor
 	Transit(command, vk::ImageLayout::eTransferDstOptimal, vk::AccessFlagBits2::eTransferWrite);
 	const vk::ImageSubresourceRange range {vk::ImageAspectFlagBits::eColor, 0, 1, 0, 1};
 	command.clearColorImage(image.image, vk::ImageLayout::eTransferDstOptimal, &color, 1, &range);
-	Transit(command, vk::ImageLayout::eTransferSrcOptimal, vk::AccessFlagBits2::eTransferRead);
 }
 
 class Swapchain final {
@@ -283,9 +265,9 @@ public:
 	[[nodiscard]] bool   NeedsResize() const;
 	[[nodiscard]] Status AcquireNextImage();
 	[[nodiscard]] bool   PrepareSystemOverlay();
-	void                 RecordPresentCommands(CommandBuffer& command, VulkanImage& source,
-	                                           bool draw_system_overlay);
-	uint64_t             Submit(CommandScheduler& scheduler);
+	void     RecordPresentCommands(CommandBuffer& command, Presenter::Frame* source,
+	                               const Presenter::Layer& overlay, bool draw_system_overlay);
+	uint64_t Submit(CommandScheduler& scheduler);
 	[[nodiscard]] Status Present();
 
 	[[nodiscard]] uint32_t ImageCount() const noexcept {
@@ -295,20 +277,25 @@ public:
 
 private:
 	void Destroy();
+	void DrawOverlay(vk::CommandBuffer command, const Presenter::Layer& layer);
 
-	WindowContext&              m_window;
-	vk::SwapchainKHR            m_handle = nullptr;
-	vk::Format                  m_format = vk::Format::eUndefined;
-	vk::Extent2D                m_extent {};
+	WindowContext&   m_window;
+	vk::SwapchainKHR m_handle = nullptr;
+	vk::Format       m_format = vk::Format::eUndefined;
+	vk::Extent2D     m_extent {};
 	// Drawable pixel size observed when this swapchain was created.
-	vk::Extent2D                m_window_extent {};
-	std::vector<vk::Image>      m_images;
-	std::vector<vk::ImageView>  m_image_views;
-	std::vector<vk::Semaphore>  m_image_acquired;
-	std::vector<vk::Semaphore>  m_render_complete;
+	vk::Extent2D                   m_window_extent {};
+	std::vector<vk::Image>         m_images;
+	std::vector<vk::ImageView>     m_image_views;
+	std::vector<vk::Semaphore>     m_image_acquired;
+	std::vector<vk::Semaphore>     m_render_complete;
 	std::unique_ptr<SystemOverlay> m_system_overlay;
-	uint32_t                    m_image_index = static_cast<uint32_t>(-1);
-	uint32_t                    m_frame_index = 0;
+	vk::DescriptorSetLayout        m_overlay_descriptors = nullptr;
+	vk::PipelineLayout             m_overlay_layout      = nullptr;
+	vk::Pipeline                   m_overlay_pipeline    = nullptr;
+	vk::Sampler                    m_overlay_sampler     = nullptr;
+	uint32_t                       m_image_index         = static_cast<uint32_t>(-1);
+	uint32_t                       m_frame_index         = 0;
 };
 
 struct Presenter::Impl {
@@ -347,12 +334,15 @@ struct Presenter::Impl {
 		cache.UpdateImage(image_id);
 		return image;
 	}
+	void Present();
 
 	RenderContext&        renderer;
 	WindowContext&        window;
 	Swapchain             swapchain;
 	CommandScheduler      present_scheduler;
 	FramePool             frames;
+	Common::Mutex         present_mutex;
+	std::array<Layer, 2>  layers {};
 	std::atomic<uint64_t> presented_overlay_revision {0};
 };
 
@@ -370,7 +360,7 @@ void Swapchain::Create() {
 	EXIT_IF(m_window_extent.width == 0);
 	EXIT_IF(m_window_extent.height == 0);
 	m_window.RefreshSurfaceCapabilities();
-	const auto&       surface = m_window.surface_capabilities;
+	const auto& surface = m_window.surface_capabilities;
 	EXIT_NOT_IMPLEMENTED(surface.formats.empty());
 
 	m_extent = surface.capabilities.currentExtent;
@@ -442,7 +432,7 @@ void Swapchain::Create() {
 		LOGF("warning: requested present mode is unavailable; falling back to Fifo\n");
 		create_info.presentMode = vk::PresentModeKHR::eFifo;
 	}
-	create_info.clipped          = VK_TRUE;
+	create_info.clipped = VK_TRUE;
 	RequireVulkanSuccess(graphics.device.createSwapchainKHR(&create_info, nullptr, &m_handle),
 	                     "vkCreateSwapchainKHR");
 	EXIT_IF(m_handle == nullptr);
@@ -483,8 +473,8 @@ void Swapchain::Create() {
 		    graphics.device.createSemaphore(&semaphore_info, nullptr, &m_render_complete[i]),
 		    "create swapchain render-complete semaphore");
 	}
-	m_image_index   = static_cast<uint32_t>(-1);
-	m_frame_index   = 0;
+	m_image_index = static_cast<uint32_t>(-1);
+	m_frame_index = 0;
 }
 
 Swapchain::~Swapchain() {
@@ -505,6 +495,14 @@ void Swapchain::Destroy() {
 	if (m_system_overlay != nullptr) {
 		m_system_overlay->ReleaseVulkan();
 	}
+	graphics.device.destroyPipeline(m_overlay_pipeline, nullptr);
+	graphics.device.destroyPipelineLayout(m_overlay_layout, nullptr);
+	graphics.device.destroyDescriptorSetLayout(m_overlay_descriptors, nullptr);
+	graphics.device.destroySampler(m_overlay_sampler, nullptr);
+	m_overlay_pipeline    = nullptr;
+	m_overlay_layout      = nullptr;
+	m_overlay_descriptors = nullptr;
+	m_overlay_sampler     = nullptr;
 
 	for (const auto semaphore: m_image_acquired) {
 		if (semaphore != nullptr) {
@@ -544,8 +542,8 @@ void Swapchain::Recreate(bool surface_lost) {
 		// Surface recreation goes through SDL_Vulkan_CreateSurface, which touches the
 		// window's view/layer and must run on the main thread on macOS.
 		EXIT_IF(!SDL_RunOnMainThread(
-		    [](void* window) { static_cast<WindowContext*>(window)->RecreateSurface(); },
-		    &m_window, true));
+		    [](void* window) { static_cast<WindowContext*>(window)->RecreateSurface(); }, &m_window,
+		    true));
 #else
 		m_window.RecreateSurface();
 #endif
@@ -592,14 +590,147 @@ bool Swapchain::PrepareSystemOverlay() {
 	return m_system_overlay->PrepareFrame(m_extent, m_format, ImageCount());
 }
 
-void Swapchain::RecordPresentCommands(CommandBuffer& command, VulkanImage& source,
-                                      bool draw_system_overlay) {
-	if (source.state.layout != vk::ImageLayout::eTransferSrcOptimal) {
-		EXIT("invalid prepared presentation image, vk_image=%p layout=%d\n",
-		     static_cast<void*>(source.image), static_cast<int>(source.state.layout));
+void Swapchain::DrawOverlay(vk::CommandBuffer command, const Presenter::Layer& layer) {
+	auto device = m_window.graphic_ctx.device;
+	if (m_overlay_pipeline == nullptr) {
+		vk::SamplerCreateInfo sampler {};
+		sampler.magFilter    = vk::Filter::eLinear;
+		sampler.minFilter    = vk::Filter::eLinear;
+		sampler.addressModeU = vk::SamplerAddressMode::eClampToEdge;
+		sampler.addressModeV = vk::SamplerAddressMode::eClampToEdge;
+		sampler.addressModeW = vk::SamplerAddressMode::eClampToEdge;
+		RequireVulkanSuccess(device.createSampler(&sampler, nullptr, &m_overlay_sampler),
+		                     "create video-out overlay sampler");
+		const vk::DescriptorSetLayoutBinding binding {0, vk::DescriptorType::eCombinedImageSampler,
+		                                              1, vk::ShaderStageFlagBits::eFragment};
+		vk::DescriptorSetLayoutCreateInfo    descriptors {};
+		descriptors.flags        = vk::DescriptorSetLayoutCreateFlagBits::ePushDescriptorKHR;
+		descriptors.bindingCount = 1;
+		descriptors.pBindings    = &binding;
+		RequireVulkanSuccess(
+		    device.createDescriptorSetLayout(&descriptors, nullptr, &m_overlay_descriptors),
+		    "create video-out overlay descriptor layout");
+		const vk::PushConstantRange alpha {vk::ShaderStageFlagBits::eFragment, 0, sizeof(uint32_t)};
+		vk::PipelineLayoutCreateInfo layout {};
+		layout.setLayoutCount         = 1;
+		layout.pSetLayouts            = &m_overlay_descriptors;
+		layout.pushConstantRangeCount = 1;
+		layout.pPushConstantRanges    = &alpha;
+		RequireVulkanSuccess(device.createPipelineLayout(&layout, nullptr, &m_overlay_layout),
+		                     "create video-out overlay pipeline layout");
+		const auto vertex   = CompileSPV(GPU_BLIT_FS_TRIANGLE_SPV, device);
+		const auto fragment = CompileSPV(GPU_VIDEO_OUT_OVERLAY_SPV, device);
+		std::array<vk::PipelineShaderStageCreateInfo, 2> stages {};
+		stages[0].stage  = vk::ShaderStageFlagBits::eVertex;
+		stages[0].module = vertex;
+		stages[0].pName  = "main";
+		stages[1].stage  = vk::ShaderStageFlagBits::eFragment;
+		stages[1].module = fragment;
+		stages[1].pName  = "main";
+		vk::PipelineVertexInputStateCreateInfo   vertex_input {};
+		vk::PipelineInputAssemblyStateCreateInfo assembly {};
+		assembly.topology = vk::PrimitiveTopology::eTriangleList;
+		vk::PipelineViewportStateCreateInfo viewport {};
+		viewport.viewportCount = 1;
+		viewport.scissorCount  = 1;
+		vk::PipelineRasterizationStateCreateInfo rasterization {};
+		rasterization.lineWidth = 1.0f;
+		vk::PipelineMultisampleStateCreateInfo multisample {};
+		multisample.rasterizationSamples = vk::SampleCountFlagBits::e1;
+		vk::PipelineColorBlendAttachmentState attachment {};
+		attachment.blendEnable         = VK_TRUE;
+		attachment.srcColorBlendFactor = vk::BlendFactor::eOne;
+		attachment.dstColorBlendFactor = vk::BlendFactor::eOneMinusSrcAlpha;
+		attachment.srcAlphaBlendFactor = vk::BlendFactor::eOne;
+		attachment.dstAlphaBlendFactor = vk::BlendFactor::eOneMinusSrcAlpha;
+		attachment.colorWriteMask = vk::ColorComponentFlagBits::eR |
+		                            vk::ColorComponentFlagBits::eG |
+		                            vk::ColorComponentFlagBits::eB | vk::ColorComponentFlagBits::eA;
+		vk::PipelineColorBlendStateCreateInfo blend {};
+		blend.attachmentCount = 1;
+		blend.pAttachments    = &attachment;
+		const std::array dynamic_states {vk::DynamicState::eViewport, vk::DynamicState::eScissor};
+		vk::PipelineDynamicStateCreateInfo dynamic {};
+		dynamic.dynamicStateCount = static_cast<uint32_t>(dynamic_states.size());
+		dynamic.pDynamicStates    = dynamic_states.data();
+		vk::PipelineRenderingCreateInfo rendering {};
+		rendering.colorAttachmentCount    = 1;
+		rendering.pColorAttachmentFormats = &m_format;
+		vk::GraphicsPipelineCreateInfo create {};
+		create.pNext               = &rendering;
+		create.stageCount          = static_cast<uint32_t>(stages.size());
+		create.pStages             = stages.data();
+		create.pVertexInputState   = &vertex_input;
+		create.pInputAssemblyState = &assembly;
+		create.pViewportState      = &viewport;
+		create.pRasterizationState = &rasterization;
+		create.pMultisampleState   = &multisample;
+		create.pColorBlendState    = &blend;
+		create.pDynamicState       = &dynamic;
+		create.layout              = m_overlay_layout;
+		RequireVulkanSuccess(
+		    device.createGraphicsPipelines(nullptr, 1, &create, nullptr, &m_overlay_pipeline),
+		    "create video-out overlay pipeline");
+		device.destroyShaderModule(fragment, nullptr);
+		device.destroyShaderModule(vertex, nullptr);
 	}
+	auto& frame = *layer.frame;
+	if (frame.view == nullptr) {
+		vk::ImageViewCreateInfo view {};
+		view.image            = frame.image.image;
+		view.viewType         = vk::ImageViewType::e2D;
+		view.format           = frame.image.format;
+		view.subresourceRange = {vk::ImageAspectFlagBits::eColor, 0, 1, 0, 1};
+		RequireVulkanSuccess(device.createImageView(&view, nullptr, &frame.view),
+		                     "create video-out overlay image view");
+	}
+	const vk::DescriptorImageInfo image {m_overlay_sampler, frame.view,
+	                                     vk::ImageLayout::eShaderReadOnlyOptimal};
+	vk::WriteDescriptorSet        write {};
+	write.dstBinding      = 0;
+	write.descriptorCount = 1;
+	write.descriptorType  = vk::DescriptorType::eCombinedImageSampler;
+	write.pImageInfo      = &image;
+	command.pushDescriptorSetKHR(vk::PipelineBindPoint::eGraphics, m_overlay_layout, 0, 1, &write);
+	const uint32_t premultiplied = layer.premultiplied_alpha;
+	command.pushConstants(m_overlay_layout, vk::ShaderStageFlagBits::eFragment, 0,
+	                      sizeof(premultiplied), &premultiplied);
+	command.bindPipeline(vk::PipelineBindPoint::eGraphics, m_overlay_pipeline);
+	vk::RenderingAttachmentInfo attachment {};
+	attachment.imageView   = m_image_views[m_image_index];
+	attachment.imageLayout = vk::ImageLayout::eColorAttachmentOptimal;
+	attachment.loadOp      = vk::AttachmentLoadOp::eLoad;
+	attachment.storeOp     = vk::AttachmentStoreOp::eStore;
+	vk::RenderingInfo rendering {};
+	rendering.renderArea.extent    = m_extent;
+	rendering.layerCount           = 1;
+	rendering.colorAttachmentCount = 1;
+	rendering.pColorAttachments    = &attachment;
+	command.beginRendering(&rendering);
+	const vk::Viewport viewport {
+	    0.0f, 0.0f, static_cast<float>(m_extent.width), static_cast<float>(m_extent.height),
+	    0.0f, 1.0f};
+	const vk::Rect2D scissor {{0, 0}, m_extent};
+	command.setViewport(0, 1, &viewport);
+	command.setScissor(0, 1, &scissor);
+	command.draw(3, 1, 0, 0);
+	command.endRendering();
+}
+
+void Swapchain::RecordPresentCommands(CommandBuffer& command, Presenter::Frame* source,
+                                      const Presenter::Layer& overlay, bool draw_system_overlay) {
 	EXIT_IF(m_image_index >= m_images.size());
-	auto vk_command = command.Handle();
+	auto       vk_command      = command.Handle();
+	const bool draw_overlay    = overlay.frame != nullptr;
+	const bool draw_attachment = draw_overlay || draw_system_overlay;
+	if (source != nullptr) {
+		source->Transit(vk_command, vk::ImageLayout::eTransferSrcOptimal,
+		                vk::AccessFlagBits2::eTransferRead);
+	}
+	if (draw_overlay) {
+		overlay.frame->Transit(vk_command, vk::ImageLayout::eShaderReadOnlyOptimal,
+		                       vk::AccessFlagBits2::eShaderRead);
+	}
 
 	vk::ImageMemoryBarrier to_transfer {};
 	to_transfer.sType                           = vk::StructureType::eImageMemoryBarrier;
@@ -619,49 +750,68 @@ void Swapchain::RecordPresentCommands(CommandBuffer& command, VulkanImage& sourc
 	                           vk::PipelineStageFlagBits::eTransfer, vk::DependencyFlags {}, 0,
 	                           nullptr, 0, nullptr, 1, &to_transfer);
 
-	vk::ImageBlit region {};
-	region.srcSubresource.aspectMask     = vk::ImageAspectFlagBits::eColor;
-	region.srcSubresource.mipLevel       = 0;
-	region.srcSubresource.baseArrayLayer = 0;
-	region.srcSubresource.layerCount     = 1;
-	region.srcOffsets[1].x               = static_cast<int>(source.extent.width);
-	region.srcOffsets[1].y               = static_cast<int>(source.extent.height);
-	region.srcOffsets[1].z               = 1;
-	region.dstSubresource.aspectMask     = vk::ImageAspectFlagBits::eColor;
-	region.dstSubresource.mipLevel       = 0;
-	region.dstSubresource.baseArrayLayer = 0;
-	region.dstSubresource.layerCount     = 1;
-	region.dstOffsets[1].x               = static_cast<int>(m_extent.width);
-	region.dstOffsets[1].y               = static_cast<int>(m_extent.height);
-	region.dstOffsets[1].z               = 1;
-	vk_command.blitImage(source.image, vk::ImageLayout::eTransferSrcOptimal,
-	                     m_images[m_image_index], vk::ImageLayout::eTransferDstOptimal, 1, &region,
-	                     vk::Filter::eLinear);
+	if (source != nullptr) {
+		vk::ImageBlit region {};
+		region.srcSubresource.aspectMask     = vk::ImageAspectFlagBits::eColor;
+		region.srcSubresource.mipLevel       = 0;
+		region.srcSubresource.baseArrayLayer = 0;
+		region.srcSubresource.layerCount     = 1;
+		region.srcOffsets[1].x               = static_cast<int>(source->image.extent.width);
+		region.srcOffsets[1].y               = static_cast<int>(source->image.extent.height);
+		region.srcOffsets[1].z               = 1;
+		region.dstSubresource.aspectMask     = vk::ImageAspectFlagBits::eColor;
+		region.dstSubresource.mipLevel       = 0;
+		region.dstSubresource.baseArrayLayer = 0;
+		region.dstSubresource.layerCount     = 1;
+		region.dstOffsets[1].x               = static_cast<int>(m_extent.width);
+		region.dstOffsets[1].y               = static_cast<int>(m_extent.height);
+		region.dstOffsets[1].z               = 1;
+		vk_command.blitImage(source->image.image, vk::ImageLayout::eTransferSrcOptimal,
+		                     m_images[m_image_index], vk::ImageLayout::eTransferDstOptimal, 1,
+		                     &region, vk::Filter::eLinear);
+	} else {
+		const vk::ClearColorValue black {};
+		vk_command.clearColorImage(m_images[m_image_index], vk::ImageLayout::eTransferDstOptimal,
+		                           &black, 1, &to_transfer.subresourceRange);
+	}
 
 	vk::ImageMemoryBarrier to_present {};
-	to_present.sType         = vk::StructureType::eImageMemoryBarrier;
-	to_present.srcAccessMask = vk::AccessFlagBits::eTransferWrite;
-	to_present.dstAccessMask = draw_system_overlay ? vk::AccessFlagBits::eColorAttachmentRead |
-	                                                     vk::AccessFlagBits::eColorAttachmentWrite
-	                                               : vk::AccessFlagBits::eMemoryRead;
-	to_present.oldLayout     = vk::ImageLayout::eTransferDstOptimal;
-	to_present.newLayout     = draw_system_overlay ? vk::ImageLayout::eColorAttachmentOptimal
-	                                               : vk::ImageLayout::ePresentSrcKHR;
-	to_present.srcQueueFamilyIndex             = VK_QUEUE_FAMILY_IGNORED;
-	to_present.dstQueueFamilyIndex             = VK_QUEUE_FAMILY_IGNORED;
-	to_present.image                           = m_images[m_image_index];
+	to_present.sType               = vk::StructureType::eImageMemoryBarrier;
+	to_present.srcAccessMask       = vk::AccessFlagBits::eTransferWrite;
+	to_present.dstAccessMask       = draw_attachment ? vk::AccessFlagBits::eColorAttachmentRead |
+	                                                       vk::AccessFlagBits::eColorAttachmentWrite
+	                                                 : vk::AccessFlagBits::eMemoryRead;
+	to_present.oldLayout           = vk::ImageLayout::eTransferDstOptimal;
+	to_present.newLayout           = draw_attachment ? vk::ImageLayout::eColorAttachmentOptimal
+	                                                 : vk::ImageLayout::ePresentSrcKHR;
+	to_present.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
+	to_present.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
+	to_present.image               = m_images[m_image_index];
 	to_present.subresourceRange.aspectMask     = vk::ImageAspectFlagBits::eColor;
 	to_present.subresourceRange.baseMipLevel   = 0;
 	to_present.subresourceRange.levelCount     = 1;
 	to_present.subresourceRange.baseArrayLayer = 0;
 	to_present.subresourceRange.layerCount     = 1;
-	vk_command.pipelineBarrier(
-	    vk::PipelineStageFlagBits::eTransfer,
-	    draw_system_overlay ? vk::PipelineStageFlagBits::eColorAttachmentOutput
-	                        : vk::PipelineStageFlagBits::eAllCommands,
-	    vk::DependencyFlagBits::eByRegion, 0, nullptr, 0, nullptr, 1, &to_present);
-	if (draw_system_overlay) {
-		m_system_overlay->Record(vk_command, m_image_views[m_image_index]);
+	vk_command.pipelineBarrier(vk::PipelineStageFlagBits::eTransfer,
+	                           draw_attachment ? vk::PipelineStageFlagBits::eColorAttachmentOutput
+	                                           : vk::PipelineStageFlagBits::eAllCommands,
+	                           vk::DependencyFlagBits::eByRegion, 0, nullptr, 0, nullptr, 1,
+	                           &to_present);
+	if (draw_attachment) {
+		if (draw_overlay) {
+			DrawOverlay(vk_command, overlay);
+			if (draw_system_overlay) {
+				to_present.srcAccessMask = vk::AccessFlagBits::eColorAttachmentWrite;
+				to_present.oldLayout     = vk::ImageLayout::eColorAttachmentOptimal;
+				vk_command.pipelineBarrier(vk::PipelineStageFlagBits::eColorAttachmentOutput,
+				                           vk::PipelineStageFlagBits::eColorAttachmentOutput,
+				                           vk::DependencyFlagBits::eByRegion, 0, nullptr, 0,
+				                           nullptr, 1, &to_present);
+			}
+		}
+		if (draw_system_overlay) {
+			m_system_overlay->Record(vk_command, m_image_views[m_image_index]);
+		}
 		to_present.srcAccessMask = vk::AccessFlagBits::eColorAttachmentWrite;
 		to_present.dstAccessMask = vk::AccessFlagBits::eMemoryRead;
 		to_present.oldLayout     = vk::ImageLayout::eColorAttachmentOptimal;
@@ -721,19 +871,18 @@ Presenter::~Presenter() = default;
 Presenter::Frame& Presenter::PrepareFrame(CommandBuffer& buffer, const ImageInfo& info) {
 	KYTY_PROFILER_FUNCTION();
 	EXIT_IF(buffer.IsInvalid());
-	auto*             frame = m_impl->frames.Acquire();
-	Common::LockGuard render_lock(m_impl->renderer.GetMutex());
-	auto&             image = m_impl->ResolveSurface(info);
-	if (image.backing.format == vk::Format::eUndefined) {
-		EXIT("unsupported presentation source, image=%p\n", static_cast<const void*>(&image));
-	}
-
 	auto frame_format = info.pixel_format;
 	switch (frame_format) {
 		case vk::Format::eR8G8B8A8Srgb: frame_format = vk::Format::eR8G8B8A8Unorm; break;
 		case vk::Format::eB8G8R8A8Srgb: frame_format = vk::Format::eB8G8R8A8Unorm; break;
 		default: break;
 	}
+	auto* frame = m_impl->frames.Acquire({info.extent.width, info.extent.height}, frame_format);
+	Common::LockGuard render_lock(m_impl->renderer.GetMutex());
+	auto&             image = m_impl->ResolveSurface(info);
+	if (image.backing.format == vk::Format::eUndefined) {
+		EXIT("unsupported presentation source, image=%p\n", static_cast<const void*>(&image));
+	}
 	frame->Configure(m_impl->window.graphic_ctx,
 	                 {image.backing.extent.width, image.backing.extent.height}, frame_format);
 	frame->CopyFrom(buffer, image);
@@ -744,7 +893,7 @@ Presenter::Frame& Presenter::PrepareBlankFrame(uint32_t width, uint32_t height,
                                                CommandBuffer* producer) {
 	KYTY_PROFILER_FUNCTION();
 	auto              format = m_impl->frames.GetFormat();
-	auto*             frame  = m_impl->frames.Acquire();
+	auto*             frame  = m_impl->frames.Acquire({width, height}, format);
 	Common::LockGuard render_lock(m_impl->renderer.GetMutex());
 	frame->Configure(m_impl->window.graphic_ctx, {width, height}, format);
 	vk::ClearColorValue clear {};
@@ -760,8 +909,13 @@ Presenter::Frame& Presenter::PrepareBlankFrame(uint32_t width, uint32_t height,
 	return *frame;
 }
 
-Presenter::Frame* Presenter::PrepareLastFrame() {
-	return m_impl->frames.AcquireLast();
+bool Presenter::PresentLastFrame() {
+	Common::LockGuard lock(m_impl->present_mutex);
+	if (m_impl->layers[0].frame == nullptr && m_impl->layers[1].frame == nullptr) {
+		return false;
+	}
+	m_impl->Present();
+	return true;
 }
 
 bool Presenter::IsGuestPaused() const noexcept {
@@ -778,44 +932,76 @@ RenderContext& Presenter::Renderer() const noexcept {
 	return m_impl->renderer;
 }
 
-void Presenter::Present(Frame& frame, bool reuse) {
+void Presenter::Present(Frame& frame) {
+	const Layer layer {&frame, 0, false};
+	Present(std::span(&layer, 1));
+}
+
+void Presenter::Present(std::span<const Layer> layers) {
+	Common::LockGuard lock(m_impl->present_mutex);
+	for (const auto& layer: layers) {
+		EXIT_IF(layer.bus < 0 || layer.bus >= static_cast<int>(m_impl->layers.size()));
+		m_impl->frames.ValidateForPresent(layer.frame);
+		auto& previous = m_impl->layers[layer.bus];
+		if (previous.frame != nullptr) {
+			m_impl->frames.Release(previous.frame);
+		}
+		previous = layer;
+	}
+	m_impl->Present();
+}
+
+void Presenter::ClearLayer(int bus) {
+	if (static_cast<size_t>(bus) >= m_impl->layers.size()) {
+		return;
+	}
+	Common::LockGuard lock(m_impl->present_mutex);
+	auto& layer = m_impl->layers[bus];
+	if (layer.frame != nullptr) {
+		m_impl->frames.Release(layer.frame);
+		layer = {};
+	}
+}
+
+void Presenter::Impl::Present() {
 	KYTY_PROFILER_FUNCTION();
-	m_impl->frames.ValidateForPresent(&frame, reuse);
 
 	const auto overlay_visual = GetSystemOverlayVisualState();
-	auto&      swapchain  = m_impl->swapchain;
 	// Some window systems keep presenting an old swapchain after a resize.
 	if (swapchain.NeedsResize()) {
-		m_impl->RecoverSwapchain(Swapchain::Status::Recreate);
+		RecoverSwapchain(Swapchain::Status::Recreate);
 	}
 	for (uint32_t attempt = 0; attempt < 2; attempt++) {
 		auto status = swapchain.AcquireNextImage();
 		if (status != Swapchain::Status::Success) {
-			m_impl->RecoverSwapchain(status);
+			RecoverSwapchain(status);
 			continue;
 		}
 		{
-			Common::LockGuard render_lock(m_impl->renderer.GetMutex());
-			auto&             command          = m_impl->present_scheduler.BeginCommand();
+			Common::LockGuard render_lock(renderer.GetMutex());
+			auto&             command = present_scheduler.BeginCommand();
 			const bool        draw_system_overlay =
 			    overlay_visual.active && swapchain.PrepareSystemOverlay();
-			swapchain.RecordPresentCommands(command, frame.image, draw_system_overlay);
-			frame.present_tick = swapchain.Submit(m_impl->present_scheduler);
+			swapchain.RecordPresentCommands(command, layers[0].frame, layers[1],
+			                                draw_system_overlay);
+			const auto tick = swapchain.Submit(present_scheduler);
+			for (const auto& layer: layers) {
+				if (layer.frame != nullptr) {
+					layer.frame->present_tick = tick;
+				}
+			}
 		}
 		status = swapchain.Present();
 		if (status != Swapchain::Status::Success) {
-			m_impl->RecoverSwapchain(status);
+			RecoverSwapchain(status);
 			continue;
 		}
 
-		m_impl->presented_overlay_revision.store(overlay_visual.revision,
-		                                         std::memory_order_release);
-		m_impl->window.UpdateTitle();
-		m_impl->frames.Release(&frame, true);
+		presented_overlay_revision.store(overlay_visual.revision, std::memory_order_release);
+		window.UpdateTitle();
 		return;
 	}
 	LOGF("Vulkan presentation retry exhausted; dropping frame\n");
-	m_impl->frames.Release(&frame, reuse);
 }
 
 void Presenter::Discard(Frame& frame) {

src/graphics/presentation/window/vulkanWindow.cpp

@@ -137,11 +137,13 @@ static uint32_t VulkanFindQueueFamily(vk::PhysicalDevice device, vk::SurfaceKHR
 	return static_cast<uint32_t>(-1);
 }
 
+// On failure out_device is null and out_rejections says why each device was skipped.
 // NOLINTNEXTLINE(readability-function-cognitive-complexity)
 static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surface,
                                      const std::vector<const char*>& device_extensions,
                                      SurfaceCapabilities&            out_capabilities,
-                                     vk::PhysicalDevice& out_device, uint32_t& out_queue_family) {
+                                     vk::PhysicalDevice& out_device, uint32_t& out_queue_family,
+                                     std::string& out_rejections) {
 	EXIT_IF(instance == nullptr);
 	EXIT_IF(surface == nullptr);
 
@@ -170,14 +172,22 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 		vk::PhysicalDeviceProperties device_properties {};
 		device.getProperties(&device_properties);
 
+		const auto reject = [&](const std::string& reason) {
+			LOGF("%s\n", reason.c_str());
+			out_rejections +=
+			    skip_device ? "; " : fmt::format("\n  {}: ", device_properties.deviceName.data());
+			out_rejections += reason;
+			skip_device = true;
+		};
+
 		LOGF("Vulkan device: %s\n", device_properties.deviceName.data());
 		if (device_properties.apiVersion < VULKAN_TARGET_API_VERSION) {
-			LOGF("Vulkan %u.%u is required, but device supports only %u.%u.%u\n",
-			     VK_VERSION_MAJOR(VULKAN_TARGET_API_VERSION),
-			     VK_VERSION_MINOR(VULKAN_TARGET_API_VERSION),
-			     VK_VERSION_MAJOR(device_properties.apiVersion),
-			     VK_VERSION_MINOR(device_properties.apiVersion),
-			     VK_VERSION_PATCH(device_properties.apiVersion));
+			reject(fmt::format("Vulkan {}.{} is required, but device supports only {}.{}.{}",
+			                   VK_VERSION_MAJOR(VULKAN_TARGET_API_VERSION),
+			                   VK_VERSION_MINOR(VULKAN_TARGET_API_VERSION),
+			                   VK_VERSION_MAJOR(device_properties.apiVersion),
+			                   VK_VERSION_MINOR(device_properties.apiVersion),
+			                   VK_VERSION_PATCH(device_properties.apiVersion)));
 			continue;
 		}
 
@@ -203,7 +213,7 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 		fragment_barycentric.pNext = &depth_clip_control;
 		features12.pNext           = &fragment_barycentric;
 #endif
-		features13.pNext = &features12;
+		features13.pNext       = &features12;
 		device_features2.pNext = &features13;
 
 		device.getFeatures2(&device_features2);
@@ -212,151 +222,82 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 
 		const auto queue_family = VulkanFindQueueFamily(device, surface);
 		if (queue_family == static_cast<uint32_t>(-1)) {
-			LOGF("No universal graphics, compute, and presentation queue\n");
-			skip_device = true;
+			reject("No universal graphics, compute, and presentation queue");
 		}
 
+		const auto check_feature = [&](vk::Bool32 supported, const char* name,
+		                               vk::Bool32 required = VK_TRUE) {
+			if (required == VK_TRUE && supported != VK_TRUE) {
+				reject(fmt::format("{} is not supported", name));
+			}
+		};
+
+#if defined(__APPLE__)
 		if (color_write_ext.colorWriteEnable != VK_TRUE) {
 			LOGF("colorWriteEnable is not supported\n");
-#if !defined(__APPLE__)
-			skip_device = true;
+		}
+#else
+		check_feature(color_write_ext.colorWriteEnable, "colorWriteEnable");
 #endif
-		}
-		if (image_view_min_lod.minLod != VK_TRUE) {
-			LOGF("image view minLod is not supported\n");
-			skip_device = true;
-		}
+		check_feature(image_view_min_lod.minLod, "image view minLod");
 
-		if (depth_clip_control.depthClipControl != VK_TRUE) {
-			LOGF("depthClipControl is not supported\n");
-			skip_device = true;
-		}
+		check_feature(depth_clip_control.depthClipControl, "depthClipControl");
+#if defined(__APPLE__)
 		if (depth_clip_enable.depthClipEnable != VK_TRUE) {
 			LOGF("depthClipEnable is not supported\n");
-#if !defined(__APPLE__)
-			skip_device = true;
+		}
+#else
+		check_feature(depth_clip_enable.depthClipEnable, "depthClipEnable");
 #endif
-		}
 #if !defined(__APPLE__)
-		if (device_features2.features.depthClamp != VK_TRUE) {
-			LOGF("depthClamp is not supported\n");
-			skip_device = true;
-		}
-		if (fragment_barycentric.fragmentShaderBarycentric != VK_TRUE) {
-			LOGF("fragmentShaderBarycentric is not supported\n");
-			skip_device = true;
-		}
+		check_feature(device_features2.features.depthClamp, "depthClamp");
+		check_feature(fragment_barycentric.fragmentShaderBarycentric, "fragmentShaderBarycentric");
 #endif
 
-		if (required_features12.samplerMirrorClampToEdge == VK_TRUE &&
-		    features12.samplerMirrorClampToEdge != VK_TRUE) {
-			LOGF("samplerMirrorClampToEdge is not supported\n");
-			skip_device = true;
-		}
-		if (required_features12.timelineSemaphore == VK_TRUE &&
-		    features12.timelineSemaphore != VK_TRUE) {
-			LOGF("timelineSemaphore is not supported\n");
-			skip_device = true;
-		}
-		if (required_features12.shaderOutputLayer == VK_TRUE &&
-		    features12.shaderOutputLayer != VK_TRUE) {
-			LOGF("shaderOutputLayer is not supported\n");
-			skip_device = true;
-		}
-		if (required_features12.shaderOutputViewportIndex == VK_TRUE &&
-		    features12.shaderOutputViewportIndex != VK_TRUE) {
-			LOGF("shaderOutputViewportIndex is not supported\n");
-			skip_device = true;
-		}
-		if (required_features12.bufferDeviceAddress == VK_TRUE &&
-		    features12.bufferDeviceAddress != VK_TRUE) {
-			LOGF("bufferDeviceAddress is not supported\n");
-			skip_device = true;
-		}
-		if (required_features12.shaderBufferInt64Atomics == VK_TRUE &&
-		    features12.shaderBufferInt64Atomics != VK_TRUE) {
-			LOGF("shaderBufferInt64Atomics is not supported\n");
-			skip_device = true;
-		}
-		if (features13.robustImageAccess != VK_TRUE) {
-			LOGF("robustImageAccess is not supported\n");
-			skip_device = true;
-		}
-		if (required_features13.dynamicRendering == VK_TRUE &&
-		    features13.dynamicRendering != VK_TRUE) {
-			LOGF("dynamicRendering is not supported\n");
-			skip_device = true;
-		}
-		if (required_features13.synchronization2 == VK_TRUE &&
-		    features13.synchronization2 != VK_TRUE) {
-			LOGF("synchronization2 is not supported\n");
-			skip_device = true;
-		}
-		if (device_features2.features.sampleRateShading != VK_TRUE) {
-			LOGF("sampleRateShading is not supported\n");
-			skip_device = true;
-		}
-		if (device_features2.features.depthBiasClamp != VK_TRUE) {
-			LOGF("depthBiasClamp is not supported\n");
-			skip_device = true;
-		}
-		if (device_features2.features.shaderClipDistance != VK_TRUE) {
-			LOGF("shaderClipDistance is not supported\n");
-			skip_device = true;
-		}
-		if (device_features2.features.shaderCullDistance != VK_TRUE) {
-			LOGF("shaderCullDistance is not supported\n");
-			skip_device = true;
-		}
-		if (device_features2.features.largePoints != VK_TRUE) {
-			LOGF("largePoints is not supported\n");
-			skip_device = true;
-		}
-		if (device_features2.features.multiViewport != VK_TRUE) {
-			LOGF("multiViewport is not supported\n");
-			skip_device = true;
-		}
-		if (device_features2.features.fillModeNonSolid != VK_TRUE) {
-			LOGF("fillModeNonSolid is not supported\n");
-			skip_device = true;
-		}
-		if (device_features2.features.fragmentStoresAndAtomics != VK_TRUE) {
-			LOGF("fragmentStoresAndAtomics is not supported\n");
-			skip_device = true;
-		}
+		check_feature(features12.samplerMirrorClampToEdge, "samplerMirrorClampToEdge",
+		              required_features12.samplerMirrorClampToEdge);
+		check_feature(features12.timelineSemaphore, "timelineSemaphore",
+		              required_features12.timelineSemaphore);
+		check_feature(features12.shaderOutputLayer, "shaderOutputLayer",
+		              required_features12.shaderOutputLayer);
+		check_feature(features12.shaderOutputViewportIndex, "shaderOutputViewportIndex",
+		              required_features12.shaderOutputViewportIndex);
+		check_feature(features12.bufferDeviceAddress, "bufferDeviceAddress",
+		              required_features12.bufferDeviceAddress);
+		check_feature(features12.shaderBufferInt64Atomics, "shaderBufferInt64Atomics",
+		              required_features12.shaderBufferInt64Atomics);
+		check_feature(features13.robustImageAccess, "robustImageAccess");
+		check_feature(features13.dynamicRendering, "dynamicRendering",
+		              required_features13.dynamicRendering);
+		check_feature(features13.synchronization2, "synchronization2",
+		              required_features13.synchronization2);
+		check_feature(device_features2.features.sampleRateShading, "sampleRateShading");
+		check_feature(device_features2.features.depthBiasClamp, "depthBiasClamp");
+		check_feature(device_features2.features.shaderClipDistance, "shaderClipDistance");
+		check_feature(device_features2.features.shaderCullDistance, "shaderCullDistance");
+		check_feature(device_features2.features.largePoints, "largePoints");
+		check_feature(device_features2.features.multiViewport, "multiViewport");
+		check_feature(device_features2.features.fillModeNonSolid, "fillModeNonSolid");
+		check_feature(device_features2.features.fragmentStoresAndAtomics,
+		              "fragmentStoresAndAtomics");
 
-		if (device_features2.features.samplerAnisotropy != VK_TRUE) {
-			LOGF("samplerAnisotropy is not supported\n");
-			skip_device = true;
-		}
-		if (device_features2.features.robustBufferAccess != VK_TRUE) {
-			LOGF("robustBufferAccess is not supported\n");
-			skip_device = true;
-		}
+		check_feature(device_features2.features.samplerAnisotropy, "samplerAnisotropy");
+		check_feature(device_features2.features.robustBufferAccess, "robustBufferAccess");
+#if defined(__APPLE__)
 		if (device_features2.features.depthBounds != VK_TRUE) {
 			LOGF("depthBounds is not supported\n");
-#if !defined(__APPLE__)
-			skip_device = true;
+		}
+#else
+		check_feature(device_features2.features.depthBounds, "depthBounds");
 #endif
-		}
-		if (device_features2.features.shaderStorageImageWriteWithoutFormat != VK_TRUE) {
-			LOGF("shaderStorageImageWriteWithoutFormat is not supported\n");
-			skip_device = true;
-		}
+		check_feature(device_features2.features.shaderStorageImageWriteWithoutFormat,
+		              "shaderStorageImageWriteWithoutFormat");
 
-		if (device_features2.features.shaderImageGatherExtended != VK_TRUE) {
-			LOGF("shaderImageGatherExtended is not supported\n");
-			skip_device = true;
-		}
+		check_feature(device_features2.features.shaderImageGatherExtended,
+		              "shaderImageGatherExtended");
 
-		if (device_features2.features.independentBlend != VK_TRUE) {
-			LOGF("independentBlend is not supported\n");
-			skip_device = true;
-		}
-		if (device_features2.features.tessellationShader != VK_TRUE) {
-			LOGF("tessellationShader is not supported\n");
-			skip_device = true;
-		}
+		check_feature(device_features2.features.independentBlend, "independentBlend");
+		check_feature(device_features2.features.tessellationShader, "tessellationShader");
 
 		if (!skip_device) {
 			auto available_extensions = EnumerateVulkan<vk::ExtensionProperties>(
@@ -368,8 +309,7 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 
 			for (const char* ext: device_extensions) {
 				if (!HasExtension(available_extensions, ext)) {
-					skip_device = true;
-					break;
+					reject(fmt::format("{} is not supported", ext));
 				}
 			}
 
@@ -387,86 +327,46 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 
 			if (!(candidate_capabilities.capabilities.supportedUsageFlags &
 			      vk::ImageUsageFlagBits::eTransferDst)) {
-				LOGF("Surface cannot be destination of blit\n");
-				skip_device = true;
+				reject("Surface cannot be destination of blit");
 			}
 		}
 
-		if (!skip_device && !CheckFormat(device, vk::Format::eD32Sfloat, true,
-		                                 vk::FormatFeatureFlagBits::eDepthStencilAttachment)) {
-			LOGF("Format vk::Format::eD32Sfloat cannot be used as depth buffer\n");
-			skip_device = true;
-		}
-
-		if (!skip_device && !CheckFormat(device, vk::Format::eD32SfloatS8Uint, true,
-		                                 vk::FormatFeatureFlagBits::eDepthStencilAttachment)) {
-			LOGF("Format vk::Format::eD32SfloatS8Uint cannot be used as depth buffer\n");
-			skip_device = true;
-		}
-
-		if (!skip_device && !CheckFormat(device, vk::Format::eD16Unorm, true,
-		                                 vk::FormatFeatureFlagBits::eDepthStencilAttachment)) {
-			LOGF("Format vk::Format::eD16Unorm cannot be used as depth buffer\n");
-			skip_device = true;
-		}
-
-		if (!skip_device && !CheckFormat(device, vk::Format::eBc3SrgbBlock, true,
-		                                 vk::FormatFeatureFlagBits::eSampledImage |
-		                                     vk::FormatFeatureFlagBits::eTransferDst)) {
-			LOGF("Format vk::Format::eBc3SrgbBlock cannot be used as texture\n");
-			skip_device = true;
-		}
-
-		if (!skip_device && !CheckFormat(device, vk::Format::eR8G8B8A8Srgb, true,
-		                                 vk::FormatFeatureFlagBits::eSampledImage |
-		                                     vk::FormatFeatureFlagBits::eTransferDst)) {
-			LOGF("Format vk::Format::eR8G8B8A8Srgb cannot be used as texture\n");
-			skip_device = true;
-		}
-
-		if (!skip_device && !CheckFormat(device, vk::Format::eR8Unorm, true,
-		                                 vk::FormatFeatureFlagBits::eSampledImage |
-		                                     vk::FormatFeatureFlagBits::eTransferDst)) {
-			LOGF("Format vk::Format::eR8Unorm cannot be used as texture\n");
-			skip_device = true;
-		}
-
-		if (!skip_device && !CheckFormat(device, vk::Format::eR8G8Unorm, true,
-		                                 vk::FormatFeatureFlagBits::eSampledImage |
-		                                     vk::FormatFeatureFlagBits::eTransferDst)) {
-			LOGF("Format vk::Format::eR8G8Unorm cannot be used as texture\n");
-			skip_device = true;
-		}
-
-		if (!skip_device && !CheckFormat(device, vk::Format::eR8G8B8A8Srgb, true,
-		                                 vk::FormatFeatureFlagBits::eStorageImage |
-		                                     vk::FormatFeatureFlagBits::eTransferDst)) {
-			LOGF("Format vk::Format::eR8G8B8A8Srgb cannot be used as texture\n");
-
-			if (!skip_device && !CheckFormat(device, vk::Format::eR8G8B8A8Unorm, true,
-			                                 vk::FormatFeatureFlagBits::eStorageImage |
-			                                     vk::FormatFeatureFlagBits::eTransferDst)) {
-				LOGF("Format vk::Format::eR8G8B8A8Unorm cannot be used as texture\n");
-				skip_device = true;
+		const auto check_format = [&](vk::Format format, vk::FormatFeatureFlags features,
+		                              const char* usage) {
+			if (!skip_device && !CheckFormat(device, format, true, features)) {
+				reject(fmt::format("Format vk::Format::e{} cannot be used as {}",
+				                   vk::to_string(format), usage));
 			}
+		};
+
+		const auto depth_features = vk::FormatFeatureFlagBits::eDepthStencilAttachment;
+		const auto texture_features =
+		    vk::FormatFeatureFlagBits::eSampledImage | vk::FormatFeatureFlagBits::eTransferDst;
+		const auto storage_features =
+		    vk::FormatFeatureFlagBits::eStorageImage | vk::FormatFeatureFlagBits::eTransferDst;
+
+		check_format(vk::Format::eD32Sfloat, depth_features, "depth buffer");
+		check_format(vk::Format::eD32SfloatS8Uint, depth_features, "depth buffer");
+		check_format(vk::Format::eD16Unorm, depth_features, "depth buffer");
+		check_format(vk::Format::eBc3SrgbBlock, texture_features, "texture");
+		check_format(vk::Format::eR8G8B8A8Srgb, texture_features, "texture");
+		check_format(vk::Format::eR8Unorm, texture_features, "texture");
+		check_format(vk::Format::eR8G8Unorm, texture_features, "texture");
+
+		if (!skip_device &&
+		    !CheckFormat(device, vk::Format::eR8G8B8A8Srgb, true, storage_features)) {
+			LOGF("Format vk::Format::eR8G8B8A8Srgb cannot be used as storage image\n");
+			check_format(vk::Format::eR8G8B8A8Unorm, storage_features, "storage image");
 		}
 
-		if (!skip_device && !CheckFormat(device, vk::Format::eB8G8R8A8Srgb, true,
-		                                 vk::FormatFeatureFlagBits::eStorageImage |
-		                                     vk::FormatFeatureFlagBits::eTransferDst)) {
-			LOGF("Format vk::Format::eB8G8R8A8Srgb cannot be used as texture\n");
-
-			if (!skip_device && !CheckFormat(device, vk::Format::eB8G8R8A8Unorm, true,
-			                                 vk::FormatFeatureFlagBits::eStorageImage |
-			                                     vk::FormatFeatureFlagBits::eTransferDst)) {
-				LOGF("Format vk::Format::eB8G8R8A8Unorm cannot be used as texture\n");
-				skip_device = true;
-			}
+		if (!skip_device &&
+		    !CheckFormat(device, vk::Format::eB8G8R8A8Srgb, true, storage_features)) {
+			LOGF("Format vk::Format::eB8G8R8A8Srgb cannot be used as storage image\n");
+			check_format(vk::Format::eB8G8R8A8Unorm, storage_features, "storage image");
 		}
 
 		if (!skip_device && device_properties.limits.maxSamplerAnisotropy < 16.0f) {
-			LOGF("maxSamplerAnisotropy < 16.0f");
-			skip_device = true;
+			reject("maxSamplerAnisotropy < 16.0f");
 		}
 
 		if (skip_device) {
@@ -520,16 +420,20 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 #endif
 	depth_clip_control.depthClipControl = VK_TRUE;
 
-	auto features12  = WindowContext::RequiredVulkan12Features();
-	features12.pNext = &depth_clip_control;
-
+	const bool workgroup_layout_extension =
+	    HasExtension(device_extensions, VK_KHR_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_EXTENSION_NAME);
+	vk::PhysicalDeviceWorkgroupMemoryExplicitLayoutFeaturesKHR supported_workgroup_layout {};
+	vk::PhysicalDeviceVulkan12Features supported_features12 {};
+	supported_features12.pNext = workgroup_layout_extension ? &supported_workgroup_layout : nullptr;
 	vk::PhysicalDeviceVulkan13Features supported_features13 {};
+	supported_features13.pNext = &supported_features12;
 
 	const auto robustness2_ext_enabled =
 	    HasExtension(device_extensions, VK_EXT_ROBUSTNESS_2_EXTENSION_NAME);
 
 	vk::PhysicalDeviceRobustness2FeaturesEXT supported_robustness2 {};
 	if (robustness2_ext_enabled) {
+		supported_robustness2.pNext = supported_features13.pNext;
 		supported_features13.pNext = &supported_robustness2;
 	}
 
@@ -557,6 +461,22 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 		supported_features2.pNext = &provoking_vertex;
 	}
 	physical_device.getFeatures2(&supported_features2);
+
+	auto features12 = WindowContext::RequiredVulkan12Features();
+	features12.shaderSharedInt64Atomics = supported_features12.shaderSharedInt64Atomics;
+	vk::PhysicalDeviceWorkgroupMemoryExplicitLayoutFeaturesKHR workgroup_layout {};
+	workgroup_layout.workgroupMemoryExplicitLayout =
+	    supported_workgroup_layout.workgroupMemoryExplicitLayout;
+	workgroup_layout.pNext = &depth_clip_control;
+	features12.pNext = workgroup_layout_extension ? static_cast<void*>(&workgroup_layout)
+	                                             : static_cast<void*>(&depth_clip_control);
+	if (!features12.shaderSharedInt64Atomics || !workgroup_layout.workgroupMemoryExplicitLayout) {
+		Log::WriteToConsoleAndLog(fmt::format(
+		    "WARNING: Native 64-bit LDS atomics are unavailable: shaderSharedInt64Atomics={}, "
+		    "workgroupMemoryExplicitLayout={}. Continuing with the selected Vulkan device.\n",
+		    features12.shaderSharedInt64Atomics != VK_FALSE,
+		    workgroup_layout.workgroupMemoryExplicitLayout != VK_FALSE));
+	}
 	graphics.mesh_shader_enabled = mesh_extension && supported_mesh.meshShader;
 
 	vk::PhysicalDeviceSubgroupSizeControlProperties subgroup_size_control {};
@@ -564,8 +484,10 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 	vk::PhysicalDeviceVulkan11Properties properties11 {};
 	properties11.pNext = &subgroup_size_control;
 
+	vk::PhysicalDeviceFloatControlsProperties float_controls {};
+	float_controls.pNext = &properties11;
 	vk::PhysicalDeviceProperties2 properties2 {};
-	properties2.pNext = &properties11;
+	properties2.pNext = &float_controls;
 
 	if (graphics.mesh_shader_enabled) {
 		subgroup_size_control.pNext = &graphics.mesh_shader_properties;
@@ -628,6 +550,15 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 	device_features.vertexPipelineStoresAndAtomics       = VK_TRUE;
 	graphics.sample_rate_shading_enabled                 = true;
 	device_features.shaderInt64 = VK_TRUE;
+	device_features.shaderFloat64 =
+	    supported_features2.features.shaderFloat64 &&
+	    float_controls.shaderSignedZeroInfNanPreserveFloat64 &&
+	    float_controls.shaderRoundingModeRTEFloat32;
+	// if (device_features.shaderFloat64 && !float_controls.shaderDenormPreserveFloat64) {
+	// 	Log::WriteToConsoleAndLog(
+	// 	    "WARNING: Vulkan device does not guarantee FP64 denormal preservation; "
+	// 	    "continuing with native FP64 arithmetic. Very small values may be flushed to zero.\n");
+	// }
 
 	vk::PhysicalDeviceRobustness2FeaturesEXT robustness2 {};
 #if defined(__APPLE__)
@@ -993,11 +924,13 @@ void WindowContext::CreateVulkan() {
 	}
 #endif
 
+	std::string rejected_devices;
 	VulkanFindPhysicalDevice(graphic_ctx.instance, surface, device_extensions, surface_capabilities,
-	                         graphic_ctx.physical_device, graphic_ctx.queue_family);
+	                         graphic_ctx.physical_device, graphic_ctx.queue_family,
+	                         rejected_devices);
 
 	if (graphic_ctx.physical_device == nullptr) {
-		EXIT("Could not find suitable device");
+		EXIT("Could not find suitable device:%s", rejected_devices.c_str());
 	}
 
 	vk::PhysicalDevicePushDescriptorProperties push_descriptor_properties {};
@@ -1038,6 +971,7 @@ void WindowContext::CreateVulkan() {
 		for (const auto* extension: {VK_EXT_ROBUSTNESS_2_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}) {
 			if (HasExtension(available_extensions, extension)) {
 				device_extensions.push_back(extension);

src/graphics/presentation/window/window.cpp

@@ -817,6 +817,7 @@ void WindowRun() {
 	EXIT_IF(g_window == nullptr);
 
 	g_window->Run();
+	Common::LockGuard lock(g_window->render_context->GetMutex());
 	g_window->render_context->GetPipelineCache().Save();
 }
 

src/graphics/shader/recompiler/ShaderRecompiler.cpp

@@ -15,12 +15,10 @@
 #include "graphics/shader/recompiler/ir/passes/ResourceMaterialization.h"
 #include "graphics/shader/recompiler/ir/passes/ResourceTracking.h"
 #include "graphics/shader/recompiler/ir/passes/ShaderInfoCollection.h"
-#include "graphics/shader/recompiler/ir/passes/SrtWalker.h"
 #include "graphics/shader/recompiler/ir/passes/SsaRewrite.h"
 
 #include <algorithm>
 #include <array>
-#include <atomic>
 #include <chrono>
 #include <fmt/format.h>
 #include <map>
@@ -524,20 +522,6 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 	     GetDumpLabel(options), StageName(options.stage), options.shader_hash,
 	     static_cast<uint64_t>(decoded.instructions.size()), phase_ms());
 
-	// Temporary workaround for games that compile ray-tracing shaders before
-	// the player can select a mode without ray tracing.
-	if (options.stage == ShaderType::Compute && decoded.has_bvh) {
-		static std::atomic_flag warned = ATOMIC_FLAG_INIT;
-		if (!warned.test_and_set(std::memory_order_relaxed)) {
-			const auto& bvh = decoded.instructions.back();
-			Log::WriteToConsoleAndLog(fmt::format(
-			    "Warning: ray tracing is not implemented; skipping compute dispatches containing "
-			    "BVH intersection instructions (shader=0x{:016x}, pc=0x{:08x}, opcode=0x{:02x}).\n",
-			    options.shader_hash, bvh.pc, bvh.opcode_id));
-		}
-		return {.skip_dispatch = true};
-	}
-
 	std::string decoded_dump;
 	if (options.dump_ir) {
 		decoded_dump = Decoder::ProgramToString(decoded);
@@ -548,30 +532,40 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 
 	LOGF("%s phase begin: stage=%s hash=0x%016" PRIx64 " CFG BuildGraph\n", GetDumpLabel(options),
 	     StageName(options.stage), options.shader_hash);
-	auto cfg = CFG::BuildGraph(decoded);
+	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
 	     " loops=%" PRIu64 " back_edges=%" PRIu64 " elapsed_ms=%" PRIu64 "\n",
 	     GetDumpLabel(options), StageName(options.stage), options.shader_hash,
-	     static_cast<uint64_t>(cfg.blocks.size()), static_cast<uint64_t>(cfg.natural_loops.size()),
-	     static_cast<uint64_t>(cfg.back_edges.size()), phase_ms());
-	if (cfg.irreducible) {
-		LogDispatcherFallback(options, cfg, "build");
+	     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);
-		if (!CFG::Structurize(cfg)) {
-			LogDispatcherFallback(options, cfg, "structurize");
+		structured_cfg = CFG::Structurize(native_cfg);
+		if (structured_cfg.unsupported) {
+			native_cfg.unsupported = true;
+			native_cfg.failure_kind = structured_cfg.failure_kind;
+			native_cfg.failure_block = structured_cfg.failure_block;
+			native_cfg.unsupported_reason = structured_cfg.unsupported_reason;
+			LogDispatcherFallback(options, native_cfg, "structurize");
 		} else {
+			selected_cfg = &structured_cfg;
 			LOGF("%s structured CFG success: blocks=%" PRIu64 "\n", GetDumpLabel(options),
-			     static_cast<uint64_t>(cfg.blocks.size()));
+			     static_cast<uint64_t>(selected_cfg->blocks.size()));
 		}
 		LOGF("%s phase end: stage=%s hash=0x%016" PRIx64 " CFG Structurize blocks=%" PRIu64
 		     " loops=%" PRIu64 " elapsed_ms=%" PRIu64 "\n",
 		     GetDumpLabel(options), StageName(options.stage), options.shader_hash,
-		     static_cast<uint64_t>(cfg.blocks.size()),
-		     static_cast<uint64_t>(cfg.natural_loops.size()), phase_ms());
+		     static_cast<uint64_t>(selected_cfg->blocks.size()),
+		     static_cast<uint64_t>(selected_cfg->natural_loops.size()), phase_ms());
 	}
 
+	const auto& cfg = *selected_cfg;
 	Frontend::EmbeddedFetchPlan embedded_fetch;
 	if ((options.stage == ShaderType::Vertex || options.stage == ShaderType::Local) &&
 	    options.input_info.vertex != nullptr && options.input_info.vertex->fetch_embedded) {
@@ -600,7 +594,7 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 	     GetDumpLabel(options), StageName(options.stage), options.shader_hash,
 	     static_cast<uint64_t>(ir.blocks.size()), phase_ms());
 	IR::RewriteToSsa(ir.blocks);
-	IR::ConstantPropagationPass(ir.blocks);
+	IR::ConstantPropagationPass(ir.blocks, ir.wave_size);
 	IR::ResolveControlFlowIdentities(ir);
 	IR::RemoveIdentities(ir.blocks);
 	IR::EliminateDeadCode(ir.blocks);
@@ -608,21 +602,27 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 	if (read_lane_stats.rewritten_reads != 0) {
 		LOGF("%s read-lane elimination: reads=%" PRIu32 "\n", GetDumpLabel(options),
 		     read_lane_stats.rewritten_reads);
-		IR::ConstantPropagationPass(ir.blocks);
+		IR::ConstantPropagationPass(ir.blocks, ir.wave_size);
 		IR::ResolveControlFlowIdentities(ir);
 		IR::RemoveIdentities(ir.blocks);
 		IR::EliminateDeadCode(ir.blocks);
 	}
 	LowerTessellationMemory(ir, options);
-	IR::BuildSrtPlan(ir);
-	IR::EliminateDeadCode(ir.blocks);
-	IR::TrackResources(ir);
+	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());
+		}
+	}
+	IR::TrackResources(ir, decoded, native_cfg);
 	IR::EliminateDeadCode(ir.blocks);
 	TranslateResult result;
 	result.program = std::move(ir);
 	if (options.dump_ir) {
 		result.decoded_dump = std::move(decoded_dump);
-		result.cfg_dump     = CFG::GraphToString(cfg);
+		result.cfg_dump     = std::move(cfg_dump);
 	}
 	return result;
 }
@@ -630,10 +630,38 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 CompileResult CompileProgram(TranslateResult translated, const CompileOptions& options,
                              const IR::ResourceSpecialization& specialization,
                              uint32_t push_data_start_dword) {
-	EXIT_IF(translated.skip_dispatch);
 	const auto emit_begin = std::chrono::steady_clock::now();
 	auto& ir = translated.program;
 	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.
+	for (auto& inst: ir.value_storage) {
+		inst.Invalidate();
+	}
+	for (auto* block: ir.blocks) {
+		for (auto& inst: *block) {
+			const auto op = inst.GetOpcode();
+			uint32_t first = 0;
+			if (op == IR::ValueOpcode::GetBufferResource) {
+				if (std::ranges::any_of(inst.Uses(), [&](const IR::Use& use) {
+					return ir.memory_info[use.user->Flags<IR::MemoryFlags>().index].kind ==
+					       IR::ResourceKind::IndirectBuffer;
+				})) {
+					continue;
+				}
+			} else if (op == IR::ValueOpcode::GetImageResource) {
+				const auto resource = inst.Flags<uint32_t>();
+				first = resource < ir.info.images.size() &&
+				                ir.info.images[resource].indirect_root == resource ? 1u : 0u;
+			} else if (op != IR::ValueOpcode::GetSamplerResource) {
+				continue;
+			}
+			for (size_t index = first; index < inst.NumArgs(); index++) {
+				inst.SetArg(index, IR::Value(0u));
+			}
+		}
+	}
+	ir.value_storage.clear();
 	IR::RemoveIdentities(ir.blocks);
 	IR::EliminateDeadCode(ir.blocks);
 

src/graphics/shader/recompiler/ShaderRecompiler.h

@@ -28,7 +28,6 @@ struct TranslateResult {
 	IR::Program program;
 	std::string decoded_dump;
 	std::string cfg_dump;
-	bool        skip_dispatch = false;
 };
 
 struct CompileResult {

src/graphics/shader/recompiler/Tessellation.cpp

@@ -217,7 +217,7 @@ void LowerTessellationMemory(IR::Program& program, const CompileOptions& options
 			    2, ActiveAddress(*block, it, it->Arg(2), it->Arg(it->NumArgs() - 1u), resolved));
 		}
 	}
-	ConstantPropagationPass(program.blocks);
+	ConstantPropagationPass(program.blocks, program.wave_size);
 	uint32_t reads = 0, writes = 0, factors = 0;
 	for (auto* block: program.blocks) {
 		for (auto it = block->begin(); it != block->end(); ++it) {
@@ -317,7 +317,7 @@ void LowerTessellationMemory(IR::Program& program, const CompileOptions& options
 			inst.Invalidate();
 		}
 	}
-	ConstantPropagationPass(program.blocks);
+	ConstantPropagationPass(program.blocks, program.wave_size);
 	RemoveIdentities(program.blocks);
 	EliminateDeadCode(program.blocks);
 	LOGF("%s tessellation lowering: reads=%u writes=%u factors=%u\n",

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

@@ -3,6 +3,7 @@
 #include "common/assert.h"
 #include "graphics/shader/recompiler/backend/spirv/spirvEmitterInternal.h"
 #include "graphics/shader/recompiler/ir/ShaderIR.h"
+#include "graphics/shader/recompiler/ir/passes/BindingLayout.h"
 
 #include <algorithm>
 #include <array>
@@ -34,9 +35,6 @@ void ValidateNativeProgram(const IR::Program& program) {
 		present[index]   = true;
 		expected[index]  = std::move(resources);
 	};
-	if (!program.info.buffers.empty()) {
-		Expect(Kind::Buffers, Dense(program.info.buffers.size()));
-	}
 	for (uint32_t i = 0; i < program.info.images.size(); i++) {
 		const auto kind = IR::DescriptorBindingForImage(program.info.images[i]);
 		if (!kind.has_value()) {
@@ -54,23 +52,9 @@ void ValidateNativeProgram(const IR::Program& program) {
 	if (!program.info.samplers.empty()) {
 		Expect(Kind::Samplers, Dense(program.info.samplers.size()));
 	}
-	bool uses_gds = false;
-	for (const auto* block: program.blocks) {
-		for (const auto& inst: *block) {
-			if (IR::SharedAccessOf(inst.GetOpcode()) == IR::SharedAccess::None) {
-				continue;
-			}
-			const auto index = inst.Flags<IR::MemoryFlags>().index;
-			if (index >= program.memory_info.size()) {
-				Fail(program, "shared operation has invalid memory metadata");
-			}
-			const auto kind = program.memory_info[index].kind;
-			if (kind != IR::ResourceKind::Lds && kind != IR::ResourceKind::Gds) {
-				Fail(program, "shared operation has invalid resource kind");
-			}
-			uses_gds |= kind == IR::ResourceKind::Gds;
-		}
-	}
+	auto& buffers = expected[static_cast<size_t>(Kind::Buffers)];
+	const bool uses_gds = IR::CollectMemoryResources(program, buffers);
+	present[static_cast<size_t>(Kind::Buffers)] = !buffers.empty();
 	if (uses_gds) {
 		Expect(Kind::Gds);
 	}
@@ -78,12 +62,7 @@ void ValidateNativeProgram(const IR::Program& program) {
 		Expect(Kind::BdaPagetable);
 		Expect(Kind::FaultBuffer);
 	}
-	const bool uses_flattened_runtime =
-	    !program.srt_reads.empty() ||
-	     std::ranges::any_of(program.info.images, [](const IR::ImageResource& image) {
-		     return image.indirect_search_iterations != 0u;
-	     });
-	if (uses_flattened_runtime) {
+	if (IR::UsesFlattenedSrt(program)) {
 		Expect(Kind::FlattenedSrt);
 	}
 	if (program.bindings.ShaderDataDwords() != 0 && !program.bindings.UsesPushData()) {
@@ -109,7 +88,7 @@ void ValidateNativeProgram(const IR::Program& program) {
 	if ((program.bindings.UsesPushData() &&
 	     !IR::PushData::CanFit(program.bindings.push_data_start_dword, shader_data_dwords)) ||
 	    program.bindings.memory_offset_dword != program.bindings.user_data_registers.size() ||
-	    program.bindings.memory_offset_count != program.info.buffers.size() ||
+	    program.bindings.memory_offset_count != buffers.size() ||
 	    has_shader_data_storage != (shader_data_dwords != 0 && !program.bindings.UsesPushData()) ||
 	    !std::is_sorted(program.bindings.user_data_registers.begin(),
 	                    program.bindings.user_data_registers.end()) ||
@@ -143,6 +122,16 @@ void ValidateNativeProgram(const IR::Program& program) {
 			              program.memory_info[index].kind == IR::ResourceKind::IndirectBuffer;
 		       });
 	};
+	const auto local_flat_handle = [&](const IR::Inst& handle) {
+		return !handle.Uses().empty() &&
+		       std::ranges::all_of(handle.Uses(), [&](const IR::Use& use) {
+			       if (IR::AddressOpcodeInfoOf(use.user->GetOpcode()).access == IR::AddressAccess::None)
+				       return false;
+			       const auto index = use.user->Flags<IR::MemoryFlags>().index;
+			       return index < program.memory_info.size() &&
+			              program.memory_info[index].kind == IR::ResourceKind::FlatLocal;
+		       });
+	};
 	for (const auto* block: program.blocks) {
 		for (const auto& inst: *block) {
 			const auto dense = inst.Flags<uint32_t>();
@@ -159,7 +148,7 @@ void ValidateNativeProgram(const IR::Program& program) {
 					if (planning_only_handle(inst)) {
 						break;
 					}
-					if (inst.NumArgs() != 2 || !program.info.uses_dma) {
+					if (inst.NumArgs() != 2 || (!program.info.uses_dma && !local_flat_handle(inst))) {
 						Fail(program, "typed address handle has invalid DMA metadata");
 					}
 					break;
@@ -198,6 +187,7 @@ Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program
 	SpirvRequirements requirements {};
 	for (const auto* block: program.blocks) {
 		for (const auto& inst: *block) {
+			requirements.float64 |= inst.GetType() == IR::Type::F64;
 			if (IR::BufferAccessOf(inst.GetOpcode()) == IR::BufferAccess::Atomic &&
 			    inst.GetType() == IR::Type::U64) {
 				requirements.buffer_int64_atomics = true;
@@ -208,11 +198,16 @@ Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program
 				if (memory_index >= program.memory_info.size()) {
 					Fail(program, "address operation has invalid memory metadata");
 				}
-				if (program.memory_info[memory_index].kind == IR::ResourceKind::Scratch) {
+				const auto kind = program.memory_info[memory_index].kind;
+				if (kind == IR::ResourceKind::Scratch || kind == IR::ResourceKind::FlatLocal) {
 					if (program.scratch_dwords == 0) {
 						Fail(program, "scratch operation has no per-thread storage");
 					}
 					requirements.function_scratch = true;
+					if (kind == IR::ResourceKind::FlatLocal && program.stage != ShaderType::Compute &&
+					    program.stage != ShaderType::Mesh) {
+						requirements.function_lds = true;
+					}
 				} else if (address_access == IR::AddressAccess::Write) {
 					Fail(program, "writable FLAT/GLOBAL addresses require GPU ownership tracking");
 				}
@@ -249,6 +244,13 @@ Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program
 				if (kind != IR::ResourceKind::Lds && kind != IR::ResourceKind::Gds) {
 					Fail(program, "shared operation has invalid resource kind");
 				}
+				if (shared_access == IR::SharedAccess::Atomic &&
+				    IR::SharedComponentCount(inst.GetOpcode()) == 2u) {
+					if (kind != IR::ResourceKind::Lds || program.stage != ShaderType::Compute) {
+						Fail(program, "64-bit shared atomics require compute LDS");
+					}
+					requirements.shared_int64_atomics = true;
+				}
 				if (program.stage != ShaderType::Compute && program.stage != ShaderType::Mesh &&
 				    kind == IR::ResourceKind::Lds) {
 					requirements.function_lds = true;
@@ -261,6 +263,7 @@ Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program
 				}
 			}
 			switch (inst.GetOpcode()) {
+				case IR::ValueOpcode::BvhIntersect: requirements.bvh = true; break;
 				case IR::ValueOpcode::Ballot: requirements.subgroup_ballot = true; break;
 				case IR::ValueOpcode::DppMoveU32:
 				case IR::ValueOpcode::ReadFirstLane:

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

@@ -291,9 +291,25 @@ uint32_t EmitConvertU32F32(EmitterState& state, uint32_t arg0) {
 	return EmitF32ToU32(state, arg0, false);
 }
 
-uint32_t EmitConvertF32S32(EmitterState& state, uint32_t arg0) {
-	const auto signed_value = Unary(state, spv::OpBitcast, TypeI32(state), arg0);
-	return EmitNative<spv::OpConvertSToF, IR::Type::F32>(state, signed_value);
+uint32_t EmitConvertF32F64(EmitterState& state, uint32_t arg0) {
+	const auto converted = Unary(state, spv::OpFConvert, TypeF32(state), arg0);
+	const auto source    = EmitNative<spv::OpCompositeExtract, IR::Type::U32>(
+	    state, Unary(state, spv::OpBitcast, TypeU64(state), arg0), 1u);
+	const auto exponent = EmitAndConstant(state, source, 0x7ff00000u);
+	const auto overflow =
+	    Binary(state, spv::OpLogicalAnd, TypeBool(state),
+	           EmitCompareU32Constant(state, spv::OpUGreaterThan, exponent, 0x47e00000u),
+	           EmitCompareU32Constant(state, spv::OpULessThan, exponent, 0x7ff00000u));
+	const auto clamped = EmitOrU32(state, EmitAndConstant(state, source, 0x80000000u),
+	                               ConstantU32(state, 0x7f7fffffu));
+	return EmitFlushF32DenormToSignedZero(
+	    state, Select(state, TypeF32(state), overflow,
+	                  Unary(state, spv::OpBitcast, TypeF32(state), clamped), converted));
+}
+
+uint32_t EmitConvertF64F32(EmitterState& state, uint32_t arg0) {
+	return EmitNative<spv::OpFConvert, IR::Type::F64>(state,
+	                                                  EmitFlushF32DenormToSignedZero(state, arg0));
 }
 
 uint32_t EmitCompositeExtractU64(EmitterState& state, uint32_t arg0, IR::Value arg1) {
@@ -428,6 +444,18 @@ uint32_t EmitUGreaterThan64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
 	return CompareOrdered64(state, arg0, arg1, spv::OpUGreaterThan, spv::OpUGreaterThan);
 }
 
+uint32_t EmitSLessThanEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
+	return CompareOrdered64(state, arg0, arg1, spv::OpSLessThan, spv::OpULessThanEqual);
+}
+
+uint32_t EmitULessThanEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
+	return CompareOrdered64(state, arg0, arg1, spv::OpULessThan, spv::OpULessThanEqual);
+}
+
+uint32_t EmitUGreaterThanEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
+	return CompareOrdered64(state, arg0, arg1, spv::OpUGreaterThan, spv::OpUGreaterThanEqual);
+}
+
 uint32_t EmitFPIsNan32(EmitterState& state, uint32_t arg0) {
 	return EmitNative<spv::OpFUnordNotEqual, IR::Type::U1>(state, arg0, arg0);
 }
@@ -453,6 +481,11 @@ uint32_t EmitFPRecip32(EmitterState& state, uint32_t arg0) {
 	return Binary(state, spv::OpFDiv, TypeF32(state), ConstantF32(state, 0x3f800000u), source);
 }
 
+uint32_t EmitFPRecip64(EmitterState& state, uint32_t arg0) {
+	const auto one = state.builder.Constant(spv::OpConstant, TypeF64(state), 0u, 0x3ff00000u);
+	return EmitNative<spv::OpFDiv, IR::Type::F64>(state, one, arg0);
+}
+
 uint32_t EmitFPRecipIFlag32(EmitterState& state, uint32_t arg0) {
 	// Integer-to-float inputs used by IFLAG cannot be denormal.
 	return Binary(state, spv::OpFDiv, TypeF32(state), ConstantF32(state, 0x3f800000u), arg0);

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

@@ -203,9 +203,7 @@ uint32_t EmitFlushF32DenormToSignedZero(EmitterState& state, uint32_t value) {
 	const auto bits      = state.builder.AllocateId();
 	const auto abs_bits  = state.builder.AllocateId();
 	const auto sign_bits = state.builder.AllocateId();
-	const auto non_zero  = state.builder.AllocateId();
 	const auto subnormal = state.builder.AllocateId();
-	const auto flush     = state.builder.AllocateId();
 	const auto selected  = state.builder.AllocateId();
 	const auto ret       = state.builder.AllocateId();
 	state.builder.AddFunction(spv::OpBitcast, TypeU32(state), bits, value);
@@ -213,12 +211,9 @@ uint32_t EmitFlushF32DenormToSignedZero(EmitterState& state, uint32_t value) {
 	                          ConstantU32(state, 0x7fffffffu));
 	state.builder.AddFunction(spv::OpBitwiseAnd, TypeU32(state), sign_bits, bits,
 	                          ConstantU32(state, 0x80000000u));
-	state.builder.AddFunction(spv::OpINotEqual, TypeBool(state), non_zero, abs_bits,
-	                          ConstantU32(state, 0));
 	state.builder.AddFunction(spv::OpULessThan, TypeBool(state), subnormal, abs_bits,
 	                          ConstantU32(state, 0x00800000u));
-	state.builder.AddFunction(spv::OpLogicalAnd, TypeBool(state), flush, non_zero, subnormal);
-	state.builder.AddFunction(spv::OpSelect, TypeU32(state), selected, flush, sign_bits, bits);
+	state.builder.AddFunction(spv::OpSelect, TypeU32(state), selected, subnormal, sign_bits, bits);
 	state.builder.AddFunction(spv::OpBitcast, TypeF32(state), ret, selected);
 	return ret;
 }

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

@@ -0,0 +1,348 @@
+#include "graphics/shader/recompiler/backend/spirv/spirvEmitterInstructions.h"
+
+namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter {
+namespace {
+
+using Vec3 = std::array<uint32_t, 3>;
+
+uint32_t Extract(EmitterState& s, uint32_t type, uint32_t vector, uint32_t index) {
+	const auto result = s.builder.AllocateId();
+	s.builder.AddFunction(spv::OpCompositeExtract, type, result, vector, index);
+	return result;
+}
+
+uint32_t Vector(EmitterState& s, uint32_t type, std::span<const uint32_t> components) {
+	const auto result = s.builder.AllocateId();
+	s.builder.AddFunction(spv::OpCompositeConstruct, type, result, components);
+	return result;
+}
+
+uint32_t NodeWord(EmitterState& s, uint32_t block, uint32_t word) {
+	const auto offset = Binary(s, spv::OpShiftLeftLogical, TypeU32(s), word, ConstantU32(s, 2));
+	const auto address = Binary(s, spv::OpIAdd, TypeScalarU64(s), block,
+	                            Unary(s, spv::OpUConvert, TypeScalarU64(s), offset));
+	const auto pointer = Unary(s, spv::OpConvertUToPtr, TypePhysicalU32Pointer(s), address);
+	const auto result = s.builder.AllocateId();
+	s.builder.AddFunction(spv::OpLoad, TypeU32(s), result, pointer, spv::MemoryAccessAlignedMask, 4u);
+	return result;
+}
+
+uint32_t Triangle(EmitterState& s, uint32_t block, uint32_t kind, uint32_t bary,
+                  Vec3 origin, Vec3 direction) {
+	const auto f = TypeF32(s), u = TypeU32(s), b = TypeBool(s);
+	const auto zero = ConstantF32Value(s, 0.0f);
+	const auto op = [&](spv::Op code, uint32_t x, uint32_t y) {
+		const auto value = Binary(s, code, f, x, y);
+		s.builder.AddAnnotation(spv::OpDecorate, value, spv::DecorationNoContraction);
+		return value;
+	};
+	const auto cmp = [&](spv::Op code, uint32_t x, uint32_t y) { return Binary(s, code, b, x, y); };
+	const auto either = [&](uint32_t x, uint32_t y) { return Binary(s, spv::OpLogicalOr, b, x, y); };
+	const auto both = [&](uint32_t x, uint32_t y) { return Binary(s, spv::OpLogicalAnd, b, x, y); };
+	const auto abs = [&](uint32_t x) { return EmitGlsl<GLSLstd450FAbs, IR::Type::F32>(s, x); };
+	const auto is_kind = [&](uint32_t n) { return cmp(spv::OpIEqual, kind, ConstantU32(s, n)); };
+	// The four triangles are (0,1,2), (1,3,2), (2,3,4), (2,4,0).
+	const std::array vertices {
+		EmitGlsl<GLSLstd450UMin, IR::Type::U32>(s, kind, ConstantU32(s, 2)),
+		Select(s, u, is_kind(0), ConstantU32(s, 1),
+		       Select(s, u, is_kind(3), ConstantU32(s, 4), ConstantU32(s, 3))),
+		Select(s, u, is_kind(2), ConstantU32(s, 4),
+		       Select(s, u, is_kind(3), ConstantU32(s, 0), ConstantU32(s, 2))) };
+	std::array<Vec3, 3> positions;
+	for (uint32_t i = 0; i < 3; ++i) {
+		const auto first = Binary(s, spv::OpIMul, u, vertices[i], ConstantU32(s, 3));
+		for (uint32_t axis = 0; axis < 3; ++axis)
+			positions[i][axis] = Unary(s, spv::OpBitcast, f,
+			    NodeWord(s, block, Binary(s, spv::OpIAdd, u, first, ConstantU32(s, axis))));
+	}
+	const auto flag = NodeWord(s, block, ConstantU32(s, 15));
+	const auto y_largest = cmp(spv::OpFOrdLessThan, abs(direction[0]), abs(direction[1]));
+	const auto z_largest = cmp(spv::OpFOrdLessThan,
+	    Select(s, f, y_largest, abs(direction[1]), abs(direction[0])), abs(direction[2]));
+	const auto rotate = [&](Vec3 value) {
+		Vec3 result;
+		for (uint32_t i = 0; i < 3; ++i)
+			result[i] = Select(s, f, z_largest, value[i],
+			    Select(s, f, y_largest, value[(i + 2) % 3], value[(i + 1) % 3]));
+		return result;
+	};
+	origin = rotate(origin);
+	direction = rotate(direction);
+	std::array<Vec3, 3> projected;
+	for (uint32_t i = 0; i < 3; ++i) {
+		const auto position = rotate(positions[i]);
+		Vec3 relative;
+		for (uint32_t axis = 0; axis < 3; ++axis)
+			relative[axis] = op(spv::OpFSub, position[axis], origin[axis]);
+		projected[i] = {op(spv::OpFSub, op(spv::OpFMul, relative[0], direction[2]),
+		                              op(spv::OpFMul, direction[0], relative[2])),
+		                op(spv::OpFSub, op(spv::OpFMul, relative[1], direction[2]),
+		                              op(spv::OpFMul, direction[1], relative[2])), relative[2]};
+	}
+	Vec3 edge, weighted;
+	for (uint32_t i = 0; i < 3; ++i) {
+		const auto& first = projected[(i + 1) % 3];
+		const auto& second = projected[(i + 2) % 3];
+		edge[i] = op(spv::OpFSub, op(spv::OpFMul, second[0], first[1]),
+		                        op(spv::OpFMul, second[1], first[0]));
+		weighted[i] = op(spv::OpFMul, edge[i], direction[2]);
+	}
+	auto numerator = op(spv::OpFAdd, op(spv::OpFAdd,
+	    op(spv::OpFMul, edge[0], projected[0][2]), op(spv::OpFMul, edge[1], projected[1][2])),
+	    op(spv::OpFMul, edge[2], projected[2][2]));
+	auto denominator = op(spv::OpFAdd, op(spv::OpFAdd, weighted[0], weighted[1]), weighted[2]);
+	uint32_t negative = ConstantBool(s, false), positive = negative;
+	for (auto value: edge) {
+		negative = either(negative, cmp(spv::OpFOrdLessThan, value, zero));
+		positive = either(positive, cmp(spv::OpFOrdGreaterThan, value, zero));
+	}
+	const auto winding = cmp(spv::OpFOrdGreaterThan, denominator, zero);
+	auto missed = either(both(negative, positive), cmp(spv::OpFOrdEqual, denominator, zero));
+	missed = either(missed, Unary(s, spv::OpIsNan, b, numerator));
+	missed = either(missed, cmp(spv::OpFOrdLessThan,
+	    Select(s, f, winding, numerator, Unary(s, spv::OpFNegate, f, numerator)), zero));
+	for (uint32_t i = 0; i < 3; ++i) {
+		const auto first = projected[(i + 1) % 3][1], second = projected[(i + 2) % 3][1];
+		const auto first_zero = cmp(spv::OpFOrdEqual, first, zero);
+		const auto horizontal = both(first_zero, cmp(spv::OpFOrdEqual, second, zero));
+		const auto below = either(cmp(spv::OpFOrdLessThan, first, zero),
+		    both(first_zero, cmp(spv::OpFOrdGreaterThan, second, zero)));
+		const auto right = Binary(s, spv::OpLogicalNotEqual, b, below, winding);
+		const auto excluded = Select(s, b, horizontal,
+		    cmp(spv::OpFOrdGreaterThan, projected[i][1], zero), right);
+		missed = either(missed, both(cmp(spv::OpFOrdEqual, edge[i], zero), excluded));
+	}
+	const auto remap = Binary(s, spv::OpShiftRightLogical, u, flag,
+	    Binary(s, spv::OpIMul, u, kind, ConstantU32(s, 8)));
+	const auto coordinate = [&](uint32_t shift) {
+		const auto index = Binary(s, spv::OpBitwiseAnd, u,
+		    Binary(s, spv::OpShiftRightLogical, u, remap, ConstantU32(s, shift)), ConstantU32(s, 3));
+		return Unary(s, spv::OpBitcast, u, Select(s, f,
+		    cmp(spv::OpIEqual, index, ConstantU32(s, 1)), weighted[1],
+		    Select(s, f, cmp(spv::OpIEqual, index, ConstantU32(s, 2)), weighted[2], weighted[0])));
+	};
+	const std::array result {
+		Unary(s, spv::OpBitcast, u, Select(s, f, missed, ConstantF32(s, 0x7f800000), numerator)),
+		Unary(s, spv::OpBitcast, u, Select(s, f, missed, ConstantF32Value(s, 1.0f), denominator)),
+		Select(s, u, bary, coordinate(0), Binary(s, spv::OpIAdd, u, flag, kind)),
+		Select(s, u, bary, coordinate(2), Select(s, u, missed, ConstantU32(s, 0), ConstantU32(s, 1))) };
+	return Vector(s, TypeU32Vector(s, 4), result);
+}
+
+std::array<uint32_t, 6> Box(EmitterState& s, uint32_t first, uint32_t second,
+                           uint32_t half, uint32_t child) {
+	const auto f = TypeF32(s);
+	const auto half_label = s.builder.AllocateId(), full_label = s.builder.AllocateId();
+	const auto merge = s.builder.AllocateId();
+	s.builder.AddFunction(spv::OpSelectionMerge, merge, spv::SelectionControlMaskNone);
+	s.builder.AddFunction(spv::OpBranchConditional, half, half_label, full_label);
+	EmitLabel(s, half_label);
+	std::array<uint32_t, 6> packed, full, result;
+	for (uint32_t pair = 0; pair < 3; ++pair) {
+		const auto word = NodeWord(s, first, ConstantU32(s, 4 + child * 3 + pair));
+		const auto unpacked = EmitGlsl<GLSLstd450UnpackHalf2x16, IR::Type::F32x2>(s, word);
+		packed[pair * 2] = Extract(s, f, unpacked, 0);
+		packed[pair * 2 + 1] = Extract(s, f, unpacked, 1);
+	}
+	s.builder.AddFunction(spv::OpBranch, merge);
+	EmitLabel(s, full_label);
+	for (uint32_t component = 0; component < 6; ++component) {
+		const auto word = 4 + child * 6 + component;
+		full[component] = Unary(s, spv::OpBitcast, f,
+		    NodeWord(s, word < 16 ? first : second, ConstantU32(s, word % 16)));
+	}
+	s.builder.AddFunction(spv::OpBranch, merge);
+	EmitLabel(s, merge);
+	for (uint32_t component = 0; component < 6; ++component) {
+		result[component] = s.builder.AllocateId();
+		s.builder.AddFunction(spv::OpPhi, f, result[component], packed[component], half_label,
+		                      full[component], full_label);
+	}
+	return result;
+}
+
+uint32_t Boxes(EmitterState& s, uint32_t first, uint32_t second, uint32_t half,
+               uint32_t grow, uint32_t sort, uint32_t extent, const Vec3& origin, const Vec3& inverse) {
+	const auto f = TypeF32(s), u = TypeU32(s), b = TypeBool(s);
+	const auto zero = ConstantF32Value(s, 0.0f);
+	std::array<uint32_t, 4> depths, indices;
+	Vec3 forward;
+	for (uint32_t axis = 0; axis < 3; ++axis)
+		forward[axis] = Binary(s, spv::OpFOrdGreaterThanEqual, b, inverse[axis], zero);
+	for (uint32_t child = 0; child < 4; ++child) {
+		const auto box = Box(s, first, second, half, child);
+		Vec3 near, far;
+		for (uint32_t axis = 0; axis < 3; ++axis) {
+			const auto lo = Binary(s, spv::OpFMul, f,
+			    Binary(s, spv::OpFSub, f, box[axis], origin[axis]), inverse[axis]);
+			const auto hi = Binary(s, spv::OpFMul, f,
+			    Binary(s, spv::OpFSub, f, box[axis + 3], origin[axis]), inverse[axis]);
+			near[axis] = Select(s, f, forward[axis], lo, hi);
+			far[axis] = Select(s, f, forward[axis], hi, lo);
+		}
+		const auto entry = EmitGlsl<GLSLstd450FMax, IR::Type::F32>(s,
+		    EmitGlsl<GLSLstd450FMax, IR::Type::F32>(s, near[0], near[1]), near[2]);
+		const auto exit = EmitGlsl<GLSLstd450FMin, IR::Type::F32>(s,
+		    EmitGlsl<GLSLstd450FMin, IR::Type::F32>(s, far[0], far[1]), far[2]);
+		// Valid hits have nonnegative exit distance. Saturate ULP growth at +infinity.
+		const auto bits = Unary(s, spv::OpBitcast, u, Select(s, f,
+		    Binary(s, spv::OpFOrdEqual, b, exit, zero), zero, exit));
+		const auto grown_bits = EmitGlsl<GLSLstd450UMin, IR::Type::U32>(s,
+		    Binary(s, spv::OpIAdd, u, bits, grow), ConstantU32(s, 0x7f800000));
+		const auto grown = Unary(s, spv::OpBitcast, f, grown_bits);
+		auto hit = Binary(s, spv::OpLogicalAnd, b,
+		    Binary(s, spv::OpFOrdGreaterThanEqual, b, exit, zero),
+		    Binary(s, spv::OpFOrdLessThan, b, entry, extent));
+		hit = Binary(s, spv::OpLogicalAnd, b, hit,
+		    Binary(s, spv::OpFOrdLessThanEqual, b, entry, grown));
+		depths[child] = Unary(s, spv::OpBitcast, u, Select(s, f, hit,
+		    Select(s, f, Binary(s, spv::OpFOrdGreaterThan, b, entry, zero), entry, zero),
+		    ConstantF32(s, 0x7f800000)));
+		indices[child] = Select(s, u, hit, NodeWord(s, first, ConstantU32(s, child)), ConstantU32(s, ~0u));
+	}
+	// A select network keeps the original child order when sorting is disabled.
+	constexpr std::array<std::pair<uint32_t, uint32_t>, 5> network {{{0,1}, {2,3}, {0,2}, {1,3}, {1,2}}};
+	for (size_t step = 0; step < network.size(); ++step) {
+		const auto [a, c] = network[step];
+		const auto swap = Binary(s, spv::OpLogicalAnd, b, sort,
+		    Binary(s, spv::OpUGreaterThan, b, depths[a], depths[c]));
+		const auto depth = depths[a], index = indices[a];
+		const auto later = [&](uint32_t position) {
+			return std::ranges::any_of(std::span(network).subspan(step + 1), [=](auto pair) {
+				return pair.first == position || pair.second == position;
+			});
+		};
+		if (later(a)) depths[a] = Select(s, u, swap, depths[c], depth);
+		if (later(c)) depths[c] = Select(s, u, swap, depth, depths[c]);
+		indices[a] = Select(s, u, swap, indices[c], index);
+		indices[c] = Select(s, u, swap, index, indices[c]);
+	}
+	return Vector(s, TypeU32Vector(s, 4), indices);
+}
+
+} // namespace
+
+void DefineBvhIntersect(EmitterState& s) {
+	if (!s.requirements.bvh) return;
+	const auto u = TypeU32(s), f = TypeF32(s), b = TypeBool(s), wide = TypeScalarU64(s);
+	const auto vec4 = TypeU32Vector(s, 4), vec3 = TypeF32Vector(s, 3);
+	const auto signature = s.builder.Type(spv::OpTypeFunction, vec4, vec4, u, f, vec3, vec3, vec3);
+	s.bvh_intersect_function = s.builder.AllocateId();
+	s.builder.AddName(s.bvh_intersect_function, "bvh_intersect");
+	s.builder.AddFunction(spv::OpFunction, vec4, s.bvh_intersect_function,
+	                      spv::FunctionControlMaskNone, signature);
+	std::array<uint32_t, 6> args;
+	const std::array types {vec4, u, f, vec3, vec3, vec3};
+	for (uint32_t i = 0; i < args.size(); ++i) {
+		args[i] = s.builder.AllocateId();
+		s.builder.AddFunction(spv::OpFunctionParameter, types[i], args[i]);
+	}
+	EmitLabel(s, s.builder.AllocateId());
+	const auto descriptor = args[0], node = args[1], extent = args[2];
+	Vec3 origin, direction, inverse;
+	for (uint32_t i = 0; i < 3; ++i) {
+		origin[i] = Extract(s, f, args[3], i);
+		direction[i] = Extract(s, f, args[4], i);
+		inverse[i] = Extract(s, f, args[5], i);
+	}
+	std::array<uint32_t, 4> words;
+	for (uint32_t i = 0; i < 4; ++i) words[i] = Extract(s, u, descriptor, i);
+	const auto and_bits = [&](uint32_t value, uint32_t mask) {
+		return Binary(s, spv::OpBitwiseAnd, u, value, ConstantU32(s, mask));
+	};
+	const auto shr = [&](uint32_t value, uint32_t shift) {
+		return Binary(s, spv::OpShiftRightLogical, u, value, ConstantU32(s, shift));
+	};
+	const auto kind = and_bits(node, 7);
+	const auto triangle = Binary(s, spv::OpULessThan, b, kind, ConstantU32(s, 4));
+	const auto full = Binary(s, spv::OpIEqual, b, kind, ConstantU32(s, 5));
+	const auto half = Binary(s, spv::OpIEqual, b, kind, ConstantU32(s, 4));
+	const auto bary = Binary(s, spv::OpINotEqual, b, and_bits(words[3], 0x01000000), ConstantU32(s, 0));
+	const auto sort = Binary(s, spv::OpINotEqual, b, and_bits(words[1], 0x80000000), ConstantU32(s, 0));
+	const auto grow = and_bits(shr(words[1], 23), 0xff);
+	const auto base = Binary(s, spv::OpShiftLeftLogical, wide,
+	    DeviceAddressFromWords(s, words[0], and_bits(words[1], 0xff)), ConstantDeviceAddress(s, 8));
+	const auto index = Unary(s, spv::OpUConvert, wide, shr(node, 3));
+	const auto last_index = Binary(s, spv::OpIAdd, wide, index,
+	    Select(s, wide, full, ConstantDeviceAddress(s, 1), ConstantDeviceAddress(s, 0)));
+	const auto address = Binary(s, spv::OpIAdd, wide, base,
+	    Binary(s, spv::OpShiftLeftLogical, wide, index, ConstantDeviceAddress(s, 6)));
+	const auto last_address = Binary(s, spv::OpIAdd, wide, base,
+	    Binary(s, spv::OpShiftLeftLogical, wide, last_index, ConstantDeviceAddress(s, 6)));
+	auto valid = Binary(s, spv::OpIEqual, b, shr(words[3], 28), ConstantU32(s, 8));
+	const auto require = [&](uint32_t condition) {
+		valid = Binary(s, spv::OpLogicalAnd, b, valid, condition);
+	};
+	require(Binary(s, spv::OpULessThanEqual, b, kind, ConstantU32(s, 5)));
+	require(Binary(s, spv::OpULessThanEqual, b, last_index, DeviceAddressFromWords(s, words[2], and_bits(words[3], 0x3ff))));
+	require(Binary(s, spv::OpULessThan, b, last_address, ConstantDeviceAddress(s, uint64_t{1} << 48)));
+	require(Unary(s, spv::OpLogicalNot, b, Unary(s, spv::OpIsNan, b, extent)));
+	for (uint32_t i = 0; i < 3; ++i) {
+		for (const auto value: {origin[i], direction[i], inverse[i]})
+			require(Unary(s, spv::OpLogicalNot, b, Unary(s, spv::OpIsNan, b, value)));
+		require(Unary(s, spv::OpLogicalNot, b, Unary(s, spv::OpIsInf, b, origin[i])));
+	}
+	const std::array triangle_invalid {ConstantU32(s, 0x7f800000), ConstantU32(s, 0x3f800000),
+	                                  ConstantU32(s, ~0u), ConstantU32(s, 0)};
+	const std::array box_invalid {ConstantU32(s, ~0u), ConstantU32(s, ~0u),
+	                             ConstantU32(s, ~0u), ConstantU32(s, ~0u)};
+	std::array<uint32_t, 4> invalid_words;
+	for (uint32_t i = 0; i < 4; ++i)
+		invalid_words[i] = Select(s, u, triangle, triangle_invalid[i], box_invalid[i]);
+	const auto invalid = Vector(s, vec4, invalid_words);
+	const auto result = EmitValueOrDefaultIfCondition(s, valid, vec4, invalid, [&] {
+		const auto first = GetBdaPointer(s, address);
+		const auto second = EmitValueOrDefaultIfCondition(s, full, wide, ConstantDeviceAddress(s, 0), [&] {
+			return GetBdaPointer(s, Binary(s, spv::OpIAdd, wide, address, ConstantDeviceAddress(s, 64)));
+		});
+		const auto present = Binary(s, spv::OpLogicalAnd, b,
+		    Binary(s, spv::OpINotEqual, b, first, ConstantDeviceAddress(s, 0)),
+		    Binary(s, spv::OpLogicalOr, b, Unary(s, spv::OpLogicalNot, b, full),
+		        Binary(s, spv::OpINotEqual, b, second, ConstantDeviceAddress(s, 0))));
+		return EmitValueOrDefaultIfCondition(s, present, vec4, invalid, [&] {
+			const auto tri_label = s.builder.AllocateId(), box_label = s.builder.AllocateId();
+			const auto merge = s.builder.AllocateId();
+			s.builder.AddFunction(spv::OpSelectionMerge, merge, spv::SelectionControlMaskNone);
+			s.builder.AddFunction(spv::OpBranchConditional, triangle, tri_label, box_label);
+			EmitLabel(s, tri_label);
+			const auto tri_value = Triangle(s, first, kind, bary, origin, direction);
+			const auto tri_exit = s.current_label;
+			s.builder.AddFunction(spv::OpBranch, merge);
+			EmitLabel(s, box_label);
+			const auto box_value = Boxes(s, first, second, half, grow, sort, extent, origin, inverse);
+			const auto box_exit = s.current_label;
+			s.builder.AddFunction(spv::OpBranch, merge);
+			EmitLabel(s, merge);
+			const auto value = s.builder.AllocateId();
+			s.builder.AddFunction(spv::OpPhi, vec4, value, tri_value, tri_exit, box_value, box_exit);
+			return value;
+		});
+	});
+	s.builder.AddFunction(spv::OpReturnValue, result);
+	s.builder.AddFunction(spv::OpFunctionEnd);
+}
+
+uint32_t EmitBvhIntersect(ValueEmitContext& ctx, const IR::Inst& inst) {
+	auto& s = ctx.state;
+	const auto* ray = ctx.ImageAddress(inst.Arg(1));
+	if (ray == nullptr || ray->NumArgs() < 11 || s.bvh_intersect_function == 0)
+		ctx.Fail(inst, "BVH intersection requires eleven address components and its shared function");
+	const auto vector = [&](uint32_t first) {
+		Vec3 values;
+		for (uint32_t i = 0; i < 3; ++i)
+			values[i] = Unary(s, spv::OpBitcast, TypeF32(s), ctx.Arg(*ray, first + i));
+		return Vector(s, TypeF32Vector(s, 3), values);
+	};
+	const auto descriptor = ctx.Arg(inst, 0), node = ctx.Arg(*ray, 0);
+	const auto extent = Unary(s, spv::OpBitcast, TypeF32(s), ctx.Arg(*ray, 1));
+	const auto origin = vector(2), direction = vector(5), inverse = vector(8);
+	return EmitValueOrDefaultIfCondition(s, ctx.Arg(inst, 2), TypeU32Vector(s, 4),
+	    ConstantU32CompositeZero(s, 4), [&] {
+		const auto result = s.builder.AllocateId();
+		s.builder.AddFunction(spv::OpFunctionCall, TypeU32Vector(s, 4), result,
+		    s.bvh_intersect_function, descriptor, node, extent, origin, direction, inverse);
+		return result;
+	});
+}
+
+} // namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter

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

@@ -452,6 +452,11 @@ void EmitSetAttribute(ValueEmitContext& ctx, const IR::Inst& inst) {
 	if (exp.kind == IR::ExportTargetKind::Null || exp.en == 0u) {
 		return;
 	}
+	// Skip dormant color exports after their valid mask; MRT1 is reserved for logical alpha.
+	if (state.program.stage == ShaderType::Pixel && exp.kind == IR::ExportTargetKind::Mrt &&
+	    exp.index != 0 && state.input_info.pixel->alpha_blend_source_remap) {
+		return;
+	}
 	EmitIfCondition(state, exec, [&]() {
 		const auto data = ctx.Arg(inst, 0);
 		if (exp.kind == IR::ExportTargetKind::Primitive) {
@@ -492,6 +497,18 @@ void EmitSetAttribute(ValueEmitContext& ctx, const IR::Inst& inst) {
 		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);
+		if (state.program.stage == ShaderType::Pixel && exp.kind == IR::ExportTargetKind::Mrt &&
+		    exp.index == 0 && !uint_output && state.input_info.pixel->alpha_blend_source_remap) {
+			// Broadcast logical alpha before swizzling the primary output.
+			const auto blend_output =
+			    OutputVariableForExport(state, {.kind = IR::ExportTargetKind::Mrt, .index = 1});
+			if (blend_output != 0) {
+				const auto alpha = state.builder.AllocateId();
+				state.builder.AddFunction(spv::OpVectorShuffle, vector_type, alpha, value, value,
+				                          3u, 3u, 3u, 3u);
+				state.builder.AddFunction(spv::OpStore, blend_output, alpha);
+			}
+		}
 		if (state.program.stage == ShaderType::Pixel && exp.kind == IR::ExportTargetKind::Mrt &&
 		    exp.index < state.input_info.pixel->target_export_mapping.size()) {
 			const auto mapping = state.input_info.pixel->target_export_mapping[exp.index];
@@ -636,6 +653,29 @@ uint32_t EmitDppUpdateU32(ValueEmitContext& ctx, const IR::Inst& inst) {
 	                                                ctx.Arg(inst, 1));
 }
 
+uint32_t EmitConditionRef(ValueEmitContext& ctx, const IR::Inst& inst) {
+	if (ctx.other_half == nullptr) return ctx.Arg(inst, 0);
+	// A native scalar branch makes one decision for both emulated wave halves.
+	if (ctx.half != 0) return ctx.other_half->Def(IR::Value(&inst));
+	const auto kind = inst.Flags<CFG::BranchCondition>();
+	if (kind == CFG::BranchCondition::ScalarInstruction) return ctx.Arg(inst, 0);
+	const auto ballot = ctx.Ballot(inst.Arg(0));
+	const auto low = ctx.state.builder.AllocateId();
+	const auto high = ctx.state.builder.AllocateId();
+	const auto combined = ctx.state.builder.AllocateId();
+	const auto result = ctx.state.builder.AllocateId();
+	ctx.state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(ctx.state), low, ballot, 0);
+	ctx.state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(ctx.state), high, ballot, 1);
+	const bool zero = kind == CFG::BranchCondition::ExecZero ||
+	                  kind == CFG::BranchCondition::VccZero || kind == CFG::BranchCondition::SccZero;
+	ctx.state.builder.AddFunction(zero ? spv::OpBitwiseAnd : spv::OpBitwiseOr,
+	                              TypeU32(ctx.state), combined, low, high);
+	ctx.state.builder.AddFunction(zero ? spv::OpIEqual : spv::OpINotEqual,
+	                              TypeBool(ctx.state), result, combined,
+	                              ConstantU32(ctx.state, zero ? ~0u : 0u));
+	return result;
+}
+
 uint32_t EmitBallot(ValueEmitContext& ctx, IR::Value predicate) {
 	return ctx.Ballot(predicate);
 }

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

@@ -539,19 +539,6 @@ uint32_t StoreTexel(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uint32_t d
 	return PackImageTexel(ctx, mem, texel);
 }
 
-spv::Op ImageAtomicOpcode(IR::ValueOpcode opcode) {
-	switch (opcode) {
-		case IR::ValueOpcode::ImageAtomicSwap32: return spv::OpAtomicExchange;
-		case IR::ValueOpcode::ImageAtomicIAdd32: return spv::OpAtomicIAdd;
-		case IR::ValueOpcode::ImageAtomicUMin32: return spv::OpAtomicUMin;
-		case IR::ValueOpcode::ImageAtomicUMax32: return spv::OpAtomicUMax;
-		case IR::ValueOpcode::ImageAtomicAnd32: return spv::OpAtomicAnd;
-		case IR::ValueOpcode::ImageAtomicOr32: return spv::OpAtomicOr;
-		case IR::ValueOpcode::ImageAtomicXor32: return spv::OpAtomicXor;
-		default: return spv::OpNop;
-	}
-}
-
 } // namespace
 
 void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
@@ -876,10 +863,9 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 		ctx.Define(inst, ResultVector(ctx, result, numeric_class, dref, mem));
 		return;
 	}
-	const auto atomic_opcode = ImageAtomicOpcode(op);
-	if (atomic_opcode != spv::OpNop) {
+	if (image_info.access == IR::ImageAccess::Atomic) {
 		const auto dimension = image.dimension;
-		ctx.Define(inst, EmitValueOrZeroIfCondition(state, ctx.Arg(inst, 3), [&]() {
+		ctx.Define(inst, EmitValueOrZeroIfCondition(state, ctx.Arg(inst, inst.NumArgs() - 1), [&]() {
 			           const auto pointer      = state.builder.AllocateId();
 			           const auto pointer_type = state.builder.Type(
 			               spv::OpTypePointer, spv::StorageClassImage, TypeU32(state));
@@ -887,12 +873,17 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 			                                     StorageImageDescriptorPointer(state, mem.resource),
 			                                     CoordU32(ctx, mem, *address, dimension),
 			                                     ConstantU32(state, 0));
-			           const auto old = state.builder.AllocateId();
-			           state.builder.AddFunction(atomic_opcode, TypeU32(state), old, pointer,
-			                                     ConstantU32(state, spv::ScopeDevice),
-			                                     ConstantU32(state, spv::MemorySemanticsMaskNone),
-			                                     ctx.Arg(inst, 2));
-			           EmitDeviceAtomicMemoryBarrier(state);
+			           if (op == IR::ValueOpcode::ImageAtomicFMin32 ||
+			               op == IR::ValueOpcode::ImageAtomicFMax32) {
+				           return AtomicUpdate(state, pointer, IR::ResourceKind::Image,
+				                               [&](uint32_t old) {
+					                               return EmitFloatAtomicReplacement(
+					                                   state, old, ctx.Arg(inst, 2),
+					                                   op == IR::ValueOpcode::ImageAtomicFMax32);
+				                               });
+			           }
+			           const auto old = EmitAtomicOperation(ctx, inst, pointer, spv::ScopeDevice);
+			           EmitAtomicMemoryBarrier(state, IR::ResourceKind::Image);
 			           return old;
 		           }));
 		return;

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

@@ -2,6 +2,8 @@
 
 #include "graphics/shader/recompiler/backend/spirv/spirvEmitterInternal.h"
 
+#include <bit>
+
 namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter {
 // Operand signatures distinguish SPIR-V IDs (uint32_t), raw literals/values
 // (IR::Value), and instructions whose metadata or operand loads must stay lazy.
@@ -13,14 +15,25 @@ inline constexpr auto EmitConvertU32U16 = EmitBitCastU16F16;
 inline constexpr auto EmitConvertU32U8  = EmitBitCastU16F16;
 inline constexpr auto EmitBitCastU32F32 = EmitBitCastU16F16;
 EMIT_NATIVE(BitCastF32U32, OpBitcast, F32, uint32_t)
+EMIT_NATIVE(BitCastU64F64, OpBitcast, U64, uint32_t)
+EMIT_NATIVE(BitCastF64U64, OpBitcast, F64, uint32_t)
 uint32_t              EmitConvertU16U32(EmitterState& state, uint32_t arg0);
 uint32_t              EmitConvertU8U32(EmitterState& state, uint32_t arg0);
 uint32_t              EmitConvertF16F32(EmitterState& state, uint32_t arg0);
 inline constexpr auto EmitConvertF32F16 = EmitF16BitsToF32;
 uint32_t              EmitConvertS32F32(EmitterState& state, uint32_t arg0);
 uint32_t              EmitConvertU32F32(EmitterState& state, uint32_t arg0);
-uint32_t              EmitConvertF32S32(EmitterState& state, uint32_t arg0);
+template <IR::Type type>
+uint32_t EmitConvertSigned32ToFloat(EmitterState& state, uint32_t arg0) {
+	const auto signed_value = Unary(state, spv::OpBitcast, TypeI32(state), arg0);
+	return EmitNative<spv::OpConvertSToF, type>(state, signed_value);
+}
+inline constexpr auto EmitConvertF32S32 = EmitConvertSigned32ToFloat<IR::Type::F32>;
+inline constexpr auto EmitConvertF64S32 = EmitConvertSigned32ToFloat<IR::Type::F64>;
+uint32_t              EmitConvertF32F64(EmitterState& state, uint32_t arg0);
+uint32_t              EmitConvertF64F32(EmitterState& state, uint32_t arg0);
 EMIT_NATIVE(ConvertF32U32, OpConvertUToF, F32, uint32_t)
+EMIT_NATIVE(ConvertF64U32, OpConvertUToF, F64, uint32_t)
 EMIT_NATIVE(CompositeConstructU64, OpCompositeConstruct, U64, uint32_t, uint32_t)
 EMIT_NATIVE(CompositeConstructU32x2, OpCompositeConstruct, U32x2, uint32_t, uint32_t)
 EMIT_NATIVE(CompositeConstructU32x3, OpCompositeConstruct, U32x3, uint32_t, uint32_t, uint32_t)
@@ -98,28 +111,55 @@ uint32_t EmitIEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1);
 uint32_t EmitINotEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1);
 uint32_t EmitULessThan64(EmitterState& state, uint32_t arg0, uint32_t arg1);
 uint32_t EmitSLessThan64(EmitterState& state, uint32_t arg0, uint32_t arg1);
+uint32_t EmitSLessThanEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1);
+uint32_t EmitULessThanEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1);
+uint32_t EmitUGreaterThanEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1);
 uint32_t EmitUGreaterThan64(EmitterState& state, uint32_t arg0, uint32_t arg1);
 EMIT_NATIVE(LogicalOr, OpLogicalOr, U1, uint32_t, uint32_t)
 EMIT_NATIVE(LogicalAnd, OpLogicalAnd, U1, uint32_t, uint32_t)
 EMIT_NATIVE(LogicalXor, OpLogicalNotEqual, U1, uint32_t, uint32_t)
 EMIT_NATIVE(LogicalNot, OpLogicalNot, U1, uint32_t)
-EMIT_NATIVE(FPOrdEqual32, OpFOrdEqual, U1, uint32_t, uint32_t)
-EMIT_NATIVE(FPUnordEqual32, OpFUnordEqual, U1, uint32_t, uint32_t)
-EMIT_NATIVE(FPOrdNotEqual32, OpFOrdNotEqual, U1, uint32_t, uint32_t)
-EMIT_NATIVE(FPUnordNotEqual32, OpFUnordNotEqual, U1, uint32_t, uint32_t)
-EMIT_NATIVE(FPOrdLessThan32, OpFOrdLessThan, U1, uint32_t, uint32_t)
-EMIT_NATIVE(FPUnordLessThan32, OpFUnordLessThan, U1, uint32_t, uint32_t)
-EMIT_NATIVE(FPOrdGreaterThan32, OpFOrdGreaterThan, U1, uint32_t, uint32_t)
-EMIT_NATIVE(FPUnordGreaterThan32, OpFUnordGreaterThan, U1, uint32_t, uint32_t)
-EMIT_NATIVE(FPOrdLessThanEqual32, OpFOrdLessThanEqual, U1, uint32_t, uint32_t)
-EMIT_NATIVE(FPUnordLessThanEqual32, OpFUnordLessThanEqual, U1, uint32_t, uint32_t)
-EMIT_NATIVE(FPOrdGreaterThanEqual32, OpFOrdGreaterThanEqual, U1, uint32_t, uint32_t)
-EMIT_NATIVE(FPUnordGreaterThanEqual32, OpFUnordGreaterThanEqual, U1, uint32_t, uint32_t)
+template <spv::Op opcode>
+uint32_t EmitFloatCompare32(ValueEmitContext& ctx, const IR::Inst& inst) {
+	// SPIR-V's DenormFlushToZero mode does not require comparison operands to flush.
+	const bool flush = inst.Flags<IR::FPCompareFlags>().flush_input_denorms;
+	const auto operand = [&](size_t index) {
+		const auto value = inst.Arg(index);
+		if (flush && value.IsImmediate()) {
+			const auto bits = std::bit_cast<uint32_t>(value.F32Value());
+			return ConstantF32(ctx.state, (bits & 0x7fffffffu) < 0x00800000u
+			                                  ? bits & 0x80000000u : bits);
+		}
+		const auto id = ctx.Arg(inst, index);
+		return flush ? EmitFlushF32DenormToSignedZero(ctx.state, id) : id;
+	};
+	return EmitNative<opcode, IR::Type::U1>(ctx.state, operand(0), operand(1));
+}
+inline constexpr auto EmitFPOrdEqual32 = EmitFloatCompare32<spv::OpFOrdEqual>;
+EMIT_NATIVE(FPOrdEqual64, OpFOrdEqual, U1, uint32_t, uint32_t)
+EMIT_NATIVE(FPOrdLessThanEqual64, OpFOrdLessThanEqual, U1, uint32_t, uint32_t)
+EMIT_NATIVE(FPOrdGreaterThanEqual64, OpFOrdGreaterThanEqual, U1, uint32_t, uint32_t)
+inline constexpr auto EmitFPUnordEqual32 = EmitFloatCompare32<spv::OpFUnordEqual>;
+inline constexpr auto EmitFPOrdNotEqual32 = EmitFloatCompare32<spv::OpFOrdNotEqual>;
+inline constexpr auto EmitFPUnordNotEqual32 = EmitFloatCompare32<spv::OpFUnordNotEqual>;
+inline constexpr auto EmitFPOrdLessThan32 = EmitFloatCompare32<spv::OpFOrdLessThan>;
+inline constexpr auto EmitFPUnordLessThan32 = EmitFloatCompare32<spv::OpFUnordLessThan>;
+inline constexpr auto EmitFPOrdGreaterThan32 = EmitFloatCompare32<spv::OpFOrdGreaterThan>;
+inline constexpr auto EmitFPUnordGreaterThan32 = EmitFloatCompare32<spv::OpFUnordGreaterThan>;
+inline constexpr auto EmitFPOrdLessThanEqual32 = EmitFloatCompare32<spv::OpFOrdLessThanEqual>;
+inline constexpr auto EmitFPUnordLessThanEqual32 = EmitFloatCompare32<spv::OpFUnordLessThanEqual>;
+inline constexpr auto EmitFPOrdGreaterThanEqual32 = EmitFloatCompare32<spv::OpFOrdGreaterThanEqual>;
+inline constexpr auto EmitFPUnordGreaterThanEqual32 = EmitFloatCompare32<spv::OpFUnordGreaterThanEqual>;
 uint32_t              EmitFPIsNan32(EmitterState& state, uint32_t arg0);
 inline constexpr auto EmitFPCmpClass32 = EmitClassMaskF32;
 EMIT_NATIVE(FPAdd32, OpFAdd, F32, uint32_t, uint32_t)
 EMIT_NATIVE(FPSub32, OpFSub, F32, uint32_t, uint32_t)
 EMIT_NATIVE(FPMul32, OpFMul, F32, uint32_t, uint32_t)
+EMIT_NATIVE(FPAdd64, OpFAdd, F64, uint32_t, uint32_t)
+EMIT_NATIVE(FPMul64, OpFMul, F64, uint32_t, uint32_t)
+inline constexpr auto EmitFPFma64 =
+    EmitGlsl<GLSLstd450Fma, IR::Type::F64, uint32_t, uint32_t, uint32_t>;
+uint32_t              EmitFPRecip64(EmitterState& state, uint32_t arg0);
 inline constexpr auto EmitFPFma32 =
     EmitGlsl<GLSLstd450Fma, IR::Type::F32, uint32_t, uint32_t, uint32_t>;
 uint32_t EmitFPMin32(EmitterState& state, uint32_t arg0, uint32_t arg1);
@@ -169,6 +209,7 @@ 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              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              EmitWriteLane(ValueEmitContext& ctx, const IR::Inst& inst);
@@ -188,6 +229,7 @@ void                  EmitLoadMemory(ValueEmitContext& ctx, const IR::Inst& inst
 void                  EmitStoreMemory(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitAtomic32(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitBufferAtomic64(ValueEmitContext& ctx, const IR::Inst& inst);
+void                  EmitSharedAtomic64(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitBufferFloatAtomic(ValueEmitContext& ctx, const IR::Inst& inst);
 void                  EmitSharedFloatAtomic(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitSharedIncDec(ValueEmitContext& ctx, const IR::Inst& inst);
@@ -223,6 +265,7 @@ inline constexpr auto EmitBufferAtomicUMin32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicSMax32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicUMax32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicAnd32     = EmitAtomic32;
+inline constexpr auto EmitBufferAtomicAnd64     = EmitBufferAtomic64;
 inline constexpr auto EmitBufferAtomicOr32      = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicOr64      = EmitBufferAtomic64;
 inline constexpr auto EmitBufferAtomicXor32     = EmitAtomic32;
@@ -244,6 +287,7 @@ inline constexpr auto EmitSharedAtomicFMin32    = EmitSharedFloatAtomic;
 inline constexpr auto EmitSharedAtomicFMax32    = EmitSharedFloatAtomic;
 inline constexpr auto EmitSharedAtomicSwap32    = EmitAtomic32;
 inline constexpr auto EmitSharedAtomicIAdd32    = EmitAtomic32;
+inline constexpr auto EmitSharedAtomicIAdd64    = EmitSharedAtomic64;
 inline constexpr auto EmitSharedAtomicISub32    = EmitAtomic32;
 inline constexpr auto EmitSharedAtomicInc32     = EmitSharedIncDec;
 inline constexpr auto EmitSharedAtomicDec32     = EmitSharedIncDec;
@@ -253,24 +297,31 @@ inline constexpr auto EmitSharedAtomicSMax32    = EmitAtomic32;
 inline constexpr auto EmitSharedAtomicUMax32    = EmitAtomic32;
 inline constexpr auto EmitSharedAtomicAnd32     = EmitAtomic32;
 inline constexpr auto EmitSharedAtomicOr32      = EmitAtomic32;
+inline constexpr auto EmitSharedAtomicOr64      = EmitSharedAtomic64;
 inline constexpr auto EmitSharedAtomicXor32     = EmitAtomic32;
 inline constexpr auto EmitDataAppend            = EmitAppendConsume;
 inline constexpr auto EmitDataConsume           = EmitAppendConsume;
 uint32_t              EmitSwizzleU32(ValueEmitContext& ctx, const IR::Inst& inst);
 void                  EmitImage(ValueEmitContext& ctx, const IR::Inst& inst);
+uint32_t              EmitBvhIntersect(ValueEmitContext& ctx, const IR::Inst& inst);
 inline constexpr auto EmitImageQueryDimensions = EmitImage;
 inline constexpr auto EmitImageQueryLod        = EmitImage;
 inline constexpr auto EmitImageRead            = EmitImage;
 inline constexpr auto EmitImageWrite           = EmitImage;
 inline constexpr auto EmitImageSampleRaw       = EmitImage;
 inline constexpr auto EmitImageGatherRaw       = EmitImage;
+inline constexpr auto EmitImageAtomicCompareSwap32 = EmitImage;
 inline constexpr auto EmitImageAtomicSwap32    = EmitImage;
 inline constexpr auto EmitImageAtomicIAdd32    = EmitImage;
+inline constexpr auto EmitImageAtomicSMin32    = EmitImage;
 inline constexpr auto EmitImageAtomicUMin32    = EmitImage;
+inline constexpr auto EmitImageAtomicSMax32    = EmitImage;
 inline constexpr auto EmitImageAtomicUMax32    = EmitImage;
 inline constexpr auto EmitImageAtomicAnd32     = EmitImage;
 inline constexpr auto EmitImageAtomicOr32      = EmitImage;
 inline constexpr auto EmitImageAtomicXor32     = EmitImage;
+inline constexpr auto EmitImageAtomicFMin32    = EmitImage;
+inline constexpr auto EmitImageAtomicFMax32    = EmitImage;
 void                  EmitUnreachable(ValueEmitContext& ctx, const IR::Inst& inst);
 inline constexpr auto EmitPhi                        = EmitUnreachable;
 inline constexpr auto EmitTessellationBase           = EmitUnreachable;

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

@@ -58,6 +58,7 @@ constexpr std::array<ImageDimensionInfo, 7> ImageDimensions {{
 const ImageDimensionInfo& ImageDimensionInfoFor(ImageDimension dimension);
 
 struct SpirvRequirements {
+	bool bvh                          = false;
 	bool subgroup_ballot              = false;
 	bool subgroup_shuffle             = false;
 	bool subgroup_local_invocation_id = false;
@@ -67,7 +68,9 @@ struct SpirvRequirements {
 	bool function_scratch             = false;
 	bool pixel_valid_mask             = false;
 	bool buffer_int64_atomics         = false;
+	bool shared_int64_atomics         = false;
 	bool coherent_buffers             = false;
+	bool float64                      = false;
 };
 
 SpirvRequirements AnalyzeProgramRequirements(const IR::Program& program);
@@ -81,6 +84,20 @@ struct EmitterState {
 	Builder                                          builder;
 	const IR::Program&                               program;
 	ShaderStageInputInfo                             input_info;
+	uint32_t                                        void_type = 0;
+	uint32_t                                        bool_type = 0;
+	uint32_t                                        u32_type = 0;
+	uint32_t                                        native_u64_type = 0;
+	uint32_t                                        i32_type = 0;
+	uint32_t                                        f32_type = 0;
+	uint32_t                                        f64_type = 0;
+	uint32_t                                        u32_pair_type = 0;
+	uint32_t                                        i32_pair_type = 0;
+	uint32_t                                        function_type = 0;
+	std::array<uint32_t, 3>                          bool_vector_types {};
+	std::array<uint32_t, 3>                          u32_vector_types {};
+	std::array<uint32_t, 3>                          i32_vector_types {};
+	std::array<uint32_t, 3>                          f32_vector_types {};
 	std::array<uint32_t, 6>                          tess_variables {};
 	uint32_t                                         tess_inner_variable = 0;
 	uint32_t                                         tess_patch_base     = 0;
@@ -94,12 +111,14 @@ struct EmitterState {
 	uint32_t                                         bda_pagetable_variable  = 0;
 	uint32_t                                         fault_buffer_variable   = 0;
 	uint32_t                                         bda_pointer_function    = 0;
+	uint32_t                                         bvh_intersect_function  = 0;
 	uint32_t                                         gds_variable            = 0;
 	uint32_t                                         gds_length              = 0;
 	uint32_t                                         push_constant_variable  = 0;
 	uint32_t                                         shader_data_storage_variable = 0;
 	uint32_t                                         flattened_srt_variable  = 0;
 	uint32_t                                         lds_variable            = 0;
+	uint32_t                                         lds_u64_variable        = 0;
 	std::array<uint32_t, 2>                          scratch_variable {};
 	std::array<uint32_t, IR::ImageBindingCount>      image_variables {};
 	uint32_t                   sampler_variable                      = 0;
@@ -139,6 +158,7 @@ uint32_t TypeU32Pair(EmitterState& state);
 uint32_t TypeI32(EmitterState& state);
 uint32_t TypeI32Pair(EmitterState& state);
 uint32_t TypeF32(EmitterState& state);
+uint32_t TypeF64(EmitterState& state);
 uint32_t TypeU32Vector(EmitterState& state, uint32_t components);
 
 uint32_t TypeU32Composite(EmitterState& state, uint32_t components);
@@ -167,6 +187,10 @@ template <spv::Op opcode, IR::Type type, typename... Args>
 uint32_t EmitNative(EmitterState& state, Args... args) {
 	const auto result = state.builder.AllocateId();
 	state.builder.AddFunction(opcode, TypeId(state, type), result, args...);
+	if constexpr (type == IR::Type::F64 &&
+	              (opcode == spv::OpFMul || opcode == spv::OpFDiv || opcode == spv::OpExtInst)) {
+		state.builder.AddAnnotation(spv::OpDecorate, result, spv::DecorationNoContraction);
+	}
 	return result;
 }
 
@@ -370,6 +394,7 @@ Prospero::BufferFormat StorageBufferFormat(const EmitterState& state, const IR::
 void EmitMemoryOffsets(EmitterState& state);
 
 uint32_t LdsDwordCount(const EmitterState& state);
+void EnsureLdsStorage(EmitterState& state);
 
 struct MemoryResourceAccess {
 	IR::ResourceKind      kind             = IR::ResourceKind::None;
@@ -411,7 +436,12 @@ uint32_t EmitUFloatToF32Bits(EmitterState& state, uint32_t raw, uint32_t bits);
 uint32_t NormalizeFormatComponent(EmitterState& state, const Format::BufferFormatInfo& info,
                                   uint32_t component, uint32_t raw);
 
-void EmitDeviceAtomicMemoryBarrier(EmitterState& state);
+spv::Op SpirvAtomicOpcode(IR::ValueOpcode opcode);
+
+uint32_t EmitAtomicOperation(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t pointer,
+                             uint32_t scope);
+
+void EmitAtomicMemoryBarrier(EmitterState& state, IR::ResourceKind kind);
 
 uint32_t EmitFloatAtomicReplacement(EmitterState& state, uint32_t old, uint32_t source,
                                     bool max_value);
@@ -478,6 +508,10 @@ uint32_t EmitF16BitsToF32(EmitterState& state, uint32_t bits);
 void EmitProgram(EmitterState& state);
 
 void DefineGetBdaPointer(EmitterState& state);
+void DefineBvhIntersect(EmitterState& state);
+uint32_t GetBdaPointer(EmitterState& state, uint32_t address);
+uint32_t ConstantDeviceAddress(EmitterState& state, uint64_t value);
+uint32_t DeviceAddressFromWords(EmitterState& state, uint32_t low, uint32_t high);
 
 // These templates accept local lambdas from several emitter translation units.
 template <typename Fn>
@@ -519,14 +553,7 @@ uint32_t EmitValueOrZeroIfCondition(EmitterState& state, uint32_t condition, Fn&
 
 template <typename Fn>
 uint32_t AtomicUpdate(EmitterState& state, uint32_t pointer, IR::ResourceKind kind, Fn&& desired) {
-	const auto scope  = kind == IR::ResourceKind::Lds ? spv::ScopeWorkgroup : spv::ScopeDevice;
-	const auto memory = [&] {
-		switch (kind) {
-			case IR::ResourceKind::Lds: return spv::MemorySemanticsWorkgroupMemoryMask;
-			case IR::ResourceKind::Image: return spv::MemorySemanticsImageMemoryMask;
-			default: return spv::MemorySemanticsUniformMemoryMask;
-		}
-	}();
+	const auto scope     = kind == IR::ResourceKind::Lds ? spv::ScopeWorkgroup : spv::ScopeDevice;
 	const auto preheader = state.builder.AllocateId();
 	const auto header    = state.builder.AllocateId();
 	const auto cont      = state.builder.AllocateId();
@@ -555,8 +582,7 @@ uint32_t AtomicUpdate(EmitterState& state, uint32_t pointer, IR::ResourceKind ki
 	EmitLabel(state, cont);
 	state.builder.AddFunction(spv::OpBranch, header);
 	EmitLabel(state, merge);
-	state.builder.AddFunction(spv::OpMemoryBarrier, ConstantU32(state, scope),
-	                          ConstantU32(state, spv::MemorySemanticsAcquireReleaseMask | memory));
+	EmitAtomicMemoryBarrier(state, kind);
 	return observed;
 }
 

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

@@ -1,6 +1,7 @@
 #include "graphics/shader/recompiler/backend/spirv/spirvEmitterInstructions.h"
 
 #include "graphics/host_gpu/renderer/cache/bufferCache.h"
+#include "graphics/shader/recompiler/frontend/decode/ShaderDecoder.h"
 
 #include <algorithm>
 
@@ -105,31 +106,23 @@ uint32_t AddU64Low(EmitterState& state, uint32_t low, uint32_t high, uint32_t ad
 	return result;
 }
 
-uint32_t ScratchByteAddress(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uint32_t low,
-                            uint32_t high) {
-	auto& state     = ctx.state;
-	auto  immediate = static_cast<int32_t>(mem.offset);
+uint32_t AddAddressOffset(EmitterState& state, const IR::MemoryInfo& mem, uint32_t low,
+                          uint32_t& high) {
+	const auto immediate = static_cast<int32_t>(mem.offset);
+	if (immediate == 0) return low;
 	const auto immediate_low  = ConstantU32(state, static_cast<uint32_t>(immediate));
 	const auto immediate_high = ConstantU32(state, immediate < 0 ? UINT32_MAX : 0u);
-	low                      = AddU64Low(state, low, high, immediate_low, immediate_high, high);
+	return AddU64Low(state, low, high, immediate_low, immediate_high, high);
+}
+
+uint32_t ScratchByteAddress(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uint32_t low,
+                            uint32_t high) {
+	auto& state = ctx.state;
+	low = AddAddressOffset(state, mem, low, high);
 	const auto valid = Binary(state, spv::OpIEqual, TypeBool(state), high, ConstantU32(state, 0));
 	return Select(state, TypeU32(state), valid, low, ConstantU32(state, UINT32_MAX));
 }
 
-uint32_t ConstantDeviceAddress(EmitterState& state, uint64_t value) {
-	return state.builder.Constant(spv::OpConstant, TypeScalarU64(state),
-	                              static_cast<uint32_t>(value),
-	                              static_cast<uint32_t>(value >> 32u));
-}
-
-uint32_t DeviceAddressFromWords(EmitterState& state, uint32_t low, uint32_t high) {
-	const auto low64  = Unary(state, spv::OpUConvert, TypeScalarU64(state), low);
-	const auto high64 = Binary(state, spv::OpShiftLeftLogical, TypeScalarU64(state),
-	                           Unary(state, spv::OpUConvert, TypeScalarU64(state), high),
-	                           ConstantDeviceAddress(state, 32));
-	return Binary(state, spv::OpBitwiseOr, TypeScalarU64(state), low64, high64);
-}
-
 uint32_t GuestAddress(ValueEmitContext& ctx, const IR::Inst& inst, const IR::MemoryInfo& mem) {
 	auto& state = ctx.state;
 	auto  low   = ctx.Arg(inst, 1);
@@ -146,7 +139,12 @@ uint32_t GuestAddress(ValueEmitContext& ctx, const IR::Inst& inst, const IR::Mem
 			ctx.Fail(inst, "has no address base pair");
 			return ConstantDeviceAddress(state, 0);
 		}
-		const auto base = DeviceAddressFromWords(state, ctx.Arg(*handle, 0), ctx.Arg(*handle, 1));
+		auto base_low = ctx.Arg(*handle, 0);
+		if (mem.kind == IR::ResourceKind::ScalarAddress) {
+			base_low = Binary(state, spv::OpBitwiseAnd, TypeU32(state), base_low,
+			                  ConstantU32(state, ~3u));
+		}
+		const auto base = DeviceAddressFromWords(state, base_low, ctx.Arg(*handle, 1));
 		address         = Binary(state, spv::OpIAdd, TypeScalarU64(state), base,
 		                         Unary(state, spv::OpUConvert, TypeScalarU64(state), low));
 	}
@@ -181,17 +179,9 @@ void RecordBdaFault(EmitterState& state, uint32_t page) {
 	                          Binary(state, spv::OpBitwiseOr, TypeU32(state), value, bit));
 }
 
-uint32_t GetBdaPointer(ValueEmitContext& ctx, uint32_t address) {
-	auto&      state  = ctx.state;
-	const auto result = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpFunctionCall, TypeScalarU64(state), result,
-	                          state.bda_pointer_function, address);
-	return result;
-}
-
 uint32_t LoadBdaDword(ValueEmitContext& ctx, uint32_t address) {
 	auto&      state   = ctx.state;
-	const auto bda     = GetBdaPointer(ctx, address);
+	const auto bda     = GetBdaPointer(state, address);
 	const auto present =
 	    Binary(state, spv::OpINotEqual, TypeBool(state), bda, ConstantDeviceAddress(state, 0));
 	return EmitValueOrZeroIfCondition(state, present, [&]() {
@@ -506,6 +496,50 @@ void StoreWordInBounds(ValueEmitContext& ctx, const MemoryResourceAccess& resour
 	                              resource.memory_access);
 }
 
+template <typename Fn>
+void ForEachLocalFlatAccess(ValueEmitContext& ctx, const IR::Inst& inst, Fn&& emit) {
+	auto& state = ctx.state;
+	auto mem = ctx.Memory(inst);
+	auto high = ctx.Arg(inst, 2);
+	const auto low = AddAddressOffset(state, mem, ctx.Arg(inst, 1), high);
+	const auto index = Binary(state, spv::OpShiftRightLogical, TypeU32(state), low, ConstantU32(state, 2));
+	const auto aligned = Binary(state, spv::OpIEqual, TypeBool(state), ConstantU32(state, 0),
+	    Binary(state, spv::OpBitwiseAnd, TypeU32(state), low, ConstantU32(state, 3)));
+	for (const auto kind: {IR::ResourceKind::Scratch, IR::ResourceKind::Lds}) {
+		mem.kind = kind;
+		const auto resource = PrepareMemoryResourceAccess(state, mem);
+		const auto aperture = kind == IR::ResourceKind::Scratch ? Decoder::PrivateApertureHigh
+		                                                       : Decoder::SharedApertureHigh;
+		const auto selected = Binary(state, spv::OpIEqual, TypeBool(state), high, ConstantU32(state, aperture));
+		const auto valid = AndCondition(state, aligned,
+		    AndCondition(state, selected, EmitMemoryElementInBounds(state, resource, index)));
+		emit(resource, index, valid);
+	}
+}
+
+uint32_t LoadLocalFlat(ValueEmitContext& ctx, const IR::Inst& inst) {
+	return EmitValueOrZeroIfCondition(ctx.state, ctx.Arg(inst, inst.NumArgs() - 1), [&]() {
+		uint32_t result = 0;
+		ForEachLocalFlatAccess(ctx, inst, [&](const auto& resource, uint32_t index, uint32_t valid) {
+			const auto value = EmitValueOrZeroIfCondition(ctx.state, valid, [&]() {
+				return LoadWordInBounds(ctx, resource, index);
+			});
+			result = result == 0 ? value : Binary(ctx.state, spv::OpBitwiseOr, TypeU32(ctx.state), result, value);
+		});
+		return result;
+	});
+}
+
+void StoreLocalFlat(ValueEmitContext& ctx, const IR::Inst& inst) {
+	EmitIfCondition(ctx.state, ctx.Arg(inst, inst.NumArgs() - 1), [&]() {
+		ForEachLocalFlatAccess(ctx, inst, [&](const auto& resource, uint32_t index, uint32_t valid) {
+			EmitIfCondition(ctx.state, valid, [&]() {
+				StoreWordInBounds(ctx, resource, index, ctx.Arg(inst, inst.NumArgs() - 2));
+			});
+		});
+	});
+}
+
 void StoreWordPrepared(ValueEmitContext& ctx, const IR::Inst& inst, const IR::MemoryInfo& mem,
                        const MemoryResourceAccess& resource, uint32_t data) {
 	const auto index = EmitMemoryElementIndex(ctx.state, resource, DwordIndex(ctx, inst, mem));
@@ -521,54 +555,6 @@ void StoreWord(ValueEmitContext& ctx, const IR::Inst& inst, IR::MemoryInfo mem)
 	});
 }
 
-spv::Op SpirvAtomicOpcode(IR::ValueOpcode opcode) {
-	switch (opcode) {
-		case IR::ValueOpcode::BufferAtomicCmpSwap32: return spv::OpAtomicCompareExchange;
-		case IR::ValueOpcode::BufferAtomicSwap32:
-		case IR::ValueOpcode::BufferAtomicSwap64:
-		case IR::ValueOpcode::SharedAtomicSwap32: return spv::OpAtomicExchange;
-		case IR::ValueOpcode::BufferAtomicIAdd32:
-		case IR::ValueOpcode::SharedAtomicIAdd32: return spv::OpAtomicIAdd;
-		case IR::ValueOpcode::BufferAtomicISub32:
-		case IR::ValueOpcode::SharedAtomicISub32: return spv::OpAtomicISub;
-		case IR::ValueOpcode::BufferAtomicSMin32:
-		case IR::ValueOpcode::SharedAtomicSMin32: return spv::OpAtomicSMin;
-		case IR::ValueOpcode::BufferAtomicUMin32:
-		case IR::ValueOpcode::SharedAtomicUMin32: return spv::OpAtomicUMin;
-		case IR::ValueOpcode::BufferAtomicSMax32:
-		case IR::ValueOpcode::SharedAtomicSMax32: return spv::OpAtomicSMax;
-		case IR::ValueOpcode::BufferAtomicUMax32:
-		case IR::ValueOpcode::SharedAtomicUMax32: return spv::OpAtomicUMax;
-		case IR::ValueOpcode::BufferAtomicAnd32:
-		case IR::ValueOpcode::SharedAtomicAnd32: return spv::OpAtomicAnd;
-		case IR::ValueOpcode::BufferAtomicOr32:
-		case IR::ValueOpcode::BufferAtomicOr64:
-		case IR::ValueOpcode::SharedAtomicOr32: return spv::OpAtomicOr;
-		case IR::ValueOpcode::BufferAtomicXor32:
-		case IR::ValueOpcode::SharedAtomicXor32: return spv::OpAtomicXor;
-		default: return spv::OpNop;
-	}
-}
-
-uint32_t EmitAtomicOperation(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t pointer,
-                             uint32_t scope) {
-	const auto old = ctx.state.builder.AllocateId();
-	if (inst.GetOpcode() == IR::ValueOpcode::BufferAtomicCmpSwap32) {
-		const auto desired    = ctx.Arg(inst, inst.NumArgs() - 3);
-		const auto comparator = ctx.Arg(inst, inst.NumArgs() - 2);
-		ctx.state.builder.AddFunction(
-		    spv::OpAtomicCompareExchange, TypeU32(ctx.state), old, pointer,
-		    ConstantU32(ctx.state, scope), ConstantU32(ctx.state, spv::MemorySemanticsMaskNone),
-		    ConstantU32(ctx.state, spv::MemorySemanticsMaskNone), desired, comparator);
-	} else {
-		const auto value = ctx.Arg(inst, inst.NumArgs() - 2);
-		ctx.state.builder.AddFunction(SpirvAtomicOpcode(inst.GetOpcode()), TypeU32(ctx.state), old,
-		                              pointer, ConstantU32(ctx.state, scope),
-		                              ConstantU32(ctx.state, spv::MemorySemanticsMaskNone), value);
-	}
-	return old;
-}
-
 template <typename Fn>
 uint32_t EmitAtomicAccess(ValueEmitContext& ctx, const IR::Inst& inst,
                           const IR::MemoryInfo& mem, Fn&& operation) {
@@ -706,26 +692,66 @@ uint32_t FormattedOutOfBoundsValue(ValueEmitContext& ctx, const IR::MemoryInfo&
 	return ConstructU32Composite(ctx.state, components, values);
 }
 
-void StoreFormattedInBounds(ValueEmitContext& ctx, const IR::MemoryInfo& mem,
-                            const PreparedFormattedMemory& plan, uint32_t component,
-                            uint32_t data) {
-	if (component >= plan.info.component_count) return;
-	const auto bits = plan.info.component_bits[component];
-	if (bits == 16u && (plan.info.type == Format::ComponentType::Snorm ||
-	                    plan.info.type == Format::ComponentType::Float)) {
+uint32_t EncodeFormattedStoreComponent(ValueEmitContext& ctx,
+                                       const Format::BufferFormatInfo& info,
+                                       uint32_t component, uint32_t data) {
+	const auto bits = info.component_bits[component];
+	if (info.type == Format::ComponentType::Unorm) {
+		const auto value = EmitBitCastF32U32(ctx.state, data);
+		const auto clamped = EmitFPMin32(
+		    ctx.state, EmitFPMax32(ctx.state, value, ConstantF32Value(ctx.state, 0.0f)),
+		    ConstantF32Value(ctx.state, 1.0f));
+		const auto scaled = EmitFPMul32(
+		    ctx.state, clamped, ConstantF32Value(ctx.state, float((1u << bits) - 1u)));
+		return Unary(ctx.state, spv::OpConvertFToU, TypeU32(ctx.state),
+		             EmitFPRoundEven32(ctx.state, scaled));
+	}
+	if (bits == 16u && (info.type == Format::ComponentType::Snorm ||
+	                    info.type == Format::ComponentType::Float)) {
 		const auto value = EmitBitCastF32U32(ctx.state, data);
 		const auto pair = EmitCompositeConstructF32x2(ctx.state, value,
 		                                               ConstantF32Value(ctx.state, 0.0f));
-		data = plan.info.type == Format::ComponentType::Float
-		           ? EmitPackHalf2x16(ctx.state, pair)
-		           : EmitPackSnorm2x16(ctx.state, pair);
-	}
-	if (bits == 8u || bits == 16u) {
-		StoreSubwordInBounds(ctx, mem, plan.resource, plan.addresses[component],
-		                     plan.indices[component], bits, data);
-	} else {
-		StoreWordInBounds(ctx, plan.resource, plan.indices[component], data);
+		return info.type == Format::ComponentType::Float ? EmitPackHalf2x16(ctx.state, pair)
+		                                               : EmitPackSnorm2x16(ctx.state, pair);
 	}
+	return data;
+}
+
+void StoreFormattedPrepared(ValueEmitContext& ctx, const IR::Inst& inst,
+                             const IR::MemoryInfo& mem, const MemoryResourceAccess& resource,
+                             const Format::BufferFormatInfo& info, uint32_t data,
+                             uint32_t components) {
+	// RDNA2 formatted stores transfer the entire format, zero-filling missing VGPRs.
+	const auto plan = PrepareFormattedMemory(ctx, inst, mem, resource, info, info.component_count,
+	                                         FormattedAccess::Store);
+	EmitIfCondition(ctx.state, plan.in_bounds, [&]() {
+		auto packed = ConstantU32(ctx.state, 0);
+		for (uint32_t component = 0; component < info.component_count; component++) {
+			auto value = ConstantU32(ctx.state, 0);
+			if (component < components) {
+				value = data;
+				if (components != 1u) {
+					value = ctx.state.builder.AllocateId();
+					ctx.state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(ctx.state),
+					                              value, data, component);
+				}
+			}
+			value = EncodeFormattedStoreComponent(ctx, info, component, value);
+			const auto bits = info.component_bits[component];
+			if (info.packed_bitfield) {
+				packed = EmitBitFieldInsert(
+				    ctx.state, packed, value,
+				    ConstantU32(ctx.state, info.component_bit_offset[component]),
+				    ConstantU32(ctx.state, bits));
+			} else if (bits == 8u || bits == 16u) {
+				StoreSubwordInBounds(ctx, mem, resource, plan.addresses[component],
+				                     plan.indices[component], bits, value);
+			} else {
+				StoreWordInBounds(ctx, resource, plan.indices[component], value);
+			}
+		}
+		if (info.packed_bitfield) StoreWordInBounds(ctx, resource, plan.indices[0], packed);
+	});
 }
 
 void FormattedStore(ValueEmitContext& ctx, const IR::Inst& inst, const IR::MemoryInfo& mem) {
@@ -737,11 +763,7 @@ void FormattedStore(ValueEmitContext& ctx, const IR::Inst& inst, const IR::Memor
 			StoreWordPrepared(ctx, inst, RebaseRawComponent(mem, 0u), resource, data);
 			return;
 		}
-		const auto plan = PrepareFormattedMemory(ctx, inst, mem, resource, info, 1u,
-		                                         FormattedAccess::Store);
-		EmitIfCondition(ctx.state, plan.in_bounds, [&]() {
-			StoreFormattedInBounds(ctx, mem, plan, 0u, data);
-		});
+		StoreFormattedPrepared(ctx, inst, mem, resource, info, data, 1u);
 	});
 }
 
@@ -859,16 +881,7 @@ void StoreWideBuffer(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t compo
 		const auto info = Format::GetFormatInfo(
 		    mem.formatted ? BufferFormat(ctx, mem) : Prospero::BufferFormat::kInvalid);
 		if (info.type != Format::ComponentType::Unknown) {
-			const auto plan = PrepareFormattedMemory(ctx, inst, mem, resource, info, components,
-			                                         FormattedAccess::Store);
-			EmitIfCondition(state, plan.in_bounds, [&]() {
-				for (uint32_t component = 0; component < components; component++) {
-					const auto data = state.builder.AllocateId();
-					state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(state), data,
-					                          composite, component);
-					StoreFormattedInBounds(ctx, mem, plan, component, data);
-				}
-			});
+			StoreFormattedPrepared(ctx, inst, mem, resource, info, composite, components);
 			return;
 		}
 		for (uint32_t component = 0; component < components; component++) {
@@ -928,6 +941,27 @@ void StoreWideShared(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t compo
 
 } // namespace
 
+uint32_t ConstantDeviceAddress(EmitterState& state, uint64_t value) {
+	return state.builder.Constant(spv::OpConstant, TypeScalarU64(state),
+	                              static_cast<uint32_t>(value),
+	                              static_cast<uint32_t>(value >> 32u));
+}
+
+uint32_t DeviceAddressFromWords(EmitterState& state, uint32_t low, uint32_t high) {
+	const auto low64  = Unary(state, spv::OpUConvert, TypeScalarU64(state), low);
+	const auto high64 = Binary(state, spv::OpShiftLeftLogical, TypeScalarU64(state),
+	                           Unary(state, spv::OpUConvert, TypeScalarU64(state), high),
+	                           ConstantDeviceAddress(state, 32));
+	return Binary(state, spv::OpBitwiseOr, TypeScalarU64(state), low64, high64);
+}
+
+uint32_t GetBdaPointer(EmitterState& state, uint32_t address) {
+	const auto result = state.builder.AllocateId();
+	state.builder.AddFunction(spv::OpFunctionCall, TypeScalarU64(state), result,
+	                          state.bda_pointer_function, address);
+	return result;
+}
+
 void DefineGetBdaPointer(EmitterState& state) {
 	if (!state.program.info.uses_dma) {
 		return;
@@ -943,7 +977,14 @@ void DefineGetBdaPointer(EmitterState& state) {
 	state.builder.AddFunction(spv::OpFunctionParameter, type, address);
 	EmitLabel(state, entry_label);
 
-	const auto page64        = Binary(state, spv::OpShiftRightLogical, type, address,
+	const auto extended = Binary(state, spv::OpUGreaterThanEqual, TypeBool(state), address,
+	                             ConstantDeviceAddress(state, LibKernel::Memory::kExtendedMemoryBase));
+	const auto packed = Select(state, type, extended,
+	                           Binary(state, spv::OpISub, type, address,
+	                                  ConstantDeviceAddress(state,
+	                                      LibKernel::Memory::kExtendedMemoryBase - LOWER_ADDRESS_SIZE)),
+	                           address);
+	const auto page64        = Binary(state, spv::OpShiftRightLogical, type, packed,
 	                                  ConstantDeviceAddress(state, BufferCache::CACHING_PAGEBITS));
 	const auto page          = Unary(state, spv::OpUConvert, TypeU32(state), page64);
 	const auto entry_pointer = state.builder.AllocateId();
@@ -984,14 +1025,7 @@ uint32_t EmitAtomic32(ValueEmitContext& ctx, const IR::Inst& inst) {
 		const auto scope =
 		    mem.kind == IR::ResourceKind::Lds ? spv::ScopeWorkgroup : spv::ScopeDevice;
 		const auto old = EmitAtomicOperation(ctx, inst, pointer, scope);
-		if (mem.kind == IR::ResourceKind::Lds) {
-			const auto semantics =
-			    spv::MemorySemanticsAcquireReleaseMask | spv::MemorySemanticsWorkgroupMemoryMask;
-			ctx.state.builder.AddFunction(spv::OpMemoryBarrier, ConstantU32(ctx.state, scope),
-			                              ConstantU32(ctx.state, semantics));
-		} else {
-			EmitDeviceAtomicMemoryBarrier(ctx.state);
-		}
+		EmitAtomicMemoryBarrier(ctx.state, mem.kind);
 		return old;
 	});
 }
@@ -1019,12 +1053,38 @@ uint32_t EmitBufferAtomic64(ValueEmitContext& ctx, const IR::Inst& inst) {
 			                                            TypeStorageBufferU64ElementPointer(state)),
 			            ConstantU32(state, spv::ScopeDevice),
 			            ConstantU32(state, spv::MemorySemanticsMaskNone), value);
-			        EmitDeviceAtomicMemoryBarrier(state);
+			        EmitAtomicMemoryBarrier(state, mem.kind);
 			        return Unary(state, spv::OpBitcast, TypeU64(state), old);
 		        });
 	    });
 }
 
+void EmitSharedAtomic64(ValueEmitContext& ctx, const IR::Inst& inst) {
+	auto& state = ctx.state;
+	const auto& mem = ctx.Memory(inst);
+	EnsureLdsStorage(state);
+	EmitIfCondition(state, ctx.Arg(inst, 2), [&]() {
+		const auto address = Binary(state, spv::OpBitwiseAnd, TypeU32(state),
+		                            ByteAddress(ctx, inst, mem), ConstantU32(state, 0xfff8u));
+		const auto index = Binary(state, spv::OpShiftRightLogical, TypeU32(state), address,
+		                          ConstantU32(state, 3u));
+		const auto in_bounds = Binary(state, spv::OpULessThan, TypeBool(state), index,
+		                              ConstantU32(state, LdsDwordCount(state) / 2u));
+		EmitIfCondition(state, in_bounds, [&]() {
+			const auto pointer = state.builder.AllocateId();
+			state.builder.AddFunction(spv::OpAccessChain,
+			                          TypePointer(state, spv::StorageClassWorkgroup, TypeScalarU64(state)),
+			                          pointer, state.lds_u64_variable, ConstantU32(state, 0), index);
+			const auto value = Unary(state, spv::OpBitcast, TypeScalarU64(state), ctx.Arg(inst, 1));
+			state.builder.AddFunction(
+			    SpirvAtomicOpcode(inst.GetOpcode()), TypeScalarU64(state), state.builder.AllocateId(),
+			    pointer, ConstantU32(state, spv::ScopeWorkgroup),
+			    ConstantU32(state, spv::MemorySemanticsAcquireReleaseMask |
+			                           spv::MemorySemanticsWorkgroupMemoryMask), value);
+		});
+	});
+}
+
 uint32_t EmitBufferFloatAtomic(ValueEmitContext& ctx, const IR::Inst& inst) {
 	const auto& mem       = ctx.Memory(inst);
 	const bool  max_value = inst.GetOpcode() == IR::ValueOpcode::BufferAtomicFMax32;
@@ -1072,16 +1132,27 @@ uint32_t EmitAppendConsume(ValueEmitContext& ctx, const IR::Inst& inst) {
 	if (ctx.half == 1) {
 		return ctx.other_half->Def(IR::Value(const_cast<IR::Inst*>(&inst)));
 	}
-	const auto m0 = ctx.Arg(inst, 0);
-	const auto base =
-	    Binary(state, spv::OpShiftRightLogical, TypeU32(state), m0, ConstantU32(state, 16));
-	const auto size =
-	    Binary(state, spv::OpBitwiseAnd, TypeU32(state), m0, ConstantU32(state, 0xffffu));
-	const auto address = Binary(state, spv::OpIAdd, TypeU32(state), base,
-	                            ConstantU32(state, ctx.Memory(inst).offset));
+	const auto mem           = ctx.Memory(inst);
+	const auto offset        = mem.offset & 0xfffcu;
+	auto       address       = ConstantU32(state, offset);
+	uint32_t   region_bounds = 0;
+	if (mem.kind == IR::ResourceKind::Gds) {
+		const auto m0 = ctx.Arg(inst, 0);
+		const auto base =
+		    Binary(state, spv::OpShiftRightLogical, TypeU32(state), m0, ConstantU32(state, 16));
+		const auto size =
+		    Binary(state, spv::OpBitwiseAnd, TypeU32(state), m0, ConstantU32(state, 0xffffu));
+		address = Binary(state, spv::OpIAdd, TypeU32(state), base, address);
+		// The entire M0 region must fit the PS5's 48 KiB GDS partition.
+		region_bounds = AndCondition(
+		    state, Binary(state, spv::OpULessThan, TypeBool(state),
+		                  ConstantU32(state, offset + 3u), size),
+		    Binary(state, spv::OpULessThanEqual, TypeBool(state),
+		           Binary(state, spv::OpIAdd, TypeU32(state), base, size),
+		           ConstantU32(state, 0xc000u)));
+	}
 	const auto raw_index =
 	    Binary(state, spv::OpShiftRightLogical, TypeU32(state), address, ConstantU32(state, 2));
-	const auto mem    = ctx.Memory(inst);
 	const auto access = PrepareMemoryResourceAccess(state, mem);
 	const auto index  = EmitMemoryElementIndex(state, access, raw_index);
 	const auto exec   = ctx.Arg(inst, 1);
@@ -1100,19 +1171,17 @@ uint32_t EmitAppendConsume(ValueEmitContext& ctx, const IR::Inst& inst) {
 	        : first;
 	const auto is_first       = Binary(state, spv::OpIEqual, TypeBool(state),
 	                                   EmitSubgroupLocalInvocationId(state), source_lane);
-	const auto storage_bounds = EmitMemoryElementInBounds(state, access, index);
-	const auto m0_bounds =
-	    mem.kind == IR::ResourceKind::Gds
-	        ? Binary(state, spv::OpINotEqual, TypeBool(state), size, ConstantU32(state, 0))
-	        : Binary(state, spv::OpULessThan, TypeBool(state),
-	                 ConstantU32(state, ctx.Memory(inst).offset + 3u), size);
+	auto bounds = EmitMemoryElementInBounds(state, access, index);
+	if (mem.kind == IR::ResourceKind::Gds) {
+		bounds = AndCondition(state, bounds, region_bounds);
+	}
 	const auto condition = AndCondition(
 	    state, is_first,
 	    AndCondition(state,
 	                 state.lane_count == 2 ? Binary(state, spv::OpINotEqual, TypeBool(state), count,
 	                                                ConstantU32(state, 0))
 	                                       : exec,
-	                 AndCondition(state, storage_bounds, m0_bounds)));
+	                 bounds));
 	const auto atomic = EmitValueOrZeroIfCondition(state, condition, [&]() {
 		const auto value = state.builder.AllocateId();
 		state.builder.AddFunction(append ? spv::OpAtomicIAdd : spv::OpAtomicISub, TypeU32(state),
@@ -1146,10 +1215,11 @@ void EmitReadConstBuffer(ValueEmitContext& ctx, const IR::Inst& inst) {
 	if (mem.planning_only) return;
 	auto& state        = ctx.state;
 	mem.kind           = IR::ResourceKind::ScalarBuffer;
-	const auto address = Binary(state, spv::OpIAdd, TypeU32(state), ctx.Arg(inst, 1),
-	                            ConstantU32(state, mem.offset));
-	const auto index =
-	    Binary(state, spv::OpShiftRightLogical, TypeU32(state), address, ConstantU32(state, 2));
+	auto index = Binary(state, spv::OpShiftRightLogical, TypeU32(state), ctx.Arg(inst, 1),
+	                    ConstantU32(state, 2));
+	if (mem.offset >= 4u) {
+		index = Binary(state, spv::OpIAdd, TypeU32(state), index, ConstantU32(state, mem.offset >> 2u));
+	}
 	const auto access    = PrepareMemoryResourceAccess(state, mem);
 	const auto element   = EmitMemoryElementIndex(state, access, index);
 	const auto condition = EmitMemoryElementInBounds(state, access, element);
@@ -1167,10 +1237,14 @@ void EmitLoadMemory(ValueEmitContext& ctx, const IR::Inst& inst) {
 	const auto shared_components = IR::SharedComponentCount(op);
 	const auto address_info      = IR::AddressOpcodeInfoOf(op);
 	uint32_t   value;
-	if (buffer_components > 1u)
+	if (mem.kind == IR::ResourceKind::FlatLocal)
+		value = LoadLocalFlat(ctx, inst);
+	else if (buffer_components > 1u)
 		value = LoadWideBuffer(ctx, inst, buffer_components);
 	else if (shared_components > 1u)
 		value = LoadWideShared(ctx, inst, shared_components);
+	else if (mem.kind == IR::ResourceKind::ScalarAddress)
+		value = LoadBdaDword(ctx, GuestAddress(ctx, inst, mem));
 	else if (address_info.access == IR::AddressAccess::Read &&
 	         mem.kind != IR::ResourceKind::Scratch)
 		value = LoadBda(ctx, GuestAddress(ctx, inst, mem), ctx.Arg(inst, inst.NumArgs() - 1),
@@ -1192,7 +1266,9 @@ void EmitStoreMemory(ValueEmitContext& ctx, const IR::Inst& inst) {
 	const auto  buffer_components = IR::BufferComponentCount(op);
 	const auto  shared_components = IR::SharedComponentCount(op);
 	const auto  type              = inst.Arg(inst.NumArgs() - 2).GetType();
-	if (buffer_components > 1u)
+	if (mem.kind == IR::ResourceKind::FlatLocal)
+		StoreLocalFlat(ctx, inst);
+	else if (buffer_components > 1u)
 		StoreWideBuffer(ctx, inst, buffer_components);
 	else if (shared_components > 1u)
 		StoreWideShared(ctx, inst, shared_components);

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

@@ -63,16 +63,34 @@ uint32_t LdsDwordCount(const EmitterState& state) {
 	return workgroup != nullptr ? workgroup->lds_size_dwords : 8192u;
 }
 
-static void EnsureLdsStorage(EmitterState& state) {
+void EnsureLdsStorage(EmitterState& state) {
 	if (state.lds_variable != 0) {
 		return;
 	}
 	if (ShaderWorkgroupInput(state.program.stage, state.input_info) == nullptr) {
 		EXIT("function LDS was not prepared before SPIR-V function emission\n");
 	}
-	state.lds_variable = state.builder.DefineGlobalVariable(
-	    TypeU32ArrayPointer(state, spv::StorageClassWorkgroup, LdsDwordCount(state)),
-	    spv::StorageClassWorkgroup);
+	const auto define = [&](uint32_t type, uint32_t bytes) {
+		const auto array = state.builder.DecoratedType(
+		    spv::OpTypeArray, {{spv::OpDecorate, {spv::DecorationArrayStride, bytes}}}, type,
+		    ConstantU32(state, std::max(LdsDwordCount(state) * 4u / bytes, 1u)));
+		const auto block = state.builder.DecoratedType(
+		    spv::OpTypeStruct, {{spv::OpMemberDecorate, {0, spv::DecorationOffset, 0}},
+		                        {spv::OpDecorate, {spv::DecorationBlock}}}, array);
+		const auto variable = state.builder.DefineGlobalVariable(
+		    TypePointer(state, spv::StorageClassWorkgroup, block), spv::StorageClassWorkgroup);
+		state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationAliased);
+		return variable;
+	};
+	if (state.requirements.shared_int64_atomics) {
+		state.lds_variable = define(TypeU32(state), 4u);
+		state.lds_u64_variable = define(TypeScalarU64(state), 8u);
+		state.builder.AddName(state.lds_u64_variable, "lds_qwords");
+	} else {
+		state.lds_variable = state.builder.DefineGlobalVariable(
+		    TypeU32ArrayPointer(state, spv::StorageClassWorkgroup, LdsDwordCount(state)),
+		    spv::StorageClassWorkgroup);
+	}
 	state.builder.AddName(state.lds_variable, "lds_dwords");
 }
 
@@ -167,6 +185,11 @@ uint32_t EmitMemoryElementPointer(EmitterState& state, const MemoryResourceAcces
 		    : ShaderWorkgroupInput(state.program.stage, state.input_info) != nullptr
 		        ? spv::StorageClassWorkgroup
 		        : spv::StorageClassFunction;
+		if (access.kind == IR::ResourceKind::Lds && state.requirements.shared_int64_atomics) {
+			state.builder.AddFunction(spv::OpAccessChain, TypeU32ElementPointer(state, storage_class),
+			                          pointer, access.object_pointer, ConstantU32(state, 0), index);
+			return pointer;
+		}
 		state.builder.AddFunction(spv::OpAccessChain, TypeU32ElementPointer(state, storage_class),
 		                          pointer, access.object_pointer, index);
 		return pointer;
@@ -283,11 +306,79 @@ uint32_t NormalizeFormatComponent(EmitterState& state, const Format::BufferForma
 	}
 }
 
-void EmitDeviceAtomicMemoryBarrier(EmitterState& state) {
-	const auto semantics =
-	    spv::MemorySemanticsAcquireReleaseMask | spv::MemorySemanticsUniformMemoryMask;
-	state.builder.AddFunction(spv::OpMemoryBarrier, ConstantU32(state, spv::ScopeDevice),
-	                          ConstantU32(state, semantics));
+spv::Op SpirvAtomicOpcode(IR::ValueOpcode opcode) {
+	switch (opcode) {
+		case IR::ValueOpcode::ImageAtomicCompareSwap32:
+		case IR::ValueOpcode::BufferAtomicCmpSwap32: return spv::OpAtomicCompareExchange;
+		case IR::ValueOpcode::ImageAtomicSwap32:
+		case IR::ValueOpcode::BufferAtomicSwap32:
+		case IR::ValueOpcode::BufferAtomicSwap64:
+		case IR::ValueOpcode::SharedAtomicSwap32: return spv::OpAtomicExchange;
+		case IR::ValueOpcode::ImageAtomicIAdd32:
+		case IR::ValueOpcode::BufferAtomicIAdd32:
+		case IR::ValueOpcode::SharedAtomicIAdd64:
+		case IR::ValueOpcode::SharedAtomicIAdd32: return spv::OpAtomicIAdd;
+		case IR::ValueOpcode::BufferAtomicISub32:
+		case IR::ValueOpcode::SharedAtomicISub32: return spv::OpAtomicISub;
+		case IR::ValueOpcode::ImageAtomicSMin32:
+		case IR::ValueOpcode::BufferAtomicSMin32:
+		case IR::ValueOpcode::SharedAtomicSMin32: return spv::OpAtomicSMin;
+		case IR::ValueOpcode::ImageAtomicUMin32:
+		case IR::ValueOpcode::BufferAtomicUMin32:
+		case IR::ValueOpcode::SharedAtomicUMin32: return spv::OpAtomicUMin;
+		case IR::ValueOpcode::ImageAtomicSMax32:
+		case IR::ValueOpcode::BufferAtomicSMax32:
+		case IR::ValueOpcode::SharedAtomicSMax32: return spv::OpAtomicSMax;
+		case IR::ValueOpcode::ImageAtomicUMax32:
+		case IR::ValueOpcode::BufferAtomicUMax32:
+		case IR::ValueOpcode::SharedAtomicUMax32: return spv::OpAtomicUMax;
+		case IR::ValueOpcode::ImageAtomicAnd32:
+		case IR::ValueOpcode::BufferAtomicAnd32:
+		case IR::ValueOpcode::BufferAtomicAnd64:
+		case IR::ValueOpcode::SharedAtomicAnd32: return spv::OpAtomicAnd;
+		case IR::ValueOpcode::ImageAtomicOr32:
+		case IR::ValueOpcode::BufferAtomicOr32:
+		case IR::ValueOpcode::BufferAtomicOr64:
+		case IR::ValueOpcode::SharedAtomicOr64:
+		case IR::ValueOpcode::SharedAtomicOr32: return spv::OpAtomicOr;
+		case IR::ValueOpcode::ImageAtomicXor32:
+		case IR::ValueOpcode::BufferAtomicXor32:
+		case IR::ValueOpcode::SharedAtomicXor32: return spv::OpAtomicXor;
+		default: return spv::OpNop;
+	}
+}
+
+uint32_t EmitAtomicOperation(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t pointer,
+                             uint32_t scope) {
+	const auto opcode = SpirvAtomicOpcode(inst.GetOpcode());
+	const auto old    = ctx.state.builder.AllocateId();
+	if (opcode == spv::OpAtomicCompareExchange) {
+		const auto desired    = ctx.Arg(inst, inst.NumArgs() - 3);
+		const auto comparator = ctx.Arg(inst, inst.NumArgs() - 2);
+		ctx.state.builder.AddFunction(
+		    spv::OpAtomicCompareExchange, TypeU32(ctx.state), old, pointer,
+		    ConstantU32(ctx.state, scope), ConstantU32(ctx.state, spv::MemorySemanticsMaskNone),
+		    ConstantU32(ctx.state, spv::MemorySemanticsMaskNone), desired, comparator);
+	} else {
+		const auto value = ctx.Arg(inst, inst.NumArgs() - 2);
+		ctx.state.builder.AddFunction(opcode, TypeU32(ctx.state), old, pointer,
+		                              ConstantU32(ctx.state, scope),
+		                              ConstantU32(ctx.state, spv::MemorySemanticsMaskNone), value);
+	}
+	return old;
+}
+
+void EmitAtomicMemoryBarrier(EmitterState& state, IR::ResourceKind kind) {
+	const auto scope  = kind == IR::ResourceKind::Lds ? spv::ScopeWorkgroup : spv::ScopeDevice;
+	const auto memory = [&] {
+		switch (kind) {
+			case IR::ResourceKind::Lds: return spv::MemorySemanticsWorkgroupMemoryMask;
+			case IR::ResourceKind::Image: return spv::MemorySemanticsImageMemoryMask;
+			default: return spv::MemorySemanticsUniformMemoryMask;
+		}
+	}();
+	state.builder.AddFunction(spv::OpMemoryBarrier, ConstantU32(state, scope),
+	                          ConstantU32(state, spv::MemorySemanticsAcquireReleaseMask | memory));
 }
 
 uint32_t EmitFloatAtomicReplacement(EmitterState& state, uint32_t old, uint32_t source,

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

@@ -6,19 +6,32 @@
 namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter {
 
 uint32_t TypeVoid(EmitterState& state) {
-	return state.builder.Type(spv::OpTypeVoid);
+	if (state.void_type == 0) {
+		state.void_type = state.builder.Type(spv::OpTypeVoid);
+	}
+	return state.void_type;
 }
 
 uint32_t TypeBool(EmitterState& state) {
-	return state.builder.Type(spv::OpTypeBool);
+	if (state.bool_type == 0) {
+		state.bool_type = state.builder.Type(spv::OpTypeBool);
+	}
+	return state.bool_type;
 }
 
 uint32_t TypeBoolVector(EmitterState& state, uint32_t components) {
-	return state.builder.Type(spv::OpTypeVector, TypeBool(state), components);
+	auto& type = state.bool_vector_types[components - 2u];
+	if (type == 0) {
+		type = state.builder.Type(spv::OpTypeVector, TypeBool(state), components);
+	}
+	return type;
 }
 
 uint32_t TypeU32(EmitterState& state) {
-	return state.builder.Type(spv::OpTypeInt, 32, 0);
+	if (state.u32_type == 0) {
+		state.u32_type = state.builder.Type(spv::OpTypeInt, 32, 0);
+	}
+	return state.u32_type;
 }
 
 uint32_t TypeU64(EmitterState& state) {
@@ -26,29 +39,55 @@ uint32_t TypeU64(EmitterState& state) {
 }
 
 uint32_t TypeScalarU64(EmitterState& state) {
-	return state.builder.Type(spv::OpTypeInt, 64, 0);
+	if (state.native_u64_type == 0) {
+		state.native_u64_type = state.builder.Type(spv::OpTypeInt, 64, 0);
+	}
+	return state.native_u64_type;
 }
 
 uint32_t TypeU32Pair(EmitterState& state) {
-	const auto element = TypeU32(state);
-	return state.builder.Type(spv::OpTypeStruct, element, element);
+	if (state.u32_pair_type == 0) {
+		const auto element = TypeU32(state);
+		state.u32_pair_type = state.builder.Type(spv::OpTypeStruct, element, element);
+	}
+	return state.u32_pair_type;
 }
 
 uint32_t TypeI32(EmitterState& state) {
-	return state.builder.Type(spv::OpTypeInt, 32, 1);
+	if (state.i32_type == 0) {
+		state.i32_type = state.builder.Type(spv::OpTypeInt, 32, 1);
+	}
+	return state.i32_type;
 }
 
 uint32_t TypeI32Pair(EmitterState& state) {
-	const auto element = TypeI32(state);
-	return state.builder.Type(spv::OpTypeStruct, element, element);
+	if (state.i32_pair_type == 0) {
+		const auto element = TypeI32(state);
+		state.i32_pair_type = state.builder.Type(spv::OpTypeStruct, element, element);
+	}
+	return state.i32_pair_type;
 }
 
 uint32_t TypeF32(EmitterState& state) {
-	return state.builder.Type(spv::OpTypeFloat, 32);
+	if (state.f32_type == 0) {
+		state.f32_type = state.builder.Type(spv::OpTypeFloat, 32);
+	}
+	return state.f32_type;
+}
+
+uint32_t TypeF64(EmitterState& state) {
+	if (state.f64_type == 0) {
+		state.f64_type = state.builder.Type(spv::OpTypeFloat, 64);
+	}
+	return state.f64_type;
 }
 
 uint32_t TypeU32Vector(EmitterState& state, uint32_t components) {
-	return state.builder.Type(spv::OpTypeVector, TypeU32(state), components);
+	auto& type = state.u32_vector_types[components - 2u];
+	if (type == 0) {
+		type = state.builder.Type(spv::OpTypeVector, TypeU32(state), components);
+	}
+	return type;
 }
 
 uint32_t TypeU32Composite(EmitterState& state, uint32_t components) {
@@ -57,11 +96,19 @@ uint32_t TypeU32Composite(EmitterState& state, uint32_t components) {
 }
 
 uint32_t TypeI32Vector(EmitterState& state, uint32_t components) {
-	return state.builder.Type(spv::OpTypeVector, TypeI32(state), components);
+	auto& type = state.i32_vector_types[components - 2u];
+	if (type == 0) {
+		type = state.builder.Type(spv::OpTypeVector, TypeI32(state), components);
+	}
+	return type;
 }
 
 uint32_t TypeF32Vector(EmitterState& state, uint32_t components) {
-	return state.builder.Type(spv::OpTypeVector, TypeF32(state), components);
+	auto& type = state.f32_vector_types[components - 2u];
+	if (type == 0) {
+		type = state.builder.Type(spv::OpTypeVector, TypeF32(state), components);
+	}
+	return type;
 }
 
 uint32_t TypePointer(EmitterState& state, spv::StorageClass storage_class, uint32_t pointee) {
@@ -69,7 +116,10 @@ uint32_t TypePointer(EmitterState& state, spv::StorageClass storage_class, uint3
 }
 
 uint32_t TypeFunction(EmitterState& state) {
-	return state.builder.Type(spv::OpTypeFunction, TypeVoid(state));
+	if (state.function_type == 0) {
+		state.function_type = state.builder.Type(spv::OpTypeFunction, TypeVoid(state));
+	}
+	return state.function_type;
 }
 
 uint32_t StorageRuntimeArrayType(EmitterState& state) {
@@ -622,10 +672,15 @@ void DefineModule(EmitterState& state) {
 		state.builder.RequireCapability(spv::CapabilityPhysicalStorageBufferAddresses);
 		state.builder.RequireExtension("SPV_KHR_physical_storage_buffer");
 	}
-	if (state.requirements.buffer_int64_atomics) {
+	if (state.requirements.buffer_int64_atomics || state.requirements.shared_int64_atomics) {
 		state.builder.RequireCapability(spv::CapabilityInt64);
 		state.builder.RequireCapability(spv::CapabilityInt64Atomics);
 	}
+	if (state.requirements.shared_int64_atomics) {
+		state.builder.RequireVersion(0x00010400u);
+		state.builder.RequireExtension("SPV_KHR_workgroup_memory_explicit_layout");
+		state.builder.RequireCapability(spv::CapabilityWorkgroupMemoryExplicitLayoutKHR);
+	}
 	if (state.clip_distance_variable != 0) {
 		state.builder.RequireCapability(spv::CapabilityClipDistance);
 	}
@@ -679,6 +734,19 @@ void DefineModule(EmitterState& state) {
 	// contract prevents host compilers from treating synthesized IEEE values as finite.
 	state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeSignedZeroInfNanPreserve,
 	                               32u);
+	if (state.requirements.float64) {
+		EXIT_NOT_IMPLEMENTED(state.program.stage == ShaderType::Compute &&
+		                     state.input_info.compute->float_mode != 0xc0);
+		// MODE=0xc0 uses round-to-nearest-even and preserves FP64 input/output denormals.
+		state.builder.RequireCapability(spv::CapabilityFloat64);
+		state.builder.RequireCapability(spv::CapabilityRoundingModeRTE);
+		state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeSignedZeroInfNanPreserve,
+		                               64u);
+		// FP64 denormal preservation is temporarily disabled.
+		// state.builder.RequireCapability(spv::CapabilityDenormPreserve);
+		// state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeDenormPreserve, 64u);
+		state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeRoundingModeRTE, 32u);
+	}
 	if (const auto* cs = ShaderWorkgroupInput(state.program.stage, state.input_info)) {
 		uint32_t    local_x = state.requirements.compute_derivatives ? 2u : 1u;
 		uint32_t    local_y = state.requirements.compute_derivatives ? 2u : 1u;

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

@@ -97,30 +97,6 @@ void EmitReturn(ValueEmitContext& ctx) {
 	ctx.state.builder.AddFunction(spv::OpReturn);
 }
 
-uint32_t BranchCondition(ValueEmitContext& ctx, const IR::BlockInfo& info) {
-	// Scalar-instruction conditions already test the full wave's raw register values.
-	if (ctx.other_half == nullptr ||
-	    info.terminator.condition == CFG::BranchCondition::ScalarInstruction ||
-	    info.terminator.condition == CFG::BranchCondition::GotoVariable) {
-		return ctx.Def(info.condition);
-	}
-	const auto ballot = ctx.Ballot(info.condition);
-	const auto low    = ctx.state.builder.AllocateId();
-	const auto high   = ctx.state.builder.AllocateId();
-	const auto result = ctx.state.builder.AllocateId();
-	ctx.state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(ctx.state), low, ballot, 0);
-	ctx.state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(ctx.state), high, ballot, 1);
-	const auto kind     = info.terminator.condition;
-	const bool zero     = kind == CFG::BranchCondition::ExecZero ||
-	                      kind == CFG::BranchCondition::VccZero ||
-	                      kind == CFG::BranchCondition::SccZero;
-	const auto combined =
-	    EmitBinaryU32(ctx.state, zero ? spv::OpBitwiseAnd : spv::OpBitwiseOr, low, high);
-	ctx.state.builder.AddFunction(zero ? spv::OpIEqual : spv::OpINotEqual, TypeBool(ctx.state),
-	                              result, combined, ConstantU32(ctx.state, zero ? ~0u : 0u));
-	return result;
-}
-
 void EmitStructuredTerminator(ValueEmitContext& ctx, const IR::Block* block,
                               const IR::BlockInfo& info) {
 	const auto& program = ctx.state.program;
@@ -160,7 +136,7 @@ void EmitStructuredTerminator(ValueEmitContext& ctx, const IR::Block* block,
 				EmitReturn(ctx);
 				return;
 			}
-			const auto condition = BranchCondition(ctx, info);
+			const auto condition = ctx.Def(info.condition);
 			emit_merge();
 			ctx.state.builder.AddFunction(spv::OpBranchConditional, condition,
 			                              ctx.Label(true_block), ctx.Label(false_block));
@@ -193,7 +169,7 @@ uint32_t EmitDispatcherNextPc(ValueEmitContext& ctx, const DispatcherFunctionSta
 			EmitDispatcherTarget(ctx, dispatcher, block, term.false_block);
 			const auto selected = ctx.state.builder.AllocateId();
 			ctx.state.builder.AddFunction(
-			    spv::OpSelect, TypeU32(ctx.state), selected, BranchCondition(ctx, info),
+			    spv::OpSelect, TypeU32(ctx.state), selected, ctx.Def(info.condition),
 			    ConstantU32(ctx.state, term.true_block), ConstantU32(ctx.state, term.false_block));
 			return selected;
 		}
@@ -267,6 +243,12 @@ void Invoke(Return (*emit)(Context&, Args...), ValueEmitContext& ctx, const IR::
 }
 
 void EmitDirectInstruction(ValueEmitContext& ctx, const IR::Inst& inst) {
+	if (ctx.half != 0 && (inst.GetOpcode() == IR::ValueOpcode::Ballot ||
+	                     inst.GetOpcode() == IR::ValueOpcode::ReadFirstLane)) {
+		// Both operations already combine both emulated halves into one whole-wave result.
+		ctx.Define(inst, ctx.other_half->Result(inst));
+		return;
+	}
 	switch (inst.GetOpcode()) {
 #define VALUE_OPCODE(name, ...)                                                                    \
 	case IR::ValueOpcode::name: return Invoke(Emit##name, ctx, inst);
@@ -440,6 +422,7 @@ uint32_t TypeId(EmitterState& state, IR::Type type) {
 		case IR::Type::U64: return TypeU64(state);
 		case IR::Type::U32x2: return TypeU32Pair(state);
 		case IR::Type::F32: return TypeF32(state);
+		case IR::Type::F64: return TypeF64(state);
 		case IR::Type::U32x3: return TypeU32Vector(state, 3);
 		case IR::Type::U32x4: return TypeU32Vector(state, 4);
 		case IR::Type::F32x2: return TypeF32Vector(state, 2);
@@ -702,6 +685,7 @@ void EmitProgram(EmitterState& state) {
 		}
 	}
 	DefineGetBdaPointer(state);
+	DefineBvhIntersect(state);
 	for (const auto* block: program.blocks) {
 		if (std::ranges::any_of(*block, [](const IR::Inst& inst) {
 			    return inst.GetOpcode() == IR::ValueOpcode::SwizzleU32 ||

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

@@ -18,10 +18,17 @@ enum class BranchCondition {
 	ExecZero,
 	ExecNonZero,
 	ScalarInstruction,
-	GotoVariable,
+	Expression,
 	Unknown
 };
 
+struct ConditionExpression {
+	enum class Op { Constant, Variable, Native, Not, Or };
+	Op       op;
+	uint32_t lhs;
+	uint32_t rhs = UINT32_MAX;
+};
+
 enum class TerminatorKind { Branch, ConditionalBranch, IndirectBranch, Return, Unsupported };
 
 enum class FailureKind {
@@ -48,8 +55,7 @@ struct Terminator {
 	std::vector<uint32_t> indirect_targets;
 	std::vector<uint32_t> indirect_selector_values;
 	std::vector<uint32_t> indirect_selector_targets;
-	uint32_t              goto_variable = UINT32_MAX;
-	int32_t               goto_value    = -1;
+	uint32_t              expression    = UINT32_MAX;
 	bool                  loop_header   = false;
 };
 
@@ -59,10 +65,11 @@ struct BasicBlock {
 	uint32_t              end_pc     = 0;
 	uint32_t              inst_begin = 0;
 	uint32_t              inst_end   = 0;
+	struct Assignment { uint32_t variable; uint32_t expression; };
+	std::vector<Assignment> assignments;
 	std::vector<uint32_t> predecessors;
 	std::vector<uint32_t> successors;
 	std::vector<uint32_t> dominators;
-	std::vector<uint32_t> post_dominators;
 	Terminator            terminator;
 };
 
@@ -73,12 +80,9 @@ struct BackEdge {
 };
 
 struct NaturalLoop {
-	uint32_t              header         = UINT32_MAX;
-	uint32_t              latch          = UINT32_MAX;
-	uint32_t              merge          = UINT32_MAX;
-	uint32_t              continue_block = UINT32_MAX;
+	uint32_t              header = UINT32_MAX;
+	uint32_t              latch  = UINT32_MAX;
 	std::vector<uint32_t> body_blocks;
-	std::vector<uint32_t> exit_blocks;
 };
 
 struct StronglyConnectedComponent {
@@ -89,6 +93,7 @@ struct StronglyConnectedComponent {
 
 struct Graph {
 	std::vector<BasicBlock>                 blocks;
+	std::vector<ConditionExpression>        expressions;
 	std::vector<BackEdge>                   back_edges;
 	std::vector<NaturalLoop>                natural_loops;
 	std::vector<StronglyConnectedComponent> components;
@@ -105,14 +110,12 @@ struct Graph {
 	const BasicBlock* FindBlockByPc(uint32_t pc) const;
 	BasicBlock*       FindBlockByPc(uint32_t pc);
 	bool              Dominates(uint32_t dominator, uint32_t block) const;
-	bool              PostDominates(uint32_t post_dominator, uint32_t block) const;
-	uint32_t          FindNearestCommonPostDominator(uint32_t block_a, uint32_t block_b) const;
 };
 
 Graph       BuildGraph(const Decoder::Program& program);
-// Commits structured control flow on success; preserves the original graph with
-// failure diagnostics on failure. failure_block is an original block ID or UINT32_MAX.
-bool        Structurize(Graph& graph);
+// 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);
 std::string BranchConditionToString(BranchCondition condition);
 std::string FailureKindToString(FailureKind kind);
 std::string GraphToString(const Graph& graph);

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

@@ -192,12 +192,17 @@ constexpr MimgGatherInfo MIMG_GATHER_OPCODE_LIST[] = {
 
 constexpr Detail::OpcodeMap MIMG_ATOMIC_OPCODE_LIST[] = {
     {0x0fu, Opcode::IMAGE_ATOMIC_SWAP},
+    {0x10u, Opcode::IMAGE_ATOMIC_CMPSWAP},
     {0x11u, Opcode::IMAGE_ATOMIC_ADD},
+    {0x14u, Opcode::IMAGE_ATOMIC_SMIN},
     {0x15u, Opcode::IMAGE_ATOMIC_UMIN},
+    {0x16u, Opcode::IMAGE_ATOMIC_SMAX},
     {0x17u, Opcode::IMAGE_ATOMIC_UMAX},
     {0x18u, Opcode::IMAGE_ATOMIC_AND},
     {0x19u, Opcode::IMAGE_ATOMIC_OR},
     {0x1au, Opcode::IMAGE_ATOMIC_XOR},
+    {0x1eu, Opcode::IMAGE_ATOMIC_FMIN},
+    {0x1fu, Opcode::IMAGE_ATOMIC_FMAX},
 };
 
 constexpr auto MIMG_SAMPLE_OPS = Detail::MakeOpcodeTable<0x100>(MIMG_SAMPLE_OPCODE_LIST);
@@ -223,6 +228,7 @@ Opcode DecodeMimgOpcode(uint32_t opcode, const MimgSampleInfo* sample, const Mim
 		case 0x09u: return Opcode::IMAGE_STORE_MIP;
 		case 0x0eu: return Opcode::IMAGE_GET_RESINFO;
 		case 0x60u: return Opcode::IMAGE_GET_LOD;
+		case 0xe6u: return Opcode::IMAGE_BVH_INTERSECT_RAY;
 		default: return Opcode::UNSUPPORTED;
 	}
 }
@@ -251,6 +257,7 @@ uint32_t DecodeMimgAddressComponents(uint32_t opcode, ImageDimension dimension,
 	}
 
 	switch (opcode) {
+		case 0xe6u: return 11u;
 		case 0x0eu: return 1u;
 		case 0x01u:
 		case 0x09u: return ImageCoordComponents(dimension) + 1u;
@@ -353,11 +360,20 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 		SetUnsupported(inst, Family::MIMG, opcode,
 		               "MIMG image gather requires exactly one dmask bit");
 	}
+	if (inst.opcode == Opcode::IMAGE_ATOMIC_CMPSWAP && inst.dmask != 0x3u) {
+		SetUnsupported(inst, Family::MIMG, opcode,
+		               "MIMG image compare-and-swap requires 32-bit DMASK 0x3");
+	}
 	const bool supports_d16 = sample != nullptr || gather != 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");
 	}
+	if (opcode == 0xe6u &&
+	    (a16 || !r128 || inst.dmask != 0xfu || (nsa_dwords != 0u && nsa_dwords != 3u))) {
+		SetUnsupported(inst, Family::MIMG, opcode,
+		               "BVH intersection requires eleven full-float ray DWORDs and R128/dmask:0xf");
+	}
 
 	DecodeVectorGpr(vdata, inst.dst);
 	DecodeVectorGpr(vaddr, inst.src0);

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

@@ -61,6 +61,7 @@ constexpr MemoryOpcodeInfo MUBUF_OPCODE_LIST[] = {
     {0x3fu, Opcode::BUFFER_ATOMIC_FMIN, 1, 32},
     {0x40u, Opcode::BUFFER_ATOMIC_FMAX, 1, 32},
     {0x50u, Opcode::BUFFER_ATOMIC_SWAP_X2, 2, 32},
+    {0x59u, Opcode::BUFFER_ATOMIC_AND_X2, 2, 32},
     {0x5au, Opcode::BUFFER_ATOMIC_OR_X2, 2, 32},
 };
 
@@ -106,6 +107,7 @@ constexpr MemoryOpcodeInfo DS_OPCODE_LIST[] = {
     {0x39u, Opcode::DS_READ_I8, 1, 8, true},     {0x3au, Opcode::DS_READ_U8, 1, 8},
     {0x3bu, Opcode::DS_READ_I16, 1, 16, true},   {0x3cu, Opcode::DS_READ_U16, 1, 16},
     {0x3du, Opcode::DS_CONSUME, 1, 32},          {0x3eu, Opcode::DS_APPEND, 1, 32},
+    {0x40u, Opcode::DS_ADD_U64, 2, 32},         {0x4au, Opcode::DS_OR_B64, 2, 32},
     {0x4du, Opcode::DS_WRITE_B64, 2, 32},        {0x4eu, Opcode::DS_WRITE2_B64, 4, 32},
     {0x4fu, Opcode::DS_WRITE2ST64_B64, 4, 32},   {0x76u, Opcode::DS_READ_B64, 2, 32},
     {0x77u, Opcode::DS_READ2_B64, 4, 32},        {0x78u, Opcode::DS_READ2ST64_B64, 4, 32},
@@ -167,6 +169,7 @@ bool IsDsWriteOpcode(Opcode opcode) {
 bool IsDsAtomicOpcode(Opcode opcode) {
 	switch (opcode) {
 		case Opcode::DS_ADD_U32:
+		case Opcode::DS_ADD_U64:
 		case Opcode::DS_ADD_RTN_U32:
 		case Opcode::DS_SUB_U32:
 		case Opcode::DS_SUB_RTN_U32:
@@ -185,6 +188,7 @@ bool IsDsAtomicOpcode(Opcode opcode) {
 		case Opcode::DS_AND_B32:
 		case Opcode::DS_AND_RTN_B32:
 		case Opcode::DS_OR_B32:
+		case Opcode::DS_OR_B64:
 		case Opcode::DS_OR_RTN_B32:
 		case Opcode::DS_XOR_B32:
 		case Opcode::DS_XOR_RTN_B32:

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

@@ -46,6 +46,7 @@ constexpr OpcodeMap SOP1_OPCODE_LIST[] = {
     {0x09u, Opcode::S_WQM_B32},
     {0x0au, Opcode::S_WQM_B64},
     {0x0bu, Opcode::S_BREV_B32},
+    {0x0cu, Opcode::S_BREV_B64},
     {0x0fu, Opcode::S_BCNT1_I32_B32},
     {0x10u, Opcode::S_BCNT1_I32_B64},
     {0x13u, Opcode::S_FF1_I32_B32},
@@ -108,6 +109,7 @@ constexpr OpcodeMap SOPP_OPCODE_LIST[] = {
     {0x12u, Opcode::S_TRAP},
     {0x16u, Opcode::S_TTRACEDATA},
     {0x17u, Opcode::S_CBRANCH_CDBGSYS},
+    {0x19u, Opcode::S_CBRANCH_CDBGSYS_OR_USER},
     {0x20u, Opcode::S_INST_PREFETCH},
     {0x23u, Opcode::S_WAITCNT_DEPCTR},
 };

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

@@ -184,8 +184,9 @@ std::string FormatExp(const Instruction& inst) {
 
 bool IsConditionalBranch(Opcode opcode) {
 	switch (opcode) {
-		// DevKit NGG should be disabled.
-		case Opcode::S_CBRANCH_CDBGSYS: return false;
+		// Conditional shader debugging is disabled.
+		case Opcode::S_CBRANCH_CDBGSYS:
+		case Opcode::S_CBRANCH_CDBGSYS_OR_USER: return false;
 		case Opcode::S_CBRANCH_SCC0:
 		case Opcode::S_CBRANCH_SCC1:
 		case Opcode::S_CBRANCH_VCCZ:
@@ -253,6 +254,8 @@ void DecodeScalarSource(uint32_t code, uint32_t pc, Operand& operand) {
 		case 125u: operand.kind = OperandKind::Null; return;
 		case 126u: operand.kind = OperandKind::ExecLo; return;
 		case 127u: operand.kind = OperandKind::ExecHi; return;
+		case 235u: operand.kind = OperandKind::SharedBase; return;
+		case 237u: operand.kind = OperandKind::PrivateBase; return;
 		case 239u: operand.kind = OperandKind::PopsExitingWaveId; return;
 		case 248u:
 			operand.kind      = OperandKind::FloatInlineConstant;
@@ -403,7 +406,6 @@ void DecodeProgram(std::span<const uint32_t> code, Program& program) {
 	program.instructions.clear();
 	program.instructions.reserve(code.size());
 	program.code = code;
-	program.has_bvh = false;
 
 	std::vector<bool> branch_targets;
 	for (uint32_t word_index = 0; word_index < code.size();) {
@@ -412,10 +414,6 @@ void DecodeProgram(std::span<const uint32_t> code, Program& program) {
 
 		const auto& inst = program.instructions.back();
 		word_index += inst.word_count;
-		if (inst.family == Family::MIMG && (inst.opcode_id == 0xe6u || inst.opcode_id == 0xe7u)) {
-			program.has_bvh = true;
-			return;
-		}
 
 		if (IsDirectBranch(inst.opcode)) {
 			const auto target_index = inst.branch_target / sizeof(uint32_t);
@@ -454,6 +452,8 @@ std::string OperandToString(const Operand& operand) {
 		case OperandKind::Scc: text = "scc"; break;
 		case OperandKind::M0: text = "m0"; break;
 		case OperandKind::PopsExitingWaveId: text = "pops_exiting_wave_id"; break;
+		case OperandKind::SharedBase: text = "shared_base"; break;
+		case OperandKind::PrivateBase: text = "private_base"; break;
 		case OperandKind::Null: text = "null"; break;
 		default: text = "unknown"; break;
 	}
@@ -504,6 +504,7 @@ std::string InstructionToString(const Instruction& inst) {
 			                                               OperandToString(inst.src0).c_str()));
 		case Opcode::S_ABS_I32:
 		case Opcode::S_BREV_B32:
+		case Opcode::S_BREV_B64:
 		case Opcode::S_BCNT1_I32_B32:
 		case Opcode::S_FLBIT_I32_B32:
 		case Opcode::S_FF1_I32_B32:
@@ -560,6 +561,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::S_CBRANCH_EXECZ:
 		case Opcode::S_CBRANCH_EXECNZ:
 		case Opcode::S_CBRANCH_CDBGSYS:
+		case Opcode::S_CBRANCH_CDBGSYS_OR_USER:
 			return WithUnsupportedReason(inst, fmt::format("0x{:08x}: {} 0x{:08x}", inst.pc,
 			                                               magic_enum::enum_name(inst.opcode),
 			                                               inst.branch_target));
@@ -572,13 +574,18 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::IMAGE_SAMPLE:
 		case Opcode::IMAGE_STORE:
 		case Opcode::IMAGE_STORE_MIP:
+		case Opcode::IMAGE_ATOMIC_CMPSWAP:
 		case Opcode::IMAGE_ATOMIC_SWAP:
 		case Opcode::IMAGE_ATOMIC_ADD:
+		case Opcode::IMAGE_ATOMIC_SMIN:
 		case Opcode::IMAGE_ATOMIC_UMIN:
+		case Opcode::IMAGE_ATOMIC_SMAX:
 		case Opcode::IMAGE_ATOMIC_UMAX:
 		case Opcode::IMAGE_ATOMIC_AND:
 		case Opcode::IMAGE_ATOMIC_OR:
 		case Opcode::IMAGE_ATOMIC_XOR:
+		case Opcode::IMAGE_ATOMIC_FMIN:
+		case Opcode::IMAGE_ATOMIC_FMAX:
 		case Opcode::IMAGE_LOAD:
 		case Opcode::IMAGE_LOAD_MIP:
 		case Opcode::IMAGE_GET_RESINFO:
@@ -639,6 +646,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::BUFFER_ATOMIC_SMAX:
 		case Opcode::BUFFER_ATOMIC_UMAX:
 		case Opcode::BUFFER_ATOMIC_AND:
+		case Opcode::BUFFER_ATOMIC_AND_X2:
 		case Opcode::BUFFER_ATOMIC_OR:
 		case Opcode::BUFFER_ATOMIC_OR_X2:
 		case Opcode::BUFFER_ATOMIC_XOR:
@@ -661,6 +669,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::FLAT_STORE_DWORDX3:
 		case Opcode::FLAT_STORE_DWORDX4:
 		case Opcode::DS_ADD_U32:
+		case Opcode::DS_ADD_U64:
 		case Opcode::DS_ADD_RTN_U32:
 		case Opcode::DS_SUB_U32:
 		case Opcode::DS_SUB_RTN_U32:
@@ -679,6 +688,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::DS_AND_B32:
 		case Opcode::DS_AND_RTN_B32:
 		case Opcode::DS_OR_B32:
+		case Opcode::DS_OR_B64:
 		case Opcode::DS_OR_RTN_B32:
 		case Opcode::DS_XOR_B32:
 		case Opcode::DS_XOR_RTN_B32:

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

@@ -10,6 +10,10 @@
 
 namespace Libs::Graphics::ShaderRecompiler::Decoder {
 
+// PS5 flat stack addresses use these SH_MEM_BASES aperture tags in VA[63:32].
+constexpr uint32_t PrivateApertureHigh = 0x70000000u;
+constexpr uint32_t SharedApertureHigh = 0x80000000u;
+
 enum class Family {
 	Unknown,
 	SOP1,
@@ -43,6 +47,7 @@ enum class Opcode {
 	S_ABS_I32,
 	S_ABSDIFF_I32,
 	S_BREV_B32,
+	S_BREV_B64,
 	S_BCNT1_I32_B32,
 	S_BCNT1_I32_B64,
 	S_FF1_I32_B32,
@@ -154,6 +159,10 @@ enum class Opcode {
 	V_CUBEMA_F32,
 	V_CNDMASK_B32,
 	V_DOT2C_F32_F16,
+	V_CVT_F64_I32,
+	V_CVT_F32_F64,
+	V_CVT_F64_F32,
+	V_CVT_F64_U32,
 	V_CVT_F32_I32,
 	V_CVT_F32_U32,
 	V_CVT_U32_F32,
@@ -173,6 +182,7 @@ enum class Opcode {
 	V_CVT_F32_UBYTE1,
 	V_CVT_F32_UBYTE2,
 	V_CVT_F32_UBYTE3,
+	V_RCP_F64,
 	V_RCP_F32,
 	V_RCP_IFLAG_F32,
 	V_FRACT_F32,
@@ -234,6 +244,9 @@ enum class Opcode {
 	V_MAD_I32_I24,
 	V_MAD_U32_U24,
 	V_MAD_U64_U32,
+	V_FMA_F64,
+	V_ADD_F64,
+	V_MUL_F64,
 	V_FMA_F32,
 	V_FMA_F16,
 	V_PACK_B32_F16,
@@ -348,6 +361,10 @@ enum class Opcode {
 	V_CMP_NEQ_F32,
 	V_CMP_NLT_F32,
 	V_CMP_TRU_F32,
+	V_CMP_EQ_F64,
+	V_CMP_LE_F64,
+	V_CMPX_LE_F64,
+	V_CMPX_GE_F64,
 	V_CMPX_LT_F32,
 	V_CMPX_EQ_F32,
 	V_CMPX_LE_F32,
@@ -377,12 +394,19 @@ enum class Opcode {
 	V_CMP_GT_I16,
 	V_CMP_NE_I16,
 	V_CMP_GE_I16,
+	V_CMPX_LT_I16,
+	V_CMPX_EQ_I16,
+	V_CMPX_LE_I16,
+	V_CMPX_GT_I16,
+	V_CMPX_NE_I16,
+	V_CMPX_GE_I16,
 	V_CMP_LT_F16,
 	V_CMP_EQ_F16,
 	V_CMP_LE_F16,
 	V_CMP_GT_F16,
 	V_CMP_LG_F16,
 	V_CMP_GE_F16,
+	V_CMP_NGE_F16,
 	V_CMP_NGT_F16,
 	V_CMP_NEQ_F16,
 	V_CMP_NLT_F16,
@@ -392,6 +416,7 @@ enum class Opcode {
 	V_CMPX_GT_F16,
 	V_CMPX_GE_F16,
 	V_CMPX_NGT_F16,
+	V_CMPX_NLE_F16,
 	V_CMPX_NEQ_F16,
 	V_CMPX_NLT_F16,
 	V_CMPX_LT_I32,
@@ -418,11 +443,16 @@ enum class Opcode {
 	V_CMP_GE_U32,
 	V_CMP_T_U32,
 	V_CMP_EQ_I64,
+	V_CMP_LT_I64,
+	V_CMP_LE_I64,
 	V_CMP_LT_U64,
 	V_CMP_EQ_U64,
+	V_CMP_LE_U64,
 	V_CMP_GT_U64,
 	V_CMP_NE_U64,
+	V_CMP_GE_U64,
 	V_CMPX_NE_I64,
+	V_CMPX_LE_U64,
 	V_CMPX_NE_U64,
 	V_CMPX_LT_U32,
 	V_CMPX_EQ_U32,
@@ -482,6 +512,7 @@ enum class Opcode {
 	BUFFER_ATOMIC_SMAX,
 	BUFFER_ATOMIC_UMAX,
 	BUFFER_ATOMIC_AND,
+	BUFFER_ATOMIC_AND_X2,
 	BUFFER_ATOMIC_OR,
 	BUFFER_ATOMIC_OR_X2,
 	BUFFER_ATOMIC_XOR,
@@ -502,6 +533,7 @@ enum class Opcode {
 	FLAT_STORE_DWORDX3,
 	FLAT_STORE_DWORDX4,
 	DS_ADD_U32,
+	DS_ADD_U64,
 	DS_ADD_RTN_U32,
 	DS_SUB_U32,
 	DS_SUB_RTN_U32,
@@ -520,6 +552,7 @@ enum class Opcode {
 	DS_AND_B32,
 	DS_AND_RTN_B32,
 	DS_OR_B32,
+	DS_OR_B64,
 	DS_OR_RTN_B32,
 	DS_XOR_B32,
 	DS_XOR_RTN_B32,
@@ -558,6 +591,7 @@ enum class Opcode {
 	DS_WRITE_B128,
 	DS_WRITE_ADDTID_B32,
 	DS_READ_ADDTID_B32,
+	IMAGE_BVH_INTERSECT_RAY,
 	IMAGE_GET_RESINFO,
 	IMAGE_GET_LOD,
 	IMAGE_LOAD,
@@ -565,12 +599,17 @@ enum class Opcode {
 	IMAGE_STORE,
 	IMAGE_STORE_MIP,
 	IMAGE_ATOMIC_SWAP,
+	IMAGE_ATOMIC_CMPSWAP,
 	IMAGE_ATOMIC_ADD,
+	IMAGE_ATOMIC_SMIN,
 	IMAGE_ATOMIC_UMIN,
+	IMAGE_ATOMIC_SMAX,
 	IMAGE_ATOMIC_UMAX,
 	IMAGE_ATOMIC_AND,
 	IMAGE_ATOMIC_OR,
 	IMAGE_ATOMIC_XOR,
+	IMAGE_ATOMIC_FMIN,
+	IMAGE_ATOMIC_FMAX,
 	IMAGE_SAMPLE,
 	IMAGE_GATHER4_L,
 	IMAGE_GATHER4_LZ,
@@ -596,6 +635,7 @@ enum class Opcode {
 	S_CBRANCH_EXECZ,
 	S_CBRANCH_EXECNZ,
 	S_CBRANCH_CDBGSYS,
+	S_CBRANCH_CDBGSYS_OR_USER,
 	S_SENDMSG,
 	S_SETREG_B32,
 	S_SLEEP,
@@ -622,6 +662,8 @@ enum class OperandKind {
 	Scc,
 	M0,
 	PopsExitingWaveId,
+	SharedBase,
+	PrivateBase,
 	Null,
 	Vgpr,
 };
@@ -733,7 +775,6 @@ struct Instruction {
 struct Program {
 	std::span<const uint32_t> code;
 	std::vector<Instruction>  instructions;
-	bool                     has_bvh = false;
 };
 
 // Code spans are trusted to contain complete instructions, valid branch targets, and 32-bit PCs.

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

@@ -87,6 +87,11 @@ constexpr OpcodeMap VOP1_OPCODE_LIST[] = {
     {0x00u, Opcode::V_NOP},
     {0x01u, Opcode::V_MOV_B32},
     {0x02u, Opcode::V_READFIRSTLANE_B32},
+    {0x04u, Opcode::V_CVT_F64_I32},
+    {0x0fu, Opcode::V_CVT_F32_F64},
+    {0x10u, Opcode::V_CVT_F64_F32},
+    {0x16u, Opcode::V_CVT_F64_U32},
+    {0x2fu, Opcode::V_RCP_F64},
     {0x05u, Opcode::V_CVT_F32_I32},
     {0x06u, Opcode::V_CVT_F32_U32},
     {0x07u, Opcode::V_CVT_U32_F32},
@@ -146,6 +151,11 @@ constexpr OpcodeMap VOP3_ENCODED_VOP1_OPCODE_LIST[] = {
     {0x00u, Opcode::V_NOP},
     {0x01u, Opcode::V_MOV_B32},
     {0x02u, Opcode::V_READFIRSTLANE_B32},
+    {0x04u, Opcode::V_CVT_F64_I32},
+    {0x0fu, Opcode::V_CVT_F32_F64},
+    {0x10u, Opcode::V_CVT_F64_F32},
+    {0x16u, Opcode::V_CVT_F64_U32},
+    {0x2fu, Opcode::V_RCP_F64},
     {0x05u, Opcode::V_CVT_F32_I32},
     {0x06u, Opcode::V_CVT_F32_U32},
     {0x07u, Opcode::V_CVT_U32_F32},
@@ -218,6 +228,8 @@ constexpr VopcOpcodeInfo VOPC_OPCODE_LIST[] = {
     {0x19u, Opcode::V_CMPX_NGE_F32},       {0x1au, Opcode::V_CMPX_NLG_F32},
     {0x1bu, Opcode::V_CMPX_NGT_F32},       {0x1cu, Opcode::V_CMPX_NLE_F32},
     {0x1du, Opcode::V_CMPX_NEQ_F32},       {0x1eu, Opcode::V_CMPX_NLT_F32},
+    {0x22u, Opcode::V_CMP_EQ_F64, false},  {0x23u, Opcode::V_CMP_LE_F64, false},
+    {0x33u, Opcode::V_CMPX_LE_F64, false}, {0x36u, Opcode::V_CMPX_GE_F64, false},
     {0x80u, Opcode::V_CMP_F_I32},          {0x81u, Opcode::V_CMP_LT_I32},
     {0x82u, Opcode::V_CMP_EQ_I32},         {0x83u, Opcode::V_CMP_LE_I32},
     {0x84u, Opcode::V_CMP_GT_I32},         {0x85u, Opcode::V_CMP_NE_I32},
@@ -229,6 +241,9 @@ constexpr VopcOpcodeInfo VOPC_OPCODE_LIST[] = {
     {0x92u, Opcode::V_CMPX_EQ_I32},        {0x93u, Opcode::V_CMPX_LE_I32},
     {0x94u, Opcode::V_CMPX_GT_I32},        {0x95u, Opcode::V_CMPX_NE_I32},
     {0x96u, Opcode::V_CMPX_GE_I32},        {0x98u, Opcode::V_CMPX_CLASS_F32},
+    {0x99u, Opcode::V_CMPX_LT_I16, false}, {0x9au, Opcode::V_CMPX_EQ_I16, false},
+    {0x9bu, Opcode::V_CMPX_LE_I16, false}, {0x9cu, Opcode::V_CMPX_GT_I16, false},
+    {0x9du, Opcode::V_CMPX_NE_I16, false}, {0x9eu, Opcode::V_CMPX_GE_I16, false},
     {0xa9u, Opcode::V_CMP_LT_U16},         {0xaau, Opcode::V_CMP_EQ_U16},
     {0xabu, Opcode::V_CMP_LE_U16},         {0xacu, Opcode::V_CMP_GT_U16},
     {0xadu, Opcode::V_CMP_NE_U16},         {0xaeu, Opcode::V_CMP_GE_U16},
@@ -238,21 +253,25 @@ constexpr VopcOpcodeInfo VOPC_OPCODE_LIST[] = {
     {0xc3u, Opcode::V_CMP_LE_U32},         {0xc4u, Opcode::V_CMP_GT_U32},
     {0xc5u, Opcode::V_CMP_NE_U32},         {0xc6u, Opcode::V_CMP_GE_U32},
     {0xc7u, Opcode::V_CMP_T_U32},          {0xa2u, Opcode::V_CMP_EQ_I64, false},
+    {0xa1u, Opcode::V_CMP_LT_I64, false},  {0xa3u, Opcode::V_CMP_LE_I64, false},
     {0xb5u, Opcode::V_CMPX_NE_I64, false}, {0xd1u, Opcode::V_CMPX_LT_U32},
     {0xd2u, Opcode::V_CMPX_EQ_U32},        {0xd3u, Opcode::V_CMPX_LE_U32},
     {0xd4u, Opcode::V_CMPX_GT_U32},        {0xd5u, Opcode::V_CMPX_NE_U32},
     {0xd6u, Opcode::V_CMPX_GE_U32},        {0xe1u, Opcode::V_CMP_LT_U64, false},
-    {0xe2u, Opcode::V_CMP_EQ_U64, false},  {0xe4u, Opcode::V_CMP_GT_U64, false},
-    {0xe5u, Opcode::V_CMP_NE_U64, false},
+    {0xe2u, Opcode::V_CMP_EQ_U64, false},  {0xe3u, Opcode::V_CMP_LE_U64, false},
+    {0xe4u, Opcode::V_CMP_GT_U64, false},  {0xe5u, Opcode::V_CMP_NE_U64, false},
+    {0xe6u, Opcode::V_CMP_GE_U64, false},  {0xf3u, Opcode::V_CMPX_LE_U64, false},
     {0xf5u, Opcode::V_CMPX_NE_U64, false}, {0xc9u, Opcode::V_CMP_LT_F16},
     {0xcau, Opcode::V_CMP_EQ_F16},         {0xcbu, Opcode::V_CMP_LE_F16},
     {0xccu, Opcode::V_CMP_GT_F16},         {0xcdu, Opcode::V_CMP_LG_F16},
-    {0xceu, Opcode::V_CMP_GE_F16},         {0xebu, Opcode::V_CMP_NGT_F16},
+    {0xceu, Opcode::V_CMP_GE_F16},         {0xe9u, Opcode::V_CMP_NGE_F16},
+    {0xebu, Opcode::V_CMP_NGT_F16},
     {0xedu, Opcode::V_CMP_NEQ_F16},        {0xeeu, Opcode::V_CMP_NLT_F16},
     {0xd9u, Opcode::V_CMPX_LT_F16},        {0xdau, Opcode::V_CMPX_EQ_F16},
     {0xdbu, Opcode::V_CMPX_LE_F16},        {0xdcu, Opcode::V_CMPX_GT_F16},
     {0xdeu, Opcode::V_CMPX_GE_F16},        {0xfbu, Opcode::V_CMPX_NGT_F16},
-    {0xfdu, Opcode::V_CMPX_NEQ_F16},       {0xfeu, Opcode::V_CMPX_NLT_F16},
+    {0xfcu, Opcode::V_CMPX_NLE_F16},       {0xfdu, Opcode::V_CMPX_NEQ_F16},
+    {0xfeu, Opcode::V_CMPX_NLT_F16},
 };
 
 constexpr auto VOPC_OPS = Detail::MakeOpcodeTable<0x100>(VOPC_OPCODE_LIST);
@@ -267,6 +286,9 @@ constexpr OpcodeMap VOP3_OPCODE_LIST[] = {
     {0x146u, Opcode::V_CUBETC_F32},
     {0x147u, Opcode::V_CUBEMA_F32},
     {0x14bu, Opcode::V_FMA_F32},
+    {0x14cu, Opcode::V_FMA_F64},
+    {0x164u, Opcode::V_ADD_F64},
+    {0x165u, Opcode::V_MUL_F64},
     {0x148u, Opcode::V_BFE_U32},
     {0x149u, Opcode::V_BFE_I32},
     {0x14au, Opcode::V_BFI_B32},
@@ -452,6 +474,9 @@ bool IsVopcCompareExec(Opcode opcode);
 
 bool IsVop1FloatSourceOpcode(Opcode opcode) {
 	switch (opcode) {
+		case Opcode::V_CVT_F32_F64:
+		case Opcode::V_CVT_F64_F32:
+		case Opcode::V_RCP_F64:
 		case Opcode::V_MOV_B32:
 		case Opcode::V_CVT_F32_F16:
 		case Opcode::V_CVT_U32_F32:
@@ -527,6 +552,7 @@ constexpr Vop1SdwaRule VOP1_SDWA_RULES[] = {
     {Opcode::V_NOT_B32, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), SdwaSelBytes() | SdwaSelWords(),
      SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), false},
     {Opcode::V_FFBL_B32, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), 0, 0, false},
+    {Opcode::V_FFBH_U32, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), 0, 0, false},
     {Opcode::V_CVT_F32_F16, SdwaSelWords() | SdwaSelFull(), 0, 0, true},
     {Opcode::V_CVT_F16_F32, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), SdwaSelWords(),
      SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), true},
@@ -709,6 +735,12 @@ void ApplyDppModifier(Operand& operand, uint32_t modifier, uint32_t encoding) {
 
 void DecodeVop1Dpp(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index,
                    uint32_t opcode, uint32_t vdst, Instruction& inst) {
+	if (inst.opcode == Opcode::V_CVT_F64_I32 || inst.opcode == Opcode::V_CVT_F32_F64 ||
+	    inst.opcode == Opcode::V_CVT_F64_F32 || inst.opcode == Opcode::V_CVT_F64_U32 ||
+	    inst.opcode == Opcode::V_RCP_F64) {
+		SetUnsupported(inst, Family::VOP1, opcode, "FP64 instructions do not support DPP");
+		return;
+	}
 	const auto modifier = code[word_index + 1u];
 	const auto src0     = modifier & 0xffu;
 	SetRawWords(inst, code, word_index, 2);
@@ -758,6 +790,7 @@ bool IsVop2FloatOpcode(Opcode opcode) {
 
 bool IsVop1FloatResultOpcode(Opcode opcode) {
 	switch (opcode) {
+		case Opcode::V_CVT_F32_F64:
 		case Opcode::V_CVT_F32_I32:
 		case Opcode::V_CVT_F32_U32:
 		case Opcode::V_CVT_F32_F16:
@@ -804,6 +837,10 @@ bool IsVopcFloatCompareOpcode(Opcode opcode) {
 		case Opcode::V_CMP_F_F32:
 		case Opcode::V_CMP_LT_F32:
 		case Opcode::V_CMP_EQ_F32:
+		case Opcode::V_CMP_EQ_F64:
+		case Opcode::V_CMP_LE_F64:
+		case Opcode::V_CMPX_LE_F64:
+		case Opcode::V_CMPX_GE_F64:
 		case Opcode::V_CMP_LE_F32:
 		case Opcode::V_CMP_GT_F32:
 		case Opcode::V_CMP_LG_F32:
@@ -836,6 +873,7 @@ bool IsVopcFloatCompareOpcode(Opcode opcode) {
 		case Opcode::V_CMP_GT_F16:
 		case Opcode::V_CMP_LG_F16:
 		case Opcode::V_CMP_GE_F16:
+		case Opcode::V_CMP_NGE_F16:
 		case Opcode::V_CMP_NGT_F16:
 		case Opcode::V_CMP_NEQ_F16:
 		case Opcode::V_CMP_NLT_F16:
@@ -845,6 +883,7 @@ bool IsVopcFloatCompareOpcode(Opcode opcode) {
 		case Opcode::V_CMPX_GT_F16:
 		case Opcode::V_CMPX_GE_F16:
 		case Opcode::V_CMPX_NGT_F16:
+		case Opcode::V_CMPX_NLE_F16:
 		case Opcode::V_CMPX_NEQ_F16:
 		case Opcode::V_CMPX_NLT_F16:
 		case Opcode::V_CMP_CLASS_F32:
@@ -1113,7 +1152,9 @@ VopcSdwaFields DecodeVopcSdwaFields(uint32_t modifier) {
 }
 
 bool SupportsVopcSdwa(Opcode opcode) {
-	return opcode != Opcode::UNSUPPORTED;
+	return opcode != Opcode::UNSUPPORTED && opcode != Opcode::V_CMP_EQ_F64 &&
+	       opcode != Opcode::V_CMP_LE_F64 && opcode != Opcode::V_CMPX_LE_F64 &&
+	       opcode != Opcode::V_CMPX_GE_F64;
 }
 
 void DecodeVopcSdwa(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index,
@@ -1178,6 +1219,8 @@ void DecodeVopcDpp(uint32_t pc, std::span<const uint32_t> code, uint32_t word_in
 
 uint32_t NativeVop3SourceCount(Opcode opcode) {
 	switch (opcode) {
+		case Opcode::V_ADD_F64:
+		case Opcode::V_MUL_F64:
 		case Opcode::V_MUL_LO_U32:
 		case Opcode::V_MUL_HI_U32:
 		case Opcode::V_MUL_LO_I32:
@@ -1328,6 +1371,9 @@ bool SupportsNativeVop3SourceModifiers(Opcode opcode) {
 		case Opcode::V_MAX_F32:
 		case Opcode::V_MAC_F32:
 		case Opcode::V_MAD_F32:
+		case Opcode::V_ADD_F64:
+		case Opcode::V_MUL_F64:
+		case Opcode::V_FMA_F64:
 		case Opcode::V_FMA_F32:
 		case Opcode::V_PACK_B32_F16:
 		case Opcode::V_CUBEID_F32:
@@ -1360,6 +1406,7 @@ bool SupportsNativeVop3ResultModifiers(Opcode opcode) {
 		case Opcode::V_MAD_F32:
 		case Opcode::V_FMA_F32:
 		case Opcode::V_FMA_F16:
+		case Opcode::V_CUBEID_F32:
 		case Opcode::V_CVT_PKRTZ_F16_F32:
 		case Opcode::V_LDEXP_F32:
 		case Opcode::V_MIN3_F32:
@@ -1430,6 +1477,8 @@ void ApplyNativeVop3SourceModifiers(Instruction& inst, uint32_t abs, uint32_t ne
 
 bool IsVopcCompareExec(Opcode opcode) {
 	switch (opcode) {
+		case Opcode::V_CMPX_LE_F64:
+		case Opcode::V_CMPX_GE_F64:
 		case Opcode::V_CMPX_LT_F32:
 		case Opcode::V_CMPX_EQ_F32:
 		case Opcode::V_CMPX_LE_F32:
@@ -1457,7 +1506,14 @@ bool IsVopcCompareExec(Opcode opcode) {
 		case Opcode::V_CMPX_NE_U32:
 		case Opcode::V_CMPX_GE_U32:
 		case Opcode::V_CMPX_NE_I64:
+		case Opcode::V_CMPX_LE_U64:
 		case Opcode::V_CMPX_NE_U64:
+		case Opcode::V_CMPX_LT_I16:
+		case Opcode::V_CMPX_EQ_I16:
+		case Opcode::V_CMPX_LE_I16:
+		case Opcode::V_CMPX_GT_I16:
+		case Opcode::V_CMPX_NE_I16:
+		case Opcode::V_CMPX_GE_I16:
 		case Opcode::V_CMPX_LT_U16:
 		case Opcode::V_CMPX_EQ_U16:
 		case Opcode::V_CMPX_GT_U16:
@@ -1467,6 +1523,7 @@ bool IsVopcCompareExec(Opcode opcode) {
 		case Opcode::V_CMPX_GT_F16:
 		case Opcode::V_CMPX_GE_F16:
 		case Opcode::V_CMPX_NGT_F16:
+		case Opcode::V_CMPX_NLE_F16:
 		case Opcode::V_CMPX_NEQ_F16:
 		case Opcode::V_CMPX_NLT_F16: return true;
 		default: return false;

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

@@ -1,5 +1,7 @@
 #include "graphics/shader/recompiler/frontend/translate/Translator.h"
 
+#include <bit>
+
 namespace Libs::Graphics::ShaderRecompiler::Frontend {
 
 void Translator::EmitCompareResult(const Decoder::Instruction& inst, IR::U1 value, bool scalar,
@@ -26,9 +28,20 @@ void Translator::EmitCompareConstant(const Decoder::Instruction& inst, bool valu
 
 void Translator::EmitIntegerCompare(const Decoder::Instruction& inst, IR::ValueOpcode opcode,
                                     IR::Type type, bool scalar, bool cmpx) {
-	const auto lhs = ReadOperand(inst.src0, type);
-	const auto rhs = ReadOperand(inst.src1, type);
-	EmitCompareResult(inst, IR::U1(ir.Emit(opcode, {lhs, rhs})), scalar, cmpx);
+	const bool signed_64 = opcode == IR::ValueOpcode::SLessThan64 ||
+	                       opcode == IR::ValueOpcode::SLessThanEqual64 ||
+	                       inst.opcode == Decoder::Opcode::V_CMP_EQ_I64 ||
+	                       inst.opcode == Decoder::Opcode::V_CMPX_NE_I64;
+	const auto read = [&](const Decoder::Operand& operand) {
+		// RDNA2 expands signed 64-bit integer literals by sign extension.
+		if (signed_64 && operand.kind == Decoder::OperandKind::LiteralConstant) {
+			return IR::Value(static_cast<uint64_t>(
+			    static_cast<int64_t>(std::bit_cast<int32_t>(operand.value))));
+		}
+		return ReadOperand(operand, type);
+	};
+	EmitCompareResult(inst, IR::U1(ir.Emit(opcode, {read(inst.src0), read(inst.src1)})), scalar,
+	                  cmpx);
 }
 
 void Translator::EmitInteger16Compare(const Decoder::Instruction& inst, IR::ValueOpcode opcode,
@@ -40,11 +53,14 @@ void Translator::EmitInteger16Compare(const Decoder::Instruction& inst, IR::Valu
 
 void Translator::EmitFloatCompare(const Decoder::Instruction& inst, IR::ValueOpcode opcode,
                                   bool half, bool cmpx) {
+	const auto type = IR::ArgTypeOf(opcode, 0);
 	const auto lhs =
-	    half ? IR::Value(ReadF16AsF32(inst.src0)) : ReadOperand(inst.src0, IR::Type::F32);
+	    half ? IR::Value(ReadF16AsF32(inst.src0)) : ReadOperand(inst.src0, type);
 	const auto rhs =
-	    half ? IR::Value(ReadF16AsF32(inst.src1)) : ReadOperand(inst.src1, IR::Type::F32);
-	EmitCompareResult(inst, IR::U1(ir.Emit(opcode, {lhs, rhs})), false, cmpx);
+	    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);
 }
 
 void Translator::EmitFloatOrderedCompare(const Decoder::Instruction& inst, bool ordered, bool cmpx) {

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

@@ -223,7 +223,7 @@ Decoder::Operand MemorySourceAt(const Decoder::Instruction& decoded, uint32_t in
 		const bool store_or_atomic = decoded.opcode == Decoder::Opcode::IMAGE_STORE ||
 		                             decoded.opcode == Decoder::Opcode::IMAGE_STORE_MIP ||
 		                             (decoded.opcode >= Decoder::Opcode::IMAGE_ATOMIC_SWAP &&
-		                              decoded.opcode <= Decoder::Opcode::IMAGE_ATOMIC_XOR);
+		                              decoded.opcode <= Decoder::Opcode::IMAGE_ATOMIC_FMAX);
 		if (store_or_atomic) {
 			return index == 0u ? decoded.dst : decoded.src0;
 		}
@@ -235,6 +235,7 @@ Decoder::Operand MemorySourceAt(const Decoder::Instruction& decoded, uint32_t in
 				return index == 0u ? decoded.src0 : MakeImmediate(decoded.offset & 0xffffu);
 			case Decoder::Opcode::DS_CONSUME:
 			case Decoder::Opcode::DS_APPEND:
+				return decoded.gds ? MakeM0Operand() : MakeImmediate(0);
 			case Decoder::Opcode::DS_READ_ADDTID_B32: return MakeM0Operand();
 			case Decoder::Opcode::DS_WRITE_ADDTID_B32:
 				return index == 0u ? decoded.src1 : MakeM0Operand();
@@ -557,9 +558,18 @@ void Translator::IMAGE_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode
 	const auto memory   = MemoryInfoFromDecoded(inst);
 	const auto resource = GetImageResource(memory);
 	const auto address  = MakeImageAddress(inst, MemorySourceAt(inst, 1));
-	const auto result =
-	    ir.Emit(opcode, {resource, address, ReadU32(MemorySourceAt(inst, 0)), ir.GetExec()},
-	            AddMemoryInfo(memory, inst.pc));
+	const auto data     = MemorySourceAt(inst, 0);
+	const auto flags    = AddMemoryInfo(memory, inst.pc);
+	IR::Value  result;
+	if (opcode == IR::ValueOpcode::ImageAtomicCompareSwap32) {
+		// VDATA supplies the replacement; VDATA+1 supplies the comparison.
+		result = ir.Emit(opcode,
+		                 {resource, address, ReadU32(data), ReadU32(OffsetOperand(data, 1u)),
+		                  ir.GetExec()},
+		                 flags);
+	} else {
+		result = ir.Emit(opcode, {resource, address, ReadU32(data), ir.GetExec()}, flags);
+	}
 	if (inst.glc) {
 		WriteOperand(inst.dst, result);
 	}
@@ -569,7 +579,10 @@ void Translator::DS_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode opc
                            bool returns_value) {
 	const auto memory  = MemoryInfoFromDecoded(inst);
 	const auto address = ReadU32(MemorySourceAt(inst, 1));
-	const auto result  = ir.Emit(opcode, {address, ReadU32(MemorySourceAt(inst, 0)), ir.GetExec()},
+	const auto data_src = MemorySourceAt(inst, 0);
+	const IR::Value value =
+	    memory.data_dwords == 2u ? IR::Value(ReadU64(data_src)) : IR::Value(ReadU32(data_src));
+	const auto result  = ir.Emit(opcode, {address, value, ir.GetExec()},
 	                             AddMemoryInfo(memory, inst.pc));
 	if (returns_value) {
 		WriteOperand(inst.dst, result);
@@ -845,8 +858,7 @@ void Translator::DS_MINMAX_F32(const Decoder::Instruction& inst, IR::ValueOpcode
 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(), ir.GetExecLo(),
-	                                ir.GetExecHi()},
+	                               {ReadU32(MemorySourceAt(inst, 0)), ir.GetExec()},
 	                               AddMemoryInfo(memory, inst.pc)));
 }
 
@@ -880,6 +892,15 @@ void Translator::DS_BPERMUTE_B32(const Decoder::Instruction& inst) {
 
 void Translator::EmitMemory(const Decoder::Instruction& inst) {
 	switch (inst.opcode) {
+		case Decoder::Opcode::IMAGE_BVH_INTERSECT_RAY: {
+			const auto result = ir.Emit(IR::ValueOpcode::BvhIntersect,
+			    {ConstructU32x4(inst.src1, 4), MakeImageAddress(inst, inst.src0), ir.GetExec()});
+			for (uint32_t component = 0; component < 4; ++component) {
+				WriteOperand(OffsetOperand(inst.dst, component),
+				    ir.Emit(IR::ValueOpcode::CompositeExtractU32x4, {result, IR::Value(component)}));
+			}
+			return;
+		}
 		case Decoder::Opcode::S_LOAD_DWORD:
 		case Decoder::Opcode::S_LOAD_DWORDX2:
 		case Decoder::Opcode::S_LOAD_DWORDX4:
@@ -954,6 +975,8 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 			return BUFFER_ATOMIC(inst, IR::ValueOpcode::BufferAtomicUMax32);
 		case Decoder::Opcode::BUFFER_ATOMIC_AND:
 			return BUFFER_ATOMIC(inst, IR::ValueOpcode::BufferAtomicAnd32);
+		case Decoder::Opcode::BUFFER_ATOMIC_AND_X2:
+			return BUFFER_ATOMIC(inst, IR::ValueOpcode::BufferAtomicAnd64);
 		case Decoder::Opcode::BUFFER_ATOMIC_OR:
 			return BUFFER_ATOMIC(inst, IR::ValueOpcode::BufferAtomicOr32);
 		case Decoder::Opcode::BUFFER_ATOMIC_OR_X2:
@@ -967,6 +990,8 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 
 		case Decoder::Opcode::DS_ADD_U32:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicIAdd32, false);
+		case Decoder::Opcode::DS_ADD_U64:
+			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicIAdd64, false);
 		case Decoder::Opcode::DS_ADD_RTN_U32:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicIAdd32, true);
 		case Decoder::Opcode::DS_SUB_U32:
@@ -1003,6 +1028,8 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicAnd32, true);
 		case Decoder::Opcode::DS_OR_B32:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicOr32, false);
+		case Decoder::Opcode::DS_OR_B64:
+			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicOr64, false);
 		case Decoder::Opcode::DS_OR_RTN_B32:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicOr32, true);
 		case Decoder::Opcode::DS_XOR_B32:
@@ -1012,12 +1039,18 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::DS_WRXCHG_RTN_B32:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicSwap32, true);
 
+		case Decoder::Opcode::IMAGE_ATOMIC_CMPSWAP:
+			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicCompareSwap32);
 		case Decoder::Opcode::IMAGE_ATOMIC_SWAP:
 			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicSwap32);
 		case Decoder::Opcode::IMAGE_ATOMIC_ADD:
 			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicIAdd32);
+		case Decoder::Opcode::IMAGE_ATOMIC_SMIN:
+			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicSMin32);
 		case Decoder::Opcode::IMAGE_ATOMIC_UMIN:
 			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicUMin32);
+		case Decoder::Opcode::IMAGE_ATOMIC_SMAX:
+			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicSMax32);
 		case Decoder::Opcode::IMAGE_ATOMIC_UMAX:
 			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicUMax32);
 		case Decoder::Opcode::IMAGE_ATOMIC_AND:
@@ -1026,6 +1059,10 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicOr32);
 		case Decoder::Opcode::IMAGE_ATOMIC_XOR:
 			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicXor32);
+		case Decoder::Opcode::IMAGE_ATOMIC_FMIN:
+			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicFMin32);
+		case Decoder::Opcode::IMAGE_ATOMIC_FMAX:
+			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicFMax32);
 
 		case Decoder::Opcode::FLAT_LOAD_UBYTE:
 		case Decoder::Opcode::FLAT_LOAD_SBYTE:

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

@@ -155,6 +155,12 @@ void Translator::EmitScalar(const Decoder::Instruction& inst) {
 		case O::S_BREV_B32:
 			return SimpleInteger(inst, IR::ValueOpcode::BitReverse32, IR::Type::U32, false, false,
 			                     false);
+		case O::S_BREV_B64: {
+			const auto value = ReadU32Pair(inst.src0);
+			WriteU32Pair(inst.dst, {IR::U32(ir.Emit(IR::ValueOpcode::BitReverse32, {value[1]})),
+			                        IR::U32(ir.Emit(IR::ValueOpcode::BitReverse32, {value[0]}))});
+			return;
+		}
 		case O::S_BCNT1_I32_B32:
 			return SimpleInteger(inst, IR::ValueOpcode::BitCount32, IR::Type::U32, false, false,
 			                     true);
@@ -231,6 +237,7 @@ void Translator::EmitScalar(const Decoder::Instruction& inst) {
 		case O::S_CBRANCH_EXECZ:
 		case O::S_CBRANCH_EXECNZ:
 		case O::S_CBRANCH_CDBGSYS:
+		case O::S_CBRANCH_CDBGSYS_OR_USER:
 		case O::S_ENDPGM: return;
 		default: return FailMissingTranslation(inst);
 	}

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

@@ -4,6 +4,7 @@
 
 #include <algorithm>
 #include <array>
+#include <bit>
 #include <unordered_map>
 #include <utility>
 
@@ -134,6 +135,8 @@ IR::U32 Translator::ReadRawU32(const Decoder::Operand& operand) {
 		case Decoder::OperandKind::IntegerInlineConstant:
 		case Decoder::OperandKind::FloatInlineConstant: return IR::U32(IR::Value(operand.value));
 		case Decoder::OperandKind::Null:
+		case Decoder::OperandKind::SharedBase:
+		case Decoder::OperandKind::PrivateBase:
 		case Decoder::OperandKind::PopsExitingWaveId: return IR::U32(IR::Value(0u));
 		case Decoder::OperandKind::Sgpr:
 			return ir.GetScalarReg(static_cast<IR::ScalarReg>(operand.reg));
@@ -225,9 +228,28 @@ IR::Value Translator::ReadOperand(const Decoder::Operand& operand, IR::Type type
 		}
 		return ir.INotEqual(ReadRawU32(operand), IR::U32(IR::Value(0u)));
 	}
-	if (type == IR::Type::U64) {
-		const auto pair = ReadU32Pair(operand);
-		return ir.ConstructU64(pair[0], pair[1]);
+	if (type == IR::Type::U64 || type == IR::Type::F64) {
+		auto pair = ReadU32Pair(operand);
+		if (type == IR::Type::F64) {
+			if (operand.kind == Decoder::OperandKind::LiteralConstant) {
+				pair = {IR::U32(IR::Value(0u)), IR::U32(IR::Value(operand.value))};
+			} else if (operand.kind == Decoder::OperandKind::FloatInlineConstant) {
+				const auto bits = operand.value == 0x3e22f983u
+				                      ? 0x3fc45f306dc9c882ull
+				                      : std::bit_cast<uint64_t>(static_cast<double>(
+				                            std::bit_cast<float>(operand.value)));
+				pair            = {IR::U32(IR::Value(static_cast<uint32_t>(bits))),
+				                   IR::U32(IR::Value(static_cast<uint32_t>(bits >> 32u)))};
+			}
+			if (operand.absolute) {
+				pair[1] = ir.BitwiseAnd(pair[1], IR::U32(IR::Value(0x7fffffffu)));
+			}
+			if (operand.negate) {
+				pair[1] = ir.BitwiseXor(pair[1], IR::U32(IR::Value(0x80000000u)));
+			}
+		}
+		const auto bits = ir.ConstructU64(pair[0], pair[1]);
+		return type == IR::Type::F64 ? ir.Emit(IR::ValueOpcode::BitCastF64U64, {bits}) : IR::Value(bits);
 	}
 	auto bits = ApplyBitSourceModifiers(operand, ReadRawU32(operand));
 	if (TypesOverlap(type, IR::Type::F32) && !TypesOverlap(type, IR::Type::U32)) {
@@ -297,6 +319,7 @@ void Translator::WriteRawU32(const Decoder::Operand& operand, IR::U32 value) {
 		case Decoder::OperandKind::Sgpr: {
 			const auto reg = static_cast<IR::ScalarReg>(operand.reg);
 			ir.SetScalarReg(reg, value);
+			program.scalar_writes.push_back({current_pc, reg});
 			ir.SetScalarMaskTag(reg, IR::U1(IR::Value(false)));
 			if (IR::RegIndex(reg) > 0u) {
 				ir.SetScalarMaskTag(static_cast<IR::ScalarReg>(IR::RegIndex(reg) - 1u),
@@ -403,6 +426,10 @@ void Translator::WriteOperand(const Decoder::Operand& operand, IR::Value value)
 		Write16Bits(operand, IR::U32(ir.Emit(IR::ValueOpcode::ConvertU32U16, {bits})));
 		return;
 	}
+	if (type == IR::Type::F64) {
+		value = ir.Emit(IR::ValueOpcode::BitCastU64F64, {value});
+		type  = IR::Type::U64;
+	}
 	if (type == IR::Type::U64) {
 		WriteU32Pair(operand, {ir.CompositeExtract(value, 0), ir.CompositeExtract(value, 1)});
 		return;
@@ -449,6 +476,12 @@ 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()};
 	}
@@ -535,6 +568,12 @@ IR::U32 Translator::ReadU16LaneRaw(const Decoder::Operand& operand, bool high_la
 	} else if (operand.sdwa_sel == 4u || operand.sdwa_sel == 5u) {
 		offset = operand.sdwa_sel == 5u ? 16u : 0u;
 	}
+	if (width == 8u && operand.sdwa_sext) {
+		// Extend the selected byte into the 16-bit operand before halfword modifiers.
+		const auto value = IR::U32(
+		    ir.Emit(IR::ValueOpcode::BitFieldSExtract, {bits, IR::Value(offset), IR::Value(width)}));
+		return ir.BitwiseAnd(value, IR::U32(IR::Value(0xffffu)));
+	}
 	return IR::U32(
 	    ir.Emit(IR::ValueOpcode::BitFieldUExtract, {bits, IR::Value(offset), IR::Value(width)}));
 }
@@ -542,7 +581,7 @@ IR::U32 Translator::ReadU16LaneRaw(const Decoder::Operand& operand, bool high_la
 IR::U32 Translator::ReadU16LaneAsU32(const Decoder::Operand& operand, bool high_lane,
                                      bool sign_extend) {
 	auto value = Read16LaneBits(operand, high_lane);
-	if (sign_extend || operand.sdwa_sext) {
+	if (sign_extend) {
 		value = IR::U32(
 		    ir.Emit(IR::ValueOpcode::BitFieldSExtract, {value, IR::Value(0u), IR::Value(16u)}));
 	}
@@ -675,11 +714,13 @@ std::array<IR::U32, 2> Translator::WriteMask(const Decoder::Operand& operand, IR
 				                    IR::U1(IR::Value(false)));
 			}
 			ir.SetScalarReg(reg, mask[0]);
+			program.scalar_writes.push_back({current_pc, reg});
 			// A wave32 VALU mask destination must not overwrite the neighboring SGPR.
 			if ((write_64 || program.wave_size == 64u) &&
 			    IR::RegIndex(reg) + 1u < IR::NumScalarRegs) {
 				const auto high = static_cast<IR::ScalarReg>(IR::RegIndex(reg) + 1u);
 				ir.SetScalarReg(high, mask[1]);
+				program.scalar_writes.push_back({current_pc, high});
 				ir.SetThreadBitScalarReg(high, IR::U1(IR::Value(false)));
 				ir.SetScalarMaskTag(high, IR::U1(IR::Value(false)));
 			}
@@ -722,55 +763,61 @@ void Translator::WriteCompareResult(const Decoder::Operand& operand, IR::U1 valu
 	WriteMask(operand, ir.LogicalAnd(ir.GetExec(), value));
 }
 
-void Translator::AddBranchCondition(const CFG::BasicBlock& source, IR::BlockInfo& info) {
-	if (source.terminator.goto_value >= 0) {
-		if (source.terminator.goto_variable == UINT32_MAX) {
-			EXIT("block %u sets an invalid goto variable", source.id);
+void Translator::AddBranchCondition(const CFG::Graph& graph, const CFG::BasicBlock& source,
+                                    IR::BlockInfo& info) {
+	const auto native_condition = [&](CFG::BranchCondition kind) -> IR::U1 {
+		IR::U1 condition;
+		switch (kind) {
+			case CFG::BranchCondition::Always: condition = IR::U1(IR::Value(true)); break;
+			case CFG::BranchCondition::SccZero: condition = ir.LogicalNot(ir.GetScc()); break;
+			case CFG::BranchCondition::SccNonZero: condition = ir.GetScc(); break;
+			case CFG::BranchCondition::VccZero: condition = ir.LogicalNot(ir.GetVcc()); break;
+			case CFG::BranchCondition::VccNonZero: condition = ir.GetVcc(); break;
+			case CFG::BranchCondition::ExecZero: condition = ir.LogicalNot(ir.GetExec()); break;
+			case CFG::BranchCondition::ExecNonZero: condition = ir.GetExec(); break;
+			case CFG::BranchCondition::ScalarInstruction:
+				EXIT_IF(instruction_branch_condition.IsEmpty());
+				condition = instruction_branch_condition;
+				break;
+			default: EXIT("block %u has an invalid native branch condition", source.id);
 		}
-		ir.SetGotoVariable(source.terminator.goto_variable,
-		                   IR::U1(IR::Value(source.terminator.goto_value != 0)));
+		return IR::U1(ir.Emit(IR::ValueOpcode::ConditionRef, {condition}, kind));
+	};
+	const auto expression = [&](auto&& self, uint32_t index) -> IR::U1 {
+		const auto& value = graph.expressions.at(index);
+		switch (value.op) {
+			case CFG::ConditionExpression::Op::Constant: return IR::U1(IR::Value(value.lhs != 0));
+			case CFG::ConditionExpression::Op::Variable: return ir.GetGotoVariable(value.lhs);
+			case CFG::ConditionExpression::Op::Native:
+				return native_condition(static_cast<CFG::BranchCondition>(value.lhs));
+			case CFG::ConditionExpression::Op::Not: return ir.LogicalNot(self(self, value.lhs));
+			case CFG::ConditionExpression::Op::Or:
+				return ir.LogicalOr(self(self, value.lhs), self(self, value.rhs));
+		}
+		EXIT("invalid CFG condition expression");
+	};
+	for (const auto& assignment: source.assignments) {
+		ir.SetGotoVariable(assignment.variable, expression(expression, assignment.expression));
 	}
-	if (source.terminator.kind == CFG::TerminatorKind::IndirectBranch) {
-		if (source.terminator.indirect_selector_code != UINT32_MAX) {
-			info.indirect_target = ReadScalarCode(source.terminator.indirect_selector_code);
-		} else if (source.terminator.indirect_pc_sgpr != UINT32_MAX) {
-			info.indirect_target =
-			    ir.GetScalarReg(static_cast<IR::ScalarReg>(source.terminator.indirect_pc_sgpr));
+	const auto& term = source.terminator;
+	if (term.kind == CFG::TerminatorKind::IndirectBranch) {
+		if (term.indirect_selector_code != UINT32_MAX) {
+			info.indirect_target = ReadScalarCode(term.indirect_selector_code);
+		} else if (term.indirect_pc_sgpr != UINT32_MAX) {
+			info.indirect_target = ir.GetScalarReg(static_cast<IR::ScalarReg>(term.indirect_pc_sgpr));
 		} else {
 			EXIT("block %u has no indirect branch selector", source.id);
 		}
 		ir.Emit(IR::ValueOpcode::ReferenceU32, {info.indirect_target});
 		return;
 	}
-	if (source.terminator.kind != CFG::TerminatorKind::ConditionalBranch) {
-		return;
+	if (term.kind == CFG::TerminatorKind::ConditionalBranch) {
+		const auto condition = term.expression != UINT32_MAX
+		                           ? expression(expression, term.expression)
+		                           : native_condition(term.condition);
+		info.condition = condition;
+		ir.Emit(IR::ValueOpcode::Reference, {condition});
 	}
-	// EXEC and VCC are invocation-local Boolean masks. Branching on that Boolean lets inactive
-	// invocations leave the region without reconstructing a host-subgroup mask.
-	IR::U1 condition;
-	switch (source.terminator.condition) {
-		case CFG::BranchCondition::Always: condition = IR::U1(IR::Value(true)); break;
-		case CFG::BranchCondition::SccZero: condition = ir.LogicalNot(ir.GetScc()); break;
-		case CFG::BranchCondition::SccNonZero: condition = ir.GetScc(); break;
-		case CFG::BranchCondition::VccZero: condition = ir.LogicalNot(ir.GetVcc()); break;
-		case CFG::BranchCondition::VccNonZero: condition = ir.GetVcc(); break;
-		case CFG::BranchCondition::ExecZero: condition = ir.LogicalNot(ir.GetExec()); break;
-		case CFG::BranchCondition::ExecNonZero: condition = ir.GetExec(); break;
-		case CFG::BranchCondition::ScalarInstruction:
-			EXIT_IF(instruction_branch_condition.IsEmpty());
-			condition = instruction_branch_condition;
-			break;
-		case CFG::BranchCondition::GotoVariable:
-			if (source.terminator.goto_variable == UINT32_MAX) {
-				EXIT("block %u reads an invalid goto variable", source.id);
-			}
-			condition = ir.GetGotoVariable(source.terminator.goto_variable);
-			break;
-		case CFG::BranchCondition::Unknown:
-			EXIT("block %u has an unknown branch condition", source.id);
-	}
-	info.condition = condition;
-	ir.Emit(IR::ValueOpcode::Reference, {condition});
 }
 
 namespace {
@@ -842,8 +889,40 @@ void IncludeInstructionVectorRegisters(const Decoder::Instruction& inst, uint32_
 	include_vector(inst.src1);
 	include_vector(inst.src2);
 	include_vector(inst.src3);
+	switch (inst.opcode) {
+		case Decoder::Opcode::V_CVT_F64_I32:
+		case Decoder::Opcode::V_CVT_F64_F32:
+		case Decoder::Opcode::V_CVT_F64_U32: include_vector(inst.dst, 2u); break;
+		case Decoder::Opcode::V_FMA_F64: include_vector(inst.src2, 2u); [[fallthrough]];
+		case Decoder::Opcode::V_ADD_F64:
+		case Decoder::Opcode::V_MUL_F64: include_vector(inst.src1, 2u); [[fallthrough]];
+		case Decoder::Opcode::V_RCP_F64: include_vector(inst.dst, 2u); [[fallthrough]];
+		case Decoder::Opcode::V_CVT_F32_F64: include_vector(inst.src0, 2u); break;
+		case Decoder::Opcode::V_CMP_EQ_F64:
+		case Decoder::Opcode::V_CMP_LE_F64:
+		case Decoder::Opcode::V_CMPX_LE_F64:
+		case Decoder::Opcode::V_CMPX_GE_F64:
+		case Decoder::Opcode::V_CMP_EQ_I64:
+		case Decoder::Opcode::V_CMP_LT_I64:
+		case Decoder::Opcode::V_CMP_LE_I64:
+		case Decoder::Opcode::V_CMP_LT_U64:
+		case Decoder::Opcode::V_CMP_EQ_U64:
+		case Decoder::Opcode::V_CMP_LE_U64:
+		case Decoder::Opcode::V_CMP_GT_U64:
+		case Decoder::Opcode::V_CMP_NE_U64:
+		case Decoder::Opcode::V_CMP_GE_U64:
+		case Decoder::Opcode::V_CMPX_NE_I64:
+		case Decoder::Opcode::V_CMPX_LE_U64:
+		case Decoder::Opcode::V_CMPX_NE_U64:
+			include_vector(inst.src0, 2u);
+			include_vector(inst.src1, 2u);
+			break;
+		default: break;
+	}
 	if (inst.family == Decoder::Family::DS) {
 		switch (inst.opcode) {
+			case Decoder::Opcode::DS_ADD_U64:
+			case Decoder::Opcode::DS_OR_B64:
 			case Decoder::Opcode::DS_WRITE_B64:
 			case Decoder::Opcode::DS_WRITE_B96:
 			case Decoder::Opcode::DS_WRITE_B128: include_vector(inst.src1, inst.data_dwords); break;
@@ -1077,12 +1156,13 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 				return entry_ir.ISub(lhs, minimum(lhs, rhs));
 			};
 			const auto local = builtin(IR::StageInputKind::LocalInvocationIndex);
-			const auto primitive_chunk = entry_ir.IMul(builtin(IR::StageInputKind::WorkgroupId, 0),
-			                                           u32(mesh.primitives_per_group));
+			const auto group = builtin(IR::StageInputKind::WorkgroupId, 0);
+			const auto primitive_chunk = mesh.fast_launch ? group :
+			    entry_ir.IMul(group, u32(mesh.primitives_per_group));
 			const auto step  = u32(mesh.InputPrimitiveStep());
 			const auto size  = u32(mesh.InputPrimitiveSize());
-			const auto chunk = entry_ir.IMul(primitive_chunk, step);
-			const auto vertices =
+			const auto chunk = mesh.fast_launch ? group : entry_ir.IMul(primitive_chunk, step);
+			const auto vertices = mesh.fast_launch ? u32(mesh.vertices_per_group) :
 			    minimum(subtract_saturate(draw(0), chunk), u32(mesh.vertices_per_group));
 			const auto primitives = entry_ir.Select(
 			    entry_ir.ULessThan(vertices, size), u32(0),
@@ -1103,56 +1183,64 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 			    entry_ir.BitwiseOr(wave_info, entry_ir.BitwiseOr(entry_ir.ShiftLeftLogical(
 			                                                         primitive_count, u32(8)),
 			                                                     vertex_count)));
-			// GS adjacency addresses local ES records in LDS. Fans retain the draw's
-			// center in every subgroup; strip winding follows the global primitive.
-			const auto vertex = entry_ir.IMul(local, step);
-			auto       first  = vertex;
-			auto       second = u32(0);
-			auto       third  = u32(0);
-			if (mesh.InputPrimitiveSize() >= 2u) {
-				second = entry_ir.IAdd(vertex, u32(1));
+			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)));
+			} else {
+				// GS adjacency addresses local ES records in LDS. Fans retain the draw's
+				// center in every subgroup; strip winding follows the global primitive.
+				const auto vertex = entry_ir.IMul(local, step);
+				auto       first  = vertex;
+				auto       second = u32(0);
+				auto       third  = u32(0);
+				if (mesh.InputPrimitiveSize() >= 2u) {
+					second = entry_ir.IAdd(vertex, u32(1));
+				}
+				if (mesh.InputPrimitiveSize() == 3u) {
+					third = entry_ir.IAdd(vertex, u32(2));
+				}
+				auto input_vertex = entry_ir.IAdd(chunk, local);
+				if (mesh.input_primitive == static_cast<uint32_t>(Prospero::PrimitiveType::kTriFan)) {
+					first = u32(0);
+					input_vertex = entry_ir.Select(entry_ir.IEqual(local, u32(0)), u32(0), input_vertex);
+				} else if (mesh.input_primitive == static_cast<uint32_t>(Prospero::PrimitiveType::kTriStrip)) {
+					const auto parity = entry_ir.BitwiseAnd(entry_ir.IAdd(primitive_chunk, local), u32(1));
+					first = entry_ir.IAdd(first, parity);
+					second = entry_ir.ISub(second, parity);
+				}
+				entry_ir.SetVectorReg(static_cast<IR::VectorReg>(0),
+				                      entry_ir.BitwiseOr(entry_ir.ShiftLeftLogical(first, u32(2)),
+				                                         entry_ir.ShiftLeftLogical(second, u32(18))));
+				entry_ir.SetVectorReg(static_cast<IR::VectorReg>(1),
+				                      entry_ir.ShiftLeftLogical(third, u32(2)));
+				const auto index_bytes  = draw(3);
+				const auto indexed      = entry_ir.INotEqual(index_bytes, u32(0));
+				const auto index_low    = draw(4);
+				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)});
+				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(
+				    IR::ValueOpcode::LoadAddressU32,
+				    {index_resource, entry_ir.BitwiseAnd(byte_offset, u32(~3u)), u32(0),
+				     entry_ir.LogicalAnd(indexed, entry_ir.ULessThan(local, vertices))},
+				    IR::MemoryFlags {.index = memory_index});
+				const auto index = IR::U32(entry_ir.Emit(
+				    IR::ValueOpcode::BitFieldUExtract,
+				    {packed_index, entry_ir.IMul(entry_ir.BitwiseAnd(byte_offset, u32(3)), u32(8)),
+				     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)));
 			}
-			if (mesh.InputPrimitiveSize() == 3u) {
-				third = entry_ir.IAdd(vertex, u32(2));
-			}
-			auto input_vertex = entry_ir.IAdd(chunk, local);
-			if (mesh.input_primitive == static_cast<uint32_t>(Prospero::PrimitiveType::kTriFan)) {
-				first = u32(0);
-				input_vertex = entry_ir.Select(entry_ir.IEqual(local, u32(0)), u32(0), input_vertex);
-			} else if (mesh.input_primitive == static_cast<uint32_t>(Prospero::PrimitiveType::kTriStrip)) {
-				const auto parity = entry_ir.BitwiseAnd(entry_ir.IAdd(primitive_chunk, local), u32(1));
-				first = entry_ir.IAdd(first, parity);
-				second = entry_ir.ISub(second, parity);
-			}
-			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(0),
-			                      entry_ir.BitwiseOr(entry_ir.ShiftLeftLogical(first, u32(2)),
-			                                         entry_ir.ShiftLeftLogical(second, u32(18))));
-			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(1),
-			                      entry_ir.ShiftLeftLogical(third, u32(2)));
-			const auto index_bytes  = draw(3);
-			const auto indexed      = entry_ir.INotEqual(index_bytes, u32(0));
-			const auto index_low    = draw(4);
-			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)});
-			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(
-			    IR::ValueOpcode::LoadAddressU32,
-			    {index_resource, entry_ir.BitwiseAnd(byte_offset, u32(~3u)), u32(0),
-			     entry_ir.LogicalAnd(indexed, entry_ir.ULessThan(local, vertices))},
-			    IR::MemoryFlags {.index = memory_index});
-			const auto index = IR::U32(entry_ir.Emit(
-			    IR::ValueOpcode::BitFieldUExtract,
-			    {packed_index, entry_ir.IMul(entry_ir.BitwiseAnd(byte_offset, u32(3)), u32(8)),
-			     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)));
 		} else if (options.stage == ShaderType::Local) {
 			entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(3), IR::U32(IR::Value(64u)));
 			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(2),
@@ -1240,9 +1328,11 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 			                      builtin(IR::StageInputKind::InstanceIndex));
 		}
 	}
+	const bool flush_f32_inputs = options.stage == ShaderType::Compute &&
+	                             (options.input_info.compute->float_mode & 0x10u) == 0;
 	for (const auto& cfg_block: cfg.blocks) {
 		const auto typed_index = block_indices.at(cfg_block.id);
-		Translator translator(result, result.blocks[typed_index], vector_limit);
+		Translator translator(result, result.blocks[typed_index], vector_limit, flush_f32_inputs);
 		for (uint32_t index = cfg_block.inst_begin; index < cfg_block.inst_end; index++) {
 			const auto& instruction = decoded.instructions[index];
 			if (IsCodeTableLoad(cfg, instruction.pc)) {
@@ -1265,7 +1355,7 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 			}
 			translator.TranslateInstruction(instruction);
 		}
-		translator.AddBranchCondition(cfg_block, result.block_info[typed_index]);
+		translator.AddBranchCondition(cfg, cfg_block, result.block_info[typed_index]);
 	}
 	IR::ValidateProgram(result, false);
 	return result;

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

@@ -9,13 +9,15 @@ namespace Libs::Graphics::ShaderRecompiler::Frontend {
 
 class Translator {
 public:
-	Translator(IR::Program& program, IR::Block* block, uint32_t vector_limit)
-	    : program(program), ir(block), current_vector_limit(vector_limit) {}
+	Translator(IR::Program& program, IR::Block* block, uint32_t vector_limit,
+	           bool flush_f32_inputs)
+	    : program(program), ir(block), current_vector_limit(vector_limit),
+	      flush_f32_inputs(flush_f32_inputs) {}
 
 	void TranslateInstruction(const Decoder::Instruction& inst);
 	void TranslateEmbeddedFetch(const Decoder::Instruction& inst, uint32_t attribute,
 	                            uint32_t component_count, const ShaderBufferResource& resource);
-	void AddBranchCondition(const CFG::BasicBlock& source, IR::BlockInfo& info);
+	void AddBranchCondition(const CFG::Graph& graph, const CFG::BasicBlock& source, IR::BlockInfo& info);
 
 private:
 	const Decoder::Operand& SourceAt(const Decoder::Instruction& inst, uint32_t index);
@@ -261,6 +263,7 @@ private:
 	Decoder::Opcode current_opcode       = Decoder::Opcode::UNKNOWN;
 	uint32_t        current_pc           = 0;
 	uint32_t        current_vector_limit = 1;
+	bool            flush_f32_inputs;
 };
 
 } // namespace Libs::Graphics::ShaderRecompiler::Frontend

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

@@ -114,18 +114,40 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 			EmitIntegerCompare(inst, IR::ValueOpcode::INotEqual64, IR::Type::U64, false, true);
 			return;
 
+		case O::V_CMP_LT_I64:
+			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThan64, IR::Type::U64, false, false);
+			return;
+		case O::V_CMP_LE_I64:
+			EmitIntegerCompare(inst, IR::ValueOpcode::SLessThanEqual64, IR::Type::U64, false, false);
+			return;
+		case O::V_CMP_LE_U64:
+			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThanEqual64, IR::Type::U64, false, false);
+			return;
+		case O::V_CMPX_LE_U64:
+			EmitIntegerCompare(inst, IR::ValueOpcode::ULessThanEqual64, IR::Type::U64, false, true);
+			return;
+		case O::V_CMP_GE_U64:
+			EmitIntegerCompare(inst, IR::ValueOpcode::UGreaterThanEqual64, IR::Type::U64, false, false);
+			return;
+
 		case O::V_CMP_EQ_U16:
 			EmitInteger16Compare(inst, IR::ValueOpcode::IEqual32, false, false);
 			return;
 		case O::V_CMP_EQ_I16:
 			EmitInteger16Compare(inst, IR::ValueOpcode::IEqual32, true, false);
 			return;
+		case O::V_CMPX_EQ_I16:
+			EmitInteger16Compare(inst, IR::ValueOpcode::IEqual32, true, true);
+			return;
 		case O::V_CMP_NE_U16:
 			EmitInteger16Compare(inst, IR::ValueOpcode::INotEqual32, false, false);
 			return;
 		case O::V_CMP_NE_I16:
 			EmitInteger16Compare(inst, IR::ValueOpcode::INotEqual32, true, false);
 			return;
+		case O::V_CMPX_NE_I16:
+			EmitInteger16Compare(inst, IR::ValueOpcode::INotEqual32, true, true);
+			return;
 		case O::V_CMP_GT_U16:
 			EmitInteger16Compare(inst, IR::ValueOpcode::UGreaterThan32, false, false);
 			return;
@@ -150,21 +172,45 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_CMP_GT_I16:
 			EmitInteger16Compare(inst, IR::ValueOpcode::SGreaterThan32, true, false);
 			return;
+		case O::V_CMPX_GT_I16:
+			EmitInteger16Compare(inst, IR::ValueOpcode::SGreaterThan32, true, true);
+			return;
 		case O::V_CMP_GE_I16:
 			EmitInteger16Compare(inst, IR::ValueOpcode::SGreaterThanEqual32, true, false);
 			return;
+		case O::V_CMPX_GE_I16:
+			EmitInteger16Compare(inst, IR::ValueOpcode::SGreaterThanEqual32, true, true);
+			return;
 		case O::V_CMP_LT_I16:
 			EmitInteger16Compare(inst, IR::ValueOpcode::SLessThan32, true, false);
 			return;
+		case O::V_CMPX_LT_I16:
+			EmitInteger16Compare(inst, IR::ValueOpcode::SLessThan32, true, true);
+			return;
 		case O::V_CMP_LE_I16:
 			EmitInteger16Compare(inst, IR::ValueOpcode::SLessThanEqual32, true, false);
 			return;
+		case O::V_CMPX_LE_I16:
+			EmitInteger16Compare(inst, IR::ValueOpcode::SLessThanEqual32, true, true);
+			return;
 
 		case O::V_CMP_F_F32: EmitCompareConstant(inst, false, false, false); return;
 		case O::V_CMP_TRU_F32: EmitCompareConstant(inst, true, false, false); return;
 		case O::V_CMP_EQ_F32:
 			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdEqual32, false, false);
 			return;
+		case O::V_CMP_EQ_F64:
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdEqual64, false, false);
+			return;
+		case O::V_CMP_LE_F64:
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdLessThanEqual64, false, false);
+			return;
+		case O::V_CMPX_LE_F64:
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdLessThanEqual64, false, true);
+			return;
+		case O::V_CMPX_GE_F64:
+			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdGreaterThanEqual64, false, true);
+			return;
 		case O::V_CMPX_EQ_F32:
 			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdEqual32, false, true);
 			return;
@@ -267,12 +313,18 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_CMPX_LE_F16:
 			EmitFloatCompare(inst, IR::ValueOpcode::FPOrdLessThanEqual32, true, true);
 			return;
+		case O::V_CMP_NGE_F16:
+			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordLessThan32, true, false);
+			return;
 		case O::V_CMP_NGT_F16:
 			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordLessThanEqual32, true, false);
 			return;
 		case O::V_CMPX_NGT_F16:
 			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordLessThanEqual32, true, true);
 			return;
+		case O::V_CMPX_NLE_F16:
+			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordGreaterThan32, true, true);
+			return;
 		case O::V_CMP_NEQ_F16:
 			EmitFloatCompare(inst, IR::ValueOpcode::FPUnordNotEqual32, true, false);
 			return;
@@ -297,6 +349,10 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_CVT_F32_UBYTE3: V_CVT_F32_UBYTE(inst, 3); return;
 		case O::V_CVT_F32_U32: V_CVT_F32_U32(inst); return;
 		case O::V_CVT_F32_I32: V_CVT_F32_I32(inst); return;
+		case O::V_CVT_F64_I32: return FloatUnary(inst, IR::ValueOpcode::ConvertF64S32);
+		case O::V_CVT_F32_F64: return FloatUnary(inst, IR::ValueOpcode::ConvertF32F64);
+		case O::V_CVT_F64_F32: return FloatUnary(inst, IR::ValueOpcode::ConvertF64F32);
+		case O::V_CVT_F64_U32: return FloatUnary(inst, IR::ValueOpcode::ConvertF64U32);
 		case O::V_CVT_U32_F32: V_CVT_U32_F32(inst); return;
 		case O::V_CVT_I32_F32: V_CVT_I32_F32(inst); return;
 		case O::V_CVT_F16_F32: V_CVT_F16_F32(inst); return;
@@ -394,6 +450,7 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 
 		case O::V_FREXP_MANT_F32: return V_FREXP_MANT_F32(inst);
 		case O::V_RCP_F32: return FloatUnary(inst, IR::ValueOpcode::FPRecip32);
+		case O::V_RCP_F64: return FloatUnary(inst, IR::ValueOpcode::FPRecip64);
 		case O::V_RCP_IFLAG_F32: return FloatUnary(inst, IR::ValueOpcode::FPRecipIFlag32);
 		case O::V_FRACT_F32: return FloatUnary(inst, IR::ValueOpcode::FPFract32);
 		case O::V_TRUNC_F32: return FloatUnary(inst, IR::ValueOpcode::FPTrunc32);
@@ -410,6 +467,8 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_SUB_F32: return FloatBinary(inst, IR::ValueOpcode::FPSub32, false);
 		case O::V_SUBREV_F32: return FloatBinary(inst, IR::ValueOpcode::FPSub32, true);
 		case O::V_MUL_F32: return FloatBinary(inst, IR::ValueOpcode::FPMul32, false);
+		case O::V_ADD_F64: return FloatBinary(inst, IR::ValueOpcode::FPAdd64, false);
+		case O::V_MUL_F64: return FloatBinary(inst, IR::ValueOpcode::FPMul64, false);
 		case O::V_MIN_F32: return FloatBinary(inst, IR::ValueOpcode::FPMin32, false);
 		case O::V_MAX_F32: return FloatBinary(inst, IR::ValueOpcode::FPMax32, false);
 		case O::V_LDEXP_F32: return FloatBinary(inst, IR::ValueOpcode::FPLdexp, false);
@@ -418,6 +477,7 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_MADAK_F32:
 		case O::V_MAD_F32:
 		case O::V_FMA_F32: return FloatTernary(inst, IR::ValueOpcode::FPFma32, false, true);
+		case O::V_FMA_F64: return FloatTernary(inst, IR::ValueOpcode::FPFma64, false, false);
 		case O::V_MIN3_F32: return FloatTernary(inst, IR::ValueOpcode::FPMinTri32, false, false);
 		case O::V_MAX3_F32: return FloatTernary(inst, IR::ValueOpcode::FPMaxTri32, false, false);
 		case O::V_MED3_F32: return FloatTernary(inst, IR::ValueOpcode::FPMedTri32, false, false);

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

@@ -81,7 +81,8 @@ bool EquivalentValue(const ResourcePlan& program, Value left, Value right,
 		    program.memory_info[li] != program.memory_info[ri]) {
 			return false;
 		}
-	} else if (lhs->Flags<uint64_t>() != rhs->Flags<uint64_t>()) {
+	} else if (lhs->GetOpcode() != ValueOpcode::ReadConst &&
+	           lhs->Flags<uint64_t>() != rhs->Flags<uint64_t>()) {
 		return false;
 	}
 	for (size_t index = 0; index < lhs->NumArgs(); index++) {
@@ -141,6 +142,7 @@ CompiledShaderInfo Program::TakeCompiledInfo() && {
 	    .user_data_base  = user_data_base,
 	    .user_data_count = user_data_count,
 	    .scratch_dwords  = scratch_dwords,
+	    .has_address_writes = has_address_writes,
 	    .info            = std::move(info),
 	    .bindings        = std::move(bindings),
 	};

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

@@ -3,6 +3,7 @@
 
 #include <array>
 #include <cstdint>
+#include <utility>
 #include <vector>
 
 namespace Libs::Graphics::ShaderRecompiler::IR {
@@ -34,6 +35,7 @@ struct ResourceSnapshot {
 	std::vector<DescriptorValue> samplers;
 	std::vector<uint32_t>        flattened_srt;
 	std::vector<uint32_t>        user_data;
+	std::vector<std::pair<uint64_t, uint64_t>> specialization_reads;
 	UniformFill                 uniform_fill;
 };
 

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

@@ -30,6 +30,7 @@ enum class ResourceKind {
 	Buffer,
 	IndirectBuffer,
 	Flat,
+	FlatLocal,
 	Global,
 	Scratch,
 	Lds,
@@ -40,7 +41,8 @@ enum class ResourceKind {
 
 [[nodiscard]] constexpr bool IsAddressResourceKind(ResourceKind kind) {
 	return kind == ResourceKind::ScalarAddress || kind == ResourceKind::Flat ||
-	       kind == ResourceKind::Global || kind == ResourceKind::Scratch;
+	       kind == ResourceKind::FlatLocal || kind == ResourceKind::Global ||
+	       kind == ResourceKind::Scratch;
 }
 
 struct MemoryInfo {
@@ -147,6 +149,7 @@ struct SamplerResource {
 	uint32_t first_use_pc          = 0;
 	bool     force_point_filtering = false;
 	bool     depth_compare         = false;
+	bool     integer_border        = false;
 
 	bool operator==(const SamplerResource& other) const = default;
 };
@@ -496,6 +499,7 @@ struct ResourceBlock {
 	Value                 condition;
 	std::vector<uint32_t> successors;
 	std::vector<uint32_t> sources;
+	std::vector<uint32_t> srt_reads;
 };
 
 // Stable shader metadata consumed by the renderer after native IR has been discarded.
@@ -507,6 +511,7 @@ struct CompiledShaderInfo {
 	uint32_t                      user_data_count     = 64;
 	uint32_t                      scratch_dwords      = 0;
 	uint32_t                      param_export_mask   = 0;
+	bool                          has_address_writes  = false;
 	ShaderInfo                    info;
 	BindingLayout                 bindings;
 };
@@ -548,7 +553,7 @@ struct ResourcePlan {
 	std::vector<SrtRead>                srt_reads;
 	std::vector<uint8_t>                clean_flat_slots;
 	bool                                requires_specialization_memory = false;
-	bool                                has_address_writes = false;
+	bool                                capture_specialization_reads = false;
 	bool                                srt_plan_complete          = false;
 	bool                                resource_tracking_complete = false;
 	ShaderInfo                          info;
@@ -561,7 +566,6 @@ struct ResourcePlan {
 	mutable std::vector<uint8_t>            visited_blocks;
 	mutable std::vector<uint32_t>           pending_blocks;
 	mutable std::vector<uint32_t>           material_keys;
-	mutable std::vector<std::pair<uint64_t, uint64_t>> specialization_reads;
 };
 
 struct Program: ResourcePlan {
@@ -582,10 +586,12 @@ struct Program: ResourcePlan {
 	CFG::FailureKind              cfg_failure_kind    = CFG::FailureKind::None;
 	std::string                   fallback_reason;
 	std::vector<BlockInfo>        block_info;
+	struct ScalarWrite { uint32_t pc; ScalarReg reg; };
+	std::vector<ScalarWrite>      scalar_writes;
 	// Typed memory and export instructions reference shader-local metadata by dense index.
 	// Decoder-only details (such as NSA register numbers) have already become IR operands.
 	std::vector<ExportInfo>       export_info;
-	std::vector<Value>            dynamic_reads;
+	bool                          has_address_writes = false;
 	bool                          shader_info_complete = false;
 	BindingLayout                 bindings;
 	bool                          binding_layout_complete = false;

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

@@ -30,6 +30,7 @@ std::string TypeName(Type type) {
 	    "ImageResource",
 	    "SamplerResource",
 	    "ImageAddress",
+	    "F64",
 	};
 	const auto bits = static_cast<uint32_t>(type);
 	if (bits == 0) {

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

@@ -27,6 +27,7 @@ enum class Type : uint32_t {
 	ImageResource   = 1u << 17u,
 	SamplerResource = 1u << 18u,
 	ImageAddress    = 1u << 19u,
+	F64             = 1u << 20u,
 };
 
 constexpr Type operator|(Type lhs, Type rhs) {

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

@@ -2,7 +2,7 @@
 
 #include <algorithm>
 #include <cstring>
-#include <limits>
+#include <memory>
 
 namespace Libs::Graphics::ShaderRecompiler::IR {
 
@@ -135,10 +135,16 @@ bool Value::operator==(const Value& other) const {
 }
 
 Inst::Inst(ValueOpcode value_opcode, uint64_t value_flags)
-    : opcode(value_opcode), flags(value_flags) {
-	const auto count = NumArgsOf(opcode);
-	if (count != std::numeric_limits<size_t>::max()) {
-		args.resize(count);
+    : opcode(value_opcode),
+      num_args(value_opcode == ValueOpcode::Phi ? PhiArity
+                                               : static_cast<uint8_t>(NumArgsOf(value_opcode))),
+      flags(value_flags) {
+	if (num_args == PhiArity) {
+		std::destroy_at(&fixed_args);
+		std::construct_at(&phi_args);
+	} else if (num_args > InlineArity) {
+		std::destroy_at(&fixed_args);
+		std::construct_at(&large_args, num_args);
 	}
 }
 
@@ -154,8 +160,8 @@ Type Inst::GetType() const {
 	if (opcode == ValueOpcode::Phi) {
 		return static_cast<Type>(flags);
 	}
-	if (opcode == ValueOpcode::Identity && !args.empty()) {
-		return args.front().GetType();
+	if (opcode == ValueOpcode::Identity && num_args != 0) {
+		return Arg(0).GetType();
 	}
 	return TypeOf(opcode);
 }
@@ -173,21 +179,24 @@ size_t Inst::UseCount() const {
 }
 
 size_t Inst::NumArgs() const {
-	return args.size();
+	return num_args == PhiArity ? phi_args.size() : num_args;
 }
 
 size_t Inst::NumPhiBlocks() const {
-	return phi_blocks.size();
+	return num_args == PhiArity ? phi_args.size() : 0;
 }
 
 Value Inst::Arg(size_t index) const {
-	EXIT_IF(index >= args.size());
-	return args[index];
+	EXIT_IF(index >= NumArgs());
+	if (num_args <= InlineArity) {
+		return fixed_args[index];
+	}
+	return num_args == PhiArity ? phi_args[index].second : large_args[index];
 }
 
 Block* Inst::PhiBlock(size_t index) const {
-	EXIT_IF(opcode != ValueOpcode::Phi || index >= phi_blocks.size());
-	return phi_blocks[index];
+	EXIT_IF(opcode != ValueOpcode::Phi || index >= phi_args.size());
+	return phi_args[index].first;
 }
 
 Block* Inst::Parent() const {
@@ -203,15 +212,17 @@ void Inst::SetParent(Block* block) {
 }
 
 void Inst::SetArg(size_t index, Value value) {
-	if (index >= args.size()) {
-		EXIT_IF(NumArgsOf(opcode) != std::numeric_limits<size_t>::max());
-		args.resize(index + 1);
-	}
-	const auto old = args[index];
+	const auto old = Arg(index);
 	if (auto* old_inst = old.TryInstruction(); old_inst != nullptr) {
 		RemoveUse(old_inst, index);
 	}
-	args[index] = value;
+	if (num_args <= InlineArity) {
+		fixed_args[index] = value;
+	} else if (num_args == PhiArity) {
+		phi_args[index].second = value;
+	} else {
+		large_args[index] = value;
+	}
 	if (auto* new_inst = value.TryInstruction(); new_inst != nullptr) {
 		AddUse(new_inst, index);
 	}
@@ -219,9 +230,8 @@ void Inst::SetArg(size_t index, Value value) {
 
 void Inst::AddPhiOperand(Block* predecessor, Value value) {
 	EXIT_IF(opcode != ValueOpcode::Phi);
-	const auto index = args.size();
-	args.push_back(value);
-	phi_blocks.push_back(predecessor);
+	const auto index = phi_args.size();
+	phi_args.emplace_back(predecessor, value);
 	if (auto* value_inst = value.TryInstruction(); value_inst != nullptr) {
 		AddUse(value_inst, index);
 	}
@@ -234,21 +244,12 @@ void Inst::ReplaceUsesWith(Value replacement, bool preserve) {
 	}
 	Invalidate();
 	if (preserve) {
-		ReplaceOpcode(ValueOpcode::Identity);
-		args.resize(1);
+		opcode = ValueOpcode::Identity;
+		num_args = 1;
 		SetArg(0, replacement);
 	}
 }
 
-void Inst::ReplaceOpcode(ValueOpcode value_opcode) {
-	opcode           = value_opcode;
-	const auto count = NumArgsOf(opcode);
-	if (count != std::numeric_limits<size_t>::max()) {
-		EXIT_IF(!args.empty() && args.size() != count);
-		args.resize(count);
-	}
-}
-
 void Inst::Invalidate() {
 	ClearArgs();
 	opcode = ValueOpcode::Void;
@@ -269,13 +270,21 @@ void Inst::RemoveUse(Inst* used, size_t operand) {
 }
 
 void Inst::ClearArgs() {
-	for (size_t index = 0; index < args.size(); index++) {
-		if (auto* value_inst = args[index].TryInstruction(); value_inst != nullptr) {
+	for (size_t index = 0; index < NumArgs(); index++) {
+		if (auto* value_inst = Arg(index).TryInstruction(); value_inst != nullptr) {
 			RemoveUse(value_inst, index);
 		}
 	}
-	args.clear();
-	phi_blocks.clear();
+	if (num_args == PhiArity) {
+		std::destroy_at(&phi_args);
+		std::construct_at(&fixed_args);
+	} else if (num_args > InlineArity) {
+		std::destroy_at(&large_args);
+		std::construct_at(&fixed_args);
+	} else {
+		fixed_args.fill(Value {});
+	}
+	num_args = 0;
 }
 
 } // namespace Libs::Graphics::ShaderRecompiler::IR

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

@@ -4,6 +4,7 @@
 #include "graphics/shader/recompiler/ir/Reg.h"
 #include "graphics/shader/recompiler/ir/opcodes/ValueOpcodes.h"
 
+#include <array>
 #include <bit>
 #include <cstdint>
 #include <cstring>
@@ -133,7 +134,6 @@ public:
 	void SetArg(size_t index, Value value);
 	void AddPhiOperand(Block* predecessor, Value value);
 	void ReplaceUsesWith(Value replacement, bool preserve = true);
-	void ReplaceOpcode(ValueOpcode opcode);
 	void Invalidate();
 
 	template <typename T>
@@ -156,13 +156,22 @@ private:
 	void RemoveUse(Inst* used, size_t operand);
 	void ClearArgs();
 
+	static constexpr uint8_t InlineArity = 4;
+	static constexpr uint8_t PhiArity = UINT8_MAX;
+
 	ValueOpcode         opcode;
+	uint8_t             num_args;
+	mutable uint32_t    evaluation_index = UINT32_MAX;
 	uint64_t            flags;
 	Block*              parent = nullptr;
-	std::vector<Value>  args;
-	std::vector<Block*> phi_blocks;
+	union {
+		std::array<Value, InlineArity> fixed_args {};
+		std::vector<Value> large_args;
+		std::vector<std::pair<Block*, Value>> phi_args;
+	};
 	std::vector<Use>    uses;
-	mutable uint32_t    evaluation_index = UINT32_MAX;
 };
 
+static_assert(sizeof(Inst) <= 112, "Inst operand storage unintentionally increased");
+
 } // namespace Libs::Graphics::ShaderRecompiler::IR