update screenshot

Co-authored-by: Behzad <behzadkhan1736@gmail.com>

.github/workflows/build.yml

@@ -113,14 +113,15 @@ jobs:
         shell: cmd
         run: |
           cmake --build _Build/windows --target launcher ^
-            audio_out2_port_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests ^
+            audio_out2_port_tests pad_haptics_tests controller_settings_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 --parallel
+            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|lru_cache|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|virtual_memory_allocation)$"
+          ctest --test-dir _Build/windows --output-on-failure -R "^(audio_out2_port|pad_haptics|controller_settings|ime_dialog|lru_cache|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|archive_file|avplayer_file|virtual_memory_allocation)$"
 
       - name: Install
         shell: cmd
@@ -201,16 +202,17 @@ jobs:
         shell: bash
         run: |
           cmake --build _Build/macos \
-            --target launcher audio_out2_port_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests \
+            --target launcher audio_out2_port_tests pad_haptics_tests controller_settings_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 \
+              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|lru_cache|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|virtual_memory_allocation)$'
+            -R '^(audio_out2_port|pad_haptics|controller_settings|ime_dialog|lru_cache|memory_tracker|page_manager|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|archive_file|avplayer_file|virtual_memory_allocation)$'
 
       - name: Install
         shell: bash
@@ -384,16 +386,16 @@ jobs:
         run: |
           cmake --build _Build/linux \
             --target launcher page_manager_tests memory_tracker_tests \
-              audio_out2_port_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests \
+              audio_out2_port_tests pad_haptics_tests controller_settings_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests \
               resource_tracking_tests scalar_provenance_tests \
-              resource_materialization_tests shader_vertex_metadata_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|lru_cache|memory_tracker|page_manager|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|virtual_memory_allocation)$'
+            -R '^(audio_out2_port|pad_haptics|controller_settings|ime_dialog|lru_cache|memory_tracker|page_manager|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|archive_file|avplayer_file|virtual_memory_allocation)$'
 
       - 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 guest instruction patcher needs a decoder/encoder and a small
+# trampoline assembler. Pin the same revisions used by shadPS4's implementation.
 set(XBYAK_BUILD_TESTS OFF CACHE BOOL "" FORCE)
 FetchContent_Declare(
 	xbyak
@@ -27,6 +48,50 @@ FetchContent_Declare(
 )
 
 FetchContent_MakeAvailable(xbyak zydis)
+target_compile_definitions(xbyak INTERFACE XBYAK_NO_EXCEPTION)
+
+# 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)
@@ -124,6 +189,7 @@ 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/cpuinfo/src/arm/mach/init.c

@@ -14,7 +14,7 @@
 #include <cpuinfo/internal-api.h>
 #include <cpuinfo/log.h>
 
-/* Polyfill recent CPUFAMILY_ARM_* values for older SDKs */
+/* Polyfill recent CPUFAMILY_ARM_* values for older platform headers */
 #ifndef CPUFAMILY_ARM_MONSOON_MISTRAL
 	#define CPUFAMILY_ARM_MONSOON_MISTRAL   0xE81E7EF6
 #endif

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

@@ -1,7 +1,7 @@
-cmake_minimum_required(VERSION 3.16...4.3)
+cmake_minimum_required(VERSION 3.19...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")
+	set(CMAKE_POLICY_VERSION_MINIMUM 3.10 CACHE STRING "Minimum policy version for bundled dependencies")
 endif()
 
 get_property(KYTY_GENERATOR_IS_MULTI_CONFIG GLOBAL PROPERTY GENERATOR_IS_MULTI_CONFIG)
@@ -244,7 +244,7 @@ file(GLOB kyty_emulator_src CONFIGURE_DEPENDS
 	"${KYTY_SOURCE_DIR}/loader/*.h"
 )
 
-find_program(KYTY_GLSLANG_VALIDATOR glslangValidator REQUIRED)
+find_program(KYTY_GLSLANG_VALIDATOR NAMES glslangValidator glslang REQUIRED)
 set(gpu_tiler_shader_dir "${KYTY_SOURCE_DIR}/graphics/host_gpu/shaders")
 set(gpu_tiler_generated_dir "${PROJECT_BINARY_DIR}/gpu_tiler_shaders")
 set(gpu_tiler_shader_names
@@ -259,7 +259,7 @@ set(gpu_tiler_shader_names
 	depth
 	promote_d16
 	demote_d16
-	swap_bgra16
+	color_transform
 )
 file(GLOB gpu_tiler_shader_includes CONFIGURE_DEPENDS "${gpu_tiler_shader_dir}/gpu_tiler_*.inc")
 foreach(shader_name IN LISTS gpu_tiler_shader_names)
@@ -332,7 +332,7 @@ if (KYTY_CLANG_CL)
 	set_source_files_properties(${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)
@@ -393,8 +393,10 @@ add_kyty_full_emulator_test(shader_cfg_tests "${KYTY_TESTS_DIR}/shaderCfgTests.c
 
 add_executable(scalar_provenance_tests EXCLUDE_FROM_ALL
 	"${KYTY_TESTS_DIR}/ScalarProvenanceTests.cpp"
+	"${KYTY_SOURCE_DIR}/graphics/guest_gpu/gpu_format.cpp"
 	"${KYTY_SOURCE_DIR}/graphics/host_gpu/hostMemory.cpp"
 	"${KYTY_SOURCE_DIR}/graphics/shader/recompiler/ir/passes/ReadLaneElimination.cpp"
+	"${KYTY_SOURCE_DIR}/graphics/shader/recompiler/ir/passes/ResourceMaterialization.cpp"
 	"${KYTY_SOURCE_DIR}/graphics/shader/recompiler/ir/passes/SrtWalker.cpp"
 )
 target_link_libraries(scalar_provenance_tests fmt::fmt)
@@ -407,6 +409,25 @@ 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 ${inc_headers})
+
+add_executable(trophy_system_tests EXCLUDE_FROM_ALL
+	"${KYTY_TESTS_DIR}/TrophySystemTests.cpp"
+)
+target_link_libraries(trophy_system_tests common)
+target_include_directories(trophy_system_tests PRIVATE ${inc_headers})
+
+add_executable(avplayer_file_tests EXCLUDE_FROM_ALL
+	"${KYTY_TESTS_DIR}/AvPlayerFileTests.cpp"
+	"${KYTY_SOURCE_DIR}/loader/timer.cpp"
+)
+target_link_libraries(avplayer_file_tests common FFmpeg::ffmpeg ZArchive::zarchive)
+target_include_directories(avplayer_file_tests PRIVATE ${inc_headers})
+
 add_executable(bit_array_tests EXCLUDE_FROM_ALL
 	"${KYTY_TESTS_DIR}/BitArrayTests.cpp"
 )
@@ -458,7 +479,6 @@ add_executable(resource_tracking_tests EXCLUDE_FROM_ALL
 	"${KYTY_SOURCE_DIR}/graphics/shader/recompiler/ir/passes/ResourceTracking.cpp"
 	"${KYTY_SOURCE_DIR}/graphics/shader/recompiler/ir/passes/ResourceMaterialization.cpp"
 	"${KYTY_SOURCE_DIR}/graphics/shader/recompiler/ir/passes/ShaderInfoCollection.cpp"
-	"${KYTY_SOURCE_DIR}/graphics/shader/recompiler/ir/passes/BindingLayout.cpp"
 )
 target_link_libraries(resource_tracking_tests fmt::fmt)
 target_include_directories(resource_tracking_tests PRIVATE ${inc_headers})
@@ -471,6 +491,10 @@ 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"
@@ -481,15 +505,25 @@ 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)
+target_link_libraries(pad_haptics_tests common fmt::fmt SDL3::SDL3-static opus)
 target_include_directories(pad_haptics_tests PRIVATE ${inc_headers})
 
+add_executable(controller_settings_tests EXCLUDE_FROM_ALL
+	"${KYTY_TESTS_DIR}/ControllerSettingsTests.cpp"
+)
+target_link_libraries(controller_settings_tests common fmt::fmt SDL3::SDL3-static)
+target_include_directories(controller_settings_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})
@@ -514,6 +548,7 @@ 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"
@@ -580,6 +615,9 @@ 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 trophy_system COMMAND $<TARGET_FILE:trophy_system_tests>)
+	add_test(NAME avplayer_file COMMAND $<TARGET_FILE:avplayer_file_tests>)
 	add_test(NAME bit_array COMMAND $<TARGET_FILE:bit_array_tests>)
 	add_test(NAME lru_cache COMMAND $<TARGET_FILE:lru_cache_tests>)
 	add_test(NAME shader_vertex_metadata COMMAND $<TARGET_FILE:shader_vertex_metadata_tests>)
@@ -589,9 +627,12 @@ 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 controller_settings COMMAND $<TARGET_FILE:controller_settings_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>)
@@ -631,16 +672,15 @@ 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.
-		add_test(NAME texture_cache_image_overlap
-			COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --image-overlap-only)
-		add_test(NAME texture_cache_htile_clear
-			COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --htile-clear-only)
-		add_test(NAME buffer_cache_ranges
-			COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --buffer-cache-range-only)
-	endif()
+	add_test(NAME texture_cache_image_overlap
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --image-overlap-only)
+	add_test(NAME texture_cache_htile_clear
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --htile-clear-only)
+	add_test(NAME buffer_cache_ranges
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --buffer-cache-range-only)
 
 	add_custom_target(kyty_tests DEPENDS
 		ime_dialog_tests
@@ -649,6 +689,9 @@ if(BUILD_TESTING)
 		image_page_table_tests
 		memory_tracker_tests
 		page_manager_tests
+		archive_file_tests
+		trophy_system_tests
+		avplayer_file_tests
 		bit_array_tests
 		lru_cache_tests
 		shader_vertex_metadata_tests
@@ -660,9 +703,12 @@ 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
+		controller_settings_tests
 		save_data_memory_tests
+		png_enc_tests
 		ces_tests
 		save_data_dialog_tests
 		http_uri_parse_tests
@@ -675,8 +721,9 @@ if(APPLE AND (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR
 	add_executable(macos_sse4a_tests EXCLUDE_FROM_ALL
 		"${KYTY_TESTS_DIR}/macosSse4aTests.cpp"
 		"${KYTY_SOURCE_DIR}/loader/x64InstructionEmulator.cpp"
+		"${KYTY_SOURCE_DIR}/loader/guestInstructionPatcher.cpp"
 	)
-	target_link_libraries(macos_sse4a_tests common fmt::fmt Zydis::Zydis)
+	target_link_libraries(macos_sse4a_tests common fmt::fmt Zydis::Zydis xbyak::xbyak)
 	target_include_directories(macos_sse4a_tests PRIVATE ${inc_headers})
 	if(BUILD_TESTING)
 		add_test(NAME macos_sse4a COMMAND $<TARGET_FILE:macos_sse4a_tests>)
@@ -688,6 +735,15 @@ endif()
 add_executable(kyty_emulator "${KYTY_SOURCE_DIR}/main.cpp" ${kyty_emulator_src})
 configure_macos_guest_address_space(kyty_emulator)
 
+set(KYTY_TROPHY_UNLOCK_SOUND "${CMAKE_CURRENT_SOURCE_DIR}/assets/sounds/trophy-unlock.wav")
+add_custom_target(kyty_trophy_sound
+	COMMAND ${CMAKE_COMMAND} -E make_directory "$<TARGET_FILE_DIR:kyty_emulator>/assets/sounds"
+	COMMAND ${CMAKE_COMMAND} -E copy_if_different "${KYTY_TROPHY_UNLOCK_SOUND}"
+		"$<TARGET_FILE_DIR:kyty_emulator>/assets/sounds/trophy-unlock.wav"
+	VERBATIM)
+add_dependencies(kyty_emulator kyty_trophy_sound)
+install(FILES "${KYTY_TROPHY_UNLOCK_SOUND}" DESTINATION assets/sounds)
+
 target_link_libraries(kyty_emulator ${kyty_emulator_link_libraries})
 if (WIN32)
 	target_link_libraries(kyty_emulator opengl32)
@@ -750,6 +806,10 @@ endif()
 install(TARGETS kyty_emulator DESTINATION .)
 if(KYTY_BUILD_LAUNCHER)
 	add_subdirectory("${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

@@ -1,8 +1,8 @@
 # KytyPS5
 
-[![Build KytyPS5 (Windows)](https://img.shields.io/github/actions/workflow/status/KytyPS5/KytyPS5/build.yml?branch=main&event=push&label=Build%20KytyPS5%20%28Windows%29)](https://github.com/KytyPS5/KytyPS5/actions/workflows/build.yml)
-[![Build KytyPS5 (Linux)](https://img.shields.io/github/actions/workflow/status/KytyPS5/KytyPS5/build.yml?branch=main&event=push&label=Build%20KytyPS5%20%28Linux%29)](https://github.com/KytyPS5/KytyPS5/actions/workflows/build.yml)
-[![Build KytyPS5 (macOS)](https://img.shields.io/github/actions/workflow/status/KytyPS5/KytyPS5/build.yml?branch=main&event=push&label=Build%20KytyPS5%20%28macOS%29)](https://github.com/KytyPS5/KytyPS5/actions/workflows/build.yml)
+[![Build KytyPS5 (Windows)](https://img.shields.io/github/check-runs/KytyPS5/KytyPS5/main?nameFilter=Build%20KytyPS5%20%28Windows%29&label=Build%20KytyPS5%20%28Windows%29)](https://github.com/KytyPS5/KytyPS5/actions/workflows/build.yml)
+[![Build KytyPS5 (Linux)](https://img.shields.io/github/check-runs/KytyPS5/KytyPS5/main?nameFilter=Build%20KytyPS5%20%28Linux%29&label=Build%20KytyPS5%20%28Linux%29)](https://github.com/KytyPS5/KytyPS5/actions/workflows/build.yml)
+[![Build KytyPS5 (macOS)](https://img.shields.io/github/check-runs/KytyPS5/KytyPS5/main?nameFilter=Build%20KytyPS5%20%28macOS%29&label=Build%20KytyPS5%20%28macOS%29)](https://github.com/KytyPS5/KytyPS5/actions/workflows/build.yml)
 [![Platform](https://img.shields.io/badge/platform-Windows%20x64%20%7C%20Linux%20x64%20%7C%20macOS%20x86__64-0078D4.svg)](#system-requirements)
 [![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)
@@ -52,8 +52,8 @@ or graphical glitches, so please include the version you tested when reporting a
       <img src="docs/screenshots/ps5-01.png" width="300" alt="Astro Bot running in KytyPS5">
     </td>
     <td align="center">
-      <strong>Dreaming Sarah</strong><br>
-      <img src="docs/screenshots/ps5-03.png" width="300" alt="Dreaming Sarah running in KytyPS5">
+      <strong>Saros</strong><br>
+      <img src="docs/screenshots/ps5-03.png" width="300" alt="Saros running in KytyPS5">
     </td>
   </tr>
   <tr>
@@ -72,8 +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>UFC 5</strong><br>
-      <img src="docs/screenshots/ps5-06.png" width="300" alt="UFC 5 running in KytyPS5">
+      <strong>UFC 6</strong><br>
+      <img src="docs/screenshots/ps5-06.jpg" width="300" alt="UFC 6 running in KytyPS5">
     </td>
   </tr>
 </table>
@@ -141,7 +141,8 @@ the Vulkan/SPIR-V validation rules.
 - Visual Studio 2022 or Build Tools 2022 with the **Desktop development with C++** workload and
   **C++ Clang tools for Windows** component
 - Qt 6 for MSVC 2022 64-bit, including Concurrent, Network, and Widgets
-- [glslang](https://github.com/KhronosGroup/glslang/releases) (`glslangValidator`) on `PATH`
+- [glslang](https://github.com/KhronosGroup/glslang/releases) (`glslang` or `glslangValidator`) on `PATH`
+- Python 3 on `PATH`; the bundled SPIRV-Tools runs it at configure time
 
 The Microsoft C++ compiler (`cl.exe`) is not supported; use `clang-cl`.
 
@@ -176,7 +177,7 @@ build has no working sound and no gamepad hotplug:
 
 ```bash
 sudo apt-get install --no-install-recommends \
-  clang lld ninja-build cmake git glslang-tools pkg-config \
+  clang lld ninja-build cmake git glslang-tools python3 pkg-config \
   libgl1-mesa-dev libx11-dev libxcursor-dev libxext-dev libxfixes-dev \
   libxi-dev libxrandr-dev libxss-dev libxtst-dev libxkbcommon-dev \
   libasound2-dev libpulse-dev libudev-dev libdbus-1-dev libwayland-dev wayland-protocols
@@ -236,9 +237,10 @@ 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
 
@@ -250,7 +252,7 @@ Requirements:
 
 - An Apple Silicon Mac with Rosetta 2 installed (`softwareupdate --install-rosetta`)
 - Xcode (or the Command Line Tools)
-- Homebrew packages: `brew install cmake ninja glslang`
+- Homebrew packages: `brew install cmake ninja glslang python`
 - Qt 6 (Concurrent, Network, Widgets) with x86-64 support. The official Qt installation is
   universal and works; Homebrew's Qt is arm64-only and will not link
 
@@ -332,17 +334,21 @@ 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 +26,8 @@ target_link_libraries(common
 		cpuinfo
 		kyty_git_version
 		spdlog::spdlog
+		ZArchive::zarchive
+		nlohmann_json::nlohmann_json
 )
 
 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

@@ -23,6 +23,7 @@ void Load(const ConfigOptions& cfg) {
 	EXIT_IF(g_config == nullptr);
 	EXIT_IF(cfg.user_name.empty() || cfg.user_name.size() > MAX_USER_NAME_LENGTH);
 	EXIT_IF(!IsConfiguredUserIdValid(cfg.user_id));
+	EXIT_IF(cfg.controller_speaker_volume > 100 || cfg.controller_vibration_intensity > 100);
 
 	*g_config = cfg;
 }
@@ -47,6 +48,18 @@ const std::string& GetAudioInputDevice() {
 	return g_config->audio_input_device;
 }
 
+const std::optional<ControllerColor>& GetControllerColor() {
+	return g_config->controller_color;
+}
+
+uint32_t GetControllerSpeakerVolume() {
+	return g_config->controller_speaker_volume;
+}
+
+uint32_t GetControllerVibrationIntensity() {
+	return g_config->controller_vibration_intensity;
+}
+
 PresentMode GetPresentMode() {
 	return g_config->present_mode;
 }
@@ -59,6 +72,10 @@ bool FullscreenEnabled() {
 	return g_config->fullscreen_enabled;
 }
 
+bool HideCursorEnabled() {
+	return g_config->hide_cursor_enabled;
+}
+
 bool VrEnabled() {
 	return g_config->vr_enabled;
 }
@@ -135,14 +152,26 @@ bool ReadbackLinearImagesEnabled() {
 	return g_config->readback_linear_images;
 }
 
+bool SyncRawImageBuffersEnabled() {
+	return g_config->sync_raw_image_buffers;
+}
+
 bool TessellationEnabled() {
 	return g_config->tessellation_enabled;
 }
 
+bool TrophyEnabled() {
+	return g_config->trophy_enabled;
+}
+
 bool PlayGoHackEnabled() {
 	return g_config->playgo_hack_enabled;
 }
 
+bool SkipNoticeScreen() {
+	return g_config->skip_notice_screen;
+}
+
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 bool RedZoneProtectionEnabled() {
 	return g_config->red_zone_protection_enabled;

src/common/emulatorConfig.h

@@ -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;
@@ -44,9 +47,13 @@ struct ConfigOptions {
 	std::string            user_name                   = "Kyty";
 	int32_t                user_id                     = DEFAULT_USER_ID;
 	std::string            audio_input_device;
+	std::optional<ControllerColor> controller_color;
+	uint32_t               controller_speaker_volume      = 50;
+	uint32_t               controller_vibration_intensity = 100;
 	PresentMode            present_mode                = PresentMode::Mailbox;
 	int32_t                gpu_index                   = -1;
 	bool                   fullscreen_enabled          = false;
+	bool                   hide_cursor_enabled         = false;
 	bool                   vr_enabled                  = false;
 	bool                   amd_cpu_enabled             = false;
 	uint32_t               vblank_frequency            = 60;
@@ -66,8 +73,11 @@ struct ConfigOptions {
 	bool                   gpu_assisted_validation_enabled = false;
 	bool                   renderdoc_enabled           = false;
 	bool                   readback_linear_images      = false;
+	bool                   sync_raw_image_buffers      = false;
 	bool                   tessellation_enabled        = false;
+	bool                   trophy_enabled              = true;
 	bool                   playgo_hack_enabled         = false;
+	bool                   skip_notice_screen          = false;
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 	bool red_zone_protection_enabled = false;
 #endif
@@ -81,9 +91,13 @@ uint32_t GetScreenHeight();
 const std::string& GetUserName();
 int32_t  GetUserId();
 const std::string& GetAudioInputDevice();
+const std::optional<ControllerColor>& GetControllerColor();
+uint32_t GetControllerSpeakerVolume();
+uint32_t GetControllerVibrationIntensity();
 PresentMode GetPresentMode();
 int32_t GetGpuIndex();
 bool     FullscreenEnabled();
+bool     HideCursorEnabled();
 bool     VrEnabled();
 bool     AmdCpuEnabled();
 uint32_t GetVblankFrequency();
@@ -111,8 +125,11 @@ bool GpuAssistedValidationEnabled();
 
 bool RenderDocEnabled();
 bool ReadbackLinearImagesEnabled();
+bool SyncRawImageBuffersEnabled();
 bool TessellationEnabled();
+bool TrophyEnabled();
 bool PlayGoHackEnabled();
+bool SkipNoticeScreen();
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 bool RedZoneProtectionEnabled();
 #endif

src/common/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

@@ -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/common/trophies.cpp

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

src/common/trophies.h

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

src/graphics/guest_gpu/command_processor/commandProcessor.h

@@ -23,6 +23,12 @@ enum class ContextStateOperation : uint32_t {
 	PushClear = 3,
 };
 
+struct IndirectDrawRegisters {
+	uint32_t vertex_offset;
+	uint32_t instance_offset;
+	uint32_t index_offset;
+};
+
 class Pm4Execution {
 public:
 	[[nodiscard]] bool MadeProgress() const noexcept { return m_made_progress; }
@@ -62,7 +68,6 @@ public:
 
 	void            BufferInit();
 	void            BufferFlush();
-	void            BufferFlushAndWait();
 	void            BufferWait();
 	HW::Context&    GetCtx() { return m_ctx; }
 	HW::UserConfig& GetUcfg() { return m_ucfg; }
@@ -80,10 +85,11 @@ public:
 	void DrawIndex(DrawIndexArgs args);
 	void DrawIndexOffset(uint32_t index_offset, uint32_t index_count);
 	void DrawIndexAuto(DrawAutoArgs args);
-	void DrawIndirect(uint32_t data_offset, uint32_t draw_initiator, bool indexed);
+	void DrawIndirect(uint32_t data_offset, IndirectDrawRegisters registers,
+	                  uint32_t draw_initiator, bool indexed);
 	void DrawIndirectMulti(uint32_t data_offset, uint32_t max_count_or_count,
 	                       const volatile uint32_t* count_addr, uint32_t stride_in_bytes,
-	                       uint32_t draw_initiator, bool indexed);
+	                       IndirectDrawRegisters registers, uint32_t draw_initiator, bool indexed);
 	void WriteAtEndOfPipe32(uint32_t cache_policy, uint32_t event_write_dest,
 	                        uint32_t eop_event_type, uint32_t cache_action, uint32_t event_index,
 	                        uint32_t event_write_source, void* dst_gpu_addr, uint32_t value,

src/graphics/guest_gpu/command_processor/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

@@ -184,7 +184,6 @@ static HW::RenderControl DecodeRenderControl(uint32_t value) {
 	    KYTY_PM4_GET(value, DB_RENDER_CONTROL, COPY_STENCIL_TO_COLOR) != 0;
 	r.copy_centroid          = KYTY_PM4_GET(value, DB_RENDER_CONTROL, COPY_CENTROID) != 0;
 	r.copy_sample            = KYTY_PM4_GET(value, DB_RENDER_CONTROL, COPY_SAMPLE);
-	EXIT_NOT_IMPLEMENTED(r.copy_depth_to_color || r.copy_stencil_to_color);
 
 	return r;
 }
@@ -1642,8 +1641,12 @@ KYTY_CP_OP_PARSER(CpOpDrawIndirect) {
 	const auto data_offset    = buffer[0];
 	const auto draw_initiator = buffer[3];
 	const bool indexed        = (cmd_id == 0xc0032500);
+	// The native indexed packet enables the high-half first-index destination in word 3.
+	const IndirectDrawRegisters registers {
+	    buffer[1] & 0xffffu, buffer[2] & 0xffffu,
+	    indexed && (buffer[2] & (1u << 28u)) != 0 ? buffer[1] >> 16u : Pm4::SH_NOP};
 
-	cp.DrawIndirect(data_offset, draw_initiator, indexed);
+	cp.DrawIndirect(data_offset, registers, draw_initiator, indexed);
 
 	return 4;
 }
@@ -1661,13 +1664,17 @@ KYTY_CP_OP_PARSER(CpOpDrawIndirectMulti) {
 	const auto stride_in_bytes = buffer[7];
 	const auto draw_initiator  = buffer[8];
 	const bool indexed         = (cmd_id == 0xc0083800);
+	// Multi-draw moves the first-index enable to word 4; the destinations stay in 2/3.
+	const IndirectDrawRegisters registers {
+	    buffer[1] & 0xffffu, buffer[2] & 0xffffu,
+	    indexed && (buffer[3] & (1u << 28u)) != 0 ? buffer[1] >> 16u : Pm4::SH_NOP};
 
 	if (count_indirect == 0) {
 		count_addr = nullptr;
 	}
 
 	cp.DrawIndirectMulti(data_offset, max_count_or_count, count_addr, stride_in_bytes,
-	                     draw_initiator, indexed);
+	                     registers, draw_initiator, indexed);
 
 	return 9;
 }
@@ -2266,7 +2273,6 @@ 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 +2280,10 @@ 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 +2321,9 @@ 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 +3629,18 @@ 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/gpu_format.cpp

@@ -20,7 +20,8 @@ struct FormatInfo {
 
 constexpr FormatInfo kFormatInfo[] = {
 	{BufferFormat::k8UNorm, 1, 0, 1, true, false},
-	{BufferFormat::k8SNorm, 0, 0, 1, false, false},
+	{BufferFormat::k8UScaled, 1, 0, 0, true, false},
+	{BufferFormat::k8SNorm, 1, 0, 1, false, false},
 	{BufferFormat::k8UInt, 1, 0, 1, true, true},
 	{BufferFormat::k16UNorm, 2, 0, 2, true, false},
 	{BufferFormat::k16SNorm, 2, 0, 2, true, false},
@@ -28,6 +29,7 @@ constexpr FormatInfo kFormatInfo[] = {
 	{BufferFormat::k16SInt, 2, 0, 2, true, false, true},
 	{BufferFormat::k16Float, 2, 0, 2, true, false},
 	{BufferFormat::k8_8UNorm, 2, 0, 2, true, false},
+	{BufferFormat::k8_8UScaled, 2, 0, 0, true, false},
 	{BufferFormat::k8_8SNorm, 2, 0, 2, true, false},
 	{BufferFormat::k8_8UInt, 2, 0, 2, true, true},
 	{BufferFormat::k8_8SInt, 2, 0, 2, true, false, true},
@@ -41,6 +43,7 @@ constexpr FormatInfo kFormatInfo[] = {
 	{BufferFormat::k16_16Float, 4, 0, 4, true, false},
 	{BufferFormat::k11_11_10UInt, 4, 0, 4, true, true},
 	{BufferFormat::k11_11_10Float, 4, 0, 4, true, false},
+	{BufferFormat::k10_11_11Float, 4, 0, 4, true, false},
 	{BufferFormat::k10_10_10_2UNorm, 4, 0, 4, true, false},
 	{BufferFormat::k10_10_10_2UInt, 4, 0, 4, true, true},
 	{BufferFormat::k8_8_8_8UNorm, 4, 0, 4, true, false},
@@ -61,13 +64,13 @@ constexpr FormatInfo kFormatInfo[] = {
 	{BufferFormat::k32_32_32_32UInt, 16, 0, 16, true, true},
 	{BufferFormat::k32_32_32_32SInt, 16, 0, 16, true, false, true},
 	{BufferFormat::k32_32_32_32Float, 16, 0, 16, true, false},
-	{BufferFormat::k8Srgb, 1, 0, 0, true, false},
-	{BufferFormat::k8_8Srgb, 2, 0, 0, true, false},
+	{BufferFormat::k8Srgb, 1, 0, 1, true, false},
+	{BufferFormat::k8_8Srgb, 2, 0, 2, true, false},
 	{BufferFormat::k8_8_8_8Srgb, 4, 0, 4, true, false},
 	{BufferFormat::k9_9_9_5Float, 4, 0, 0, true, false},
 	{BufferFormat::k5_6_5UNorm, 2, 0, 2, true, false},
 	{BufferFormat::k5_5_5_1UNorm, 2, 0, 2, true, false},
-	{BufferFormat::k1_5_5_5UNorm, 0, 0, 2, false, false},
+	{BufferFormat::k1_5_5_5UNorm, 2, 0, 2, false, false},
 	{BufferFormat::k4_4_4_4UNorm, 2, 0, 2, true, false},
 	{BufferFormat::kFmask8_S2_F1, 1, 0, 1, true, false},
 	{BufferFormat::kFmask8_S4_F1, 1, 0, 1, true, false},
@@ -231,7 +234,12 @@ TextureNumericClass SampledTextureNumericClass(BufferFormat format) {
 }
 
 BufferFormat RemapTextureFormat(BufferFormat format) {
-	return format == BufferFormat::k11_11_10UInt ? BufferFormat::k32UInt : format;
+	switch (format) {
+		case BufferFormat::k8UScaled: return BufferFormat::k8UNorm;
+		case BufferFormat::k8_8UScaled: return BufferFormat::k8_8UNorm;
+		case BufferFormat::k11_11_10UInt: return BufferFormat::k32UInt;
+		default: return format;
+	}
 }
 
 } // namespace Libs::Graphics::Prospero

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;
@@ -712,7 +689,7 @@ void CommandProcessor::ProcessPm4(Pm4Execution& execution) {
 		const auto        opcode        = (packet_header >> 8u) & 0xffu;
 		EXIT_NOT_IMPLEMENTED(remaining_dw > total_dw);
 
-		if (packet_header == 0x80000000u) {
+		if (packet_header == 0x80000000u || Pm4::PacketSizeDw(packet_header) == 1u) {
 			cursor.offset_dw++;
 			execution.m_made_progress = true;
 			continue;
@@ -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;
@@ -900,65 +875,33 @@ void CommandProcessor::DrawIndexOffset(uint32_t index_offset, uint32_t index_cou
 	DrawIndex({.index_count = index_count, .index_addr = index_addr});
 }
 
-void CommandProcessor::DrawIndirect(uint32_t data_offset, uint32_t draw_initiator, bool indexed) {
-	EXIT_NOT_IMPLEMENTED((draw_initiator & ~0x20u) != 2u);
-	EXIT_NOT_IMPLEMENTED(m_draw_indirect_args_base_addr == 0);
-
-	const auto* args_addr =
-	    reinterpret_cast<const void*>(m_draw_indirect_args_base_addr + data_offset);
-
-	if (!indexed) {
-		DrawIndirectArgs args {};
-		std::memcpy(&args, args_addr, sizeof(args));
-		m_num_instances = args.instance_count;
-		DrawIndexAuto({.vertex_count   = args.vertex_count_per_instance,
-		               .instance_count = args.instance_count,
-		               .first_vertex   = args.start_vertex_location,
-		               .first_instance = args.start_instance_location,
-		               .offset_source  = DrawOffsetSource::IndirectArgs});
-		return;
-	}
-
-	DrawIndexedIndirectArgs args {};
-	std::memcpy(&args, args_addr, sizeof(args));
-
-	uint64_t index_size = 0;
-	switch (m_index_type_and_size) {
-		case 0: index_size = 2; break;
-		case 1: index_size = 4; break;
-		case 2: index_size = 1; break;
-		default: EXIT("unknown index_type_and_size: %u\n", m_index_type_and_size);
-	}
-
-	auto* index_addr = reinterpret_cast<const void*>(
-	    m_index_base_addr + static_cast<uint64_t>(args.start_index_location) * index_size);
-
-	const uint32_t index_count =
-	    (m_index_buffer_size != 0 ? std::min(args.index_count_per_instance, m_index_buffer_size)
-	                              : args.index_count_per_instance);
-	if (GraphicsRunDebugDumpEnabled() && index_count != args.index_count_per_instance) {
-		static std::atomic<uint32_t> log_count {0};
-		if (log_count.fetch_add(1, std::memory_order_relaxed) < 64) {
-			LOGF("\t DrawIndexIndirect: clamped index_count from %" PRIu32 " to %" PRIu32
-			     " using INDEX_BUFFER_SIZE\n",
-			     args.index_count_per_instance, index_count);
+static void PatchIndirectDrawOffsets(CommandProcessor& cp, IndirectDrawRegisters registers,
+                                     uint32_t vertex_offset, uint32_t instance_offset,
+                                     uint32_t first_index) {
+	const auto write = [&](uint32_t location, uint32_t value) {
+		if (location == Pm4::SH_NOP) {
+			return;
 		}
-	}
+		EXIT_NOT_IMPLEMENTED(location >= Pm4::SH_NUM ||
+		                     g_hw_sh_indirect_func[location] == nullptr);
+		g_hw_sh_indirect_func[location](cp, location, value);
+	};
+	write(registers.vertex_offset, vertex_offset);
+	write(registers.instance_offset, instance_offset);
+	write(registers.index_offset, first_index);
+}
 
-	m_num_instances = args.instance_count;
-	DrawIndex({.index_count    = index_count,
-	           .index_addr     = index_addr,
-	           .instance_count = args.instance_count,
-	           .base_vertex    = static_cast<int32_t>(args.base_vertex_location),
-	           .first_instance = args.start_instance_location,
-	           .offset_source  = DrawOffsetSource::IndirectArgs});
+void CommandProcessor::DrawIndirect(uint32_t data_offset, IndirectDrawRegisters registers,
+                                    uint32_t draw_initiator, bool indexed) {
+	const auto args_size = indexed ? sizeof(DrawIndexedIndirectArgs) : sizeof(DrawIndirectArgs);
+	DrawIndirectMulti(data_offset, 1, nullptr, args_size, registers, draw_initiator, indexed);
 }
 
 void CommandProcessor::DrawIndirectMulti(uint32_t data_offset, uint32_t max_count_or_count,
                                          const volatile uint32_t* count_addr,
-                                         uint32_t stride_in_bytes, uint32_t draw_initiator,
-                                         bool indexed) {
-	EXIT_NOT_IMPLEMENTED((draw_initiator & ~0x20u) != 2u);
+                                         uint32_t stride_in_bytes, IndirectDrawRegisters registers,
+                                         uint32_t draw_initiator, bool indexed) {
+	EXIT_NOT_IMPLEMENTED((draw_initiator & ~0x20u) != (indexed ? 0u : 2u));
 	EXIT_NOT_IMPLEMENTED(m_draw_indirect_args_base_addr == 0);
 
 	uint32_t draw_count = max_count_or_count;
@@ -992,6 +935,8 @@ void CommandProcessor::DrawIndirectMulti(uint32_t data_offset, uint32_t max_coun
 
 		if (!indexed) {
 			auto* args = reinterpret_cast<const DrawIndirectArgs*>(args_addr);
+			PatchIndirectDrawOffsets(*this, registers, args->start_vertex_location,
+			                         args->start_instance_location, 0);
 			m_num_instances = args->instance_count;
 			DrawIndexAuto({.vertex_count   = args->vertex_count_per_instance,
 			               .instance_count = args->instance_count,
@@ -1002,6 +947,8 @@ void CommandProcessor::DrawIndirectMulti(uint32_t data_offset, uint32_t max_coun
 		}
 
 		auto* args = reinterpret_cast<const DrawIndexedIndirectArgs*>(args_addr);
+		PatchIndirectDrawOffsets(*this, registers, args->base_vertex_location,
+		                         args->start_instance_location, args->start_index_location);
 
 		auto* index_addr = reinterpret_cast<const void*>(
 		    m_index_base_addr + static_cast<uint64_t>(args->start_index_location) * index_size);
@@ -1034,9 +981,6 @@ 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 +1002,9 @@ 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);
+		                                              thread_group_y, thread_group_z, mode, IsAsyncComputeQueue());
 	}
-
-	/*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) {
@@ -1097,7 +1015,8 @@ void CommandProcessor::DispatchIndirect(uint64_t args_addr, uint32_t mode) {
 		return;
 	}
 	m_sh_ctx.SetCsWaveSize(Pm4::ComputeWaveSize(mode));
-	m_renderer.GetRenderExecutor().DispatchIndirect(m_submit_id, CurrentBuffer(), args_addr, mode);
+	m_renderer.GetRenderExecutor().DispatchIndirect(m_submit_id, CurrentBuffer(), args_addr, mode,
+	                                                IsAsyncComputeQueue());
 }
 
 void CommandProcessor::DrawIndexAuto(DrawAutoArgs args) {

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.cpp

@@ -120,15 +120,15 @@ void DumpPm4PacketStream(Common::File* file, const uint32_t* cmd_buffer, uint32_
 		uint32_t len = 0;
 
 		const auto packet_type = static_cast<PacketType>(cmd_id >> 30u);
-		// Type-2 packets are header-only padding; every other packet type requires a body.
-		EXIT_NOT_IMPLEMENTED(dw < 2 && packet_type != PacketType::Type2);
+		EXIT_NOT_IMPLEMENTED(dw < 2 && packet_type != PacketType::Type2 &&
+		                     PacketSizeDw(cmd_id) != 1u);
 
 		switch (packet_type) {
 			case PacketType::Type3: {
 				const bool    sh_gx = (cmd_id & 0x2u) == 0;
 				const uint8_t op    = ((cmd_id >> 8u) & 0xffu);
 				const auto    r     = KYTY_PM4_R(cmd_id);
-				len                 = ((cmd_id >> 16u) & 0x3fffu) + 1;
+				len                 = PacketSizeDw(cmd_id) - 1u;
 
 				EXIT_NOT_IMPLEMENTED(len >= dw);
 

src/graphics/guest_gpu/pm4.h

@@ -19,6 +19,10 @@ class File;
 
 namespace Libs::Graphics::Pm4 {
 
+constexpr uint32_t PacketSizeDw(uint32_t header) {
+	return (header & 0x3fffff00u) == 0x3fff1000u ? 1u : KYTY_PM4_LEN(header);
+}
+
 constexpr uint32_t IT_NOP                       = 0x10;
 constexpr uint32_t IT_SET_BASE                  = 0x11;
 constexpr uint32_t IT_CLEAR_STATE               = 0x12;
@@ -697,6 +701,8 @@ 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/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                               shader_image_int64_atomics_enabled    = false;
 	// bool fp64_denorm_preserve = false; // Temporarily disabled.
 	bool                               attachment_feedback_loop_enabled      = false;
 	bool                               provoking_vertex_last_enabled         = false;

src/graphics/host_gpu/memoryTracker.h

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

src/graphics/host_gpu/pageManager.cpp

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

src/graphics/host_gpu/regionManager.h

@@ -9,6 +9,12 @@
 #include <mutex>
 #include <utility>
 
+#if defined(_MSC_VER)
+#include <intrin.h>
+#elif defined(__x86_64__)
+#include <xmmintrin.h>
+#endif
+
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 #ifndef NOMINMAX
 #define NOMINMAX
@@ -30,21 +36,20 @@ 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");
-			}
-			std::atomic_signal_fence(std::memory_order_seq_cst);
+			EXIT_NOT_IMPLEMENTED(m_owner.load(std::memory_order_relaxed) == thread);
+#if defined(__x86_64__) || defined(_M_X64)
+			_mm_pause();
+#elif defined(_M_ARM64)
+			__yield();
+#elif defined(__aarch64__)
+			asm volatile("yield");
+#endif
 		}
 		m_owner.store(thread, std::memory_order_relaxed);
 	}
 	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);
 	}
@@ -55,7 +60,9 @@ private:
 		return GetCurrentThreadId();
 #elif defined(__APPLE__)
 		// mach thread port is a nonzero per-thread id (0 is the "no owner" sentinel).
-		return static_cast<uint32_t>(pthread_mach_thread_np(pthread_self()));
+		static thread_local const uint32_t tid =
+		    static_cast<uint32_t>(pthread_mach_thread_np(pthread_self()));
+		return tid;
 #elif defined(__linux__)
 		static thread_local const uint32_t tid = static_cast<uint32_t>(::syscall(SYS_gettid));
 		return tid;
@@ -88,8 +95,7 @@ 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>
@@ -122,6 +128,9 @@ public:
 	void ForEachModifiedRange(uint64_t vaddr, uint64_t size, Func&& func) {
 		const auto [start, end] = GetPageRange(vaddr, size);
 		auto&      bits         = GetBits<source>();
+		if (bits.FirstRangeFrom(start).first >= end) {
+			return;
+		}
 		RegionBits mask(bits, start, end);
 		if constexpr (clear) {
 			bits.UnsetRange(start, end);

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

@@ -34,9 +34,36 @@ void BufferCache::WriteDataBuffer(Buffer& buffer, uint64_t address, const void*
 	auto* bytes = static_cast<const uint8_t*>(source);
 	while (size != 0) {
 		const auto chunk  = std::min(size, m_staging_buffer.Size());
+		// Copy waits for ring reuse and flushes the fresh host data before submission.
 		const auto offset = m_staging_buffer.Copy(bytes, chunk, 4);
-		buffer.CopyFrom(m_scheduler.Current(), m_staging_buffer, offset, buffer.Offset(address),
-		                chunk, vk::AccessFlagBits::eHostWrite);
+		const auto destination_offset = buffer.Offset(address);
+		EXIT_IF(destination_offset > buffer.Size() || chunk > buffer.Size() - destination_offset);
+		m_scheduler.EndRendering();
+		const auto command = m_scheduler.Current().Handle();
+		vk::BufferMemoryBarrier2 before {};
+		before.srcStageMask  = vk::PipelineStageFlagBits2::eAllCommands;
+		before.srcAccessMask = vk::AccessFlagBits2::eMemoryRead | vk::AccessFlagBits2::eMemoryWrite;
+		before.dstStageMask  = vk::PipelineStageFlagBits2::eTransfer;
+		before.dstAccessMask = vk::AccessFlagBits2::eTransferWrite;
+		before.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
+		before.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
+		before.buffer        = buffer.Handle();
+		before.offset        = destination_offset;
+		before.size          = chunk;
+		vk::DependencyInfo dependency {};
+		dependency.dependencyFlags          = vk::DependencyFlagBits::eByRegion;
+		dependency.bufferMemoryBarrierCount = 1;
+		dependency.pBufferMemoryBarriers    = &before;
+		command.pipelineBarrier2(dependency);
+		const vk::BufferCopy copy {offset, destination_offset, chunk};
+		command.copyBuffer(m_staging_buffer.Handle(), buffer.Handle(), 1, &copy);
+		auto after          = before;
+		after.srcStageMask  = vk::PipelineStageFlagBits2::eTransfer;
+		after.srcAccessMask = vk::AccessFlagBits2::eTransferWrite;
+		after.dstStageMask  = vk::PipelineStageFlagBits2::eAllCommands;
+		after.dstAccessMask = vk::AccessFlagBits2::eMemoryRead | vk::AccessFlagBits2::eMemoryWrite;
+		dependency.pBufferMemoryBarriers = &after;
+		command.pipelineBarrier2(dependency);
 		bytes += chunk;
 		address += chunk;
 		size -= chunk;
@@ -110,6 +137,7 @@ 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;
@@ -129,11 +157,16 @@ 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;
 	}
@@ -152,27 +185,35 @@ 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;
 }
 
@@ -244,18 +285,8 @@ void BufferCache::ReadMemory(uint64_t vaddr, uint64_t size, bool is_write) {
 		}
 		auto& buffer = m_slot_buffers[FindBuffer(vaddr, size)];
 
-		// Widen nearby CPU reads so they share one GPU drain.
-		constexpr uint64_t WindowSize   = 512 * 1024;
-		const auto         buffer_begin = buffer.CpuAddress();
-		const auto         buffer_end   = buffer_begin + buffer.Size();
-		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);
-			m_memory_tracker.UnmarkRegionAsGpuModified(window_begin, window_end - window_begin);
+		if (DownloadBufferMemory<false>(buffer, vaddr, size)) {
+			m_memory_tracker.UnmarkRegionAsGpuModified(vaddr, size);
 		}
 		if (is_write) {
 			m_memory_tracker.MarkRegionAsCpuModified(vaddr, size);
@@ -454,7 +485,8 @@ 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};
 		}
@@ -490,8 +522,7 @@ 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();
@@ -609,7 +640,7 @@ 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

@@ -117,7 +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/samplerCache.cpp

@@ -123,8 +123,11 @@ vk::Sampler SamplerCache::GetSampler(const ShaderSamplerResource& r, bool intege
 	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/textureCache.cpp

@@ -195,6 +195,9 @@ bool TextureCache::SameBacking(const ImageInfo& cached, const ImageInfo& request
 	if (cached.tile_mode != requested.tile_mode) {
 		return false;
 	}
+	if (cached.GetColorTransform() != requested.GetColorTransform()) {
+		return false;
+	}
 	if (!ImageViewOps::FormatsCompatible(cached.pixel_format, requested.pixel_format) ||
 	    cached.type != requested.type) {
 		return false;
@@ -222,14 +225,6 @@ 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()) {
@@ -591,7 +586,7 @@ 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,
@@ -642,6 +637,7 @@ 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.info.GetColorTransform() == destination.info.GetColorTransform() &&
 	    (source.backing.image_type == destination.backing.image_type ||
 	     (source.backing.image_type != vk::ImageType::e1D &&
 	      destination.backing.image_type != vk::ImageType::e1D)) &&
@@ -655,7 +651,7 @@ void TextureCache::CopyImage(ImageId destination_id, ImageId source_id) {
 			EXIT("TextureCache: cross-format multisample image copy is unsupported\n");
 		}
 		auto& copy_buffer = m_buffer_cache.GetUtilityBuffer(MemoryUsage::DeviceLocal);
-		destination.CopyImageWithBuffer(source, copy_buffer);
+		destination.CopyImageWithBuffer(source, copy_buffer, m_tiler);
 	}
 	if (source.IsGpuModified()) {
 		destination.MarkGpuModified();
@@ -698,7 +694,9 @@ ImageId TextureCache::ResolveDepthOverlap(const ImageInfo& requested, BindingTyp
 	    requested.type == cached.info.type && requested.pitch == cached.info.pitch &&
 	    !requested.HasStencil() && !cached.info.HasStencil() && !requested.HasMetadata() &&
 	    !cached.info.HasMetadata();
-	// PPSA04264
+	// PPSA04264, PPSA04288
+	// A partial view retains the entire matching array layout. HTile belongs to
+	// the depth source, independently of the data slices copied into color storage.
 	const bool retain_cached_layout =
 	    requested.samples == 1 && cached.info.samples == 1 && cached.backing.samples == 1 &&
 	    requested.bytes_per_block == cached.info.bytes_per_block &&
@@ -717,8 +715,10 @@ ImageId TextureCache::ResolveDepthOverlap(const ImageInfo& requested, BindingTyp
 	    requested.data.size / requested.resources.layers ==
 	        cached.info.data.size / cached.info.resources.layers &&
 	    !requested.HasStencil() && !cached.info.HasStencil() && !requested.HasMetadata() &&
-	    !cached.info.HasMetadata();
-	bool recreate = cached.info.resources < requested.resources;
+	    (cached.info.metadata.kind == ImageMetadataKind::None ||
+	     cached.info.metadata.kind == ImageMetadataKind::Htile);
+	bool recreate = cached.info.resources < requested.resources ||
+	                requested.IsVolume() != cached.info.IsVolume();
 	switch (binding) {
 		case BindingType::Texture:
 			recreate |= requested.IsDepth() && !cached.info.IsDepth();
@@ -741,8 +741,6 @@ ImageId TextureCache::ResolveDepthOverlap(const ImageInfo& requested, BindingTyp
 		info.data       = cached.info.data;
 		info.resources  = cached.info.resources;
 		info.mip_layout = cached.info.mip_layout;
-	} else {
-		info.resources = std::max(requested.resources, cached.info.resources);
 	}
 	info.htile_clear_mask     = 0;
 	const auto replacement_id = InsertImage(info);
@@ -799,13 +797,6 @@ TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& reques
 		if (const auto depth_id = ResolveDepthOverlap(requested, binding, cached_id)) {
 			return {depth_id};
 		}
-		if (requested.IsBlock() && !cached.info.IsBlock()) {
-			return {ExpandImage(requested, cached_id)};
-		}
-		if (requested.data.size == cached.info.data.size &&
-		    (requested.IsVolume() || cached.info.IsVolume())) {
-			return {ExpandImage(requested, cached_id)};
-		}
 		// Equal pitch does not imply equal mip placement: a changed extent can move
 		// a level into or out of the mip tail. These are separate guest layouts.
 		if (requested.tile_mode != cached.info.tile_mode ||
@@ -816,6 +807,22 @@ TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& reques
 			}
 			return {merged_id};
 		}
+		if (requested.GetColorTransform() != cached.info.GetColorTransform()) {
+			return {ExpandImage(requested, cached_id)};
+		}
+		if (requested.IsBlock() && !cached.info.IsBlock()) {
+			return {ExpandImage(requested, cached_id)};
+		}
+		// Volume depth is not an array-layer count. A larger depth can retain the
+		// same block-slice layout while requiring a larger native image.
+		if ((requested.IsVolume() || cached.info.IsVolume()) &&
+		    (requested.data.size == cached.info.data.size ||
+		     (requested.type == cached.info.type && requested.resources == cached.info.resources &&
+		      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)};
+		}
 		// PPSA08394
 		// A view cannot change the native image type or grow its extent.
 		if (requested.data.size == cached.info.data.size &&
@@ -838,6 +845,9 @@ TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& reques
 		      requested.resources.layers == cached.info.resources.layers))) {
 			return {ExpandImage(requested, cached_id)};
 		}
+		if (requested.type != cached.info.type) {
+			return {merged_id};
+		}
 		if (requested.pixel_format != cached.info.pixel_format ||
 		    requested.data.size <= cached.info.data.size) {
 			const auto result_id = merged_id ? merged_id : cached_id;
@@ -886,9 +896,7 @@ TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& reques
 ImageId TextureCache::ExpandImage(const ImageInfo& info, ImageId source_id) {
 	RefreshCopySource(source_id);
 	const auto expanded_id = InsertImage(info);
-	auto&      expanded    = m_slot_images[expanded_id];
 	auto&      source      = m_slot_images[source_id];
-	expanded.usage         = source.usage;
 	if (source.binding.is_bound || source.binding.is_target) {
 		source.binding.needs_rebind = true;
 	}
@@ -909,8 +917,8 @@ struct TextureCache::TextureTransfer {
 	TextureUploadLayout              layout;
 	std::vector<vk::BufferImageCopy> regions;
 	std::vector<GpuTileInfo>         tiles;
-	bool                             swap_bgra16 = false;
-	bool                             valid       = false;
+	ColorTransform                   color_transform = ColorTransform::None;
+	bool                             valid           = false;
 
 	[[nodiscard]] uint64_t LinearSize() const {
 		uint64_t size = 0;
@@ -934,22 +942,12 @@ TextureCache::BuildTextureTransfer(const Image& image, BindingType binding,
 	const bool  upload           = direction == TransferDirection::Upload;
 	const bool  render_target    = binding == BindingType::RenderTarget;
 	const bool  video_out        = binding == BindingType::VideoOut;
-	auto        format           = info.guest_format;
 	uint32_t    layers           = info.TransferLayers();
 	bool        volume           = info.IsVolume();
-	bool        allow_depth_tile = upload;
 	const char* owner            = "TextureCache readback";
 
 	TextureTransfer transfer;
-	transfer.swap_bgra16 = info.bgra16 && (!upload || render_target || video_out);
-	if (render_target) {
-		format = ImageOps::RenderTargetTransferFormat(info.bytes_per_block);
-	}
-	if (video_out) {
-		allow_depth_tile = false;
-	} else if (render_target || binding == BindingType::Storage) {
-		allow_depth_tile = true;
-	}
+	transfer.color_transform = info.GetColorTransform();
 	if (upload) {
 		if ((render_target || video_out) &&
 		    (info.resources.layers == 0 || info.data.size % info.resources.layers != 0 ||
@@ -971,9 +969,9 @@ TextureCache::BuildTextureTransfer(const Image& image, BindingType binding,
 		}
 	}
 
-	transfer.layout  = TextureCalcUploadLayout(format, info.extent.width, info.extent.height,
-	                                       info.resources.levels, layers, info.tile_mode,
-	                                       info.data.size, allow_depth_tile, volume, owner);
+	transfer.layout = TextureCalcUploadLayout(info.guest_format, info.extent.width,
+	                                         info.extent.height, info.resources.levels, layers,
+	                                         info.tile_mode, info.data.size, volume, owner);
 	transfer.regions = TextureBuildImageCopies(transfer.layout);
 	if (info.IsDepth()) {
 		for (auto& region: transfer.regions) {
@@ -1037,10 +1035,9 @@ void TextureCache::UploadImage(Image& image, Buffer& source, uint64_t source_off
 		TileManager::Result linear {source.Handle(), source_offset, info.data.size};
 		if (!transfer.tiles.empty()) {
 			linear = m_tiler.Detile(source.Handle(), source_offset, info.data.size,
-			                        transfer.LinearSize(), transfer.tiles);
-		}
-		if (transfer.swap_bgra16) {
-			linear = m_tiler.SwapBgra16(linear);
+			                        transfer.LinearSize(), transfer.tiles, transfer.color_transform);
+		} else if (transfer.color_transform != ColorTransform::None) {
+			linear = m_tiler.TransformColor(linear, transfer.color_transform, true);
 		}
 		upload(transfer.regions, linear);
 		return;
@@ -1051,6 +1048,8 @@ void TextureCache::UploadImage(Image& image, Buffer& source, uint64_t source_off
 	auto info = destination.info;
 	if (image.depth_id) {
 		info.data            = image.info.data;
+		info.resources       = {image.stencil_subresources.level_count,
+		                        image.stencil_subresources.layer_count};
 		info.guest_format    = Prospero::BufferFormat::k8UInt;
 		info.bytes_per_block = 1;
 		if (info.IsTiled()) info.pitch = TileGetDepthPitch(info.extent.width, 1, 0);
@@ -1065,6 +1064,11 @@ void TextureCache::UploadImage(Image& image, Buffer& source, uint64_t source_off
 	auto copies = BuildDepthCopies(info, full_slice_size, image.depth_id
 	                                                        ? vk::ImageAspectFlagBits::eStencil
 	                                                        : vk::ImageAspectFlagBits::eDepth);
+	if (image.depth_id) {
+		for (auto& copy: copies) {
+			copy.imageSubresource.baseArrayLayer += image.stencil_subresources.base_layer;
+		}
+	}
 	TileManager::Result linear {source.Handle(), source_offset, source.Size() - source_offset};
 	if (info.IsTiled()) {
 		const auto tiles = BuildDepthTiles(info);
@@ -1079,7 +1083,7 @@ 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,
@@ -1141,7 +1145,7 @@ void TextureCache::MaterializeColorClear(ImageId id, const ImageDesc& desc,
 		image.info.metadata = desc.info.metadata;
 		// 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) {
+		if (range.size == 0 || desc.info.resources.levels != 1) {
 			return;
 		}
 	}
@@ -1189,10 +1193,12 @@ void TextureCache::MaterializeColorClear(ImageId id, const ImageDesc& desc,
 			           {vk::ImageAspectFlagBits::eColor, view.base_level, view.level_count,
 			            image_first + slice, 1}, clear);
 		}
-		// Native expanded keys own consumption. Existing buffer tracking publishes this CPU
-		// write to future GPU readers; FillBuffer can fault and must run outside the texture lock.
+		// Publish the conversion's expanded keys without treating them as guest writes
+		// to overlapping image data. Invalidate the buffer before updating its backing.
 		if (desc.type != BindingType::VideoOut) {
-			m_buffer_cache.FillBuffer(address, slice_size, UINT32_MAX, false);
+			std::fill(bytes.begin(), bytes.end(), uint8_t {0xff});
+			m_buffer_cache.InvalidateMemory(address, slice_size);
+			LibKernel::Memory::WriteBacking(address, bytes.data(), bytes.size());
 		}
 	}
 }
@@ -1251,7 +1257,8 @@ ImageId TextureCache::AssociateStencil(ImageId depth_id, GuestRange stencil) {
 	}
 	auto& record = m_slot_images[association];
 	TouchImage(record);
-	record.depth_id = depth_id;
+	record.depth_id             = depth_id;
+	record.stencil_subresources = depth.stencil_subresources;
 	return association;
 }
 
@@ -1360,7 +1367,7 @@ 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);
@@ -1457,6 +1464,18 @@ vk::ImageView TextureCache::FindDepthTarget(ImageId id, const ImageDesc& desc) {
 	TouchImage(image);
 	image.MarkGpuModified();
 	image.usage.depth_target = true;
+	if (desc.info.HasStencil()) {
+		const auto layers = image.info.resources.layers;
+		EXIT_IF(layers == 0 || image.info.data.size % layers != 0 ||
+		        desc.info.data.address < image.info.data.address);
+		const auto slice_size = image.info.data.size / layers;
+		const auto offset     = desc.info.data.address - image.info.data.address;
+		EXIT_IF(slice_size == 0 || offset % slice_size != 0 || offset / slice_size >= layers ||
+		        desc.info.resources.layers > layers - offset / slice_size);
+		image.stencil_subresources = {0, desc.info.resources.levels,
+		                             static_cast<uint32_t>(offset / slice_size),
+		                             desc.info.resources.layers};
+	}
 	image.info.stencil = desc.info.stencil;
 	image.info.metadata = desc.info.metadata;
 	if (desc.info.HasMetadata()) {
@@ -1503,9 +1522,11 @@ bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address
 	if (command.IsInvalid() || !GuestRange {address, size}.Valid()) {
 		EXIT("TextureCache: invalid image clear\n");
 	}
-	std::scoped_lock     lock {m_lock};
-	ImageId              selected {};
-	vk::ImageAspectFlags aspect {};
+	std::scoped_lock      lock {m_lock};
+	ImageId               selected {};
+	ImageId               stencil_id {};
+	vk::ImageAspectFlags  aspect {};
+	ImageSubresourceRange subresources;
 	for (const auto id: FindImagesInRegion(address, size, false)) {
 		auto owner = m_slot_images.try_get(id);
 		if (owner == nullptr) {
@@ -1513,10 +1534,13 @@ bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address
 		}
 		vk::ImageAspectFlags candidate {};
 		ImageId              candidate_id = id;
+		ImageSubresourceRange candidate_subresources {
+		    0, owner->info.resources.levels, 0, owner->info.TransferLayers()};
 		if (owner->depth_id && owner->info.data.address == address &&
 		    owner->info.data.size == size) {
-			candidate    = vk::ImageAspectFlagBits::eStencil;
-			candidate_id = owner->depth_id;
+			candidate              = vk::ImageAspectFlagBits::eStencil;
+			candidate_id           = owner->depth_id;
+			candidate_subresources = owner->stencil_subresources;
 			owner        = m_slot_images.try_get(candidate_id);
 			if (owner == nullptr || owner->backing.image == nullptr || !owner->info.HasStencil()) {
 				continue;
@@ -1529,11 +1553,13 @@ bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address
 		if (!candidate) {
 			continue;
 		}
-		if (selected && selected != candidate_id) {
+		if (selected && (selected != candidate_id || subresources != candidate_subresources)) {
 			return false;
 		}
-		selected = candidate_id;
-		aspect   = candidate;
+		selected     = candidate_id;
+		stencil_id   = candidate == vk::ImageAspectFlagBits::eStencil ? id : ImageId {};
+		aspect       = candidate;
+		subresources = candidate_subresources;
 	}
 	if (!selected) {
 		return false;
@@ -1541,7 +1567,7 @@ 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 {
@@ -1555,7 +1581,12 @@ bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address
 		clear.depthStencil.stencil = stencil_clear;
 	}
 	ClearImage(command, selected, image.backing.format,
-	           {aspect, 0, image.info.resources.levels, 0, image.info.TransferLayers()}, clear);
+	           {aspect, subresources.base_level, subresources.level_count,
+	            subresources.base_layer, subresources.layer_count}, clear);
+	if (stencil_id) {
+		TrackImage(stencil_id);
+		CommitGpuWrite(m_slot_images[stencil_id]);
+	}
 	return true;
 }
 
@@ -1687,7 +1718,7 @@ 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,
@@ -1721,14 +1752,14 @@ void TextureCache::DownloadImage(Image& image, Buffer& destination, uint64_t des
 	}
 
 	auto&      texture   = transfer.texture;
-	const auto transform = texture.swap_bgra16 ? TileManager::ColorTransform::SwapBgra16
-	                                           : TileManager::ColorTransform::None;
+	const auto transform = texture.color_transform;
 	if (texture.tiles.empty()) {
-		if (transform == TileManager::ColorTransform::SwapBgra16) {
+		if (transform != ColorTransform::None) {
 			auto linear = m_tiler.GetScratchBuffer(destination_size);
 			image.Download(texture.regions, linear.buffer, 0, linear.size);
-			m_tiler.SwapBgra16(linear,
-			                   {destination.Handle(), destination_offset, destination_size});
+			m_tiler.TransformColor(linear,
+			                       {destination.Handle(), destination_offset, destination_size},
+			                       transform, false);
 			return;
 		}
 		for (auto& copy: texture.regions) {
@@ -1751,7 +1782,7 @@ 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)) {
@@ -1761,14 +1792,12 @@ bool BufferCache::SynchronizeBufferFromImage(Buffer& buffer, uint64_t vaddr, uin
 	const auto available  = buffer.Size() - buf_offset;
 	uint32_t   levels     = 0;
 	uint64_t   copy_size  = 0;
-	if (image.info.IsVolume()) {
-		// Volume mips contain strided block slices, so a mip's linear span cannot prove that
-		// every retained slice fits. Keep volume synchronization whole-image only.
-		if (!buffer.IsInBounds(image.info.data.address, image.info.data.size)) {
-			return false;
-		}
+	if (buffer.IsInBounds(image.info.data.address, image.info.data.size)) {
 		levels    = image.info.resources.levels;
 		copy_size = image.info.data.size;
+	} else if (image.info.IsVolume() || image.info.resources.layers > 1) {
+		// Array and volume slices are strided across the full mip chain.
+		return false;
 	} else {
 		for (; levels < image.info.resources.levels; ++levels) {
 			const auto& mip = image.info.mip_layout[levels];
@@ -1814,7 +1843,7 @@ 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;
@@ -1858,7 +1887,7 @@ void TextureCache::InvalidateMemoryFromGPU(uint64_t address, uint64_t size) {
 	std::scoped_lock lock {m_lock};
 	for (const auto id: FindImagesInRegion(address, size, true)) {
 		auto& image = m_slot_images[id];
-		if (!image.Overlaps(address, size)) {
+		if (image.info.data.address != address) {
 			continue;
 		}
 		if (image.IsGpuModified()) {
@@ -1868,20 +1897,6 @@ void TextureCache::InvalidateMemoryFromGPU(uint64_t address, uint64_t size) {
 	}
 }
 
-bool TextureCache::IsRegionGpuModified(uint64_t address, uint64_t size) {
-	if (!GuestRange {address, size}.Valid()) {
-		return false;
-	}
-	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()) {
-			return true;
-		}
-	}
-	return false;
-}
-
 void TextureCache::InvalidateCpuAliases(uint64_t address, uint64_t size) {
 	const auto page_begin = Common::AlignDown(address, TRACKER_PAGE_SIZE);
 	const auto page_end   = Common::AlignUp(address + size, TRACKER_PAGE_SIZE);
@@ -1946,6 +1961,11 @@ bool TextureCache::TouchMeta(uint64_t address, uint32_t slice, bool is_clear) {
 	return true;
 }
 
+bool TextureCache::IsRegionRegistered(uint64_t address, uint64_t size) {
+	std::scoped_lock lock {m_lock};
+	return !FindImagesInRegion(address, size, false).empty();
+}
+
 void TextureCache::UnmapMemory(uint64_t address, uint64_t size) {
 	if (!GuestRange {address, size}.Valid()) {
 		EXIT("TextureCache: invalid unmap range\n");
@@ -2001,7 +2021,7 @@ 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

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

src/graphics/host_gpu/renderer/colorRenderTarget.cpp

@@ -21,8 +21,6 @@
 
 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:
@@ -262,25 +260,11 @@ 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;
-	desc.info.guest_format = transfer_format;
+	desc.info.guest_format = target_format.guest_format;
 	desc.info.type         = image_type;
 	desc.info.extent       = {width, height, depth};
 	desc.info.resources    = {levels, volume ? 1u : view.image_layers};
@@ -348,7 +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/commandScheduler.cpp

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

src/graphics/host_gpu/renderer/commandScheduler.h

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

src/graphics/host_gpu/renderer/depthRenderTarget.cpp

@@ -160,8 +160,7 @@ static TextureCache::ImageDesc MakeDepthTargetDesc(const CommandBuffer& buffer,
 			           z.depth_view.slice_max);
 	}
 	// EXPCLEAR permits an HTile acceleration state; the host attachment is already expanded.
-	if (z.z_info.partially_resident ||
-	    z.stencil_info.partially_resident || z.z_info.max_mip_level != 0 ||
+	if (z.z_info.max_mip_level != 0 ||
 	    z.depth_view.current_mip_level != 0 || unsupported_shading_rate_encoding ||
 	    depth_address == 0 || (depth_address & 0xffffu) != 0) {
 		DepthFatal("unsupported depth register state");
@@ -316,8 +315,6 @@ void RenderExecutor::ResolveRenderDepthTarget(CommandBuffer& buffer, RenderDepth
 	r.depth_compare_op        = static_cast<vk::CompareOp>(dc.zfunc);
 
 	r.depth_bounds_test_enable = dc.depth_bounds_enable;
-	r.depth_min_bounds         = hw.GetDepthBoundsMin();
-	r.depth_max_bounds         = hw.GetDepthBoundsMax();
 
 	r.stencil_clear_enable =
 	    has_stencil && rc.stencil_clear_enable && !z.depth_view.stencil_write_disable;

src/graphics/host_gpu/renderer/depthRenderTarget.h

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

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

@@ -1,16 +1,20 @@
 #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"
 #include "graphics/host_gpu/renderer/image/imageView.h"
+#include "graphics/host_gpu/renderer/image/tiler.h"
 #include "graphics/host_gpu/renderer/renderTarget.h"
 #include "kernel/memory.h"
 
 #include <algorithm>
 #include <array>
+#include <atomic>
 #include <cstdint>
+#include <fmt/format.h>
 #include <xxhash.h>
 
 namespace Libs::Graphics {
@@ -48,12 +52,29 @@ 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;
 	}
@@ -320,7 +341,9 @@ void Image::CopyImage(Image& source) {
 	EXIT_IF(source.backing.samples != backing.samples);
 	m_scheduler.EndRendering();
 	const uint32_t levels     = std::min(source.backing.mip_levels, backing.mip_levels);
-	const uint32_t base_depth = backing.image_type == vk::ImageType::e3D
+	const uint32_t base_depth = source.backing.image_type == backing.image_type
+	                                ? std::min(source.backing.extent.depth, backing.extent.depth)
+	                            : backing.image_type == vk::ImageType::e3D
 	                                ? backing.extent.depth
 	                                : source.backing.extent.depth;
 	const auto     source_aspect =
@@ -434,7 +457,7 @@ uint32_t Image::CopyRows(uint64_t row_size, uint32_t rows, uint64_t capacity) no
 	return static_cast<uint32_t>(std::min<uint64_t>(rows, capacity / row_size));
 }
 
-void Image::CopyImageWithBuffer(Image& source, Buffer& buffer) {
+void Image::CopyImageWithBuffer(Image& source, Buffer& buffer, TileManager& tiler) {
 	EXIT_IF(buffer.Handle() == nullptr || source.backing.samples != 1 || backing.samples != 1);
 	m_scheduler.EndRendering();
 	const uint32_t levels = std::min(source.backing.mip_levels, backing.mip_levels);
@@ -452,6 +475,8 @@ void Image::CopyImageWithBuffer(Image& source, Buffer& buffer) {
 	const uint32_t destination_block = info.IsBlock() ? 4u : 1u;
 	EXIT_IF(levels == 0 || source_bytes == 0 || source_bytes != destination_bytes ||
 	        source_block != destination_block);
+	const auto source_transform = source.info.GetColorTransform();
+	const auto target_transform = info.GetColorTransform();
 
 	vk::BufferMemoryBarrier2 barrier {};
 	barrier.srcStageMask        = vk::PipelineStageFlagBits2::eTransfer;
@@ -513,9 +538,19 @@ void Image::CopyImageWithBuffer(Image& source, Buffer& buffer) {
 				command.copyImageToBuffer(source.backing.image,
 				                          vk::ImageLayout::eTransferSrcOptimal, buffer.Handle(),
 				                          source_copy);
-				barrier.srcAccessMask = vk::AccessFlagBits2::eTransferWrite;
-				barrier.dstAccessMask = vk::AccessFlagBits2::eTransferRead;
-				command.pipelineBarrier2(dependency);
+				if (source_transform != target_transform) {
+					const TileManager::Result bytes {buffer.Handle(), 0, copy_size};
+					if (source_transform != ColorTransform::None) {
+						tiler.TransformColor(bytes, bytes, source_transform, false);
+					}
+					if (target_transform != ColorTransform::None) {
+						tiler.TransformColor(bytes, bytes, target_transform, true);
+					}
+				} else {
+					barrier.srcAccessMask = vk::AccessFlagBits2::eTransferWrite;
+					barrier.dstAccessMask = vk::AccessFlagBits2::eTransferRead;
+					command.pipelineBarrier2(dependency);
+				}
 				command.copyBufferToImage(buffer.Handle(), backing.image,
 				                          vk::ImageLayout::eTransferDstOptimal, destination_copy);
 			}
@@ -651,7 +686,9 @@ Prospero::BufferFormat RenderTargetTransferFormat(uint32_t bytes_per_element) {
 } // namespace ImageOps
 
 Image::Image(GraphicContext& graphics, CommandScheduler& scheduler, const ImageInfo& image_info)
-    : info(image_info), m_graphics(graphics), m_scheduler(scheduler) {
+    : info(image_info),
+      stencil_subresources {0, image_info.resources.levels, 0, image_info.resources.layers},
+      m_graphics(graphics), m_scheduler(scheduler) {
 	KYTY_PROFILER_FUNCTION();
 	ImageOps::Validate(info);
 	m_cpu_dirty =

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

@@ -18,6 +18,7 @@ namespace Libs::Graphics {
 
 class Buffer;
 class CommandScheduler;
+class TileManager;
 struct ImageTestAccess;
 
 using ImageId = Common::SlotId;
@@ -66,7 +67,7 @@ public:
 	void CopyImage(Image& source);
 	void Resolve(Image& source, const ImageSubresourceRange& source_range,
 	             const ImageSubresourceRange& destination_range);
-	void CopyImageWithBuffer(Image& source, Buffer& buffer);
+	void CopyImageWithBuffer(Image& source, Buffer& buffer, TileManager& tiler);
 	void CopyMip(Image& source, uint32_t mip, uint32_t layer);
 
 	void InvalidateCpuWrite(uint64_t vaddr, uint64_t size) {
@@ -148,6 +149,8 @@ public:
 	uint64_t         track_addr     = 0;
 	uint64_t         track_addr_end = 0;
 	ImageId          depth_id {};
+	// The current stencil plane's mapping into the depth image, also retained by its association.
+	ImageSubresourceRange stencil_subresources;
 	uint64_t         tick_accessed_last = 0;
 	size_t           lru_id             = 0;
 

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

@@ -15,6 +15,8 @@
 
 namespace Libs::Graphics {
 
+enum class ColorTransform : uint32_t { None, SwapBgra16, Reverse10_11_11 };
+
 enum class VideoOutCompression : uint8_t { Uncompressed, Dcc256_256_0, Dcc256_64_64, Unsupported };
 
 enum class ImageMetadataKind : uint8_t { None, Htile, Dcc, Cmask };
@@ -70,6 +72,14 @@ struct ImageInfo {
 	bool                         bgra16          = false;
 	std::array<ImageMipInfo, 16> mip_layout {};
 
+	[[nodiscard]] ColorTransform GetColorTransform() const noexcept {
+		if (bgra16) return ColorTransform::SwapBgra16;
+		switch (guest_format) {
+			case Prospero::BufferFormat::k10_11_11Float:
+				return ColorTransform::Reverse10_11_11;
+			default: return ColorTransform::None;
+		}
+	}
 	[[nodiscard]] constexpr bool HasStencil() const noexcept { return !stencil.Empty(); }
 	[[nodiscard]] constexpr bool HasMetadata() const noexcept {
 		return metadata.kind != ImageMetadataKind::None;
@@ -91,7 +101,8 @@ struct ImageInfo {
 	}
 	[[nodiscard]] bool IsCompatible(const ImageInfo& other) const noexcept {
 		return pixel_format == other.pixel_format && samples == other.samples &&
-		       bytes_per_block == other.bytes_per_block;
+		       bytes_per_block == other.bytes_per_block &&
+		       GetColorTransform() == other.GetColorTransform();
 	}
 	[[nodiscard]] int32_t MipOf(const ImageInfo& container) const noexcept {
 		if (container.resources.levels > container.mip_layout.size()) {
@@ -379,7 +390,28 @@ inline constexpr std::array<VideoOutFormatPolicy, 7> VIDEO_OUT_FORMAT_POLICIES {
 		const auto encoded = static_cast<float>(value & 0xffu) / 255.0f;
 		return encoded <= 0.04045f ? encoded / 12.92f : std::pow((encoded + 0.055f) / 1.055f, 2.4f);
 	};
+	const auto float16 = [](uint32_t value) {
+		const auto sign = (value & 0x8000u) << 16u;
+		const auto exponent = (value >> 10u) & 0x1fu;
+		const auto mantissa = value & 0x3ffu;
+		if (exponent == 0) {
+			return std::copysign(static_cast<float>(mantissa) * 0x1p-24f,
+			                     std::bit_cast<float>(sign));
+		}
+		return std::bit_cast<float>(sign | ((exponent == 31 ? 255u : exponent + 112u) << 23u) |
+		                            (mantissa << 13u));
+	};
 	switch (format) {
+		case vk::Format::eR16G16Sfloat:
+			next.float32[1] = float16(packed >> 16u);
+			[[fallthrough]];
+		case vk::Format::eR16Sfloat: next.float32[0] = float16(packed); break;
+		case vk::Format::eR16G16Unorm:
+			next.float32[1] = static_cast<float>(packed >> 16u) / 65535.0f;
+			[[fallthrough]];
+		case vk::Format::eR16Unorm:
+			next.float32[0] = static_cast<float>(packed & 0xffffu) / 65535.0f;
+			break;
 		// A single-plane float target carries its clear as raw float bits, the same encoding the
 		// depth decoder below uses. Without this the clear is discarded and the target keeps stale
 		// contents.
@@ -428,6 +460,24 @@ 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::eR16Sfloat || format == vk::Format::eR16Unorm) &&
+	    (packed & 0xffffu) != (packed >> 16u)) {
+		return false;
+	}
+	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.cpp

@@ -358,10 +358,8 @@ vk::ImageView Image::FindView(const ImageViewInfo& view_info) {
 	}
 
 	vk::ImageViewUsageCreateInfo usage {};
-	usage.usage = image.usage;
-	if (!is_storage) {
-		usage.usage &= ~vk::ImageUsageFlagBits::eStorage;
-	}
+	usage.usage = is_storage ? vk::ImageUsageFlagBits::eStorage
+	                         : image.usage & ~vk::ImageUsageFlagBits::eStorage;
 	vk::ImageViewMinLodCreateInfoEXT min_lod {};
 	if (normalized.min_lod != 0) {
 		min_lod.minLod = static_cast<float>(normalized.base_level) +

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
@@ -103,8 +90,11 @@ IsSupportedSampledDepthResource(const ShaderRecompiler::IR::ImageResource& resou
 	if (resource.numeric_class != Prospero::TextureNumericClass::Float && resource.depth_compare) {
 		return false;
 	}
-	return resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::None && resource.read &&
-	       !resource.written && !resource.atomic;
+	return ((resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::None &&
+	         resource.mip_count == 1u) ||
+	        (resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::Dynamic &&
+	         resource.mip_count != 0u)) &&
+	       resource.read && !resource.written && !resource.atomic;
 }
 
 inline void ValidateStorageColorView(vk::Format image_format, vk::Format view_format,
@@ -127,7 +117,7 @@ IsSupportedStorageImageResource(const ShaderRecompiler::IR::ImageResource& resou
 	        resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim2DArray) &&
 	       ((resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::None &&
 	         resource.mip_count == 1u) ||
-	        (resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::DynamicStorage &&
+	        (resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::Dynamic &&
 	         resource.mip_count != 0u)) &&
 	       resource.written &&
 	       (!resource.atomic ||

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

@@ -35,6 +35,9 @@ struct HostFormatInfo {
 
 HostFormatInfo ResolveHostFormat(Prospero::BufferFormat guest_format,
                                  Prospero::ChannelOrder order) {
+	if (guest_format == Prospero::BufferFormat::k10_11_11Float) {
+		return {vk::Format::eB10G11R11UfloatPack32, Prospero::ColorMappingBgra};
+	}
 	if (order == Prospero::ChannelOrder::kAlt) {
 		switch (guest_format) {
 			case Prospero::BufferFormat::k8_8_8_8UNorm:
@@ -131,7 +134,8 @@ RenderTargetFormatInfo TextureGetRenderTargetFormat(Prospero::ChannelLayout layo
 		if (host_format.format != vk::Format::eUndefined && bytes != 0) {
 			const auto order_mapping =
 			    kRenderTargetColorMappings[static_cast<size_t>(order)][encoding.components - 1u];
-			return {host_format.format, bytes, host_format.host_to_storage.Then(order_mapping)};
+			return {host_format.format, bytes, host_format.host_to_storage.Then(order_mapping),
+			        encoding.buffer_format};
 		}
 	}
 	EXIT("unsupported render-target format combination: layout=%u type=%u order=%u\n",
@@ -175,8 +179,7 @@ SurfaceFormatInfo TextureGetSurfaceFormatInfo(Prospero::BufferFormat format) {
 TextureUploadLayout TextureCalcUploadLayout(Prospero::BufferFormat format, uint32_t width,
                                             uint32_t height, uint32_t levels, uint32_t depth,
                                             Prospero::TileMode tile_mode, uint64_t upload_size,
-                                            bool allow_depth_tile, bool volume_texture,
-                                            const char* owner) {
+                                            bool volume_texture, const char* owner) {
 	TextureUploadLayout layout {};
 	layout.surface.description = {
 	    format,
@@ -204,8 +207,7 @@ TextureUploadLayout TextureCalcUploadLayout(Prospero::BufferFormat format, uint3
 			     static_cast<uint32_t>(format));
 		}
 	} else {
-		if ((tile_mode == Prospero::TileMode::kDepth && !allow_depth_tile) ||
-		    !TileGetTiledTextureLayout(description, layout.surface)) {
+		if (!TileGetTiledTextureLayout(description, layout.surface)) {
 			EXIT("%s: unsupported typed tiled upload: fmt=%u tile=%u "
 			     "size=%" PRIu64 " extent=%ux%u levels=%u\n",
 			     owner, static_cast<uint32_t>(format), static_cast<uint32_t>(tile_mode),

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

@@ -15,6 +15,7 @@ struct RenderTargetFormatInfo {
 	vk::Format                      format            = vk::Format::eUndefined;
 	uint32_t                        bytes_per_element = 0;
 	Prospero::ColorComponentMapping export_mapping;
+	Prospero::BufferFormat          guest_format = Prospero::BufferFormat::kInvalid;
 };
 
 struct SurfaceFormatInfo {
@@ -47,8 +48,7 @@ RenderTargetFormatInfo TextureGetRenderTargetFormat(Prospero::ChannelLayout layo
 TextureUploadLayout    TextureCalcUploadLayout(Prospero::BufferFormat format, uint32_t width,
                                                uint32_t height, uint32_t levels, uint32_t depth,
                                                Prospero::TileMode tile_mode, uint64_t upload_size,
-                                               bool allow_depth_tile, bool volume_texture,
-                                               const char* owner);
+                                               bool volume_texture, const char* owner);
 std::vector<vk::BufferImageCopy> TextureBuildImageCopies(const TextureUploadLayout& layout);
 bool TextureBuildGpuTileInfos(uint64_t tiled_size, const std::vector<vk::BufferImageCopy>& regions,
                               const TextureUploadLayout& layout, uint32_t levels,

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

@@ -13,7 +13,7 @@
 #include "gpu_tiler_shaders/gpu_tiler_standard4_spv.h"
 #include "gpu_tiler_shaders/gpu_tiler_standard64_3d_spv.h"
 #include "gpu_tiler_shaders/gpu_tiler_standard64_spv.h"
-#include "gpu_tiler_shaders/gpu_tiler_swap_bgra16_spv.h"
+#include "gpu_tiler_shaders/gpu_tiler_color_transform_spv.h"
 #include "graphics/host_gpu/graphicContext.h"
 #include "graphics/host_gpu/renderer/cache/streamBuffer.h"
 #include "graphics/host_gpu/renderer/commandScheduler.h"
@@ -31,7 +31,7 @@ TileManager::TileManager(GraphicContext& graphics, CommandScheduler& scheduler,
                          StreamBuffer& stream_buffer)
     : m_graphics(graphics), m_scheduler(scheduler), m_stream_buffer(stream_buffer) {
 	static_assert(FamilyCount == 9);
-	static_assert(sizeof(Push) == 52);
+	static_assert(sizeof(Push) == 60);
 	std::array<vk::DescriptorSetLayoutBinding, 3> bindings {};
 	for (uint32_t index = 0; index < 2; index++) {
 		bindings[index] = {index, vk::DescriptorType::eStorageBuffer, 1,
@@ -77,8 +77,8 @@ TileManager::~TileManager() {
 	if (m_d32_to_d16 != nullptr) {
 		m_graphics.device.destroyPipeline(m_d32_to_d16, nullptr);
 	}
-	if (m_swap_bgra16 != nullptr) {
-		m_graphics.device.destroyPipeline(m_swap_bgra16, nullptr);
+	if (m_color_transform != nullptr) {
+		m_graphics.device.destroyPipeline(m_color_transform, nullptr);
 	}
 	if (m_pipeline_layout != nullptr) {
 		m_graphics.device.destroyPipelineLayout(m_pipeline_layout, nullptr);
@@ -88,33 +88,10 @@ 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) {
+                          uint64_t target_base, std::vector<Dispatch>& dispatches,
+                          ColorTransform transform) {
 	EXIT_IF(infos.empty() || tiled_capacity == 0 || linear_capacity == 0);
 	const auto& limits = m_graphics.GetPhysicalDeviceProperties().limits;
 	EXIT_NOT_IMPLEMENTED(tiled_capacity > UINT32_MAX || linear_capacity > UINT32_MAX);
@@ -132,6 +109,8 @@ void TileManager::Prepare(bool tile, uint64_t tiled_capacity, uint64_t linear_ca
 	dispatches.clear();
 	dispatches.reserve(infos.size());
 	for (const auto& info: infos) {
+		EXIT_IF((transform == ColorTransform::SwapBgra16 && info.bytes_per_element != 8u) ||
+		        (transform == ColorTransform::Reverse10_11_11 && info.bytes_per_element != 4u));
 		TileBlockLayout block {};
 		const uint32_t  tiled_width  = info.tiled_width != 0 ? info.tiled_width : info.pitch;
 		const uint32_t  tiled_height = info.tiled_height != 0 ? info.tiled_height : info.height;
@@ -217,6 +196,7 @@ void TileManager::Prepare(bool tile, uint64_t tiled_capacity, uint64_t linear_ca
 		dispatch.push.tail_x           = info.tail_x;
 		dispatch.push.tail_y           = info.tail_y;
 		dispatch.push.tail             = info.tail;
+		dispatch.push.color_transform  = static_cast<uint32_t>(transform);
 		dispatches.push_back(dispatch);
 	}
 
@@ -362,15 +342,15 @@ void TileManager::Record(vk::Buffer source, uint64_t source_offset,
 
 TileManager::Result TileManager::Detile(vk::Buffer tiled, uint64_t tiled_offset,
                                         uint64_t tiled_capacity, uint64_t linear_capacity,
-                                        std::span<const GpuTileInfo> infos) {
+                                        std::span<const GpuTileInfo> infos,
+                                        ColorTransform transform) {
 	const auto&    limits = m_graphics.GetPhysicalDeviceProperties().limits;
 	const uint64_t descriptor_alignment =
 	    std::max<uint64_t>(limits.minStorageBufferOffsetAlignment, 4);
 	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);
+	Prepare(false, tiled_capacity, linear_capacity, infos, source_base, 0, dispatches, transform);
+	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};
@@ -400,24 +380,27 @@ 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.
-	Prepare(true, tiled_capacity, linear_capacity, infos, 0, target_base, dispatches);
-	auto linear = AllocateScratch(Common::AlignUp(linear_capacity, 4));
-	DeferDestroy(linear);
+	// Reserve stream parameters before recording the image download.
+	Prepare(true, tiled_capacity, linear_capacity, infos, 0, target_base, dispatches, transform);
+	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) {
-		source = SwapBgra16(source);
-	}
-	Record(source.buffer, source.offset, linear_capacity, tiled, tiled_offset, tiled_capacity,
+	Record(linear.buffer, 0, linear_capacity, tiled, tiled_offset, tiled_capacity,
 	       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 {
@@ -600,9 +583,20 @@ void TileManager::ConvertD16(Result source, Result target, D16Direction directio
 	                        &barriers[1], 0, nullptr);
 }
 
-void TileManager::SwapBgra16(Result input, Result output, uint32_t pixels) {
-	if (m_swap_bgra16 == nullptr) {
-		const auto module = CompileSPV(GPU_TILER_SWAP_BGRA16_SPV, m_graphics.device);
+void TileManager::TransformColor(Result input, Result output, ColorTransform transform,
+                                 bool to_host) {
+	uint32_t element_bytes = 0;
+	switch (transform) {
+		case ColorTransform::SwapBgra16: element_bytes = 8; break;
+		case ColorTransform::Reverse10_11_11: element_bytes = 4; break;
+		default: EXIT("invalid color transform\n");
+	}
+	EXIT_IF(input.size == 0 || input.size % element_bytes != 0 ||
+	        input.size / element_bytes > UINT32_MAX ||
+	        output.size < input.size);
+	const auto pixels = static_cast<uint32_t>(input.size / element_bytes);
+	if (m_color_transform == nullptr) {
+		const auto module = CompileSPV(GPU_TILER_COLOR_TRANSFORM_SPV, m_graphics.device);
 		vk::PipelineShaderStageCreateInfo stage {};
 		stage.stage  = vk::ShaderStageFlagBits::eCompute;
 		stage.module = module;
@@ -611,14 +605,12 @@ void TileManager::SwapBgra16(Result input, Result output, uint32_t pixels) {
 		create.stage  = stage;
 		create.layout = m_pipeline_layout;
 		const auto result =
-		    m_graphics.device.createComputePipelines(nullptr, 1, &create, nullptr, &m_swap_bgra16);
+		    m_graphics.device.createComputePipelines(nullptr, 1, &create, nullptr, &m_color_transform);
 		m_graphics.device.destroyShaderModule(module, nullptr);
-		RequireVulkanSuccess(result, "create BGRA16 swap pipeline");
+		RequireVulkanSuccess(result, "create color transform pipeline");
 	}
-	const uint64_t bytes = static_cast<uint64_t>(pixels) * 8u;
-	EXIT_IF(pixels == 0);
-	const auto input_binding  = BindStorage(input, bytes);
-	const auto output_binding = BindStorage(output, bytes);
+	const auto input_binding  = BindStorage(input, input.size);
+	const auto output_binding = BindStorage(output, input.size);
 
 	const vk::DescriptorBufferInfo infos[] {
 	    input_binding.info,
@@ -649,13 +641,15 @@ void TileManager::SwapBgra16(Result input, Result output, uint32_t pixels) {
 	command.pipelineBarrier(
 	    vk::PipelineStageFlagBits::eAllCommands | vk::PipelineStageFlagBits::eHost,
 	    vk::PipelineStageFlagBits::eComputeShader, {}, 0, nullptr, 2, barriers, 0, nullptr);
-	command.bindPipeline(vk::PipelineBindPoint::eCompute, m_swap_bgra16);
+	command.bindPipeline(vk::PipelineBindPoint::eCompute, m_color_transform);
 	command.pushDescriptorSetKHR(vk::PipelineBindPoint::eCompute, m_pipeline_layout, 0,
 	                             static_cast<uint32_t>(writes.size()), writes.data());
 	Push push {};
-	push.src_base = input_binding.base;
-	push.dst_base = output_binding.base;
-	push.width    = pixels;
+	push.src_base        = input_binding.base;
+	push.dst_base        = output_binding.base;
+	push.width           = pixels;
+	push.color_transform = static_cast<uint32_t>(transform);
+	push.to_host         = to_host;
 	command.pushConstants(m_pipeline_layout, vk::ShaderStageFlagBits::eCompute, 0, sizeof(push),
 	                      &push);
 	command.dispatch((pixels + 63u) / 64u, 1, 1);
@@ -666,19 +660,11 @@ void TileManager::SwapBgra16(Result input, Result output, uint32_t pixels) {
 	                        0, nullptr);
 }
 
-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};
-	SwapBgra16(input, result, static_cast<uint32_t>(input.size / 8u));
+TileManager::Result TileManager::TransformColor(Result input, ColorTransform transform,
+                                                bool to_host) {
+	auto result = GetScratchBuffer(input.size, input.buffer);
+	TransformColor(input, result, transform, to_host);
 	return result;
 }
 
-void TileManager::SwapBgra16(Result input, Result output) {
-	EXIT_NOT_IMPLEMENTED(input.size == 0 || input.size % 8u != 0 || input.size / 8u > UINT32_MAX ||
-	                     output.size < input.size);
-	SwapBgra16(input, output, static_cast<uint32_t>(input.size / 8u));
-}
-
 } // namespace Libs::Graphics

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

@@ -3,16 +3,18 @@
 
 #include "common/common.h"
 #include "graphics/guest_gpu/tile.h"
+#include "graphics/host_gpu/renderer/image/imageInfo.h"
 #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;
@@ -41,7 +43,6 @@ struct GpuTileInfo {
 class TileManager final {
 public:
 	enum class D16Direction { Promote, Demote };
-	enum class ColorTransform { None, SwapBgra16 };
 
 	struct Result {
 		vk::Buffer buffer = nullptr;
@@ -62,20 +63,21 @@ 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);
+	                            uint64_t linear_capacity, std::span<const GpuTileInfo> infos,
+	                            ColorTransform transform = ColorTransform::None);
 	void Tile(vk::Buffer linear, uint64_t linear_offset, uint64_t linear_capacity, vk::Buffer tiled,
 	          uint64_t tiled_offset, uint64_t tiled_capacity, std::span<const GpuTileInfo> infos);
 	void TileImage(Image& image, std::span<const vk::BufferImageCopy> regions, vk::Buffer tiled,
 	               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);
-	void                 SwapBgra16(Result input, Result output);
+	[[nodiscard]] Result TransformColor(Result input, ColorTransform transform, bool to_host);
+	void TransformColor(Result input, Result output, ColorTransform transform, bool to_host);
 
 private:
 	friend struct TileManagerTestAccess;
@@ -99,49 +101,44 @@ private:
 		uint32_t tail_x;
 		uint32_t tail_y;
 		uint32_t tail;
+		uint32_t color_transform;
+		uint32_t to_host;
 	};
 	struct Dispatch {
 		Push     push {};
 		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);
+	             std::vector<Dispatch>& dispatches, ColorTransform transform = ColorTransform::None);
 	void Record(vk::Buffer source, uint64_t source_offset, uint64_t source_capacity,
 	            vk::Buffer target, uint64_t target_offset, uint64_t target_capacity,
 	            std::span<Dispatch> dispatches, bool clear_target);
 	[[nodiscard]] vk::Pipeline GetPipeline(uint32_t slot);
-	void                       SwapBgra16(Result input, Result output, uint32_t pixels);
 
 	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 {};
-	vk::Pipeline                            m_d16_to_d24  = nullptr;
-	vk::Pipeline                            m_d16_to_d32  = nullptr;
-	vk::Pipeline                            m_d24_to_d16  = nullptr;
-	vk::Pipeline                            m_d32_to_d16  = nullptr;
-	vk::Pipeline                            m_swap_bgra16 = nullptr;
+	vk::Pipeline                            m_d16_to_d24      = nullptr;
+	vk::Pipeline                            m_d16_to_d32      = nullptr;
+	vk::Pipeline                            m_d24_to_d16      = nullptr;
+	vk::Pipeline                            m_d32_to_d16      = nullptr;
+	vk::Pipeline                            m_color_transform = nullptr;
 };
 
 } // namespace Libs::Graphics

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

@@ -0,0 +1,70 @@
+#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;
+	}
+	if (blend.separate_alpha_blend &&
+	    blend.color_srcblend == static_cast<uint8_t>(Prospero::BlendFactor::kSrcAlpha) &&
+	    blend.color_destblend == static_cast<uint8_t>(Prospero::BlendFactor::kOneMinusSrcAlpha) &&
+	    blend.alpha_destblend == blend.color_destblend &&
+	    blend.color_comb_fcn == static_cast<uint8_t>(Prospero::BlendOp::kAdd) &&
+	    blend.alpha_comb_fcn == blend.color_comb_fcn) {
+		switch (static_cast<Prospero::BlendFactor>(blend.alpha_srcblend)) {
+			case Prospero::BlendFactor::kZero: return BlendMappingSupport::SourceAlphaZero;
+			case Prospero::BlendFactor::kOne: return BlendMappingSupport::SourceAlphaOne;
+			default: break;
+		}
+	}
+	// 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 +1,29 @@
+#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.
+	SourceAlphaOne, // Source alpha for RGB, one for alpha, with a shared destination equation.
+	SourceAlphaZero, // Source alpha for RGB, zero for alpha, with a shared destination equation.
+	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

@@ -3,6 +3,7 @@
 #include "common/alignment.h"
 #include "common/assert.h"
 #include "common/common.h"
+#include "common/emulatorConfig.h"
 #include "common/file.h"
 #include "common/logging/log.h"
 #include "common/profiler.h"
@@ -25,7 +26,6 @@
 #include "graphics/host_gpu/renderer/renderContext.h"
 #include "graphics/host_gpu/vulkanCommon.h"
 #include "graphics/shader/recompiler/ir/ShaderIR.h"
-#include "graphics/shader/recompiler/ir/passes/BindingLayout.h"
 #include "graphics/shader/recompiler/ir/passes/ResourceMaterialization.h"
 #include "graphics/shader/shader.h"
 #include "kernel/memory.h"
@@ -68,16 +68,13 @@ vk::DescriptorType NativeDescriptorType(BindingKind kind) {
 		case BindingKind::BdaPagetable:
 		case BindingKind::FaultBuffer:
 		case BindingKind::FlattenedSrt:
-		case BindingKind::ShaderData: return vk::DescriptorType::eStorageBuffer;
+		case BindingKind::ShaderData:
+		case BindingKind::SharedMemory: return vk::DescriptorType::eStorageBuffer;
 		case BindingKind::Count: EXIT("invalid native descriptor binding kind");
 	}
 	EXIT("invalid native descriptor binding kind");
 }
 
-uint32_t NativeDescriptorCount(const ShaderRecompiler::IR::DescriptorBinding& binding) {
-	return binding.resources.empty() ? 1u : static_cast<uint32_t>(binding.resources.size());
-}
-
 vk::DescriptorImageInfo MakeImageInfo(const TextureBinding& texture, uint32_t element) {
 	vk::ImageView view = nullptr;
 	if (texture.mip_views.empty()) {
@@ -91,19 +88,6 @@ 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 +110,7 @@ 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;
@@ -139,12 +122,13 @@ NativeStorageBuffer(RenderContext& context, const PreparedBindings::BufferSource
 	if (size > graphics.GetPhysicalDeviceProperties().limits.maxStorageBufferRange) {
 		EXIT("storage buffer range is unsupported\n");
 	}
-	auto [buffer, offset] = context.GetBufferCache().ObtainBuffer(address, size, resource.written,
-	                                                              resource.formatted, id);
+	auto [buffer, offset] = context.GetBufferCache().ObtainBuffer(
+	    address, size, resource.written,
+	    resource.formatted || (resource.read && Config::SyncRawImageBuffersEnabled()), id);
 	const auto aligned_offset = Common::AlignDown(offset, alignment);
 	const auto adjustment     = offset - aligned_offset;
 	const auto max_range      = graphics.GetPhysicalDeviceProperties().limits.maxStorageBufferRange;
-	if (adjustment % sizeof(uint32_t) != 0 || adjustment >= 256 || size > max_range - adjustment) {
+	if (adjustment >= 256 || size > max_range - adjustment) {
 		EXIT("storage buffer offset adjustment is unsupported\n");
 	}
 	buffer_offset = static_cast<uint32_t>(adjustment);
@@ -152,16 +136,6 @@ 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;
 }
 
@@ -336,13 +310,14 @@ void ValidateStorageTexture(const ShaderRecompiler::IR::ImageResource& 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;
+	                              resource.atomic && !resource.atomic64;
 	const bool format_ok =
 	    raw_sint_storage || raw_float_atomic ||
 	    (numeric_class != Prospero::TextureNumericClass::Unsupported &&
 	     numeric_class != Prospero::TextureNumericClass::Sint &&
 	     uint_resource == (numeric_class == Prospero::TextureNumericClass::Uint) &&
-	     (!resource.atomic || format == Prospero::BufferFormat::k32UInt));
+	     (!resource.atomic || format == (resource.atomic64 ? Prospero::BufferFormat::k32_32UInt
+	                                                       : Prospero::BufferFormat::k32UInt)));
 	if (resource_ok && descriptor_ok && encoding_ok && format_ok && size != 0) {
 		return;
 	}
@@ -378,7 +353,8 @@ static TextureCache::ImageDesc NullTextureDesc(const ShaderRecompiler::IR::Image
 			desc.info.guest_format = Prospero::BufferFormat::k32Float;
 			break;
 		case Prospero::TextureNumericClass::Uint:
-			desc.info.guest_format = Prospero::BufferFormat::k32UInt;
+			desc.info.guest_format = resource.atomic64 ? Prospero::BufferFormat::k32_32UInt
+			                                         : Prospero::BufferFormat::k32UInt;
 			break;
 		case Prospero::TextureNumericClass::Sint:
 			desc.info.guest_format = Prospero::BufferFormat::k32SInt;
@@ -389,10 +365,10 @@ static TextureCache::ImageDesc NullTextureDesc(const ShaderRecompiler::IR::Image
 	desc.info.type            = Prospero::ImageType::kColor2D;
 	desc.info.extent          = {1, 1, 1};
 	desc.info.resources       = {1, 1};
-	desc.info.bytes_per_block = 4;
+	desc.info.bytes_per_block = Prospero::NumBytesPerElement(desc.info.guest_format);
 	desc.info.samples         = 1;
 	desc.info.mip_layout[0]   = {0, 0, 1, 1};
-	desc.view_info.format     = desc.info.pixel_format;
+	desc.view_info.format     = resource.atomic64 ? vk::Format::eR64Uint : desc.info.pixel_format;
 	desc.view_info.type       = vk::ImageViewType::e2D;
 	desc.view_info.aspect     = vk::ImageAspectFlagBits::eColor;
 	desc.view_info.usage      = binding == TextureCache::BindingType::Storage
@@ -562,6 +538,9 @@ static bool ResolveTextureMipView(const TileSurfaceDescription& description, boo
 
 TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageResource&   resource,
                                               const ShaderRecompiler::IR::DescriptorValue& value) {
+	if (resource.atomic64 && !m_context.GetGraphics().shader_image_int64_atomics_enabled) {
+		EXIT("64-bit image atomics require shaderImageInt64Atomics\n");
+	}
 	auto descriptor = DecodeNativeDescriptor<ShaderTextureResource>(value);
 	const bool storage = resource.written;
 	if (storage) {
@@ -587,7 +566,7 @@ TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageR
 	const auto physical_levels = multisampled ? 1u : static_cast<uint32_t>(max_mip) + 1u;
 	// IMAGE_STORE addresses BASE_LEVEL; only IMAGE_STORE_MIP selects other view mips.
 	const bool single_storage_mip =
-	    storage && resource.mip_mode != ShaderRecompiler::IR::ImageMipMode::DynamicStorage;
+	    storage && resource.mip_mode != ShaderRecompiler::IR::ImageMipMode::Dynamic;
 	const auto view_levels = multisampled || single_storage_mip
 	                             ? 1u
 	                             : static_cast<uint32_t>(last_level - base_level) + 1u;
@@ -674,8 +653,9 @@ TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageR
 			pixel_format = depth_format->depth_attachment_format;
 		}
 	}
-	const auto storage_view_format = storage && (resource.atomic ||
-	                                            format == Prospero::BufferFormat::k32SInt)
+	const auto storage_view_format = resource.atomic64 ? vk::Format::eR64Uint
+	                                 : storage && (resource.atomic ||
+	                                               format == Prospero::BufferFormat::k32SInt)
 	                                     ? vk::Format::eR32Uint
 	                                     : SrgbStorageViewFormat(pixel_format);
 	const auto view_format         = storage && storage_view_format != vk::Format::eUndefined
@@ -797,8 +777,7 @@ void RenderExecutor::PrepareBindings(const ShaderStageRuntime& runtime,
 	prepared.gds = {nullptr, 0, VK_WHOLE_SIZE};
 	prepared.flattened_srt = {};
 	prepared.shader_data_buffer = {};
-	prepared.buffer_sources.clear();
-	prepared.buffers.clear();
+	prepared.shared_memory = {};
 	prepared.images.resize(program.info.images.size());
 	prepared.samplers.clear();
 	prepared.shader_data.clear();
@@ -811,43 +790,66 @@ void RenderExecutor::PrepareBindings(const ShaderStageRuntime& runtime,
 	}
 	prepared.samplers.reserve(program.info.samplers.size());
 	for (uint32_t i = 0; i < program.info.samplers.size(); i++) {
-		prepared.samplers.push_back(NativeSampler(m_context, program, i, snapshot.samplers[i]));
+		prepared.samplers.push_back(NativeSampler(
+		    m_context, program, i, snapshot.samplers[program.info.samplers[i].snapshot_index]));
 	}
 	prepared.shader_data.reserve(program.bindings.ShaderDataDwords());
-	for (const auto reg: program.bindings.user_data_registers) {
+	for (const auto reg: program.info.user_data_registers) {
 		prepared.shader_data.push_back(snapshot.user_data[reg - program.user_data_base]);
 	}
 	prepared.shader_data.resize(program.bindings.ShaderDataDwords());
-	if (ShaderRecompiler::IR::FindBinding(
-	        program.bindings, ShaderRecompiler::IR::DescriptorBindingKind::Gds) != nullptr) {
+	if (program.bindings.descriptor_counts[static_cast<size_t>(BindingKind::Gds)] != 0) {
 		prepared.gds.buffer = m_context.GetBufferCache().GetGdsBuffer()->Handle();
 	}
 }
 
-void RenderExecutor::FindBuffers(PreparedBindings& prepared) {
+void RenderExecutor::FindBuffers(std::span<PreparedBindings* const> stages) {
 	KYTY_PROFILER_FUNCTION();
-	EXIT_IF(prepared.runtime == nullptr || !*prepared.runtime);
-	const auto& program  = *prepared.runtime->program;
-	const auto& snapshot = *prepared.runtime->resources;
-	auto&       cache    = m_context.GetBufferCache();
-
-	prepared.buffer_sources.clear();
-	const auto& layout = program.bindings;
-	if (layout.memory_offset_count == 0) {
-		return;
-	}
-	const auto& resources = layout.descriptors.front().resources;
-	prepared.buffer_sources.reserve(resources.size());
-	for (const auto resource: resources) {
-		auto descriptor = DecodeNativeDescriptor<ShaderBufferResource>(snapshot.buffers[resource]);
-		const auto address = descriptor.Base48();
-		const auto requested_size = descriptor.GetSize();
-		if (address == 0 || requested_size == 0) {
-			prepared.buffer_sources.push_back({});
+	auto& cache = m_context.GetBufferCache();
+	for (auto* stage: stages) {
+		auto& prepared = *stage;
+		EXIT_IF(prepared.runtime == nullptr || !*prepared.runtime);
+		const auto& program  = *prepared.runtime->program;
+		const auto& snapshot = *prepared.runtime->resources;
+		const auto& layout = program.bindings;
+		prepared.buffer_sources.resize(layout.descriptor_counts[static_cast<size_t>(BindingKind::Buffers)]);
+		if (layout.descriptor_counts[static_cast<size_t>(BindingKind::Buffers)] == 0) {
 			continue;
 		}
-		const auto size = Libs::LibKernel::Memory::ClampRangeSize(address, requested_size);
-		prepared.buffer_sources.push_back({address, size, cache.FindBuffer(address, size)});
+		for (uint32_t resource = 0; resource < program.info.buffers.size(); ++resource) {
+			const auto& metadata = program.info.buffers[resource];
+			if (metadata.descriptor_index == UINT32_MAX) continue;
+			auto& source = prepared.buffer_sources[metadata.descriptor_index];
+			const auto descriptor = DecodeNativeDescriptor<ShaderBufferResource>(snapshot.buffers[resource]);
+			const auto address = descriptor.Base48();
+			auto size = descriptor.GetSize();
+			if (address == 0 || size == 0) {
+				source = {};
+				continue;
+			}
+			if (descriptor.NumRecords() == UINT32_MAX) {
+				if (program.info.buffers[resource].written) {
+					// Sentinel write ranges end before tables captured by any bound stage.
+					for (const auto* reader: stages) {
+						for (const auto [read_address, read_size]: reader->runtime->resources->specialization_reads) {
+							if (read_size == 0) continue;
+							if (read_address > address) {
+								size = std::min(size, read_address - address);
+							} else if (address - read_address < read_size) {
+								EXIT("scalar resource reads overlap a shader buffer write\n");
+							}
+						}
+					}
+				} else {
+					const auto& graphics = m_context.GetGraphics();
+					const auto limit = uint64_t {graphics.GetPhysicalDeviceProperties().limits.maxStorageBufferRange} -
+					                   (graphics.StorageMinAlignment() - 1);
+					size = std::min({size, uint64_t {256} * 1024 * 1024, limit});
+				}
+			}
+			size = Libs::LibKernel::Memory::ClampRangeSize(address, size);
+			source = {address, size, cache.FindBuffer(address, size)};
+		}
 	}
 }
 
@@ -857,10 +859,9 @@ 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() != layout.memory_offset_count);
+	EXIT_IF(prepared.buffer_sources.size() != layout.descriptor_counts[static_cast<size_t>(BindingKind::Buffers)]);
 
-	prepared.buffers.clear();
-	prepared.buffers.reserve(layout.memory_offset_count);
+	prepared.buffers.resize(layout.descriptor_counts[static_cast<size_t>(BindingKind::Buffers)]);
 	EXIT_IF(prepared.shader_data.size() != layout.ShaderDataDwords());
 	std::fill(prepared.shader_data.begin() + layout.memory_offset_dword,
 	          prepared.shader_data.end(), 0);
@@ -869,20 +870,18 @@ void RenderExecutor::RebindBuffers(PreparedBindings& prepared) {
 		const auto shift = (index % 4u) * 8u;
 		prepared.shader_data[dword] |= offset << shift;
 	};
-	for (uint32_t i = 0; i < layout.memory_offset_count; i++) {
-		const auto resource = layout.descriptors.front().resources[i];
+	for (const auto& resource: program.info.buffers) {
+		const auto index = resource.descriptor_index;
+		if (index == UINT32_MAX) continue;
 		uint32_t buffer_offset = 0;
-		prepared.buffers.push_back(NativeStorageBuffer(m_context, prepared.buffer_sources[i],
-		                                               program.info.buffers[resource],
-		                                               program.stage, resource, buffer_offset));
-		pack_memory_offset(i, buffer_offset);
+		prepared.buffers[index] = NativeStorageBuffer(m_context, prepared.buffer_sources[index],
+		                                              resource, buffer_offset);
+		pack_memory_offset(index, buffer_offset);
 	}
-	if (ShaderRecompiler::IR::FindBinding(
-	        layout, ShaderRecompiler::IR::DescriptorBindingKind::FlattenedSrt) != nullptr) {
+	if (layout.descriptor_counts[static_cast<size_t>(BindingKind::FlattenedSrt)] != 0) {
 		prepared.flattened_srt = NativeUpload(m_context, snapshot.flattened_srt);
 	}
-	if (ShaderRecompiler::IR::FindBinding(
-	        program.bindings, ShaderRecompiler::IR::DescriptorBindingKind::ShaderData) != nullptr) {
+	if (layout.descriptor_counts[static_cast<size_t>(BindingKind::ShaderData)] != 0) {
 		prepared.shader_data_buffer = NativeUpload(m_context, prepared.shader_data);
 	}
 }
@@ -895,6 +894,7 @@ void RenderExecutor::RebindImages(PreparedBindings& prepared) {
 	auto&       images   = prepared.images;
 	EXIT_IF(images.size() != program.info.images.size());
 	auto& texture_cache = m_context.GetTextureCache();
+	// Acquire each view before a later overlapping descriptor can replace its image.
 	for (uint32_t i = 0; i < program.info.images.size(); i++) {
 		const auto old_image = texture_cache.m_slot_images.try_get(images[i].image_id);
 		if (old_image == nullptr || (!old_image->registered && !old_image->info.data.Empty()) ||
@@ -906,12 +906,10 @@ void RenderExecutor::RebindImages(PreparedBindings& prepared) {
 			BindImage(images[i].image_id,
 			          images[i].desc.type == TextureCache::BindingType::Storage);
 		}
-	}
-	for (uint32_t i = 0; i < program.info.images.size(); i++) {
 		auto& binding = images[i];
 		binding.mip_views.clear();
 		const auto& resource = program.info.images[i];
-		if (resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::DynamicStorage) {
+		if (resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::Dynamic) {
 			EXIT_IF(resource.mip_count == 0u ||
 			        resource.mip_count != binding.desc.view_info.level_count);
 			binding.mip_views.reserve(resource.mip_count);
@@ -919,6 +917,8 @@ void RenderExecutor::RebindImages(PreparedBindings& prepared) {
 				auto desc = binding.desc;
 				desc.view_info.base_level += mip;
 				desc.view_info.level_count = 1;
+				// The shader selects the mip after applying the guest minimum LOD.
+				desc.view_info.min_lod = 0;
 				binding.mip_views.push_back(texture_cache.FindTexture(binding.image_id, desc));
 			}
 			binding.image_view = binding.mip_views.front();
@@ -935,8 +935,8 @@ void RenderExecutor::RebindImages(PreparedBindings& prepared) {
 void RenderExecutor::PrepareGraphicsBindings(std::span<PreparedBindings* const> stages,
                                              std::span<RenderColorInfo> colors) {
 	bool uses_dma = false;
+	FindBuffers(stages);
 	for (auto* stage: stages) {
-		FindBuffers(*stage);
 		uses_dma |= stage->runtime->program->info.uses_dma;
 	}
 	if (uses_dma) {
@@ -973,7 +973,7 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
                                     std::span<PreparedBindings* const> prepared_bindings) {
 	KYTY_PROFILER_FUNCTION();
 	auto   vk_buffer        = buffer.Handle();
-	size_t descriptor_count = 0;
+	size_t image_descriptor_count = 0;
 	size_t write_count      = 0;
 	ShaderRecompiler::IR::PushData push_data;
 	bool                           has_push_data = false;
@@ -987,9 +987,14 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 	for (const auto* prepared: prepared_bindings) {
 		EXIT_IF(prepared == nullptr || prepared->runtime == nullptr || !*prepared->runtime);
 		const auto& program = *prepared->runtime->program;
-		write_count += program.bindings.descriptors.size();
-		for (const auto& binding: program.bindings.descriptors) {
-			descriptor_count += NativeDescriptorCount(binding);
+		write_count += std::popcount(program.bindings.descriptor_mask);
+		for (auto mask = program.bindings.descriptor_mask; mask != 0; mask &= mask - 1) {
+			const auto index = std::countr_zero(mask);
+			const auto count = program.bindings.descriptor_counts[index];
+			if (index >= ShaderRecompiler::IR::FirstImageBinding &&
+			    index <= static_cast<size_t>(BindingKind::Samplers)) {
+				image_descriptor_count += count;
+			}
 		}
 		const auto shader_stage = NativeShaderStage(program.stage);
 		push_stages |= shader_stage;
@@ -998,12 +1003,45 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 		        (pipeline_bind_point == vk::PipelineBindPoint::eCompute &&
 		         shader_stage != vk::ShaderStageFlagBits::eCompute));
 	}
-	m_descriptor_buffers.clear();
-	m_descriptor_images.clear();
+	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 (const auto& resource: program.info.buffers) {
+					if (!resource.written || resource.descriptor_index == UINT32_MAX) continue;
+					const auto& written = writer->buffer_sources[resource.descriptor_index];
+					if (written.size != 0 && ImageRangeOverlaps(address, size,
+					                                          written.address, written.size)) {
+						EXIT("scalar resource reads overlap a shader buffer write\n");
+					}
+				}
+			}
+		}
+	}
+	m_descriptor_images.resize(image_descriptor_count);
 	m_descriptor_writes.clear();
-	m_descriptor_buffers.reserve(descriptor_count);
-	m_descriptor_images.reserve(descriptor_count);
 	m_descriptor_writes.reserve(write_count);
+	size_t image_cursor = 0;
+	std::array<vk::DescriptorBufferInfo, 2> address_buffers;
 
 	for (auto* prepared: prepared_bindings) {
 		const auto& program       = *prepared->runtime->program;
@@ -1068,77 +1106,79 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 			binding.layout = image.backing.state.layout;
 		}
 
-		m_image_occurrences.assign(descriptors.images.size(), 0);
-		for (const auto& binding: program.bindings.descriptors) {
+		std::array<size_t, ShaderRecompiler::IR::ImageBindingCount> image_starts {};
+		for (auto mask = program.bindings.descriptor_mask; mask != 0; mask &= mask - 1) {
+			const auto index = std::countr_zero(mask);
+			const auto kind = static_cast<BindingKind>(index);
 			vk::WriteDescriptorSet write {};
-			write.dstBinding     = ShaderRecompiler::IR::NativeBinding(program.stage, binding.kind);
-			write.descriptorType = NativeDescriptorType(binding.kind);
-			write.descriptorCount   = NativeDescriptorCount(binding);
-			const auto buffer_start = m_descriptor_buffers.size();
-			const auto image_start  = m_descriptor_images.size();
-			if (ShaderRecompiler::IR::ImageBindingResourceClass(binding.kind) !=
+			write.dstBinding = ShaderRecompiler::IR::NativeBinding(program.stage, kind);
+			write.descriptorType = NativeDescriptorType(kind);
+			write.descriptorCount = program.bindings.descriptor_counts[index];
+			const auto image_start = image_cursor;
+			if (ShaderRecompiler::IR::ImageBindingResourceClass(kind) !=
 			    ShaderRecompiler::IR::ImageResourceClass::None) {
-				for (const auto resource: binding.resources) {
-					m_descriptor_images.push_back(MakeImageInfo(
-					    descriptors.images.at(resource), m_image_occurrences.at(resource)++));
-				}
+				image_starts[ShaderRecompiler::IR::ImageBindingIndex(kind)] = image_start;
+				image_cursor += write.descriptorCount;
 			} else {
-				switch (binding.kind) {
+				switch (kind) {
 					case BindingKind::Buffers:
-						EXIT_IF(descriptors.buffers.size() != binding.resources.size());
+						EXIT_IF(descriptors.buffers.size() != write.descriptorCount);
 						for (const auto& view: descriptors.buffers) {
 							EXIT_IF(view.buffer == nullptr);
-							m_descriptor_buffers.push_back(view);
 						}
+						write.pBufferInfo = descriptors.buffers.data();
 						break;
 					case BindingKind::BdaPagetable:
 					case BindingKind::FaultBuffer: {
-						auto&       cache      = m_context.GetBufferCache();
-						const auto* bda_buffer = binding.kind == BindingKind::BdaPagetable
+						auto& cache = m_context.GetBufferCache();
+						const auto* bda_buffer = kind == BindingKind::BdaPagetable
 						                             ? cache.GetBdaPageTableBuffer()
 						                             : cache.GetFaultBuffer();
-						m_descriptor_buffers.emplace_back(bda_buffer->Handle(), 0,
-						                                  bda_buffer->Size());
+						auto& view = address_buffers[kind == BindingKind::FaultBuffer];
+						view = {bda_buffer->Handle(), 0, bda_buffer->Size()};
+						write.pBufferInfo = &view;
 						break;
 					}
 					case BindingKind::FlattenedSrt:
 					case BindingKind::ShaderData:
+					case BindingKind::SharedMemory:
 					case BindingKind::Gds: {
 						const vk::DescriptorBufferInfo* view = &descriptors.gds;
-						if (binding.kind == BindingKind::FlattenedSrt) {
+						if (kind == BindingKind::FlattenedSrt) {
 							view = &descriptors.flattened_srt;
-						} else if (binding.kind == BindingKind::ShaderData) {
+						} else if (kind == BindingKind::ShaderData) {
 							view = &descriptors.shader_data_buffer;
+						} else if (kind == BindingKind::SharedMemory) {
+							view = &descriptors.shared_memory;
 						}
 						EXIT_IF(view->buffer == nullptr);
-						m_descriptor_buffers.push_back(*view);
+						write.pBufferInfo = view;
 						break;
 					}
 					case BindingKind::Samplers:
-						for (const auto resource: binding.resources) {
-							const auto sampler = descriptors.samplers.at(resource);
+						for (const auto sampler: descriptors.samplers) {
 							EXIT_IF(sampler == nullptr);
-							m_descriptor_images.emplace_back(sampler, nullptr,
-							                                 vk::ImageLayout::eUndefined);
+							m_descriptor_images[image_cursor++] = {sampler, nullptr, vk::ImageLayout::eUndefined};
 						}
 						break;
 					case BindingKind::Count: EXIT("invalid descriptor binding kind");
 				}
 			}
-			if (m_descriptor_buffers.size() != buffer_start) {
-				write.pBufferInfo = m_descriptor_buffers.data() + buffer_start;
-			}
-			if (m_descriptor_images.size() != image_start) {
+			if (image_cursor != image_start) {
 				write.pImageInfo = m_descriptor_images.data() + image_start;
 			}
 			m_descriptor_writes.push_back(write);
 		}
-		for (uint32_t i = 0; i < descriptors.images.size(); i++) {
-			const auto expected =
-			    descriptors.images[i].mip_views.empty()
-			        ? 1u
-			        : static_cast<uint32_t>(descriptors.images[i].mip_views.size());
-			EXIT_IF(m_image_occurrences[i] != expected);
+		for (uint32_t i = 0; i < program.info.images.size(); ++i) {
+			const auto& resource = program.info.images[i];
+			const auto& texture = descriptors.images[i];
+			EXIT_IF(resource.mip_count != (texture.mip_views.empty() ? 1u : texture.mip_views.size()));
+			const auto kind = *ShaderRecompiler::IR::DescriptorBindingForImage(resource);
+			const auto start = image_starts[ShaderRecompiler::IR::ImageBindingIndex(kind)] +
+			                   resource.descriptor_index;
+			for (uint32_t mip = 0; mip < resource.mip_count; ++mip) {
+				m_descriptor_images[start + mip] = MakeImageInfo(texture, mip);
+			}
 		}
 
 		const auto shader_data_dwords = program.bindings.ShaderDataDwords();

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

@@ -44,13 +44,12 @@ struct PreparedBindings {
 	vk::DescriptorBufferInfo              gds {nullptr, 0, VK_WHOLE_SIZE};
 	vk::DescriptorBufferInfo              flattened_srt;
 	vk::DescriptorBufferInfo              shader_data_buffer;
+	vk::DescriptorBufferInfo              shared_memory;
 	std::vector<uint32_t>                 shader_data;
 };
 
 [[nodiscard]] vk::DescriptorType
 NativeDescriptorType(ShaderRecompiler::IR::DescriptorBindingKind kind);
-[[nodiscard]] uint32_t
-NativeDescriptorCount(const ShaderRecompiler::IR::DescriptorBinding& binding);
 [[nodiscard]] vk::DescriptorImageInfo MakeImageInfo(const TextureBinding& texture,
                                                     uint32_t              element = 0);
 

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"
@@ -37,6 +39,16 @@ namespace Libs::Graphics {
 
 namespace {
 
+uint8_t RemapSourceAlphaFactor(uint8_t factor) {
+	switch (static_cast<Prospero::BlendFactor>(factor)) {
+		case Prospero::BlendFactor::kSrcAlpha:
+			return static_cast<uint8_t>(Prospero::BlendFactor::kSrc1Alpha);
+		case Prospero::BlendFactor::kOneMinusSrcAlpha:
+			return static_cast<uint8_t>(Prospero::BlendFactor::kOneMinusSrc1Alpha);
+		default: return factor;
+	}
+}
+
 vk::PolygonMode ResolvePolygonMode(const HW::ModeControl& mode, bool cull_front, bool cull_back) {
 	// CxPrimitiveSetup::PolygonMode disables both per-face modes when it is zero.
 	if (mode.poly_mode == 0) {
@@ -92,8 +104,10 @@ void PipelineCacheLog(fmt::format_string<Args...> format, Args&&... args) {
 }
 
 bool ReadShaderGuestMemory(void*, uint64_t address, std::span<uint32_t> values) {
+	// Scalar and unformatted buffer dependencies use the same backing as native raw loads.
+	// Image synchronization belongs to formatted buffer bindings, not these reads.
 	return !values.empty() &&
-	       Libs::LibKernel::Memory::TryReadGpuCleanBacking(address, values.data(), values.size_bytes());
+	       Libs::LibKernel::Memory::TryReadBufferBacking(address, values.data(), values.size_bytes());
 }
 
 void DumpShaderSpirv(const char* stage_name, uint64_t shader_hash,
@@ -165,6 +179,32 @@ 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;
@@ -177,22 +217,57 @@ struct PipelineCache::ProgramCache {
 	};
 
 	struct Permutation {
+		std::vector<uint32_t>                       function_code;
 		ShaderRecompiler::IR::ResourceSpecialization specialization;
 		ShaderRecompiler::IR::CompiledShaderInfo     program;
 		ShaderProgram                                handle;
 	};
 
 	struct SourceEntry {
-		explicit SourceEntry(ShaderRecompiler::IR::ResourcePlan plan)
-		    : resource_plan(std::move(plan)) {
+		SourceEntry(ShaderRecompiler::IR::ResourcePlan plan,
+		            std::unique_ptr<ShaderRecompiler::ShaderSource> call_source)
+		    : call_source(std::move(call_source)), resource_plan(std::move(plan)) {
 			permutations.reserve(8);
 		}
 
+		std::span<const uint32_t> FunctionCode() const {
+			return call_source
+			    ? std::span<const uint32_t>(call_source->linked->code).subspan(call_source->code.size())
+			    : std::span<const uint32_t>{};
+		}
+
+		bool RefreshCallSource(const ShaderRecompiler::IR::SrtRuntime& runtime) {
+			const auto revision = call_source->revision;
+			ShaderRecompiler::RefreshShaderSource(*call_source, runtime);
+			resource_plan.source_reads = call_source->reads;
+			return revision != call_source->revision;
+		}
+
+		template <bool HasCalls>
+		Permutation* FindPermutation(const ShaderRecompiler::IR::SrtRuntime& runtime,
+		                             uint32_t push_data_cursor) {
+			EXIT_IF(!ShaderRecompiler::IR::MaterializeResources(
+			    resource_plan, runtime, resources, specialization));
+			for (auto& candidate: permutations) {
+				const auto& layout = candidate.program.bindings;
+				if (layout.push_data_start_dword != ShaderRecompiler::IR::PushData::StartFor(
+				        push_data_cursor, layout.ShaderDataDwords()) ||
+				    candidate.specialization != specialization) {
+					continue;
+				}
+				if constexpr (HasCalls) {
+					if (!std::ranges::equal(candidate.function_code, FunctionCode())) continue;
+				}
+				return &candidate;
+			}
+			return nullptr;
+		}
+
+		std::unique_ptr<ShaderRecompiler::ShaderSource> call_source;
 		ShaderRecompiler::IR::ResourcePlan           resource_plan;
 		ShaderRecompiler::IR::ResourceSnapshot       resources;
 		ShaderRecompiler::IR::ResourceSpecialization specialization;
 		std::vector<Permutation>                    permutations;
-		bool                                        skip_dispatch = false;
 	};
 
 	struct ProgramKeyHash {
@@ -206,18 +281,19 @@ 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 char*                                  stage_name,
 	                               const ShaderRecompiler::CompileOptions&      options,
 	                               ShaderRecompiler::TranslateResult            translated,
 	                               ShaderRecompiler::IR::ResourceSpecialization specialization,
-	                               uint32_t push_data_start_dword) {
+	                               uint32_t push_data_start_dword,
+	                               std::span<const uint32_t> function_code) {
 		auto result = ShaderRecompiler::CompileProgram(std::move(translated), options,
 		                                               specialization, push_data_start_dword);
 		if (!ValidateShaderSpirv(options.dump_label, options.shader_hash, result.spirv)) {
@@ -234,6 +310,7 @@ struct PipelineCache::ProgramCache {
 			     static_cast<uint64_t>(result.spirv.size()), options.wave_size);
 		}
 		return {
+		    .function_code  = {function_code.begin(), function_code.end()},
 		    .specialization = std::move(specialization),
 		    .program        = std::move(result.program).TakeCompiledInfo(),
 		    .handle         = {.id = ++next_shader_id, .module = module},
@@ -260,29 +337,26 @@ 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 {
+		ShaderRecompiler::IR::SrtRuntime             runtime {
 		    .user_data                  = user_data,
 		    .shader_base                = params.Base(),
 		    .read_specialization_memory = ReadShaderGuestMemory,
 		};
+		if constexpr (std::is_same_v<InputInfo, ShaderComputeInputInfo>) {
+			runtime.workgroup_counts = input_info.workgroup_counts;
+		}
+		bool source_changed = false;
 		if (entry != programs.end()) {
-			EXIT_IF(!ShaderRecompiler::IR::MaterializeResources(
-			    entry->second.resource_plan, runtime, entry->second.resources,
-			    entry->second.specialization));
-			if (const auto permutation = std::ranges::find_if(
-			        entry->second.permutations, [&](const Permutation& candidate) {
-				        const auto& layout = candidate.program.bindings;
-				        return layout.push_data_start_dword ==
-				                   ShaderRecompiler::IR::PushData::StartFor(
-				                       push_data_cursor, layout.ShaderDataDwords()) &&
-				               candidate.specialization == entry->second.specialization;
-			        });
-			    permutation != entry->second.permutations.end()) {
-				input_info.stage = {.program   = &permutation->program,
-				                    .resources = &entry->second.resources};
+			auto& cached = entry->second;
+			Permutation* permutation = nullptr;
+			if (cached.call_source) {
+				source_changed = cached.RefreshCallSource(runtime);
+				if (!source_changed) permutation = cached.FindPermutation<true>(runtime, push_data_cursor);
+			} else {
+				permutation = cached.FindPermutation<false>(runtime, push_data_cursor);
+			}
+			if (permutation != nullptr) {
+				input_info.stage = {.program = &permutation->program, .resources = &entry->second.resources};
 				permutation->program.bindings.AdvancePushData(push_data_cursor);
 				return permutation->handle;
 			}
@@ -329,21 +403,39 @@ struct PipelineCache::ProgramCache {
 			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 {};
+		ShaderRecompiler::ShaderSource source;
+		if (entry == programs.end()) source = ShaderRecompiler::PrepareShaderSource(params.code, options);
+		auto* current = entry == programs.end() ? &source : entry->second.call_source.get();
+		ShaderRecompiler::TranslateResult translated;
+		if (current != nullptr) {
+			if (current->call) {
+				if (entry == programs.end()) ShaderRecompiler::RefreshShaderSource(*current, runtime);
+				translated = ShaderRecompiler::TranslateProgram(current->linked->decoded, options);
+				translated.program.source_reads = current->reads;
+			} else {
+				translated = ShaderRecompiler::TranslateProgram(current->decoded, options);
+			}
+		} else {
+			translated = ShaderRecompiler::TranslateProgram(params.code, options);
 		}
 		if (entry == programs.end()) {
+			auto retained = source.call ? std::make_unique<ShaderRecompiler::ShaderSource>(std::move(source)) : nullptr;
 			entry = programs.try_emplace(lookup_key,
-			    ShaderRecompiler::IR::ExtractResourcePlan(translated.program)).first;
+			    ShaderRecompiler::IR::ExtractResourcePlan(translated.program), std::move(retained)).first;
 			EXIT_IF(!ShaderRecompiler::IR::MaterializeResources(
 			    entry->second.resource_plan, runtime, entry->second.resources,
 			    entry->second.specialization));
+		} else if (source_changed) {
+			entry->second.resource_plan = ShaderRecompiler::IR::ExtractResourcePlan(translated.program);
+			if (const auto* permutation = entry->second.FindPermutation<true>(runtime, push_data_cursor)) {
+				input_info.stage = {.program = &permutation->program, .resources = &entry->second.resources};
+				permutation->program.bindings.AdvancePushData(push_data_cursor);
+				return permutation->handle;
+			}
 		}
 		entry->second.permutations.push_back(CompilePermutation(
-		    stage_name, options, std::move(translated), entry->second.specialization, push_data_cursor));
+		    stage_name, options, std::move(translated), entry->second.specialization, push_data_cursor,
+		    entry->second.FunctionCode()));
 		const auto& permutation = entry->second.permutations.back();
 		input_info.stage = {.program = &permutation.program, .resources = &entry->second.resources};
 		permutation.program.bindings.AdvancePushData(push_data_cursor);
@@ -492,7 +584,6 @@ void PipelineCache::InitializeDriverCache() {
 }
 
 void PipelineCache::Save() {
-	Common::LockGuard lock(m_mutex);
 	if (m_driver_cache == nullptr) {
 		return;
 	}
@@ -583,21 +674,47 @@ 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);
+		if (!mesh_active && !tess_active && Prospero::IsRectList(user_config.GetPrimType())) {
+			pixel_info.parameter_mode = ShaderPixelParameterMode::Rectangle;
+		} else if (context.GetModeControl().provoking_vtx_last) {
+			pixel_info.parameter_mode = ShaderPixelParameterMode::LastVertex;
+		} else {
+			pixel_info.parameter_mode = ShaderPixelParameterMode::FirstVertex;
+		}
+		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; })) {
+			switch (ClassifyBlendMapping(blend, pixel_info.target_export_mapping[0])) {
+				case BlendMappingSupport::SourceAlpha:
+					pixel_info.alpha_blend_source = ShaderAlphaBlendSource::SourceAlpha;
+					break;
+				case BlendMappingSupport::SourceAlphaOne:
+					pixel_info.alpha_blend_source = ShaderAlphaBlendSource::SourceAlphaOne;
+					break;
+				case BlendMappingSupport::SourceAlphaZero:
+					pixel_info.alpha_blend_source = ShaderAlphaBlendSource::SourceAlphaZero;
+					break;
+				default: break;
+			}
+			if (pixel_info.alpha_blend_source != ShaderAlphaBlendSource::None) {
+				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) {
@@ -614,7 +731,6 @@ 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;
@@ -632,7 +748,8 @@ 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);
+	input_info.lds_storage = input_info.lds_size_dwords * 4u >
+	    m_graphics.GetPhysicalDeviceProperties().limits.maxComputeSharedMemorySize;
 	uint32_t          push_data_cursor = 0;
 	return m_program_cache->Get(params, input_info, push_data_cursor);
 }
@@ -657,7 +774,6 @@ 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();
@@ -690,14 +806,51 @@ 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;
+		auto alpha_source = ShaderAlphaBlendSource::None;
+		if (slot == 0 && ps_input_info != nullptr) {
+			alpha_source = ps_input_info->alpha_blend_source;
+		}
+		static_params.blend_enable[slot] = bc.enable && !rt.info.blend_bypass;
+		if (static_params.blend_enable[slot] && alpha_source == ShaderAlphaBlendSource::None &&
+		    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 (static_params.blend_enable[slot]) {
+			auto blend = bc;
+			switch (alpha_source) {
+				case ShaderAlphaBlendSource::SourceAlpha:
+					blend.color_srcblend  = RemapSourceAlphaFactor(blend.color_srcblend);
+					blend.color_destblend = RemapSourceAlphaFactor(blend.color_destblend);
+					blend.separate_alpha_blend = false;
+					break;
+				case ShaderAlphaBlendSource::SourceAlphaOne:
+				case ShaderAlphaBlendSource::SourceAlphaZero:
+					// The second source carries the mapped source factor; its alpha stays logical Sa.
+					blend.color_srcblend = static_cast<uint8_t>(Prospero::BlendFactor::kSrc1Color);
+					blend.color_destblend =
+					    static_cast<uint8_t>(Prospero::BlendFactor::kOneMinusSrc1Alpha);
+					blend.separate_alpha_blend = false;
+					break;
+				case ShaderAlphaBlendSource::None: break;
+			}
+			static_params.color_srcblend[slot]       = blend.color_srcblend;
+			static_params.color_comb_fcn[slot]       = blend.color_comb_fcn;
+			static_params.color_destblend[slot]      = blend.color_destblend;
+			static_params.separate_alpha_blend[slot] = blend.separate_alpha_blend;
+			if (blend.separate_alpha_blend) {
+				static_params.alpha_srcblend[slot]  = blend.alpha_srcblend;
+				static_params.alpha_comb_fcn[slot]  = blend.alpha_comb_fcn;
+				static_params.alpha_destblend[slot] = blend.alpha_destblend;
+			}
+		}
 	}
 	const bool with_depth =
 	    depth.desc.view_info.format != vk::Format::eUndefined && static_cast<bool>(depth.image_id);
@@ -743,9 +896,6 @@ PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(
 	if (static_params.sample_shading_enable && !m_graphics.sample_rate_shading_enabled) {
 		EXIT("Pipeline: sample-rate shading is required but unsupported by the host\n");
 	}
-	static_params.depth_bounds_test_enable = depth.depth_bounds_test_enable;
-	static_params.depth_min_bounds         = depth.depth_min_bounds;
-	static_params.depth_max_bounds         = depth.depth_max_bounds;
 	const bool rect_list = Prospero::IsRectList(command.GetUserConfig().GetPrimType());
 	static_params.cull_back  = !rect_list && mc.cull_back;
 	static_params.cull_front = !rect_list && mc.cull_front;
@@ -761,24 +911,20 @@ 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()) {
@@ -817,8 +963,6 @@ 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"
@@ -39,9 +38,6 @@ struct PipelineStaticParameters {
 	bool                       primitive_restart_enable = false;
 	uint32_t                   samples                  = 1;
 	bool                       sample_shading_enable    = false;
-	bool                       depth_bounds_test_enable = false;
-	float                      depth_min_bounds         = 0.0f;
-	float                      depth_max_bounds         = 0.0f;
 	uint32_t                   color_mask[RENDER_COLOR_ATTACHMENTS_MAX]           = {};
 	bool                       cull_front                                         = false;
 	bool                       cull_back                                          = false;
@@ -65,7 +61,7 @@ struct PipelineStaticParameters {
 static_assert(std::is_trivially_copyable_v<PipelineStaticParameters>);
 static_assert(std::is_standard_layout_v<PipelineStaticParameters>);
 static_assert(alignof(PipelineStaticParameters) == 1);
-static_assert(sizeof(PipelineStaticParameters) == 125);
+static_assert(sizeof(PipelineStaticParameters) == 116);
 
 struct PipelineRenderingState {
 	std::array<vk::Format, RENDER_COLOR_ATTACHMENTS_MAX> color_formats {};
@@ -103,6 +99,7 @@ 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 +163,10 @@ 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 +176,6 @@ 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

@@ -19,6 +19,7 @@
 #include "graphics/shader/shader.h"
 
 #include <algorithm>
+#include <bit>
 #include <limits>
 #include <span>
 #include <vector>
@@ -180,10 +181,12 @@ static vk::BlendOp GetBlendOp(uint32_t op) {
 static void AddLayoutBindings(std::vector<vk::DescriptorSetLayoutBinding>& descriptor_bindings,
                               const ShaderRecompiler::IR::CompiledShaderInfo& program,
                               vk::ShaderStageFlagBits              stage) {
-	for (const auto& binding: program.bindings.descriptors) {
+	for (auto mask = program.bindings.descriptor_mask; mask != 0; mask &= mask - 1) {
+		const auto index = std::countr_zero(mask);
+		const auto kind = static_cast<ShaderRecompiler::IR::DescriptorBindingKind>(index);
 		descriptor_bindings.push_back(
-		    {ShaderRecompiler::IR::NativeBinding(program.stage, binding.kind),
-		     NativeDescriptorType(binding.kind), NativeDescriptorCount(binding), stage, nullptr});
+		    {ShaderRecompiler::IR::NativeBinding(program.stage, kind), NativeDescriptorType(kind),
+		     program.bindings.descriptor_counts[index], stage, nullptr});
 	}
 }
 
@@ -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);
 	}
 
@@ -403,8 +403,9 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 		    GetBlendFactor(static_params.color_destblend[i]);
 		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])
+		         : color_blend_attachment[i].srcColorBlendFactor);
 		color_blend_attachment[i].dstAlphaBlendFactor =
 		    (static_params.separate_alpha_blend[i]
 		         ? GetBlendFactor(static_params.alpha_destblend[i])
@@ -473,14 +474,6 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	EXIT_NOT_IMPLEMENTED(pipeline.pipeline_layout == nullptr);
 
 	vk::PipelineDepthStencilStateCreateInfo depth_stencil_info {};
-	depth_stencil_info.depthBoundsTestEnable =
-#if defined(__APPLE__)
-	    VK_FALSE; // MoltenVK lacks the depthBounds feature; depth-bounds testing is disabled
-#else
-	    (static_params.depth_bounds_test_enable ? VK_TRUE : VK_FALSE);
-#endif
-	depth_stencil_info.minDepthBounds    = static_params.depth_min_bounds;
-	depth_stencil_info.maxDepthBounds    = static_params.depth_max_bounds;
 
 	std::vector<vk::DynamicState> dynamic_states {
 	    vk::DynamicState::eViewportWithCount,
@@ -499,6 +492,8 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	    vk::DynamicState::eBlendConstants,
 	};
 #if !defined(__APPLE__)
+	dynamic_states.push_back(vk::DynamicState::eDepthBoundsTestEnable);
+	dynamic_states.push_back(vk::DynamicState::eDepthBounds);
 	if (rendering.color_count != 0) {
 		dynamic_states.push_back(vk::DynamicState::eColorWriteEnableEXT);
 	}

src/graphics/host_gpu/renderer/render.h

@@ -156,12 +156,12 @@ public:
 	KYTY_CLASS_NO_COPY(RenderExecutor);
 
 	void DispatchDirect(uint64_t submit_id, CommandBuffer& buffer, uint32_t thread_group_x,
-	                    uint32_t thread_group_y, uint32_t thread_group_z, uint32_t mode);
+	                    uint32_t thread_group_y, uint32_t thread_group_z, uint32_t mode, bool async_compute);
 	void DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer, uint64_t args_addr,
-	                      uint32_t mode);
+	                      uint32_t mode, bool async_compute);
 
 	void PrepareBindings(const ShaderStageRuntime& runtime, PreparedBindings& prepared);
-	void                           FindBuffers(PreparedBindings& bindings);
+	void                           FindBuffers(std::span<PreparedBindings* const> stages);
 	void                           RebindBuffers(PreparedBindings& bindings);
 	void                           RebindImages(PreparedBindings& bindings);
 	void CommitBindings(CommandBuffer& buffer, vk::PipelineBindPoint pipeline_bind_point,
@@ -213,10 +213,8 @@ private:
 	GraphicsBindings                     m_graphics_bindings;
 	PreparedBindings                     m_compute_bindings;
 	std::vector<ImageId>                  m_bound_images;
-	std::vector<vk::DescriptorBufferInfo> m_descriptor_buffers;
 	std::vector<vk::DescriptorImageInfo>  m_descriptor_images;
 	std::vector<vk::WriteDescriptorSet>   m_descriptor_writes;
-	std::vector<uint32_t>                 m_image_occurrences;
 
 	friend class CommandProcessor;
 	friend struct RenderExecutorTestAccess;

src/graphics/host_gpu/renderer/renderCompute.cpp

@@ -195,9 +195,38 @@ bool RenderExecutor::TryConsumeComputeImageClear(const ShaderComputeInputInfo& i
 	return true;
 }
 
+static void BindSharedMemory(RenderContext& context, ShaderComputeInputInfo& input,
+                             PreparedBindings& bindings, uint64_t indirect_args = 0) {
+	if (input.stage.program->bindings.descriptor_counts[
+	        static_cast<size_t>(ShaderRecompiler::IR::DescriptorBindingKind::SharedMemory)] == 0) {
+		return;
+	}
+	auto& cache = context.GetBufferCache();
+	if (indirect_args != 0) {
+		cache.ReadMemory(indirect_args, sizeof(vk::DispatchIndirectCommand));
+		std::memcpy(input.workgroup_counts, reinterpret_cast<const void*>(indirect_args),
+		            sizeof(input.workgroup_counts));
+	}
+	// LDS has no contents to preserve between dispatches. The existing shader hazard
+	// barriers also order other users of this GPU-only utility buffer.
+	auto& storage = cache.GetUtilityBuffer(MemoryUsage::DeviceLocal);
+	const auto limit = std::min<uint64_t>(storage.Size(),
+	    context.GetGraphics().GetPhysicalDeviceProperties().limits.maxStorageBufferRange);
+	uint64_t size = sizeof(uint32_t);
+	if (std::ranges::find(input.workgroup_counts, 0u) == std::end(input.workgroup_counts)) {
+		size = uint64_t {input.lds_size_dwords} * sizeof(uint32_t);
+		EXIT_IF(size == 0 || size > limit);
+		for (const auto count: input.workgroup_counts) {
+			EXIT_IF(size > limit / count);
+			size *= count;
+		}
+	}
+	bindings.shared_memory = {storage.Handle(), 0, size};
+}
+
 void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
                                     uint32_t thread_group_x, uint32_t thread_group_y,
-                                    uint32_t thread_group_z, uint32_t mode) {
+                                    uint32_t thread_group_z, uint32_t mode, bool async_compute) {
 	EXIT_IF(buffer.IsInvalid());
 	m_context.GetCommandScheduler().PopPendingOperations();
 	auto& ctx    = buffer.GetRegisters();
@@ -252,14 +281,23 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 	const auto& sh_regs = ctx.GetShaderRegisters();
 
 	ShaderComputeInputInfo input_info {};
+	input_info.async_compute = async_compute;
 	const bool use_thread_dimensions = (mode & DISPATCH_INITIATOR_USE_THREAD_DIMENSIONS) != 0;
 	input_info.dispatch_thread_dimensions = use_thread_dimensions;
+	input_info.workgroup_counts[0] = thread_group_x;
+	input_info.workgroup_counts[1] = thread_group_y;
+	input_info.workgroup_counts[2] = thread_group_z;
+	if (use_thread_dimensions) {
+		const uint32_t group_sizes[] = {cs_regs.cs_regs.num_thread_x, cs_regs.cs_regs.num_thread_y,
+		                                cs_regs.cs_regs.num_thread_z};
+		for (uint32_t axis = 0; axis < 3u; ++axis) {
+			const auto size = std::max(group_sizes[axis], 1u);
+			const auto threads = input_info.workgroup_counts[axis];
+			input_info.workgroup_counts[axis] = threads / size + (threads % size != 0u);
+		}
+	}
 	const auto compute_program =
 	    m_context.GetPipelineCache().GetComputeProgram(cs_regs, sh_regs, input_info);
-	if (!compute_program) {
-		// 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;
@@ -268,31 +306,83 @@ 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)) {
+	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;
 	}
-	if (TryConsumeComputeImageClear(input_info, buffer, thread_group_x, thread_group_y,
-	                                thread_group_z, mode)) {
-		ResetBindings();
-		return;
+
+	if (use_thread_dimensions) {
+		const uint32_t old_x = thread_group_x;
+		const uint32_t old_y = thread_group_y;
+		const uint32_t old_z = thread_group_z;
+		thread_group_x       = input_info.workgroup_counts[0];
+		thread_group_y       = input_info.workgroup_counts[1];
+		thread_group_z       = input_info.workgroup_counts[2];
+
+		static std::atomic<uint32_t> log_count {0};
+		if (log_count.fetch_add(1, std::memory_order_relaxed) < 32) {
+			LOGF("GraphicsRenderDispatchDirect: use-thread-dimensions %ux%ux%u / %ux%ux%u -> "
+			     "groups %ux%ux%u\n",
+			     old_x, old_y, old_z, std::max(cs_regs.cs_regs.num_thread_x, 1u),
+			     std::max(cs_regs.cs_regs.num_thread_y, 1u),
+			     std::max(cs_regs.cs_regs.num_thread_z, 1u), thread_group_x, thread_group_y,
+			     thread_group_z);
+		}
 	}
-	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();
-	static std::atomic<uint32_t> dispatch_log_count {0};
-	if ((large_workgroup || has_sampler) &&
-	    dispatch_log_count.fetch_add(1, std::memory_order_relaxed) < 512) {
+
+	buffer.EndRendering();
+	auto& pipeline =
+	    m_context.GetPipelineCache().GetComputePipeline(input_info, compute_program);
+	auto& bindings = m_compute_bindings;
+	PrepareBindings(input_info.stage, bindings);
+	if (program.bindings.dispatch_thread_dword != ShaderRecompiler::IR::PushData::NoStart) {
+		std::copy(std::begin(input_info.dispatch_threads_num), std::end(input_info.dispatch_threads_num),
+		          bindings.shader_data.begin() + program.bindings.dispatch_thread_dword);
+	}
+	PreparedBindings* descriptor_stage = &bindings;
+	FindBuffers(std::span {&descriptor_stage, 1u});
+	if (program.info.uses_dma) {
+		m_context.PrepareBda();
+	}
+	RebindImages(bindings);
+	BindSharedMemory(m_context, input_info, bindings);
+	RebindBuffers(bindings);
+
+	auto              vk_buffer        = buffer.Handle();
+	CommitBindings(buffer, vk::PipelineBindPoint::eCompute, pipeline,
+	               std::span {&descriptor_stage, 1u});
+	bool has_storage_writes = bindings.shared_memory.buffer != nullptr ||
+	    HasShaderBufferWrites(input_info.stage);
+	has_storage_writes =
+	    std::any_of(program.info.images.begin(), program.info.images.end(),
+	                [](const auto& image) {
+		                return image.written &&
+		                       image.resource_class ==
+		                           ShaderRecompiler::IR::ImageResourceClass::Storage;
+	                }) ||
+	    has_storage_writes;
+	if (has_storage_writes) {
+		// A host fence used to serialize every dispatch. Preserve its read-before-write ordering
+		// while allowing the queue to execute asynchronously.
+		ShaderWriteHazardBarrier(vk_buffer, vk::PipelineStageFlagBits::eComputeShader);
+	}
+	if (Config::GraphicsDebugDumpEnabled()) {
 		const auto sampled_images = std::count_if(
 		    program.info.images.begin(), program.info.images.end(), [](const auto& image) {
 			    return image.resource_class == ShaderRecompiler::IR::ImageResourceClass::Sampled;
 		    });
 		const uint32_t frame_num = static_cast<uint32_t>(m_context.GetGpu().GetFrameNum());
 		LOGF("GraphicsRenderDispatchDirect: frame=%u shader=0x%016" PRIx64
+		     " hash=0x%016" PRIx64 " tick=%" PRIu64
 		     " groups=%ux%ux%u mode=0x%08" PRIx32 " local=%ux%ux%u "
 		     "buffers=%zu textures=%zu sampled=%zu storage=%zu samplers=%zu push=%u\n",
-		     frame_num, sh_ctx.GetCs().cs_regs.data_addr, thread_group_x, thread_group_y,
-		     thread_group_z, mode, input_info.threads_num[0], input_info.threads_num[1],
+		     frame_num, sh_ctx.GetCs().cs_regs.data_addr, program.shader_hash,
+		     m_context.GetCommandScheduler().CurrentTick(),
+		     thread_group_x, thread_group_y, thread_group_z, mode,
+		     input_info.threads_num[0], input_info.threads_num[1],
 		     input_info.threads_num[2], program.info.buffers.size(), program.info.images.size(),
 		     sampled_images, program.info.images.size() - sampled_images,
 		     program.info.samplers.size(),
@@ -326,7 +416,8 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 			     static_cast<uint32_t>(r.TileMode()));
 		}
 		for (uint32_t i = 0; i < program.info.samplers.size(); i++) {
-			const auto r = DecodeNativeDescriptor<ShaderSamplerResource>(resources.samplers[i]);
+			const auto r = DecodeNativeDescriptor<ShaderSamplerResource>(
+			    resources.samplers[program.info.samplers[i].snapshot_index]);
 			LOGF("  CS sampler[%u]: source=%u clamp=%u/%u/%u filter=%u/%u/%u mip=%u "
 			     "lod=%u-%u bias=%d\n",
 			     i, program.info.samplers[i].source, static_cast<uint32_t>(r.ClampX()),
@@ -338,61 +429,6 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 		}
 	}
 
-	if (use_thread_dimensions) {
-		auto groups_from_threads = [](uint32_t threads, uint32_t group_size) {
-			return (threads == 0
-			            ? 0u
-			            : (threads + std::max(group_size, 1u) - 1u) / std::max(group_size, 1u));
-		};
-
-		const uint32_t old_x = thread_group_x;
-		const uint32_t old_y = thread_group_y;
-		const uint32_t old_z = thread_group_z;
-		thread_group_x       = groups_from_threads(thread_group_x, cs_regs.cs_regs.num_thread_x);
-		thread_group_y       = groups_from_threads(thread_group_y, cs_regs.cs_regs.num_thread_y);
-		thread_group_z       = groups_from_threads(thread_group_z, cs_regs.cs_regs.num_thread_z);
-
-		static std::atomic<uint32_t> log_count {0};
-		if (log_count.fetch_add(1, std::memory_order_relaxed) < 32) {
-			LOGF("GraphicsRenderDispatchDirect: use-thread-dimensions %ux%ux%u / %ux%ux%u -> "
-			     "groups %ux%ux%u\n",
-			     old_x, old_y, old_z, std::max(cs_regs.cs_regs.num_thread_x, 1u),
-			     std::max(cs_regs.cs_regs.num_thread_y, 1u),
-			     std::max(cs_regs.cs_regs.num_thread_z, 1u), thread_group_x, thread_group_y,
-			     thread_group_z);
-		}
-	}
-
-	buffer.EndRendering();
-	auto& pipeline =
-	    m_context.GetPipelineCache().GetComputePipeline(input_info, compute_program);
-	auto& bindings = m_compute_bindings;
-	PrepareBindings(input_info.stage, bindings);
-	FindBuffers(bindings);
-	if (program.info.uses_dma) {
-		m_context.PrepareBda();
-	}
-	RebindImages(bindings);
-	RebindBuffers(bindings);
-
-	auto              vk_buffer        = buffer.Handle();
-	PreparedBindings* descriptor_stage = &bindings;
-	CommitBindings(buffer, vk::PipelineBindPoint::eCompute, pipeline,
-	               std::span {&descriptor_stage, 1u});
-	bool has_storage_writes = HasShaderBufferWrites(input_info.stage);
-	has_storage_writes =
-	    std::any_of(program.info.images.begin(), program.info.images.end(),
-	                [](const auto& image) {
-		                return image.written &&
-		                       image.resource_class ==
-		                           ShaderRecompiler::IR::ImageResourceClass::Storage;
-	                }) ||
-	    has_storage_writes;
-	if (has_storage_writes) {
-		// A host fence used to serialize every dispatch. Preserve its read-before-write ordering
-		// while allowing the queue to execute asynchronously.
-		ShaderWriteHazardBarrier(vk_buffer, vk::PipelineStageFlagBits::eComputeShader);
-	}
 	vk_buffer.bindPipeline(vk::PipelineBindPoint::eCompute, pipeline.pipeline);
 	vk_buffer.dispatch(thread_group_x, thread_group_y, thread_group_z);
 
@@ -402,7 +438,7 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 }
 
 void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
-                                      uint64_t args_addr, uint32_t mode) {
+                                      uint64_t args_addr, uint32_t mode, bool async_compute) {
 	EXIT_IF(buffer.IsInvalid() || args_addr == 0 || (args_addr & 3u) != 0 ||
 	        (mode & Pm4::COMPUTE_DISPATCH_INITIATOR_USE_THREAD_DIMENSIONS) != 0);
 	m_context.GetCommandScheduler().PopPendingOperations();
@@ -415,32 +451,31 @@ void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
 		return;
 	}
 	ShaderComputeInputInfo input_info {};
+	input_info.async_compute = async_compute;
 	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;
 	PrepareBindings(input_info.stage, bindings);
-	FindBuffers(bindings);
+	PreparedBindings* descriptor_stage = &bindings;
+	FindBuffers(std::span {&descriptor_stage, 1u});
 	const auto& program = *input_info.stage.program;
 	if (program.info.uses_dma) {
 		m_context.PrepareBda();
 	}
+	BindSharedMemory(m_context, input_info, bindings, args_addr);
 	RebindImages(bindings);
 	// Acquiring arguments can merge cache buffers; finalize shader bindings afterward.
 	const auto [args_buffer, args_offset] = m_context.GetBufferCache().ObtainBuffer(
 	    args_addr, sizeof(vk::DispatchIndirectCommand), false);
 	EXIT_IF(args_buffer == nullptr || (args_offset & 3u) != 0);
 	RebindBuffers(bindings);
-	PreparedBindings* descriptor_stage = &bindings;
 	CommitBindings(buffer, vk::PipelineBindPoint::eCompute, pipeline,
 	               std::span {&descriptor_stage, 1u});
 	const auto vk_buffer = buffer.Handle();
-	const bool has_storage_writes = HasShaderBufferWrites(input_info.stage) ||
+	const bool has_storage_writes = bindings.shared_memory.buffer != nullptr ||
+	    HasShaderBufferWrites(input_info.stage) ||
 	    std::any_of(program.info.images.begin(), program.info.images.end(), [](const auto& image) {
 		    return image.written && image.resource_class ==
 		                                ShaderRecompiler::IR::ImageResourceClass::Storage;

src/graphics/host_gpu/renderer/renderContext.cpp

@@ -99,7 +99,12 @@ void RenderContext::UnmapMemory(uint64_t vaddr, uint64_t size) {
 		     vaddr, size);
 	}
 	const auto unmap = [this, vaddr, size] {
-		if (m_command_scheduler.Active()) {
+		// Check cache ownership on the GPU thread. Guest-memory callbacks can still
+		// access a range with no cached data, so they must finish before it is unmapped.
+		if (m_command_scheduler.Active() &&
+		    (m_buffer_cache.IsRegionRegistered(vaddr, size) ||
+		     m_texture_cache.IsRegionRegistered(vaddr, size) ||
+		     m_command_scheduler.HasPendingPriorityOperations())) {
 			const auto tick = m_command_scheduler.CurrentTick();
 			m_command_scheduler.Finish();
 			m_command_scheduler.WaitPriorityOperations(tick);
@@ -119,6 +124,10 @@ 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]);
 		}
 	}
 }
@@ -375,6 +379,12 @@ static void SetGraphicsDynamicParams(const CommandBuffer& buffer, vk::CommandBuf
 	vk_buffer.setDepthTestEnable(depth.depth_test_enable ? VK_TRUE : VK_FALSE);
 	vk_buffer.setDepthWriteEnable(depth.depth_write_enable ? VK_TRUE : VK_FALSE);
 	vk_buffer.setDepthCompareOp(depth.depth_compare_op);
+#if !defined(__APPLE__)
+	vk_buffer.setDepthBoundsTestEnable(depth.depth_bounds_test_enable ? VK_TRUE : VK_FALSE);
+	if (depth.depth_bounds_test_enable) {
+		vk_buffer.setDepthBounds(ctx.GetDepthBoundsMin(), ctx.GetDepthBoundsMax());
+	}
+#endif
 
 	const auto& mode              = ctx.GetModeControl();
 	const auto& poly_offset       = ctx.GetPolyOffset();
@@ -579,8 +589,12 @@ static uint32_t DrawColorOutputMask(const HW::Context& ctx) {
 	const auto& sh_regs     = ctx.GetShaderRegisters();
 	const auto  write_mask  = ctx.GetRenderTargetMask() & sh_regs.m_cbShaderMask;
 	uint32_t    output_mask = 0;
-	for (uint32_t slot = 0; slot < RENDER_COLOR_ATTACHMENTS_MAX; slot++) {
-		if (sh_regs.target_output_mode[slot] != 0 &&
+	for (uint32_t index = 0; index < RENDER_COLOR_ATTACHMENTS_MAX; index++) {
+		const auto slot = ShaderPixelExportTarget(sh_regs.m_cbShaderMask, index);
+		if (slot >= RENDER_COLOR_ATTACHMENTS_MAX) {
+			break;
+		}
+		if (sh_regs.target_output_mode[index] != 0 &&
 		    render_target_mask_slot(write_mask, slot) != 0) {
 			output_mask |= 1u << slot;
 		}
@@ -630,19 +644,22 @@ 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);
@@ -679,7 +696,7 @@ 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;
@@ -718,9 +735,6 @@ 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.
@@ -752,10 +766,6 @@ 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;
@@ -784,6 +794,7 @@ static bool GetDrawTopology(const HW::UserConfig& ucfg, vk::PrimitiveTopology& t
 			topology = vk::PrimitiveTopology::eTriangleList;
 			break;
 		case Prospero::PrimitiveType::kTriFan:
+		case Prospero::PrimitiveType::kPolygon:
 			topology = vk::PrimitiveTopology::eTriangleFan;
 			break;
 		case Prospero::PrimitiveType::kTriStrip:
@@ -895,7 +906,7 @@ bool RenderExecutor::PrepareDrawRenderState(CommandBuffer& buffer, const DrawCal
 	if (state.ps_active) {
 		for (const auto& output: state.ps_input_info.stage.program->info.outputs) {
 			if (output.kind == ShaderRecompiler::IR::StageOutputKind::Mrt) {
-				mrt_mask |= 1u << output.index;
+				mrt_mask |= 1u << output.location;
 			}
 		}
 	}
@@ -937,17 +948,11 @@ 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;
 }
@@ -1000,6 +1005,7 @@ static void EmitDrawPrimitives(const HW::UserConfig& ucfg, vk::CommandBuffer vk_
 		case Prospero::PrimitiveType::kTriFan:
 		case Prospero::PrimitiveType::kTriStrip:
 		case Prospero::PrimitiveType::kRectList:
+		case Prospero::PrimitiveType::kPolygon:
 		case Prospero::PrimitiveType::kRectListLegacy:
 		case Prospero::PrimitiveType::kPatch:
 			if (draw.IsIndexed()) {
@@ -1038,6 +1044,7 @@ 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; "
@@ -1277,8 +1284,10 @@ void RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
 	             : ResolveDrawOffsets(ucfg.GetIndexOffset(), state.vertex_info[0]);
 
 	DrawEmitInfo emit {};
-	emit.vertex_offset  = vertex_offset + args.base_vertex;
-	emit.first_instance = instance_offset;
+	// Native fetches consume the patched SGPRs; rewritten fetches need Vulkan offsets.
+	const bool native_indirect = indirect && !state.vertex_info[0].fetch_embedded;
+	emit.vertex_offset  = native_indirect ? 0 : vertex_offset + args.base_vertex;
+	emit.first_instance = native_indirect ? 0 : instance_offset;
 
 	ExecutePreparedDraw(submit_id, buffer, draw, state, topology, emit, index_source,
 	                    primitive_restart);
@@ -1366,8 +1375,10 @@ void RenderExecutor::DrawAuto(uint64_t submit_id, CommandBuffer& buffer, const D
 	    indirect ? std::pair<int32_t, uint32_t> {0, args.first_instance}
 	             : ResolveDrawOffsets(ucfg.GetIndexOffset(), state.vertex_info[0]);
 	DrawEmitInfo emit {};
-	emit.first_vertex = static_cast<uint32_t>(vertex_offset + static_cast<int32_t>(args.first_vertex));
-	emit.first_instance = instance_offset;
+	const bool native_indirect = indirect && !state.vertex_info[0].fetch_embedded;
+	emit.first_vertex = native_indirect ? 0 :
+	    static_cast<uint32_t>(vertex_offset + static_cast<int32_t>(args.first_vertex));
+	emit.first_instance = native_indirect ? 0 : instance_offset;
 
 	DrawIndexBufferSource index_source {};
 	ExecutePreparedDraw(submit_id, buffer, draw, state, topology, emit, index_source, false);

src/graphics/host_gpu/shaders/gpu_tiler_color.inc

@@ -0,0 +1,20 @@
+// ColorTransform values match imageInfo.h. Preserve the floating-point bit fields exactly.
+uvec2 transform_color(uvec2 value, uint transform, bool to_host) {
+	switch (transform) {
+		case 1u:
+			return uvec2((value.y & 0xffffu) | (value.x & 0xffff0000u),
+			             (value.x & 0xffffu) | (value.y & 0xffff0000u));
+		case 2u: {
+			uint packed = value.x;
+			if (to_host) {
+				value.x = (packed >> 21u) | (((packed >> 10u) & 0x7ffu) << 11u) |
+				          ((packed & 0x3ffu) << 22u);
+			} else {
+				value.x = (packed >> 22u) | (((packed >> 11u) & 0x7ffu) << 10u) |
+				          ((packed & 0x7ffu) << 21u);
+			}
+			break;
+		}
+	}
+	return value;
+}

src/graphics/host_gpu/shaders/gpu_tiler_color_transform.comp

@@ -1,4 +1,5 @@
 #version 450
+#extension GL_GOOGLE_include_directive : require
 
 layout(local_size_x = 64) in;
 
@@ -19,17 +20,23 @@ layout(push_constant, std430) uniform Push {
 	uint tail_x;
 	uint tail_y;
 	uint tail;
+	uint color_transform;
+	uint to_host;
 } params;
 
+#include "gpu_tiler_color.inc"
+
 void main() {
 	uint pixel = gl_GlobalInvocationID.x;
 	if (pixel >= params.width) {
 		return;
 	}
-	uint src = (params.src_base >> 2u) + pixel * 2u;
-	uint dst = (params.dst_base >> 2u) + pixel * 2u;
-	uint bg = input_buffer.data[src];
-	uint ra = input_buffer.data[src + 1u];
-	output_buffer.data[dst] = (ra & 0x0000ffffu) | (bg & 0xffff0000u);
-	output_buffer.data[dst + 1u] = (bg & 0x0000ffffu) | (ra & 0xffff0000u);
+	uint words = params.color_transform == 1u ? 2u : 1u;
+	uint src = (params.src_base >> 2u) + pixel * words;
+	uint dst = (params.dst_base >> 2u) + pixel * words;
+	uvec2 value = uvec2(input_buffer.data[src], 0u);
+	if (words == 2u) value.y = input_buffer.data[src + 1u];
+	value = transform_color(value, params.color_transform, params.to_host != 0u);
+	output_buffer.data[dst] = value.x;
+	if (words == 2u) output_buffer.data[dst + 1u] = value.y;
 }

src/graphics/host_gpu/shaders/gpu_tiler_common.inc

@@ -17,10 +17,19 @@ layout(set = 0, binding = 2, std140) uniform Push {
 	uint tail_x;
 	uint tail_y;
 	uint tail;
+	uint color_transform;
 } params;
 
+#include "gpu_tiler_color.inc"
+
 void copy_element(uint src, uint dst) {
-	if (ELEMENT_BYTES >= 4) {
+	if ((ELEMENT_BYTES == 4u || ELEMENT_BYTES == 8u) && params.color_transform != 0u) {
+		uvec2 value = uvec2(input_buffer.data[src >> 2], 0u);
+		if (ELEMENT_BYTES == 8u) value.y = input_buffer.data[(src >> 2) + 1u];
+		value = transform_color(value, params.color_transform, TILE == 0u);
+		output_buffer.data[dst >> 2] = value.x;
+		if (ELEMENT_BYTES == 8u) output_buffer.data[(dst >> 2) + 1u] = value.y;
+	} else if (ELEMENT_BYTES >= 4) {
 		for (uint i = 0; i < ELEMENT_BYTES; i += 4) {
 			output_buffer.data[(dst + i) >> 2] = input_buffer.data[(src + i) >> 2];
 		}

src/graphics/host_gpu/vulkanCommon.cpp

@@ -48,6 +48,7 @@ constexpr FormatMapping kFormatMappings[] = {
     {Prospero::BufferFormat::k11_11_10Float, vk::Format::eB10G11R11UfloatPack32},
     {Prospero::BufferFormat::k10_10_10_2UNorm, vk::Format::eA2B10G10R10UnormPack32},
     {Prospero::BufferFormat::k10_10_10_2SNorm, vk::Format::eA2B10G10R10SnormPack32},
+    {Prospero::BufferFormat::k10_10_10_2UScaled, vk::Format::eA2B10G10R10UscaledPack32},
     {Prospero::BufferFormat::k10_10_10_2UInt, vk::Format::eA2B10G10R10UintPack32},
     {Prospero::BufferFormat::k8_8_8_8UNorm, vk::Format::eR8G8B8A8Unorm},
     {Prospero::BufferFormat::k8_8_8_8SNorm, vk::Format::eR8G8B8A8Snorm},

src/graphics/presentation/systemOverlay.cpp

@@ -3,14 +3,17 @@
 #include <SDL3/SDL.h>
 
 #include "common/assert.h"
+#include "common/logging/log.h"
 #include "common/stringUtils.h"
 #include "graphics/host_gpu/graphicContext.h"
 #include "imgui.h"
+#include "imgui_internal.h"
 #include "imgui_impl_vulkan.h"
 #include "libs/controller.h"
 #include "libs/dialog.h"
 #include "libs/ime.h"
 #include "libs/imeDialog.h"
+#include "stb_image.h"
 
 #include <algorithm>
 #include <array>
@@ -18,20 +21,25 @@
 #include <chrono>
 #include <cmath>
 #include <cstdint>
+#include <cstring>
 #include <deque>
+#include <limits>
+#include <memory>
 #include <mutex>
+#include <optional>
 #include <span>
 #include <string>
 #include <string_view>
 #include <utility>
+#include <vector>
 
 namespace Libs::Graphics {
 
 namespace {
 
-namespace CoreIme   = Libs::Ime;
-namespace DialogIme = Libs::Dialog::ImeDialog;
-namespace ErrorDialog = Libs::Dialog::ErrorDialog;
+namespace CoreIme      = Libs::Ime;
+namespace DialogIme    = Libs::Dialog::ImeDialog;
+namespace SystemDialog = Libs::Dialog::SystemDialog;
 
 namespace Ime {
 
@@ -104,7 +112,7 @@ bool HostQueueExternalInput(uint64_t generation, ExternalInput input) {
 
 } // namespace Ime
 
-enum class OverlayKind : uint8_t { None, Ime, Error };
+enum class OverlayKind : uint8_t { None, Ime, Dialog };
 
 struct OverlaySession {
 	OverlayKind kind       = OverlayKind::None;
@@ -114,14 +122,14 @@ struct OverlaySession {
 };
 
 struct OverlaySnapshot {
-	OverlaySession            session;
-	Ime::HostSnapshot         ime;
-	ErrorDialog::HostSnapshot error;
+	OverlaySession             session;
+	Ime::HostSnapshot          ime;
+	SystemDialog::HostSnapshot dialog;
 };
 
 bool GetOverlaySnapshot(OverlaySnapshot* snapshot) {
-	if (ErrorDialog::GetHostSnapshot(&snapshot->error)) {
-		snapshot->session = {OverlayKind::Error, snapshot->error.generation};
+	if (SystemDialog::GetHostSnapshot(&snapshot->dialog)) {
+		snapshot->session = {OverlayKind::Dialog, snapshot->dialog.generation};
 		return true;
 	}
 	if (Ime::GetHostSnapshot(&snapshot->ime)) {
@@ -178,6 +186,45 @@ bool                         g_input_lifecycle_active     = false;
 bool                         g_controller_captured        = false;
 OverlaySession               g_session;
 SDL_Window*                  g_input_window               = nullptr;
+struct TrophyNotification {
+	std::string            name;
+	int32_t                grade;
+	std::vector<std::byte> icon_png;
+};
+std::mutex                     g_trophy_notification_mutex;
+std::deque<TrophyNotification> g_trophy_notifications;
+std::atomic<bool>              g_trophy_notification_active {false};
+
+std::vector<std::string_view> WrapTrophyTitle(std::string_view text, float font_size, float width) {
+	std::vector<std::string_view> lines;
+	auto* const                  font = ImGui::GetFont();
+	for (;;) {
+		const auto newline   = text.find('\n');
+		const auto paragraph = text.substr(0, newline);
+		auto       line      = paragraph.substr(0, 0);
+		for (auto start = paragraph.find_first_not_of(" \t\r"); start != std::string_view::npos;) {
+			const auto end  = std::min(paragraph.find_first_of(" \t\r", start), paragraph.size());
+			const auto word = paragraph.substr(start, end - start);
+			const std::string_view candidate(line.empty() ? word.begin() : line.begin(),
+			                                 word.end());
+			// Keep an oversized word intact; the toast's clip rectangle handles overflow.
+			if (!line.empty() && font->CalcTextSizeA(font_size, std::numeric_limits<float>::max(),
+			                                       0.0f, candidate.data(),
+			                                       candidate.data() + candidate.size()).x > width) {
+				lines.push_back(line);
+				line = word;
+			} else {
+				line = candidate;
+			}
+			start = paragraph.find_first_not_of(" \t\r", end);
+		}
+		lines.push_back(line);
+		if (newline == std::string_view::npos) {
+			return lines;
+		}
+		text.remove_prefix(newline + 1);
+	}
+}
 
 void ClearInputEvents() {
 	std::scoped_lock lock(g_input_mutex);
@@ -234,7 +281,7 @@ void RefreshVisibility() {
 	bool capture_keyboard   = false;
 	bool text_input         = false;
 	bool multiline          = false;
-	if (snapshot.session.kind == OverlayKind::Error) {
+	if (snapshot.session.kind == OverlayKind::Dialog) {
 		capture_controller = true;
 		capture_keyboard   = true;
 	} else if (snapshot.session.kind == OverlayKind::Ime) {
@@ -429,14 +476,14 @@ void InitializeSystemOverlayInput(SDL_Window* window) {
 	}
 	CoreIme::SetVisibilityCallback(OnCoreVisibilityChanged);
 	DialogIme::SetVisibilityCallback(OnDialogVisibilityChanged);
-	ErrorDialog::SetVisibilityCallback(RefreshVisibility);
+	SystemDialog::SetVisibilityCallback(RefreshVisibility);
 	RefreshVisibility();
 }
 
 void ShutdownSystemOverlayInput() {
 	CoreIme::SetVisibilityCallback(nullptr);
 	DialogIme::SetVisibilityCallback(nullptr);
-	ErrorDialog::SetVisibilityCallback(nullptr);
+	SystemDialog::SetVisibilityCallback(nullptr);
 	{
 		std::scoped_lock lock(g_visibility_mutex);
 		g_input_lifecycle_active = false;
@@ -466,9 +513,10 @@ void ShutdownSystemOverlayInput() {
 SystemOverlayVisualState GetSystemOverlayVisualState() noexcept {
 	const auto core   = CoreIme::GetVisualState();
 	const auto dialog = DialogIme::GetVisualState();
-	const auto error  = ErrorDialog::GetVisualState();
-	return {core.active || dialog.active || error.active,
-	        core.revision + dialog.revision + error.revision};
+	const auto system = SystemDialog::GetVisualState();
+	const bool trophy_active = g_trophy_notification_active.load(std::memory_order_acquire);
+	return {core.active || dialog.active || system.active || trophy_active,
+	        core.revision + dialog.revision + system.revision + (trophy_active ? 1u : 0u)};
 }
 
 bool ProcessSystemOverlayInput(const SDL_Event& event) {
@@ -521,10 +569,11 @@ bool ProcessSystemOverlayInput(const SDL_Event& event) {
 	if (keyboard_event && !g_input_keyboard) {
 		return false;
 	}
-	if (keyboard_event && session.kind == OverlayKind::Error) {
+	if (keyboard_event && session.kind == OverlayKind::Dialog) {
 		if (event.type == SDL_EVENT_KEY_DOWN && !event.key.repeat &&
-		    (event.key.key == SDLK_RETURN || event.key.key == SDLK_KP_ENTER)) {
-			ErrorDialog::HostAccept(generation);
+		    (event.key.key == SDLK_RETURN || event.key.key == SDLK_KP_ENTER ||
+		     event.key.key == SDLK_ESCAPE)) {
+			SystemDialog::HostClose(generation);
 		}
 		return true;
 	}
@@ -598,12 +647,34 @@ bool ProcessSystemOverlayInput(const SDL_Event& event) {
 	}
 }
 
+void NotifyTrophyUnlocked(std::string_view name, int32_t grade,
+                          std::span<const std::byte> icon_png) {
+	std::scoped_lock lock(g_trophy_notification_mutex);
+	g_trophy_notifications.push_back(
+	    {std::string(name), grade, std::vector<std::byte>(icon_png.begin(), icon_png.end())});
+	g_trophy_notification_active.store(true, std::memory_order_release);
+}
+
 struct SystemOverlay::Impl {
+	using Clock = std::chrono::steady_clock;
+
 	explicit Impl(GraphicContext& context): graphics(context) {}
 
 	~Impl() {
+		SDL_DestroyAudioStream(trophy_sound_stream);
+		SDL_free(trophy_sound_data);
+		if (trophy_sound_initialized) {
+			SDL_QuitSubSystem(SDL_INIT_AUDIO);
+		}
 		ReleaseVulkan();
 		if (imgui_context != nullptr) {
+			ImGui::SetCurrentContext(imgui_context);
+			if (trophy_notification_image) {
+				ImGui::UnregisterUserTexture(trophy_notification_image.get());
+			}
+			for (const auto& texture: retired_trophy_images) {
+				ImGui::UnregisterUserTexture(texture.get());
+			}
 			ImGui::DestroyContext(imgui_context);
 		}
 	}
@@ -630,6 +701,75 @@ struct SystemOverlay::Impl {
 		style.ItemSpacing    = {8.0f, 8.0f};
 	}
 
+	void PlayTrophySound() {
+		if (!trophy_sound_attempted) {
+			trophy_sound_attempted   = true;
+			trophy_sound_initialized = SDL_InitSubSystem(SDL_INIT_AUDIO);
+			if (!trophy_sound_initialized) {
+				LOGF("[Trophy] could not initialize audio: %s\n", SDL_GetError());
+				return;
+			}
+			const char* base_path = SDL_GetBasePath();
+			if (base_path == nullptr) {
+				LOGF("[Trophy] could not locate sound: %s\n", SDL_GetError());
+				return;
+			}
+			const std::string path = std::string(base_path) + "assets/sounds/trophy-unlock.wav";
+			SDL_AudioSpec     spec {};
+			if (!SDL_LoadWAV(path.c_str(), &spec, &trophy_sound_data, &trophy_sound_size)) {
+				LOGF("[Trophy] could not load %s: %s\n", path.c_str(), SDL_GetError());
+				return;
+			}
+			if (trophy_sound_size > static_cast<uint32_t>(std::numeric_limits<int>::max())) {
+				LOGF("[Trophy] sound is too large: %s\n", path.c_str());
+				return;
+			}
+			trophy_sound_stream = SDL_OpenAudioDeviceStream(SDL_AUDIO_DEVICE_DEFAULT_PLAYBACK,
+			                                                &spec, nullptr, nullptr);
+			if (trophy_sound_stream == nullptr) {
+				LOGF("[Trophy] could not open audio stream: %s\n", SDL_GetError());
+				return;
+			}
+		}
+		if (trophy_sound_stream != nullptr &&
+		    (!SDL_PutAudioStreamData(trophy_sound_stream, trophy_sound_data,
+		                             static_cast<int>(trophy_sound_size)) ||
+		     !SDL_ResumeAudioStreamDevice(trophy_sound_stream))) {
+			LOGF("[Trophy] could not play sound: %s\n", SDL_GetError());
+		}
+	}
+
+	std::unique_ptr<ImTextureData> LoadTrophyIcon(std::span<const std::byte> encoded) {
+		if (encoded.empty() ||
+		    encoded.size() > static_cast<std::size_t>(std::numeric_limits<int>::max())) {
+			return nullptr;
+		}
+
+		const auto* data   = reinterpret_cast<const stbi_uc*>(encoded.data());
+		const int   size   = static_cast<int>(encoded.size());
+		int         width  = 0;
+		int         height = 0;
+		if (!stbi_info_from_memory(data, size, &width, &height, nullptr) || width <= 0 ||
+		    height <= 0 || width > 512 || height > 512) {
+			LOGF("[Trophy] invalid icon or dimensions exceed 512x512\n");
+			return nullptr;
+		}
+		auto* pixels = stbi_load_from_memory(data, size, &width, &height, nullptr, 4);
+		if (pixels == nullptr) {
+			LOGF("[Trophy] could not decode trophy icon: %s\n", stbi_failure_reason());
+			return nullptr;
+		}
+
+		auto texture = std::make_unique<ImTextureData>();
+		texture->Create(ImTextureFormat_RGBA32, width, height);
+		std::memcpy(texture->Pixels, pixels,
+		            static_cast<std::size_t>(width) * static_cast<std::size_t>(height) * 4);
+		texture->UseColors = true;
+		stbi_image_free(pixels);
+		ImGui::RegisterUserTexture(texture.get());
+		return texture;
+	}
+
 	void EnsureVulkan(vk::Format format, uint32_t image_count) {
 		EnsureContext();
 		if (vulkan_initialized) {
@@ -904,7 +1044,7 @@ struct SystemOverlay::Impl {
 		ImGui::End();
 	}
 
-	void DrawError(const ErrorDialog::HostSnapshot& snapshot, vk::Extent2D extent) {
+	void DrawDialog(const SystemDialog::HostSnapshot& snapshot, vk::Extent2D extent) {
 		const ImVec2 display(static_cast<float>(extent.width), static_cast<float>(extent.height));
 		const float  scale = std::max(std::min(display.x / 1280.0f, display.y / 720.0f), 0.5f);
 		ImGui::GetBackgroundDrawList()->AddRectFilled({0.0f, 0.0f}, display,
@@ -922,17 +1062,23 @@ struct SystemOverlay::Impl {
 		constexpr ImGuiWindowFlags flags = ImGuiWindowFlags_NoDecoration | ImGuiWindowFlags_NoMove |
 		                                   ImGuiWindowFlags_NoSavedSettings |
 		                                   ImGuiWindowFlags_AlwaysAutoResize;
-		ImGui::Begin("##SystemError", nullptr, flags);
-		ImGui::TextUnformatted("Error");
+		const bool                 signin = snapshot.kind == SystemDialog::Kind::Signin;
+		ImGui::Begin("##SystemDialog", nullptr, flags);
+		ImGui::TextUnformatted(signin ? "Sign in to PlayStation Network" : "Error");
 		ImGui::Separator();
 		const auto error_code = static_cast<uint32_t>(snapshot.error_code);
-		ImGui::TextWrapped("%s", error_code == 0x80550006u
+		ImGui::TextWrapped("%s", signin ? "PlayStation Network sign-in is unavailable."
+		                         : error_code == 0x80550006u
 		                             ? "You are not signed in to PlayStation Network."
 		                             : "An error has occurred.");
-		ImGui::TextDisabled("Error code: 0x%08X", error_code);
+		if (!signin) {
+			ImGui::TextDisabled("Error code: 0x%08X", error_code);
+		}
 		const float button_width = std::min(140.0f * scale, ImGui::GetContentRegionAvail().x);
 		ImGui::SetCursorPosX((ImGui::GetWindowSize().x - button_width) * 0.5f);
-		const bool accepted = ImGui::Button("OK", {button_width, 44.0f * scale});
+		const bool accepted =
+		    ImGui::Button(signin ? "Cancel" : "OK", {button_width, 44.0f * scale}) ||
+		    (signin && ImGui::IsKeyPressed(ImGuiKey_GamepadFaceRight, false));
 		if (focus_pending) {
 			ImGui::SetItemDefaultFocus();
 			focus_pending = false;
@@ -941,13 +1087,135 @@ struct SystemOverlay::Impl {
 		ImGui::PopFont();
 		ImGui::PopStyleVar(2);
 		if (accepted) {
-			ErrorDialog::HostAccept(snapshot.generation);
+			SystemDialog::HostClose(snapshot.generation);
 		}
 	}
 
+	void DrawTrophyNotification(vk::Extent2D frame_extent, Clock::time_point now) {
+		std::erase_if(retired_trophy_images, [](const auto& texture) {
+			if (texture->Status != ImTextureStatus_Destroyed) {
+				return false;
+			}
+			ImGui::UnregisterUserTexture(texture.get());
+			return true;
+		});
+		constexpr float TOAST_LIFETIME = 5.0f;
+		if (trophy_notification &&
+		    std::chrono::duration<float>(now - trophy_notification_born).count() >=
+		        TOAST_LIFETIME) {
+			trophy_notification.reset();
+			if (trophy_notification_image) {
+				// Keep the texture registered until the backend retires in-flight frames.
+				trophy_notification_image->WantDestroyNextFrame = true;
+				retired_trophy_images.push_back(std::move(trophy_notification_image));
+			}
+		}
+		if (!trophy_notification) {
+			{
+				std::scoped_lock lock(g_trophy_notification_mutex);
+				if (g_trophy_notifications.empty()) {
+					g_trophy_notification_active.store(!retired_trophy_images.empty(),
+					                                   std::memory_order_release);
+					return;
+				}
+				trophy_notification = std::move(g_trophy_notifications.front());
+				g_trophy_notifications.pop_front();
+			}
+			trophy_notification_born  = now;
+			trophy_notification_image = LoadTrophyIcon(trophy_notification->icon_png);
+			PlayTrophySound();
+		}
+		const float age = std::chrono::duration<float>(now - trophy_notification_born).count();
+		const float screen_scale =
+		    std::clamp(std::min(static_cast<float>(frame_extent.width) / 1920.0f,
+		                        static_cast<float>(frame_extent.height) / 1080.0f),
+		               0.65f, 1.25f);
+		const float width = std::max(
+		    0.0f, std::min(520.0f * screen_scale, static_cast<float>(frame_extent.width) - 32.0f));
+		constexpr std::array GRADE_COLORS {
+		    IM_COL32(161, 225, 245, 255), IM_COL32(255, 207, 82, 255), IM_COL32(211, 220, 232, 255),
+		    IM_COL32(211, 142, 91, 255)};
+		const int   grade        = trophy_notification->grade;
+		const bool  has_grade    = grade > 0 && grade <= static_cast<int>(GRADE_COLORS.size());
+		const float pad          = 10.0f * screen_scale;
+		const float base_height  = 104.0f * screen_scale;
+		const float icon_size    = base_height - 2.0f * pad;
+		const float text_offset  = trophy_notification_image != nullptr
+		                               ? pad + icon_size + 18.0f * screen_scale
+		                               : 20.0f * screen_scale;
+		const float title_offset = text_offset + (has_grade ? 30.0f * screen_scale : 0.0f);
+		const float title_size   = 26.0f * screen_scale;
+		const float line_height  = 30.0f * screen_scale;
+		const float text_padding = 16.0f * screen_scale;
+		const auto  title_lines  = WrapTrophyTitle(trophy_notification->name, title_size,
+		                                           width - title_offset - text_padding);
+		const float block_height =
+		    static_cast<float>(title_lines.size()) * line_height + 28.0f * screen_scale;
+		const float  height   = std::max(base_height, block_height + 2.0f * text_padding);
+		const float  rounding = 14.0f * screen_scale;
+		const float  fade_in  = std::clamp(age / 0.22f, 0.0f, 1.0f);
+		const float  fade_out = std::clamp((TOAST_LIFETIME - age) / 0.4f, 0.0f, 1.0f);
+		const float  alpha    = std::min(fade_in, fade_out);
+		const float  slide    = (1.0f - fade_in) * (width + 24.0f);
+		const ImVec2 top_left {static_cast<float>(frame_extent.width) - 32.0f - width + slide,
+		                       32.0f};
+		const ImVec2 bottom_right {top_left.x + width, top_left.y + height};
+		auto*        draw = ImGui::GetForegroundDrawList();
+		const auto   fade = [alpha](int r, int g, int b, int a) {
+            return IM_COL32(r, g, b, static_cast<int>(static_cast<float>(a) * alpha));
+		};
+		draw->AddRectFilled({top_left.x + 3.0f, top_left.y + 5.0f},
+		                    {bottom_right.x + 3.0f, bottom_right.y + 5.0f}, fade(0, 0, 0, 90),
+		                    rounding);
+		draw->AddRectFilled(top_left, bottom_right, fade(24, 26, 32, 235), rounding);
+		draw->AddRect(top_left, bottom_right, fade(255, 255, 255, 28), rounding,
+		              ImDrawFlags_RoundCornersAll, 1.0f);
+
+		const float text_left = top_left.x + text_offset;
+		if (trophy_notification_image != nullptr) {
+			const ImVec2 icon_min {top_left.x + pad, top_left.y + (height - icon_size) * 0.5f};
+			const ImVec2 icon_max {icon_min.x + icon_size, icon_min.y + icon_size};
+			draw->AddImageRounded(trophy_notification_image->GetTexRef(), icon_min, icon_max,
+			                      {0, 0}, {1, 1}, fade(255, 255, 255, 255), 8.0f * screen_scale);
+			draw->AddRect(icon_min, icon_max, fade(255, 255, 255, 40), 8.0f * screen_scale,
+			              ImDrawFlags_RoundCornersAll, 1.0f);
+		}
+
+		const float title_y = top_left.y + (height - block_height) * 0.5f;
+		const float title_x = top_left.x + title_offset;
+		if (has_grade) {
+			// Small cup glyph tinted with the trophy grade.
+			const ImU32 cup = ImGui::GetColorU32(GRADE_COLORS[grade - 1], alpha);
+			const float u   = screen_scale;
+			const float cx  = text_left + 9.0f * u;
+			const float cy  = title_y + 15.0f * u;
+			draw->AddRectFilled({cx - 8.0f * u, cy - 11.0f * u}, {cx + 8.0f * u, cy + 2.0f * u},
+			                    cup, 2.0f * u, ImDrawFlags_RoundCornersBottom);
+			draw->AddRect({cx - 11.0f * u, cy - 9.0f * u}, {cx - 6.0f * u, cy - 3.0f * u}, cup,
+			              2.0f * u, ImDrawFlags_RoundCornersLeft, 1.8f * u);
+			draw->AddRect({cx + 6.0f * u, cy - 9.0f * u}, {cx + 11.0f * u, cy - 3.0f * u}, cup,
+			              2.0f * u, ImDrawFlags_RoundCornersRight, 1.8f * u);
+			draw->AddRectFilled({cx - 1.5f * u, cy + 2.0f * u}, {cx + 1.5f * u, cy + 7.0f * u}, cup);
+			draw->AddRectFilled({cx - 5.0f * u, cy + 7.0f * u}, {cx + 5.0f * u, cy + 10.0f * u}, cup,
+			                    1.5f * u);
+		}
+		draw->PushClipRect(top_left, {bottom_right.x - text_padding, bottom_right.y}, true);
+		float line_y = title_y;
+		for (const auto line : title_lines) {
+			draw->AddText(nullptr, title_size, {title_x, line_y}, fade(255, 255, 255, 255),
+			              line.data(), line.data() + line.size());
+			line_y += line_height;
+		}
+		draw->AddText(nullptr, 20.0f * screen_scale, {text_left, line_y + 8.0f * screen_scale},
+		              fade(176, 180, 192, 255), "Trophy earned!");
+		draw->PopClipRect();
+	}
+
 	bool PrepareFrame(vk::Extent2D frame_extent, vk::Format format, uint32_t image_count) {
 		OverlaySnapshot snapshot;
-		if (!GetOverlaySnapshot(&snapshot)) {
+		const bool      has_overlay = GetOverlaySnapshot(&snapshot);
+		const auto now = Clock::now();
+		if (!has_overlay && !g_trophy_notification_active.load(std::memory_order_acquire)) {
 			return false;
 		}
 		const auto prepared_session = snapshot.session;
@@ -955,7 +1223,8 @@ struct SystemOverlay::Impl {
 		if (session != snapshot.session) {
 			session       = snapshot.session;
 			focus_pending = true;
-			shift         = (snapshot.ime.option & Ime::OPTION_NO_AUTO_CAPITALIZE) == 0;
+			shift         = session.kind == OverlayKind::Ime &&
+			                (snapshot.ime.option & Ime::OPTION_NO_AUTO_CAPITALIZE) == 0;
 			symbol_mode   = false;
 			panel_offset  = {};
 			right_stick   = {};
@@ -969,7 +1238,6 @@ struct SystemOverlay::Impl {
 		auto& io       = ImGui::GetIO();
 		io.DisplaySize = {static_cast<float>(frame_extent.width),
 		                  static_cast<float>(frame_extent.height)};
-		const auto now = std::chrono::steady_clock::now();
 		io.DeltaTime   = last_frame == std::chrono::steady_clock::time_point {}
 		                     ? 1.0f / 60.0f
 		                     : std::clamp(std::chrono::duration<float>(now - last_frame).count(),
@@ -977,15 +1245,19 @@ struct SystemOverlay::Impl {
 		last_frame     = now;
 		ImGui_ImplVulkan_NewFrame();
 		ImGui::NewFrame();
-		if (!GetOverlaySnapshot(&snapshot) || snapshot.session != prepared_session) {
+		GetOverlaySnapshot(&snapshot);
+		if (snapshot.session != prepared_session) {
 			ImGui::EndFrame();
 			return false;
 		}
-		if (snapshot.session.kind == OverlayKind::Error) {
-			DrawError(snapshot.error, frame_extent);
-		} else {
-			DrawIme(snapshot.ime, frame_extent);
+		if (has_overlay) {
+			if (snapshot.session.kind == OverlayKind::Dialog) {
+				DrawDialog(snapshot.dialog, frame_extent);
+			} else {
+				DrawIme(snapshot.ime, frame_extent);
+			}
 		}
+		DrawTrophyNotification(frame_extent, now);
 		ImGui::Render();
 		extent = frame_extent;
 		return true;
@@ -1032,9 +1304,18 @@ struct SystemOverlay::Impl {
 	float                                 button_height      = 42.0f;
 	ImVec2                                panel_offset {};
 	ImVec2                                right_stick {};
-	OverlaySession                        session;
-	vk::Extent2D                          extent {};
-	std::chrono::steady_clock::time_point last_frame;
+	OverlaySession                              session;
+	std::unique_ptr<ImTextureData>              trophy_notification_image;
+	std::optional<TrophyNotification>           trophy_notification;
+	Clock::time_point                           trophy_notification_born;
+	std::vector<std::unique_ptr<ImTextureData>> retired_trophy_images;
+	SDL_AudioStream*                            trophy_sound_stream      = nullptr;
+	uint8_t*                                    trophy_sound_data        = nullptr;
+	uint32_t                                    trophy_sound_size        = 0;
+	bool                                        trophy_sound_attempted   = false;
+	bool                                        trophy_sound_initialized = false;
+	vk::Extent2D                                extent {};
+	std::chrono::steady_clock::time_point       last_frame;
 };
 
 SystemOverlay::SystemOverlay(GraphicContext& graphics): m_impl(std::make_unique<Impl>(graphics)) {}

src/graphics/presentation/systemOverlay.h

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

src/graphics/presentation/videoOut.cpp

@@ -1814,6 +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

@@ -90,6 +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/hostInput.cpp

@@ -25,6 +25,8 @@ struct ControlInfo {
 	Controller::Axis axis     = Controller::Axis::AxisMax;
 	bool             positive = false;
 	float            touch_x  = 0.0f;
+
+	Controller::Setting setting {};
 };
 
 static constexpr std::array CONTROL_INFO = {
@@ -55,6 +57,10 @@ static constexpr std::array CONTROL_INFO = {
     ControlInfo {"RightStickRight", 0, Controller::Axis::RightX, true},
     ControlInfo {"RightStickUp", 0, Controller::Axis::RightY, false},
     ControlInfo {"RightStickDown", 0, Controller::Axis::RightY, true},
+    ControlInfo {.name = "SpeakerVolume", .setting = Controller::Setting::SpeakerVolume},
+    ControlInfo {.name = "VibrationIntensity", .setting = Controller::Setting::VibrationIntensity},
+    ControlInfo {.name    = "TriggerEffectIntensity",
+                 .setting = Controller::Setting::TriggerEffectIntensity},
 };
 
 constexpr std::size_t INVALID_CONTROL = CONTROL_INFO.size();
@@ -66,6 +72,7 @@ struct Binding {
 };
 
 constexpr int              MOUSE_POLL_INTERVAL_MS = 33;
+constexpr uint64_t         CURSOR_IDLE_HIDE_MS    = 2000;
 constexpr std::string_view MOUSE_SENSITIVITY      = "MouseSensitivity=";
 
 struct MouseJoystickState {
@@ -75,7 +82,8 @@ struct MouseJoystickState {
 };
 
 MouseJoystickState g_mouse;
-SDL_Window*        g_mouse_window = nullptr;
+SDL_Window*        g_mouse_window   = nullptr;
+uint64_t           g_cursor_hide_at = 0;
 
 std::size_t ControlFromName(std::string_view name) {
 	const auto info = std::find_if(CONTROL_INFO.begin(), CONTROL_INFO.end(),
@@ -242,6 +250,12 @@ void SetStickAxis(Controller::Axis axis, bool negative, bool positive) {
 	Controller::SetAxis(Controller::HOST_INPUT_CONTROLLER_ID, axis, value);
 }
 
+void CycleSettingOnPress(Controller::Setting setting, bool down) {
+	if (down) {
+		Controller::CycleSetting(setting);
+	}
+}
+
 void SetControl(std::size_t control, bool down) {
 	if (control == INVALID_CONTROL) {
 		return;
@@ -256,6 +270,10 @@ void SetControl(std::size_t control, bool down) {
 		SetButton(info.button, down);
 		return;
 	}
+	if (info.axis == Controller::Axis::AxisMax) {
+		CycleSettingOnPress(info.setting, down);
+		return;
+	}
 
 	struct AxisKeys {
 		bool negative = false;
@@ -280,6 +298,9 @@ void DefaultKeyboardInput(int key_code, bool down) {
 	switch (NormalizeKey(static_cast<SDL_Keycode>(key_code))) {
 		case SDLK_BACKSPACE: SetTouchPad(0.25f, down); return;
 		case SDLK_TAB: SetTouchPad(0.75f, down); return;
+		case SDLK_1: CycleSettingOnPress(Controller::Setting::SpeakerVolume, down); return;
+		case SDLK_2: CycleSettingOnPress(Controller::Setting::VibrationIntensity, down); return;
+		case SDLK_3: CycleSettingOnPress(Controller::Setting::TriggerEffectIntensity, down); return;
 		case SDLK_A:
 			left.left = down;
 			SetStickAxis(Controller::Axis::LeftX, left.left, left.right);
@@ -361,11 +382,31 @@ int PollMouse(uint64_t now_ms) {
 	return MOUSE_POLL_INTERVAL_MS;
 }
 
+bool IsCursorActivity(const SDL_Event& event) {
+	SDL_MouseID which;
+	switch (event.type) {
+		case SDL_EVENT_MOUSE_MOTION:
+			if (event.motion.xrel == 0.0f && event.motion.yrel == 0.0f) {
+				return false;
+			}
+			which = event.motion.which;
+			break;
+		case SDL_EVENT_MOUSE_BUTTON_DOWN:
+		case SDL_EVENT_MOUSE_BUTTON_UP: which = event.button.which; break;
+		case SDL_EVENT_MOUSE_WHEEL: which = event.wheel.which; break;
+		default: return false;
+	}
+	return which != SDL_TOUCH_MOUSEID && which != SDL_PEN_MOUSEID;
+}
+
 } // namespace
 
 void HostInputInit(SDL_Window* window) {
 	GetInputMap();
 	g_mouse_window = window;
+	if (Config::HideCursorEnabled()) {
+		g_cursor_hide_at = SDL_GetTicks() + CURSOR_IDLE_HIDE_MS;
+	}
 }
 
 void HostInputShutdown() {
@@ -374,7 +415,11 @@ void HostInputShutdown() {
 		CenterMouseStick();
 		g_mouse = {};
 	}
-	g_mouse_window = nullptr;
+	g_mouse_window   = nullptr;
+	g_cursor_hide_at = 0;
+	if (Config::HideCursorEnabled()) {
+		SDL_ShowCursor();
+	}
 }
 
 void HostInputKey(int key_code, bool down) {
@@ -409,21 +454,38 @@ void HostInputToggleMouseToJoystick() {
 	LOGF("Mouse to right stick: enabled (F7 to release)\n");
 }
 
-bool HostInputWaitEvent(SDL_Event* event) {
-	bool has_event;
+bool HostInputWaitEvent(SDL_Event* event, int max_wait_ms) {
+	int timeout = -1;
 	if (!g_mouse.enabled || SDL_GetKeyboardFocus() != g_mouse_window) {
 		g_mouse.next_poll = 0;
 		CenterMouseStick();
-		has_event = SDL_WaitEvent(event);
-		if (!has_event) {
-			EXIT("%s\n", SDL_GetError());
-		}
 	} else {
 		if (g_mouse.next_poll == 0) {
 			SDL_GetRelativeMouseState(nullptr, nullptr);
 			g_mouse.next_poll = SDL_GetTicks() + MOUSE_POLL_INTERVAL_MS;
 		}
-		has_event = SDL_WaitEventTimeout(event, PollMouse(SDL_GetTicks()));
+		timeout = PollMouse(SDL_GetTicks());
+	}
+
+	if (g_cursor_hide_at != 0) {
+		const auto now_ms = SDL_GetTicks();
+		const int  cursor_timeout =
+		    now_ms < g_cursor_hide_at ? static_cast<int>(g_cursor_hide_at - now_ms) : 0;
+		timeout = timeout < 0 ? cursor_timeout : std::min(timeout, cursor_timeout);
+	}
+	timeout              = timeout < 0 ? max_wait_ms : std::min(timeout, max_wait_ms);
+	const bool has_event = SDL_WaitEventTimeout(event, timeout);
+
+	if (Config::HideCursorEnabled()) {
+		const auto now_ms = SDL_GetTicks();
+		if (has_event && !g_mouse.enabled && IsCursorActivity(*event) &&
+		    SDL_GetWindowFromEvent(event) == g_mouse_window) {
+			SDL_ShowCursor();
+			g_cursor_hide_at = now_ms + CURSOR_IDLE_HIDE_MS;
+		} else if (g_cursor_hide_at != 0 && now_ms >= g_cursor_hide_at) {
+			SDL_HideCursor();
+			g_cursor_hide_at = 0;
+		}
 	}
 
 	if (has_event && event->type == SDL_EVENT_WINDOW_FOCUS_LOST &&

src/graphics/presentation/window/hostInput.h

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

src/graphics/presentation/window/swapchain.cpp

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

src/graphics/presentation/window/vulkanWindow.cpp

@@ -47,6 +47,12 @@ struct VulkanExtensions {
 	std::vector<vk::LayerProperties>     available_layers;
 };
 
+vk::PhysicalDeviceVulkan11Features WindowContext::RequiredVulkan11Features() noexcept {
+	vk::PhysicalDeviceVulkan11Features features {};
+	features.storageBuffer16BitAccess = VK_TRUE;
+	return features;
+}
+
 vk::PhysicalDeviceVulkan12Features WindowContext::RequiredVulkan12Features() noexcept {
 	vk::PhysicalDeviceVulkan12Features features {};
 	features.samplerMirrorClampToEdge  = VK_TRUE;
@@ -55,6 +61,8 @@ vk::PhysicalDeviceVulkan12Features WindowContext::RequiredVulkan12Features() noe
 	features.shaderOutputViewportIndex = VK_TRUE;
 	features.bufferDeviceAddress       = VK_TRUE;
 	features.shaderBufferInt64Atomics  = VK_TRUE;
+	features.storageBuffer8BitAccess   = VK_TRUE;
+	features.shaderSampledImageArrayNonUniformIndexing = VK_TRUE;
 	return features;
 }
 
@@ -137,11 +145,13 @@ 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 +180,22 @@ 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;
 		}
 
@@ -196,6 +214,7 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 		depth_clip_control.pNext = &depth_clip_enable;
 
 		vk::PhysicalDeviceVulkan12Features features12 {};
+		vk::PhysicalDeviceVulkan11Features features11 {};
 #if defined(__APPLE__)
 		features12.pNext = &depth_clip_control;
 #else
@@ -203,7 +222,9 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 		fragment_barycentric.pNext = &depth_clip_control;
 		features12.pNext           = &fragment_barycentric;
 #endif
-		features13.pNext = &features12;
+		features11.pNext       = features12.pNext;
+		features12.pNext       = &features11;
+		features13.pNext       = &features12;
 		device_features2.pNext = &features13;
 
 		device.getFeatures2(&device_features2);
@@ -212,151 +233,89 @@ 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(features11.storageBuffer16BitAccess, "storageBuffer16BitAccess",
+		              WindowContext::RequiredVulkan11Features().storageBuffer16BitAccess);
+		check_feature(features12.storageBuffer8BitAccess, "storageBuffer8BitAccess",
+		              required_features12.storageBuffer8BitAccess);
+		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(features12.shaderSampledImageArrayNonUniformIndexing,
+		              "shaderSampledImageArrayNonUniformIndexing",
+		              required_features12.shaderSampledImageArrayNonUniformIndexing);
+		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 +327,7 @@ 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 +345,46 @@ 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 +438,20 @@ 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;
 	}
 
@@ -556,7 +478,34 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 		provoking_vertex.pNext = supported_features2.pNext;
 		supported_features2.pNext = &provoking_vertex;
 	}
+	const bool image_atomic_int64_extension =
+	    HasExtension(device_extensions, VK_EXT_SHADER_IMAGE_ATOMIC_INT64_EXTENSION_NAME);
+	vk::PhysicalDeviceShaderImageAtomicInt64FeaturesEXT image_atomic_int64 {};
+	if (image_atomic_int64_extension) {
+		image_atomic_int64.pNext = supported_features2.pNext;
+		supported_features2.pNext = &image_atomic_int64;
+	}
 	physical_device.getFeatures2(&supported_features2);
+	graphics.shader_image_int64_atomics_enabled = image_atomic_int64.shaderImageInt64Atomics;
+
+	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);
+	auto features11 = WindowContext::RequiredVulkan11Features();
+	features11.pNext = features12.pNext;
+	features12.pNext = &features11;
+	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 {};
@@ -632,8 +581,7 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 	device_features.shaderInt64 = VK_TRUE;
 	device_features.shaderFloat64 =
 	    supported_features2.features.shaderFloat64 &&
-	    float_controls.shaderSignedZeroInfNanPreserveFloat64 &&
-	    float_controls.shaderRoundingModeRTEFloat32;
+	    float_controls.shaderSignedZeroInfNanPreserveFloat64;
 	// if (device_features.shaderFloat64 && !float_controls.shaderDenormPreserveFloat64) {
 	// 	Log::WriteToConsoleAndLog(
 	// 	    "WARNING: Vulkan device does not guarantee FP64 denormal preservation; "
@@ -685,6 +633,11 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 		provoking_vertex.transformFeedbackPreservesProvokingVertex = VK_FALSE;
 		create_info.pNext = &provoking_vertex;
 	}
+	if (graphics.shader_image_int64_atomics_enabled) {
+		image_atomic_int64.pNext = const_cast<void*>(create_info.pNext);
+		image_atomic_int64.sparseImageInt64Atomics = VK_FALSE;
+		create_info.pNext = &image_atomic_int64;
+	}
 	create_info.pQueueCreateInfos       = &queue_create_info;
 	create_info.queueCreateInfoCount    = 1;
 	create_info.enabledExtensionCount   = static_cast<uint32_t>(device_extensions.size());
@@ -1004,11 +957,13 @@ 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 {};
@@ -1047,8 +1002,10 @@ void WindowContext::CreateVulkan() {
 			graphic_ctx.memory_budget_ext_enabled = true;
 		}
 		for (const auto* extension: {VK_EXT_ROBUSTNESS_2_EXTENSION_NAME,
+		                             VK_EXT_SHADER_IMAGE_ATOMIC_INT64_EXTENSION_NAME,
 		                             VK_EXT_PROVOKING_VERTEX_EXTENSION_NAME,
 		                             VK_EXT_MESH_SHADER_EXTENSION_NAME,
+		                             VK_KHR_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_EXTENSION_NAME,
 		                             VK_EXT_DEPTH_RANGE_UNRESTRICTED_EXTENSION_NAME}) {
 			if (HasExtension(available_extensions, extension)) {
 				device_extensions.push_back(extension);

src/graphics/presentation/window/window.cpp

@@ -306,14 +306,20 @@ static void GameEventController([[maybe_unused]] const EventController& f) {
 
 	if (f.added) {
 		auto* pad = SDL_OpenGamepad(f.id);
-		EXIT_NOT_IMPLEMENTED(pad == nullptr);
+		if (pad == nullptr) {
+			LOGF("Controller: ignoring gamepad %d that could not be opened: %s\n", f.id,
+			     SDL_GetError());
+			return;
+		}
 		int id = SDL_GetJoystickID(SDL_GetGamepadJoystick(pad));
 		Controller::Connect(id);
 	}
 
 	if (f.removed) {
-		Controller::Disconnect(f.id);
-		SDL_CloseGamepad(SDL_GetGamepadFromID(f.id));
+		if (auto* pad = SDL_GetGamepadFromID(f.id); pad != nullptr) {
+			Controller::Disconnect(f.id);
+			SDL_CloseGamepad(pad);
+		}
 	}
 
 	if (f.down || f.up) {
@@ -356,11 +362,13 @@ static void GameEventDidEnterForeground(WindowLoopState& game) {
 	SetPause(game, false);
 }
 
-void WindowContext::Resize(uint32_t new_width, uint32_t new_height) {
-	EXIT_IF(new_width == 0 || new_height == 0);
+void WindowContext::Resize(int new_width, int new_height) {
+	if (new_width <= 0 || new_height <= 0) {
+		return;
+	}
 	Common::LockGuard lock(mutex);
-	graphic_ctx.screen_width  = new_width;
-	graphic_ctx.screen_height = new_height;
+	graphic_ctx.screen_width  = static_cast<uint32_t>(new_width);
+	graphic_ctx.screen_height = static_cast<uint32_t>(new_height);
 }
 
 void WindowContext::ProcessWindowEvent(const SDL_WindowEvent& event) {
@@ -714,6 +722,10 @@ void WindowContext::Run() {
 	loop.need_exit = false;
 	loop.paused.store(false, std::memory_order_release);
 
+	constexpr uint64_t title_interval_ms = 1000;
+	auto               title_time        = SDL_GetTicks();
+	auto               title_frames      = loop.presented_frames.load(std::memory_order_relaxed);
+
 	while (!loop.need_exit) {
 		if (loop.paused.load(std::memory_order_acquire)) {
 			if (!timer.IsPaused()) {
@@ -723,7 +735,20 @@ void WindowContext::Run() {
 			timer.Resume();
 		}
 
-		if (!HostInputWaitEvent(&loop.event)) {
+		// Refresh the title on the main thread without making presentation wait for it.
+		const auto now     = SDL_GetTicks();
+		const auto elapsed = now - title_time;
+		if (elapsed >= title_interval_ms) {
+			const auto frames = loop.presented_frames.load(std::memory_order_relaxed);
+			if (frames != 0) {
+				UpdateTitle(frames, static_cast<double>(frames - title_frames) * 1000.0 /
+				                        static_cast<double>(elapsed));
+			}
+			title_time   = now;
+			title_frames = frames;
+		}
+		const auto wait_ms = static_cast<int>(title_interval_ms - (now - title_time));
+		if (!HostInputWaitEvent(&loop.event, wait_ms)) {
 			continue;
 		}
 		ProcessEvent(timer.GetTimeS());
@@ -817,6 +842,7 @@ void WindowRun() {
 	EXIT_IF(g_window == nullptr);
 
 	g_window->Run();
+	Common::LockGuard lock(g_window->render_context->GetMutex());
 	g_window->render_context->GetPipelineCache().Save();
 }
 
@@ -904,7 +930,7 @@ void WindowContext::UpdateIcon() {
 	}
 }
 
-void WindowContext::UpdateTitle() {
+void WindowContext::UpdateTitle(uint64_t frame_num, double current_fps) {
 	static char title[128];
 	static char title_id[12];
 	static char app_ver[12];
@@ -914,10 +940,6 @@ void WindowContext::UpdateTitle() {
 	static bool has_app_ver =
 	    Loader::SystemContentParamSfoGetString("APP_VER", app_ver, sizeof(app_ver));
 	static const std::string processor_name = Common::GetSystemInfo().ProcessorName;
-	static uint64_t fps_start   = Common::Timer::QueryPerformanceCounter();
-	static uint64_t frame_num   = 0;
-	static uint64_t fps_frames  = 0;
-	static double   current_fps = 0.0;
 
 #if KYTY_BUILD == KYTY_BUILD_DEBUG
 	static constexpr auto build_type = "Debug";
@@ -927,17 +949,6 @@ void WindowContext::UpdateTitle() {
 	static constexpr auto build_type = "Unknown";
 #endif
 
-	const auto now       = Common::Timer::QueryPerformanceCounter();
-	const auto frequency = Common::Timer::QueryPerformanceFrequency();
-	frame_num++;
-	fps_frames++;
-	if (now - fps_start >= frequency) {
-		current_fps = static_cast<double>(fps_frames) * static_cast<double>(frequency) /
-		              static_cast<double>(now - fps_start);
-		fps_start   = now;
-		fps_frames  = 0;
-	}
-
 	const auto* device_name = graphic_ctx.GetPhysicalDeviceProperties().deviceName.data();
 	auto text = fmt::format(
 	    "[{} | {}] {}{}{}{}{}{}[{}] [{}], frame: {}, fps: {:.0f}", KYTY_BUILD_LABEL, build_type,
@@ -945,16 +956,7 @@ void WindowContext::UpdateTitle() {
 	    (has_title_id ? ", " : ""), (has_app_ver ? app_ver : ""), (has_app_ver ? " " : ""),
 	    device_name, processor_name, frame_num, current_fps);
 
-	struct TitleUpdate {
-		SDL_Window*  window;
-		std::string* text;
-	} update {window, &text};
-	EXIT_IF(!SDL_RunOnMainThread(
-	    [](void* data) {
-		    auto& title = *static_cast<TitleUpdate*>(data);
-		    SDL_SetWindowTitle(title.window, title.text->c_str());
-	    },
-	    &update, true));
+	SDL_SetWindowTitle(window, text.c_str());
 }
 
 } // namespace Libs::Graphics

src/graphics/presentation/window/windowInternal.h

@@ -27,6 +27,7 @@ struct WindowLoopState {
 	SDL_Event        event {};
 	bool             need_exit = false;
 	std::atomic_bool paused    = false;
+	std::atomic<uint64_t> presented_frames {0};
 };
 
 struct WindowContext {
@@ -34,6 +35,7 @@ struct WindowContext {
 	~WindowContext();
 	KYTY_CLASS_NO_COPY(WindowContext);
 
+	[[nodiscard]] static vk::PhysicalDeviceVulkan11Features RequiredVulkan11Features() noexcept;
 	[[nodiscard]] static vk::PhysicalDeviceVulkan12Features RequiredVulkan12Features() noexcept;
 	[[nodiscard]] static vk::PhysicalDeviceVulkan13Features RequiredVulkan13Features() noexcept;
 	[[nodiscard]] static uint32_t InitialWindowFlags(bool fullscreen) noexcept;
@@ -41,8 +43,8 @@ struct WindowContext {
 	void                                                    RecreateSurface();
 	void                                                    RefreshSurfaceCapabilities();
 	void                                                    UpdateIcon();
-	void                                                    UpdateTitle();
-	void                                                    Resize(uint32_t width, uint32_t height);
+	void UpdateTitle(uint64_t frame_num, double current_fps);
+	void                                                    Resize(int width, int height);
 	void ProcessWindowEvent(const SDL_WindowEvent& event);
 	void ProcessDisplayEvent(const SDL_DisplayEvent& event);
 	void ProcessEvent(double time_seconds);

src/graphics/shader/recompiler/ShaderRecompiler.cpp

@@ -7,8 +7,8 @@
 #include "graphics/shader/recompiler/frontend/cfg/ShaderCFG.h"
 #include "graphics/shader/recompiler/frontend/decode/ShaderDecoder.h"
 #include "graphics/shader/recompiler/frontend/translate/Translate.h"
+#include "graphics/shader/recompiler/frontend/translate/Translator.h"
 #include "graphics/shader/recompiler/ir/ShaderIR.h"
-#include "graphics/shader/recompiler/ir/passes/BindingLayout.h"
 #include "graphics/shader/recompiler/ir/passes/ConstantPropagation.h"
 #include "graphics/shader/recompiler/ir/passes/DeadCodeElimination.h"
 #include "graphics/shader/recompiler/ir/passes/ReadLaneElimination.h"
@@ -19,9 +19,9 @@
 
 #include <algorithm>
 #include <array>
-#include <atomic>
 #include <chrono>
 #include <fmt/format.h>
+#include <functional>
 #include <map>
 #include <span>
 #include <utility>
@@ -468,6 +468,7 @@ Decoder::Program DecodeFusedProgram(std::span<const uint32_t> front, std::span<c
 	Decoder::DecodeInstruction(joined_code, front_words - 1u, result.instructions.back());
 	Decoder::Program back_program;
 	Decoder::DecodeProgram(back, back_program);
+	result.has_swap_pc |= back_program.has_swap_pc;
 	const auto back_pc = front_words * sizeof(uint32_t);
 	for (auto& inst: back_program.instructions) {
 		// A back-stage PC-relative data reference requires its guest code address.
@@ -480,11 +481,227 @@ Decoder::Program DecodeFusedProgram(std::span<const uint32_t> front, std::span<c
 	return result;
 }
 
+void PrepareCallTarget(ShaderSource& source, const CompileOptions& options) {
+	using namespace IR;
+	for (uint32_t i = 0; i < source.decoded.instructions.size(); ++i) {
+		if (source.decoded.instructions[i].opcode != Decoder::Opcode::S_SWAPPC_B64) continue;
+		EXIT_NOT_IMPLEMENTED(source.call.has_value() || options.stage != ShaderType::Compute);
+		source.call = ShaderSource::Call {.instruction = i};
+	}
+	EXIT_IF(options.input_info.compute == nullptr);
+	const auto& call = source.decoded.instructions[source.call->instruction];
+	EXIT_NOT_IMPLEMENTED(call.src0.kind != Decoder::OperandKind::Sgpr ||
+	                     call.dst.kind != Decoder::OperandKind::Sgpr ||
+	                     call.src0.reg != call.dst.reg || call.dst.reg >= 105u);
+	const auto graph = CFG::BuildGraph(source.decoded);
+	Program query;
+	query.stage = options.stage;
+	query.user_data_base = options.user_data_base;
+	query.user_data_count = static_cast<uint32_t>(options.user_data.size());
+	query.wave_size = options.wave_size;
+	auto& block = *query.block_storage.emplace_back(std::make_unique<Block>());
+	query.blocks.push_back(&block);
+	struct Producer { bool ready = false; std::map<uint32_t, Value> values; };
+	std::map<uint32_t, Producer> producers;
+	std::function<Value(uint32_t, uint32_t)> Register;
+	std::function<Value(uint32_t, Value)> Resolve = [&](uint32_t before, Value value) -> Value {
+		value = value.Resolve();
+		auto* inst = value.TryInstruction();
+		if (inst == nullptr) return value;
+		uint32_t code = UINT32_MAX;
+		switch (inst->GetOpcode()) {
+			case ValueOpcode::GetScalarRegister: code = static_cast<uint32_t>(inst->Arg(0).ScalarRegister()); break;
+			case ValueOpcode::GetVccLo: code = 106u; break;
+			case ValueOpcode::GetVccHi: code = 107u; break;
+			default: break;
+		}
+		if (code != UINT32_MAX) {
+			const auto result = Register(before, code);
+			inst->ReplaceUsesWith(result);
+			return result;
+		}
+		for (size_t i = 0; i < inst->NumArgs(); ++i) inst->SetArg(i, Resolve(before, inst->Arg(i)));
+		return value;
+	};
+	Register = [&](uint32_t before, uint32_t code) -> Value {
+		const auto index = CFG::FindScalarDefinition(source.decoded, graph, before, code);
+		if (index == UINT32_MAX) {
+			EXIT_IF(code < query.user_data_base || code - query.user_data_base >= query.user_data_count);
+			return IREmitter(&block).GetUserData(static_cast<ScalarReg>(code));
+		}
+		auto [entry, inserted] = producers.try_emplace(index);
+		auto& producer = entry->second;
+		if (inserted) {
+			Block translated;
+			Frontend::Translator translator(query, &translated, 1u,
+			    (options.input_info.compute->float_mode & 0x10u) == 0u,
+			    !options.input_info.compute->async_compute);
+			translator.TranslateInstruction(source.decoded.instructions[index]);
+			for (const auto& inst: translated) {
+				uint32_t destination = UINT32_MAX;
+				Value value;
+				switch (inst.GetOpcode()) {
+					case ValueOpcode::SetScalarRegister:
+						destination = static_cast<uint32_t>(inst.Arg(0).ScalarRegister()); value = inst.Arg(1); break;
+					case ValueOpcode::SetVccLo: destination = 106u; value = inst.Arg(0); break;
+					case ValueOpcode::SetVccHi: destination = 107u; value = inst.Arg(0); break;
+					default: break;
+				}
+				if (destination != UINT32_MAX) producer.values.insert_or_assign(destination, value);
+			}
+			for (auto& [destination, value]: producer.values) value = Resolve(index, value);
+			for (auto& inst: translated) inst.SetParent(&block);
+			block.Instructions().splice(block.end(), translated.Instructions());
+			producer.ready = true;
+		}
+		EXIT_IF(!producer.ready || !producer.values.contains(code));
+		return producer.values.at(code).Resolve();
+	};
+	for (uint32_t i = 0; i < 2u; ++i) {
+		const auto value = Register(source.call->instruction, call.src0.reg + i);
+		EXIT_IF(!ValidateRuntimeValue(query, value, RuntimeValueType::Integer));
+		auto& root = query.value_storage.emplace_back(ValueOpcode::ReferenceU32);
+		root.SetArg(0, value);
+		source.call->target[i] = Value(&root);
+	}
+	RewriteToSsa(query.blocks);
+	ConstantPropagationPass(query.blocks, query.wave_size);
+	RemoveIdentities(query.blocks);
+	EliminateDeadCode(query.blocks);
+	for (auto& value: source.call->target) value = value.Instruction()->Arg(0).Resolve();
+	query.value_storage.clear();
+	for (auto& inst: block) {
+		inst.SetParent(nullptr);
+		if (inst.GetOpcode() == ValueOpcode::LoadAddressU32 || inst.GetOpcode() == ValueOpcode::ReadConstBuffer)
+			inst.SetFlags(SrtReadFlags {.index = inst.Flags<MemoryFlags>().index});
+	}
+	query.value_storage.splice(query.value_storage.end(), block.Instructions());
+	source.call_targets = std::move(static_cast<ResourcePlan&>(query));
+}
+
 } // namespace
 
+ShaderSource PrepareShaderSource(std::span<const uint32_t> code, const CompileOptions& options) {
+	EXIT_IF(code.empty());
+	ShaderSource source;
+	if (!options.back_code.empty()) {
+		source.decoded = DecodeFusedProgram(code, options.back_code, source.code);
+	} else if (options.stage == ShaderType::Local) {
+		source.decoded = Decoder::DecodeFrontProgram(code);
+		auto& handoff = source.decoded.instructions.back();
+		handoff.opcode = Decoder::Opcode::S_ENDPGM;
+		handoff.src_count = 0;
+	} else {
+		Decoder::DecodeProgram(code, source.decoded);
+	}
+	if (source.decoded.has_swap_pc) {
+		PrepareCallTarget(source, options);
+		source.code.assign(code.begin(), code.end());
+		source.decoded.code = source.code;
+	}
+	return source;
+}
+
+const Decoder::Program& RefreshShaderSource(ShaderSource& source, const IR::SrtRuntime& runtime) {
+	auto& reads = source.reads;
+	reads.clear();
+	IR::SrtReadCapture capture(runtime, reads);
+	const auto clean = IR::CleanRuntime(capture.ObservedRuntime());
+	IR::SrtWalker walker(source.call_targets, clean);
+	uint32_t low = 0, high = 0;
+	const auto& call = source.decoded.instructions[source.call->instruction];
+	if (!walker.Evaluate(source.call->target[0], low) || !walker.Evaluate(source.call->target[1], high))
+		EXIT("shader call at pc 0x%08x has an unavailable scalar target", call.pc);
+	// SWAPPC reads the old aliased pair and ignores the target's low two bits.
+	const auto address = ((uint64_t {high} << 32u) | low) & ~uint64_t {3};
+	EXIT_IF(address == 0u);
+	const auto query_reads = reads.size();
+	if (source.linked) {
+		auto& linked = *source.linked;
+		const auto function = std::span(linked.code).subspan(source.code.size());
+		if (clean.read_memory(clean.userdata, address, linked.observed_function) &&
+		    std::ranges::equal(function, linked.observed_function)) {
+			return linked.decoded;
+		}
+		reads.resize(query_reads);
+	}
+	ShaderSource::Linked linked {.code = source.code,
+	                             .decoded = source.decoded};
+	const auto& last = source.decoded.instructions.back();
+	EXIT_NOT_IMPLEMENTED(last.opcode != Decoder::Opcode::S_ENDPGM);
+	// Keep the caller's data/footer bytes at their native offsets. The leaf's virtual CFG
+	// PCs follow decoded instructions; its physical code is independently read and owned.
+	const auto function_pc = last.pc + last.word_count * sizeof(uint32_t);
+	for (uint32_t index = 0;;) {
+		// The observed queue helper is a scalar leaf. Scalar encodings need at most one literal.
+		EXIT_IF(index >= 64u || address > UINT64_MAX - (uint64_t {index} + 2u) * sizeof(uint32_t));
+		std::array<uint32_t, 2> words {};
+		if (!clean.read_memory(clean.userdata, address + uint64_t {index} * sizeof(uint32_t),
+		                       std::span(words).first(1)))
+			EXIT("shader call at pc 0x%08x cannot read leaf code", call.pc);
+		const auto family = Decoder::GetInstructionFamily(words[0]);
+		EXIT_NOT_IMPLEMENTED(family != Decoder::Family::SOP1 && family != Decoder::Family::SOP2 &&
+		                     family != Decoder::Family::SOPC && family != Decoder::Family::SOPK &&
+		                     family != Decoder::Family::SOPP);
+		Decoder::Instruction inst;
+		Decoder::DecodeInstruction(words, 0u, inst);
+		if (inst.word_count == 2u) {
+			if (!clean.read_memory(clean.userdata, address + uint64_t {index + 1u} * sizeof(uint32_t),
+			                       std::span(words).last(1)))
+				EXIT("shader call at pc 0x%08x cannot read leaf literal", call.pc);
+			inst = {};
+			Decoder::DecodeInstruction(words, 0u, inst);
+		}
+		inst.pc = index * sizeof(uint32_t);
+		if (Decoder::IsDirectBranch(inst.opcode)) inst.branch_target += inst.pc;
+		linked.code.insert(linked.code.end(), inst.raw, inst.raw + inst.word_count);
+		index += inst.word_count;
+		EXIT_NOT_IMPLEMENTED(inst.opcode == Decoder::Opcode::S_SWAPPC_B64 ||
+		                     inst.opcode == Decoder::Opcode::S_GETPC_B64 ||
+		                     inst.opcode == Decoder::Opcode::S_ENDPGM);
+		EXIT_IF(CFG::MayWriteScalarRegister(inst, call.dst.reg) || CFG::MayWriteScalarRegister(inst, call.dst.reg + 1u));
+		const bool returns = inst.opcode == Decoder::Opcode::S_SETPC_B64;
+		if (returns) {
+			EXIT_NOT_IMPLEMENTED(inst.src0.kind != Decoder::OperandKind::Sgpr || inst.src0.reg != call.dst.reg);
+			inst.opcode = Decoder::Opcode::S_BRANCH;
+			inst.branch_target = call.pc + call.word_count * sizeof(uint32_t);
+		} else if (Decoder::IsDirectBranch(inst.opcode)) {
+			EXIT_IF(inst.branch_target > UINT32_MAX - function_pc);
+			inst.branch_target += function_pc;
+		}
+		inst.pc += function_pc;
+		linked.decoded.instructions.push_back(std::move(inst));
+		if (returns) break;
+	}
+	const auto function_end = linked.decoded.instructions.back().pc +
+	                          linked.decoded.instructions.back().word_count * sizeof(uint32_t);
+	for (size_t i = source.decoded.instructions.size(); i + 1u < linked.decoded.instructions.size(); ++i) {
+		const auto& inst = linked.decoded.instructions[i];
+		if (Decoder::IsDirectBranch(inst.opcode))
+			EXIT_IF(inst.branch_target < function_pc || inst.branch_target >= function_end);
+	}
+	reads.resize(query_reads);
+	reads.emplace_back(address, (linked.code.size() - source.code.size()) * sizeof(uint32_t));
+	linked.decoded.instructions[source.call->instruction].branch_target = function_pc;
+	linked.decoded.code = linked.code;
+	linked.observed_function.resize(linked.code.size() - source.code.size());
+	source.linked = std::move(linked);
+	++source.revision;
+	return source.linked->decoded;
+}
+
 TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOptions& options) {
-	if (code.empty()) {
-		EXIT("shader recompiler input is empty\n");
+	auto source = PrepareShaderSource(code, options);
+	return TranslateProgram(source.decoded, options);
+}
+
+TranslateResult TranslateProgram(const Decoder::Program& decoded, const CompileOptions& options) {
+	EXIT_IF(decoded.instructions.empty());
+	if (decoded.has_swap_pc) {
+		for (const auto& inst: decoded.instructions) {
+			if (inst.opcode == Decoder::Opcode::S_SWAPPC_B64 && inst.branch_target == UINT32_MAX)
+				EXIT("shader call at pc 0x%08x must be linked before translation", inst.pc);
+		}
 	}
 	if (options.stage != ShaderType::Compute && options.stage != ShaderType::Vertex &&
 	    options.stage != ShaderType::Pixel && options.stage != ShaderType::Mesh &&
@@ -495,48 +712,11 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 	}
 
 	const auto compile_begin = std::chrono::steady_clock::now();
-	const auto phase_ms      = [&compile_begin]() {
+	const auto phase_ms = [&compile_begin]() {
 		return static_cast<uint64_t>(std::chrono::duration_cast<std::chrono::milliseconds>(
-		                                 std::chrono::steady_clock::now() - compile_begin)
-		                                 .count());
+		    std::chrono::steady_clock::now() - compile_begin).count());
 	};
 
-	LOGF("%s phase begin: stage=%s hash=0x%016" PRIx64 " code_words=%" PRIu64 " decode\n",
-	     GetDumpLabel(options), StageName(options.stage), options.shader_hash,
-	     static_cast<uint64_t>(code.size()));
-
-	Decoder::Program decoded;
-	std::vector<uint32_t> joined_code;
-	if (!options.back_code.empty()) {
-		decoded = DecodeFusedProgram(code, options.back_code, joined_code);
-	} else if (options.stage == ShaderType::Local) {
-		decoded = Decoder::DecodeFrontProgram(code);
-		// The separately compiled hull half runs in the next Vulkan stage.
-		auto& handoff     = decoded.instructions.back();
-		handoff.opcode    = Decoder::Opcode::S_ENDPGM;
-		handoff.src_count = 0;
-	} else {
-		Decoder::DecodeProgram(code, decoded);
-	}
-	LOGF("%s phase end: stage=%s hash=0x%016" PRIx64 " decode instructions=%" PRIu64
-	     " elapsed_ms=%" PRIu64 "\n",
-	     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);
@@ -609,7 +789,7 @@ 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);
@@ -617,11 +797,17 @@ 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);
 	}
+	if (decoded.has_swap_pc) {
+		// Leaf returns are resolved CFG edges; native PC values must not escape into GPU data.
+		for (const auto* block: ir.blocks)
+			for (const auto& inst: *block)
+				EXIT_NOT_IMPLEMENTED(inst.GetOpcode() == IR::ValueOpcode::GetShaderBase);
+	}
 	LowerTessellationMemory(ir, options);
 	std::string cfg_dump;
 	if (options.dump_ir) {
@@ -632,7 +818,6 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 		}
 	}
 	IR::TrackResources(ir, decoded, native_cfg);
-	IR::EliminateDeadCode(ir.blocks);
 	TranslateResult result;
 	result.program = std::move(ir);
 	if (options.dump_ir) {
@@ -645,7 +830,6 @@ 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);
@@ -665,11 +849,20 @@ CompileResult CompileProgram(TranslateResult translated, const CompileOptions& o
 				})) {
 					continue;
 				}
+				const auto resource = inst.Flags<uint32_t>();
+				first               = resource < ir.info.buffers.size() &&
+				                              ir.info.buffers[resource].indirect_root == resource
+				                          ? 1u
+				                          : 0u;
 			} 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) {
+			} else if (op == IR::ValueOpcode::GetSamplerResource) {
+				const auto resource = inst.Flags<uint32_t>();
+				first = resource < ir.info.samplers.size() &&
+				                !ir.info.samplers[resource].indirect_resources.empty() ? 1u : 0u;
+			} else {
 				continue;
 			}
 			for (size_t index = first; index < inst.NumArgs(); index++) {
@@ -682,18 +875,17 @@ CompileResult CompileProgram(TranslateResult translated, const CompileOptions& o
 	IR::EliminateDeadCode(ir.blocks);
 
 	IR::CollectShaderInfo(ir, options.input_info);
-	IR::AllocateBindings(ir, push_data_start_dword);
 	std::string ir_dump;
 	if (options.dump_ir) {
 		ir_dump = MakeIrDump(translated.cfg_dump, ir);
 		if (options.early_dump) {
-			LOGF("%s native IR and bindings (early):\n%s", GetDumpLabel(options), ir_dump.c_str());
+			LOGF("%s native IR (early):\n%s", GetDumpLabel(options), ir_dump.c_str());
 		}
 	}
 
 	LOGF("%s phase begin: stage=%s hash=0x%016" PRIx64 " SPIR-V EmitProgram\n",
 	     GetDumpLabel(options), StageName(ir.stage), ir.shader_hash);
-	auto spirv = Spirv::EmitProgram(ir, options.input_info);
+	auto spirv = Spirv::EmitProgram(ir, options.input_info, push_data_start_dword);
 	LOGF("%s phase end: stage=%s hash=0x%016" PRIx64 " SPIR-V EmitProgram words=%" PRIu64
 	     " elapsed_ms=%" PRIu64 "\n",
 	     GetDumpLabel(options), StageName(ir.stage), ir.shader_hash,

src/graphics/shader/recompiler/ShaderRecompiler.h

@@ -4,8 +4,11 @@
 #include "common/common.h"
 #include "common/stringUtils.h"
 #include "graphics/shader/recompiler/ir/passes/ResourceMaterialization.h"
+#include "graphics/shader/recompiler/frontend/decode/ShaderDecoder.h"
 #include "graphics/shader/shader.h"
 
+#include <array>
+#include <optional>
 #include <span>
 #include <vector>
 
@@ -28,7 +31,6 @@ struct TranslateResult {
 	IR::Program program;
 	std::string decoded_dump;
 	std::string cfg_dump;
-	bool        skip_dispatch = false;
 };
 
 struct CompileResult {
@@ -38,6 +40,32 @@ struct CompileResult {
 	IR::Program            program;
 };
 
+// Decoded source and typed scalar call queries are retained once per cache source.
+struct ShaderSource {
+	struct Call {
+		uint32_t instruction;
+		std::array<IR::Value, 2> target;
+	};
+	struct Linked {
+		std::vector<uint32_t> code;
+		Decoder::Program decoded;
+		std::vector<uint32_t> observed_function;
+	};
+	std::vector<uint32_t> code;
+	Decoder::Program decoded;
+	IR::ResourcePlan call_targets;
+	std::optional<Call> call;
+	std::optional<Linked> linked;
+	uint64_t revision = 0;
+	std::vector<std::pair<uint64_t, uint64_t>> reads;
+};
+
+[[nodiscard]] ShaderSource PrepareShaderSource(std::span<const uint32_t> code,
+	                                           const CompileOptions& options);
+const Decoder::Program& RefreshShaderSource(ShaderSource& source, const IR::SrtRuntime& runtime);
+[[nodiscard]] TranslateResult TranslateProgram(const Decoder::Program& decoded,
+	                                           const CompileOptions& options);
+
 [[nodiscard]] TranslateResult TranslateProgram(std::span<const uint32_t> code,
                                                const CompileOptions& options);
 [[nodiscard]] CompileResult CompileProgram(TranslateResult translated,

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/SpirvBuilder.cpp

@@ -171,10 +171,12 @@ void Builder::AddFunction(std::span<const uint32_t> words) {
 }
 
 DeferredPhi Builder::AddDeferredPhi(uint32_t type, uint32_t result, size_t incoming_count) {
-	std::vector<uint32_t> words {spv::OpPhi, type, result};
-	words.resize(words.size() + incoming_count * 2u);
 	const DeferredPhi phi {m_functions.size()};
-	AddFunction(words);
+	const auto word_count = static_cast<uint32_t>(3u + incoming_count * 2u);
+	m_functions.push_back((word_count << spv::WordCountShift) | spv::OpPhi);
+	m_functions.push_back(type);
+	m_functions.push_back(result);
+	m_functions.resize(m_functions.size() + incoming_count * 2u);
 	m_unpatched_phi_incomings += incoming_count;
 	return phi;
 }

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

@@ -3,7 +3,6 @@
 #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>
@@ -18,86 +17,15 @@ namespace {
 	std::abort();
 }
 
-void ValidateNativeProgram(const IR::Program& program) {
-	using Kind                                             = IR::DescriptorBindingKind;
-	constexpr auto                               KindCount = static_cast<size_t>(Kind::Count);
-	std::array<std::vector<uint32_t>, KindCount> expected;
-	std::array<bool, KindCount>                  present {};
-	const auto                                   Dense = [](size_t size) {
-		std::vector<uint32_t> values(size);
-		for (uint32_t i = 0; i < values.size(); i++) {
-			values[i] = i;
-		}
-		return values;
+void ValidateNativeProgram(const IR::Program& program, bool lds_storage) {
+	uint64_t live_buffers = 0;
+	bool uses_lds = false;
+	bool uses_gds = false;
+	const auto uses_mapping = [](const auto& resource) {
+		return resource.indirect_root != UINT32_MAX;
 	};
-	auto Expect = [&](Kind kind, std::vector<uint32_t> resources = {}) {
-		const auto index = static_cast<size_t>(kind);
-		present[index]   = true;
-		expected[index]  = std::move(resources);
-	};
-	for (uint32_t i = 0; i < program.info.images.size(); i++) {
-		const auto kind = IR::DescriptorBindingForImage(program.info.images[i]);
-		if (!kind.has_value()) {
-			Fail(program, "native shader plan has an invalid image class");
-		}
-		present[static_cast<size_t>(*kind)] = true;
-		const auto dynamic = program.info.images[i].mip_mode == IR::ImageMipMode::DynamicStorage;
-		const auto count   = dynamic ? program.info.images[i].mip_count : 1u;
-		if (count == 0u || (!dynamic && program.info.images[i].mip_count != 1u)) {
-			Fail(program, "native shader plan has an invalid image mip descriptor count");
-		}
-		expected[static_cast<size_t>(*kind)].insert(expected[static_cast<size_t>(*kind)].end(),
-		                                            count, i);
-	}
-	if (!program.info.samplers.empty()) {
-		Expect(Kind::Samplers, Dense(program.info.samplers.size()));
-	}
-	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);
-	}
-	if (program.info.uses_dma) {
-		Expect(Kind::BdaPagetable);
-		Expect(Kind::FaultBuffer);
-	}
-	if (IR::UsesFlattenedSrt(program)) {
-		Expect(Kind::FlattenedSrt);
-	}
-	if (program.bindings.ShaderDataDwords() != 0 && !program.bindings.UsesPushData()) {
-		Expect(Kind::ShaderData);
-	}
-
-	std::array<bool, KindCount> seen {};
-	for (const auto& binding: program.bindings.descriptors) {
-		const auto kind = static_cast<size_t>(binding.kind);
-		if (kind >= KindCount || seen[kind] || !present[kind] ||
-		    binding.resources != expected[kind]) {
-			Fail(program, "native descriptor groups do not match shader topology");
-		}
-		seen[kind] = true;
-	}
-	for (size_t i = 0; i < KindCount; i++) {
-		if (present[i] != seen[i]) {
-			Fail(program, "native shader plan is missing a required descriptor group");
-		}
-	}
-	const auto has_shader_data_storage = present[static_cast<size_t>(Kind::ShaderData)];
-	const auto shader_data_dwords = program.bindings.ShaderDataDwords();
-	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 != 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()) ||
-	    std::adjacent_find(program.bindings.user_data_registers.begin(),
-	                       program.bindings.user_data_registers.end()) !=
-	        program.bindings.user_data_registers.end()) {
-		Fail(program, "native shader-data layout is inconsistent");
-	}
-
+	bool uses_flattened_srt = std::ranges::any_of(program.info.buffers, uses_mapping) ||
+	                          std::ranges::any_of(program.info.images, uses_mapping);
 	const auto planning_only_handle = [&](const IR::Inst& handle) {
 		return !handle.Uses().empty() &&
 		       std::ranges::all_of(handle.Uses(), [&](const IR::Use& use) {
@@ -122,8 +50,51 @@ 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 op = inst.GetOpcode();
+			uses_flattened_srt |= op == IR::ValueOpcode::ReadConst;
+			if (IR::BufferAccessOf(op) != IR::BufferAccess::None ||
+			    IR::SharedAccessOf(op) != IR::SharedAccess::None ||
+			    IR::AddressOpcodeInfoOf(op).access != IR::AddressAccess::None) {
+				const auto index = inst.Flags<IR::MemoryFlags>().index;
+				if (index >= program.memory_info.size()) {
+					Fail(program, "typed shader contains invalid memory metadata");
+				}
+				const auto& memory = program.memory_info[index];
+				if (!memory.planning_only) {
+					uses_lds |= memory.kind == IR::ResourceKind::FlatLocal;
+					if (IR::SharedAccessOf(op) != IR::SharedAccess::None) {
+						if (memory.kind != IR::ResourceKind::Lds && memory.kind != IR::ResourceKind::Gds) {
+							Fail(program, "typed shader contains invalid shared-memory metadata");
+						}
+						uses_gds |= memory.kind == IR::ResourceKind::Gds;
+						uses_lds |= memory.kind == IR::ResourceKind::Lds;
+					} else if (memory.kind == IR::ResourceKind::Buffer ||
+					           memory.kind == IR::ResourceKind::ScalarBuffer) {
+						if (memory.resource >= program.info.buffers.size()) {
+							Fail(program, "typed shader contains an invalid buffer resource");
+						}
+						live_buffers |= uint64_t{1} << memory.resource;
+						for (const auto child: program.info.buffers[memory.resource].indirect_resources) {
+							if (child >= program.info.buffers.size()) {
+								Fail(program, "typed shader contains an invalid indirect buffer resource");
+							}
+							live_buffers |= uint64_t{1} << child;
+						}
+					}
+				}
+			}
 			const auto dense = inst.Flags<uint32_t>();
 			switch (inst.GetOpcode()) {
 				case IR::ValueOpcode::GetBufferResource:
@@ -138,7 +109,7 @@ 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;
@@ -169,136 +140,64 @@ void ValidateNativeProgram(const IR::Program& program) {
 			}
 		}
 	}
+	using Kind = IR::DescriptorBindingKind;
+	std::array<uint32_t, static_cast<size_t>(Kind::Count)> expected {};
+	auto& buffer_count = expected[static_cast<size_t>(Kind::Buffers)];
+	for (uint32_t index = 0; index < program.info.buffers.size(); ++index) {
+		const auto slot = (live_buffers & (uint64_t{1} << index)) != 0 ? buffer_count++ : UINT32_MAX;
+		if (program.info.buffers[index].descriptor_index != slot) {
+			Fail(program, "native buffer descriptor index does not match shader topology");
+		}
+	}
+	for (const auto& image: program.info.images) {
+		const auto kind = IR::DescriptorBindingForImage(image);
+		if (!kind.has_value()) {
+			Fail(program, "native shader plan has an invalid image class");
+		}
+		if (image.mip_count == 0u ||
+		    (image.mip_mode != IR::ImageMipMode::Dynamic && image.mip_count != 1u)) {
+			Fail(program, "native shader plan has an invalid image mip descriptor count");
+		}
+		auto& count = expected[static_cast<size_t>(*kind)];
+		if (image.descriptor_index != count) {
+			Fail(program, "native image descriptor index does not match shader topology");
+		}
+		count += image.mip_count;
+	}
+	expected[static_cast<size_t>(Kind::Samplers)] = static_cast<uint32_t>(program.info.samplers.size());
+	expected[static_cast<size_t>(Kind::Gds)] = uses_gds;
+	expected[static_cast<size_t>(Kind::SharedMemory)] = uses_lds && lds_storage;
+	expected[static_cast<size_t>(Kind::BdaPagetable)] = program.info.uses_dma;
+	expected[static_cast<size_t>(Kind::FaultBuffer)] = program.info.uses_dma;
+	expected[static_cast<size_t>(Kind::FlattenedSrt)] = uses_flattened_srt;
+	expected[static_cast<size_t>(Kind::ShaderData)] =
+	    program.bindings.ShaderDataDwords() != 0 && !program.bindings.UsesPushData();
+	uint64_t descriptor_mask = 0;
+	for (uint32_t index = 0; index < expected.size(); ++index) {
+		if (expected[index] != 0) descriptor_mask |= uint64_t{1} << index;
+	}
+	if (program.bindings.descriptor_counts != expected ||
+	    program.bindings.descriptor_mask != descriptor_mask) {
+		Fail(program, "native descriptor counts do not match shader topology");
+	}
+	const auto shader_data_dwords = program.bindings.ShaderDataDwords();
+	const auto& registers = program.info.user_data_registers;
+	const bool has_dispatch_threads = program.bindings.dispatch_thread_dword != IR::PushData::NoStart;
+	if ((program.bindings.UsesPushData() &&
+	     !IR::PushData::CanFit(program.bindings.push_data_start_dword, shader_data_dwords)) ||
+	    (has_dispatch_threads && (program.stage != ShaderType::Compute ||
+	                              program.bindings.dispatch_thread_dword != registers.size())) ||
+	    program.bindings.memory_offset_dword != registers.size() + (has_dispatch_threads ? 3u : 0u) ||
+	    !std::is_sorted(registers.begin(), registers.end()) ||
+	    std::adjacent_find(registers.begin(), registers.end()) != registers.end()) {
+		Fail(program, "native shader-data layout is inconsistent");
+	}
 }
 
 } // namespace
 
-Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program& 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;
-			}
-			const auto address_access = IR::AddressOpcodeInfoOf(inst.GetOpcode()).access;
-			if (address_access != IR::AddressAccess::None) {
-				const auto memory_index = inst.Flags<IR::MemoryFlags>().index;
-				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) {
-					if (program.scratch_dwords == 0) {
-						Fail(program, "scratch operation has no per-thread storage");
-					}
-					requirements.function_scratch = true;
-				} else if (address_access == IR::AddressAccess::Write) {
-					Fail(program, "writable FLAT/GLOBAL addresses require GPU ownership tracking");
-				}
-			}
-			if (IR::BufferAccessOf(inst.GetOpcode()) != IR::BufferAccess::None) {
-				const auto memory_index = inst.Flags<IR::MemoryFlags>().index;
-				if (memory_index >= program.memory_info.size()) {
-					Fail(program, "buffer operation has invalid memory metadata");
-				}
-				const auto& memory = program.memory_info[memory_index];
-				if (memory.kind == IR::ResourceKind::IndirectBuffer) {
-					requirements.subgroup_local_invocation_id = true;
-				}
-				if (memory.kind == IR::ResourceKind::Buffer) {
-					requirements.coherent_buffers |= memory.coherent;
-					if (memory.resource >= program.info.buffers.size()) {
-						Fail(program, "buffer operation has invalid resource metadata");
-					}
-					if ((program.info.buffers[memory.resource].packed_stride & (1u << 20u)) != 0u) {
-						if (program.stage != ShaderType::Compute) {
-							Fail(program, "buffer ADD_TID is only valid for compute shaders");
-						}
-						requirements.subgroup_local_invocation_id = true;
-					}
-				}
-			}
-			const auto shared_access = IR::SharedAccessOf(inst.GetOpcode());
-			if (shared_access != IR::SharedAccess::None) {
-				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");
-				}
-				if (program.stage != ShaderType::Compute && program.stage != ShaderType::Mesh &&
-				    kind == IR::ResourceKind::Lds) {
-					requirements.function_lds = true;
-				}
-				if (shared_access == IR::SharedAccess::Append ||
-				    shared_access == IR::SharedAccess::Consume) {
-					requirements.subgroup_ballot              = true;
-					requirements.subgroup_shuffle             = true;
-					requirements.subgroup_local_invocation_id = true;
-				}
-			}
-			switch (inst.GetOpcode()) {
-				case IR::ValueOpcode::Ballot: requirements.subgroup_ballot = true; break;
-				case IR::ValueOpcode::DppMoveU32:
-				case IR::ValueOpcode::ReadFirstLane:
-				case IR::ValueOpcode::ReadLane: {
-					requirements.subgroup_ballot  = true;
-					requirements.subgroup_shuffle = true;
-					if (inst.GetOpcode() == IR::ValueOpcode::DppMoveU32) {
-						requirements.subgroup_local_invocation_id = true;
-					}
-					break;
-				}
-				case IR::ValueOpcode::DppUpdateU32:
-				case IR::ValueOpcode::WriteLane: {
-					requirements.subgroup_ballot              = true;
-					requirements.subgroup_local_invocation_id = true;
-					break;
-				}
-				case IR::ValueOpcode::Permlane16U32: {
-					requirements.subgroup_ballot              = true;
-					requirements.subgroup_shuffle             = true;
-					requirements.subgroup_local_invocation_id = true;
-					break;
-				}
-				case IR::ValueOpcode::SwizzleU32:
-				case IR::ValueOpcode::BpermuteU32: {
-					requirements.subgroup_ballot              = true;
-					requirements.subgroup_shuffle             = true;
-					requirements.subgroup_local_invocation_id = true;
-					break;
-				}
-				case IR::ValueOpcode::LaneId:
-					requirements.subgroup_local_invocation_id |=
-					    program.stage != ShaderType::TessellationControl;
-					break;
-				case IR::ValueOpcode::ImageQueryLod: requirements.compute_derivatives = true; break;
-				case IR::ValueOpcode::ImageGatherRaw:
-					requirements.image_gather_extended = true;
-					break;
-				case IR::ValueOpcode::SetAttribute: {
-					const auto index = inst.Flags<IR::ExportFlags>().index;
-					if (index >= program.export_info.size()) {
-						Fail(program, "attribute export has invalid metadata");
-					}
-					if (program.stage == ShaderType::Pixel &&
-					    program.export_info[index].vm) {
-						requirements.pixel_valid_mask = true;
-					}
-					break;
-				}
-				default: break;
-			}
-		}
-	}
-	return requirements;
-}
-
-std::vector<uint32_t> EmitProgram(const IR::Program& program,
-                                  ShaderStageInputInfo input_info) {
+std::vector<uint32_t> EmitProgram(IR::Program& program, ShaderStageInputInfo input_info,
+                                  uint32_t push_data_start_dword) {
 	using namespace Emitter;
 
 	if (program.stage != ShaderType::Compute && program.stage != ShaderType::Vertex &&
@@ -308,11 +207,14 @@ std::vector<uint32_t> EmitProgram(const IR::Program& program,
 		Fail(program, "binary SPIR-V emitter received an unsupported shader stage");
 	}
 	if (!program.srt_plan_complete || !program.resource_tracking_complete ||
-	    !program.shader_info_complete || !program.binding_layout_complete) {
+	    !program.shader_info_complete) {
 		Fail(program, "SPIR-V emitter requires a fully planned native shader program");
 	}
-	ValidateNativeProgram(program);
+	if (program.info.buffers.size() > IR::ShaderInfo::MaxBuffers) {
+		Fail(program, "native shader exceeds the buffer resource limit");
+	}
 	IR::ValidateProgram(program, true);
+	program.bindings = {.push_data_start_dword = push_data_start_dword};
 	EmitterState state(program, input_info);
 	const auto* workgroup = ShaderWorkgroupInput(program.stage, input_info);
 	state.lane_count =
@@ -320,6 +222,8 @@ std::vector<uint32_t> EmitProgram(const IR::Program& program,
 	        ? 2u
 	        : 1u;
 	DefineModule(state);
+	ValidateNativeProgram(program, program.stage == ShaderType::Compute &&
+	                                   input_info.compute != nullptr && input_info.compute->lds_storage);
 	EmitProgram(state);
 	state.builder.AddEntryPoint(ExecutionModelForStage(state.program.stage), state.main_func,
 	                            "main", state.interface_variables);

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

@@ -9,8 +9,9 @@
 
 namespace Libs::Graphics::ShaderRecompiler::Spirv {
 
-std::vector<uint32_t> EmitProgram(const IR::Program& program,
-                                  ShaderStageInputInfo input_info);
+// Resets and writes backend IDs and descriptor indices; do not emit the same program concurrently.
+std::vector<uint32_t> EmitProgram(IR::Program& program, ShaderStageInputInfo input_info,
+                                  uint32_t push_data_start_dword = 0);
 
 } // namespace Libs::Graphics::ShaderRecompiler::Spirv
 

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

@@ -17,28 +17,26 @@ Pair ExtractPair(EmitterState& state, uint32_t value) {
 	return result;
 }
 
-uint32_t MakePair(EmitterState& state, uint32_t low, uint32_t high) {
-	const auto result = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpCompositeConstruct, TypeU64(state), result, low, high);
-	return result;
-}
+uint32_t EmitMinMaxF64(EmitterState& state, uint32_t lhs, uint32_t rhs, bool max_value) {
+	const auto bits_type = TypeU32Vector(state, 2);
+	const auto lhs_bits = Unary(state, spv::OpBitcast, bits_type, lhs);
+	const auto rhs_bits = Unary(state, spv::OpBitcast, bits_type, rhs);
+	const auto ordered = Binary(state, max_value ? spv::OpFOrdGreaterThan : spv::OpFOrdLessThan,
+	                            TypeBool(state), lhs, rhs);
+	auto result = Select(state, TypeF64(state), ordered, lhs, rhs);
 
-uint32_t CompareEqual64(EmitterState& state, uint32_t lhs_value, uint32_t rhs_value,
-                        bool not_equal) {
-	const auto compare = Binary(state, not_equal ? spv::OpINotEqual : spv::OpIEqual,
-	                            TypeBoolVector(state, 2), lhs_value, rhs_value);
-	return Unary(state, not_equal ? spv::OpAny : spv::OpAll, TypeBool(state), compare);
-}
+	// Equal values have identical bits except signed zero: min chooses -0, max chooses +0.
+	const auto equal = Binary(state, spv::OpFOrdEqual, TypeBool(state), lhs, rhs);
+	const auto equal_bits = Binary(state, max_value ? spv::OpBitwiseAnd : spv::OpBitwiseOr,
+	                               bits_type, lhs_bits, rhs_bits);
+	result = Select(state, TypeF64(state), equal,
+	                Unary(state, spv::OpBitcast, TypeF64(state), equal_bits), result);
 
-uint32_t CompareOrdered64(EmitterState& state, uint32_t lhs_value, uint32_t rhs_value,
-                          spv::Op high_compare, spv::Op low_compare) {
-	const auto lhs         = ExtractPair(state, lhs_value);
-	const auto rhs         = ExtractPair(state, rhs_value);
-	const auto high_equal  = Binary(state, spv::OpIEqual, TypeBool(state), lhs.high, rhs.high);
-	const auto high_result = Binary(state, high_compare, TypeBool(state), lhs.high, rhs.high);
-	const auto low_result  = Binary(state, low_compare, TypeBool(state), lhs.low, rhs.low);
-	const auto low_path = Binary(state, spv::OpLogicalAnd, TypeBool(state), high_equal, low_result);
-	return Binary(state, spv::OpLogicalOr, TypeBool(state), high_result, low_path);
+	// Non-IEEE mode selects the other operand for NaN, including rhs when both are NaN.
+	result = Select(state, TypeF64(state), Unary(state, spv::OpIsNan, TypeBool(state), rhs),
+	                lhs, result);
+	return Select(state, TypeF64(state), Unary(state, spv::OpIsNan, TypeBool(state), lhs),
+	              rhs, result);
 }
 
 uint32_t EmitMulHigh(EmitterState& state, uint32_t lhs, uint32_t rhs, bool signed_value) {
@@ -58,113 +56,14 @@ uint32_t EmitMulHigh(EmitterState& state, uint32_t lhs, uint32_t rhs, bool signe
 	return signed_value ? Unary(state, spv::OpBitcast, TypeU32(state), high) : high;
 }
 
-uint32_t EmitShift64(EmitterState& state, spv::Op opcode, uint32_t value, uint32_t shift) {
-	const auto pair       = ExtractPair(state, value);
-	const auto amount     = EmitAndConstant(state, shift, 63u);
-	const auto word_shift = EmitAndConstant(state, amount, 31u);
-	const auto at_least_32 =
-	    Binary(state, spv::OpUGreaterThanEqual, TypeBool(state), amount, ConstantU32(state, 32));
-	const auto nonzero =
-	    Binary(state, spv::OpINotEqual, TypeBool(state), amount, ConstantU32(state, 0));
-	const auto carry_count = EmitAndConstant(
-	    state, Binary(state, spv::OpISub, TypeU32(state), ConstantU32(state, 32), word_shift), 31u);
-	if (opcode == spv::OpShiftLeftLogical) {
-		const auto low = Binary(state, opcode, TypeU32(state), pair.low, word_shift);
-		const auto carry =
-		    Select(state, TypeU32(state), nonzero,
-		           Binary(state, spv::OpShiftRightLogical, TypeU32(state), pair.low, carry_count),
-		           ConstantU32(state, 0));
-		const auto high =
-		    Binary(state, spv::OpBitwiseOr, TypeU32(state),
-		           Binary(state, opcode, TypeU32(state), pair.high, word_shift), carry);
-		return MakePair(state,
-		                Select(state, TypeU32(state), at_least_32, ConstantU32(state, 0), low),
-		                Select(state, TypeU32(state), at_least_32, low, high));
-	}
-	const auto high = Binary(state, opcode, TypeU32(state), pair.high, word_shift);
-	const auto carry =
-	    Select(state, TypeU32(state), nonzero,
-	           Binary(state, spv::OpShiftLeftLogical, TypeU32(state), pair.high, carry_count),
-	           ConstantU32(state, 0));
-	const auto low = Binary(
-	    state, spv::OpBitwiseOr, TypeU32(state),
-	    Binary(state, spv::OpShiftRightLogical, TypeU32(state), pair.low, word_shift), carry);
-	const auto fill = opcode == spv::OpShiftRightArithmetic
-	                      ? Binary(state, opcode, TypeU32(state), pair.high, ConstantU32(state, 31))
-	                      : ConstantU32(state, 0);
-	return MakePair(state, Select(state, TypeU32(state), at_least_32, high, low),
-	                Select(state, TypeU32(state), at_least_32, fill, high));
-}
-
-uint32_t EmitConstantShift64(EmitterState& state, spv::Op opcode, uint32_t value, uint32_t shift) {
-	shift &= 63u;
-	if (shift == 0u) {
-		return value;
-	}
-	const auto pair = ExtractPair(state, value);
-	if (opcode == spv::OpShiftLeftLogical) {
-		if (shift < 32u) {
-			const auto low  = Binary(state, spv::OpShiftLeftLogical, TypeU32(state), pair.low,
-			                         ConstantU32(state, shift));
-			const auto high = Binary(state, spv::OpBitwiseOr, TypeU32(state),
-			                         Binary(state, spv::OpShiftLeftLogical, TypeU32(state),
-			                                pair.high, ConstantU32(state, shift)),
-			                         Binary(state, spv::OpShiftRightLogical, TypeU32(state),
-			                                pair.low, ConstantU32(state, 32u - shift)));
-			return MakePair(state, low, high);
-		}
-		return MakePair(state, ConstantU32(state, 0),
-		                shift == 32u ? pair.low
-		                             : Binary(state, spv::OpShiftLeftLogical, TypeU32(state),
-		                                      pair.low, ConstantU32(state, shift - 32u)));
-	}
-	if (shift < 32u) {
-		const auto low = Binary(state, spv::OpBitwiseOr, TypeU32(state),
-		                        Binary(state, spv::OpShiftRightLogical, TypeU32(state), pair.low,
-		                               ConstantU32(state, shift)),
-		                        Binary(state, spv::OpShiftLeftLogical, TypeU32(state), pair.high,
-		                               ConstantU32(state, 32u - shift)));
-		const auto high =
-		    Binary(state, opcode, TypeU32(state), pair.high, ConstantU32(state, shift));
-		return MakePair(state, low, high);
-	}
-	const auto high = opcode == spv::OpShiftRightArithmetic
-	                      ? Binary(state, spv::OpShiftRightArithmetic, TypeU32(state), pair.high,
-	                               ConstantU32(state, 31u))
-	                      : ConstantU32(state, 0);
-	const auto low  = shift == 32u ? pair.high
-	                               : Binary(state, opcode, TypeU32(state), pair.high,
-	                                        ConstantU32(state, shift - 32u));
-	return MakePair(state, low, high);
-}
-
-uint32_t EmitMinMax3(EmitterState& state, uint32_t a, uint32_t b, uint32_t c, bool signed_value,
-                     bool max_value) {
-	const auto ab = signed_value ? EmitMinMaxI32Value(state, a, b, max_value)
-	                             : EmitMinMaxU32Value(state, a, b, max_value);
-	return signed_value ? EmitMinMaxI32Value(state, ab, c, max_value)
-	                    : EmitMinMaxU32Value(state, ab, c, max_value);
-}
-
-uint32_t EmitMed3(EmitterState& state, uint32_t a, uint32_t b, uint32_t c, bool signed_value) {
-	const auto minimum = EmitMinMax3(state, a, b, c, signed_value, false);
-	const auto maximum = EmitMinMax3(state, a, b, c, signed_value, true);
-	const auto ab      = Binary(state, spv::OpIAdd, TypeU32(state), a, b);
-	const auto abc     = Binary(state, spv::OpIAdd, TypeU32(state), ab, c);
-	return Binary(state, spv::OpISub, TypeU32(state),
-	              Binary(state, spv::OpISub, TypeU32(state), abc, minimum), maximum);
-}
-
 uint32_t EmitFMinMax3(EmitterState& state, uint32_t a, uint32_t b, uint32_t c, bool max_value) {
 	return EmitMinMaxF32Value(state, EmitMinMaxF32Value(state, a, b, max_value), c, max_value);
 }
 
 uint32_t EmitExt(EmitterState& state, uint32_t type, uint32_t opcode,
                  std::initializer_list<uint32_t> args) {
-	const auto            result = state.builder.AllocateId();
-	std::vector<uint32_t> words {spv::OpExtInst, type, result, GlslStd450(state), opcode};
-	words.insert(words.end(), args.begin(), args.end());
-	state.builder.AddFunction(words);
+	const auto result = state.builder.AllocateId();
+	state.builder.AddFunction(spv::OpExtInst, type, result, GlslStd450(state), opcode, args);
 	return result;
 }
 
@@ -201,6 +100,28 @@ uint32_t EmitF32ToU32(EmitterState& state, uint32_t src, bool signed_value) {
 }
 
 } // namespace
+
+uint32_t EmitFPFma32(EmitterState& state, uint32_t a, uint32_t b, uint32_t c) {
+	const auto result = EmitGlsl<GLSLstd450Fma, IR::Type::F32>(state, a, b, c);
+	state.builder.AddAnnotation(spv::OpDecorate, result, spv::DecorationNoContraction);
+	return result;
+}
+
+uint32_t EmitFPMad32(EmitterState& state, uint32_t a, uint32_t b, uint32_t c) {
+	// Explicit denormal checks typically drop 3DMiniGolf menu FPS from 28-29 to 21-24
+	// on NVIDIA RTX 5080 Laptop GPU, so leave them disabled for this experiment.
+	// a = EmitFlushF32DenormToSignedZero(state, a);
+	// b = EmitFlushF32DenormToSignedZero(state, b);
+	// c = EmitFlushF32DenormToSignedZero(state, c);
+	const auto product = EmitFPMul32(state, a, b);
+	state.builder.AddAnnotation(spv::OpDecorate, product, spv::DecorationNoContraction);
+	// const auto sum = EmitFPAdd32(state, EmitFlushF32DenormToSignedZero(state, product), c);
+	const auto sum = EmitFPAdd32(state, product, c);
+	state.builder.AddAnnotation(spv::OpDecorate, sum, spv::DecorationNoContraction);
+	// return EmitFlushF32DenormToSignedZero(state, sum);
+	return sum;
+}
+
 uint32_t EmitFPMedTri32(EmitterState& state, uint32_t a, uint32_t b, uint32_t c) {
 	const auto min_ab   = EmitMinMaxF32Value(state, a, b, false);
 	const auto min3     = EmitMinMaxF32Value(state, min_ab, c, false);
@@ -215,57 +136,15 @@ uint32_t EmitFPMedTri32(EmitterState& state, uint32_t a, uint32_t b, uint32_t c)
 }
 
 uint32_t EmitFindUMsb64(EmitterState& state, uint32_t value) {
-	const auto pair   = ExtractPair(state, value);
-	const auto high_i = state.builder.AllocateId();
-	const auto low_i  = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpExtInst, TypeI32(state), high_i, GlslStd450(state),
-	                          GLSLstd450FindUMsb, pair.high);
-	state.builder.AddFunction(spv::OpExtInst, TypeI32(state), low_i, GlslStd450(state),
-	                          GLSLstd450FindUMsb, pair.low);
-	const auto high = Unary(state, spv::OpBitcast, TypeU32(state), high_i);
-	const auto low  = Unary(state, spv::OpBitcast, TypeU32(state), low_i);
-	const auto high_nonzero =
-	    Binary(state, spv::OpINotEqual, TypeBool(state), pair.high, ConstantU32(state, 0));
-	return Select(state, TypeU32(state), high_nonzero,
-	              Binary(state, spv::OpIAdd, TypeU32(state), high, ConstantU32(state, 32)), low);
-}
-
-uint32_t EmitIMul64(EmitterState& state, uint32_t lhs_value, uint32_t rhs_value) {
-	const auto lhs   = ExtractPair(state, lhs_value);
-	const auto rhs   = ExtractPair(state, rhs_value);
-	const auto low   = Binary(state, spv::OpIMul, TypeU32(state), lhs.low, rhs.low);
-	const auto high0 = EmitMulHigh(state, lhs.low, rhs.low, false);
-	const auto high1 = Binary(state, spv::OpIMul, TypeU32(state), lhs.low, rhs.high);
-	const auto high2 = Binary(state, spv::OpIMul, TypeU32(state), lhs.high, rhs.low);
-	return MakePair(state, low,
-	                Binary(state, spv::OpIAdd, TypeU32(state),
-	                       Binary(state, spv::OpIAdd, TypeU32(state), high0, high1), high2));
-}
-
-uint32_t EmitISub64(EmitterState& state, uint32_t lhs_value, uint32_t rhs_value) {
-	const auto lhs    = ExtractPair(state, lhs_value);
-	const auto rhs    = ExtractPair(state, rhs_value);
-	const auto low    = Binary(state, spv::OpISub, TypeU32(state), lhs.low, rhs.low);
-	const auto borrow = Binary(state, spv::OpULessThan, TypeBool(state), lhs.low, rhs.low);
-	const auto borrow_u32 =
-	    Select(state, TypeU32(state), borrow, ConstantU32(state, 1), ConstantU32(state, 0));
-	const auto high0 = Binary(state, spv::OpISub, TypeU32(state), lhs.high, rhs.high);
-	const auto high  = Binary(state, spv::OpISub, TypeU32(state), high0, borrow_u32);
-	return MakePair(state, low, high);
-}
-
-uint32_t EmitIAdd64(EmitterState& state, uint32_t lhs_value, uint32_t rhs_value) {
-	const auto lhs      = ExtractPair(state, lhs_value);
-	const auto rhs      = ExtractPair(state, rhs_value);
-	const auto low_pair = state.builder.AllocateId();
-	const auto low      = state.builder.AllocateId();
-	const auto carry    = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpIAddCarry, TypeU32Pair(state), low_pair, lhs.low, rhs.low);
-	state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(state), low, low_pair, 0);
-	state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(state), carry, low_pair, 1);
-	const auto high0 = Binary(state, spv::OpIAdd, TypeU32(state), lhs.high, rhs.high);
-	const auto high  = Binary(state, spv::OpIAdd, TypeU32(state), high0, carry);
-	return MakePair(state, low, high);
+	// GLSL FindUMsb is restricted to 32-bit components.
+	const auto unpacked = Unary(state, spv::OpBitcast, TypeU32Vector(state, 2), value);
+	const auto pair =
+	    ExtractPair(state, EmitExt(state, TypeU32Vector(state, 2), GLSLstd450FindUMsb, {unpacked}));
+	const auto high_found =
+	    Binary(state, spv::OpINotEqual, TypeBool(state), pair.high, ConstantU32(state, UINT32_MAX));
+	return Select(state, TypeU32(state), high_found,
+	              Binary(state, spv::OpIAdd, TypeU32(state), pair.high, ConstantU32(state, 32)),
+	              pair.low);
 }
 
 uint32_t EmitConvertU16U32(EmitterState& state, uint32_t arg0) {
@@ -294,7 +173,7 @@ uint32_t EmitConvertU32F32(EmitterState& state, uint32_t arg0) {
 uint32_t EmitConvertF32F64(EmitterState& state, uint32_t arg0) {
 	const auto converted = Unary(state, spv::OpFConvert, TypeF32(state), arg0);
 	const auto source    = EmitNative<spv::OpCompositeExtract, IR::Type::U32>(
-	    state, Unary(state, spv::OpBitcast, TypeU64(state), arg0), 1u);
+	    state, Unary(state, spv::OpBitcast, TypeU32Vector(state, 2), arg0), 1u);
 	const auto exponent = EmitAndConstant(state, source, 0x7ff00000u);
 	const auto overflow =
 	    Binary(state, spv::OpLogicalAnd, TypeBool(state),
@@ -312,7 +191,21 @@ uint32_t EmitConvertF64F32(EmitterState& state, uint32_t arg0) {
 	                                                  EmitFlushF32DenormToSignedZero(state, arg0));
 }
 
+uint32_t EmitCompositeConstructU64(ValueEmitContext& ctx, uint32_t arg0, IR::Value arg1) {
+	arg1 = arg1.Resolve();
+	return arg1.IsImmediate() && arg1.U32() == 0u
+	           ? Unary(ctx.state, spv::OpUConvert, TypeU64(ctx.state), arg0)
+	           : PackU64(ctx.state, arg0, ctx.Def(arg1));
+}
+
 uint32_t EmitCompositeExtractU64(EmitterState& state, uint32_t arg0, IR::Value arg1) {
+	const auto value = arg1.U32() == 0u ? arg0
+	                                    : Binary(state, spv::OpShiftRightLogical, TypeU64(state),
+	                                             arg0, ConstantU32(state, 32));
+	return Unary(state, spv::OpUConvert, TypeU32(state), value);
+}
+
+uint32_t EmitCompositeExtractU32x2(EmitterState& state, uint32_t arg0, IR::Value arg1) {
 	return EmitNative<spv::OpCompositeExtract, IR::Type::U32>(state, arg0, arg1.U32());
 }
 
@@ -336,116 +229,38 @@ uint32_t EmitUMulHi(EmitterState& state, uint32_t arg0, uint32_t arg1) {
 	return EmitMulHigh(state, arg0, arg1, false);
 }
 
-uint32_t EmitIAbs32(EmitterState& state, uint32_t arg0) {
-	const auto value = arg0;
-	const auto neg   = Unary(state, spv::OpSNegate, TypeU32(state), value);
-	const auto negative =
-	    Binary(state, spv::OpSLessThan, TypeBool(state), value, ConstantU32(state, 0));
-	return Select(state, TypeU32(state), negative, neg, value);
-}
-
-uint32_t EmitShiftLeftLogical64(ValueEmitContext& ctx, uint32_t arg0, IR::Value arg1) {
-	auto& state = ctx.state;
-	return arg1.Resolve().IsImmediate()
-	           ? EmitConstantShift64(state, spv::OpShiftLeftLogical, arg0, arg1.Resolve().U32())
-	           : EmitShift64(state, spv::OpShiftLeftLogical, arg0, ctx.Def(arg1));
-}
-
-uint32_t EmitShiftRightLogical64(ValueEmitContext& ctx, uint32_t arg0, IR::Value arg1) {
-	auto& state = ctx.state;
-	return arg1.Resolve().IsImmediate()
-	           ? EmitConstantShift64(state, spv::OpShiftRightLogical, arg0, arg1.Resolve().U32())
-	           : EmitShift64(state, spv::OpShiftRightLogical, arg0, ctx.Def(arg1));
-}
-
-uint32_t EmitShiftRightArithmetic64(ValueEmitContext& ctx, uint32_t arg0, IR::Value arg1) {
-	auto& state = ctx.state;
-	return arg1.Resolve().IsImmediate()
-	           ? EmitConstantShift64(state, spv::OpShiftRightArithmetic, arg0, arg1.Resolve().U32())
-	           : EmitShift64(state, spv::OpShiftRightArithmetic, arg0, ctx.Def(arg1));
-}
-
-uint32_t EmitBitwiseAnd64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
-	return Binary(state, spv::OpBitwiseAnd, TypeU64(state), arg0, arg1);
-}
-
 uint32_t EmitBitCount64(EmitterState& state, uint32_t arg0) {
-	const auto pair = ExtractPair(state, Unary(state, spv::OpBitCount, TypeU64(state), arg0));
+	// Vulkan bit-count operands remain 32-bit, including when counting a U64.
+	const auto unpacked = Unary(state, spv::OpBitcast, TypeU32Vector(state, 2), arg0);
+	const auto pair =
+	    ExtractPair(state, Unary(state, spv::OpBitCount, TypeU32Vector(state, 2), unpacked));
 	return Binary(state, spv::OpIAdd, TypeU32(state), pair.low, pair.high);
 }
 
-uint32_t EmitFindILsb32(EmitterState& state, uint32_t arg0) {
-	const auto value = EmitExt(state, TypeI32(state), GLSLstd450FindILsb, {arg0});
-	return Unary(state, spv::OpBitcast, TypeU32(state), value);
-}
-
-uint32_t EmitFindUMsb32(EmitterState& state, uint32_t arg0) {
-	const auto value = EmitExt(state, TypeI32(state), GLSLstd450FindUMsb, {arg0});
-	return Unary(state, spv::OpBitcast, TypeU32(state), value);
-}
-
-uint32_t EmitSMin32(EmitterState& state, uint32_t arg0, uint32_t arg1) {
-	return EmitMinMaxI32Value(state, arg0, arg1, false);
-}
-
-uint32_t EmitSMax32(EmitterState& state, uint32_t arg0, uint32_t arg1) {
-	return EmitMinMaxI32Value(state, arg0, arg1, true);
-}
-
-uint32_t EmitUMin32(EmitterState& state, uint32_t arg0, uint32_t arg1) {
-	return EmitMinMaxU32Value(state, arg0, arg1, false);
-}
-
-uint32_t EmitUMax32(EmitterState& state, uint32_t arg0, uint32_t arg1) {
-	return EmitMinMaxU32Value(state, arg0, arg1, true);
-}
-
 uint32_t EmitSMinTri32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32_t arg2) {
-	return EmitMinMax3(state, arg0, arg1, arg2, true, false);
+	return EmitSMin32(state, arg0, EmitSMin32(state, arg1, arg2));
 }
 
 uint32_t EmitSMaxTri32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32_t arg2) {
-	return EmitMinMax3(state, arg0, arg1, arg2, true, true);
+	return EmitSMax32(state, arg0, EmitSMax32(state, arg1, arg2));
 }
 
 uint32_t EmitUMinTri32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32_t arg2) {
-	return EmitMinMax3(state, arg0, arg1, arg2, false, false);
+	return EmitUMin32(state, arg0, EmitUMin32(state, arg1, arg2));
 }
 
 uint32_t EmitUMaxTri32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32_t arg2) {
-	return EmitMinMax3(state, arg0, arg1, arg2, false, true);
+	return EmitUMax32(state, arg0, EmitUMax32(state, arg1, arg2));
 }
 
 uint32_t EmitSMedTri32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32_t arg2) {
-	return EmitMed3(state, arg0, arg1, arg2, true);
+	const auto high_min = EmitSMin32(state, EmitSMax32(state, arg0, arg1), arg2);
+	return EmitSMax32(state, EmitSMin32(state, arg0, arg1), high_min);
 }
 
 uint32_t EmitUMedTri32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32_t arg2) {
-	return EmitMed3(state, arg0, arg1, arg2, false);
-}
-
-uint32_t EmitIEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
-	return CompareEqual64(state, arg0, arg1, false);
-}
-
-uint32_t EmitINotEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
-	return CompareEqual64(state, arg0, arg1, true);
-}
-
-uint32_t EmitULessThan64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
-	return CompareOrdered64(state, arg0, arg1, spv::OpULessThan, spv::OpULessThan);
-}
-
-uint32_t EmitSLessThan64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
-	return CompareOrdered64(state, arg0, arg1, spv::OpSLessThan, spv::OpULessThan);
-}
-
-uint32_t EmitUGreaterThan64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
-	return CompareOrdered64(state, arg0, arg1, spv::OpUGreaterThan, spv::OpUGreaterThan);
-}
-
-uint32_t EmitFPIsNan32(EmitterState& state, uint32_t arg0) {
-	return EmitNative<spv::OpFUnordNotEqual, IR::Type::U1>(state, arg0, arg0);
+	const auto high_min = EmitUMin32(state, EmitUMax32(state, arg0, arg1), arg2);
+	return EmitUMax32(state, EmitUMin32(state, arg0, arg1), high_min);
 }
 
 uint32_t EmitFPMin32(EmitterState& state, uint32_t arg0, uint32_t arg1) {
@@ -456,6 +271,14 @@ uint32_t EmitFPMax32(EmitterState& state, uint32_t arg0, uint32_t arg1) {
 	return EmitMinMaxF32Value(state, arg0, arg1, true);
 }
 
+uint32_t EmitFPMin64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
+	return EmitMinMaxF64(state, arg0, arg1, false);
+}
+
+uint32_t EmitFPMax64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
+	return EmitMinMaxF64(state, arg0, arg1, true);
+}
+
 uint32_t EmitFPMinTri32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32_t arg2) {
 	return EmitFMinMax3(state, arg0, arg1, arg2, false);
 }

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

@@ -82,35 +82,12 @@ uint32_t EmitF32ToF16RtzBits(EmitterState& state, uint32_t f32) {
 	return EmitAndConstant(state, EmitSelectValueU32(state, exp_eq_255, special, finite2), 0xffffu);
 }
 
-uint32_t EmitMinMaxU32Value(EmitterState& state, uint32_t lhs, uint32_t rhs, bool max_value) {
-	const auto cond = state.builder.AllocateId();
-	const auto ret  = state.builder.AllocateId();
-	state.builder.AddFunction(max_value ? spv::OpUGreaterThan : spv::OpULessThan, TypeBool(state),
-	                          cond, lhs, rhs);
-	state.builder.AddFunction(spv::OpSelect, TypeU32(state), ret, cond, lhs, rhs);
-	return ret;
-}
-
-uint32_t EmitMinMaxI32Value(EmitterState& state, uint32_t lhs, uint32_t rhs, bool max_value) {
-	const auto cond = state.builder.AllocateId();
-	const auto ret  = state.builder.AllocateId();
-	state.builder.AddFunction(max_value ? spv::OpSGreaterThan : spv::OpSLessThan, TypeBool(state),
-	                          cond, lhs, rhs);
-	state.builder.AddFunction(spv::OpSelect, TypeU32(state), ret, cond, lhs, rhs);
-	return ret;
-}
-
 F32Class EmitClassifyF32Bits(EmitterState& state, uint32_t bits) {
 	F32Class cls;
-	cls.bits                 = bits;
-	const auto abs_bits      = EmitAndConstant(state, cls.bits, 0x7fffffffu);
-	const auto exponent_bits = EmitAndConstant(state, abs_bits, 0x7f800000u);
-	const auto mantissa_bits = EmitAndConstant(state, abs_bits, 0x007fffffu);
-	const auto exponent_max =
-	    EmitCompareU32Constant(state, spv::OpIEqual, exponent_bits, 0x7f800000u);
-	const auto mantissa_nonzero = EmitCompareU32Constant(state, spv::OpINotEqual, mantissa_bits, 0);
-	cls.nan  = EmitLogicalAndBool(state, exponent_max, mantissa_nonzero);
-	cls.zero                    = EmitCompareU32Constant(state, spv::OpIEqual, abs_bits, 0);
+	cls.bits            = bits;
+	const auto abs_bits = EmitAndConstant(state, cls.bits, 0x7fffffffu);
+	cls.nan  = EmitCompareU32Constant(state, spv::OpUGreaterThan, abs_bits, 0x7f800000u);
+	cls.zero = EmitCompareU32Constant(state, spv::OpIEqual, abs_bits, 0);
 	return cls;
 }
 
@@ -125,20 +102,23 @@ uint32_t EmitClassMaskBitMatch(EmitterState& state, uint32_t mask, uint32_t bit,
 	return EmitLogicalAndBool(state, selected, class_match);
 }
 
-uint32_t EmitClassMaskF32(EmitterState& state, uint32_t value, uint32_t mask) {
-	const auto bits          = EmitBitcastF32ToU32(state, value);
-	const auto sign_bits     = EmitAndConstant(state, bits, 0x80000000u);
-	const auto abs_bits      = EmitAndConstant(state, bits, 0x7fffffffu);
-	const auto exponent_bits = EmitAndConstant(state, abs_bits, 0x7f800000u);
-	const auto mantissa_bits = EmitAndConstant(state, abs_bits, 0x007fffffu);
-	const auto quiet_bits    = EmitAndConstant(state, mantissa_bits, 0x00400000u);
+static uint32_t EmitClassMaskBits(EmitterState& state, uint32_t bits, uint32_t mask, bool half) {
+	const uint32_t sign_mask     = half ? 0x8000u : 0x80000000u;
+	const uint32_t exponent_mask = half ? 0x7c00u : 0x7f800000u;
+	const uint32_t mantissa_mask = half ? 0x03ffu : 0x007fffffu;
+	const uint32_t quiet_mask    = half ? 0x0200u : 0x00400000u;
+	const auto     sign_bits     = EmitAndConstant(state, bits, sign_mask);
+	const auto     abs_bits      = EmitAndConstant(state, bits, sign_mask - 1u);
+	const auto     exponent_bits = EmitAndConstant(state, abs_bits, exponent_mask);
+	const auto     mantissa_bits = EmitAndConstant(state, abs_bits, mantissa_mask);
+	const auto     quiet_bits    = EmitAndConstant(state, mantissa_bits, quiet_mask);
 
 	const auto sign             = EmitCompareU32Constant(state, spv::OpINotEqual, sign_bits, 0);
 	const auto positive         = EmitLogicalNotBool(state, sign);
 	const auto exponent_zero    = EmitCompareU32Constant(state, spv::OpIEqual, exponent_bits, 0);
 	const auto exponent_nonzero = EmitLogicalNotBool(state, exponent_zero);
 	const auto exponent_inf =
-	    EmitCompareU32Constant(state, spv::OpIEqual, exponent_bits, 0x7f800000u);
+	    EmitCompareU32Constant(state, spv::OpIEqual, exponent_bits, exponent_mask);
 	const auto finite_exponent  = EmitLogicalNotBool(state, exponent_inf);
 	const auto mantissa_zero    = EmitCompareU32Constant(state, spv::OpIEqual, mantissa_bits, 0);
 	const auto mantissa_nonzero = EmitLogicalNotBool(state, mantissa_zero);
@@ -178,6 +158,14 @@ uint32_t EmitClassMaskF32(EmitterState& state, uint32_t value, uint32_t mask) {
 	    EmitClassMaskBitMatch(state, mask, 9, EmitLogicalAndBool(state, inf, positive)));
 }
 
+uint32_t EmitClassMaskF32(EmitterState& state, uint32_t value, uint32_t mask) {
+	return EmitClassMaskBits(state, EmitBitcastF32ToU32(state, value), mask, false);
+}
+
+uint32_t EmitClassMaskF16(EmitterState& state, uint32_t bits, uint32_t mask) {
+	return EmitClassMaskBits(state, bits, mask, true);
+}
+
 uint32_t EmitMinMaxF32Value(EmitterState& state, uint32_t lhs, uint32_t rhs, bool max_value) {
 	const auto lhs_class = EmitClassifyF32(state, lhs);
 	const auto rhs_class = EmitClassifyF32(state, rhs);

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

@@ -59,21 +59,6 @@ uint32_t          ConstantU32(EmitterState& state, uint32_t value);
 	std::abort();
 }
 
-uint32_t ResourceForDescriptor(const EmitterState& state, IR::DescriptorBindingKind kind,
-                               uint32_t resource) {
-	const auto* descriptor = IR::FindBinding(state.program.bindings, kind);
-	if (descriptor == nullptr) {
-		ExitDescriptorBindingFailure(state, kind, resource, "descriptor group was not allocated");
-	}
-	const auto found =
-	    std::find(descriptor->resources.begin(), descriptor->resources.end(), resource);
-	if (found == descriptor->resources.end()) {
-		ExitDescriptorBindingFailure(state, kind, resource,
-		                             "resource is absent from descriptor group");
-	}
-	return static_cast<uint32_t>(found - descriptor->resources.begin());
-}
-
 uint32_t DescriptorElementPointer(EmitterState& state, uint32_t result_ptr_type,
                                   uint32_t variable_id, uint32_t array_index,
                                   IR::DescriptorBindingKind kind, uint32_t resource,
@@ -83,7 +68,7 @@ uint32_t DescriptorElementPointer(EmitterState& state, uint32_t result_ptr_type,
 	}
 	const auto pointer = state.builder.AllocateId();
 	state.builder.AddFunction(spv::OpAccessChain, result_ptr_type, pointer, variable_id,
-	                          ConstantU32(state, array_index));
+	                          array_index);
 	return pointer;
 }
 
@@ -124,13 +109,21 @@ uint32_t ImageType(EmitterState& state, const IR::ImageResource& image) {
 		sampled = 2;
 		if (image.atomic) {
 			EXIT_IF(image.numeric_class != Prospero::TextureNumericClass::Uint);
-			format = spv::ImageFormatR32ui;
+			format = image.atomic64 ? spv::ImageFormatR64ui : spv::ImageFormatR32ui;
+			if (image.atomic64) {
+				state.builder.RequireExtension("SPV_EXT_shader_image_int64");
+				state.builder.RequireCapability(spv::CapabilityInt64);
+				state.builder.RequireCapability(spv::CapabilityInt64Atomics);
+				state.builder.RequireCapability(spv::CapabilityInt64ImageEXT);
+			}
 		}
 	} else {
 		EXIT("invalid image resource class");
 	}
 	const auto& info = ImageDimensionInfoFor(image.dimension);
-	return state.builder.Type(spv::OpTypeImage, ImageScalarType(state, image.numeric_class),
+	const auto  scalar_type =
+	    image.atomic64 ? TypeU64(state) : ImageScalarType(state, image.numeric_class);
+	return state.builder.Type(spv::OpTypeImage, scalar_type,
 	                          info.spirv_dimension, image.depth_compare ? 1u : 0u, info.arrayed,
 	                          info.multisampled, sampled, format);
 }
@@ -144,59 +137,75 @@ uint32_t ImageViewSizeType(EmitterState& state, ImageDimension dimension) {
 	}
 }
 
-uint32_t LoadSampledImageDescriptor(EmitterState& state, uint32_t resource) {
-	const auto& image_resource = state.program.info.images.at(resource);
-	EXIT_IF(image_resource.resource_class != IR::ImageResourceClass::Sampled);
-	const auto kind = IR::DescriptorBindingForImage(image_resource);
-	EXIT_IF(!kind.has_value());
-	const auto array_index  = ResourceForDescriptor(state, *kind, resource);
-	const auto variable     = state.image_variables[IR::ImageBindingIndex(*kind)];
-	const auto pointer_type = state.builder.Type(
-	    spv::OpTypePointer, spv::StorageClassUniformConstant, ImageType(state, image_resource));
-	const auto pointer =
-	    DescriptorElementPointer(state, pointer_type, variable, array_index, *kind, resource,
-	                             "sampled image descriptor array was not emitted");
+namespace {
+void DecorateNonUniformDescriptor(EmitterState& state, uint32_t id) {
+	state.builder.RequireExtension("SPV_EXT_descriptor_indexing");
+	state.builder.RequireCapability(spv::CapabilityShaderNonUniform);
+	state.builder.RequireCapability(spv::CapabilitySampledImageArrayNonUniformIndexing);
+	state.builder.AddAnnotation(spv::OpDecorate, id, spv::DecorationNonUniform);
+}
+} // namespace
+
+uint32_t LoadImageDescriptor(EmitterState& state, uint32_t resource, uint32_t mip,
+                             uint32_t array_index) {
+	const auto pointer = ImageDescriptorPointer(state, resource, mip, array_index);
 	const auto image = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpLoad, ImageType(state, image_resource), image, pointer);
+	const auto kind = *IR::DescriptorBindingForImage(state.program.info.images[resource]);
+	state.builder.AddFunction(spv::OpLoad, state.images[IR::ImageBindingIndex(kind)].type,
+	                          image, pointer);
+	if (array_index != 0u) {
+		DecorateNonUniformDescriptor(state, image);
+	}
 	return image;
 }
 
-uint32_t LoadSamplerDescriptor(EmitterState& state, uint32_t sampler) {
-	const auto array_index =
-	    ResourceForDescriptor(state, IR::DescriptorBindingKind::Samplers, sampler);
-	const auto sampler_type = state.builder.Type(spv::OpTypeSampler);
-	const auto pointer_type =
-	    state.builder.Type(spv::OpTypePointer, spv::StorageClassUniformConstant, sampler_type);
+uint32_t LoadSamplerDescriptor(EmitterState& state, uint32_t sampler, uint32_t array_index) {
+	if (sampler >= state.program.info.samplers.size()) {
+		ExitDescriptorBindingFailure(state, IR::DescriptorBindingKind::Samplers, sampler,
+		                             "sampler resource index is out of range");
+	}
 	const auto pointer = DescriptorElementPointer(
-	    state, pointer_type, state.sampler_variable, array_index,
+	    state, state.sampler_pointer_type, state.sampler_variable,
+	    array_index == 0u ? ConstantU32(state, sampler) : array_index,
 	    IR::DescriptorBindingKind::Samplers, sampler, "sampler descriptor array was not emitted");
 	const auto sampler_id = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpLoad, sampler_type, sampler_id, pointer);
+	state.builder.AddFunction(spv::OpLoad, state.sampler_type, sampler_id, pointer);
+	if (array_index != 0u) DecorateNonUniformDescriptor(state, sampler_id);
 	return sampler_id;
 }
 
-uint32_t MakeSampledImage(EmitterState& state, uint32_t resource, uint32_t sampler) {
+uint32_t MakeSampledImage(EmitterState& state, uint32_t resource, uint32_t sampler_id, uint32_t mip,
+                          uint32_t array_index, bool sampler_dynamic) {
 	const auto& image_resource = state.program.info.images.at(resource);
-	const auto  image          = LoadSampledImageDescriptor(state, resource);
-	const auto  sampler_id     = LoadSamplerDescriptor(state, sampler);
+	EXIT_IF(image_resource.resource_class != IR::ImageResourceClass::Sampled);
+	const auto  image          = LoadImageDescriptor(state, resource, mip, array_index);
 	const auto  sampled_image = state.builder.AllocateId();
-	const auto  sampled_type =
-	    state.builder.Type(spv::OpTypeSampledImage, ImageType(state, image_resource));
+	const auto kind = *IR::DescriptorBindingForImage(image_resource);
+	auto& sampled_type = state.images[IR::ImageBindingIndex(kind)].sampled_type;
+	if (sampled_type == 0) {
+		sampled_type = state.builder.Type(
+		    spv::OpTypeSampledImage, state.images[IR::ImageBindingIndex(kind)].type);
+	}
 	state.builder.AddFunction(spv::OpSampledImage, sampled_type, sampled_image, image, sampler_id);
+	if (array_index != 0u || sampler_dynamic) {
+		DecorateNonUniformDescriptor(state, sampled_image);
+	}
 	return sampled_image;
 }
 
-uint32_t StorageImageDescriptorPointer(EmitterState& state, uint32_t resource) {
+uint32_t ImageDescriptorPointer(EmitterState& state, uint32_t resource, uint32_t mip,
+                                uint32_t array_index) {
 	const auto& image = state.program.info.images.at(resource);
-	EXIT_IF(image.resource_class != IR::ImageResourceClass::Storage);
+	EXIT_IF(mip >= image.mip_count);
 	const auto kind = IR::DescriptorBindingForImage(image);
 	EXIT_IF(!kind.has_value());
-	const auto array_index  = ResourceForDescriptor(state, *kind, resource);
-	const auto pointer_type = state.builder.Type(
-	    spv::OpTypePointer, spv::StorageClassUniformConstant, ImageType(state, image));
-	const auto variable = state.image_variables[IR::ImageBindingIndex(*kind)];
-	return DescriptorElementPointer(state, pointer_type, variable, array_index, *kind, resource,
-	                                "storage image descriptor array was not emitted");
+	if (array_index == 0u) {
+		array_index = ConstantU32(state, image.descriptor_index + mip);
+	}
+	const auto& definition = state.images[IR::ImageBindingIndex(*kind)];
+	return DescriptorElementPointer(state, definition.pointer_type, definition.variable,
+	                                array_index, *kind, resource,
+	                                "image descriptor array was not emitted");
 }
 
 void EmitStorageImageWrite(EmitterState& state, uint32_t resource, uint32_t mip_lod, uint32_t coord,
@@ -206,46 +215,15 @@ void EmitStorageImageWrite(EmitterState& state, uint32_t resource, uint32_t mip_
 	if (!image.atomic) {
 		state.builder.RequireCapability(spv::CapabilityStorageImageWriteWithoutFormat);
 	}
-	const auto kind = IR::DescriptorBindingForImage(image);
-	EXIT_IF(!kind.has_value());
-	const auto array_index = ResourceForDescriptor(state, *kind, resource);
-	const auto image_type  = ImageType(state, image);
-	const auto pointer_type =
-	    state.builder.Type(spv::OpTypePointer, spv::StorageClassUniformConstant, image_type);
-	const auto variable = state.image_variables[IR::ImageBindingIndex(*kind)];
-	const auto LoadAt   = [&](uint32_t array_index) {
-		const auto pointer =
-		    DescriptorElementPointer(state, pointer_type, variable, array_index, *kind, resource,
-		                             "storage image descriptor array was not emitted");
-		const auto descriptor = state.builder.AllocateId();
-		state.builder.AddFunction(spv::OpLoad, image_type, descriptor, pointer);
-		return descriptor;
+	const auto EmitWrite = [&](uint32_t mip) {
+		state.builder.AddFunction(spv::OpImageWrite, LoadImageDescriptor(state, resource, mip),
+		                          coord, texel);
 	};
-	if (image.mip_mode != IR::ImageMipMode::DynamicStorage) {
-		state.builder.AddFunction(spv::OpImageWrite, LoadAt(array_index), coord, texel);
+	if (image.mip_mode != IR::ImageMipMode::Dynamic) {
+		EmitWrite(0);
 		return;
 	}
-	if (image.mip_count == 0u) {
-		ExitDescriptorBindingFailure(state, *kind, resource,
-		                             "dynamic storage image has no mip descriptors");
-	}
-
-	const auto            merge_label = state.builder.AllocateId();
-	std::vector<uint32_t> labels(image.mip_count);
-	std::vector<uint32_t> words {spv::OpSwitch, mip_lod, merge_label};
-	for (uint32_t mip = 0; mip < image.mip_count; mip++) {
-		labels[mip] = state.builder.AllocateId();
-		words.push_back(mip);
-		words.push_back(labels[mip]);
-	}
-	state.builder.AddFunction(spv::OpSelectionMerge, merge_label, spv::SelectionControlMaskNone);
-	state.builder.AddFunction(words);
-	for (uint32_t mip = 0; mip < image.mip_count; mip++) {
-		EmitLabel(state, labels[mip]);
-		state.builder.AddFunction(spv::OpImageWrite, LoadAt(array_index + mip), coord, texel);
-		state.builder.AddFunction(spv::OpBranch, merge_label);
-	}
-	EmitLabel(state, merge_label);
+	EmitIndexSwitch(state, mip_lod, image.mip_count, 0, EmitWrite);
 }
 
 spv::ExecutionModel ExecutionModelForStage(ShaderType stage) {

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

@@ -0,0 +1,355 @@
+#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, TypeU64(s), block, Unary(s, spv::OpUConvert, TypeU64(s), offset));
+	const auto pointer = Unary(s, spv::OpConvertUToPtr, TypePhysicalU32Pointer(s), address);
+	const auto result = s.builder.AllocateId();
+	s.builder.AddFunction(spv::OpLoad, TypeU32(s), result, pointer, spv::MemoryAccessAlignedMask, 4u);
+	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.program.info.bvh) return;
+	const auto u = TypeU32(s), f = TypeF32(s), b = TypeBool(s), wide = TypeU64(s);
+	const auto vec4 = TypeU32Vector(s, 4), vec3 = TypeF32Vector(s, 3);
+	const auto signature = s.builder.Type(spv::OpTypeFunction, vec4, vec4, wide, f, vec3, vec3, vec3);
+	s.bvh_intersect_function = s.builder.AllocateId();
+	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, wide, 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(Unary(s, spv::OpUConvert, u, 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,
+	                          PackU64(s, words[0], and_bits(words[1], 0xff)), ConstantU32(s, 8));
+	const auto index = Binary(s, spv::OpShiftRightLogical, wide, node, ConstantU32(s, 3));
+	const auto last_index = Binary(s, spv::OpIAdd, wide, index,
+	                               Select(s, wide, full, ConstantU64(s, 1), ConstantU64(s, 0)));
+	const auto address = Binary(s, spv::OpIAdd, wide, base,
+	                            Binary(s, spv::OpShiftLeftLogical, wide, index, ConstantU32(s, 6)));
+	const auto last_address =
+	    Binary(s, spv::OpIAdd, wide, base,
+	           Binary(s, spv::OpShiftLeftLogical, wide, last_index, ConstantU32(s, 6)));
+	auto valid = Binary(s, spv::OpIEqual, b, shr(words[3], 28), ConstantU32(s, 8));
+	const auto require = [&](uint32_t condition) {
+		valid = Binary(s, spv::OpLogicalAnd, b, valid, condition);
+	};
+	require(Binary(s, spv::OpULessThanEqual, b, kind, ConstantU32(s, 5)));
+	require(Binary(s, spv::OpULessThanEqual, b, last_index,
+	               PackU64(s, words[2], and_bits(words[3], 0x3ff))));
+	require(Binary(s, spv::OpULessThan, b, last_address, ConstantU64(s, uint64_t {1} << 48)));
+	require(Unary(s, spv::OpLogicalNot, b, Unary(s, spv::OpIsNan, b, extent)));
+	for (uint32_t i = 0; i < 3; ++i) {
+		for (const auto value: {origin[i], direction[i], inverse[i]})
+			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, ConstantU64(s, 0), [&] {
+			return GetBdaPointer(s, Binary(s, spv::OpIAdd, wide, address, ConstantU64(s, 64)));
+		});
+		const auto present = Binary(
+		    s, spv::OpLogicalAnd, b, Binary(s, spv::OpINotEqual, b, first, ConstantU64(s, 0)),
+		    Binary(s, spv::OpLogicalOr, b, Unary(s, spv::OpLogicalNot, b, full),
+		           Binary(s, spv::OpINotEqual, b, second, ConstantU64(s, 0))));
+		return EmitValueOrDefaultIfCondition(s, present, vec4, invalid, [&] {
+			const auto tri_label = s.builder.AllocateId(), box_label = s.builder.AllocateId();
+			const auto merge = s.builder.AllocateId();
+			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));
+	const auto node_words = inst.Flags<uint32_t>();
+	if (ray == nullptr || (node_words != 1u && node_words != 2u) ||
+	    ray->NumArgs() < node_words + 10u || s.bvh_intersect_function == 0)
+		ctx.Fail(inst, "BVH intersection requires its node pointer, ray components and 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);
+	const auto node   = node_words == 2u ? PackU64(s, ctx.Arg(*ray, 0), ctx.Arg(*ray, 1))
+	                                     : Unary(s, spv::OpUConvert, TypeU64(s), ctx.Arg(*ray, 0));
+	const auto extent = Unary(s, spv::OpBitcast, TypeF32(s), ctx.Arg(*ray, node_words));
+	const auto origin = vector(node_words + 1u), direction = vector(node_words + 4u),
+	           inverse = vector(node_words + 7u);
+	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

@@ -10,7 +10,7 @@ namespace {
 
 bool UserDataDwordIndex(const EmitterState& state, IR::ScalarReg reg, uint32_t& dword_index) {
 	const auto register_index = IR::RegIndex(reg);
-	const auto& registers = state.program.bindings.user_data_registers;
+	const auto& registers = state.program.info.user_data_registers;
 	const auto  found     = std::lower_bound(registers.begin(), registers.end(), register_index);
 	if (found == registers.end() || *found != register_index) {
 		return false;
@@ -42,8 +42,9 @@ uint32_t EmitBuiltinU32(EmitterState& state, IR::StageInputKind kind, uint32_t c
 		                  EmitBinaryU32(state, spv::OpIMul, group, ConstantU32(state, size)));
 	}
 	const bool centroid = kind == IR::StageInputKind::BaryCoordSmoothCentroid;
+	const bool sample = kind == IR::StageInputKind::BaryCoordSmoothSample;
 	const auto variable = InputVariableForKind(
-	    state, centroid ? IR::StageInputKind::BaryCoordSmooth : kind);
+	    state, centroid || sample ? IR::StageInputKind::BaryCoordSmooth : kind);
 	if (variable == 0) {
 		return ConstantU32(state, 0);
 	}
@@ -76,15 +77,24 @@ uint32_t EmitBuiltinU32(EmitterState& state, IR::StageInputKind kind, uint32_t c
 		state.builder.AddFunction(spv::OpBitcast, TypeU32(state), bits, value);
 		return bits;
 	}
-	if (centroid || kind == IR::StageInputKind::BaryCoordSmooth ||
+	if (centroid || sample || kind == IR::StageInputKind::BaryCoordSmooth ||
 	    kind == IR::StageInputKind::BaryCoordNoPerspective) {
 		const auto value   = state.builder.AllocateId();
 		const auto bits    = state.builder.AllocateId();
-		if (centroid) {
+		if (centroid || sample) {
 			const auto coordinates = state.builder.AllocateId();
 			state.builder.RequireCapability(spv::CapabilityInterpolationFunction);
-			state.builder.AddFunction(spv::OpExtInst, TypeF32Vector(state, 3), coordinates,
-			                          GlslStd450(state), GLSLstd450InterpolateAtCentroid, variable);
+			if (sample) {
+				const auto sample_id = state.builder.AllocateId();
+				state.builder.AddFunction(spv::OpLoad, TypeI32(state), sample_id,
+				                          InputVariableForKind(state, IR::StageInputKind::SampleId));
+				state.builder.AddFunction(spv::OpExtInst, TypeF32Vector(state, 3), coordinates,
+				                          GlslStd450(state), GLSLstd450InterpolateAtSample,
+				                          variable, sample_id);
+			} else {
+				state.builder.AddFunction(spv::OpExtInst, TypeF32Vector(state, 3), coordinates,
+				                          GlslStd450(state), GLSLstd450InterpolateAtCentroid, variable);
+			}
 			state.builder.AddFunction(spv::OpCompositeExtract, TypeF32(state), value,
 			                          coordinates, component + 1u);
 		} else {
@@ -167,6 +177,13 @@ uint32_t EmitAttribute(EmitterState& state, uint32_t attr, uint32_t chan) {
 		return value;
 	};
 	if (input->per_vertex) {
+		if (PixelParameterIsFlat(state, attr)) {
+			const auto value = load_per_vertex(
+			    state.input_info.pixel->parameter_mode == ShaderPixelParameterMode::LastVertex ? 2u : 0u);
+			const auto bits  = state.builder.AllocateId();
+			state.builder.AddFunction(spv::OpBitcast, TypeU32(state), bits, value);
+			return bits;
+		}
 		const auto barycentric_kind = state.input_info.pixel->ps_no_perspective
 		                                  ? IR::StageInputKind::BaryCoordNoPerspective
 		                                  : IR::StageInputKind::BaryCoordSmooth;
@@ -221,8 +238,12 @@ uint32_t EmitInterpolationParameter(ValueEmitContext& ctx, uint32_t attr, uint32
 		return value;
 	};
 
-	const auto selected_vertex = (mode + 1u) % 3u;
-	uint32_t   value           = load_vertex(selected_vertex);
+	uint32_t selected_vertex = (mode + 1u) % 3u;
+	if (mode == 2u && PixelParameterIsFlat(state, attr)) {
+		selected_vertex =
+		    state.input_info.pixel->parameter_mode == ShaderPixelParameterMode::LastVertex ? 2u : 0u;
+	}
+	uint32_t value = load_vertex(selected_vertex);
 	if (!PixelParameterIsCustom(state, attr) && mode < 2u) {
 		const auto delta = state.builder.AllocateId();
 		state.builder.AddFunction(spv::OpFSub, TypeF32(state), delta, value, load_vertex(0));
@@ -252,8 +273,12 @@ uint32_t ExportVector(ValueEmitContext& ctx, uint32_t data, const IR::ExportInfo
                       bool uint_output) {
 	auto& state = ctx.state;
 	if (exp.compr && !uint_output) {
-		const auto unpack =
-		    MrtOutputMode(state, exp) == 5u ? GLSLstd450UnpackUnorm2x16 : GLSLstd450UnpackHalf2x16;
+		auto unpack = GLSLstd450UnpackHalf2x16;
+		switch (MrtOutputMode(state, exp)) {
+			case 5: unpack = GLSLstd450UnpackUnorm2x16; break;
+			case 6: unpack = GLSLstd450UnpackSnorm2x16; break;
+			default: break;
+		}
 		uint32_t f32[4] = {ConstantF32(state, 0), ConstantF32(state, 0), ConstantF32(state, 0),
 		                   ConstantF32(state, 0x3f800000u)};
 		for (uint32_t pair = 0; pair < 2u; pair++) {
@@ -279,50 +304,46 @@ uint32_t ExportVector(ValueEmitContext& ctx, uint32_t data, const IR::ExportInfo
 		                          f32[0], f32[1], f32[2], f32[3]);
 		return vector;
 	}
-	uint32_t raw[4] = {
-	    ConstantU32(state, 0),
-	    ConstantU32(state, 0),
-	    ConstantU32(state, 0),
-	    ConstantU32(state, uint_output ? 1u : 0x3f800000u),
-	};
-	if (exp.compr) {
-		for (uint32_t pair = 0; pair < 2u; pair++) {
-			if ((exp.en & (3u << (pair * 2u))) == 0u) {
-				continue;
-			}
-			const auto packed = ExportRawComponent(ctx, data, pair);
-			for (uint32_t lane = 0; lane < 2u; lane++) {
-				const auto component = pair * 2u + lane;
-				if (((exp.en >> component) & 1u) == 0u) {
+	if (exp.compr || exp.en != 0xfu) {
+		uint32_t raw[4] = {
+		    ConstantU32(state, 0),
+		    ConstantU32(state, 0),
+		    ConstantU32(state, 0),
+		    ConstantU32(state, uint_output ? 1u : 0x3f800000u),
+		};
+		if (exp.compr) {
+			for (uint32_t pair = 0; pair < 2u; pair++) {
+				if ((exp.en & (3u << (pair * 2u))) == 0u) {
 					continue;
 				}
-				raw[component] = state.builder.AllocateId();
-				state.builder.AddFunction(spv::OpBitFieldUExtract, TypeU32(state), raw[component],
-				                          packed, ConstantU32(state, lane * 16u),
-				                          ConstantU32(state, 16));
+				const auto packed = ExportRawComponent(ctx, data, pair);
+				for (uint32_t lane = 0; lane < 2u; lane++) {
+					const auto component = pair * 2u + lane;
+					if (((exp.en >> component) & 1u) == 0u) {
+						continue;
+					}
+					raw[component] = state.builder.AllocateId();
+					state.builder.AddFunction(spv::OpBitFieldUExtract, TypeU32(state), raw[component],
+					                          packed, ConstantU32(state, lane * 16u),
+					                          ConstantU32(state, 16));
+				}
 			}
-		}
-	} else {
-		for (uint32_t component = 0; component < 4u; component++) {
-			if (((exp.en >> component) & 1u) != 0u) {
-				raw[component] = ExportRawComponent(ctx, data, component);
+		} else {
+			for (uint32_t component = 0; component < 4u; component++) {
+				if (((exp.en >> component) & 1u) != 0u) {
+					raw[component] = ExportRawComponent(ctx, data, component);
+				}
 			}
 		}
+		data = state.builder.AllocateId();
+		state.builder.AddFunction(spv::OpCompositeConstruct, TypeU32Vector(state, 4), data,
+		                          raw[0], raw[1], raw[2], raw[3]);
 	}
 	if (uint_output) {
-		const auto vector = state.builder.AllocateId();
-		state.builder.AddFunction(spv::OpCompositeConstruct, TypeU32Vector(state, 4), vector,
-		                          raw[0], raw[1], raw[2], raw[3]);
-		return vector;
-	}
-	uint32_t f32[4] {};
-	for (uint32_t component = 0; component < 4u; component++) {
-		f32[component] = state.builder.AllocateId();
-		state.builder.AddFunction(spv::OpBitcast, TypeF32(state), f32[component], raw[component]);
+		return data;
 	}
 	const auto vector = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpCompositeConstruct, TypeF32Vector(state, 4), vector, f32[0],
-	                          f32[1], f32[2], f32[3]);
+	state.builder.AddFunction(spv::OpBitcast, TypeF32Vector(state, 4), vector, data);
 	return vector;
 }
 
@@ -366,18 +387,23 @@ void EmitAuxPositionExport(ValueEmitContext& ctx, uint32_t data, const IR::Expor
 			state.builder.AddFunction(spv::OpStore, state.point_size_variable, f32);
 			continue;
 		}
-		auto StoreDistance = [&](uint32_t variable, uint32_t index) {
+		auto StoreDistance = [&](IR::StageOutputKind kind, uint32_t variable, uint32_t index) {
 			if (index == UINT32_MAX) {
 				return;
 			}
-			const auto pointer = state.builder.AllocateId();
-			state.builder.AddFunction(spv::OpAccessChain,
-			                          TypePointer(state, spv::StorageClassOutput, TypeF32(state)),
-			                          pointer, variable, ConstantU32(state, index));
+			uint32_t pointer;
+			if (state.program.stage == ShaderType::Mesh) {
+				pointer = MeshOutputPointer(state, kind, index);
+			} else {
+				pointer = state.builder.AllocateId();
+				state.builder.AddFunction(spv::OpAccessChain,
+				                          TypePointer(state, spv::StorageClassOutput, TypeF32(state)),
+				                          pointer, variable, ConstantU32(state, index));
+			}
 			state.builder.AddFunction(spv::OpStore, pointer, f32);
 		};
-		StoreDistance(state.clip_distance_variable, output.clip_distance);
-		StoreDistance(state.cull_distance_variable, output.cull_distance);
+		StoreDistance(IR::StageOutputKind::ClipDistance, state.clip_distance_variable, output.clip_distance);
+		StoreDistance(IR::StageOutputKind::CullDistance, state.cull_distance_variable, output.cull_distance);
 	}
 }
 
@@ -425,16 +451,16 @@ uint32_t EmitWqmU64(EmitterState& state, uint32_t value) {
 	const auto quad_bits   = state.builder.AllocateId();
 	const auto result      = state.builder.AllocateId();
 	state.builder.AddFunction(spv::OpShiftRightLogical, TypeU64(state), shifted_one, value,
-	                          ConstantU64(state, 0x0000000100000001ull));
+	                          ConstantU32(state, 1));
 	state.builder.AddFunction(spv::OpBitwiseOr, TypeU64(state), merged_one, value, shifted_one);
 	state.builder.AddFunction(spv::OpShiftRightLogical, TypeU64(state), shifted_two, merged_one,
-	                          ConstantU64(state, 0x0000000200000002ull));
+	                          ConstantU32(state, 2));
 	state.builder.AddFunction(spv::OpBitwiseOr, TypeU64(state), merged_two, merged_one,
 	                          shifted_two);
 	state.builder.AddFunction(spv::OpBitwiseAnd, TypeU64(state), quad_bits, merged_two,
 	                          ConstantU64(state, 0x1111111111111111ull));
 	state.builder.AddFunction(spv::OpIMul, TypeU64(state), result, quad_bits,
-	                          ConstantU64(state, 0x0000000f0000000full));
+	                          ConstantU64(state, 15));
 	return result;
 }
 
@@ -442,7 +468,7 @@ void EmitSetAttribute(ValueEmitContext& ctx, const IR::Inst& inst) {
 	auto&       state = ctx.state;
 	const auto& exp   = ctx.Export(inst);
 	const auto  exec  = ctx.Arg(inst, 1);
-	if (state.program.stage == ShaderType::Pixel && exp.vm && state.requirements.pixel_valid_mask &&
+	if (state.program.stage == ShaderType::Pixel && exp.vm && state.program.info.pixel_valid_mask &&
 	    state.pixel_valid_mask_variable != 0) {
 		const auto value = state.builder.AllocateId();
 		state.builder.AddFunction(spv::OpSelect, TypeU32(state), value, exec, ConstantU32(state, 1),
@@ -452,6 +478,11 @@ 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 != ShaderAlphaBlendSource::None) {
+		return;
+	}
 	EmitIfCondition(state, exec, [&]() {
 		const auto data = ctx.Arg(inst, 0);
 		if (exp.kind == IR::ExportTargetKind::Primitive) {
@@ -492,6 +523,40 @@ 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 != ShaderAlphaBlendSource::None) {
+			// Broadcast logical alpha before swizzling the primary output.
+			const auto blend_output =
+			    OutputVariableForExport(state, {.kind = IR::ExportTargetKind::Mrt, .index = 1});
+			if (blend_output != 0) {
+				auto alpha = state.builder.AllocateId();
+				state.builder.AddFunction(spv::OpVectorShuffle, vector_type, alpha, value, value,
+				                          3u, 3u, 3u, 3u);
+				uint32_t alpha_factor = 0;
+				switch (state.input_info.pixel->alpha_blend_source) {
+					case ShaderAlphaBlendSource::SourceAlphaOne:
+						alpha_factor = ConstantF32Value(state, 1.0f);
+						break;
+					case ShaderAlphaBlendSource::SourceAlphaZero:
+						alpha_factor = ConstantF32Value(state, 0.0f);
+						break;
+					default: break;
+				}
+				if (alpha_factor != 0) {
+					const auto mapping = state.input_info.pixel->target_export_mapping[0];
+					for (uint32_t component = 0; component < 4; component++) {
+						if (mapping.Map(component) != 3u) continue;
+						const auto factors = state.builder.AllocateId();
+						state.builder.AddFunction(spv::OpCompositeInsert, vector_type, factors,
+						                          alpha_factor, alpha, component);
+						alpha = factors;
+						break;
+					}
+				}
+				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];
@@ -557,6 +622,14 @@ uint32_t EmitIdentity(ValueEmitContext&, uint32_t value) {
 
 void EmitVoid(ValueEmitContext&) {}
 
+void EmitStoreCompletion(EmitterState& state) {
+	state.builder.AddFunction(
+	    spv::OpControlBarrier, ConstantU32(state, spv::ScopeSubgroup),
+	    ConstantU32(state, spv::ScopeDevice),
+	    ConstantU32(state, spv::MemorySemanticsReleaseMask | spv::MemorySemanticsUniformMemoryMask |
+	                          spv::MemorySemanticsImageMemoryMask));
+}
+
 void EmitBarrier(EmitterState& state) {
 	const auto tessellation = state.program.stage == ShaderType::TessellationControl;
 	if (!tessellation && ShaderWorkgroupInput(state.program.stage, state.input_info) == nullptr) {
@@ -564,9 +637,11 @@ void EmitBarrier(EmitterState& state) {
 		return;
 	}
 	const auto memory_scope = tessellation ? spv::ScopeInvocation : spv::ScopeWorkgroup;
-	const auto semantics    = tessellation ? spv::MemorySemanticsMaskNone
-	                                       : spv::MemorySemanticsAcquireReleaseMask |
-	                                             spv::MemorySemanticsWorkgroupMemoryMask;
+	const auto memory = state.lds_storage_class == spv::StorageClassStorageBuffer
+	                        ? spv::MemorySemanticsUniformMemoryMask
+	                        : spv::MemorySemanticsWorkgroupMemoryMask;
+	const auto semantics = tessellation ? spv::MemorySemanticsMaskNone
+	                                    : spv::MemorySemanticsAcquireReleaseMask | memory;
 	state.builder.AddFunction(spv::OpControlBarrier, ConstantU32(state, spv::ScopeWorkgroup),
 	                          ConstantU32(state, memory_scope), ConstantU32(state, semantics));
 }
@@ -606,6 +681,14 @@ uint32_t EmitGetBuiltin(ValueEmitContext& ctx, IR::Value kind, IR::Value index)
 	return EmitBuiltinU32(ctx.state, static_cast<IR::StageInputKind>(kind.U32()), index.U32());
 }
 
+uint32_t EmitGetDispatchThreadExtent(ValueEmitContext& ctx, const IR::Inst& inst) {
+	const auto start = ctx.state.program.bindings.dispatch_thread_dword;
+	if (start == IR::PushData::NoStart || !inst.Arg(0).IsImmediate() || inst.Arg(0).U32() >= 3u) {
+		ctx.Fail(inst, "invalid dispatch thread extent");
+	}
+	return EmitShaderDataDwordLoad(ctx.state, start + inst.Arg(0).U32());
+}
+
 uint32_t EmitUndefU1(EmitterState& state, const IR::Inst& inst) {
 	const auto result = state.builder.AllocateId();
 	state.builder.AddFunction(spv::OpUndef, TypeId(state, inst.GetType()), result);
@@ -637,26 +720,18 @@ uint32_t EmitDppUpdateU32(ValueEmitContext& ctx, const IR::Inst& inst) {
 }
 
 uint32_t EmitConditionRef(ValueEmitContext& ctx, const IR::Inst& inst) {
-	if (ctx.other_half == nullptr) return ctx.Arg(inst, 0);
-	// A native scalar branch makes one decision for both emulated wave halves.
-	if (ctx.half != 0) return ctx.other_half->Def(IR::Value(&inst));
+	// Scalar branches include lanes disabled by EXEC, which the branch body may restore.
+	if (ctx.other_half != nullptr && ctx.half != 0) return ctx.other_half->Def(IR::Value(&inst));
 	const auto kind = inst.Flags<CFG::BranchCondition>();
 	if (kind == CFG::BranchCondition::ScalarInstruction) return ctx.Arg(inst, 0);
-	const auto ballot = ctx.Ballot(inst.Arg(0));
-	const auto low = ctx.state.builder.AllocateId();
-	const auto high = ctx.state.builder.AllocateId();
-	const auto combined = ctx.state.builder.AllocateId();
-	const auto result = ctx.state.builder.AllocateId();
-	ctx.state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(ctx.state), low, ballot, 0);
-	ctx.state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(ctx.state), high, ballot, 1);
 	const bool zero = kind == CFG::BranchCondition::ExecZero ||
 	                  kind == CFG::BranchCondition::VccZero || kind == CFG::BranchCondition::SccZero;
-	ctx.state.builder.AddFunction(zero ? spv::OpBitwiseAnd : spv::OpBitwiseOr,
-	                              TypeU32(ctx.state), combined, low, high);
-	ctx.state.builder.AddFunction(zero ? spv::OpIEqual : spv::OpINotEqual,
-	                              TypeBool(ctx.state), result, combined,
-	                              ConstantU32(ctx.state, zero ? ~0u : 0u));
-	return result;
+	const auto ballot = ctx.Ballot(inst.Arg(0));
+	const auto reference = zero ? ctx.Ballot(IR::Value(true))
+	                            : ConstantU32CompositeZero(ctx.state, 4);
+	const auto compare = Binary(ctx.state, zero ? spv::OpIEqual : spv::OpINotEqual,
+	                            TypeBoolVector(ctx.state, 4), ballot, reference);
+	return Unary(ctx.state, zero ? spv::OpAll : spv::OpAny, TypeBool(ctx.state), compare);
 }
 
 uint32_t EmitBallot(ValueEmitContext& ctx, IR::Value predicate) {

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

@@ -3,9 +3,7 @@
 #include "graphics/shader/recompiler/frontend/decode/ImageOps.h"
 
 #include <algorithm>
-#include <atomic>
 #include <bit>
-#include <cstdio>
 
 namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter {
 namespace {
@@ -86,6 +84,46 @@ uint32_t ZeroF32(EmitterState& state) {
 	return ConstantF32(state, 0);
 }
 
+uint32_t GatherMip(ValueEmitContext& ctx, const IR::Inst& inst, const IR::MemoryInfo& mem,
+                   const IR::Inst& address, const ImageSampleLayout& layout, uint32_t mip_count) {
+	auto& state = ctx.state;
+	const auto Extract = [&](uint32_t word, uint32_t first, uint32_t count) {
+		return EmitBitFieldUExtract(state, word, ConstantU32(state, first),
+		                           ConstantU32(state, count));
+	};
+	const auto LodFixed = [&](uint32_t word, uint32_t first) {
+		return Binary(state, spv::OpFMul, TypeF32(state),
+		              Unary(state, spv::OpConvertUToF, TypeF32(state), Extract(word, first, 12)),
+		              ConstantF32Value(state, 1.0f / 256.0f));
+	};
+	const auto sampler_control = ctx.Arg(inst, 6);
+	const auto preclamp = Binary(state, spv::OpINotEqual, TypeBool(state),
+	                            Extract(sampler_control, 28, 1), ConstantU32(state, 0));
+	const auto half = ConstantF32Value(state, 0.5f);
+	const auto zero = ZeroF32(state);
+	const auto base = Unary(state, spv::OpConvertUToF, TypeF32(state),
+	                        Extract(ctx.Arg(inst, 4), 12, 4));
+	const auto minimum = Binary(state, spv::OpFSub, TypeF32(state),
+	                            LodFixed(ctx.Arg(inst, 3), 8), base);
+	const auto last = ConstantF32Value(state, static_cast<float>(mip_count - 1u));
+	auto lod = Binary(state, spv::OpFAdd, TypeF32(state),
+	                  AddressF32(ctx, mem, address, layout.lod),
+	                  Select(state, TypeF32(state), preclamp, half, zero));
+	lod = EmitGlsl<GLSLstd450FClamp, IR::Type::F32>(
+	    state, lod, LodFixed(ctx.Arg(inst, 5), 0), LodFixed(ctx.Arg(inst, 5), 12));
+	// T# minimum LOD is in physical mip space and applies after the S# view-relative clamp.
+	lod = EmitGlsl<GLSLstd450FMax, IR::Type::F32>(state, lod, minimum);
+	lod = EmitGlsl<GLSLstd450FClamp, IR::Type::F32>(state, lod, zero, last);
+	lod = Binary(state, spv::OpFAdd, TypeF32(state), lod,
+	             Select(state, TypeF32(state), preclamp, zero, half));
+	// The clamped value is nonnegative, so conversion performs the integer mip rounding.
+	const auto mip = EmitUMin32(state, Unary(state, spv::OpConvertFToU, TypeU32(state), lod),
+	                           ConstantU32(state, mip_count - 1u));
+	const auto mip_none = Binary(state, spv::OpIEqual, TypeBool(state),
+	                            Extract(sampler_control, 26, 2), ConstantU32(state, 0));
+	return Select(state, TypeU32(state), mip_none, ConstantU32(state, 0), mip);
+}
+
 uint32_t CubeAxis(EmitterState& state, uint32_t value) {
 	return Binary(state, spv::OpFSub, TypeF32(state), value, ConstantF32(state, 0x3f800000u));
 }
@@ -252,7 +290,7 @@ uint32_t QueryDimensions(ValueEmitContext& ctx, const IR::MemoryInfo& mem,
                          const IR::Inst& address) {
 	const auto  dimension = ctx.state.program.info.images.at(mem.resource).dimension;
 	const auto& info      = ImageDimensionInfoFor(dimension);
-	const auto  image     = LoadSampledImageDescriptor(ctx.state, mem.resource);
+	const auto  image     = LoadImageDescriptor(ctx.state, mem.resource);
 	const auto  size      = ctx.state.builder.AllocateId();
 	if (info.multisampled != 0u) {
 		ctx.state.builder.AddFunction(spv::OpImageQuerySize,
@@ -353,28 +391,50 @@ Format::BufferFormatInfo ImageConversionFormat(const EmitterState&   state,
 	const auto format = state.program.info.images[mem.resource].conversion_format;
 	if (format == Prospero::BufferFormat::kInvalid) return {};
 	const auto info = Format::GetFormatInfo(format);
-	EXIT_IF(Prospero::SampledTextureNumericClass(format) != Prospero::TextureNumericClass::Uint ||
-	        Prospero::RemapTextureFormat(format) == format ||
-	        info.type != Format::ComponentType::Uint || !info.packed_bitfield ||
-	        info.byte_size != sizeof(uint32_t) || info.component_count == 0u ||
+	EXIT_IF(Prospero::RemapTextureFormat(format) == format || info.component_count == 0u ||
 	        info.component_count > 4u);
+	EXIT_IF(info.type != Format::ComponentType::Uscaled &&
+	        (info.type != Format::ComponentType::Uint || !info.packed_bitfield ||
+	         info.byte_size != sizeof(uint32_t)));
 	return info;
 }
 
+uint32_t ImageGatherSource(const EmitterState& state, const IR::MemoryInfo& mem) {
+	const auto component = ImageGatherComponent(mem.dmask);
+	const auto info      = ImageConversionFormat(state, mem);
+	if (info.format == Prospero::BufferFormat::kInvalid) return component;
+	if (info.packed_bitfield) return 0u;
+	const auto selector =
+	    (state.program.info.images[mem.resource].shader_swizzle >> (component * 3u)) & 7u;
+	return selector >= 4u ? (selector - 4u) % info.component_count : 0u;
+}
+
 uint32_t UnpackImageTexel(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uint32_t texel) {
 	const auto info = ImageConversionFormat(ctx.state, mem);
 	if (info.format == Prospero::BufferFormat::kInvalid) return texel;
 
-	const auto packed = ctx.state.builder.AllocateId();
-	ctx.state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(ctx.state), packed, texel, 0u);
-	uint32_t components[4] = {ConstantU32(ctx.state, 0), ConstantU32(ctx.state, 0),
-	                          ConstantU32(ctx.state, 0), ConstantU32(ctx.state, 0)};
+	const auto numeric_class = ctx.state.program.info.images[mem.resource].numeric_class;
+	const auto scalar_type   = ImageScalarType(ctx.state, numeric_class);
+	uint32_t   packed        = 0;
+	if (info.packed_bitfield) {
+		packed = ctx.state.builder.AllocateId();
+		ctx.state.builder.AddFunction(spv::OpCompositeExtract, scalar_type, packed, texel, 0u);
+	}
+	uint32_t components[4] {};
 	for (uint32_t component = 0; component < info.component_count; component++) {
 		components[component] = ctx.state.builder.AllocateId();
-		ctx.state.builder.AddFunction(spv::OpBitFieldUExtract, TypeU32(ctx.state),
-		                              components[component], packed,
-		                              ConstantU32(ctx.state, info.component_bit_offset[component]),
-		                              ConstantU32(ctx.state, info.component_bits[component]));
+		if (info.packed_bitfield) {
+			ctx.state.builder.AddFunction(spv::OpBitFieldUExtract, scalar_type,
+			                              components[component], packed,
+			                              ConstantU32(ctx.state, info.component_bit_offset[component]),
+			                              ConstantU32(ctx.state, info.component_bits[component]));
+		} else {
+			ctx.state.builder.AddFunction(spv::OpCompositeExtract, scalar_type,
+			                              components[component], texel, component);
+			// UNorm backing preserves the guest's filtering; scaling precedes swizzle constants.
+			components[component] = Binary(ctx.state, spv::OpFMul, scalar_type, components[component],
+			                               ConstantF32Value(ctx.state, 255.0f));
+		}
 	}
 	for (uint32_t component = info.component_count; component < 4u; component++) {
 		components[component] = components[component % info.component_count];
@@ -385,16 +445,19 @@ uint32_t UnpackImageTexel(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uint
 	for (uint32_t component = 0; component < 4u; component++) {
 		const auto selector = (swizzle >> (component * 3u)) & 7u;
 		if (selector == 1u) {
-			selected[component] = ConstantU32(ctx.state, 1u);
+			selected[component] = info.packed_bitfield ? ConstantU32(ctx.state, 1u)
+			                                          : ConstantF32Value(ctx.state, 1.0f);
 		} else if (selector >= 4u) {
 			selected[component] = components[selector - 4u];
 		} else {
-			selected[component] = ConstantU32(ctx.state, 0u);
+			selected[component] = info.packed_bitfield ? ConstantU32(ctx.state, 0u)
+			                                          : ZeroF32(ctx.state);
 		}
 	}
 	const auto result = ctx.state.builder.AllocateId();
-	ctx.state.builder.AddFunction(spv::OpCompositeConstruct, TypeU32Vector(ctx.state, 4), result,
-	                              selected[0], selected[1], selected[2], selected[3]);
+	ctx.state.builder.AddFunction(spv::OpCompositeConstruct,
+	                              ImageVectorType(ctx.state, numeric_class, 4), result, selected[0],
+	                              selected[1], selected[2], selected[3]);
 	return result;
 }
 
@@ -405,10 +468,24 @@ uint32_t UnpackImageGather(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uin
 	const auto component = ImageGatherComponent(mem.dmask);
 	const auto selector =
 	    (ctx.state.program.info.images[mem.resource].shader_swizzle >> (component * 3u)) & 7u;
+	const auto numeric_class = ctx.state.program.info.images[mem.resource].numeric_class;
+	const auto vector_type   = ImageVectorType(ctx.state, numeric_class, 4);
 	if (selector < 4u) {
-		const auto value = ConstantU32(ctx.state, selector == 1u ? 1u : 0u);
-		return ctx.state.builder.Constant(spv::OpConstantComposite, TypeU32Vector(ctx.state, 4),
-		                                  value, value, value, value);
+		uint32_t value;
+		switch (info.type) {
+			case Format::ComponentType::Uscaled:
+				value = ConstantF32Value(ctx.state, selector == 1u ? 1.0f : 0.0f);
+				break;
+			default: value = ConstantU32(ctx.state, selector == 1u ? 1u : 0u); break;
+		}
+		return ctx.state.builder.Constant(spv::OpConstantComposite, vector_type, value, value,
+		                                  value, value);
+	}
+	if (!info.packed_bitfield) {
+		const auto scale  = ConstantF32Value(ctx.state, 255.0f);
+		const auto scales = ctx.state.builder.Constant(spv::OpConstantComposite, vector_type,
+		                                                scale, scale, scale, scale);
+		return Binary(ctx.state, spv::OpFMul, vector_type, gathered, scales);
 	}
 
 	uint32_t values[4] {};
@@ -433,7 +510,7 @@ uint32_t EmitOneDimensionalGatherLz(ValueEmitContext& ctx, const IR::MemoryInfo&
                                     uint32_t coord, Prospero::TextureNumericClass numeric_class) {
 	auto& state = ctx.state;
 	state.builder.RequireCapability(spv::CapabilityImageQuery);
-	const auto image = LoadSampledImageDescriptor(state, mem.resource);
+	const auto image = LoadImageDescriptor(state, mem.resource);
 	const auto width = state.builder.AllocateId();
 	state.builder.AddFunction(spv::OpImageQuerySizeLod, TypeU32(state), width, image,
 	                          ConstantU32(state, 0));
@@ -446,13 +523,11 @@ uint32_t EmitOneDimensionalGatherLz(ValueEmitContext& ctx, const IR::MemoryInfo&
 	                                 Binary(state, spv::OpFMul, TypeF32(state), coord, width_f32),
 	                                 ConstantF32(state, 0x3f000000u)));
 
-	const auto sampled     = MakeSampledImage(state, mem.resource, mem.sampler);
+	const auto sampled = MakeSampledImage(state, mem.resource,
+	                                     LoadSamplerDescriptor(state, mem.sampler));
 	const auto vector_type = ImageVectorType(state, numeric_class, 4);
 	const auto scalar_type = ImageScalarType(state, numeric_class);
-	const auto component =
-	    ImageConversionFormat(state, mem).format == Prospero::BufferFormat::kInvalid
-	        ? ImageGatherComponent(mem.dmask)
-	        : 0u;
+	const auto component = ImageGatherSource(state, mem);
 	uint32_t values[2] {};
 	for (uint32_t index = 0; index < 2u; index++) {
 		const auto sample_coord =
@@ -476,6 +551,7 @@ uint32_t EmitOneDimensionalGatherLz(ValueEmitContext& ctx, const IR::MemoryInfo&
 uint32_t PackImageTexel(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uint32_t texel) {
 	const auto info = ImageConversionFormat(ctx.state, mem);
 	if (info.format == Prospero::BufferFormat::kInvalid) return texel;
+	EXIT_IF(info.type != Format::ComponentType::Uint || !info.packed_bitfield);
 
 	auto packed = ConstantU32(ctx.state, 0u);
 	for (uint32_t component = 0; component < info.component_count; component++) {
@@ -539,17 +615,88 @@ 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;
+template <typename Emit>
+uint32_t EmitImageAccess(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t result_type,
+                         const Emit& emit) {
+	auto& state = ctx.state;
+	const auto& mem = ctx.Memory(inst);
+	const auto& image = state.program.info.images.at(mem.resource);
+	if (image.indirect_root != mem.resource) return emit(mem.resource, 0u);
+	const auto* handle = inst.Arg(0).ResolveInstruction();
+	const auto* source = image.source < state.program.descriptor_sources.size()
+	                         ? &state.program.descriptor_sources[image.source]
+	                         : nullptr;
+	if (handle == nullptr || source == nullptr || !source->indirect_descriptor.has_value() ||
+	    handle->NumArgs() == 0u) {
+		ctx.Fail(inst, "has invalid indirect image key provenance");
 	}
+	const auto key = ctx.Def(handle->Arg(0));
+	if (state.flattened_srt_variable == 0 ||
+	    image.indirect_resources.size() < 2u) {
+		ctx.Fail(inst, "has no indirect image runtime mapping");
+	}
+	const auto selected = EmitIndirectResourceIndex(
+	    state, key, image.indirect_mapping_offset, image.indirect_search_iterations, 0u);
+	struct ImageRun {
+		uint32_t first;
+		uint32_t count;
+		uint32_t resource;
+		uint32_t slot_bias;
+	};
+	std::vector<ImageRun> runs;
+	for (uint32_t ordinal = 0; ordinal < image.indirect_resources.size(); ++ordinal) {
+		const auto resource = image.indirect_resources[ordinal];
+		const auto& candidate = state.program.info.images[resource];
+		const auto kind = *IR::DescriptorBindingForImage(candidate);
+		if (!runs.empty()) {
+			auto& run = runs.back();
+			const auto& first = state.program.info.images[run.resource];
+			const auto next_slot = run.slot_bias + ordinal;
+			if (candidate.dimension == first.dimension && candidate.cube == first.cube &&
+			    IR::DescriptorBindingForImage(first) == kind &&
+			    candidate.mip_count == 1u && first.mip_count == 1u &&
+			    (ordinal == 1u || candidate.descriptor_index == next_slot)) {
+				// Native roots precede appended children; only the root may have a slot gap.
+				if (ordinal == 1u)
+					run.slot_bias = candidate.descriptor_index - ordinal;
+				++run.count;
+				continue;
+			}
+		}
+		runs.push_back({ordinal, 1u, resource,
+		                candidate.descriptor_index - ordinal});
+	}
+	const auto EmitRun = [&](uint32_t index) {
+		const auto& run = runs[index];
+		if (run.count == 1u) return emit(run.resource, 0u);
+		auto slot = Binary(state, spv::OpIAdd, TypeU32(state), selected,
+		                   ConstantU32(state, run.slot_bias));
+		if (run.first == 0u) {
+			const auto root_slot = image.descriptor_index;
+			if (root_slot != run.slot_bias) {
+				const auto is_root = Binary(state, spv::OpIEqual, TypeBool(state), selected,
+				                            ConstantU32(state, 0u));
+				const auto root_or_child = state.builder.AllocateId();
+				state.builder.AddFunction(spv::OpSelect, TypeU32(state), root_or_child,
+				                          is_root, ConstantU32(state, root_slot), slot);
+				slot = root_or_child;
+			}
+		}
+		return emit(run.resource, slot);
+	};
+	auto run_index = ConstantU32(state, 0u);
+	for (uint32_t index = 1; index < runs.size(); ++index) {
+		const auto at_or_after = Binary(state, spv::OpUGreaterThanEqual, TypeBool(state),
+		                               selected, ConstantU32(state, runs[index].first));
+		const auto next = state.builder.AllocateId();
+		state.builder.AddFunction(spv::OpSelect, TypeU32(state), next, at_or_after,
+		                          ConstantU32(state, index), run_index);
+		run_index = next;
+	}
+	return runs.size() == 1u
+	           ? EmitRun(0u)
+	           : EmitIndexSwitch(state, run_index, static_cast<uint32_t>(runs.size()),
+	                             result_type, EmitRun);
 }
 
 } // namespace
@@ -571,7 +718,8 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 	if (op == IR::ValueOpcode::ImageQueryLod) {
 		state.builder.RequireCapability(spv::CapabilityImageQuery);
 		const auto dimension = image.dimension;
-		const auto sampled   = MakeSampledImage(state, mem.resource, mem.sampler);
+		const auto sampled = MakeSampledImage(state, mem.resource,
+		                                     LoadSamplerDescriptor(state, mem.sampler));
 		const auto lod       = state.builder.AllocateId();
 		state.builder.AddFunction(
 		    spv::OpImageQueryLod, TypeF32Vector(state, 2), lod, sampled,
@@ -592,29 +740,35 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 		return;
 	}
 	if (op == IR::ValueOpcode::ImageRead) {
-		const auto  dimension      = image.dimension;
-		const auto& dimension_info = ImageDimensionInfoFor(dimension);
-		const auto  numeric_class  = image.numeric_class;
-		const auto  condition      = ctx.Arg(inst, 2);
+		const auto numeric_class = image.numeric_class;
+		const auto condition = ctx.Arg(inst, 2);
 		ctx.Define(
 		    inst,
 		    EmitValueOrDefaultIfCondition(
 		        state, condition, TypeU32Vector(state, 4), ConstantU32CompositeZero(state, 4),
 		        [&]() {
-			        const auto descriptor = LoadSampledImageDescriptor(state, mem.resource);
-			        const auto color      = state.builder.AllocateId();
-			        const auto coord      = CoordU32(ctx, mem, *address, dimension);
-			        if (dimension_info.multisampled != 0u) {
-				        state.builder.AddFunction(
-				            spv::OpImageFetch, ImageVectorType(state, numeric_class, 4), color,
-				            descriptor, coord, spv::ImageOperandsSampleMask,
-				            AddressU32(ctx, mem, *address, dimension_info.coordinate_components));
-			        } else {
-				        state.builder.AddFunction(spv::OpImageFetch,
-				                                  ImageVectorType(state, numeric_class, 4), color,
-				                                  descriptor, coord, spv::ImageOperandsLodMask,
-				                                  LodU32(ctx, mem, *address, dimension));
-			        }
+			        const auto color = EmitImageAccess(
+			            ctx, inst, ImageVectorType(state, numeric_class, 4),
+			            [&](uint32_t resource, uint32_t array_index) {
+				        const auto& candidate = state.program.info.images[resource];
+				        const auto dimension = candidate.dimension;
+				        const auto& dimension_info = ImageDimensionInfoFor(dimension);
+				        const auto descriptor = LoadImageDescriptor(state, resource, 0u, array_index);
+				        const auto fetched = state.builder.AllocateId();
+				        const auto coord = CoordU32(ctx, mem, *address, dimension);
+				        if (dimension_info.multisampled != 0u) {
+					        state.builder.AddFunction(
+					            spv::OpImageFetch, ImageVectorType(state, candidate.numeric_class, 4), fetched,
+					            descriptor, coord, spv::ImageOperandsSampleMask,
+					            AddressU32(ctx, mem, *address, dimension_info.coordinate_components));
+				        } else {
+					        state.builder.AddFunction(
+					            spv::OpImageFetch, ImageVectorType(state, candidate.numeric_class, 4), fetched,
+					            descriptor, coord, spv::ImageOperandsLodMask,
+					            LodU32(ctx, mem, *address, mem.image_dimension));
+				        }
+				        return fetched;
+			            });
 			        return ResultVector(ctx, UnpackImageTexel(ctx, mem, color), numeric_class,
 			                            false, mem);
 		        }));
@@ -625,7 +779,7 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 		const auto dimension  = image.dimension;
 		EmitIfCondition(state, ctx.Arg(inst, 3), [&]() {
 			const auto mip_lod =
-			    state.program.info.images[mem.resource].mip_mode == IR::ImageMipMode::DynamicStorage
+			    state.program.info.images[mem.resource].mip_mode == IR::ImageMipMode::Dynamic
 			        ? LodU32(ctx, mem, *address, dimension)
 			        : 0u;
 			const auto coord = CoordU32(ctx, mem, *address, dimension);
@@ -642,19 +796,12 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 		const bool  dref           = HasFlag(mem, Decoder::ImageSampleFlagCompare);
 		if (dref && state.program.info.images[mem.resource].conversion_format !=
 		                Prospero::BufferFormat::kInvalid) {
-			ctx.Fail(inst, "uses depth comparison with a packed integer image");
+			ctx.Fail(inst, "uses depth comparison with a converted image");
 			return;
 		}
 		if (op == IR::ValueOpcode::ImageGatherRaw) {
 			const auto coord = CoordF32(ctx, mem, *address, layout.coord,
 			                            dimension_info.coordinate_components, image.cube);
-			if (HasFlag(mem, Decoder::ImageSampleFlagLod)) {
-				static std::atomic_flag warned = ATOMIC_FLAG_INIT;
-				if (!warned.test_and_set(std::memory_order_relaxed)) {
-					std::fputs("Warning: approximating IMAGE_GATHER4_L at mip level 0; explicit LOD is ignored.\n",
-					           stderr);
-				}
-			}
 			if (dimension == ImageDimension::Dim1D) {
 				if (dref || !HasFlag(mem, Decoder::ImageSampleFlagLevelZero) ||
 				    HasFlag(mem, Decoder::ImageSampleFlagOffset) ||
@@ -671,44 +818,43 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 				ctx.Fail(inst, "has an unsupported 1D-array gather");
 				return;
 			}
-			const auto            sampled = MakeSampledImage(state, mem.resource, mem.sampler);
-			const auto            sample  = state.builder.AllocateId();
-			std::vector<uint32_t> words;
+			const auto result_numeric_class = dref ? Prospero::TextureNumericClass::Float
+			                                       : numeric_class;
+			const auto result_type = ImageVectorType(state, result_numeric_class, 4);
+			uint32_t component_or_dref;
 			if (dref) {
-				auto dref_value = ZeroF32(state);
-				if (layout.dref != NoImageComponent) {
-					dref_value = AddressF32(ctx, mem, *address, layout.dref);
-				}
-				words = {spv::OpImageDrefGather,
-				         TypeF32Vector(state, 4),
-				         sample,
-				         sampled,
-				         coord,
-				         dref_value};
+				component_or_dref = layout.dref != NoImageComponent
+				                        ? AddressF32(ctx, mem, *address, layout.dref)
+				                        : ZeroF32(state);
 			} else {
-				uint32_t component = 0;
-				if (ImageConversionFormat(state, mem).format == Prospero::BufferFormat::kInvalid) {
-					component = ImageGatherComponent(mem.dmask);
-				}
-				words = {spv::OpImageGather,
-				         ImageVectorType(state, numeric_class, 4),
-				         sample,
-				         sampled,
-				         coord,
-				         ConstantU32(state, component)};
+				component_or_dref = ConstantU32(state, ImageGatherSource(state, mem));
 			}
+			uint32_t operand_mask = 0;
+			uint32_t offset = 0;
 			if (HasFlag(mem, Decoder::ImageSampleFlagGatherHorizontal)) {
-				words.push_back(spv::ImageOperandsConstOffsetsMask);
-				words.push_back(HorizontalOffsets(state, dimension));
+				operand_mask = spv::ImageOperandsConstOffsetsMask;
+				offset = HorizontalOffsets(state, dimension);
 			} else if (layout.offset != NoImageComponent) {
-				words.push_back(spv::ImageOperandsOffsetMask);
-				words.push_back(PackedOffset(ctx, mem, *address, layout, dimension));
-			}
-			state.builder.AddFunction(words);
-			auto result_numeric_class = numeric_class;
-			if (dref) {
-				result_numeric_class = Prospero::TextureNumericClass::Float;
+				operand_mask = spv::ImageOperandsOffsetMask;
+				offset = PackedOffset(ctx, mem, *address, layout, dimension);
 			}
+			const auto sampler_id = LoadSamplerDescriptor(state, mem.sampler);
+			const auto EmitGather = [&](uint32_t mip) {
+				const auto sampled = MakeSampledImage(state, mem.resource, sampler_id, mip);
+				const auto sample = state.builder.AllocateId();
+				const std::array operands {operand_mask, offset};
+				state.builder.AddFunction(
+				    dref ? spv::OpImageDrefGather : spv::OpImageGather, result_type,
+				    sample, sampled, coord, component_or_dref,
+				    std::span(operands).first(operand_mask != 0u ? 2u : 0u));
+				return sample;
+			};
+			const auto sample = image.mip_mode == IR::ImageMipMode::Dynamic && image.mip_count > 1u
+			                        ? EmitIndexSwitch(
+			                              state, GatherMip(ctx, inst, mem, *address, layout,
+			                                               image.mip_count),
+			                              image.mip_count, result_type, EmitGather)
+			                        : EmitGather(0);
 			ctx.Define(inst, ResultVector(ctx, UnpackImageGather(ctx, mem, sample),
 			                              result_numeric_class, false, mem, true));
 			return;
@@ -732,159 +878,70 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 				dref_value = AddressF32(ctx, mem, *address, layout.dref);
 			}
 		}
-		uint32_t              operand_mask = 0;
-		std::vector<uint32_t> operands;
+		std::array<uint32_t, 3> operands {};
+		size_t operand_count = 0;
 		if (HasFlag(mem, Decoder::ImageSampleFlagDerivative)) {
-			operand_mask |= spv::ImageOperandsGradMask;
-			operands.push_back(
-			    CoordF32(ctx, mem, *address, layout.grad_x, dimension_info.spatial_components));
-			operands.push_back(
-			    CoordF32(ctx, mem, *address, layout.grad_y, dimension_info.spatial_components));
+			operands[0] = spv::ImageOperandsGradMask;
+			operands[1] = CoordF32(ctx, mem, *address, layout.grad_x, dimension_info.spatial_components);
+			operands[2] = CoordF32(ctx, mem, *address, layout.grad_y, dimension_info.spatial_components);
+			operand_count = 3;
 		} else if (explicit_lod) {
-			operand_mask |= spv::ImageOperandsLodMask;
 			auto lod = ZeroF32(state);
 			if (HasFlag(mem, Decoder::ImageSampleFlagLod) && layout.lod != NoImageComponent) {
 				lod = AddressF32(ctx, mem, *address, layout.lod);
 			}
-			operands.push_back(lod);
+			operands[0] = spv::ImageOperandsLodMask;
+			operands[1] = lod;
+			operand_count = 2;
 		} else if (layout.bias != NoImageComponent) {
-			operand_mask |= spv::ImageOperandsBiasMask;
-			operands.push_back(AddressF32(ctx, mem, *address, layout.bias));
+			operands[0] = spv::ImageOperandsBiasMask;
+			operands[1] = AddressF32(ctx, mem, *address, layout.bias);
+			operand_count = 2;
 		}
-		const auto EmitSample = [&](uint32_t resource) {
+		const auto& samplers = state.program.info.samplers.at(mem.sampler).indirect_resources;
+		uint32_t sampler_index = 0;
+		if (!samplers.empty()) {
+			const auto* handle = inst.Arg(1).ResolveInstruction();
+			if (handle == nullptr || handle->GetOpcode() != IR::ValueOpcode::GetSamplerResource)
+				ctx.Fail(inst, "has invalid finite sampler key provenance");
+			sampler_index = EmitIndexSwitch(
+			    state, ctx.Def(handle->Arg(0)), static_cast<uint32_t>(samplers.size()),
+			    TypeU32(state), [&](uint32_t ordinal) { return ConstantU32(state, samplers[ordinal]); });
+		}
+		const auto sampler_id = LoadSamplerDescriptor(state, mem.sampler, sampler_index);
+		const auto EmitSample = [&](uint32_t resource, uint32_t array_index) {
 			const auto& candidate = state.program.info.images[resource];
 			const auto coord =
 			    CoordF32(ctx, mem, *address, layout.coord,
 			             ImageDimensionInfoFor(candidate.dimension).coordinate_components,
 			             candidate.cube);
-			const auto            sampled = MakeSampledImage(state, resource, mem.sampler);
-			const auto            sample  = state.builder.AllocateId();
-			std::vector<uint32_t> sample_operands {result_type, sample, sampled, coord};
-			if (dref) {
-				sample_operands.push_back(dref_value);
-			}
-			if (operand_mask != 0u) {
-				sample_operands.push_back(operand_mask);
-				sample_operands.insert(sample_operands.end(), operands.begin(), operands.end());
-			}
-			state.builder.AddFunction(opcode, sample_operands);
+			const auto sampled = MakeSampledImage(state, resource, sampler_id, 0u, array_index,
+			                                      sampler_index != 0u);
+			const auto sample = state.builder.AllocateId();
+			state.builder.AddFunction(opcode, result_type, sample, sampled, coord,
+			                          std::span(&dref_value, dref ? 1u : 0u),
+			                          std::span<const uint32_t>(operands).first(operand_count));
 			return sample;
 		};
-		if (image.indirect_root != mem.resource) {
-			const auto sample = EmitSample(mem.resource);
-			auto       result = sample;
-			if (!dref) {
-				result = UnpackImageTexel(ctx, mem, sample);
-			}
-			ctx.Define(inst, ResultVector(ctx, result, numeric_class, dref, mem));
-			return;
-		}
-		const auto* handle = image_arg.ResolveInstruction();
-		const auto* source = image.source < state.program.descriptor_sources.size()
-		                         ? &state.program.descriptor_sources[image.source]
-		                         : nullptr;
-		if (handle == nullptr || source == nullptr || !source->indirect_image.has_value() ||
-		    handle->NumArgs() == 0u) {
-			ctx.Fail(inst, "has invalid indirect image key provenance");
-			return;
-		}
-		const auto key = ctx.Def(handle->Arg(0));
-		if (state.flattened_srt_variable == 0 || image.indirect_search_iterations == 0u ||
-		    image.indirect_resources.size() < 2u) {
-			ctx.Fail(inst, "has no indirect image runtime mapping");
-			return;
-		}
-		const auto LoadMapping = [&](uint32_t index) {
-			const auto pointer = state.builder.AllocateId();
-			state.builder.AddFunction(spv::OpAccessChain, TypeStorageBufferElementPointer(state),
-			                          pointer, state.flattened_srt_variable, ConstantU32(state, 0),
-			                          index);
-			const auto value = state.builder.AllocateId();
-			state.builder.AddFunction(spv::OpLoad, TypeU32(state), value, pointer);
-			return value;
-		};
-		const auto mapping  = ConstantU32(state, image.indirect_mapping_offset);
-		auto       low      = ConstantU32(state, 0u);
-		auto       high     = LoadMapping(mapping);
-		auto       selected = ConstantU32(state, 0u);
-		for (uint32_t iteration = 0; iteration < image.indirect_search_iterations;
-		     iteration++) {
-			const auto active = Binary(state, spv::OpULessThan, TypeBool(state), low, high);
-			const auto mid    = Binary(state, spv::OpShiftRightLogical, TypeU32(state),
-			                           Binary(state, spv::OpIAdd, TypeU32(state), low, high),
-			                           ConstantU32(state, 1u));
-			const auto probe  = state.builder.AllocateId();
-			state.builder.AddFunction(spv::OpSelect, TypeU32(state), probe, active, mid,
-			                          ConstantU32(state, 0u));
-			const auto entry = Binary(state, spv::OpIAdd, TypeU32(state), mapping,
-			                          Binary(state, spv::OpIAdd, TypeU32(state),
-			                                 Binary(state, spv::OpShiftLeftLogical, TypeU32(state),
-			                                        probe, ConstantU32(state, 1u)),
-			                                 ConstantU32(state, 1u)));
-			const auto mapped_key = LoadMapping(entry);
-			const auto candidate  = LoadMapping(
-			    Binary(state, spv::OpIAdd, TypeU32(state), entry, ConstantU32(state, 1u)));
-			const auto equal = Binary(state, spv::OpIEqual, TypeBool(state), mapped_key, key);
-			const auto match = Binary(state, spv::OpLogicalAnd, TypeBool(state), active, equal);
-			const auto next_selected = state.builder.AllocateId();
-			state.builder.AddFunction(spv::OpSelect, TypeU32(state), next_selected, match,
-			                          candidate, selected);
-			selected              = next_selected;
-			const auto less = Binary(state, spv::OpULessThan, TypeBool(state), mapped_key, key);
-			const auto take_upper = Binary(state, spv::OpLogicalAnd, TypeBool(state), active, less);
-			const auto take_lower = Binary(state, spv::OpLogicalAnd, TypeBool(state), active,
-			                               Unary(state, spv::OpLogicalNot, TypeBool(state), less));
-			const auto next_low   = state.builder.AllocateId();
-			state.builder.AddFunction(
-			    spv::OpSelect, TypeU32(state), next_low, take_upper,
-			    Binary(state, spv::OpIAdd, TypeU32(state), mid, ConstantU32(state, 1u)), low);
-			low                  = next_low;
-			const auto next_high = state.builder.AllocateId();
-			state.builder.AddFunction(spv::OpSelect, TypeU32(state), next_high, take_lower, mid,
-			                          high);
-			high = next_high;
-		}
-		const auto            default_label = state.builder.AllocateId();
-		const auto            merge_label   = state.builder.AllocateId();
-		std::vector<uint32_t> labels(image.indirect_resources.size() - 1u);
-		std::vector<uint32_t> switch_words {spv::OpSwitch, selected, default_label};
-		for (uint32_t candidate = 1; candidate < image.indirect_resources.size(); candidate++) {
-			labels[candidate - 1u] = state.builder.AllocateId();
-			switch_words.push_back(candidate);
-			switch_words.push_back(labels[candidate - 1u]);
-		}
-		state.builder.AddFunction(spv::OpSelectionMerge, merge_label,
-		                          spv::SelectionControlMaskNone);
-		state.builder.AddFunction(switch_words);
-		std::vector<uint32_t> phi_words {spv::OpPhi, result_type, state.builder.AllocateId()};
-		EmitLabel(state, default_label);
-		phi_words.push_back(EmitSample(image.indirect_resources[0]));
-		phi_words.push_back(default_label);
-		state.builder.AddFunction(spv::OpBranch, merge_label);
-		for (uint32_t candidate = 1; candidate < image.indirect_resources.size(); candidate++) {
-			EmitLabel(state, labels[candidate - 1u]);
-			phi_words.push_back(EmitSample(image.indirect_resources[candidate]));
-			phi_words.push_back(labels[candidate - 1u]);
-			state.builder.AddFunction(spv::OpBranch, merge_label);
-		}
-		EmitLabel(state, merge_label);
-		state.builder.AddFunction(phi_words);
-		auto result = phi_words[2];
+		auto result = EmitImageAccess(ctx, inst, result_type, EmitSample);
 		if (!dref) {
 			result = UnpackImageTexel(ctx, mem, result);
 		}
 		ctx.Define(inst, ResultVector(ctx, result, numeric_class, dref, mem));
 		return;
 	}
-	const auto atomic_opcode = ImageAtomicOpcode(op);
 	if (image_info.access == IR::ImageAccess::Atomic) {
 		const auto dimension = image.dimension;
-		ctx.Define(inst, EmitValueOrZeroIfCondition(state, ctx.Arg(inst, 3), [&]() {
+		const auto result_type = TypeId(state, inst.GetType());
+		const auto zero = image.atomic64 ? ConstantU64(state, 0) : ConstantU32(state, 0);
+		ctx.Define(inst, EmitValueOrDefaultIfCondition(
+		                     state, ctx.Arg(inst, inst.NumArgs() - 1), result_type, zero, [&]() {
 			           const auto pointer      = state.builder.AllocateId();
-			           const auto pointer_type = state.builder.Type(
-			               spv::OpTypePointer, spv::StorageClassImage, TypeU32(state));
+			           const auto pointer_type =
+			               state.builder.Type(spv::OpTypePointer, spv::StorageClassImage,
+			                                  image.atomic64 ? TypeU64(state) : TypeU32(state));
 			           state.builder.AddFunction(spv::OpImageTexelPointer, pointer_type, pointer,
-			                                     StorageImageDescriptorPointer(state, mem.resource),
+			                                     ImageDescriptorPointer(state, mem.resource),
 			                                     CoordU32(ctx, mem, *address, dimension),
 			                                     ConstantU32(state, 0));
 			           if (op == IR::ValueOpcode::ImageAtomicFMin32 ||
@@ -896,12 +953,8 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 					                                   op == IR::ValueOpcode::ImageAtomicFMax32);
 				                               });
 			           }
-			           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);
+			           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

@@ -23,27 +23,22 @@ 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);
-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>;
+EMIT_NATIVE(ConvertF32S32, OpConvertSToF, F32, uint32_t)
+EMIT_NATIVE(ConvertF64S32, OpConvertSToF, F64, uint32_t)
 uint32_t              EmitConvertF32F64(EmitterState& state, uint32_t arg0);
 uint32_t              EmitConvertF64F32(EmitterState& state, uint32_t arg0);
 EMIT_NATIVE(ConvertF32U32, OpConvertUToF, F32, uint32_t)
 EMIT_NATIVE(ConvertF64U32, OpConvertUToF, F64, uint32_t)
-EMIT_NATIVE(CompositeConstructU64, OpCompositeConstruct, U64, uint32_t, uint32_t)
+uint32_t EmitCompositeConstructU64(ValueEmitContext& ctx, uint32_t arg0, IR::Value arg1);
 EMIT_NATIVE(CompositeConstructU32x2, OpCompositeConstruct, U32x2, uint32_t, uint32_t)
 EMIT_NATIVE(CompositeConstructU32x3, OpCompositeConstruct, U32x3, uint32_t, uint32_t, uint32_t)
 EMIT_NATIVE(CompositeConstructF32x2, OpCompositeConstruct, F32x2, uint32_t, uint32_t)
 EMIT_NATIVE(CompositeConstructU32x4, OpCompositeConstruct, U32x4, uint32_t, uint32_t, uint32_t,
             uint32_t)
 uint32_t              EmitCompositeExtractU64(EmitterState& state, uint32_t arg0, IR::Value arg1);
-inline constexpr auto EmitCompositeExtractU32x2 = EmitCompositeExtractU64;
-inline constexpr auto EmitCompositeExtractU32x3 = EmitCompositeExtractU64;
-inline constexpr auto EmitCompositeExtractU32x4 = EmitCompositeExtractU64;
+uint32_t              EmitCompositeExtractU32x2(EmitterState& state, uint32_t arg0, IR::Value arg1);
+inline constexpr auto EmitCompositeExtractU32x3 = EmitCompositeExtractU32x2;
+inline constexpr auto EmitCompositeExtractU32x4 = EmitCompositeExtractU32x2;
 inline constexpr auto EmitPackHalf2x16 = EmitGlsl<GLSLstd450PackHalf2x16, IR::Type::U32, uint32_t>;
 inline constexpr auto EmitPackSnorm2x16 =
     EmitGlsl<GLSLstd450PackSnorm2x16, IR::Type::U32, uint32_t>;
@@ -63,34 +58,46 @@ EMIT_NATIVE(IAdd32, OpIAdd, U32, uint32_t, uint32_t)
 EMIT_NATIVE(ISub32, OpISub, U32, uint32_t, uint32_t)
 EMIT_NATIVE(IMul32, OpIMul, U32, uint32_t, uint32_t)
 EMIT_NATIVE(UDiv32, OpUDiv, U32, uint32_t, uint32_t)
-uint32_t EmitIAdd64(EmitterState& state, uint32_t arg0, uint32_t arg1);
-uint32_t EmitISub64(EmitterState& state, uint32_t arg0, uint32_t arg1);
-uint32_t EmitIMul64(EmitterState& state, uint32_t arg0, uint32_t arg1);
+EMIT_NATIVE(IAdd64, OpIAdd, U64, uint32_t, uint32_t)
+EMIT_NATIVE(ISub64, OpISub, U64, uint32_t, uint32_t)
+EMIT_NATIVE(IMul64, OpIMul, U64, uint32_t, uint32_t)
 EMIT_NATIVE(IAddCarry32, OpIAddCarry, U32x2, uint32_t, uint32_t)
 uint32_t EmitSMulHi(EmitterState& state, uint32_t arg0, uint32_t arg1);
 uint32_t EmitUMulHi(EmitterState& state, uint32_t arg0, uint32_t arg1);
-uint32_t EmitIAbs32(EmitterState& state, uint32_t arg0);
+inline constexpr auto EmitIAbs32 = EmitGlsl<GLSLstd450SAbs, IR::Type::U32, uint32_t>;
 EMIT_NATIVE(ShiftLeftLogical32, OpShiftLeftLogical, U32, uint32_t, uint32_t)
 EMIT_NATIVE(ShiftRightLogical32, OpShiftRightLogical, U32, uint32_t, uint32_t)
 EMIT_NATIVE(ShiftRightArithmetic32, OpShiftRightArithmetic, U32, uint32_t, uint32_t)
-uint32_t EmitShiftLeftLogical64(ValueEmitContext& ctx, uint32_t arg0, IR::Value arg1);
-uint32_t EmitShiftRightLogical64(ValueEmitContext& ctx, uint32_t arg0, IR::Value arg1);
-uint32_t EmitShiftRightArithmetic64(ValueEmitContext& ctx, uint32_t arg0, IR::Value arg1);
+template <spv::Op opcode>
+uint32_t EmitShift64(ValueEmitContext& ctx, uint32_t value, IR::Value shift) {
+	shift = shift.Resolve();
+	if (shift.IsImmediate()) {
+		const auto amount = shift.U32() & 63u;
+		return amount == 0u ? value
+		                    : Binary(ctx.state, opcode, TypeU64(ctx.state), value,
+		                             ConstantU32(ctx.state, amount));
+	}
+	return Binary(ctx.state, opcode, TypeU64(ctx.state), value,
+	              EmitAndConstant(ctx.state, ctx.Def(shift), 63u));
+}
+inline constexpr auto EmitShiftLeftLogical64     = EmitShift64<spv::OpShiftLeftLogical>;
+inline constexpr auto EmitShiftRightLogical64    = EmitShift64<spv::OpShiftRightLogical>;
+inline constexpr auto EmitShiftRightArithmetic64 = EmitShift64<spv::OpShiftRightArithmetic>;
 EMIT_NATIVE(BitwiseAnd32, OpBitwiseAnd, U32, uint32_t, uint32_t)
 EMIT_NATIVE(BitwiseOr32, OpBitwiseOr, U32, uint32_t, uint32_t)
 EMIT_NATIVE(BitwiseXor32, OpBitwiseXor, U32, uint32_t, uint32_t)
 EMIT_NATIVE(BitwiseNot32, OpNot, U32, uint32_t)
-uint32_t EmitBitwiseAnd64(EmitterState& state, uint32_t arg0, uint32_t arg1);
+EMIT_NATIVE(BitwiseAnd64, OpBitwiseAnd, U64, uint32_t, uint32_t)
 EMIT_NATIVE(BitReverse32, OpBitReverse, U32, uint32_t)
 EMIT_NATIVE(BitCount32, OpBitCount, U32, uint32_t)
 uint32_t EmitBitCount64(EmitterState& state, uint32_t arg0);
-uint32_t EmitFindILsb32(EmitterState& state, uint32_t arg0);
-uint32_t EmitFindUMsb32(EmitterState& state, uint32_t arg0);
+inline constexpr auto EmitFindILsb32 = EmitGlsl<GLSLstd450FindILsb, IR::Type::U32, uint32_t>;
+inline constexpr auto EmitFindUMsb32 = EmitGlsl<GLSLstd450FindUMsb, IR::Type::U32, uint32_t>;
 uint32_t EmitFindUMsb64(EmitterState& state, uint32_t arg0);
-uint32_t EmitSMin32(EmitterState& state, uint32_t arg0, uint32_t arg1);
-uint32_t EmitSMax32(EmitterState& state, uint32_t arg0, uint32_t arg1);
-uint32_t EmitUMin32(EmitterState& state, uint32_t arg0, uint32_t arg1);
-uint32_t EmitUMax32(EmitterState& state, uint32_t arg0, uint32_t arg1);
+inline constexpr auto EmitSMin32 = EmitGlsl<GLSLstd450SMin, IR::Type::U32, uint32_t, uint32_t>;
+inline constexpr auto EmitSMax32 = EmitGlsl<GLSLstd450SMax, IR::Type::U32, uint32_t, uint32_t>;
+inline constexpr auto EmitUMin32 = EmitGlsl<GLSLstd450UMin, IR::Type::U32, uint32_t, uint32_t>;
+inline constexpr auto EmitUMax32 = EmitGlsl<GLSLstd450UMax, IR::Type::U32, uint32_t, uint32_t>;
 uint32_t EmitSMinTri32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32_t arg2);
 uint32_t EmitSMaxTri32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32_t arg2);
 uint32_t EmitUMinTri32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32_t arg2);
@@ -107,11 +114,14 @@ EMIT_NATIVE(UGreaterThan32, OpUGreaterThan, U1, uint32_t, uint32_t)
 EMIT_NATIVE(INotEqual32, OpINotEqual, U1, uint32_t, uint32_t)
 EMIT_NATIVE(SGreaterThanEqual32, OpSGreaterThanEqual, U1, uint32_t, uint32_t)
 EMIT_NATIVE(UGreaterThanEqual32, OpUGreaterThanEqual, U1, uint32_t, uint32_t)
-uint32_t EmitIEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1);
-uint32_t EmitINotEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1);
-uint32_t EmitULessThan64(EmitterState& state, uint32_t arg0, uint32_t arg1);
-uint32_t EmitSLessThan64(EmitterState& state, uint32_t arg0, uint32_t arg1);
-uint32_t EmitUGreaterThan64(EmitterState& state, uint32_t arg0, uint32_t arg1);
+EMIT_NATIVE(IEqual64, OpIEqual, U1, uint32_t, uint32_t)
+EMIT_NATIVE(INotEqual64, OpINotEqual, U1, uint32_t, uint32_t)
+EMIT_NATIVE(ULessThan64, OpULessThan, U1, uint32_t, uint32_t)
+EMIT_NATIVE(SLessThan64, OpSLessThan, U1, uint32_t, uint32_t)
+EMIT_NATIVE(SLessThanEqual64, OpSLessThanEqual, U1, uint32_t, uint32_t)
+EMIT_NATIVE(ULessThanEqual64, OpULessThanEqual, U1, uint32_t, uint32_t)
+EMIT_NATIVE(UGreaterThanEqual64, OpUGreaterThanEqual, U1, uint32_t, uint32_t)
+EMIT_NATIVE(UGreaterThan64, OpUGreaterThan, U1, uint32_t, uint32_t)
 EMIT_NATIVE(LogicalOr, OpLogicalOr, U1, uint32_t, uint32_t)
 EMIT_NATIVE(LogicalAnd, OpLogicalAnd, U1, uint32_t, uint32_t)
 EMIT_NATIVE(LogicalXor, OpLogicalNotEqual, U1, uint32_t, uint32_t)
@@ -133,6 +143,9 @@ uint32_t EmitFloatCompare32(ValueEmitContext& ctx, const IR::Inst& inst) {
 	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>;
@@ -144,19 +157,23 @@ inline constexpr auto EmitFPOrdLessThanEqual32 = EmitFloatCompare32<spv::OpFOrdL
 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);
+EMIT_NATIVE(FPIsNan32, OpIsNan, U1, uint32_t)
 inline constexpr auto EmitFPCmpClass32 = EmitClassMaskF32;
+inline constexpr auto EmitFPCmpClass16 = EmitClassMaskF16;
 EMIT_NATIVE(FPAdd32, OpFAdd, F32, uint32_t, uint32_t)
 EMIT_NATIVE(FPSub32, OpFSub, F32, uint32_t, uint32_t)
 EMIT_NATIVE(FPMul32, OpFMul, F32, uint32_t, uint32_t)
+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 EmitFPFma32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32_t arg2);
+uint32_t EmitFPMad32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32_t arg2);
 uint32_t EmitFPMin32(EmitterState& state, uint32_t arg0, uint32_t arg1);
 uint32_t EmitFPMax32(EmitterState& state, uint32_t arg0, uint32_t arg1);
+uint32_t EmitFPMin64(EmitterState& state, uint32_t arg0, uint32_t arg1);
+uint32_t EmitFPMax64(EmitterState& state, uint32_t arg0, uint32_t arg1);
 uint32_t EmitFPMinTri32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32_t arg2);
 uint32_t EmitFPMaxTri32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32_t arg2);
 uint32_t EmitFPMedTri32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32_t arg2);
@@ -171,7 +188,11 @@ inline constexpr auto EmitFPRoundEven32 = EmitGlsl<GLSLstd450RoundEven, IR::Type
 inline constexpr auto EmitFPFloor32     = EmitGlsl<GLSLstd450Floor, IR::Type::F32, uint32_t>;
 inline constexpr auto EmitFPCeil32      = EmitGlsl<GLSLstd450Ceil, IR::Type::F32, uint32_t>;
 inline constexpr auto EmitFPTrunc32     = EmitGlsl<GLSLstd450Trunc, IR::Type::F32, uint32_t>;
+inline constexpr auto EmitFPFloor64     = EmitGlsl<GLSLstd450Floor, IR::Type::F64, uint32_t>;
+inline constexpr auto EmitFPCeil64      = EmitGlsl<GLSLstd450Ceil, IR::Type::F64, uint32_t>;
+inline constexpr auto EmitFPTrunc64     = EmitGlsl<GLSLstd450Trunc, IR::Type::F64, uint32_t>;
 inline constexpr auto EmitFPFract32     = EmitGlsl<GLSLstd450Fract, IR::Type::F32, uint32_t>;
+inline constexpr auto EmitFPFract64     = EmitGlsl<GLSLstd450Fract, IR::Type::F64, uint32_t>;
 uint32_t              EmitFPSin(EmitterState& state, uint32_t arg0);
 uint32_t              EmitFPCos(EmitterState& state, uint32_t arg0);
 #undef EMIT_NATIVE
@@ -181,16 +202,17 @@ void                  EmitVoid(ValueEmitContext&);
 inline constexpr auto EmitReference    = EmitVoid;
 inline constexpr auto EmitReferenceU32 = EmitVoid;
 inline constexpr auto EmitControlNop   = EmitVoid;
-inline constexpr auto EmitWaitcnt      = EmitVoid;
 inline constexpr auto EmitSendmsg      = EmitVoid;
 inline constexpr auto EmitTtraceData   = EmitVoid;
 inline constexpr auto EmitInstPrefetch = EmitVoid;
 void                  EmitBarrier(EmitterState& state);
+void                  EmitStoreCompletion(EmitterState& state);
 void                  EmitMeshAllocate(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitMeshDrawParameter(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitGetTessellationAttribute(ValueEmitContext& ctx, const IR::Inst& inst);
 void                  EmitSetTessellationAttribute(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitGetUserData(EmitterState& state, IR::ScalarReg reg);
+uint32_t              EmitGetDispatchThreadExtent(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitGetBuiltin(ValueEmitContext& ctx, IR::Value kind, IR::Value index);
 uint32_t              EmitUndefU1(EmitterState& state, const IR::Inst& inst);
 inline constexpr auto EmitUndefU8  = EmitUndefU1;
@@ -222,10 +244,13 @@ 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);
+void                  EmitSharedAtomicMaskedOr32(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);
 uint32_t              EmitAppendConsume(ValueEmitContext& ctx, const IR::Inst& inst);
+uint32_t              EmitPermuteU32(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitBpermuteU32(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitReadConst(ValueEmitContext& ctx, const IR::Inst& inst);
 void                  EmitReadConstBuffer(ValueEmitContext& ctx, const IR::Inst& inst);
@@ -257,6 +282,7 @@ 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;
@@ -278,6 +304,7 @@ 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;
@@ -287,24 +314,36 @@ 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 EmitImageAtomicSwap64    = EmitImage;
 inline constexpr auto EmitImageAtomicIAdd32    = EmitImage;
+inline constexpr auto EmitImageAtomicIAdd64    = EmitImage;
+inline constexpr auto EmitImageAtomicSMin32    = EmitImage;
 inline constexpr auto EmitImageAtomicUMin32    = EmitImage;
+inline constexpr auto EmitImageAtomicUMin64    = EmitImage;
+inline constexpr auto EmitImageAtomicSMax32    = EmitImage;
 inline constexpr auto EmitImageAtomicUMax32    = EmitImage;
+inline constexpr auto EmitImageAtomicUMax64    = EmitImage;
 inline constexpr auto EmitImageAtomicAnd32     = EmitImage;
+inline constexpr auto EmitImageAtomicAnd64     = EmitImage;
 inline constexpr auto EmitImageAtomicOr32      = EmitImage;
+inline constexpr auto EmitImageAtomicOr64      = EmitImage;
 inline constexpr auto EmitImageAtomicXor32     = EmitImage;
+inline constexpr auto EmitImageAtomicXor64     = EmitImage;
 inline constexpr auto EmitImageAtomicFMin32    = EmitImage;
 inline constexpr auto EmitImageAtomicFMax32    = EmitImage;
 void                  EmitUnreachable(ValueEmitContext& ctx, const IR::Inst& inst);

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

@@ -6,7 +6,6 @@
 #include "graphics/shader/recompiler/BufferFormat.h"
 #include "graphics/shader/recompiler/backend/spirv/SpirvBuilder.h"
 #include "graphics/shader/recompiler/ir/ShaderIR.h"
-#include "graphics/shader/recompiler/ir/passes/BindingLayout.h"
 #include "graphics/shader/recompiler/ir/passes/ResourceMaterialization.h"
 
 #include <algorithm>
@@ -19,6 +18,7 @@
 #include <map>
 #include <spirv/unified1/GLSL.std.450.h>
 #include <string>
+#include <type_traits>
 #include <unordered_map>
 #include <utility>
 #include <vector>
@@ -57,53 +57,77 @@ constexpr std::array<ImageDimensionInfo, 7> ImageDimensions {{
 
 const ImageDimensionInfo& ImageDimensionInfoFor(ImageDimension dimension);
 
-struct SpirvRequirements {
-	bool subgroup_ballot              = false;
-	bool subgroup_shuffle             = false;
-	bool subgroup_local_invocation_id = false;
-	bool compute_derivatives          = false;
-	bool image_gather_extended        = false;
-	bool function_lds                 = false;
-	bool function_scratch             = false;
-	bool pixel_valid_mask             = false;
-	bool buffer_int64_atomics         = false;
-	bool coherent_buffers             = false;
-	bool float64                      = false;
+struct BufferDefinition {
+	uint32_t variable             = 0;
+	uint32_t type                 = 0;
+	uint32_t element_type         = 0;
+	uint32_t pointer_type         = 0;
+	uint32_t element_pointer_type = 0;
 };
 
-SpirvRequirements AnalyzeProgramRequirements(const IR::Program& program);
+struct ImageDefinition {
+	uint32_t variable     = 0;
+	uint32_t type         = 0;
+	uint32_t pointer_type = 0;
+	uint32_t sampled_type = 0;
+};
 
 struct EmitterState {
-	EmitterState(const IR::Program& program_, ShaderStageInputInfo input_info_)
+	EmitterState(IR::Program& program_, ShaderStageInputInfo input_info_)
 	    : builder(program_.stage == ShaderType::Mesh ? 0x00010400u : 0x00010300u),
-	      program(program_), input_info(input_info_),
-	      requirements(AnalyzeProgramRequirements(program_)) {}
+	      program(program_), input_info(input_info_) {
+		if (ShaderWorkgroupInput(program.stage, input_info) != nullptr) {
+			lds_storage_class = spv::StorageClassWorkgroup;
+		}
+		if (program.info.uses_lds && program.stage == ShaderType::Compute &&
+		    input_info.compute != nullptr && input_info.compute->lds_storage) {
+			lds_storage_class = spv::StorageClassStorageBuffer;
+		}
+	}
 
 	Builder                                          builder;
-	const IR::Program&                               program;
+	IR::Program&                                     program;
 	ShaderStageInputInfo                             input_info;
+	uint32_t                                        void_type = 0;
+	uint32_t                                        bool_type = 0;
+	uint32_t                                        u32_type = 0;
+	uint32_t                                        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;
 
-	const SpirvRequirements                          requirements;
 	uint32_t                                         lane_count              = 1;
 	uint32_t                                         lane_half               = 0;
-	uint32_t                                         storage_buffer_variable = 0;
-	uint32_t                                         storage_buffer_u64_variable = 0;
+	std::array<BufferDefinition, 4>                  storage_buffers {};
 	std::array<uint32_t, IR::ShaderInfo::MaxBuffers> memory_byte_offsets {};
 	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;
+	spv::StorageClass                                lds_storage_class = spv::StorageClassFunction;
+	uint32_t                                         lds_base_dwords         = 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 {};
+	std::array<ImageDefinition, IR::ImageBindingCount> images {};
 	uint32_t                   sampler_variable                      = 0;
+	uint32_t                   sampler_type                          = 0;
+	uint32_t                   sampler_pointer_type                  = 0;
 	uint32_t                   main_func                             = 0;
 	uint32_t                   mesh_guest_func                       = 0;
 	uint32_t                   mesh_allocation                       = 0;
@@ -127,7 +151,6 @@ struct EmitterState {
 	std::vector<InputBinding>  inputs;
 	std::vector<OutputBinding> outputs;
 	std::vector<uint32_t>      interface_variables;
-	std::unordered_map<const IR::Block*, uint32_t> labels;
 };
 
 uint32_t TypeVoid(EmitterState& state);
@@ -135,7 +158,6 @@ uint32_t TypeBool(EmitterState& state);
 uint32_t TypeBoolVector(EmitterState& state, uint32_t components);
 uint32_t TypeU32(EmitterState& state);
 uint32_t TypeU64(EmitterState& state);
-uint32_t TypeScalarU64(EmitterState& state);
 uint32_t TypeU32Pair(EmitterState& state);
 uint32_t TypeI32(EmitterState& state);
 uint32_t TypeI32Pair(EmitterState& state);
@@ -148,10 +170,7 @@ uint32_t TypeI32Vector(EmitterState& state, uint32_t components);
 uint32_t TypeF32Vector(EmitterState& state, uint32_t components);
 uint32_t TypePointer(EmitterState& state, spv::StorageClass storage_class, uint32_t pointee);
 uint32_t TypeFunction(EmitterState& state);
-uint32_t TypeStorageBufferPointer(EmitterState& state);
-uint32_t TypeStorageBufferElementPointer(EmitterState& state);
-uint32_t TypeStorageBufferU64Pointer(EmitterState& state);
-uint32_t TypeStorageBufferU64ElementPointer(EmitterState& state);
+uint32_t TypeStorageBufferElementPointer(EmitterState& state, uint32_t bits = 32);
 uint32_t TypePhysicalU32Pointer(EmitterState& state);
 uint32_t TypePushConstantElementPointer(EmitterState& state);
 uint32_t TypeU32ArrayPointer(EmitterState& state, spv::StorageClass storage_class, uint32_t dwords);
@@ -189,6 +208,12 @@ inline uint32_t Unary(EmitterState& state, spv::Op opcode, uint32_t type, uint32
 	return result;
 }
 
+inline uint32_t PackU64(EmitterState& state, uint32_t low, uint32_t high) {
+	const auto pair = state.builder.AllocateId();
+	state.builder.AddFunction(spv::OpCompositeConstruct, TypeU32Vector(state, 2), pair, low, high);
+	return Unary(state, spv::OpBitcast, TypeU64(state), pair);
+}
+
 inline uint32_t Binary(EmitterState& state, spv::Op opcode, uint32_t type, uint32_t lhs,
                        uint32_t rhs) {
 	const auto result = state.builder.AllocateId();
@@ -218,12 +243,10 @@ struct ValueEmitContext {
 	const IR::Inst*       ImageAddress(IR::Value value);
 	const IR::MemoryInfo& Memory(const IR::Inst& inst) const;
 	const IR::ExportInfo& Export(const IR::Inst& inst) const;
-	uint32_t              Label(const IR::Block* block) const;
 	[[noreturn]] void     Fail(const char* reason) const;
 	[[noreturn]] void     Fail(const IR::Inst& inst, const char* reason) const;
 
 	EmitterState&                                                      state;
-	std::unordered_map<const IR::Inst*, uint32_t>                      definitions;
 	const std::unordered_map<const IR::Inst*, uint32_t>*               dispatcher_spills = nullptr;
 	std::unordered_map<const IR::Inst*, std::pair<uint32_t, uint32_t>> dispatcher_block_loads;
 	uint32_t                                                           scratch_u32_variable = 0;
@@ -279,9 +302,6 @@ uint32_t EmitSubgroupLocalInvocationId(EmitterState& state);
                                                IR::DescriptorBindingKind kind, uint32_t resource,
                                                const char* reason);
 
-uint32_t ResourceForDescriptor(const EmitterState& state, IR::DescriptorBindingKind kind,
-                               uint32_t resource);
-
 uint32_t DescriptorElementPointer(EmitterState& state, uint32_t result_ptr_type,
                                   uint32_t variable_id, uint32_t array_index,
                                   IR::DescriptorBindingKind kind, uint32_t resource,
@@ -296,13 +316,17 @@ uint32_t ImageType(EmitterState& state, const IR::ImageResource& image);
 
 uint32_t ImageViewSizeType(EmitterState& state, ImageDimension dimension);
 
-uint32_t LoadSampledImageDescriptor(EmitterState& state, uint32_t resource);
+uint32_t ImageDescriptorPointer(EmitterState& state, uint32_t resource, uint32_t mip = 0,
+                                uint32_t array_index = 0);
 
-uint32_t LoadSamplerDescriptor(EmitterState& state, uint32_t sampler);
+uint32_t LoadImageDescriptor(EmitterState& state, uint32_t resource, uint32_t mip = 0,
+                             uint32_t array_index = 0);
 
-uint32_t MakeSampledImage(EmitterState& state, uint32_t resource, uint32_t sampler);
+uint32_t LoadSamplerDescriptor(EmitterState& state, uint32_t sampler, uint32_t array_index = 0);
 
-uint32_t StorageImageDescriptorPointer(EmitterState& state, uint32_t resource);
+uint32_t MakeSampledImage(EmitterState& state, uint32_t resource, uint32_t sampler_id,
+                          uint32_t mip = 0, uint32_t array_index = 0,
+                          bool sampler_dynamic = false);
 
 void EmitStorageImageWrite(EmitterState& state, uint32_t resource, uint32_t mip_lod, uint32_t coord,
                            uint32_t texel);
@@ -328,7 +352,8 @@ uint32_t DefineInterfaceVariable(EmitterState& state, uint32_t type, spv::Storag
 void     DefineModule(EmitterState& state);
 void     DefineTessellationInterfaces(EmitterState& state);
 void     DefineTessellationExecutionModes(EmitterState& state);
-void     DefineMeshOutputs(EmitterState& state);
+void     DefineMeshOutputs(EmitterState& state, uint32_t clip_distance_count,
+                           uint32_t cull_distance_count);
 void     EmitMeshEntryPoint(EmitterState& state);
 void     EmitMeshAllocate(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t MeshOutputPointer(EmitterState& state, IR::StageOutputKind kind, uint32_t index = 0);
@@ -376,25 +401,24 @@ 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;
 	uint32_t              object_pointer   = 0;
 	uint32_t              length           = 0;
-	uint32_t              index_offset     = 0;
 	uint32_t              byte_offset      = 0;
-	bool                  add_index_offset = false;
+	uint32_t              element_bits     = 32;
 	spv::MemoryAccessMask memory_access    = spv::MemoryAccessMaskNone;
 };
 
+uint32_t EmitIndirectResourceIndex(EmitterState& state, uint32_t key, uint32_t mapping_offset,
+                                   uint32_t search_iterations, uint32_t default_resource);
 MemoryResourceAccess PrepareMemoryResourceAccess(EmitterState& state, const IR::MemoryInfo& mem);
 
 MemoryResourceAccess PrepareStorageBufferResourceAccess(EmitterState&         state,
                                                         const IR::MemoryInfo& mem,
-                                                        uint32_t variable, uint32_t pointer_type);
-
-uint32_t EmitMemoryElementIndex(EmitterState& state, const MemoryResourceAccess& access,
-                                uint32_t raw_index);
+                                                        BufferDefinition& buffer);
 
 uint32_t EmitMemoryElementInBounds(EmitterState& state, const MemoryResourceAccess& access,
                                    uint32_t index);
@@ -417,11 +441,19 @@ 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);
 
+uint32_t EmitDsFloatAtomicReplacement(EmitterState& state, uint32_t old, uint32_t source,
+                                      bool max_value);
+
 uint32_t EmitDsSwizzleTargetLane(EmitterState& state, uint32_t subid, uint32_t control);
 
 inline constexpr auto EmitSelectValueU32 =
@@ -440,10 +472,6 @@ uint32_t EmitSubConstantMinusU32(EmitterState& state, uint32_t constant, uint32_
 
 uint32_t EmitF32ToF16RtzBits(EmitterState& state, uint32_t f32);
 
-uint32_t EmitMinMaxU32Value(EmitterState& state, uint32_t lhs, uint32_t rhs, bool max_value);
-
-uint32_t EmitMinMaxI32Value(EmitterState& state, uint32_t lhs, uint32_t rhs, bool max_value);
-
 inline constexpr auto EmitBitcastF32ToU32 = EmitNative<spv::OpBitcast, IR::Type::U32, uint32_t>;
 
 inline constexpr auto EmitBitcastU32ToF32 = EmitNative<spv::OpBitcast, IR::Type::F32, uint32_t>;
@@ -466,6 +494,7 @@ uint32_t EmitClassMaskBitMatch(EmitterState& state, uint32_t mask, uint32_t bit,
                                uint32_t class_match);
 
 uint32_t EmitClassMaskF32(EmitterState& state, uint32_t value, uint32_t mask);
+uint32_t EmitClassMaskF16(EmitterState& state, uint32_t bits, uint32_t mask);
 
 uint32_t EmitMinMaxF32Value(EmitterState& state, uint32_t lhs, uint32_t rhs, bool max_value);
 
@@ -484,8 +513,49 @@ 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);
 
 // These templates accept local lambdas from several emitter translation units.
+template <typename Fn>
+auto EmitIndexSwitch(EmitterState& state, uint32_t index, uint32_t count,
+                        uint32_t result_type, Fn&& emit) {
+	constexpr bool has_result = !std::is_void_v<std::invoke_result_t<Fn, uint32_t>>;
+	EXIT_IF(count == 0u);
+	const auto merge_label = state.builder.AllocateId();
+	std::vector<uint32_t> labels(count);
+	std::vector<uint32_t> words {spv::OpSwitch, index, merge_label};
+	for (uint32_t item = 0; item < count; item++) {
+		labels[item] = state.builder.AllocateId();
+		words.push_back(item);
+		words.push_back(labels[item]);
+	}
+	if constexpr (has_result) {
+		words[2] = labels.front();
+	}
+	state.builder.AddFunction(spv::OpSelectionMerge, merge_label, spv::SelectionControlMaskNone);
+	state.builder.AddFunction(words);
+	std::vector<uint32_t> phi;
+	if constexpr (has_result) {
+		phi = {spv::OpPhi, result_type, state.builder.AllocateId()};
+	}
+	for (uint32_t item = 0; item < count; item++) {
+		EmitLabel(state, labels[item]);
+		if constexpr (has_result) {
+			phi.push_back(emit(item));
+			phi.push_back(state.current_label);
+		} else {
+			emit(item);
+		}
+		state.builder.AddFunction(spv::OpBranch, merge_label);
+	}
+	EmitLabel(state, merge_label);
+	if constexpr (has_result) {
+		state.builder.AddFunction(phi);
+		return phi[2];
+	}
+}
+
 template <typename Fn>
 void EmitIfCondition(EmitterState& state, uint32_t condition, Fn&& fn) {
 	const auto then_label  = state.builder.AllocateId();
@@ -525,14 +595,7 @@ 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();
@@ -561,8 +624,7 @@ 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;
 }