ngs2: let CLEAR reopen a closed waveform block list

Co-authored-by: Phr3d13 <phr3d13@gmail.com>

.github/workflows/build.yml

@@ -113,7 +113,7 @@ jobs:
         shell: cmd
         run: |
           cmake --build _Build/windows --target launcher ^
-            audio_out2_port_tests pad_haptics_tests controller_settings_tests host_input_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests ^
+            audio_out2_port_tests ngs2_sampler_tests pad_haptics_tests controller_settings_tests host_input_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests ^
             shader_recompiler_compute_tests ^
             resource_tracking_tests scalar_provenance_tests ^
             resource_materialization_tests shader_vertex_metadata_tests archive_file_tests avplayer_file_tests --parallel
@@ -121,7 +121,7 @@ jobs:
       - name: Test
         shell: cmd
         run: |
-          ctest --test-dir _Build/windows --output-on-failure -R "^(audio_out2_port|pad_haptics|controller_settings|host_input|ime_dialog|lru_cache|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|archive_file|avplayer_file|virtual_memory_allocation)$"
+          ctest --test-dir _Build/windows --output-on-failure -R "^(audio_out2_port|ngs2_sampler|pad_haptics|controller_settings|host_input|ime_dialog|lru_cache|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|archive_file|avplayer_file|virtual_memory_allocation)$"
 
       - name: Install
         shell: cmd
@@ -202,7 +202,7 @@ jobs:
         shell: bash
         run: |
           cmake --build _Build/macos \
-            --target launcher audio_out2_port_tests pad_haptics_tests controller_settings_tests host_input_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests \
+            --target launcher audio_out2_port_tests ngs2_sampler_tests pad_haptics_tests controller_settings_tests host_input_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests \
               memory_tracker_tests page_manager_tests shader_recompiler_compute_tests \
               resource_tracking_tests scalar_provenance_tests \
               resource_materialization_tests shader_vertex_metadata_tests archive_file_tests avplayer_file_tests \
@@ -212,7 +212,7 @@ jobs:
         shell: bash
         run: |
           ctest --test-dir _Build/macos --output-on-failure \
-            -R '^(audio_out2_port|pad_haptics|controller_settings|host_input|ime_dialog|lru_cache|memory_tracker|page_manager|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|archive_file|avplayer_file|virtual_memory_allocation)$'
+            -R '^(audio_out2_port|ngs2_sampler|pad_haptics|controller_settings|host_input|ime_dialog|lru_cache|memory_tracker|page_manager|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|archive_file|avplayer_file|virtual_memory_allocation)$'
 
       - name: Install
         shell: bash
@@ -388,7 +388,7 @@ jobs:
         run: |
           cmake --build _Build/linux \
             --target launcher page_manager_tests memory_tracker_tests \
-              audio_out2_port_tests pad_haptics_tests controller_settings_tests host_input_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests \
+              audio_out2_port_tests ngs2_sampler_tests pad_haptics_tests controller_settings_tests host_input_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests \
               resource_tracking_tests scalar_provenance_tests \
               resource_materialization_tests shader_vertex_metadata_tests archive_file_tests avplayer_file_tests \
             --parallel
@@ -397,7 +397,7 @@ jobs:
         shell: bash
         run: |
           ctest --test-dir _Build/linux --output-on-failure \
-            -R '^(audio_out2_port|pad_haptics|controller_settings|host_input|ime_dialog|lru_cache|memory_tracker|page_manager|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|archive_file|avplayer_file|virtual_memory_allocation)$'
+            -R '^(audio_out2_port|ngs2_sampler|pad_haptics|controller_settings|host_input|ime_dialog|lru_cache|memory_tracker|page_manager|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|archive_file|avplayer_file|virtual_memory_allocation)$'
 
       - name: Install
         shell: bash

3rdparty/CMakeLists.txt

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

CMakeLists.txt

@@ -325,6 +325,7 @@ add_gpu_blit_shader(gpu_blit_fs_triangle gpu_blit_fs_triangle.vert)
 add_gpu_blit_shader(gpu_blit_color_to_ms_depth gpu_blit_color_to_ms_depth.frag)
 add_gpu_blit_shader(gpu_video_out_overlay gpu_video_out_overlay.frag)
 add_gpu_blit_shader(gpu_imgui_hdr gpu_video_out_overlay.frag -DIMGUI_HDR=1)
+add_gpu_blit_shader(gpu_pssr_linear gpu_pssr_linear.comp)
 list(APPEND kyty_emulator_src ${gpu_blit_shader_headers})
 
 list(APPEND kyty_emulator_src
@@ -407,6 +408,8 @@ add_executable(scalar_provenance_tests EXCLUDE_FROM_ALL
 target_link_libraries(scalar_provenance_tests fmt::fmt xbyak::xbyak)
 target_include_directories(scalar_provenance_tests PRIVATE ${inc_headers})
 
+add_executable(build_flags_tests EXCLUDE_FROM_ALL "${KYTY_TESTS_DIR}/BuildFlagsTests.cpp")
+
 add_executable(page_manager_tests EXCLUDE_FROM_ALL
 	"${KYTY_TESTS_DIR}/PageManagerTests.cpp"
 	"${KYTY_SOURCE_DIR}/graphics/host_gpu/pageManager.cpp"
@@ -520,6 +523,13 @@ add_executable(ngs2_sampler_tests EXCLUDE_FROM_ALL
 target_link_libraries(ngs2_sampler_tests common fmt::fmt LibAtrac9)
 target_include_directories(ngs2_sampler_tests PRIVATE ${inc_headers})
 
+add_executable(ajm_resample_tests EXCLUDE_FROM_ALL
+	"${KYTY_TESTS_DIR}/AjmResampleTests.cpp"
+	"${KYTY_SOURCE_DIR}/loader/timer.cpp"
+)
+target_link_libraries(ajm_resample_tests common fmt::fmt FFmpeg::ffmpeg LibAtrac9 opus)
+target_include_directories(ajm_resample_tests PRIVATE ${inc_headers})
+
 add_executable(pad_haptics_tests EXCLUDE_FROM_ALL
 	"${KYTY_TESTS_DIR}/PadHapticsTests.cpp"
 )
@@ -622,6 +632,19 @@ target_sources(shader_recompiler_compute_tests PRIVATE
 	"${gpu_test_ms_depth_header}")
 
 add_kyty_full_emulator_test(virtual_memory_allocation_tests "${KYTY_TESTS_DIR}/VirtualMemoryAllocationTests.cpp")
+add_library(kyty_test_gtest STATIC EXCLUDE_FROM_ALL
+	"${KYTY_THIRD_PARTY_DIR}/fmt/test/gtest/gmock-gtest-all.cc")
+target_include_directories(kyty_test_gtest PUBLIC "${KYTY_THIRD_PARTY_DIR}/fmt/test/gtest")
+target_sources(virtual_memory_allocation_tests PRIVATE "${KYTY_TESTS_DIR}/TrinityTests.cpp")
+target_link_libraries(virtual_memory_allocation_tests kyty_test_gtest)
+
+add_executable(psml_tests EXCLUDE_FROM_ALL
+	"${KYTY_TESTS_DIR}/PsmlTests.cpp"
+	"${KYTY_SOURCE_DIR}/libs/libPsml.cpp"
+	"${KYTY_SOURCE_DIR}/loader/symbolDatabase.cpp"
+	"${KYTY_SOURCE_DIR}/loader/timer.cpp")
+target_link_libraries(psml_tests common kyty_test_gtest)
+target_include_directories(psml_tests PRIVATE ${inc_headers})
 target_compile_definitions(virtual_memory_allocation_tests PRIVATE
 	KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS=1)
 
@@ -650,6 +673,7 @@ if(BUILD_TESTING)
 	add_test(NAME image_page_table COMMAND $<TARGET_FILE:image_page_table_tests>)
 	add_test(NAME memory_tracker COMMAND $<TARGET_FILE:memory_tracker_tests>)
 	add_test(NAME page_manager COMMAND $<TARGET_FILE:page_manager_tests>)
+	add_test(NAME build_flags COMMAND $<TARGET_FILE:build_flags_tests>)
 	add_test(NAME archive_file COMMAND $<TARGET_FILE:archive_file_tests>)
 	add_test(NAME trophy_system COMMAND $<TARGET_FILE:trophy_system_tests>)
 	add_test(NAME avplayer_file COMMAND $<TARGET_FILE:avplayer_file_tests>)
@@ -665,6 +689,7 @@ if(BUILD_TESTING)
 	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 ajm_resample COMMAND $<TARGET_FILE:ajm_resample_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 host_input COMMAND $<TARGET_FILE:host_input_tests>)
@@ -681,6 +706,16 @@ if(BUILD_TESTING)
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --tessellation-only)
 	add_test(NAME shader_recompiler_float_min_max
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --float-min-max-only)
+	add_test(NAME shader_recompiler_vop3_conversions
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --vop3-conversions-only)
+	add_test(NAME shader_recompiler_quadmask
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --quadmask-only)
+	add_test(NAME shader_recompiler_integer16_ternary
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --integer16-ternary-only)
+	add_test(NAME shader_recompiler_mad_u16
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --mad-u16-only)
+	add_test(NAME shader_recompiler_mul_hi24
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --mul-hi24-only)
 	add_test(NAME shader_recompiler_alignbyte
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --alignbyte-only)
 	add_test(NAME graphics_pipeline_rasterization
@@ -689,6 +724,10 @@ if(BUILD_TESTING)
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --draw-offset-only)
 	add_test(NAME virtual_memory_allocation
 		COMMAND $<TARGET_FILE:virtual_memory_allocation_tests>)
+	add_test(NAME trinity_mode
+		COMMAND $<TARGET_FILE:virtual_memory_allocation_tests> --trinity-only)
+	add_test(NAME psml COMMAND $<TARGET_FILE:psml_tests>)
+	add_test(NAME pssr COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --pssr-only)
 	add_test(NAME guest_red_zone_patcher
 		COMMAND $<TARGET_FILE:virtual_memory_allocation_tests> --red-zone-patcher-only)
 	add_test(NAME command_scheduler_timeline
@@ -697,8 +736,6 @@ if(BUILD_TESTING)
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --stream-buffer-only)
 	add_test(NAME gpu_command_lane
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --gpu-command-lane-only)
-	add_test(NAME wait_until_safe_for_rendering
-		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --wait-until-safe-for-rendering-only)
 	add_test(NAME pm4_flip
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --flip-only)
 	add_test(NAME depth_bits
@@ -743,6 +780,7 @@ if(BUILD_TESTING)
 		image_page_table_tests
 		memory_tracker_tests
 		page_manager_tests
+		build_flags_tests
 		archive_file_tests
 		trophy_system_tests
 		avplayer_file_tests
@@ -760,6 +798,7 @@ if(BUILD_TESTING)
 		audio_out2_port_tests
 		audio_out_timing_tests
 		ngs2_sampler_tests
+		ajm_resample_tests
 		pad_haptics_tests
 		controller_settings_tests
 		host_input_tests
@@ -811,6 +850,19 @@ add_custom_target(kyty_touch_cursors
 add_dependencies(kyty_emulator kyty_touch_cursors)
 install(DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/assets/cursors" DESTINATION assets)
 
+set(KYTY_NOTO_SANS_FONTS
+	"${KYTY_NOTO_SANS_DIR}/NotoSans-Regular.ttf"
+	"${KYTY_NOTO_SANS_DIR}/NotoSans-Bold.ttf"
+	"${KYTY_NOTO_SANS_DIR}/NotoSans-Italic.ttf"
+	"${KYTY_NOTO_SANS_DIR}/NotoSans-BoldItalic.ttf")
+add_custom_target(kyty_noto_sans_fonts
+	COMMAND ${CMAKE_COMMAND} -E make_directory "$<TARGET_FILE_DIR:kyty_emulator>/assets/fonts"
+	COMMAND ${CMAKE_COMMAND} -E copy_if_different ${KYTY_NOTO_SANS_FONTS}
+		"$<TARGET_FILE_DIR:kyty_emulator>/assets/fonts"
+	VERBATIM)
+add_dependencies(kyty_emulator kyty_noto_sans_fonts)
+install(FILES ${KYTY_NOTO_SANS_FONTS} DESTINATION assets/fonts)
+
 target_link_libraries(kyty_emulator ${kyty_emulator_link_libraries})
 if (WIN32)
 	target_link_libraries(kyty_emulator opengl32)

src/common/emulatorConfig.cpp

@@ -105,10 +105,18 @@ bool HideCursorEnabled() {
 	return g_config->hide_cursor_enabled;
 }
 
+bool StretchEnabled() {
+	return g_config->stretch_enabled;
+}
+
 bool VrEnabled() {
 	return g_config->vr_enabled;
 }
 
+bool Ps5ProEnabled() {
+	return g_config->pro_enabled;
+}
+
 bool AmdCpuEnabled() {
 	return g_config->amd_cpu_enabled;
 }
@@ -129,6 +137,10 @@ bool IsNetworkConnected() {
 	return g_config->network_connected;
 }
 
+bool PsnEmulationEnabled() {
+	return g_config->psn_emulation_enabled;
+}
+
 bool ShaderValidationEnabled() {
 	return g_config->shader_validation_enabled;
 }

src/common/emulatorConfig.h

@@ -64,10 +64,13 @@ struct ConfigOptions {
 	bool                   fullscreen_enabled          = false;
 	bool                   hdr_enabled                 = false;
 	bool                   hide_cursor_enabled         = false;
+	bool                   stretch_enabled             = false;
 	bool                   splash_enabled              = true;
 	bool                   vr_enabled                  = false;
+	bool                   pro_enabled                 = false;
 	bool                   amd_cpu_enabled             = false;
 	bool                   network_connected           = false;
+	bool                   psn_emulation_enabled       = false;
 	uint32_t               vblank_frequency            = 60;
 	uint32_t               console_language            = DEFAULT_CONSOLE_LANGUAGE;
 	bool                   vulkan_validation_enabled   = false;
@@ -125,9 +128,12 @@ int32_t GetGpuIndex();
 bool     FullscreenEnabled();
 bool     HdrEnabled();
 bool     HideCursorEnabled();
+bool     StretchEnabled();
 bool     VrEnabled();
+bool     Ps5ProEnabled();
 bool     AmdCpuEnabled();
 bool     IsNetworkConnected();
+bool     PsnEmulationEnabled();
 uint32_t GetVblankFrequency();
 uint32_t GetConsoleLanguage();
 bool     VulkanValidationEnabled();

src/emulator.cpp

@@ -99,10 +99,6 @@ static void Init(const Config::ConfigOptions& cfg, const std::filesystem::path&
 
 	if (Common::File::IsFileExisting(param_json)) {
 		Loader::SystemContentLoadParamSfo(param_json);
-		if (const auto flexible_memory_size = Loader::SystemContentGetFlexibleMemorySize();
-		    flexible_memory_size != 0) {
-			Libs::LibKernel::Memory::SetFlexibleMemorySize(flexible_memory_size);
-		}
 	}
 	if (cfg.splash_enabled) {
 		Libs::Graphics::InitializeSplashScreen(param_json.parent_path());

src/graphics/guest_gpu/command_processor/commandProcessor.h

@@ -3,6 +3,7 @@
 
 #include "common/assert.h"
 #include "graphics/guest_gpu/hardwareContext.h"
+#include "graphics/guest_gpu/pssr.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
 
@@ -92,6 +93,7 @@ public:
 	                    uint32_t mode);
 	void DispatchIndirect(uint64_t args_addr, uint32_t mode);
 	void WaitFlipDone(uint32_t video_out_handle, uint32_t display_buffer_index);
+	void ProcessPssr(const Pssr::Request& request);
 	void TriggerEvent(uint32_t event_type, uint32_t event_index, uint64_t event_address = 0);
 
 	void WaitForRewind(bool valid);

src/graphics/guest_gpu/command_processor/pm4Dispatch.cpp

@@ -226,6 +226,7 @@ constexpr auto MakeCustomOpcodeDispatchTable() {
 	g_cp_op_custom_func[Pm4::R_POP_MARKER]     = CpOpPopMarker;
 	g_cp_op_custom_func[Pm4::R_FLIP]           = CpOpFlip;
 	g_cp_op_custom_func[Pm4::R_CONTEXT_STATE]  = CpOpContextState;
+	g_cp_op_custom_func[Pm4::R_PSSR]           = CpOpPssr;
 
 	return g_cp_op_custom_func;
 }

src/graphics/guest_gpu/command_processor/pm4Dispatch.h

@@ -151,6 +151,7 @@ uint32_t CpOpPushMarker(CommandProcessor&, uint32_t, const uint32_t*, uint32_t,
 uint32_t CpOpPopMarker(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpWaitRegMem64(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 uint32_t CpOpFlip(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
+uint32_t CpOpPssr(CommandProcessor&, uint32_t, const uint32_t*, uint32_t, uint32_t);
 
 } // namespace Libs::Graphics
 

src/graphics/guest_gpu/command_processor/pm4Handlers.cpp

@@ -1706,6 +1706,17 @@ KYTY_CP_OP_PARSER(CpOpFlip) {
 	return 5;
 }
 
+KYTY_CP_OP_PARSER(CpOpPssr) {
+	KYTY_PROFILER_FUNCTION();
+
+	const auto packet_dw = KYTY_PM4_LEN(cmd_id) - 1;
+	EXIT_NOT_IMPLEMENTED(packet_dw < Pssr::RequestDwords || packet_dw > dw);
+	Pssr::Request request {};
+	std::memcpy(&request, buffer, sizeof(request));
+	cp.ProcessPssr(request);
+	return packet_dw;
+}
+
 KYTY_CP_OP_PARSER(CpOpIndexType) {
 	KYTY_PROFILER_FUNCTION();
 

src/graphics/guest_gpu/gpu_format.cpp

@@ -46,6 +46,7 @@ constexpr FormatInfo kFormatInfo[] = {
 	{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::k2_10_10_10UNorm, 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},

src/graphics/guest_gpu/graphicsRun.cpp

@@ -30,7 +30,6 @@
 #include <mutex>
 #include <semaphore>
 #include <thread>
-#include <utility>
 #include <vector>
 
 namespace Libs::Graphics {
@@ -92,12 +91,6 @@ void GuestGpu::SendCommand(Common::UniqueFunction<void>&& command) {
 	m_work_available.Signal();
 }
 
-void GuestGpu::Wake() {
-	Common::LockGuard lock(m_queue_mutex);
-	m_wake_pending = true;
-	m_work_available.Signal();
-}
-
 void GuestGpu::ProcessCommands() {
 	EXIT_IF(!IsGpuThread());
 	while (m_pending_commands.load(std::memory_order_acquire) != 0) {
@@ -443,10 +436,7 @@ void GuestGpu::ThreadRun(void* data) {
 				}
 				if (selected_queue < 0) {
 					gpu->m_processing = false;
-					// A wake that came before this wait must not be lost.
-					if (!std::exchange(gpu->m_wake_pending, false)) {
-						gpu->m_work_available.WaitFor(&gpu->m_queue_mutex, 100);
-					}
+					gpu->m_work_available.WaitFor(&gpu->m_queue_mutex, 100);
 					for (auto& queue: gpu->m_queues) {
 						if (!queue.empty()) {
 							queue.front().blocked = false;
@@ -814,6 +804,27 @@ static void PatchIndirectDrawOffsets(CommandProcessor& cp, IndirectDrawRegisters
 	write(registers.draw_index, draw_index);
 }
 
+// Writes the GS user-data slots the packet patches; false when the shaders cannot read the
+// offsets from GPU records: a first-index or subdraw-index destination, or another stage's
+// user data.
+static bool IndirectDrawOffsets(IndirectDrawRegisters registers, ShaderIndirectOffsets& offsets) {
+	if (registers.index_offset != Pm4::SH_NOP || registers.draw_index != Pm4::SH_NOP) {
+		return false;
+	}
+	for (const auto [location, slot]: {std::pair {registers.vertex_offset, &offsets.vertex},
+	                                   std::pair {registers.instance_offset, &offsets.instance}}) {
+		if (location == Pm4::SH_NOP) {
+			continue;
+		}
+		const uint32_t index = location - Pm4::SPI_SHADER_USER_DATA_GS_0;
+		if (index >= HW::UserSgprInfo::SGPRS_MAX) {
+			return false;
+		}
+		*slot = static_cast<uint8_t>(index);
+	}
+	return true;
+}
+
 void CommandProcessor::DrawIndirect(uint32_t data_offset, IndirectDrawRegisters registers,
                                     uint32_t draw_initiator, bool indexed) {
 	const auto args_size =
@@ -845,13 +856,10 @@ void CommandProcessor::DrawIndirectMulti(uint32_t data_offset, uint32_t max_coun
 	// NUM_INSTANCES keeps the last draw's instance count.
 	m_num_instances_addr = args_addr + static_cast<uint64_t>(draw_count - 1) * stride_in_bytes +
 	                       offsetof(vk::DrawIndirectCommand, instanceCount);
-	// With no SGPRs to patch, the GPU reads the records if the shaders allow. Shaders apply
-	// the records' offsets only through those SGPRs; Vulkan applies them, which matches
-	// while they are zero.
-	const DrawIndirectArgs records {args_addr, draw_count, stride_in_bytes};
-	if (registers.vertex_offset == Pm4::SH_NOP && registers.instance_offset == Pm4::SH_NOP &&
-	    registers.index_offset == Pm4::SH_NOP && registers.draw_index == Pm4::SH_NOP &&
-	    (!indexed || m_index_buffer_size != 0) &&
+	// The GPU reads the records if the shaders allow; the VS prolog removes Vulkan's bases
+	// from the index VGPRs and copies them only to the user SGPRs the packet patches.
+	DrawIndirectArgs records {args_addr, draw_count, stride_in_bytes};
+	if (IndirectDrawOffsets(registers, records.offsets) && (!indexed || m_index_buffer_size != 0) &&
 	    (indexed ? executor.DrawIndex(m_submit_id, CurrentBuffer(),
 	                                  {.index_count         = m_index_buffer_size,
 	                                   .index_addr          = IndexAddress(0),
@@ -921,6 +929,11 @@ void CommandProcessor::DispatchDirect(uint32_t thread_group_x, uint32_t thread_g
 	}
 }
 
+void CommandProcessor::ProcessPssr(const Pssr::Request& request) {
+	EXIT_NOT_IMPLEMENTED(IsAsyncComputeQueue());
+	m_renderer.GetRenderExecutor().ProcessPssr(CurrentBuffer(), request);
+}
+
 void CommandProcessor::DispatchIndirect(uint64_t args_addr, uint32_t mode) {
 	EXIT_NOT_IMPLEMENTED(args_addr == 0 || (args_addr & 3u) != 0);
 	if ((mode & Pm4::COMPUTE_DISPATCH_INITIATOR_USE_THREAD_DIMENSIONS) != 0) {
@@ -941,11 +954,10 @@ void CommandProcessor::DrawIndexAuto(DrawAutoArgs args) {
 }
 
 void CommandProcessor::WaitFlipDone(uint32_t video_out_handle, uint32_t display_buffer_index) {
-	// Stall only this command processor, not the host command lane or other queues.
-	if (!m_renderer.GetVideoOut().IsSafeForRendering(static_cast<int>(video_out_handle),
-	                                                 static_cast<int>(display_buffer_index))) {
-		SuspendPm4();
-	}
+	BufferFlush();
+
+	m_renderer.GetVideoOut().WaitFlipDone(static_cast<int>(video_out_handle),
+	                                      static_cast<int>(display_buffer_index));
 }
 
 void CommandProcessor::ReleaseMem(uint32_t data_sel, uint32_t interrupt_selector, void* address,
@@ -976,16 +988,28 @@ void CommandProcessor::ReleaseMem(uint32_t data_sel, uint32_t interrupt_selector
 	// Dumped occlusion counters are written before any end-of-pipe signal reaches the guest.
 	m_renderer.GetOcclusionQueries().Publish();
 
+	// A label beside GPU-written bytes goes to the backing store, so they are not downloaded.
+	// Other labels go through the guest mapping: its fault keeps a clean cached copy current and
+	// downloads GPU-written bytes under the label before the store.
+	const auto write_label = [this, address](const void* value, uint64_t size) {
+		const auto vaddr   = reinterpret_cast<uint64_t>(address);
+		auto&      buffers = m_renderer.GetBufferCache();
+		if (!GuestRange {vaddr, size}.Valid() || !buffers.IsRegionGpuModified(vaddr, size) ||
+		    buffers.HasGpuDirtyBytes(vaddr, size) ||
+		    !LibKernel::Memory::TryWriteBacking(vaddr, value, size)) {
+			std::memcpy(address, value, size);
+		}
+	};
 	// EndOfPipeData: kNone, kImmediate32b, kImmediate64b, kRefClk; 5 reads GDS.
 	switch (data_sel) {
 		case 0: break;
 		case 1: {
 			const auto value = static_cast<uint32_t>(data);
-			std::memcpy(address, &value, sizeof(value));
+			write_label(&value, sizeof(value));
 			break;
 		}
 		case 3: data = Sync::ReadReferenceClock(); [[fallthrough]];
-		case 2: std::memcpy(address, &data, sizeof(data)); break;
+		case 2: write_label(&data, sizeof(data)); break;
 		case 5: {
 			GetScheduler().Finish();
 			const auto gds = static_cast<uint32_t>(data);

src/graphics/guest_gpu/graphicsRun.h

@@ -31,8 +31,6 @@ public:
 	[[nodiscard]] bool IsStopping();
 	void               SendCommand(Common::UniqueFunction<void>&& command);
 	void               SendCommandSync(Common::UniqueFunction<void>&& command);
-	// Retries blocked submissions now instead of after the 100 ms stall timeout.
-	void               Wake();
 
 	// Submitted command memory is borrowed and must remain valid until GPU execution completes.
 	void Submit(std::span<const uint32_t> commands);
@@ -84,7 +82,6 @@ private:
 	bool                                           m_processing        = false;
 	bool                                           m_accepting         = true;
 	bool                                           m_stopping          = false;
-	bool                                           m_wake_pending      = false;
 	bool                                           m_shutdown_complete = false;
 	// Completion callbacks can outlive GuestGpu during renderer shutdown.
 	std::shared_ptr<std::binary_semaphore> m_suspend_point_ready =

src/graphics/guest_gpu/pm4.h

@@ -81,6 +81,7 @@ constexpr uint32_t R_WRITE_DATA     = 0x15;
 constexpr uint32_t R_FLIP           = 0x17;
 constexpr uint32_t R_DMA_DATA       = 0x19;
 constexpr uint32_t R_CONTEXT_STATE  = 0x1A;
+constexpr uint32_t R_PSSR           = 0x1B;
 
 constexpr uint32_t R_NUM = 0x3F + 1;
 

src/graphics/guest_gpu/pssr.h

@@ -0,0 +1,20 @@
+#ifndef EMULATOR_SRC_GRAPHICS_GUEST_GPU_PSSR_H_
+#define EMULATOR_SRC_GRAPHICS_GUEST_GPU_PSSR_H_
+
+#include <cstdint>
+
+namespace Libs::Graphics::Pssr {
+
+struct Request {
+	uint32_t output[8];
+	uint32_t input[8];
+	uint32_t width;
+	uint32_t height;
+};
+
+constexpr uint32_t RequestDwords = sizeof(Request) / sizeof(uint32_t);
+static_assert(sizeof(Request) == 72);
+
+} // namespace Libs::Graphics::Pssr
+
+#endif // EMULATOR_SRC_GRAPHICS_GUEST_GPU_PSSR_H_

src/graphics/host_gpu/rangeSet.h

@@ -5,6 +5,7 @@
 
 #include <algorithm>
 #include <cstdint>
+#include <iterator>
 #include <map>
 
 namespace Libs::Graphics {
@@ -17,14 +18,19 @@ public:
 		if (it != m_ranges.begin() && std::prev(it)->second >= address) {
 			it = std::prev(it);
 		}
-		uint64_t begin = address;
-		uint64_t last  = end;
-		while (it != m_ranges.end() && it->first <= last) {
-			begin = std::min(begin, it->first);
-			last  = std::max(last, it->second);
-			it    = m_ranges.erase(it);
+		if (it == m_ranges.end() || it->first > address) {
+			it = m_ranges.emplace_hint(it, address, end);
+		} else {
+			if (it->second >= end) {
+				return;
+			}
+			it->second = end;
+		}
+		auto next = std::next(it);
+		while (next != m_ranges.end() && next->first <= it->second) {
+			it->second = std::max(it->second, next->second);
+			next       = m_ranges.erase(next);
 		}
-		m_ranges.emplace(begin, last);
 	}
 
 	void Subtract(uint64_t address, uint64_t size) {

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

@@ -488,16 +488,15 @@ void TextureCache::RefreshCopySource(ImageId id) {
 }
 
 bool TextureCache::CopyD16(Image& destination, Image& source) {
-	const bool source_depth      = source.info.IsDepth();
-	const bool destination_depth = destination.info.IsDepth();
-	if (source_depth == destination_depth) {
+	// Converts guest D16 between a four-byte host depth aspect and two-byte color or D16 depth.
+	const bool source_wide = DepthAspectTransferBytes(source.backing.format) == sizeof(uint32_t);
+	if (source_wide == (DepthAspectTransferBytes(destination.backing.format) == sizeof(uint32_t))) {
 		return false;
 	}
-	auto&      depth          = source_depth ? source : destination;
-	auto&      color          = source_depth ? destination : source;
-	const auto transfer_bytes = DepthAspectTransferBytes(depth.backing.format);
+	auto& depth  = source_wide ? source : destination;
+	auto& narrow = source_wide ? destination : source;
 	if (depth.info.bytes_per_block != sizeof(uint16_t) ||
-	    color.info.bytes_per_block != sizeof(uint16_t) || transfer_bytes != sizeof(uint32_t)) {
+	    narrow.info.bytes_per_block != sizeof(uint16_t)) {
 		return false;
 	}
 	EXIT_IF(source.backing.samples != 1 || destination.backing.samples != 1 ||
@@ -507,42 +506,42 @@ bool TextureCache::CopyD16(Image& destination, Image& source) {
 
 	const auto     layers = depth.info.resources.layers;
 	const uint64_t depth_slice =
-	    static_cast<uint64_t>(depth.info.pitch) * depth.info.extent.height * transfer_bytes;
-	const uint64_t color_slice =
-	    static_cast<uint64_t>(color.info.pitch) * color.info.extent.height * sizeof(uint16_t);
-	EXIT_IF(layers == 0 || depth_slice > UINT64_MAX / layers || color_slice > UINT64_MAX / layers);
-	const auto                       depth_size = depth_slice * layers;
-	const auto                       color_size = color_slice * layers;
-	const std::array                 depth_copies {
+	    static_cast<uint64_t>(depth.info.pitch) * depth.info.extent.height * sizeof(uint32_t);
+	const uint64_t narrow_slice =
+	    static_cast<uint64_t>(narrow.info.pitch) * narrow.info.extent.height * sizeof(uint16_t);
+	EXIT_IF(layers == 0 || depth_slice > UINT64_MAX / layers || narrow_slice > UINT64_MAX / layers);
+	const std::array depth_copies {
 	    vk::BufferImageCopy {0,
 	                         depth.info.pitch,
 	                         depth.info.extent.height,
 	                         {vk::ImageAspectFlagBits::eDepth, 0, 0, layers},
 	                         {},
 	                         depth.info.extent}};
-	const std::array color_copies {
+	const std::array narrow_copies {
 	    vk::BufferImageCopy {0,
-	                         color.info.pitch,
-	                         color.info.extent.height,
-	                         {vk::ImageAspectFlagBits::eColor, 0, 0, layers},
+	                         narrow.info.pitch,
+	                         narrow.info.extent.height,
+	                         {narrow.info.IsDepth() ? vk::ImageAspectFlagBits::eDepth
+	                                                : vk::ImageAspectFlagBits::eColor,
+	                          0, 0, layers},
 	                         {},
-	                         color.info.extent}};
+	                         narrow.info.extent}};
 
-	auto                         depth_buffer = m_tiler.GetScratchBuffer(depth_size);
-	auto                         color_buffer = m_tiler.GetScratchBuffer(color_size, depth_buffer.buffer);
+	auto depth_buffer  = m_tiler.GetScratchBuffer(depth_slice * layers);
+	auto narrow_buffer = m_tiler.GetScratchBuffer(narrow_slice * layers, depth_buffer.buffer);
 	const TileManager::D16Layout promote_layout {
 	    .width               = depth.info.extent.width,
 	    .height              = depth.info.extent.height,
 	    .layers              = layers,
-	    .source_row_stride   = static_cast<uint64_t>(color.info.pitch) * sizeof(uint16_t),
-	    .target_row_stride   = static_cast<uint64_t>(depth.info.pitch) * transfer_bytes,
-	    .source_slice_stride = color_slice,
+	    .source_row_stride   = static_cast<uint64_t>(narrow.info.pitch) * sizeof(uint16_t),
+	    .target_row_stride   = static_cast<uint64_t>(depth.info.pitch) * sizeof(uint32_t),
+	    .source_slice_stride = narrow_slice,
 	    .target_slice_stride = depth_slice,
 	};
 	const bool d32 = DepthAspectTransferFormat(depth.backing.format) == vk::Format::eD32Sfloat;
-	if (source_depth) {
+	if (source_wide) {
 		source.Download(depth_copies, depth_buffer.buffer, depth_buffer.offset, depth_buffer.size);
-		m_tiler.ConvertD16(depth_buffer, color_buffer, TileManager::D16Direction::Demote, d32,
+		m_tiler.ConvertD16(depth_buffer, narrow_buffer, TileManager::D16Direction::Demote, d32,
 		                   {.width               = promote_layout.width,
 		                    .height              = promote_layout.height,
 		                    .layers              = promote_layout.layers,
@@ -550,11 +549,12 @@ bool TextureCache::CopyD16(Image& destination, Image& source) {
 		                    .target_row_stride   = promote_layout.source_row_stride,
 		                    .source_slice_stride = promote_layout.target_slice_stride,
 		                    .target_slice_stride = promote_layout.source_slice_stride});
-		destination.Upload(color_copies, color_buffer.buffer, color_buffer.offset,
-		                   color_buffer.size);
+		destination.Upload(narrow_copies, narrow_buffer.buffer, narrow_buffer.offset,
+		                   narrow_buffer.size);
 	} else {
-		source.Download(color_copies, color_buffer.buffer, color_buffer.offset, color_buffer.size);
-		m_tiler.ConvertD16(color_buffer, depth_buffer, TileManager::D16Direction::Promote, d32,
+		source.Download(narrow_copies, narrow_buffer.buffer, narrow_buffer.offset,
+		                narrow_buffer.size);
+		m_tiler.ConvertD16(narrow_buffer, depth_buffer, TileManager::D16Direction::Promote, d32,
 		                   promote_layout);
 		destination.Upload(depth_copies, depth_buffer.buffer, depth_buffer.offset,
 		                   depth_buffer.size);

src/graphics/host_gpu/renderer/dynamicState.cpp

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

src/graphics/host_gpu/renderer/dynamicState.h

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

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

@@ -17,9 +17,15 @@
 
 namespace Libs::Graphics {
 
-enum class ColorTransform : uint32_t { None, SwapBgra16, Reverse10_11_11 };
+enum class ColorTransform : uint32_t { None, SwapBgra16, Reverse10_11_11, Rotate2_10_10_10 };
 
-enum class VideoOutCompression : uint8_t { Uncompressed, Dcc256_256_0, Dcc256_64_64, Unsupported };
+enum class VideoOutCompression : uint8_t {
+	Uncompressed,
+	Dcc256_256_0,
+	Dcc256_64_64,
+	Dcc128_128_128,
+	Unsupported
+};
 
 enum class ImageMetadataKind : uint8_t { None, Htile, Dcc, Cmask };
 
@@ -99,6 +105,8 @@ struct ImageInfo {
 		switch (guest_format) {
 			case Prospero::BufferFormat::k10_11_11Float:
 				return ColorTransform::Reverse10_11_11;
+			case Prospero::BufferFormat::k2_10_10_10UNorm:
+				return ColorTransform::Rotate2_10_10_10;
 			default: return ColorTransform::None;
 		}
 	}
@@ -287,8 +295,9 @@ inline bool ImageInfo::IsDepth() const noexcept {
 [[nodiscard]] inline VideoOutCompression
 ClassifyVideoOutCompression(bool compressed, uint64_t metadata_address, uint32_t dcc_control,
                             uint64_t dcc_clear_color) noexcept {
-	constexpr uint32_t DCC_256_256_0 = 0x00000048u;
-	constexpr uint32_t DCC_256_64_64 = 0x00000208u;
+	constexpr uint32_t DCC_256_256_0   = 0x00000048u;
+	constexpr uint32_t DCC_256_64_64   = 0x00000208u;
+	constexpr uint32_t DCC_128_128_128 = 0x00100024u;
 	if (!compressed) {
 		return metadata_address == 0 && dcc_control == 0 && dcc_clear_color == 0
 		           ? VideoOutCompression::Uncompressed
@@ -300,6 +309,7 @@ ClassifyVideoOutCompression(bool compressed, uint64_t metadata_address, uint32_t
 	switch (dcc_control) {
 		case DCC_256_256_0: return VideoOutCompression::Dcc256_256_0;
 		case DCC_256_64_64: return VideoOutCompression::Dcc256_64_64;
+		case DCC_128_128_128: return VideoOutCompression::Dcc128_128_128;
 		default: return VideoOutCompression::Unsupported;
 	}
 }

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

@@ -32,8 +32,12 @@ 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};
+	switch (guest_format) {
+		case Prospero::BufferFormat::k10_11_11Float:
+			return {vk::Format::eB10G11R11UfloatPack32, Prospero::ColorMappingBgra};
+		case Prospero::BufferFormat::k2_10_10_10UNorm:
+			return {vk::Format::eA2R10G10B10UnormPack32, Prospero::ColorMappingAbgr};
+		default: break;
 	}
 	if (order == Prospero::ChannelOrder::kAlt) {
 		switch (guest_format) {

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

@@ -104,7 +104,8 @@ TileManager::Tiling TileManager::Prepare(bool tile, uint64_t tiled_capacity,
 	    block.block_depth == 0 || block.block_size == 0 || info.tiling.texel_width == 0 ||
 	    info.tiling.texel_height == 0 || info.tiled_slice_stride == 0);
 	EXIT_IF((transform == ColorTransform::SwapBgra16 && block.bytes_per_element != 8u) ||
-	        (transform == ColorTransform::Reverse10_11_11 && block.bytes_per_element != 4u));
+	        ((transform == ColorTransform::Reverse10_11_11 ||
+	          transform == ColorTransform::Rotate2_10_10_10) && block.bytes_per_element != 4u));
 	const auto checked_multiply = [](uint64_t left, uint64_t right, uint64_t& result) {
 		return (left == 0 || right <= UINT64_MAX / left) && (result = left * right, true);
 	};
@@ -484,7 +485,7 @@ void TileManager::ConvertD16(Result source, Result target, D16Direction directio
 	                            vk::AccessFlagBits::eHostWrite |
 	                            vk::AccessFlagBits::eTransferWrite |
 	                            vk::AccessFlagBits::eShaderWrite;
-	barriers[1].dstAccessMask = vk::AccessFlagBits::eShaderRead | vk::AccessFlagBits::eShaderWrite;
+	barriers[1].dstAccessMask = vk::AccessFlagBits::eShaderWrite;
 	command.pipelineBarrier(
 	    vk::PipelineStageFlagBits::eAllCommands | vk::PipelineStageFlagBits::eHost,
 	    vk::PipelineStageFlagBits::eComputeShader, {}, 0, nullptr, 2, barriers, 0, nullptr);
@@ -560,7 +561,8 @@ void TileManager::TransformColor(Result input, Result output, ColorTransform tra
 	uint32_t element_bytes = 0;
 	switch (transform) {
 		case ColorTransform::SwapBgra16: element_bytes = 8; break;
-		case ColorTransform::Reverse10_11_11: element_bytes = 4; break;
+		case ColorTransform::Reverse10_11_11:
+		case ColorTransform::Rotate2_10_10_10: element_bytes = 4; break;
 		default: EXIT("invalid color transform\n");
 	}
 	EXIT_IF(input.size == 0 || input.size % element_bytes != 0 ||

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

@@ -20,7 +20,7 @@ namespace {
 
 namespace IR = ShaderRecompiler::IR;
 
-// Smaller tables are cheap to materialize on each draw and are often transient.
+// Smaller tables are often written for one draw, so they enter on a second draw of a frame.
 constexpr uint64_t MinTableBytes = 64ull * 1024ull;
 constexpr uint32_t ElementMask   = (1u << IR::ImageHeapKindShift) - 1u;
 
@@ -133,12 +133,16 @@ ImageHeap::~ImageHeap() {
 bool ImageHeap::Acquire(const IR::ImageHeapRequest& request, IR::ImageHeapView& result) {
 	KYTY_PROFILER_FUNCTION();
 	const auto size = request.size & ~uint64_t {3};
-	if (size < MinTableBytes || request.count == 0u || request.step == 0u ||
+	if (request.count == 0u || request.step == 0u ||
 	    ((request.base | request.first | request.step) & 3u) != 0u) {
 		return false;
 	}
 	auto memory = std::ranges::find_if(
 	    m_memories, [&](const auto& candidate) { return candidate->base == request.base; });
+	if (memory == m_memories.end() && size < MinTableBytes &&
+	    m_sightings.try_emplace(request.base, m_draw).first->second == m_draw) {
+		return false;
+	}
 	// A table follows its descriptor's size, as the first program of a draw acquires it.
 	if (memory == m_memories.end() || ((*memory)->size != size && (*memory)->draw != m_draw)) {
 		if (!m_context.IsMapped(request.base, size)) return false;
@@ -593,26 +597,42 @@ void ImageHeap::UnmapMemory(uint64_t address, uint64_t size) {
 	} else {
 		m_unmapped.push_back({address, size});
 	}
-	auto&           scheduler = m_context.GetCommandScheduler();
 	std::scoped_lock lock {m_mutex};
 	std::erase_if(m_memories, [&](const std::unique_ptr<Memory>& memory) {
 		if (address >= memory->base + memory->size || memory->base >= address + size) {
 			return false;
 		}
-		Unwatch(*memory);
-		for (auto& view: memory->views) {
-			std::erase(m_active, view.get());
-			for (const auto& [value, index]: view->elements) {
-				Unreference(m_elements[index]);
-				view->released.push_back(index);
-			}
-			Retire(*view);
-			if (scheduler.Active()) {
-				scheduler.DeferOperation([buffer = std::move(view->buffer)] {});
-			}
-		}
+		Drop(*memory);
 		return true;
 	});
 }
 
+void ImageHeap::RunGarbageCollector() {
+	// A smaller table rebuilt each frame moves to a new base. Larger ones keep their dead words.
+	std::scoped_lock lock {m_mutex};
+	std::erase_if(m_memories, [&](const std::unique_ptr<Memory>& memory) {
+		if (memory->size >= MinTableBytes || memory->draw > m_suspend_draw) return false;
+		Drop(*memory);
+		return true;
+	});
+	m_suspend_draw = m_draw;
+	m_sightings.clear();
+}
+
+void ImageHeap::Drop(Memory& memory) {
+	auto& scheduler = m_context.GetCommandScheduler();
+	Unwatch(memory);
+	for (auto& view: memory.views) {
+		std::erase(m_active, view.get());
+		for (const auto& [value, index]: view->elements) {
+			Unreference(m_elements[index]);
+			view->released.push_back(index);
+		}
+		Retire(*view);
+		if (scheduler.Active()) {
+			scheduler.DeferOperation([buffer = std::move(view->buffer)] {});
+		}
+	}
+}
+
 } // namespace Libs::Graphics

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

@@ -26,7 +26,7 @@ struct ImageHeapTestAccess;
 // of a table owns device slot words, one per table dword, that select an array element. CPU
 // writes to a table fault in its snapshot, which each draw refreshes once; a view takes the
 // changes when a draw first acquires it. Changes to the images behind elements bump
-// Image::heap_epoch.
+// Image::heap_epoch. A smaller table leaves at a suspend point if the frame did not acquire it.
 class ImageHeap final: public ShaderRecompiler::IR::ImageHeapSource {
 public:
 	ImageHeap(RenderContext& context, PageManager& page_manager);
@@ -47,6 +47,8 @@ public:
 	// CPU write fault, on any thread.
 	void InvalidateMemory(uint64_t address, uint64_t size);
 	void UnmapMemory(uint64_t address, uint64_t size);
+	// At a suspend point.
+	void RunGarbageCollector();
 
 private:
 	using DescriptorValue = ShaderRecompiler::IR::DescriptorValue;
@@ -107,6 +109,7 @@ private:
 	template <bool track>
 	void     Watch(const Memory& memory, size_t begin, size_t end);
 	void     Unwatch(const Memory& memory);
+	void     Drop(Memory& memory);
 	void     UpdateProbes(View& view, uint64_t begin, uint64_t end);
 	void     UpdateProbe(View& view, uint64_t probe);
 	void     Assign(View& view, size_t slot, uint32_t entry);
@@ -125,9 +128,11 @@ private:
 	std::mutex                           m_mutex; // Fault-visible Memory state and m_memories.
 	std::vector<std::unique_ptr<Memory>> m_memories;
 	std::vector<View*>                   m_active; // Acquired since the last commit.
-	uint64_t                             m_draw = 1;
+	uint64_t                             m_draw         = 1;
+	uint64_t                             m_suspend_draw = 0; // m_draw at the last suspend point.
 	std::vector<Element>                 m_elements;
 	std::vector<uint32_t>                m_free;
+	std::unordered_map<uint64_t, uint64_t>    m_sightings; // Smaller tables: base, first draw.
 	std::deque<std::pair<uint64_t, uint32_t>> m_retired; // Tick, element.
 	std::vector<std::pair<size_t, size_t>>    m_runs;
 	std::vector<uint32_t>                     m_scratch;

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

@@ -687,6 +687,10 @@ void PipelineCache::GetGraphicsPrograms(const CommandBuffer& command, DrawRender
 		vertex_params[0] = PrepareProgram(vertex_regs, context, user_config, vertex_info[0]);
 	}
 	const bool mesh_active = vertex_info[0].logical_stage == ShaderType::Mesh;
+	if (!tess_active && !mesh_active) {
+		vertex_info[0].indirect_offsets = state.indirect_offsets;
+		vertex_info[0].indirect_draw    = state.indirect_draw;
+	}
 	if (object_primitive_id) {
 		// CxObjectIdMode: register source plus a 32-bit pipelined primitive-ID addition.
 		EXIT_NOT_IMPLEMENTED(sh.m_paClObjprimIdCntl != 2u || mesh_active || tess_active ||
@@ -1003,15 +1007,11 @@ PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(const CommandBuffer&
 	}
 	static_params.alpha_to_coverage_enable = ps_active && ctx.GetShaderRegisters().AlphaToMask();
 	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;
-	static_params.face       = mc.face;
 	static_params.provoking_vtx_last = mc.provoking_vtx_last;
 	static_params.polygon_mode =
-	    ResolvePolygonMode(mc, static_params.cull_front, static_params.cull_back);
+	    ResolvePolygonMode(mc, !rect_list && mc.cull_front, !rect_list && mc.cull_back);
 	if (tessellation) {
-		static_params.patch_control_points =
-		    static_cast<uint8_t>(vs_input_info.tess.input_control_points);
+		static_params.patch_control_points = vs_input_info.tess.input_control_points;
 	} else if (!mesh && static_params.topology == vk::PrimitiveTopology::ePatchList) {
 		// Rectangle lists expand through generated tessellation stages.
 		static_params.patch_control_points = 3;

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

@@ -54,15 +54,12 @@ struct PipelineStaticParameters {
 	bool                       depth_clip_enable        = true;
 	vk::PrimitiveTopology      topology                 = vk::PrimitiveTopology::ePointList;
 	// 0 without tessellation, 3 for a rectangle list, else the input control points.
-	uint8_t                    patch_control_points     = 0;
+	uint32_t                   patch_control_points     = 0;
 	bool                       primitive_restart_enable = false;
 	uint32_t                   samples                  = 1;
 	bool                       sample_shading_enable    = false;
 	bool                       alpha_to_coverage_enable = false;
 	uint32_t                   color_mask[RENDER_COLOR_ATTACHMENTS_MAX]           = {};
-	bool                       cull_front                                         = false;
-	bool                       cull_back                                          = false;
-	bool                       face                                               = false;
 	bool                       provoking_vtx_last                                 = false;
 	vk::PolygonMode            polygon_mode                                       = vk::PolygonMode::eFill;
 	uint8_t                    color_srcblend[RENDER_COLOR_ATTACHMENTS_MAX]       = {};
@@ -140,6 +137,9 @@ struct DrawRenderState {
 	bool                                 ps_active                                = false;
 	std::array<ShaderVertexInputInfo, 3> vertex_info;
 	ShaderPixelInputInfo                 ps_input_info;
+	// Input: the user-data slots a GPU-read indirect draw patches.
+	ShaderIndirectOffsets                indirect_offsets;
+	bool                                 indirect_draw = false;
 
 	[[nodiscard]] std::span<const ShaderVertexInputInfo> VertexStages() const {
 		return {vertex_info.data(), vertex_info[1].stage ? 3u : 1u};

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

@@ -219,17 +219,6 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	vk::PipelineViewportStateCreateInfo viewport_state {};
 	viewport_state.pNext = &depth_clip_control;
 
-	vk::CullModeFlags cull_mode = vk::CullModeFlagBits::eNone;
-	if (static_params.cull_back) {
-		cull_mode |= vk::CullModeFlagBits::eBack;
-	}
-	if (static_params.cull_front) {
-		cull_mode |= vk::CullModeFlagBits::eFront;
-	}
-
-	vk::FrontFace front_face =
-	    (static_params.face ? vk::FrontFace::eClockwise : vk::FrontFace::eCounterClockwise);
-
 	vk::PipelineRasterizationDepthClipStateCreateInfoEXT clip_ext {};
 	clip_ext.depthClipEnable = static_params.depth_clip_enable ? VK_TRUE : VK_FALSE;
 
@@ -250,8 +239,6 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 		provoking_vertex.pNext = rasterizer.pNext;
 		rasterizer.pNext = &provoking_vertex;
 	}
-	rasterizer.cullMode  = cull_mode;
-	rasterizer.frontFace = front_face;
 	rasterizer.polygonMode = static_params.polygon_mode;
 	rasterizer.lineWidth = 1.0f;
 
@@ -347,6 +334,8 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	    vk::DynamicState::eViewportWithCount,
 	    vk::DynamicState::eScissorWithCount,
 	    vk::DynamicState::eLineWidth,
+	    vk::DynamicState::eCullMode,
+	    vk::DynamicState::eFrontFace,
 	    vk::DynamicState::eDepthTestEnable,
 	    vk::DynamicState::eDepthWriteEnable,
 	    vk::DynamicState::eDepthCompareOp,

src/graphics/host_gpu/renderer/render.h

@@ -45,9 +45,10 @@ enum class DrawOffsetSource : uint8_t {
 
 // Nonzero address: the GPU reads count Vulkan indirect draw records, stride bytes apart.
 struct DrawIndirectArgs {
-	uint64_t address = 0;
-	uint32_t count   = 0;
-	uint32_t stride  = 0;
+	uint64_t              address = 0;
+	uint32_t              count   = 0;
+	uint32_t              stride  = 0;
+	ShaderIndirectOffsets offsets;
 };
 
 struct DrawIndexArgs {
@@ -149,15 +150,21 @@ private:
 	friend class CommandScheduler;
 };
 
+namespace Pssr {
+struct Request;
+}
+
 class RenderExecutor {
 public:
 	explicit RenderExecutor(RenderContext& context): m_context(context) {}
+	~RenderExecutor();
 	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, bool async_compute);
 	void DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer, uint64_t args_addr,
 	                      uint32_t mode, bool async_compute);
+	void ProcessPssr(CommandBuffer& buffer, const Pssr::Request& request);
 
 	void PrepareBindings(const ShaderStageRuntime& runtime, PreparedBindings& prepared);
 	void                           FindBuffers(std::span<PreparedBindings* const> stages);
@@ -218,6 +225,9 @@ private:
 	std::vector<ImageId>                  m_bound_images;
 	std::vector<vk::DescriptorImageInfo>  m_descriptor_images;
 	std::vector<vk::WriteDescriptorSet>   m_descriptor_writes;
+	vk::DescriptorSetLayout               m_pssr_descriptor_layout = nullptr;
+	vk::PipelineLayout                    m_pssr_pipeline_layout   = nullptr;
+	vk::Pipeline                          m_pssr_pipeline          = nullptr;
 
 	friend class CommandProcessor;
 	friend struct RenderExecutorTestAccess;

src/graphics/host_gpu/renderer/renderContext.cpp

@@ -147,6 +147,7 @@ void RenderContext::RunGarbageCollector() {
 		m_fault_process_pending = false;
 		m_buffer_cache.ProcessFaultBuffer();
 	}
+	m_image_heap.RunGarbageCollector();
 	m_texture_cache.ProcessDownloadImages();
 	m_texture_cache.RunGarbageCollector();
 	m_buffer_cache.RunGarbageCollector();

src/graphics/host_gpu/renderer/renderDraw.cpp

@@ -52,7 +52,7 @@ std::pair<int32_t, uint32_t> ResolveDrawOffsets(uint32_t index_offset,
 	                                           const ShaderVertexInputInfo& vs_input_info) {
 	auto     vertex_offset   = static_cast<int32_t>(index_offset);
 	uint32_t instance_offset = 0;
-	if (!vs_input_info.fetch_embedded) {
+	if (!vs_input_info.has_dynamic_vertex_fetch) {
 		return {vertex_offset, instance_offset};
 	}
 
@@ -403,6 +403,20 @@ static void UpdateGraphicsDynamicState(const CommandBuffer& buffer, DynamicState
 		dynamic_state.SetDepthBias(constant_factor, poly_offset.clamp, slope_factor);
 	}
 
+	// Rect lists are drawn without culling.
+	vk::CullModeFlags cull_mode {};
+	if (!Prospero::IsRectList(buffer.GetUserConfig().GetPrimType())) {
+		if (mode.cull_front) {
+			cull_mode |= vk::CullModeFlagBits::eFront;
+		}
+		if (mode.cull_back) {
+			cull_mode |= vk::CullModeFlagBits::eBack;
+		}
+	}
+	dynamic_state.SetCullMode(cull_mode);
+	dynamic_state.SetFrontFace(mode.face ? vk::FrontFace::eClockwise
+	                                     : vk::FrontFace::eCounterClockwise);
+
 	dynamic_state.SetStencilTestEnabled(depth.stencil_test_enable);
 	if (depth.stencil_test_enable) {
 		const auto& front = depth.stencil_front;
@@ -1236,24 +1250,36 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 	}
 }
 
-// Vulkan offsets for a prepared draw, or nullopt when its indirect records need CPU counts, as
-// mesh and quad-list draws do. Indirect records carry their own offsets.
+// Vulkan offsets for a prepared draw, or nullopt when its indirect records need the CPU: mesh
+// and quad-list draws for counts, tessellation draws for the index/SGPR prolog. Indirect records
+// carry their own offsets.
 static std::optional<DrawEmitInfo> ResolveDrawEmit(const HW::UserConfig&        ucfg,
                                                    const ShaderVertexInputInfo& vs,
                                                    DrawOffsetSource             source,
                                                    int32_t                      first_vertex,
                                                    uint32_t                     first_instance,
                                                    const DrawIndirectArgs&      indirect) {
+	if ((indirect.address != 0 || source == DrawOffsetSource::IndirectArgs) &&
+	    ucfg.GetIndexOffset() != 0) {
+		static std::once_flag warning_once;
+		std::call_once(warning_once, [&] {
+			Log::WriteToConsoleAndLog(fmt::format(
+			    "GPU: indirect draw ignores GE_INDX_OFFSET=0x{:08x}; geometry may be incorrect.\n",
+			    ucfg.GetIndexOffset()));
+		});
+	}
 	if (indirect.address != 0) {
-		if (vs.stage.module->info.stage == ShaderType::Mesh ||
-		    ucfg.GetPrimType() == Prospero::PrimitiveType::kQuadListLegacy) {
+		if (!vs.indirect_draw ||
+		    vs.stage.module->info.stage == ShaderType::Mesh ||
+		    ucfg.GetPrimType() == Prospero::PrimitiveType::kQuadListLegacy ||
+		    vs.indirect_offsets != indirect.offsets) {
 			return std::nullopt;
 		}
 		return DrawEmitInfo {.indirect = indirect};
 	}
 	if (source == DrawOffsetSource::IndirectArgs) {
 		// Native fetches consume the patched SGPRs; rewritten fetches need Vulkan offsets.
-		return vs.fetch_embedded ? DrawEmitInfo {first_vertex, first_instance} : DrawEmitInfo {};
+		return vs.has_dynamic_vertex_fetch ? DrawEmitInfo {first_vertex, first_instance} : DrawEmitInfo {};
 	}
 	const auto [vertex_offset, instance_offset] = ResolveDrawOffsets(ucfg.GetIndexOffset(), vs);
 	return DrawEmitInfo {vertex_offset + first_vertex, instance_offset};
@@ -1310,7 +1336,8 @@ bool RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
 
 	hw_check(buffer);
 
-	DrawRenderState state {};
+	DrawRenderState state {.indirect_offsets = args.indirect.offsets,
+	                       .indirect_draw    = args.indirect.address != 0};
 	if (!GetDrawTopology(ucfg, state.topology)) {
 		return true;
 	}
@@ -1415,7 +1442,8 @@ bool RenderExecutor::DrawAuto(uint64_t submit_id, CommandBuffer& buffer, const D
 	const DrawCallInfo draw {CommandBufferDebugOp::DrawIndexAuto,
 	                         args.vertex_count, args.instance_count, args.first_instance};
 
-	DrawRenderState state {};
+	DrawRenderState state {.indirect_offsets = args.indirect.offsets,
+	                       .indirect_draw    = args.indirect.address != 0};
 	if (!GetDrawTopology(ucfg, state.topology)) {
 		ResetBindings();
 		return true;

src/graphics/host_gpu/renderer/renderPssr.cpp

@@ -0,0 +1,184 @@
+#include "common/assert.h"
+#include "common/threads.h"
+#include "gpu_blit_shaders/gpu_pssr_linear_spv.h"
+#include "graphics/guest_gpu/pssr.h"
+#include "graphics/host_gpu/graphicContext.h"
+#include "graphics/host_gpu/renderer/image/image.h"
+#include "graphics/host_gpu/renderer/pipeline/shaderResourceBarrier.h"
+#include "graphics/host_gpu/renderer/render.h"
+#include "graphics/host_gpu/renderer/renderContext.h"
+#include "graphics/shader/shaderBindings.h"
+
+#include <algorithm>
+#include <array>
+
+namespace Libs::Graphics {
+namespace {
+
+struct PssrParameters {
+	uint32_t input_size[2];
+	uint32_t output_size[2];
+	uint32_t output_channels[4];
+};
+
+bool IsInputFormat(Prospero::BufferFormat format) {
+	return format == Prospero::BufferFormat::k11_11_10Float ||
+	       format == Prospero::BufferFormat::k10_11_11Float ||
+	       format == Prospero::BufferFormat::k16_16_16_16Float;
+}
+
+} // namespace
+
+RenderExecutor::~RenderExecutor() {
+	const auto device = m_context.GetGraphics().device;
+	device.destroyPipeline(m_pssr_pipeline, nullptr);
+	device.destroyPipelineLayout(m_pssr_pipeline_layout, nullptr);
+	device.destroyDescriptorSetLayout(m_pssr_descriptor_layout, nullptr);
+}
+
+void RenderExecutor::ProcessPssr(CommandBuffer& buffer, const Pssr::Request& request) {
+	EXIT_IF(buffer.IsInvalid());
+	m_context.GetCommandScheduler().PopPendingOperations();
+	Common::LockGuard lock(m_context.GetMutex());
+	ShaderTextureResource input_descriptor {}, output_descriptor {};
+	std::copy_n(request.input, 8, input_descriptor.fields);
+	std::copy_n(request.output, 8, output_descriptor.fields);
+	const auto is_2d = [](const ShaderTextureResource& descriptor) {
+		return !descriptor.IsNull() &&
+		       (descriptor.Type() == Prospero::ImageType::kColor2D ||
+		        descriptor.Type() == Prospero::ImageType::kColor2DArray) &&
+		       descriptor.TileMode() != Prospero::TileMode::kDepth;
+	};
+	EXIT_IF(!is_2d(input_descriptor) || !is_2d(output_descriptor) ||
+	        !IsInputFormat(input_descriptor.Format()) ||
+	        output_descriptor.Format() != Prospero::BufferFormat::k11_11_10Float ||
+	        output_descriptor.TileMode() != Prospero::TileMode::kStandard256B ||
+	        output_descriptor.BaseLevel() != 0 || output_descriptor.LastLevel() != 0 ||
+	        output_descriptor.MaxMip() != 0 || output_descriptor.BaseArray5() != 0 ||
+	        output_descriptor.Depth() != 0 || request.width == 0 || request.height == 0);
+	const auto input_width = std::max(1u, (input_descriptor.Width5() + 1u) >>
+	                                         input_descriptor.BaseLevel());
+	const auto input_height = std::max(1u, (input_descriptor.Height5() + 1u) >>
+	                                          input_descriptor.BaseLevel());
+	EXIT_IF(request.width > input_width || request.height > input_height);
+	PssrParameters parameters {{request.width, request.height},
+	                           {output_descriptor.Width5() + 1u, output_descriptor.Height5() + 1u},
+	                           {4, 4, 4, 4}};
+	for (uint32_t component = 0; component < 3; ++component) {
+		for (uint32_t source = 0; source < 4; ++source) {
+			if (GetDstSel(output_descriptor.DstSelXYZW(), source) == 4u + component) {
+				parameters.output_channels[component] = source;
+				break;
+			}
+		}
+	}
+	const auto resolve = [&](const uint32_t* words, bool output) {
+		ShaderRecompiler::IR::ImageResource resource {};
+		resource.resource_class = output ? ShaderRecompiler::IR::ImageResourceClass::Storage
+		                                 : ShaderRecompiler::IR::ImageResourceClass::Sampled;
+		resource.numeric_class = Prospero::TextureNumericClass::Float;
+		resource.dimension = ShaderRecompiler::Decoder::ImageDimension::Dim2D;
+		resource.read = !output;
+		resource.written = output;
+		ShaderRecompiler::IR::DescriptorValue value {};
+		std::copy_n(words, 8, value.dwords.begin());
+		value.dword_count = 8;
+		auto binding = ResolveTexture(resource, value);
+		binding.desc.view_info.level_count = 1;
+		binding.desc.view_info.min_lod = 0;
+		return binding;
+	};
+	auto& cache = m_context.GetTextureCache();
+	auto input = resolve(request.input, false);
+	BindImage(input.image_id, false);
+	auto output = resolve(request.output, true);
+	BindImage(output.image_id, true);
+	EXIT_IF(input.desc.info.data.address < output.desc.info.data.End() &&
+	        output.desc.info.data.address < input.desc.info.data.End());
+	const auto acquire = [&](TextureBinding& binding, const uint32_t* words, bool storage) {
+		auto& previous = cache.GetImage(binding.image_id);
+		if (!previous.registered || previous.binding.needs_rebind) {
+			previous.binding = {};
+			binding = resolve(words, storage);
+			BindImage(binding.image_id, storage);
+		}
+		binding.image_view = cache.FindTexture(binding.image_id, binding.desc);
+	};
+	acquire(input, request.input, false);
+	acquire(output, request.output, true);
+	auto& source = cache.GetImage(input.image_id);
+	auto& destination = cache.GetImage(output.image_id);
+	const auto& source_view = input.desc.view_info;
+	const auto& destination_view = output.desc.view_info;
+	EXIT_IF(source.info.IsDepth() || destination.info.IsDepth() ||
+	        source.backing.samples != 1 || destination.backing.samples != 1 ||
+	        (input.image_id == output.image_id &&
+	         source_view.base_level == destination_view.base_level &&
+	         source_view.base_layer == destination_view.base_layer));
+
+	auto& graphics = m_context.GetGraphics();
+	if (m_pssr_pipeline == nullptr) {
+		const std::array bindings {
+		    vk::DescriptorSetLayoutBinding {0, vk::DescriptorType::eSampledImage, 1,
+		                                    vk::ShaderStageFlagBits::eCompute},
+		    vk::DescriptorSetLayoutBinding {1, vk::DescriptorType::eStorageImage, 1,
+		                                    vk::ShaderStageFlagBits::eCompute}};
+		vk::DescriptorSetLayoutCreateInfo descriptor_info {};
+		descriptor_info.flags = vk::DescriptorSetLayoutCreateFlagBits::ePushDescriptorKHR;
+		descriptor_info.bindingCount = bindings.size();
+		descriptor_info.pBindings = bindings.data();
+		RequireVulkanSuccess(graphics.device.createDescriptorSetLayout(
+		                         &descriptor_info, nullptr, &m_pssr_descriptor_layout),
+		                     "create upscaling descriptor layout");
+		const vk::PushConstantRange push {vk::ShaderStageFlagBits::eCompute, 0,
+		                                  sizeof(PssrParameters)};
+		vk::PipelineLayoutCreateInfo layout_info {};
+		layout_info.setLayoutCount = 1;
+		layout_info.pSetLayouts = &m_pssr_descriptor_layout;
+		layout_info.pushConstantRangeCount = 1;
+		layout_info.pPushConstantRanges = &push;
+		RequireVulkanSuccess(graphics.device.createPipelineLayout(
+		                         &layout_info, nullptr, &m_pssr_pipeline_layout),
+		                     "create upscaling pipeline layout");
+		const auto shader = CompileSPV(GPU_PSSR_LINEAR_SPV, graphics.device);
+		vk::ComputePipelineCreateInfo create {};
+		create.stage.stage = vk::ShaderStageFlagBits::eCompute;
+		create.stage.module = shader;
+		create.stage.pName = "main";
+		create.layout = m_pssr_pipeline_layout;
+		const auto result = graphics.device.createComputePipelines(
+		    nullptr, 1, &create, nullptr, &m_pssr_pipeline);
+		graphics.device.destroyShaderModule(shader, nullptr);
+		RequireVulkanSuccess(result, "create upscaling pipeline");
+	}
+	buffer.EndRendering();
+	const auto command = buffer.Handle();
+	source.Transit(vk::ImageLayout::eGeneral, vk::AccessFlagBits2::eShaderRead,
+	               ImageSubresourceRange {source_view.base_level, 1, source_view.base_layer, 1},
+	               command);
+	destination.Transit(vk::ImageLayout::eGeneral, vk::AccessFlagBits2::eShaderWrite,
+	                    ImageSubresourceRange {destination_view.base_level, 1,
+	                                           destination_view.base_layer, 1}, command);
+	const std::array images {
+	    vk::DescriptorImageInfo {nullptr, input.image_view, vk::ImageLayout::eGeneral},
+	    vk::DescriptorImageInfo {nullptr, output.image_view, vk::ImageLayout::eGeneral}};
+	std::array<vk::WriteDescriptorSet, 2> writes {};
+	for (uint32_t i = 0; i < writes.size(); ++i) {
+		writes[i].dstBinding = i;
+		writes[i].descriptorCount = 1;
+		writes[i].descriptorType = i == 0 ? vk::DescriptorType::eSampledImage
+		                                 : vk::DescriptorType::eStorageImage;
+		writes[i].pImageInfo = &images[i];
+	}
+	command.pushDescriptorSetKHR(vk::PipelineBindPoint::eCompute, m_pssr_pipeline_layout,
+	                             0, writes.size(), writes.data());
+	command.bindPipeline(vk::PipelineBindPoint::eCompute, m_pssr_pipeline);
+	command.pushConstants(m_pssr_pipeline_layout, vk::ShaderStageFlagBits::eCompute,
+	                      0, sizeof(parameters), &parameters);
+	command.dispatch((parameters.output_size[0] + 7) / 8,
+	                 (parameters.output_size[1] + 7) / 8, 1);
+	ShaderAccessBarrier(command, vk::PipelineStageFlagBits::eComputeShader);
+	ResetBindings();
+}
+
+} // namespace Libs::Graphics

src/graphics/host_gpu/shaders/gpu_pssr_linear.comp

@@ -0,0 +1,33 @@
+#version 450 core
+#extension GL_EXT_samplerless_texture_functions : require
+
+layout(local_size_x = 8, local_size_y = 8) in;
+layout(binding = 0) uniform texture2D source;
+layout(binding = 1) writeonly uniform image2D destination;
+layout(push_constant) uniform Parameters {
+	uvec2 input_size;
+	uvec2 output_size;
+	uvec4 output_channels;
+} parameters;
+
+vec4 fetch_color(ivec2 coordinate) {
+	return texelFetch(source, clamp(coordinate, ivec2(0), ivec2(parameters.input_size) - 1), 0);
+}
+
+void main() {
+	uvec2 output_coordinate = gl_GlobalInvocationID.xy;
+	if (any(greaterThanEqual(output_coordinate, parameters.output_size))) return;
+	vec2 input_coordinate = (vec2(output_coordinate) + 0.5) *
+	                        vec2(parameters.input_size) / vec2(parameters.output_size) - 0.5;
+	ivec2 lower = ivec2(floor(input_coordinate));
+	vec2 weight = fract(input_coordinate);
+	vec4 color = mix(mix(fetch_color(lower), fetch_color(lower + ivec2(1, 0)), weight.x),
+	                 mix(fetch_color(lower + ivec2(0, 1)), fetch_color(lower + ivec2(1, 1)), weight.x),
+	                 weight.y);
+	vec4 stored = vec4(0.0);
+	for (uint channel = 0; channel < 3; ++channel) {
+		uint selected = parameters.output_channels[channel];
+		if (selected < 4) stored[channel] = color[selected];
+	}
+	imageStore(destination, ivec2(output_coordinate), stored);
+}

src/graphics/host_gpu/shaders/gpu_tiler_color.inc

@@ -1,4 +1,4 @@
-// ColorTransform values match imageInfo.h. Preserve the floating-point bit fields exactly.
+// ColorTransform values match imageInfo.h. Preserve the component bit fields exactly.
 uvec2 transform_color(uvec2 value, uint transform, bool to_host) {
 	switch (transform) {
 		case 1u:
@@ -15,6 +15,13 @@ uvec2 transform_color(uvec2 value, uint transform, bool to_host) {
 			}
 			break;
 		}
+		case 3u:
+			if (to_host) {
+				value.x = (value.x >> 2u) | (value.x << 30u);
+			} else {
+				value.x = (value.x << 2u) | (value.x >> 30u);
+			}
+			break;
 	}
 	return value;
 }

src/graphics/host_gpu/shaders/gpu_tiler_demote_d16.comp

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

src/graphics/host_gpu/vma.cpp

@@ -151,7 +151,8 @@ bool GraphicContext::CreateImage(const vk::ImageCreateInfo& image_info, VulkanIm
 	EXIT_IF(allocator == nullptr || image.image != nullptr || image.allocation != nullptr);
 
 	VmaAllocationCreateInfo alloc_info {};
-	alloc_info.requiredFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
+	alloc_info.flags = VMA_ALLOCATION_CREATE_WITHIN_BUDGET_BIT;
+	alloc_info.usage = VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE;
 
 	vk::Image::CType native_image = VK_NULL_HANDLE;
 	const auto        result       = static_cast<vk::Result>(

src/graphics/presentation/performanceOverlay.cpp

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

src/graphics/presentation/videoOut.cpp

@@ -221,7 +221,7 @@ public:
 	void WaitForSubmitSlot(VideoOutConfig& cfg);
 	bool Flip(uint32_t micros);
 	void GetFlipStatus(VideoOutConfig& cfg, VideoOutFlipStatus& out);
-	bool IsSafeForRendering(VideoOutConfig& cfg, int index);
+	void Wait(VideoOutConfig& cfg, int index);
 
 private:
 	enum class RequestState { Reserved, Recording, Ready, Presenting };
@@ -240,7 +240,6 @@ private:
 		Graphics::Presenter::Frame* frame;
 		bool                        premultiplied_alpha;
 	};
-	void SignalDoneLocked();
 
 	Graphics::Presenter& m_presenter;
 	Common::Mutex        m_mutex;
@@ -251,7 +250,6 @@ private:
 	std::list<Request>   m_cpu_requests;
 	std::list<Request>   m_cancelled_requests;
 	bool                 m_processing      = false;
-	bool                 m_gpu_waiting     = false;
 	uint64_t             m_next_request_id = 1;
 };
 
@@ -955,7 +953,7 @@ bool FlipQueue::Reserve(VideoOutConfig& cfg, int index, int64_t flip_arg, FlipRe
 				}
 			}
 		}
-		SignalDoneLocked();
+		m_done_cond_var.SignalAll();
 	}
 	if (cfg.flip_status.flipPendingNum >= VIDEO_OUT_FLIP_QUEUE_CAPACITY) {
 		Common::DebugCounters::Add(Common::DebugCounters::Counter::FlipsRejected);
@@ -1035,7 +1033,7 @@ void FlipQueue::Cancel(VideoOutConfig& cfg) {
 			it = queue->erase(it);
 		}
 	}
-	SignalDoneLocked();
+	m_done_cond_var.SignalAll();
 	m_submit_slot_cond_var.SignalAll();
 	m_submit_cond_var.SignalAll();
 	m_mutex.Unlock();
@@ -1207,24 +1205,17 @@ void FlipQueue::WaitForSubmitSlot(VideoOutConfig& cfg) {
 	}
 }
 
-void FlipQueue::SignalDoneLocked() {
-	m_done_cond_var.SignalAll();
-	// A stalled GPU cannot drain and shut down without re-checking under m_mutex, so it is alive here.
-	if (m_gpu_waiting) {
-		m_gpu_waiting = false;
-		m_presenter.Renderer().GetGpu().Wake();
-	}
-}
-
-// A display buffer is safe for rendering once it is neither displayed nor waiting to be.
-bool FlipQueue::IsSafeForRendering(VideoOutConfig& cfg, int index) {
+void FlipQueue::Wait(VideoOutConfig& cfg, int index) {
 	Common::LockGuard lock(m_mutex);
-	auto matches = [&cfg, index](const auto& r) { return r.cfg == &cfg && r.index == index; };
-	const bool safe = cfg.flip_status.currentBuffer != index &&
-	                  std::none_of(m_requests.begin(), m_requests.end(), matches) &&
-	                  std::none_of(m_cpu_requests.begin(), m_cpu_requests.end(), matches);
-	m_gpu_waiting = !safe;
-	return safe;
+
+	auto has_request = [this, &cfg, index] {
+		auto matches = [&cfg, index](const auto& r) { return r.cfg == &cfg && r.index == index; };
+		return std::any_of(m_requests.begin(), m_requests.end(), matches) ||
+		       std::any_of(m_cpu_requests.begin(), m_cpu_requests.end(), matches);
+	};
+	while (has_request()) {
+		m_done_cond_var.Wait(&m_mutex);
+	}
 }
 
 bool FlipQueue::Flip(uint32_t micros) {
@@ -1320,7 +1311,7 @@ bool FlipQueue::Flip(uint32_t micros) {
 		it = m_requests.erase(it);
 	}
 	m_processing = false;
-	SignalDoneLocked();
+	m_done_cond_var.SignalAll();
 	m_submit_slot_cond_var.SignalAll();
 	m_mutex.Unlock();
 	for (size_t i = count; i != 0; i--) {
@@ -1703,12 +1694,12 @@ void VideoOutDriver::WaitForSubmitSlot(int handle) {
 	m_impl->GetFlipQueue().WaitForSubmitSlot(*cfg);
 }
 
-bool VideoOutDriver::IsSafeForRendering(int handle, int index) {
+void VideoOutDriver::WaitFlipDone(int handle, int index) {
 	auto* ctx = m_impl->Get(handle);
 	EXIT_IF(ctx == nullptr);
 
 	EXIT_NOT_IMPLEMENTED(!IsValidBufferIndex(index));
-	return m_impl->GetFlipQueue().IsSafeForRendering(*ctx, index);
+	m_impl->GetFlipQueue().Wait(*ctx, index);
 }
 
 KYTY_SYSV_ABI int VideoOutGetFlipStatus(int handle, VideoOutFlipStatus* status) {

src/graphics/presentation/videoOut.h

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

src/graphics/presentation/window/swapchain.cpp

@@ -290,7 +290,8 @@ public:
 private:
 	vk::SurfaceFormatKHR SelectFormat();
 	void Destroy();
-	void DrawOverlay(vk::CommandBuffer command, const Presenter::Layer& layer);
+	void DrawOverlay(vk::CommandBuffer command, const Presenter::Layer& layer,
+	                 const vk::Rect2D& rect);
 
 	WindowContext&   m_window;
 	vk::SwapchainKHR m_handle = nullptr;
@@ -637,7 +638,23 @@ bool Swapchain::PrepareSystemOverlay() {
 	return m_system_overlay->PrepareFrame(m_extent, m_format, ImageCount());
 }
 
-void Swapchain::DrawOverlay(vk::CommandBuffer command, const Presenter::Layer& layer) {
+static vk::Rect2D FitAspect(const vk::Extent3D& source, const vk::Extent2D& target) {
+	const uint64_t src_w = std::max(source.width, 1u);
+	const uint64_t src_h = std::max(source.height, 1u);
+	uint32_t       w     = target.width;
+	uint32_t       h     = target.height;
+	if (src_w * h > src_h * w) {
+		h = std::max<uint32_t>(1, static_cast<uint32_t>(src_h * w / src_w));
+	} else {
+		w = std::max<uint32_t>(1, static_cast<uint32_t>(src_w * h / src_h));
+	}
+	return {{static_cast<int32_t>((target.width - w) / 2),
+	         static_cast<int32_t>((target.height - h) / 2)},
+	        {w, h}};
+}
+
+void Swapchain::DrawOverlay(vk::CommandBuffer command, const Presenter::Layer& layer,
+                            const vk::Rect2D& rect) {
 	auto device = m_window.graphic_ctx.device;
 	if (m_overlay_pipeline == nullptr) {
 		vk::SamplerCreateInfo sampler {};
@@ -756,12 +773,14 @@ void Swapchain::DrawOverlay(vk::CommandBuffer command, const Presenter::Layer& l
 	rendering.colorAttachmentCount = 1;
 	rendering.pColorAttachments    = &attachment;
 	command.beginRendering(&rendering);
-	const vk::Viewport viewport {
-	    0.0f, 0.0f, static_cast<float>(m_extent.width), static_cast<float>(m_extent.height),
-	    0.0f, 1.0f};
-	const vk::Rect2D scissor {{0, 0}, m_extent};
+	const vk::Viewport viewport {static_cast<float>(rect.offset.x),
+	                             static_cast<float>(rect.offset.y),
+	                             static_cast<float>(rect.extent.width),
+	                             static_cast<float>(rect.extent.height),
+	                             0.0f,
+	                             1.0f};
 	command.setViewport(0, 1, &viewport);
-	command.setScissor(0, 1, &scissor);
+	command.setScissor(0, 1, &rect);
 	command.draw(3, 1, 0, 0);
 	command.endRendering();
 }
@@ -799,6 +818,24 @@ void Swapchain::RecordPresentCommands(CommandBuffer& command, Presenter::Frame*
 	                           vk::PipelineStageFlagBits::eTransfer, vk::DependencyFlags {}, 0,
 	                           nullptr, 0, nullptr, 1, &to_transfer);
 
+	const auto* aspect_source = source != nullptr ? source : overlay.frame;
+	const auto  game_rect     = !Config::StretchEnabled() && aspect_source != nullptr
+	                               ? FitAspect(aspect_source->image.extent, m_extent)
+	                               : vk::Rect2D {{0, 0}, m_extent};
+	if (source == nullptr || game_rect.extent != m_extent) {
+		const vk::ClearColorValue black {};
+		vk_command.clearColorImage(m_images[m_image_index], vk::ImageLayout::eTransferDstOptimal,
+		                           &black, 1, &to_transfer.subresourceRange);
+		if (source != nullptr) {
+			vk::ImageMemoryBarrier after_clear = to_transfer;
+			after_clear.srcAccessMask          = vk::AccessFlagBits::eTransferWrite;
+			after_clear.oldLayout              = vk::ImageLayout::eTransferDstOptimal;
+			vk_command.pipelineBarrier(vk::PipelineStageFlagBits::eTransfer,
+			                           vk::PipelineStageFlagBits::eTransfer,
+			                           vk::DependencyFlags {}, 0, nullptr, 0, nullptr, 1,
+			                           &after_clear);
+		}
+	}
 	if (source != nullptr) {
 		vk::ImageBlit region {};
 		region.srcSubresource.aspectMask     = vk::ImageAspectFlagBits::eColor;
@@ -812,16 +849,14 @@ void Swapchain::RecordPresentCommands(CommandBuffer& command, Presenter::Frame*
 		region.dstSubresource.mipLevel       = 0;
 		region.dstSubresource.baseArrayLayer = 0;
 		region.dstSubresource.layerCount     = 1;
-		region.dstOffsets[1].x               = static_cast<int>(m_extent.width);
-		region.dstOffsets[1].y               = static_cast<int>(m_extent.height);
+		region.dstOffsets[0].x = game_rect.offset.x;
+		region.dstOffsets[0].y = game_rect.offset.y;
+		region.dstOffsets[1].x = game_rect.offset.x + static_cast<int>(game_rect.extent.width);
+		region.dstOffsets[1].y = game_rect.offset.y + static_cast<int>(game_rect.extent.height);
 		region.dstOffsets[1].z               = 1;
 		vk_command.blitImage(source->image.image, vk::ImageLayout::eTransferSrcOptimal,
 		                     m_images[m_image_index], vk::ImageLayout::eTransferDstOptimal, 1,
 		                     &region, vk::Filter::eLinear);
-	} else {
-		const vk::ClearColorValue black {};
-		vk_command.clearColorImage(m_images[m_image_index], vk::ImageLayout::eTransferDstOptimal,
-		                           &black, 1, &to_transfer.subresourceRange);
 	}
 
 	vk::ImageMemoryBarrier to_present {};
@@ -848,7 +883,7 @@ void Swapchain::RecordPresentCommands(CommandBuffer& command, Presenter::Frame*
 	                           &to_present);
 	if (draw_attachment) {
 		if (draw_overlay) {
-			DrawOverlay(vk_command, overlay);
+			DrawOverlay(vk_command, overlay, game_rect);
 			if (draw_system_overlay) {
 				to_present.srcAccessMask = vk::AccessFlagBits::eColorAttachmentWrite;
 				to_present.oldLayout     = vk::ImageLayout::eColorAttachmentOptimal;

src/graphics/presentation/window/vulkanWindow.cpp

@@ -51,6 +51,7 @@ struct VulkanExtensions {
 vk::PhysicalDeviceVulkan11Features WindowContext::RequiredVulkan11Features() noexcept {
 	vk::PhysicalDeviceVulkan11Features features {};
 	features.storageBuffer16BitAccess = VK_TRUE;
+	features.shaderDrawParameters     = VK_TRUE;
 	return features;
 }
 
@@ -278,6 +279,8 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 		              required_features12.samplerMirrorClampToEdge);
 		check_feature(features11.storageBuffer16BitAccess, "storageBuffer16BitAccess",
 		              WindowContext::RequiredVulkan11Features().storageBuffer16BitAccess);
+		check_feature(features11.shaderDrawParameters, "shaderDrawParameters",
+		              WindowContext::RequiredVulkan11Features().shaderDrawParameters);
 		check_feature(features12.storageBuffer8BitAccess, "storageBuffer8BitAccess",
 		              required_features12.storageBuffer8BitAccess);
 		check_feature(features12.timelineSemaphore, "timelineSemaphore",

src/graphics/shader/recompiler/ShaderRecompiler.cpp

@@ -814,7 +814,7 @@ TranslateResult TranslateProgram(const Decoder::Program& decoded, const CompileO
 	const auto& cfg = *selected_cfg;
 	Frontend::EmbeddedFetchPlan embedded_fetch;
 	if ((options.stage == ShaderType::Vertex || options.stage == ShaderType::Local) &&
-	    options.input_info.vertex->fetch_embedded) {
+	    options.input_info.vertex->has_dynamic_vertex_fetch) {
 		embedded_fetch = DetectEmbeddedVertexFetch(
 		    decoded, *options.input_info.vertex, options.user_data_base,
 		    static_cast<uint32_t>(options.user_data.size()), options.wave_size);

src/graphics/shader/recompiler/Tessellation.cpp

@@ -156,11 +156,14 @@ void ReflectTessellationLayout(const Decoder::Program& program, bool local,
 						if (rhs.constant == 0u && third.constant == 8u) value = constant(0u);
 					}
 					break;
-				case Opcode::V_AND_B32:
-					if ((lhs.kind == Kind::PackedControlPoint && rhs == constant(0xffu)) ||
-					    (rhs.kind == Kind::PackedControlPoint && lhs == constant(0xffu)))
-						value = constant(0u);
+				case Opcode::V_AND_B32: {
+					const auto mask = lhs.kind == Kind::PackedControlPoint ? rhs
+					                : rhs.kind == Kind::PackedControlPoint ? lhs
+					                                                       : Address {};
+					if (mask == constant(0xffu)) value = constant(0u);
+					if (mask == constant(0x1f00u)) value = {Kind::Affine, 256u, 0u};
 					break;
+				}
 				case Opcode::V_ADD_NC_U32: value = add(lhs, rhs); break;
 				case Opcode::V_MUL_LO_U32: value = multiply(lhs, rhs); break;
 				case Opcode::V_MUL_U32_U24: value = multiply(low24(lhs), low24(rhs)); break;

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

@@ -30,16 +30,7 @@ uint32_t NodeWord(EmitterState& s, uint32_t block, uint32_t word) {
 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 {
@@ -56,6 +47,28 @@ uint32_t Triangle(EmitterState& s, uint32_t block, uint32_t kind, uint32_t bary,
 			    NodeWord(s, block, Binary(s, spv::OpIAdd, u, first, ConstantU32(s, axis))));
 	}
 	const auto flag = NodeWord(s, block, ConstantU32(s, 15));
+	const auto remap = Binary(s, spv::OpShiftRightLogical, u, flag,
+	    Binary(s, spv::OpIMul, u, kind, ConstantU32(s, 8)));
+	return EmitBvhTriangle(s, positions, origin, direction, remap,
+	    Binary(s, spv::OpIAdd, u, flag, kind), bary);
+}
+
+} // namespace
+
+uint32_t EmitBvhTriangle(EmitterState& s, const std::array<Vec3, 3>& positions,
+                         Vec3 origin, Vec3 direction, uint32_t remap,
+                         uint32_t triangle_id, uint32_t bary) {
+	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 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]));
@@ -112,8 +125,6 @@ uint32_t Triangle(EmitterState& s, uint32_t block, uint32_t kind, uint32_t bary,
 		    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));
@@ -124,11 +135,13 @@ uint32_t Triangle(EmitterState& s, uint32_t block, uint32_t kind, uint32_t bary,
 	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(0), triangle_id),
 		Select(s, u, bary, coordinate(2), Select(s, u, missed, ConstantU32(s, 0), ConstantU32(s, 1))) };
 	return Vector(s, TypeU32Vector(s, 4), result);
 }
 
+namespace {
+
 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);

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

@@ -63,7 +63,8 @@ uint32_t EmitBuiltinU32(EmitterState& state, IR::StageInputKind kind, uint32_t c
 	}
 	if (kind == IR::StageInputKind::VertexIndex || kind == IR::StageInputKind::InstanceIndex ||
 	    kind == IR::StageInputKind::InvocationId || kind == IR::StageInputKind::PrimitiveId ||
-	    kind == IR::StageInputKind::Layer || kind == IR::StageInputKind::SampleId) {
+	    kind == IR::StageInputKind::Layer || kind == IR::StageInputKind::SampleId ||
+	    kind == IR::StageInputKind::BaseVertex || kind == IR::StageInputKind::BaseInstance) {
 		const auto value = state.builder.AllocateId();
 		const auto bits  = state.builder.AllocateId();
 		state.builder.AddFunction(spv::OpLoad, TypeI32(state), value, variable);

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

@@ -338,6 +338,9 @@ 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);
+uint32_t              EmitBvh8Intersect(ValueEmitContext& ctx, const IR::Inst& inst);
+uint32_t              EmitBvh8Extract(ValueEmitContext& ctx, const IR::Inst& inst);
+uint32_t              EmitBvhStack(ValueEmitContext& ctx, const IR::Inst& inst);
 inline constexpr auto EmitImageQueryDimensions = EmitImage;
 inline constexpr auto EmitImageQueryLod        = EmitImage;
 inline constexpr auto EmitImageRead            = EmitImage;

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

@@ -118,8 +118,11 @@ struct EmitterState {
 	uint32_t                                         bda_pointer_function    = 0;
 	uint32_t                                         indirect_format_function = 0;
 	uint32_t                                         indirect_format_geometry_function = 0;
-	std::map<std::array<uint32_t, 3>, uint32_t>        indirect_buffer_load_functions;
+	std::map<std::array<uint32_t, 4>, uint32_t>        indirect_buffer_load_functions;
 	uint32_t                                         bvh_intersect_function  = 0;
+	uint32_t                                         bvh8_result_type       = 0;
+	uint32_t                                         bvh8_intersect_function = 0;
+	uint32_t                                         bvh_stack_function      = 0;
 	uint32_t                                         gds_variable            = 0;
 	uint32_t                                         gds_length              = 0;
 	uint32_t                                         push_constant_variable  = 0;
@@ -169,6 +172,7 @@ uint32_t TypeBoolVector(EmitterState& state, uint32_t components);
 uint32_t TypeU32(EmitterState& state);
 uint32_t TypeU64(EmitterState& state);
 uint32_t TypeU32Pair(EmitterState& state);
+uint32_t TypeBvh8Result(EmitterState& state);
 uint32_t TypeI32(EmitterState& state);
 uint32_t TypeI32Pair(EmitterState& state);
 uint32_t TypeF32(EmitterState& state);
@@ -540,9 +544,13 @@ uint32_t EmitF32BitsToF16RtzBits(EmitterState& state, uint32_t bits);
 void EmitProgram(EmitterState& state);
 
 void DefineGetBdaPointer(EmitterState& state);
-void DefineIndirectBufferFormats(ValueEmitContext& ctx);
 void DefineIndirectBufferLoads(ValueEmitContext& ctx);
 void DefineBvhIntersect(EmitterState& state);
+void DefineTrinityBvhIntersect(EmitterState& state);
+uint32_t EmitBvhTriangle(EmitterState& state, const std::array<std::array<uint32_t, 3>, 3>& positions,
+                         std::array<uint32_t, 3> origin, std::array<uint32_t, 3> direction,
+                         uint32_t remap, uint32_t triangle_id, uint32_t bary);
+void DefineTrinityFunctions(EmitterState& state);
 uint32_t GetBdaPointer(EmitterState& state, uint32_t address);
 
 // These templates accept local lambdas from several emitter translation units.

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

@@ -1026,53 +1026,50 @@ uint32_t LoadIndirectBuffer(ValueEmitContext& ctx, const IndirectBufferAccess& b
 	return ConstructU32Composite(state, components, values);
 }
 
-uint32_t LoadIndirectFormattedX(ValueEmitContext& ctx, const IR::Inst& inst) {
+uint32_t LoadIndirectFormattedX(ValueEmitContext& ctx, const IndirectBufferAccess& buffer) {
 	auto& state = ctx.state;
-	return EmitValueOrZeroIfCondition(state, ctx.Arg(inst, inst.NumArgs() - 1), [&]() {
-		const auto buffer = PrepareIndirectBuffer(state, IndirectBufferArgs(ctx, inst), true);
-		const auto geometry = state.builder.AllocateId();
-		state.builder.AddFunction(spv::OpFunctionCall, TypeU32Composite(state, 4), geometry,
-		                          state.indirect_format_geometry_function, buffer.format);
-		const auto field = [&](uint32_t index) {
-			const auto value = state.builder.AllocateId();
-			state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(state), value, geometry, index);
-			return value;
-		};
-		const auto bytes = field(0);
-		const auto component = Binary(
-		    state, spv::OpUMod, TypeU32(state),
-		    Binary(state, spv::OpISub, TypeU32(state), buffer.selector, ConstantU32(state, 4)), field(1));
-		const auto alignment = Unary(state, spv::OpUConvert, TypeU64(state), field(3));
-		const auto base = Binary(state, spv::OpBitwiseAnd, TypeU64(state), buffer.address,
-		                         Unary(state, spv::OpNot, TypeU64(state), alignment));
-		const auto address = Binary(
-		    state, spv::OpIAdd, TypeU64(state), base,
-		    Unary(state, spv::OpUConvert, TypeU64(state),
-		          Binary(state, spv::OpIMul, TypeU32(state), component, field(2))));
-		const auto memory = AndCondition(
-		    state, Binary(state, spv::OpINotEqual, TypeBool(state), bytes, ConstantU32(state, 0)),
-		    Binary(state, spv::OpUGreaterThanEqual, TypeBool(state), buffer.selector,
-		           ConstantU32(state, 4)));
-		const auto raw = EmitValueOrZeroIfCondition(
-		    state, AndCondition(state, memory, IndirectBufferInBounds(state, buffer, 0u, bytes, true)),
-		    [&]() {
-			    // Format alignment and component spacing keep every selected component within one DWORD.
-			    const auto word = LoadBdaDword(
-			        ctx, Binary(state, spv::OpBitwiseAnd, TypeU64(state), address,
-			                    ConstantU64(state, ~uint64_t {3})));
-			    const auto byte = Binary(
-			        state, spv::OpBitwiseAnd, TypeU32(state),
-			        Unary(state, spv::OpUConvert, TypeU32(state), address), ConstantU32(state, 3));
-			    return Binary(state, spv::OpShiftRightLogical, TypeU32(state), word,
-			                  Binary(state, spv::OpShiftLeftLogical, TypeU32(state), byte,
-			                         ConstantU32(state, 3)));
-		    });
-		const auto result = state.builder.AllocateId();
-		state.builder.AddFunction(spv::OpFunctionCall, TypeU32(state), result,
-		                          state.indirect_format_function, buffer.format, buffer.selector,
-		                          component, raw);
-		return result;
-	});
+	const auto geometry = state.builder.AllocateId();
+	state.builder.AddFunction(spv::OpFunctionCall, TypeU32Composite(state, 4), geometry,
+	                          state.indirect_format_geometry_function, buffer.format);
+	const auto field = [&](uint32_t index) {
+		const auto value = state.builder.AllocateId();
+		state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(state), value, geometry, index);
+		return value;
+	};
+	const auto bytes = field(0);
+	const auto component = Binary(
+	    state, spv::OpUMod, TypeU32(state),
+	    Binary(state, spv::OpISub, TypeU32(state), buffer.selector, ConstantU32(state, 4)), field(1));
+	const auto alignment = Unary(state, spv::OpUConvert, TypeU64(state), field(3));
+	const auto base = Binary(state, spv::OpBitwiseAnd, TypeU64(state), buffer.address,
+	                         Unary(state, spv::OpNot, TypeU64(state), alignment));
+	const auto address = Binary(
+	    state, spv::OpIAdd, TypeU64(state), base,
+	    Unary(state, spv::OpUConvert, TypeU64(state),
+	          Binary(state, spv::OpIMul, TypeU32(state), component, field(2))));
+	const auto memory = AndCondition(
+	    state, Binary(state, spv::OpINotEqual, TypeBool(state), bytes, ConstantU32(state, 0)),
+	    Binary(state, spv::OpUGreaterThanEqual, TypeBool(state), buffer.selector,
+	           ConstantU32(state, 4)));
+	const auto raw = EmitValueOrZeroIfCondition(
+	    state, AndCondition(state, memory, IndirectBufferInBounds(state, buffer, 0u, bytes, true)),
+	    [&]() {
+		    // Format alignment and component spacing keep every selected component within one DWORD.
+		    const auto word = LoadBdaDword(
+		        ctx, Binary(state, spv::OpBitwiseAnd, TypeU64(state), address,
+		                    ConstantU64(state, ~uint64_t {3})));
+		    const auto byte = Binary(
+		        state, spv::OpBitwiseAnd, TypeU32(state),
+		        Unary(state, spv::OpUConvert, TypeU32(state), address), ConstantU32(state, 3));
+		    return Binary(state, spv::OpShiftRightLogical, TypeU32(state), word,
+		                  Binary(state, spv::OpShiftLeftLogical, TypeU32(state), byte,
+		                         ConstantU32(state, 3)));
+	    });
+	const auto result = state.builder.AllocateId();
+	state.builder.AddFunction(spv::OpFunctionCall, TypeU32(state), result,
+	                          state.indirect_format_function, buffer.format, buffer.selector,
+	                          component, raw);
+	return result;
 }
 
 uint32_t LoadRawBuffer(ValueEmitContext& ctx, const IR::Inst& inst, const IR::MemoryInfo& mem,
@@ -1135,15 +1132,12 @@ uint32_t LoadRawBuffer(ValueEmitContext& ctx, const IR::Inst& inst, const IR::Me
 uint32_t LoadBuffer(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t components) {
 	auto& state = ctx.state;
 	const auto mem = ctx.Memory(inst);
-	if (mem.kind == IR::ResourceKind::IndirectBuffer && mem.formatted) {
-		return LoadIndirectFormattedX(ctx, inst);
-	}
 	return EmitValueOrDefaultIfCondition(
 	    state, ctx.Arg(inst, inst.NumArgs() - 1), TypeU32Composite(state, components),
 	    ConstantU32CompositeZero(state, components), [&]() {
 		    if (mem.kind == IR::ResourceKind::IndirectBuffer) {
 			    const auto function = state.indirect_buffer_load_functions.at(
-			        {mem.data_bits, components, uint32_t(mem.coherent)});
+			        {mem.data_bits, components, uint32_t(mem.coherent), uint32_t(mem.formatted)});
 			    const auto args = IndirectBufferArgs(ctx, inst);
 			    const auto result = state.builder.AllocateId();
 			    state.builder.AddFunction(spv::OpFunctionCall, TypeU32Composite(state, components),
@@ -1231,62 +1225,8 @@ void StoreWideShared(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t compo
 	});
 }
 
-} // namespace
-
-uint32_t GetBdaPointer(EmitterState& state, uint32_t address) {
-	const auto result = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpFunctionCall, TypeU64(state), result,
-	                          state.bda_pointer_function, address);
-	return result;
-}
-
-void DefineIndirectBufferLoads(ValueEmitContext& ctx) {
-	auto& state = ctx.state;
-	for (const auto* block: state.program.blocks) {
-		for (const auto& inst: *block) {
-			const auto components = IR::BufferComponentCount(inst.GetOpcode());
-			if (components == 0u || inst.GetOpcode() == IR::ValueOpcode::ReadConstBuffer) continue;
-			const auto& memory = ctx.Memory(inst);
-			if (memory.kind != IR::ResourceKind::IndirectBuffer || memory.formatted ||
-			    !memory.SupportsIndirectBufferLoad(inst.GetOpcode())) continue;
-			auto [entry, inserted] = state.indirect_buffer_load_functions.try_emplace(
-			    std::array {memory.data_bits, components, uint32_t(memory.coherent)}, 0u);
-			if (!inserted) continue;
-			const auto function = entry->second = state.builder.AllocateId();
-			const auto type = TypeU32(state);
-			const auto result_type = TypeU32Composite(state, components);
-			const auto signature = state.builder.Type(
-			    spv::OpTypeFunction, result_type, type, type, type, type, type, type, type, type);
-			const auto name = "load_indirect_buffer_" + std::to_string(memory.data_bits) + "x" +
-			                  std::to_string(components) + (memory.coherent ? "_coherent" : "");
-			state.builder.AddName(function, name.c_str());
-			state.builder.AddFunction(spv::OpFunction, result_type, function,
-			                          spv::FunctionControlMaskNone, signature);
-			std::array<uint32_t, 8> args;
-			for (auto& arg: args) {
-				arg = state.builder.AllocateId();
-				state.builder.AddFunction(spv::OpFunctionParameter, type, arg);
-			}
-			EmitLabel(state, state.builder.AllocateId());
-			const auto result = LoadIndirectBuffer(ctx, PrepareIndirectBuffer(state, args, false),
-			                                       memory.data_bits, components, memory.coherent);
-			state.builder.AddFunction(spv::OpReturnValue, result);
-			state.builder.AddFunction(spv::OpFunctionEnd);
-		}
-	}
-}
-
 void DefineIndirectBufferFormats(ValueEmitContext& ctx) {
 	auto& state = ctx.state;
-	const auto needed = std::ranges::any_of(state.program.blocks, [&](const IR::Block* block) {
-		return std::ranges::any_of(*block, [&](const IR::Inst& inst) {
-			if (inst.GetOpcode() != IR::ValueOpcode::LoadBufferU32) return false;
-			const auto& memory = ctx.Memory(inst);
-			return memory.kind == IR::ResourceKind::IndirectBuffer && memory.formatted;
-		});
-	});
-	if (!needed) return;
-
 	const auto type = TypeU32(state);
 	constexpr auto format_count =
 	    static_cast<uint32_t>(Prospero::BufferFormat::k32_32_32_32Float) + 1u;
@@ -1358,6 +1298,57 @@ void DefineIndirectBufferFormats(ValueEmitContext& ctx) {
 	state.builder.AddFunction(spv::OpFunctionEnd);
 }
 
+} // namespace
+
+uint32_t GetBdaPointer(EmitterState& state, uint32_t address) {
+	const auto result = state.builder.AllocateId();
+	state.builder.AddFunction(spv::OpFunctionCall, TypeU64(state), result,
+	                          state.bda_pointer_function, address);
+	return result;
+}
+
+void DefineIndirectBufferLoads(ValueEmitContext& ctx) {
+	auto& state = ctx.state;
+	for (const auto* block: state.program.blocks) {
+		for (const auto& inst: *block) {
+			const auto components = IR::BufferComponentCount(inst.GetOpcode());
+			if (components == 0u || inst.GetOpcode() == IR::ValueOpcode::ReadConstBuffer) continue;
+			const auto& memory = ctx.Memory(inst);
+			if (memory.kind != IR::ResourceKind::IndirectBuffer ||
+			    !memory.SupportsIndirectBufferLoad(inst.GetOpcode())) continue;
+			auto [entry, inserted] = state.indirect_buffer_load_functions.try_emplace(
+			    std::array {memory.data_bits, components, uint32_t(memory.coherent),
+			                uint32_t(memory.formatted)}, 0u);
+			if (!inserted) continue;
+			if (memory.formatted && state.indirect_format_function == 0u) {
+				DefineIndirectBufferFormats(ctx);
+			}
+			const auto function = entry->second = state.builder.AllocateId();
+			const auto type = TypeU32(state);
+			const auto result_type = TypeU32Composite(state, components);
+			const auto signature = state.builder.Type(
+			    spv::OpTypeFunction, result_type, type, type, type, type, type, type, type, type);
+			const auto name = "load_indirect_buffer_" + std::to_string(memory.data_bits) + "x" +
+			                  std::to_string(components) + (memory.coherent ? "_coherent" : "") +
+			                  (memory.formatted ? "_formatted" : "");
+			state.builder.AddName(function, name.c_str());
+			state.builder.AddFunction(spv::OpFunction, result_type, function,
+			                          spv::FunctionControlMaskNone, signature);
+			std::array<uint32_t, 8> args;
+			for (auto& arg: args) {
+				arg = state.builder.AllocateId();
+				state.builder.AddFunction(spv::OpFunctionParameter, type, arg);
+			}
+			EmitLabel(state, state.builder.AllocateId());
+			const auto buffer = PrepareIndirectBuffer(state, args, memory.formatted);
+			const auto result = memory.formatted ? LoadIndirectFormattedX(ctx, buffer)
+			    : LoadIndirectBuffer(ctx, buffer, memory.data_bits, components, memory.coherent);
+			state.builder.AddFunction(spv::OpReturnValue, result);
+			state.builder.AddFunction(spv::OpFunctionEnd);
+		}
+	}
+}
+
 void DefineGetBdaPointer(EmitterState& state) {
 	if (!state.program.info.uses_dma) {
 		return;

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

@@ -476,6 +476,8 @@ uint32_t BuiltInForInput(IR::StageInputKind kind) {
 		case IR::StageInputKind::LocalInvocationId: return spv::BuiltInLocalInvocationId;
 		case IR::StageInputKind::LocalInvocationIndex: return spv::BuiltInLocalInvocationIndex;
 		case IR::StageInputKind::GlobalInvocationId: return spv::BuiltInGlobalInvocationId;
+		case IR::StageInputKind::BaseVertex: return spv::BuiltInBaseVertex;
+		case IR::StageInputKind::BaseInstance: return spv::BuiltInBaseInstance;
 		default: return UINT32_MAX;
 	}
 }
@@ -526,7 +528,9 @@ void DefineInputs(EmitterState& state) {
 			case IR::StageInputKind::PrimitiveId:
 			case IR::StageInputKind::InstanceIndex:
 			case IR::StageInputKind::Layer:
-			case IR::StageInputKind::SampleId: type = TypeI32(state); break;
+			case IR::StageInputKind::SampleId:
+			case IR::StageInputKind::BaseVertex:
+			case IR::StageInputKind::BaseInstance: type = TypeI32(state); break;
 			case IR::StageInputKind::WorkgroupId:
 			case IR::StageInputKind::NumWorkgroups:
 			case IR::StageInputKind::LocalInvocationId:
@@ -802,6 +806,10 @@ void DefineModule(EmitterState& state) {
 	if (InputVariableForKind(state, IR::StageInputKind::SampleId) != 0) {
 		state.builder.RequireCapability(spv::CapabilitySampleRateShading);
 	}
+	if (InputVariableForKind(state, IR::StageInputKind::BaseVertex) != 0 ||
+	    InputVariableForKind(state, IR::StageInputKind::BaseInstance) != 0) {
+		state.builder.RequireCapability(spv::CapabilityDrawParameters);
+	}
 	if (state.program.info.image_gather_extended) {
 		state.builder.RequireCapability(spv::CapabilityImageGatherExtended);
 	}

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

@@ -414,6 +414,7 @@ uint32_t TypeId(EmitterState& state, IR::Type type) {
 		case IR::Type::U32:
 		case IR::Type::F16: return TypeU32(state);
 		case IR::Type::U64: return TypeU64(state);
+		case IR::Type::Bvh8Result: return TypeBvh8Result(state);
 		case IR::Type::U32x2: return TypeU32Pair(state);
 		case IR::Type::F32: return TypeF32(state);
 		case IR::Type::F64: return TypeF64(state);
@@ -675,8 +676,8 @@ void EmitProgram(EmitterState& state) {
 	}
 	DefineGetBdaPointer(state);
 	DefineIndirectBufferLoads(ctx);
-	DefineIndirectBufferFormats(ctx);
 	DefineBvhIntersect(state);
+	DefineTrinityFunctions(state);
 	if (program.info.uses_swizzle) {
 		ctx.scratch_u32_variable = state.builder.AllocateId();
 		if (state.lane_count == 2) {

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

@@ -29,12 +29,14 @@ uint32_t TessellationPointer(ValueEmitContext& ctx, const IR::Inst& inst) {
 	const auto pointer = state.builder.AllocateId();
 	if (kind == Attribute::Factor) {
 		EXIT_NOT_IMPLEMENTED(!inst.Arg(1).IsImmediate());
-		const auto index = inst.Arg(1).U32() / 4u;
-		const auto outer = index < 3u;
-		EXIT_NOT_IMPLEMENTED(index >= 4u);
+		const auto index       = inst.Arg(1).U32() / 4u;
+		const bool quad        = tess.domain == 2u;
+		const auto outer_count = quad ? 4u : 3u;
+		const bool outer       = index < outer_count;
+		EXIT_NOT_IMPLEMENTED(index >= (quad ? 6u : 4u));
 		state.builder.AddFunction(spv::OpAccessChain, TypePointer(state, storage, TypeF32(state)),
 		                          pointer, outer ? variable : state.tess_inner_variable,
-		                          ConstantU32(state, outer ? index : index - 3u));
+		                          ConstantU32(state, outer ? index : index - outer_count));
 		return pointer;
 	}
 	auto address = ctx.Arg(inst, 1);
@@ -128,20 +130,25 @@ void DefineTessellationInterfaces(EmitterState& state) {
 
 void DefineTessellationExecutionModes(EmitterState& state) {
 	const auto& tess = state.input_info.vertex->tess;
-	// CxVgtTfParam: triangles, integer or fractional-odd partitioning, clockwise output.
-	EXIT_NOT_IMPLEMENTED(tess.domain != 1u ||
-	                     (tess.partitioning != 0u && tess.partitioning != 2u) ||
-	                     tess.output_topology != 2u);
+	// Triangle or quad domain, integer or fractional partitioning, triangle output.
+	EXIT_NOT_IMPLEMENTED((tess.domain != 1u && tess.domain != 2u) ||
+	                     (tess.partitioning != 0u && tess.partitioning != 2u && tess.partitioning != 3u) ||
+	                     (tess.output_topology != 2u && tess.output_topology != 3u));
 	state.builder.RequireCapability(spv::CapabilityTessellation);
 	if (state.program.stage == ShaderType::TessellationControl) {
 		state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeOutputVertices,
 		                               tess.output_control_points);
 	} else {
-		state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeTriangles);
+		state.builder.AddExecutionMode(state.main_func, tess.domain == 1u
+		                                                   ? spv::ExecutionModeTriangles
+		                                                   : spv::ExecutionModeQuads);
 		state.builder.AddExecutionMode(
 		    state.main_func, tess.partitioning == 0u ? spv::ExecutionModeSpacingEqual
-		                                             : spv::ExecutionModeSpacingFractionalOdd);
-		state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeVertexOrderCw);
+		                     : tess.partitioning == 2u ? spv::ExecutionModeSpacingFractionalOdd
+		                                                : spv::ExecutionModeSpacingFractionalEven);
+		state.builder.AddExecutionMode(state.main_func, tess.output_topology == 2u
+		                                                   ? spv::ExecutionModeVertexOrderCw
+		                                                   : spv::ExecutionModeVertexOrderCcw);
 	}
 }
 

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

@@ -0,0 +1,208 @@
+#include "graphics/shader/recompiler/backend/spirv/spirvEmitterInstructions.h"
+
+namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter {
+namespace {
+
+uint32_t Bits(EmitterState& s, uint32_t value, uint32_t offset, uint32_t count) {
+	const auto result = s.builder.AllocateId();
+	s.builder.AddFunction(spv::OpBitFieldUExtract, TypeU32(s), result, value,
+	                      ConstantU32(s, offset), ConstantU32(s, count));
+	return result;
+}
+
+uint32_t SetBits(EmitterState& s, uint32_t value, uint32_t field, uint32_t offset, uint32_t count) {
+	const auto result = s.builder.AllocateId();
+	s.builder.AddFunction(spv::OpBitFieldInsert, TypeU32(s), result, value, field,
+	                      ConstantU32(s, offset), ConstantU32(s, count));
+	return result;
+}
+
+uint32_t Pair(EmitterState& s, uint32_t first, uint32_t second) {
+	const auto result = s.builder.AllocateId();
+	s.builder.AddFunction(spv::OpCompositeConstruct, TypeU32Pair(s), result, first, second);
+	return result;
+}
+
+uint32_t StackIndex(EmitterState& s, uint32_t control, uint32_t ring) {
+	return EmitAddU32(s, Bits(s, control, 15, 14),
+	    Binary(s, spv::OpShiftLeftLogical, TypeU32(s), ring, ConstantU32(s, 5)));
+}
+
+uint32_t StackPush(EmitterState& s, const MemoryResourceAccess& memory,
+                   uint32_t control, uint32_t node, uint32_t capacity) {
+	const auto u = TypeU32(s), b = TypeBool(s), zero = ConstantU32(s, 0), one = ConstantU32(s, 1);
+	const auto count = Bits(s, control, 0, 5), parent = Bits(s, control, 5, 5), ring = Bits(s, control, 10, 5);
+	const auto index = StackIndex(s, control, ring);
+	EmitIfCondition(s, EmitMemoryElementInBounds(s, memory, index), [&] {
+		s.builder.AddFunction(spv::OpStore, EmitMemoryElementPointer(s, memory, index), node);
+	});
+	control = Binary(s, spv::OpBitwiseOr, u, control,
+	    Select(s, u, Binary(s, spv::OpIEqual, b, count, capacity), ConstantU32(s, 1u << 30), zero));
+	control = Select(s, u, Binary(s, spv::OpIEqual, b, parent, capacity),
+	    Binary(s, spv::OpBitwiseAnd, u, control, ConstantU32(s, ~(1u << 29))), control);
+	control = SetBits(s, control, EmitGlsl<GLSLstd450UMin, IR::Type::U32>(s,
+	    EmitAddU32(s, count, one), capacity), 0, 5);
+	control = SetBits(s, control, EmitGlsl<GLSLstd450UMin, IR::Type::U32>(s,
+	    EmitAddU32(s, parent, one), capacity), 5, 5);
+	const auto next = EmitAddU32(s, ring, one);
+	return SetBits(s, control, Select(s, u, Binary(s, spv::OpIEqual, b, next, capacity), zero, next), 10, 5);
+}
+
+uint32_t StackPop(EmitterState& s, const MemoryResourceAccess& memory,
+                  uint32_t control, uint32_t capacity) {
+	const auto u = TypeU32(s), b = TypeBool(s), zero = ConstantU32(s, 0), one = ConstantU32(s, 1);
+	const auto count = Bits(s, control, 0, 5), parent = Bits(s, control, 5, 5), ring = Bits(s, control, 10, 5);
+	const auto has_parent = Binary(s, spv::OpINotEqual, b, Bits(s, control, 29, 1), zero);
+	const auto returning = Binary(s, spv::OpLogicalAnd, b, has_parent,
+	    Binary(s, spv::OpIEqual, b, parent, zero));
+	control = SetBits(s, control, Select(s, u, returning, one, zero), 31, 1);
+	control = Select(s, u, returning,
+	    Binary(s, spv::OpBitwiseAnd, u, control, ConstantU32(s, ~(1u << 29))), control);
+	const auto overflow = Binary(s, spv::OpINotEqual, b, Bits(s, control, 30, 1), zero);
+	const auto empty = Pair(s,
+	    Binary(s, spv::OpBitwiseAnd, u, control, ConstantU32(s, 0x5fffffff)),
+	    Select(s, u, overflow, ConstantU32(s, 0xffffffff), ConstantU32(s, 0xfffffffe)));
+	return EmitValueOrDefaultIfCondition(s, Binary(s, spv::OpINotEqual, b, count, zero),
+	    TypeU32Pair(s), empty, [&] {
+		control = SetBits(s, control, Binary(s, spv::OpISub, u, count, one), 0, 5);
+		control = SetBits(s, control, Binary(s, spv::OpISub, u, parent, one), 5, 5);
+		const auto previous = Select(s, u, Binary(s, spv::OpIEqual, b, ring, zero),
+		    Binary(s, spv::OpISub, u, capacity, one), Binary(s, spv::OpISub, u, ring, one));
+		control = SetBits(s, control, previous, 10, 5);
+		const auto index = StackIndex(s, control, previous);
+		const auto node = EmitValueOrDefaultIfCondition(s,
+		    EmitMemoryElementInBounds(s, memory, index), u, zero, [&] {
+			const auto value = s.builder.AllocateId();
+			s.builder.AddFunction(spv::OpLoad, u, value, EmitMemoryElementPointer(s, memory, index));
+			return value;
+		});
+		const auto instance = Binary(s, spv::OpIEqual, b, Bits(s, node, 0, 3), ConstantU32(s, 6));
+		control = EmitValueOrDefaultIfCondition(s, instance, u, control, [&] {
+			const auto nested = Binary(s, spv::OpINotEqual, b, Bits(s, control, 29, 1), zero);
+			const auto discarded = Binary(s, spv::OpBitwiseOr, u,
+			    SetBits(s, control, Bits(s, control, 5, 5), 0, 5), ConstantU32(s, 1u << 30));
+			return Binary(s, spv::OpBitwiseOr, u,
+			    SetBits(s, Select(s, u, nested, discarded, control), zero, 5, 5), ConstantU32(s, 1u << 29));
+		});
+		return Pair(s, control, node);
+	});
+}
+
+void DefineBvhStack(EmitterState& s) {
+	if (!s.program.info.bvh_stack) return;
+	const auto u = TypeU32(s), b = TypeBool(s), vec4 = TypeU32Vector(s, 4), pair = TypeU32Pair(s);
+	const auto zero = ConstantU32(s, 0), one = ConstantU32(s, 1);
+	const bool local = s.lds_storage_class == spv::StorageClassFunction;
+	const bool buffer = s.lds_storage_class == spv::StorageClassStorageBuffer;
+	std::vector<uint32_t> types {u, u, vec4, vec4, u};
+	if (local) types.push_back(TypeU32ArrayPointer(s, spv::StorageClassFunction, LdsDwordCount(s)));
+	if (buffer) types.push_back(u);
+	const auto signature = s.builder.Type(spv::OpTypeFunction, pair, types);
+	s.bvh_stack_function = s.builder.AllocateId();
+	s.builder.AddName(s.bvh_stack_function, "trinity_bvh_stack");
+	s.builder.AddFunction(spv::OpFunction, pair, s.bvh_stack_function, spv::FunctionControlMaskNone, signature);
+	std::vector<uint32_t> args;
+	for (const auto type: types) {
+		args.push_back(s.builder.AllocateId());
+		s.builder.AddFunction(spv::OpFunctionParameter, type, args.back());
+	}
+	const auto old_lds = s.lds_variable, old_base = s.lds_base_dwords;
+	if (local) s.lds_variable = args.back();
+	if (buffer) s.lds_base_dwords = args.back();
+	const auto memory = PrepareMemoryResourceAccess(s, {.kind = IR::ResourceKind::Lds});
+	const auto entry = s.builder.AllocateId();
+	EmitLabel(s, entry);
+	const auto last = args[1], capacity = args[4];
+	const auto walkback = Binary(s, spv::OpINotEqual, b, Bits(s, last, 3, 29), ConstantU32(s, 0x1fffffff));
+	const auto last4 = s.builder.AllocateId();
+	s.builder.AddFunction(spv::OpCompositeConstruct, vec4, last4, last, last, last, last);
+	const auto found = [&](uint32_t values) {
+		return Unary(s, spv::OpAny, b, Binary(s, spv::OpIEqual, TypeBoolVector(s, 4), values, last4));
+	};
+	const auto initial_valid = Binary(s, spv::OpLogicalOr, b, Unary(s, spv::OpLogicalNot, b, walkback),
+	    Binary(s, spv::OpLogicalOr, b, found(args[2]), found(args[3])));
+	const auto header = s.builder.AllocateId(), body = s.builder.AllocateId();
+	const auto cont = s.builder.AllocateId(), merge = s.builder.AllocateId();
+	s.builder.AddFunction(spv::OpBranch, header);
+	EmitLabel(s, header);
+	const auto index = s.builder.AllocateId(), control = s.builder.AllocateId();
+	const auto valid = s.builder.AllocateId(), skip_low = s.builder.AllocateId(), skip_high = s.builder.AllocateId();
+	const auto index_phi = s.builder.AddDeferredPhi(u, index, 2);
+	const auto control_phi = s.builder.AddDeferredPhi(u, control, 2);
+	const auto valid_phi = s.builder.AddDeferredPhi(b, valid, 2);
+	const auto low_phi = s.builder.AddDeferredPhi(b, skip_low, 2);
+	const auto high_phi = s.builder.AddDeferredPhi(b, skip_high, 2);
+	const auto more = Binary(s, spv::OpLogicalAnd, b, valid,
+	    Binary(s, spv::OpULessThan, b, index, ConstantU32(s, 8)));
+	s.builder.AddFunction(spv::OpLoopMerge, merge, cont, spv::LoopControlDontUnrollMask);
+	s.builder.AddFunction(spv::OpBranchConditional, more, body, merge);
+	EmitLabel(s, body);
+	const auto high = Binary(s, spv::OpUGreaterThanEqual, b, index, ConstantU32(s, 4));
+	const auto high4 = s.builder.AllocateId();
+	s.builder.AddFunction(spv::OpCompositeConstruct, TypeBoolVector(s, 4), high4, high, high, high, high);
+	const auto values = Select(s, vec4, high4, args[3], args[2]);
+	const auto node = s.builder.AllocateId();
+	s.builder.AddFunction(spv::OpVectorExtractDynamic, u, node, values, Bits(s, index, 0, 2));
+	const auto sentinel = Binary(s, spv::OpIEqual, b, Bits(s, node, 3, 29), ConstantU32(s, 0x1fffffff));
+	const auto enabled = Unary(s, spv::OpLogicalNot, b, Select(s, b, high, skip_high, skip_low));
+	const auto marker = Binary(s, spv::OpLogicalAnd, b, sentinel, enabled);
+	const auto next_low = Binary(s, spv::OpLogicalOr, b, skip_low,
+	    Binary(s, spv::OpLogicalAnd, b, marker, Binary(s, spv::OpIEqual, b, Bits(s, node, 1, 1), zero)));
+	const auto next_high = Binary(s, spv::OpLogicalOr, b, skip_high,
+	    Binary(s, spv::OpLogicalAnd, b, high,
+	        Binary(s, spv::OpLogicalAnd, b, marker, Binary(s, spv::OpIEqual, b, Bits(s, node, 2, 1), zero))));
+	const auto next_valid = Unary(s, spv::OpLogicalNot, b,
+	    Binary(s, spv::OpLogicalAnd, b, walkback, Binary(s, spv::OpIEqual, b, node, last)));
+	const auto push = Binary(s, spv::OpLogicalAnd, b, next_valid,
+	    Unary(s, spv::OpLogicalNot, b,
+	        Binary(s, spv::OpLogicalOr, b, sentinel, Select(s, b, high, next_high, next_low))));
+	const auto next_control = EmitValueOrDefaultIfCondition(s, push, u, control, [&] {
+		return StackPush(s, memory, control, node, capacity);
+	});
+	s.builder.AddFunction(spv::OpBranch, cont);
+	EmitLabel(s, cont);
+	const auto next_index = Binary(s, spv::OpISub, u, index, one);
+	s.builder.AddFunction(spv::OpBranch, header);
+	s.builder.PatchDeferredPhi(index_phi, 0, ConstantU32(s, 7), entry);
+	s.builder.PatchDeferredPhi(index_phi, 1, next_index, cont);
+	s.builder.PatchDeferredPhi(control_phi, 0, args[0], entry);
+	s.builder.PatchDeferredPhi(control_phi, 1, next_control, cont);
+	s.builder.PatchDeferredPhi(valid_phi, 0, initial_valid, entry);
+	s.builder.PatchDeferredPhi(valid_phi, 1, next_valid, cont);
+	s.builder.PatchDeferredPhi(low_phi, 0, ConstantBool(s, false), entry);
+	s.builder.PatchDeferredPhi(low_phi, 1, next_low, cont);
+	s.builder.PatchDeferredPhi(high_phi, 0, ConstantBool(s, false), entry);
+	s.builder.PatchDeferredPhi(high_phi, 1, next_high, cont);
+	EmitLabel(s, merge);
+	s.builder.AddFunction(spv::OpReturnValue, StackPop(s, memory, control, capacity));
+	s.builder.AddFunction(spv::OpFunctionEnd);
+	if (local) s.lds_variable = old_lds;
+	if (buffer) s.lds_base_dwords = old_base;
+}
+
+} // namespace
+
+void DefineTrinityFunctions(EmitterState& s) {
+	DefineBvhStack(s);
+	DefineTrinityBvhIntersect(s);
+}
+
+uint32_t EmitBvhStack(ValueEmitContext& ctx, const IR::Inst& inst) {
+	auto& s = ctx.state;
+	const auto capacity = ctx.Memory(inst).offset;
+	if (s.bvh_stack_function == 0 || capacity < 8 || capacity > 31)
+		ctx.Fail(inst, "invalid BVH stack function or capacity");
+	const std::array values {ctx.Arg(inst, 0), ctx.Arg(inst, 1), ctx.Arg(inst, 2), ctx.Arg(inst, 3)};
+	return EmitValueOrDefaultIfCondition(s, ctx.Arg(inst, 4), TypeU32Pair(s),
+	    ConstantU32CompositeZero(s, 2), [&] {
+		const auto result = s.builder.AllocateId();
+		std::vector<uint32_t> call {spv::OpFunctionCall, TypeU32Pair(s), result,
+		    s.bvh_stack_function, values[0], values[1], values[2], values[3], ConstantU32(s, capacity)};
+		if (s.lds_storage_class == spv::StorageClassFunction) call.push_back(s.lds_variable);
+		if (s.lds_storage_class == spv::StorageClassStorageBuffer) call.push_back(s.lds_base_dwords);
+		s.builder.AddFunction(call);
+		return result;
+	});
+}
+
+} // namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter

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

@@ -0,0 +1,511 @@
+#include "graphics/shader/recompiler/backend/spirv/spirvEmitterInstructions.h"
+
+namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter {
+
+uint32_t TypeBvh8Result(EmitterState& s) {
+	if (s.bvh8_result_type == 0) {
+		const auto v = TypeU32Vector(s, 4);
+		s.bvh8_result_type = s.builder.Type(spv::OpTypeStruct, v, v, v, v, TypeU32(s));
+	}
+	return s.bvh8_result_type;
+}
+
+namespace {
+using Vec3 = std::array<uint32_t, 3>;
+
+struct Bvh8Emitter {
+	EmitterState& s;
+	const uint32_t u = TypeU32(s), f = TypeF32(s), b = TypeBool(s), w = TypeU64(s);
+	const uint32_t v4 = TypeU32Vector(s, 4), v3 = TypeF32Vector(s, 3);
+	uint32_t triangle_function = 0, boxes_function = 0, decompress_function = 0;
+
+	uint32_t C(uint32_t x) { return ConstantU32(s, x); }
+	uint32_t And(uint32_t x, uint32_t y) { return Binary(s, spv::OpBitwiseAnd, u, x, y); }
+	uint32_t Or(uint32_t x, uint32_t y) { return Binary(s, spv::OpBitwiseOr, u, x, y); }
+	uint32_t Add(uint32_t x, uint32_t y) { return EmitAddU32(s, x, y); }
+	uint32_t Sub(uint32_t x, uint32_t y) { return Binary(s, spv::OpISub, u, x, y); }
+	uint32_t Shl(uint32_t x, uint32_t n) { return Binary(s, spv::OpShiftLeftLogical, u, x, C(n)); }
+	uint32_t Shr(uint32_t x, uint32_t n) { return Binary(s, spv::OpShiftRightLogical, u, x, C(n)); }
+	uint32_t Eq(uint32_t x, uint32_t y) { return Binary(s, spv::OpIEqual, b, x, y); }
+	uint32_t Both(uint32_t x, uint32_t y) { return Binary(s, spv::OpLogicalAnd, b, x, y); }
+	uint32_t Either(uint32_t x, uint32_t y) { return Binary(s, spv::OpLogicalOr, b, x, y); }
+	uint32_t Not(uint32_t x) { return Unary(s, spv::OpLogicalNot, b, x); }
+	uint32_t Bit(uint32_t x, uint32_t n) { return Not(Eq(And(x, C(1u << n)), C(0))); }
+	uint32_t Choose(uint32_t condition, uint32_t x, uint32_t y) { return Select(s, u, condition, x, y); }
+	uint32_t Float(uint32_t x) { return Unary(s, spv::OpBitcast, f, x); }
+	uint32_t Raw(uint32_t x) { return Unary(s, spv::OpBitcast, u, x); }
+	uint32_t Fop(spv::Op op, uint32_t x, uint32_t y) { return Binary(s, op, f, x, y); }
+	uint32_t Fcmp(spv::Op op, uint32_t x, uint32_t y) { return Binary(s, op, b, x, y); }
+	uint32_t Min(uint32_t x, uint32_t y) { return EmitGlsl<GLSLstd450FMin, IR::Type::F32>(s, x, y); }
+	uint32_t Max(uint32_t x, uint32_t y) { return EmitGlsl<GLSLstd450FMax, IR::Type::F32>(s, x, y); }
+	uint32_t Extract(uint32_t type, uint32_t value, uint32_t index) {
+		const auto id = s.builder.AllocateId();
+		s.builder.AddFunction(spv::OpCompositeExtract, type, id, value, index);
+		return id;
+	}
+	uint32_t Make(uint32_t type, std::span<const uint32_t> values) {
+		const auto id = s.builder.AllocateId();
+		s.builder.AddFunction(spv::OpCompositeConstruct, type, id, values);
+		return id;
+	}
+	uint32_t Make4(uint32_t x, uint32_t y, uint32_t z, uint32_t t) {
+		return Make(v4, std::array{x, y, z, t});
+	}
+	Vec3 Unpack3(uint32_t value) {
+		return {Extract(f, value, 0), Extract(f, value, 1), Extract(f, value, 2)};
+	}
+	uint32_t Word(uint32_t block, uint32_t index) {
+		const auto offset = Unary(s, spv::OpUConvert, w, Shl(index, 2));
+		const auto address = Binary(s, spv::OpIAdd, w, block, offset);
+		const auto pointer = Unary(s, spv::OpConvertUToPtr, TypePhysicalU32Pointer(s), address);
+		const auto id = s.builder.AllocateId();
+		s.builder.AddFunction(spv::OpLoad, u, id, pointer, spv::MemoryAccessAlignedMask, 4u);
+		return id;
+	}
+	uint32_t Call(uint32_t type, uint32_t function, std::span<const uint32_t> args) {
+		const auto id = s.builder.AllocateId();
+		std::vector<uint32_t> call {spv::OpFunctionCall, type, id, function};
+		call.insert(call.end(), args.begin(), args.end());
+		s.builder.AddFunction(call);
+		return id;
+	}
+	std::vector<uint32_t> Begin(uint32_t& function, const char* name, uint32_t result,
+	                            std::span<const uint32_t> types) {
+		function = s.builder.AllocateId();
+		s.builder.AddName(function, name);
+		s.builder.AddFunction(spv::OpFunction, result, function, spv::FunctionControlMaskNone,
+		    s.builder.Type(spv::OpTypeFunction, result, types));
+		std::vector<uint32_t> args;
+		for (auto type: types) {
+			args.push_back(s.builder.AllocateId());
+			s.builder.AddFunction(spv::OpFunctionParameter, type, args.back());
+		}
+		EmitLabel(s, s.builder.AllocateId());
+		return args;
+	}
+	void End(uint32_t result) {
+		s.builder.AddFunction(spv::OpReturnValue, result);
+		s.builder.AddFunction(spv::OpFunctionEnd);
+	}
+	uint32_t Result(const std::array<uint32_t, 10>& dst, const Vec3& org, const Vec3& dir,
+	                uint32_t instance) {
+		const std::array groups {Make4(dst[0], dst[1], dst[2], dst[3]),
+		    Make4(dst[4], dst[5], dst[6], dst[7]), Make4(dst[8], dst[9], Raw(org[0]), Raw(org[1])),
+		    Make4(Raw(org[2]), Raw(dir[0]), Raw(dir[1]), Raw(dir[2])), instance};
+		return Make(TypeBvh8Result(s), groups);
+	}
+	uint32_t NormalizeFlags(uint32_t flags) {
+		flags = Choose(Bit(flags, 2), And(flags, C(~0x30u)), flags);
+		return Or(flags, Choose(Eq(And(flags, C(0x30)), C(0x30)), C(0x100), C(0)));
+	}
+
+	void DefineTriangle() {
+		const auto a = Begin(triangle_function, "trinity_triangle", v4,
+		    std::array{w, u, u, u, v3, v3});
+		const auto block = a[0], slot = a[1], flags = a[2], descriptor = a[3];
+		const auto origin = Unpack3(a[4]), direction = Unpack3(a[5]);
+		const auto flag_word = Word(block, C(15));
+		const auto remap = Binary(s, spv::OpShiftRightLogical, u, flag_word, Shl(slot, 3));
+		const auto bary = Bit(descriptor, 24);
+		const auto pointer_flags = Both(bary, Bit(descriptor, 23));
+		const auto tri_flags = Choose(pointer_flags, remap, C(0));
+		const auto procedural = Bit(tri_flags, 6);
+		const auto raw = EmitValueOrDefaultIfCondition(s, Not(procedural), v4,
+		    Make4(C(0), C(0x3f800000), C(0x3f800000), C(0x3f800000)), [&] {
+			std::array<Vec3, 3> positions;
+			const std::array vertices {slot, Choose(Eq(slot, C(0)), C(1), C(3)), C(2)};
+			for (uint32_t vertex = 0; vertex < 3; ++vertex)
+				for (uint32_t axis = 0; axis < 3; ++axis)
+					positions[vertex][axis] = Float(Word(block,
+					    Add(Binary(s, spv::OpIMul, u, vertices[vertex], C(3)), C(axis))));
+			return EmitBvhTriangle(s, positions, origin, direction, remap, Add(flag_word, slot), bary);
+		});
+		std::array<uint32_t, 4> result;
+		for (uint32_t i = 0; i < 4; ++i) result[i] = Extract(u, raw, i);
+		auto missed = Eq(result[0], C(0x7f800000));
+		const auto flip = Both(Both(Not(procedural), Not(missed)),
+		    Binary(s, spv::OpLogicalNotEqual, b, Bit(flags, 3), Bit(tri_flags, 5)));
+		for (uint32_t i = 0; i < 4; ++i)
+			result[i] = Binary(s, spv::OpBitwiseXor, u, result[i],
+			    Choose(i < 2 ? flip : Both(flip, bary), C(0x80000000), C(0)));
+		const auto front = Fcmp(spv::OpFOrdGreaterThan, Float(result[1]), ConstantF32Value(s, 0.0f));
+		const auto opaque = Both(Either(Bit(tri_flags, 7), Bit(flags, 0)), Not(Bit(flags, 1)));
+		missed = Either(missed, Either(Both(opaque, Bit(flags, 6)), Both(Not(opaque), Bit(flags, 7))));
+		missed = Either(missed, Both(Both(Not(Bit(tri_flags, 4)), Not(procedural)),
+		    Select(s, b, front, Bit(flags, 5), Bit(flags, 4))));
+		missed = Either(missed, Select(s, b, procedural, Bit(flags, 9), Bit(flags, 8)));
+		const auto sign = [&](uint32_t value, uint32_t enabled) {
+			return Choose(pointer_flags, Or(And(value, C(0x7fffffff)),
+			    Choose(enabled, C(0x80000000), C(0))), value);
+		};
+		result[2] = sign(result[2], Either(Not(opaque), procedural));
+		result[3] = sign(result[3], Not(opaque));
+		const std::array miss {C(0x7f800000), C(0x3f800000), Choose(bary, C(0), C(~0u)), C(0)};
+		for (uint32_t i = 0; i < 4; ++i) result[i] = Choose(missed, miss[i], result[i]);
+		End(Make(v4, result));
+	}
+
+	uint32_t Instance(uint32_t block, uint32_t flags, uint32_t mask, Vec3 org, Vec3 dir) {
+		const auto lo = Word(block, C(12)), hi = Word(block, C(13));
+		const auto inst_flags = Shr(hi, 22);
+		auto merged = Or(And(Or(flags, inst_flags), C(0x3f0)), And(inst_flags, C(4)));
+		merged = Or(merged, And(Binary(s, spv::OpBitwiseXor, u, flags, inst_flags), C(8)));
+		merged = Or(merged, Choose(Either(Bit(flags, 0), Both(Bit(inst_flags, 0), Not(Bit(flags, 1)))), C(1), C(0)));
+		merged = Or(merged, Choose(Either(Bit(flags, 1), Both(Bit(inst_flags, 1), Not(Bit(flags, 0)))), C(2), C(0)));
+		const auto user = Word(block, C(15));
+		std::array<uint32_t, 10> dst;
+		dst.fill(C(~0u));
+		dst[2] = lo;
+		dst[3] = Or(And(hi, C(0x3fffff)), Shl(merged, 22));
+		dst[6] = And(user, C(0xffffff));
+		dst[7] = Choose(Eq(And(mask, Shr(user, 24)), C(0)), C(~0u), Word(block, C(14)));
+		Vec3 transformed_org, transformed_dir;
+		for (uint32_t axis = 0; axis < 3; ++axis) {
+			Vec3 row;
+			for (uint32_t i = 0; i < 3; ++i) row[i] = Float(Word(block, C(axis * 4 + i)));
+			const auto dot = [&](const Vec3& value) {
+				return Fop(spv::OpFAdd, Fop(spv::OpFAdd, Fop(spv::OpFMul, row[0], value[0]),
+				    Fop(spv::OpFMul, row[1], value[1])), Fop(spv::OpFMul, row[2], value[2]));
+			};
+			transformed_org[axis] = Fop(spv::OpFAdd, dot(org), Float(Word(block, C(axis * 4 + 3))));
+			transformed_dir[axis] = dot(dir);
+		}
+		return Result(dst, transformed_org, transformed_dir, C(1));
+	}
+
+	void DefineDecompress();
+	void DefineBoxes();
+	uint32_t Boxes(uint32_t first, uint32_t second, uint32_t full, uint32_t compressed,
+	               uint32_t flags, uint32_t descriptor, uint32_t extent, uint32_t org, uint32_t dir);
+	void DefineIntersect();
+};
+
+void Bvh8Emitter::DefineDecompress() {
+	const auto a = Begin(decompress_function, "trinity_shared_exponent", f, std::array{u, u, b});
+	const auto exponent = a[0], packed = a[1], minimum = a[2];
+	const auto negative = Bit(packed, 17), mantissa = And(packed, C(0x1ffff));
+	const auto outward = Eq(Choose(negative, C(1), C(0)), Choose(minimum, C(1), C(0)));
+	const auto nonzero = Not(Eq(mantissa, C(0)));
+	const auto highest = EmitGlsl<GLSLstd450FindUMsb, IR::Type::U32>(s, mantissa);
+	const auto shift = EmitGlsl<GLSLstd450UMin, IR::Type::U32>(s, Sub(C(16), highest), exponent);
+	const auto decoded_exp = Choose(nonzero, Sub(exponent, shift),
+	    Choose(Both(outward, Binary(s, spv::OpUGreaterThan, b, exponent, C(17))), Sub(exponent, C(17)), C(0)));
+	// Zero mantissas take a separate result, but their unused shifts must still be defined.
+	const auto amount = Add(C(7), Choose(nonzero, shift, C(0)));
+	const auto trail = Choose(Both(outward, Not(Eq(decoded_exp, C(255)))),
+	    Sub(Binary(s, spv::OpShiftLeftLogical, u, C(1), amount), C(1)), C(0));
+	const auto fraction = Choose(nonzero,
+	    And(Or(Binary(s, spv::OpShiftLeftLogical, u, mantissa, amount), trail), C(0x7fffff)),
+	    Choose(outward, C(0x7fffff), C(0)));
+	End(Float(Or(Or(fraction, Shl(decoded_exp, 23)), Choose(negative, C(0x80000000), C(0)))));
+}
+
+void Bvh8Emitter::DefineBoxes() {
+	const auto result_type = s.builder.Type(spv::OpTypeStruct, v4, v4);
+	const auto a = Begin(boxes_function, "trinity_box_slot", result_type,
+	    std::array{w, w, b, b, u, u, f, v3, v3});
+	const auto first = a[0], second = a[1], full = a[2], compressed = a[3];
+	const auto flags = a[4], descriptor = a[5], extent = a[6];
+	const auto org = Unpack3(a[7]), dir = Unpack3(a[8]);
+	const auto zero = ConstantF32Value(s, 0.0f), infinity = ConstantF32(s, 0x7f800000);
+	Vec3 inverse;
+	for (uint32_t i = 0; i < 3; ++i) inverse[i] = Fop(spv::OpFDiv, ConstantF32Value(s, 1.0f), dir[i]);
+	const auto packed = [&](uint32_t offset, uint32_t width) {
+		const auto word = Shr(offset, 5), shift = And(offset, C(31));
+		const auto lo = Binary(s, spv::OpShiftRightLogical, u, Word(first, word), shift);
+		const auto next = EmitGlsl<GLSLstd450UMin, IR::Type::U32>(s, Add(word, C(1)), C(15));
+		const auto hi = Binary(s, spv::OpShiftLeftLogical, u, Word(first, next), And(Sub(C(32), shift), C(31)));
+		return And(Or(lo, Choose(Eq(shift, C(0)), C(0), hi)), C((1u << width) - 1));
+	};
+	const auto format = Choose(compressed, C(0), Choose(full, C(1), C(2)));
+	const auto entry_label = s.current_label;
+	const auto header = s.builder.AllocateId(), body = s.builder.AllocateId();
+	const auto cont = s.builder.AllocateId(), merge = s.builder.AllocateId();
+	const auto initial_nodes = Make4(C(~0u), C(~0u), C(~0u), C(~0u));
+	const auto initial_keys = Make4(C(0x7f800000), C(0x7f800000), C(0x7f800000), C(0x7f800000));
+	const auto base = And(Word(first, C(0)), C(0x1fffffff));
+	s.builder.AddFunction(spv::OpBranch, header);
+	EmitLabel(s, header);
+	const auto child = s.builder.AllocateId(), prefix = s.builder.AllocateId();
+	const auto nodes = s.builder.AllocateId(), keys = s.builder.AllocateId();
+	const auto child_phi = s.builder.AddDeferredPhi(u, child, 2);
+	const auto prefix_phi = s.builder.AddDeferredPhi(u, prefix, 2);
+	const auto nodes_phi = s.builder.AddDeferredPhi(v4, nodes, 2);
+	const auto keys_phi = s.builder.AddDeferredPhi(v4, keys, 2);
+	const auto more = Binary(s, spv::OpULessThan, b, child, C(4));
+	s.builder.AddFunction(spv::OpLoopMerge, merge, cont, spv::LoopControlDontUnrollMask);
+	s.builder.AddFunction(spv::OpBranchConditional, more, body, merge);
+	EmitLabel(s, body);
+	const auto kind = packed(Add(C(29), Binary(s, spv::OpIMul, u, child, C(3))), 3);
+	const auto direct_node = Word(first, child);
+	const auto node = Choose(compressed, Or(Shl(prefix, 3), kind), direct_node);
+	const auto child_flags = Choose(compressed,
+	    And(packed(Add(C(41), Shl(child, 2)), 4), Choose(Eq(child, C(3)), C(7), C(15))), C(0));
+	const auto step = Choose(Eq(kind, C(2)), C(4),
+	    Choose(Either(Either(Eq(kind, C(3)), Eq(kind, C(5))), Eq(kind, C(6))), C(2), C(1)));
+	const auto next_prefix = Add(prefix, step);
+	const auto bounds_type = s.builder.Type(spv::OpTypeStruct, v3, v3);
+	const auto box = EmitIndexSwitch(s, format, 3, bounds_type, [&](uint32_t encoding) {
+		std::array<uint32_t, 6> bounds;
+		if (encoding == 0) {
+			const auto yz = Word(first, C(2));
+			const std::array exponents {Shr(Word(first, C(1)), 24), And(yz, C(255)), And(Shr(yz, 8), C(255))};
+			for (uint32_t component = 0; component < 6; ++component) {
+				const auto offset = Add(C(80 + component * 18), Binary(s, spv::OpIMul, u, child, C(108)));
+				bounds[component] = Call(f, decompress_function,
+				    std::array{exponents[component % 3], packed(offset, 18), ConstantBool(s, component < 3)});
+			}
+		} else if (encoding == 1) {
+			for (uint32_t component = 0; component < 6; ++component) {
+				const auto word = Add(C(4 + component), Binary(s, spv::OpIMul, u, child, C(6)));
+				const auto block = Select(s, w, Binary(s, spv::OpULessThan, b, word, C(16)), first, second);
+				bounds[component] = Float(Word(block, And(word, C(15))));
+			}
+		} else {
+			for (uint32_t pair = 0; pair < 3; ++pair) {
+				const auto word = Add(C(4 + pair), Binary(s, spv::OpIMul, u, child, C(3)));
+				const auto value = EmitGlsl<GLSLstd450UnpackHalf2x16, IR::Type::F32x2>(s, Word(first, word));
+				bounds[pair * 2] = Extract(f, value, 0);
+				bounds[pair * 2 + 1] = Extract(f, value, 1);
+			}
+		}
+		return Make(bounds_type, std::array{Make(v3, std::span(bounds).first(3)),
+		    Make(v3, std::span(bounds).last(3))});
+	});
+	const auto lo_bound = Unpack3(Extract(v3, box, 0)), hi_bound = Unpack3(Extract(v3, box, 1));
+	Vec3 near, far;
+	auto valid = Not(Eq(And(child_flags, C(3)), C(3)));
+	for (uint32_t axis = 0; axis < 3; ++axis) {
+		valid = Both(valid, Fcmp(spv::OpFOrdLessThanEqual, lo_bound[axis], hi_bound[axis]));
+		const auto lo = Fop(spv::OpFMul, Fop(spv::OpFSub, lo_bound[axis], org[axis]), inverse[axis]);
+		const auto hi = Fop(spv::OpFMul, Fop(spv::OpFSub, hi_bound[axis], org[axis]), inverse[axis]);
+		const auto forward = Fcmp(spv::OpFOrdGreaterThanEqual, inverse[axis], zero);
+		near[axis] = Select(s, f, forward, lo, hi);
+		far[axis] = Select(s, f, forward, hi, lo);
+	}
+	const auto entry = Max(Max(near[0], near[1]), near[2]);
+	const auto exit = Min(Min(far[0], far[1]), far[2]);
+	const auto exit_bits = Raw(Select(s, f, Fcmp(spv::OpFOrdEqual, exit, zero), zero, exit));
+	const auto grown = Float(EmitGlsl<GLSLstd450UMin, IR::Type::U32>(s,
+	    Add(exit_bits, And(Shr(descriptor, 23), C(255))), C(0x7f800000)));
+	valid = Both(valid, Both(Fcmp(spv::OpFOrdGreaterThanEqual, exit, zero),
+	    Both(Fcmp(spv::OpFOrdLessThan, entry, extent), Fcmp(spv::OpFOrdLessThanEqual, entry, grown))));
+	const auto opaque = Both(Either(Bit(child_flags, 0), Bit(flags, 0)), Not(Bit(flags, 1)));
+	const auto nonopaque = Both(Either(Bit(child_flags, 1), Bit(flags, 1)), Not(Bit(flags, 0)));
+	const auto culled = Either(Either(Both(opaque, Bit(flags, 6)), Both(nonopaque, Bit(flags, 7))),
+	    Either(Both(Bit(child_flags, 2), Bit(flags, 8)), Both(Bit(child_flags, 3), Bit(flags, 9))));
+	valid = Both(valid, Not(culled));
+	const auto heuristic = And(Shr(descriptor, 21), C(3));
+	auto key = Max(entry, zero);
+	key = Select(s, f, Eq(heuristic, C(1)), Unary(s, spv::OpFNegate, f, Fop(spv::OpFSub, exit, entry)), key);
+	key = Select(s, f, Eq(heuristic, C(2)), Fop(spv::OpFAdd, entry, exit), key);
+	const auto next_nodes = s.builder.AllocateId(), next_keys = s.builder.AllocateId();
+	s.builder.AddFunction(spv::OpVectorInsertDynamic, v4, next_nodes, nodes, Choose(valid, node, C(~0u)), child);
+	s.builder.AddFunction(spv::OpVectorInsertDynamic, v4, next_keys, keys, Raw(Select(s, f, valid, key, infinity)), child);
+	s.builder.AddFunction(spv::OpBranch, cont);
+	EmitLabel(s, cont);
+	const auto next_child = Add(child, C(1));
+	s.builder.AddFunction(spv::OpBranch, header);
+	s.builder.PatchDeferredPhi(child_phi, 0, C(0), entry_label);
+	s.builder.PatchDeferredPhi(child_phi, 1, next_child, cont);
+	s.builder.PatchDeferredPhi(prefix_phi, 0, base, entry_label);
+	s.builder.PatchDeferredPhi(prefix_phi, 1, next_prefix, cont);
+	s.builder.PatchDeferredPhi(nodes_phi, 0, initial_nodes, entry_label);
+	s.builder.PatchDeferredPhi(nodes_phi, 1, next_nodes, cont);
+	s.builder.PatchDeferredPhi(keys_phi, 0, initial_keys, entry_label);
+	s.builder.PatchDeferredPhi(keys_phi, 1, next_keys, cont);
+	EmitLabel(s, merge);
+	End(Make(result_type, std::array{nodes, keys}));
+}
+
+uint32_t Bvh8Emitter::Boxes(uint32_t first, uint32_t second, uint32_t full, uint32_t compressed,
+                            uint32_t flags, uint32_t descriptor, uint32_t extent,
+                            uint32_t org, uint32_t dir) {
+	return Call(s.builder.Type(spv::OpTypeStruct, v4, v4), boxes_function,
+	    std::array{first, second, full, compressed, flags, descriptor, extent, org, dir});
+}
+
+void Bvh8Emitter::DefineIntersect() {
+	const auto result_type = TypeBvh8Result(s);
+	const auto a = Begin(s.bvh8_intersect_function, "trinity_bvh_intersect", result_type,
+	    std::array{v4, w, u, f, u, v3, v3});
+	const auto descriptor = a[0], encoded_base = a[1], node = a[2], extent = a[3], mask = a[4];
+	const auto org = Unpack3(a[5]), dir = Unpack3(a[6]);
+	std::array<uint32_t, 4> d;
+	for (uint32_t i = 0; i < 4; ++i) d[i] = Extract(u, descriptor, i);
+	const auto kind = And(node, C(7));
+	const auto triangle = Binary(s, spv::OpULessThan, b, kind, C(2));
+	const auto instance = Eq(kind, C(6));
+	const auto full = Either(Eq(kind, C(2)), Eq(kind, C(5)));
+	const auto big = Either(Eq(kind, C(2)), Eq(kind, C(3)));
+	const auto compressed = Both(Not(full), Bit(d[3], 20));
+	const auto flags = NormalizeFlags(Choose(Bit(d[3], 23),
+	    Unary(s, spv::OpUConvert, u, Binary(s, spv::OpShiftRightLogical, w, encoded_base, C(54))), C(0)));
+	const auto base = Binary(s, spv::OpShiftLeftLogical, w, PackU64(s, d[0], And(d[1], C(255))), C(8));
+	const auto lane_base = Binary(s, spv::OpShiftLeftLogical, w,
+	    Binary(s, spv::OpBitwiseAnd, w, encoded_base, ConstantU64(s, (uint64_t{1} << 54) - 1)), C(3));
+	const auto offset = Binary(s, spv::OpShiftLeftLogical, w, Unary(s, spv::OpUConvert, w, Shr(node, 3)), C(6));
+	const auto address = Binary(s, spv::OpIAdd, w, base, Binary(s, spv::OpIAdd, w, lane_base, offset));
+	const auto capacity = Binary(s, spv::OpIAdd, w,
+	    PackU64(s, d[2], And(d[3], C(0x3ff))), ConstantU64(s, 1));
+	const auto limit = Binary(s, spv::OpIAdd, w, base,
+	    Binary(s, spv::OpShiftLeftLogical, w, capacity, C(6)));
+	auto valid = Both(Eq(Shr(d[3], 28), C(8)),
+	    Eq(And(Unary(s, spv::OpUConvert, u, encoded_base), C(7)), C(0)));
+	valid = Both(valid, Both(Not(Eq(kind, C(7))), Either(Not(instance), Bit(d[3], 22))));
+	valid = Both(valid, Not(Unary(s, spv::OpIsNan, b, extent)));
+	for (uint32_t i = 0; i < 3; ++i) {
+		valid = Both(valid, Not(Either(Unary(s, spv::OpIsNan, b, org[i]), Unary(s, spv::OpIsInf, b, org[i]))));
+		valid = Both(valid, Not(Unary(s, spv::OpIsNan, b, dir[i])));
+	}
+	const auto opaque = Bit(flags, 0), nonopaque = Bit(flags, 1);
+	valid = Both(valid, Not(Either(Either(Both(opaque, Bit(flags, 6)), Both(nonopaque, Bit(flags, 7))),
+	    Both(Bit(flags, 6), Bit(flags, 7)))));
+	const auto fetch_size = Select(s, w, full, ConstantU64(s, 128), ConstantU64(s, 64));
+	const auto fetch = [&](uint32_t at, uint32_t enabled) {
+		const auto end = Binary(s, spv::OpIAdd, w, at, fetch_size);
+		const auto in_bounds = Both(Binary(s, spv::OpULessThanEqual, b, end, limit),
+		    Binary(s, spv::OpULessThanEqual, b, end, ConstantU64(s, (uint64_t{1} << 48) - 1)));
+		return EmitValueOrDefaultIfCondition(s, Both(Both(valid, enabled), in_bounds), w,
+		    ConstantU64(s, 0), [&] { return GetBdaPointer(s, at); });
+	};
+	const auto first = fetch(address, ConstantBool(s, true));
+	const auto present = [&](uint32_t pointer) { return Not(Binary(s, spv::OpIEqual, b, pointer, ConstantU64(s, 0))); };
+	const auto tail = [&](uint32_t at, uint32_t pointer) {
+		return EmitValueOrDefaultIfCondition(s, Both(full, present(pointer)), w, ConstantU64(s, 0), [&] {
+			return GetBdaPointer(s, Binary(s, spv::OpIAdd, w, at, ConstantU64(s, 64)));
+		});
+	};
+	const auto second = tail(address, first);
+	const auto available = Both(present(first), Either(Not(full), present(second)));
+	std::array<uint32_t, 10> invalid;
+	invalid.fill(C(~0u));
+	for (uint32_t slot = 0; slot < 2; ++slot) {
+		invalid[slot * 4] = Choose(triangle, C(0x7f800000), C(~0u));
+		invalid[slot * 4 + 1] = Choose(triangle, C(0x3f800000), C(~0u));
+		invalid[slot * 4 + 3] = Choose(triangle, C(0), C(~0u));
+	}
+	const auto default_result = Result(invalid, org, dir, C(0));
+	const auto value = EmitValueOrDefaultIfCondition(s, Both(instance, available), result_type, default_result, [&] {
+		return Instance(first, flags, mask, org, dir);
+	});
+	const auto result = EmitValueOrDefaultIfCondition(s, Both(Not(instance), valid), result_type, value, [&] {
+		const auto second_address = Binary(s, spv::OpIAdd, w, address, fetch_size);
+		const auto next_first = fetch(second_address, big);
+		const auto next_second = tail(second_address, next_first);
+		const auto next_available = Both(present(next_first), Either(Not(full), present(next_second)));
+		const auto triangles = EmitValueOrDefaultIfCondition(s, triangle, result_type, default_result, [&] {
+			auto dst = invalid;
+			for (uint32_t slot = 0; slot < 2; ++slot) {
+				const auto active = Both(available, slot == 0 ? ConstantBool(s, true) : Eq(kind, C(1)));
+				const auto absent = Both(available, Both(Eq(kind, C(0)), Bit(d[3], 24)));
+				const auto miss = Make4(C(0x7f800000), C(0x3f800000),
+				    slot == 1 ? Choose(absent, C(0), C(~0u)) : C(~0u), C(0));
+				const auto hit = EmitValueOrDefaultIfCondition(s, active, v4, miss, [&] {
+					return Call(v4, triangle_function, std::array{first, C(slot), flags, d[3], a[5], a[6]});
+				});
+				for (uint32_t i = 0; i < 4; ++i) dst[slot * 4 + i] = Extract(u, hit, i);
+				dst[8 + slot] = EmitValueOrDefaultIfCondition(s, active, u, C(~0u), [&] { return Word(first, C(14)); });
+			}
+			return Result(dst, org, dir, C(0));
+		});
+		return EmitValueOrDefaultIfCondition(s, Not(triangle), result_type, triangles, [&] {
+			auto dst = invalid;
+			const auto slot_type = s.builder.Type(spv::OpTypeStruct, v4, v4);
+			const auto miss = Make(slot_type, std::array{Make4(C(~0u), C(~0u), C(~0u), C(~0u)),
+			    Make4(C(0x7f800000), C(0x7f800000), C(0x7f800000), C(0x7f800000))});
+			std::array<uint32_t, 8> keys;
+			for (uint32_t slot = 0; slot < 2; ++slot) {
+				const auto hit = EmitValueOrDefaultIfCondition(s, slot == 0 ? available : next_available,
+				    slot_type, miss, [&] {
+					return Boxes(slot == 0 ? first : next_first, slot == 0 ? second : next_second,
+					    full, compressed, flags, d[1], extent, a[5], a[6]);
+				});
+				const auto indices = Extract(v4, hit, 0), depths = Extract(v4, hit, 1);
+				for (uint32_t i = 0; i < 4; ++i) {
+					dst[slot * 4 + i] = Extract(u, indices, i);
+					keys[slot * 4 + i] = Extract(u, depths, i);
+				}
+			}
+			const auto sort = Bit(d[1], 31), wide = Both(sort, Bit(d[3], 21)), tri_first = Bit(d[1], 20);
+			constexpr std::array<std::pair<uint32_t, uint32_t>, 19> network {{{0,1}, {2,3}, {0,2}, {1,3}, {1,2},
+			    {4,5}, {6,7}, {4,6}, {5,7}, {5,6}, {0,4}, {1,5}, {1,4}, {2,6}, {3,7}, {3,6}, {2,4}, {3,5}, {3,4}}};
+			const auto unsorted = Make(slot_type, std::array{Make(v4, std::span(dst).first(4)),
+			    Make(v4, std::span(dst).subspan(4, 4))});
+			const auto sorted_result = EmitValueOrDefaultIfCondition(s, sort, slot_type, unsorted, [&] {
+				auto sorted_dst = dst;
+				for (size_t step = 0; step < network.size(); ++step) {
+					const auto [x, y] = network[step];
+					const auto is_tri = [&](uint32_t index) { return Binary(s, spv::OpULessThan, b, And(index, C(7)), C(2)); };
+					const auto xt = is_tri(sorted_dst[x]), yt = is_tri(sorted_dst[y]);
+					auto swap = Fcmp(spv::OpFOrdGreaterThan, Float(keys[x]), Float(keys[y]));
+					swap = Select(s, b, Both(tri_first, Binary(s, spv::OpLogicalNotEqual, b, xt, yt)), yt, swap);
+					if (step >= 10) swap = Both(wide, swap);
+					const auto key = keys[x], index = sorted_dst[x];
+					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(x)) keys[x] = Choose(swap, keys[y], key);
+					if (later(y)) keys[y] = Choose(swap, key, keys[y]);
+					sorted_dst[x] = Choose(swap, sorted_dst[y], index);
+					sorted_dst[y] = Choose(swap, index, sorted_dst[y]);
+				}
+				return Make(slot_type, std::array{Make(v4, std::span(sorted_dst).first(4)),
+				    Make(v4, std::span(sorted_dst).subspan(4, 4))});
+			});
+			for (uint32_t slot = 0; slot < 2; ++slot) {
+				const auto indices = Extract(v4, sorted_result, slot);
+				for (uint32_t i = 0; i < 4; ++i) dst[slot * 4 + i] = Extract(u, indices, i);
+			}
+			return Result(dst, org, dir, C(0));
+		});
+	});
+	End(result);
+}
+
+} // namespace
+
+void DefineTrinityBvhIntersect(EmitterState& s) {
+	if (!s.program.info.bvh8) return;
+	Bvh8Emitter e {s};
+	e.DefineDecompress();
+	e.DefineTriangle();
+	e.DefineBoxes();
+	e.DefineIntersect();
+}
+
+uint32_t EmitBvh8Extract(ValueEmitContext& ctx, const IR::Inst& inst) {
+	const auto index = inst.Arg(1).U32();
+	if (index > 16) ctx.Fail(inst, "invalid BVH8 result component");
+	const auto result = ctx.state.builder.AllocateId();
+	if (index == 16)
+		ctx.state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(ctx.state), result,
+		    ctx.Arg(inst, 0), 4u);
+	else
+		ctx.state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(ctx.state), result,
+		    ctx.Arg(inst, 0), index / 4, index % 4);
+	return result;
+}
+
+uint32_t EmitBvh8Intersect(ValueEmitContext& ctx, const IR::Inst& inst) {
+	auto& s = ctx.state;
+	const auto* ray = ctx.ImageAddress(inst.Arg(1));
+	if (ray == nullptr || ray->NumArgs() < 11 || s.bvh8_intersect_function == 0)
+		ctx.Fail(inst, "BVH8 intersection requires the complete ray and shared function");
+	Bvh8Emitter e {s};
+	const auto vector = [&](uint32_t first) {
+		return e.Make(e.v3, Vec3{e.Float(ctx.Arg(*ray, first)), e.Float(ctx.Arg(*ray, first + 1)),
+		    e.Float(ctx.Arg(*ray, first + 2))});
+	};
+	const std::array args {ctx.Arg(inst, 0), PackU64(s, ctx.Arg(*ray, 0), ctx.Arg(*ray, 1)),
+	    ctx.Arg(*ray, 10), e.Float(ctx.Arg(*ray, 2)), ctx.Arg(*ray, 3), vector(4), vector(7)};
+	const auto type = TypeBvh8Result(s);
+	return EmitValueOrDefaultIfCondition(s, ctx.Arg(inst, 2), type,
+	    s.builder.Constant(spv::OpConstantNull, type), [&] {
+		return e.Call(type, s.bvh8_intersect_function, args);
+	});
+}
+
+} // namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter

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

@@ -281,6 +281,7 @@ Opcode DecodeMimgOpcode(uint32_t opcode, const MimgSampleInfo* sample, const Mim
 		case 0x09u: return Opcode::IMAGE_STORE_MIP;
 		case 0x0eu: return Opcode::IMAGE_GET_RESINFO;
 		case 0x60u: return Opcode::IMAGE_GET_LOD;
+		case 0xe5u: return Opcode::IMAGE_BVH8_INTERSECT_RAY;
 		case 0xe6u:
 		case 0xe7u: return Opcode::IMAGE_BVH_INTERSECT_RAY;
 		default: return Opcode::UNSUPPORTED;
@@ -297,6 +298,7 @@ uint32_t DecodeMimgAddressComponents(uint32_t opcode, ImageDimension dimension,
 	}
 
 	switch (opcode) {
+		case 0xe5u:
 		case 0xe6u: return 11u;
 		case 0xe7u: return 12u;
 		case 0x0eu: return 1u;
@@ -427,6 +429,36 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 		               "BVH intersection requires full-float ray DWORDs and R128/dmask:0xf");
 	}
 
+	if (opcode == 0xe5u) {
+		inst.data_components = inst.data_dwords = 10;
+		if (a16 || d16 || !r128 || inst.dmask != 0xfu ||
+		    (nsa_dwords != 0u && nsa_dwords != 2u) || vdata > 246u ||
+		    (nsa_dwords == 0u && vaddr > 245u)) {
+			SetUnsupported(inst, Family::MIMG, opcode, "invalid BVH8 ray or destination encoding");
+		}
+		if (nsa_dwords == 2u) {
+			const auto origin = inst.image_nsa_addr[3], direction = inst.image_nsa_addr[4];
+			const auto node = inst.image_nsa_addr[5];
+			if (origin > 253u || direction > 253u)
+				SetUnsupported(inst, Family::MIMG, opcode, "BVH8 ray registers exceed VGPR range");
+			for (uint32_t i = 0; i < 3; ++i) {
+				inst.image_nsa_addr[3 + i] = origin + i;
+				inst.image_nsa_addr[6 + i] = direction + i;
+			}
+			inst.image_nsa_addr[9] = node;
+		}
+		const auto origin = nsa_dwords != 0 ? inst.image_nsa_addr[3] : vaddr + 4;
+		const auto direction = nsa_dwords != 0 ? inst.image_nsa_addr[6] : vaddr + 7;
+		bool overlapping = origin < direction + 3 && direction < origin + 3;
+		for (const auto output : {2u, 3u, 6u, 7u}) {
+			const auto reg = vdata + output;
+			overlapping |= (reg >= origin && reg < origin + 3) ||
+			               (reg >= direction && reg < direction + 3);
+		}
+		if (overlapping)
+			SetUnsupported(inst, Family::MIMG, opcode, "overlapping BVH8 instance outputs are unsupported");
+	}
+
 	DecodeVectorGpr(vdata, inst.dst);
 	DecodeVectorGpr(vaddr, inst.src0);
 	DecodeScalarSource(srsrc * 4u, pc, inst.src1);

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

@@ -142,6 +142,7 @@ constexpr MemoryOpcodeInfo DS_OPCODE_LIST[] = {
     {0xb0u, Opcode::DS_WRITE_ADDTID_B32, 1, 32}, {0xb1u, Opcode::DS_READ_ADDTID_B32, 1, 32},
     {0xb2u, Opcode::DS_PERMUTE_B32, 1, 32},     {0xb3u, Opcode::DS_BPERMUTE_B32, 1, 32},
     {0xdeu, Opcode::DS_WRITE_B96, 3, 32},        {0xdfu, Opcode::DS_WRITE_B128, 4, 32},
+    {0xe1u, Opcode::DS_BVH_STACK_PUSH8_POP1_RTN_B32, 1, 32},
     {0xfeu, Opcode::DS_READ_B96, 3, 32},         {0xffu, Opcode::DS_READ_B128, 4, 32},
 };
 
@@ -232,6 +233,7 @@ bool IsDsAtomicOpcode(Opcode opcode) {
 
 uint32_t DsSourceCount(Opcode opcode) {
 	switch (opcode) {
+		case Opcode::DS_BVH_STACK_PUSH8_POP1_RTN_B32: return 3u;
 		case Opcode::DS_WRITE2_B32:
 		case Opcode::DS_WRITE2ST64_B32:
 		case Opcode::DS_WRITE2_B64:
@@ -473,6 +475,13 @@ void DecodeDs(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index,
 	if (inst.opcode == Opcode::DS_SWIZZLE_B32 && inst.offset >= 0xe000u) {
 		SetUnsupported(inst, Family::DS, opcode, "DS swizzle FFT mode is not implemented");
 	}
+	if (inst.opcode == Opcode::DS_BVH_STACK_PUSH8_POP1_RTN_B32 &&
+	    (inst.gds || inst.offset < 8u || inst.offset > 31u || data1 > 248u)) {
+		SetUnsupported(inst, Family::DS, opcode, "invalid BVH stack operands or capacity");
+	}
+	if (inst.opcode == Opcode::DS_BVH_STACK_PUSH8_POP1_RTN_B32 && vdst == addr) {
+		SetUnsupported(inst, Family::DS, opcode, "overlapping BVH stack outputs are not implemented");
+	}
 	if (inst.gds && inst.opcode == Opcode::DS_CMPST_RTN_B32) {
 		SetUnsupported(inst, Family::DS, opcode, "DS compare-store return B32 GDS variant is not implemented");
 	}

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

@@ -70,6 +70,7 @@ constexpr OpcodeMap SOP1_OPCODE_LIST[] = {
     {0x21u, Opcode::S_SWAPPC_B64},
     {0x24u, Opcode::S_AND_SAVEEXEC_B64},
     {0x28u, Opcode::S_ORN2_SAVEEXEC_B64},
+    {0x2cu, Opcode::S_QUADMASK_B32},
     {0x2du, Opcode::S_QUADMASK_B64},
     {0x34u, Opcode::S_ABS_I32},
     {0x37u, Opcode::S_ANDN1_SAVEEXEC_B64},

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

@@ -517,6 +517,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::S_NOT_B64:
 		case Opcode::S_WQM_B32:
 		case Opcode::S_WQM_B64:
+		case Opcode::S_QUADMASK_B32:
 		case Opcode::S_QUADMASK_B64:
 		case Opcode::S_AND_SAVEEXEC_B32:
 		case Opcode::S_ORN2_SAVEEXEC_B32:
@@ -739,6 +740,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::DS_READ2ST64_B64:
 		case Opcode::DS_READ_B96:
 		case Opcode::DS_READ_B128:
+		case Opcode::DS_BVH_STACK_PUSH8_POP1_RTN_B32:
 		case Opcode::DS_WRITE_B8:
 		case Opcode::DS_WRITE_B16:
 		case Opcode::DS_WRITE_B8_D16_HI:

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

@@ -63,6 +63,7 @@ enum class Opcode {
 	S_FLBIT_I32,
 	S_FLBIT_I32_I64,
 	S_BITREPLICATE_B64_B32,
+	S_QUADMASK_B32,
 	S_QUADMASK_B64,
 	S_GETPC_B64,
 	S_SETPC_B64,
@@ -248,7 +249,9 @@ enum class Opcode {
 	V_MIN_I32,
 	V_MAX_I32,
 	V_MUL_I32_I24,
+	V_MUL_HI_I32_I24,
 	V_MUL_U32_U24,
+	V_MUL_HI_U32_U24,
 	V_MUL_LO_U32,
 	V_MUL_HI_U32,
 	V_MUL_LO_I32,
@@ -257,6 +260,7 @@ enum class Opcode {
 	V_CVT_PK_U8_F32,
 	V_MAD_F32,
 	V_MAD_I16,
+	V_MAD_U16,
 	V_MAD_I32_I24,
 	V_MAD_U32_U24,
 	V_MAD_U64_U32,
@@ -277,16 +281,20 @@ enum class Opcode {
 	V_MIN3_I32,
 	V_MIN3_U32,
 	V_MIN3_F16,
+	V_MIN3_I16,
 	V_MIN3_U16,
 	V_MAX3_F32,
 	V_MAX3_I32,
 	V_MAX3_U32,
 	V_MAX3_F16,
+	V_MAX3_I16,
+	V_MAX3_U16,
 	V_MED3_F32,
 	V_MED3_I32,
 	V_MED3_U32,
 	V_MED3_F16,
 	V_MED3_I16,
+	V_MED3_U16,
 	V_SAD_U32,
 	V_ADD3_U32,
 	V_LSHL_ADD_U32,
@@ -354,6 +362,7 @@ enum class Opcode {
 	V_ASHRREV_I32,
 	V_LSHLREV_B64,
 	V_LSHRREV_B64,
+	V_ASHRREV_I64,
 	V_LSHLREV_B16,
 	V_LSHRREV_B16,
 	V_ASHRREV_I16,
@@ -676,6 +685,7 @@ enum class Opcode {
 	DS_READ2ST64_B64,
 	DS_READ_B96,
 	DS_READ_B128,
+	DS_BVH_STACK_PUSH8_POP1_RTN_B32,
 	DS_WRITE_B8,
 	DS_WRITE_B16,
 	DS_WRITE_B8_D16_HI,
@@ -691,6 +701,7 @@ enum class Opcode {
 	DS_WRITE_ADDTID_B32,
 	DS_READ_ADDTID_B32,
 	IMAGE_BVH_INTERSECT_RAY,
+	IMAGE_BVH8_INTERSECT_RAY,
 	IMAGE_GET_RESINFO,
 	IMAGE_GET_LOD,
 	IMAGE_LOAD,

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

@@ -35,7 +35,9 @@ constexpr Vop2OpcodeInfo VOP2_OPCODE_LIST[] = {
     {0x05u, Opcode::V_SUBREV_F32},
     {0x08u, Opcode::V_MUL_F32, Vop2SdwaProfile::Float32},
     {0x09u, Opcode::V_MUL_I32_I24, Vop2SdwaProfile::IntegerPartialDestination},
+    {0x0au, Opcode::V_MUL_HI_I32_I24, Vop2SdwaProfile::IntegerPartialDestination},
     {0x0bu, Opcode::V_MUL_U32_U24, Vop2SdwaProfile::IntegerPartialDestination},
+    {0x0cu, Opcode::V_MUL_HI_U32_U24, Vop2SdwaProfile::IntegerPartialDestination},
     {0x0fu, Opcode::V_MIN_F32},
     {0x10u, Opcode::V_MAX_F32},
     {0x11u, Opcode::V_MIN_I32, Vop2SdwaProfile::IntegerFullDestination},
@@ -157,6 +159,10 @@ constexpr OpcodeMap VOP3_ENCODED_VOP1_OPCODE_LIST[] = {
     {0x04u, Opcode::V_CVT_F64_I32},
     {0x0fu, Opcode::V_CVT_F32_F64},
     {0x10u, Opcode::V_CVT_F64_F32},
+    {0x11u, Opcode::V_CVT_F32_UBYTE0},
+    {0x12u, Opcode::V_CVT_F32_UBYTE1},
+    {0x13u, Opcode::V_CVT_F32_UBYTE2},
+    {0x14u, Opcode::V_CVT_F32_UBYTE3},
     {0x16u, Opcode::V_CVT_F64_U32},
     {0x17u, Opcode::V_TRUNC_F64},
     {0x18u, Opcode::V_CEIL_F64},
@@ -167,6 +173,7 @@ constexpr OpcodeMap VOP3_ENCODED_VOP1_OPCODE_LIST[] = {
     {0x07u, Opcode::V_CVT_U32_F32},
     {0x08u, Opcode::V_CVT_I32_F32},
     {0x0au, Opcode::V_CVT_F16_F32},
+    {0x0bu, Opcode::V_CVT_F32_F16},
     {0x0cu, Opcode::V_CVT_RPI_I32_F32},
     {0x0du, Opcode::V_CVT_FLR_I32_F32},
     {0x0eu, Opcode::V_CVT_OFF_F32_I4},
@@ -334,16 +341,20 @@ constexpr OpcodeMap VOP3_OPCODE_LIST[] = {
     {0x152u, Opcode::V_MIN3_I32},
     {0x153u, Opcode::V_MIN3_U32},
     {0x351u, Opcode::V_MIN3_F16},
+    {0x352u, Opcode::V_MIN3_I16},
     {0x353u, Opcode::V_MIN3_U16},
     {0x154u, Opcode::V_MAX3_F32},
     {0x155u, Opcode::V_MAX3_I32},
     {0x156u, Opcode::V_MAX3_U32},
     {0x354u, Opcode::V_MAX3_F16},
+    {0x355u, Opcode::V_MAX3_I16},
+    {0x356u, Opcode::V_MAX3_U16},
     {0x157u, Opcode::V_MED3_F32},
     {0x158u, Opcode::V_MED3_I32},
     {0x159u, Opcode::V_MED3_U32},
     {0x357u, Opcode::V_MED3_F16},
     {0x358u, Opcode::V_MED3_I16},
+    {0x359u, Opcode::V_MED3_U16},
     {0x15du, Opcode::V_SAD_U32},
     {0x15eu, Opcode::V_CVT_PK_U8_F32},
     {0x178u, Opcode::V_XOR3_B32},
@@ -354,6 +365,7 @@ constexpr OpcodeMap VOP3_OPCODE_LIST[] = {
     {0x16cu, Opcode::V_MUL_HI_I32},
     {0x2ffu, Opcode::V_LSHLREV_B64},
     {0x300u, Opcode::V_LSHRREV_B64},
+    {0x301u, Opcode::V_ASHRREV_I64},
     {0x303u, Opcode::V_ADD_NC_U16},
     {0x304u, Opcode::V_SUB_NC_U16},
     {0x305u, Opcode::V_MUL_LO_U16},
@@ -370,6 +382,7 @@ constexpr OpcodeMap VOP3_OPCODE_LIST[] = {
     {0x311u, Opcode::V_PACK_B32_F16},
     {0x314u, Opcode::V_LSHLREV_B16},
     {0x319u, Opcode::V_SUBREV_I32},
+    {0x340u, Opcode::V_MAD_U16},
     {0x345u, Opcode::V_XAD_U32},
     {0x346u, Opcode::V_LSHL_ADD_U32},
     {0x347u, Opcode::V_ADD_LSHL_U32},
@@ -488,8 +501,10 @@ bool IsNativeVop3F16TernaryOpcode(Opcode opcode) {
 }
 
 bool IsNativeVop3B16TernaryOpcode(Opcode opcode) {
-	return opcode == Opcode::V_MIN3_U16 || opcode == Opcode::V_MED3_I16 ||
-	       opcode == Opcode::V_MAD_I16;
+	return opcode == Opcode::V_MIN3_I16 || opcode == Opcode::V_MIN3_U16 ||
+	       opcode == Opcode::V_MAX3_I16 || opcode == Opcode::V_MAX3_U16 ||
+	       opcode == Opcode::V_MED3_I16 || opcode == Opcode::V_MED3_U16 ||
+	       opcode == Opcode::V_MAD_I16 || opcode == Opcode::V_MAD_U16;
 }
 
 bool IsNativeVop3B16BinaryOpcode(Opcode opcode) {
@@ -1335,6 +1350,7 @@ uint32_t NativeVop3SourceCount(Opcode opcode) {
 		case Opcode::V_SUB_NC_I16:
 		case Opcode::V_LSHLREV_B64:
 		case Opcode::V_LSHRREV_B64:
+		case Opcode::V_ASHRREV_I64:
 		case Opcode::V_LSHLREV_B16:
 		case Opcode::V_LSHRREV_B16:
 		case Opcode::V_ASHRREV_I16:
@@ -1523,7 +1539,10 @@ bool SupportsNativeVop3ResultModifiers(Opcode opcode) {
 
 bool SupportsNativeVop3Clamp(Opcode opcode) {
 	return SupportsNativeVop3ResultModifiers(opcode) || UsesInexactClampControl(opcode) ||
-	       opcode == Opcode::V_MAD_I16;
+	       opcode == Opcode::V_MAD_I16 || opcode == Opcode::V_MAD_U16 ||
+	       opcode == Opcode::V_MIN3_I16 || opcode == Opcode::V_MAX3_I16 ||
+	       opcode == Opcode::V_MAX3_U16 || opcode == Opcode::V_MED3_U16 ||
+	       opcode == Opcode::V_MUL_HI_I32_I24 || opcode == Opcode::V_MUL_HI_U32_U24;
 }
 
 bool HasUnsupportedNativeVop3Modifiers(Opcode opcode, bool permlane, bool mad_mix,
@@ -1554,7 +1573,17 @@ bool HasUnsupportedNativeVop3Modifiers(Opcode opcode, bool permlane, bool mad_mi
 		return clamp != 0u || omod != 0u || neg != 0u;
 	}
 	switch (opcode) {
+		case Opcode::V_MUL_HI_I32_I24:
+		case Opcode::V_MUL_HI_U32_U24:
+			// The high product fits in 32 bits, so clamp has no effect.
+			return abs != 0u || op_sel != 0u || omod != 0u || neg != 0u;
 		case Opcode::V_LDEXP_F32: return (abs & ~1u) != 0u || op_sel != 0u || (neg & ~1u) != 0u;
+		case Opcode::V_CVT_F32_UBYTE0:
+		case Opcode::V_CVT_F32_UBYTE1:
+		case Opcode::V_CVT_F32_UBYTE2:
+		case Opcode::V_CVT_F32_UBYTE3:
+			// Byte conversions ignore input modifiers and support all output modifiers.
+			return op_sel != 0u;
 		case Opcode::V_INTERP_P1LL_F16:
 		case Opcode::V_INTERP_P2_F16:
 			return abs != 0u || (op_sel & 0x7u) != 0u || clamp != 0u || omod != 0u || neg != 0u;

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

@@ -20,12 +20,12 @@ void Translator::Integer16Binary(const Decoder::Instruction& inst, IR::ValueOpco
 	Write16Bits(DestinationOperand(inst), result);
 }
 
-void Translator::V_MAD_I16(const Decoder::Instruction& inst) {
-	const auto lhs    = ReadU16AsU32(inst.src0, true);
-	const auto rhs    = ReadU16AsU32(inst.src1, true);
-	auto       result = ir.IAdd(ir.IMul(lhs, rhs), ReadU16AsU32(inst.src2, true));
+void Translator::Integer16Mad(const Decoder::Instruction& inst, bool sign) {
+	const auto lhs    = ReadU16AsU32(inst.src0, sign);
+	const auto rhs    = ReadU16AsU32(inst.src1, sign);
+	auto       result = ir.IAdd(ir.IMul(lhs, rhs), ReadU16AsU32(inst.src2, sign));
 	if (inst.dst.clamp) {
-		result = Saturate16(result, true);
+		result = Saturate16(result, sign);
 	}
 	Write16Bits(DestinationOperand(inst), result);
 }
@@ -147,6 +147,12 @@ void Translator::SimpleInteger(const Decoder::Instruction& inst, IR::ValueOpcode
 		const auto arg_type = IR::ArgTypeOf(opcode, index);
 		const auto& operand = SourceAt(inst, reverse && index < 2u ? 1u - index : index);
 		args[index]         = ReadOperand(operand, arg_type == IR::Type::Void ? type : arg_type);
+		if (opcode == IR::ValueOpcode::ShiftRightArithmetic64 && index == 0u &&
+		    operand.kind == Decoder::OperandKind::LiteralConstant) {
+			args[index] = ir.ConstructU64(
+			    ReadU32(operand),
+			    IR::U32(IR::Value((operand.value & 0x80000000u) ? 0xffffffffu : 0u)));
+		}
 		if (mask_shift_count && index == 1u) {
 			args[index] = ir.BitwiseAnd(IR::U32(args[index]), IR::U32(IR::Value(31u)));
 		}
@@ -263,7 +269,7 @@ void Translator::V_FFBH_32(const Decoder::Instruction& inst, bool sign) {
 	WriteOperand(DestinationOperand(inst), result);
 }
 
-void Translator::Integer24(const Decoder::Instruction& inst, bool sign, bool addend) {
+void Translator::Integer24(const Decoder::Instruction& inst, bool sign, bool addend, bool high) {
 	const auto extract24 = [&](IR::U32 value) {
 		return IR::U32(
 		    ir.Emit(sign ? IR::ValueOpcode::BitFieldSExtract : IR::ValueOpcode::BitFieldUExtract,
@@ -271,7 +277,9 @@ void Translator::Integer24(const Decoder::Instruction& inst, bool sign, bool add
 	};
 	const auto lhs    = extract24(ReadU32(inst.src0));
 	const auto rhs    = extract24(ReadU32(inst.src1));
-	auto       result = ir.IMul(lhs, rhs);
+	const auto multiply = high ? (sign ? IR::ValueOpcode::SMulHi : IR::ValueOpcode::UMulHi)
+	                           : IR::ValueOpcode::IMul32;
+	auto result = IR::U32(ir.Emit(multiply, {lhs, rhs}));
 	if (addend) {
 		result = ir.IAdd(result, ReadU32(inst.src2));
 	}
@@ -369,7 +377,7 @@ void Translator::S_BITREPLICATE_B64_B32(const Decoder::Instruction& inst) {
 	WriteOperand(DestinationOperand(inst), ir.ConstructU64(replicate(low), replicate(high)));
 }
 
-void Translator::S_QUADMASK_B64(const Decoder::Instruction& inst) {
+void Translator::S_QUADMASK(const Decoder::Instruction& inst, bool is64) {
 	const auto compact = [&](IR::U32 value) {
 		auto bits = ir.BitwiseOr(value, ir.ShiftRightLogical(value, IR::U32(IR::Value(1u))));
 		bits      = ir.BitwiseOr(bits, ir.ShiftRightLogical(bits, IR::U32(IR::Value(2u))));
@@ -382,12 +390,17 @@ void Translator::S_QUADMASK_B64(const Decoder::Instruction& inst) {
 		    ir.BitwiseOr(bits, ir.ShiftRightLogical(bits, IR::U32(IR::Value(12u)))),
 		    IR::U32(IR::Value(0xffu)));
 	};
-	const auto source = ReadU32Pair(inst.src0);
-	const auto quads  = ir.BitwiseOr(
-	    compact(source[0]), ir.ShiftLeftLogical(compact(source[1]), IR::U32(IR::Value(8u))));
-	const auto result = ir.ConstructU64(quads, IR::U32(IR::Value(0u)));
-	WriteOperand(DestinationOperand(inst), result);
-	ir.SetScc(IR::U1(ir.Emit(IR::ValueOpcode::INotEqual64, {result, IR::Value(uint64_t {0})})));
+	IR::U32 quads;
+	if (is64) {
+		const auto source = ReadU32Pair(inst.src0);
+		quads = ir.BitwiseOr(
+		    compact(source[0]), ir.ShiftLeftLogical(compact(source[1]), IR::U32(IR::Value(8u))));
+		WriteOperand(DestinationOperand(inst), ir.ConstructU64(quads, IR::U32(IR::Value(0u))));
+	} else {
+		quads = compact(ReadU32(inst.src0));
+		WriteOperand(DestinationOperand(inst), quads);
+	}
+	ir.SetScc(ir.INotEqual(quads, IR::U32(IR::Value(0u))));
 }
 
 void Translator::BFM_B32(const Decoder::Instruction& inst) {

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

@@ -326,7 +326,8 @@ IR::Value Translator::MakeImageAddress(const Decoder::Instruction& inst,
 	    inst.image_sample_flags, inst.image_dimension, inst.image_address_components);
 	EXIT_IF(count > components.size());
 	const auto nsa_components =
-	    std::min(inst.image_nsa_dwords * 4u, Decoder::MaxImageNsaAddressComponents);
+	    inst.opcode == Decoder::Opcode::IMAGE_BVH8_INTERSECT_RAY && inst.image_nsa_dwords != 0u
+	        ? 10u : std::min(inst.image_nsa_dwords * 4u, Decoder::MaxImageNsaAddressComponents);
 	for (uint32_t index = 0; index < count; index++) {
 		if (index != 0u && index - 1u < nsa_components) {
 			components[index] =
@@ -781,6 +782,15 @@ void Translator::WriteSharedU32(uint32_t width, IR::U32 address,
 	}
 }
 
+void Translator::DS_BVH_STACK(const Decoder::Instruction& inst) {
+	const auto result = ir.Emit(IR::ValueOpcode::BvhStack,
+	    {ReadU32(inst.src0), ReadU32(inst.src1), ConstructU32x4(inst.src2, 4),
+	     ConstructU32x4(OffsetOperand(inst.src2, 4), 4), ir.GetExec()},
+	    AddMemoryInfo(MemoryInfoFromDecoded(inst), inst.pc));
+	WriteOperand(inst.src0, ir.Emit(IR::ValueOpcode::CompositeExtractU32x2, {result, IR::Value(0u)}));
+	WriteOperand(inst.dst, ir.Emit(IR::ValueOpcode::CompositeExtractU32x2, {result, IR::Value(1u)}));
+}
+
 void Translator::DS_READ(const Decoder::Instruction& inst) {
 	const auto memory  = MemoryInfoFromDecoded(inst);
 	const auto address = ReadU32(MemorySourceAt(inst, 0));
@@ -939,6 +949,31 @@ void Translator::DS_PERMUTE(const Decoder::Instruction& inst, bool backward) {
 	                               {ReadU32(inst.src1), address, ir.GetExec()}));
 }
 
+void Translator::IMAGE_BVH8(const Decoder::Instruction& inst) {
+	std::array<IR::U32, 10> previous;
+	for (uint32_t i = 0; i < previous.size(); ++i)
+		previous[i] = ReadU32(OffsetOperand(inst.dst, i));
+	const auto result = ir.Emit(IR::ValueOpcode::Bvh8Intersect,
+	    {ConstructU32x4(inst.src1, 4), MakeImageAddress(inst, inst.src0), ir.GetExec()});
+	const auto get = [&](uint32_t i) {
+		return IR::U32(ir.Emit(IR::ValueOpcode::Bvh8Extract, {result, IR::Value(i)}));
+	};
+	const auto instance = ir.INotEqual(get(16), IR::U32(IR::Value(0u)));
+	for (uint32_t i = 0; i < 10; ++i) {
+		auto value = get(i);
+		if (i == 0 || i == 1 || i == 4 || i == 5 || i == 8 || i == 9)
+			value = ir.Select(instance, previous[i], value);
+		WriteOperand(OffsetOperand(inst.dst, i), value);
+	}
+	for (uint32_t i = 0; i < 6; ++i) {
+		auto reg = inst.image_nsa_dwords != 0 ? inst.image_nsa_addr[3 + i] : inst.src0.reg + 4 + i;
+		auto target = inst.src0;
+		target.reg = reg;
+		// Non-instance results may alias the ray inputs.
+		WriteOperand(target, ir.Select(instance, get(10 + i), ReadU32(target)));
+	}
+}
+
 void Translator::EmitMemory(const Decoder::Instruction& inst) {
 	switch (inst.opcode) {
 		// Translated scalar loads have no cache of their own to invalidate.
@@ -947,6 +982,7 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::BUFFER_GL1_INV:
 			ir.Emit(IR::ValueOpcode::DataCacheInvalidate);
 			return;
+		case Decoder::Opcode::IMAGE_BVH8_INTERSECT_RAY: return IMAGE_BVH8(inst);
 		case Decoder::Opcode::IMAGE_BVH_INTERSECT_RAY: {
 			const auto result = ir.Emit(IR::ValueOpcode::BvhIntersect,
 			    {ConstructU32x4(inst.src1, 4), MakeImageAddress(inst, inst.src0), ir.GetExec()},
@@ -1186,6 +1222,7 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 			return DS_APPEND_CONSUME(inst, IR::ValueOpcode::DataAppend);
 		case Decoder::Opcode::DS_WRITE_ADDTID_B32: return DS_ADDTID(inst, true);
 		case Decoder::Opcode::DS_READ_ADDTID_B32: return DS_ADDTID(inst, false);
+		case Decoder::Opcode::DS_BVH_STACK_PUSH8_POP1_RTN_B32: return DS_BVH_STACK(inst);
 		case Decoder::Opcode::DS_READ2_B32:
 		case Decoder::Opcode::DS_READ2ST64_B32:
 		case Decoder::Opcode::DS_READ2_B64:

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

@@ -248,7 +248,8 @@ void Translator::EmitScalar(const Decoder::Instruction& inst) {
 		case O::S_BITSET0_B64: return S_BITSET_B64(inst, false);
 		case O::S_BITSET1_B64: return S_BITSET_B64(inst, true);
 		case O::S_BITREPLICATE_B64_B32: return S_BITREPLICATE_B64_B32(inst);
-		case O::S_QUADMASK_B64: return S_QUADMASK_B64(inst);
+		case O::S_QUADMASK_B32: return S_QUADMASK(inst, false);
+		case O::S_QUADMASK_B64: return S_QUADMASK(inst, true);
 		case O::S_BFM_B32: return BFM_B32(inst);
 		case O::S_BFM_B64: return S_BFM_B64(inst);
 		case O::S_BFE_U32: return S_BFE_U32(inst, false);

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

@@ -1165,6 +1165,14 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 		                                               : IR::U32(IR::Value(0u));
 		entry_ir.SetExecLo(initial_lo);
 		entry_ir.SetExecHi(initial_hi);
+		const auto active_lane_count = [&] {
+			auto count = IR::U32(entry_ir.Emit(IR::ValueOpcode::BitCount32, {initial_lo}));
+			if (options.wave_size == 64u) {
+				count = entry_ir.IAdd(count, IR::U32(entry_ir.Emit(
+				    IR::ValueOpcode::BitCount32, {initial_hi})));
+			}
+			return count;
+		};
 		if (options.stage == ShaderType::Compute) {
 			const auto* cs = options.input_info.compute;
 			const auto  thread_ids =
@@ -1302,7 +1310,7 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 				                      builtin(IR::StageInputKind::WorkgroupId, 1));
 			}
 		} else if (options.stage == ShaderType::Local) {
-			entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(3), IR::U32(IR::Value(64u)));
+			entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(3), active_lane_count());
 			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(2),
 			                      builtin(IR::StageInputKind::VertexIndex));
 			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(3), IR::U32(IR::Value(0u)));
@@ -1380,19 +1388,32 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 		} else if (options.stage == ShaderType::Vertex) {
 			// Vulkan owns primitive assembly; each vertex subgroup is one NGG wave.
 			// ESVertCount describes input vertices, independently of lane ordinals.
-			auto vertex_count = IR::U32(entry_ir.Emit(IR::ValueOpcode::BitCount32, {initial_lo}));
-			if (options.wave_size == 64u) {
-				vertex_count = entry_ir.IAdd(vertex_count, IR::U32(entry_ir.Emit(
-				    IR::ValueOpcode::BitCount32, {initial_hi})));
-			}
 			entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(2),
 			                      IR::U32(IR::Value(options.wave_size << 12u)));
 			entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(3),
-			                      entry_ir.BitwiseOr(vertex_count, IR::U32(IR::Value(1u << 28u))));
+			                      entry_ir.BitwiseOr(active_lane_count(), IR::U32(IR::Value(1u << 28u))));
+			// Vulkan adds an indirect record's bases to the indices; the hardware passes them
+			// only in the user SGPRs the packet patches. SH_NOP leaves those SGPRs unchanged;
+			// it does not make the record's bases zero.
+			const auto& offsets = options.input_info.vertex->indirect_offsets;
+			const auto  index   = [&](uint8_t slot, IR::StageInputKind kind,
+			                          IR::StageInputKind base_kind) {
+				if (!options.input_info.vertex->indirect_draw) {
+					return builtin(kind);
+				}
+				const auto base = builtin(base_kind);
+				if (slot < options.user_data_count) {
+					entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(options.user_data_base + slot),
+					                      base);
+				}
+				return entry_ir.ISub(builtin(kind), base);
+			};
 			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(5),
-			                      builtin(IR::StageInputKind::VertexIndex));
+			                      index(offsets.vertex, IR::StageInputKind::VertexIndex,
+			                            IR::StageInputKind::BaseVertex));
 			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(8),
-			                      builtin(IR::StageInputKind::InstanceIndex));
+			                      index(offsets.instance, IR::StageInputKind::InstanceIndex,
+			                            IR::StageInputKind::BaseInstance));
 		}
 	}
 	const bool flush_f32_inputs = options.stage == ShaderType::Compute &&

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

@@ -105,6 +105,8 @@ private:
 	void WriteSharedU32(uint32_t width, IR::U32 address, const std::array<IR::Value, 4>& values,
 	                    const IR::MemoryInfo& memory, uint32_t pc);
 	void DS_READ(const Decoder::Instruction& inst);
+	void IMAGE_BVH8(const Decoder::Instruction& inst);
+	void DS_BVH_STACK(const Decoder::Instruction& inst);
 	void DS_READ2(const Decoder::Instruction& inst);
 	void DS_WRITE(const Decoder::Instruction& inst);
 	void DS_WRITE2(const Decoder::Instruction& inst);
@@ -167,7 +169,7 @@ private:
 	void Integer16Shift(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool arithmetic);
 	void Integer16Binary(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool sign);
 	void Integer16Ternary(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool sign);
-	void V_MAD_I16(const Decoder::Instruction& inst);
+	void Integer16Mad(const Decoder::Instruction& inst, bool sign);
 	void PackedInteger16Shift(const Decoder::Instruction& inst, IR::ValueOpcode opcode,
 	                          bool arithmetic);
 	void PackedInteger16Binary(const Decoder::Instruction& inst, IR::ValueOpcode opcode);
@@ -187,7 +189,7 @@ private:
 	void V_XOR3_B32(const Decoder::Instruction& inst);
 	void S_FF_I32_B64(const Decoder::Instruction& inst, bool find_zero, bool from_msb, bool sign);
 	void V_FFBH_32(const Decoder::Instruction& inst, bool sign);
-	void Integer24(const Decoder::Instruction& inst, bool sign, bool addend);
+	void Integer24(const Decoder::Instruction& inst, bool sign, bool addend, bool high = false);
 	void V_MAD_U64_U32(const Decoder::Instruction& inst);
 	void V_SAD_U32(const Decoder::Instruction& inst);
 	void V_ADD3_U32(const Decoder::Instruction& inst);
@@ -196,7 +198,7 @@ private:
 	void V_BCNT_U32_B32(const Decoder::Instruction& inst);
 	void V_MBCNT_U32_B32(const Decoder::Instruction& inst, bool low);
 	void S_BITREPLICATE_B64_B32(const Decoder::Instruction& inst);
-	void S_QUADMASK_B64(const Decoder::Instruction& inst);
+	void S_QUADMASK(const Decoder::Instruction& inst, bool is64);
 	void BFM_B32(const Decoder::Instruction& inst);
 	IR::U32 RightMask32(IR::U32 count);
 	void    S_BFM_B64(const Decoder::Instruction& inst);

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

@@ -375,9 +375,14 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_SUB_NC_U16:
 		case O::V_SUB_NC_I16: return Integer16Binary(inst, IR::ValueOpcode::ISub32, false);
 		case O::V_MUL_LO_U16: return Integer16Binary(inst, IR::ValueOpcode::IMul32, false);
-		case O::V_MAD_I16: return V_MAD_I16(inst);
+		case O::V_MAD_I16: return Integer16Mad(inst, true);
+		case O::V_MAD_U16: return Integer16Mad(inst, false);
 		case O::V_MED3_I16: return Integer16Ternary(inst, IR::ValueOpcode::SMedTri32, true);
 		case O::V_MIN3_U16: return Integer16Ternary(inst, IR::ValueOpcode::UMinTri32, false);
+		case O::V_MIN3_I16: return Integer16Ternary(inst, IR::ValueOpcode::SMinTri32, true);
+		case O::V_MAX3_I16: return Integer16Ternary(inst, IR::ValueOpcode::SMaxTri32, true);
+		case O::V_MAX3_U16: return Integer16Ternary(inst, IR::ValueOpcode::UMaxTri32, false);
+		case O::V_MED3_U16: return Integer16Ternary(inst, IR::ValueOpcode::UMedTri32, false);
 		case O::V_MIN_I16: return Integer16Binary(inst, IR::ValueOpcode::SMin32, true);
 		case O::V_MAX_I16: return Integer16Binary(inst, IR::ValueOpcode::SMax32, true);
 		case O::V_MIN_U16: return Integer16Binary(inst, IR::ValueOpcode::UMin32, false);
@@ -565,6 +570,9 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_LSHRREV_B64:
 			return SimpleInteger(inst, IR::ValueOpcode::ShiftRightLogical64, IR::Type::U64, true,
 			                     false, false);
+		case O::V_ASHRREV_I64:
+			return SimpleInteger(inst, IR::ValueOpcode::ShiftRightArithmetic64, IR::Type::U64, true,
+			                     false, false);
 
 		case O::V_XNOR_B32:
 			return ComposedIntegerBinary(inst, IR::ValueOpcode::BitwiseXor32, false, true, false);
@@ -578,6 +586,8 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_MAD_U32_U24: return Integer24(inst, false, true);
 		case O::V_MUL_I32_I24: return Integer24(inst, true, false);
 		case O::V_MUL_U32_U24: return Integer24(inst, false, false);
+		case O::V_MUL_HI_I32_I24: return Integer24(inst, true, false, true);
+		case O::V_MUL_HI_U32_U24: return Integer24(inst, false, false, true);
 		case O::V_MAD_U64_U32: return V_MAD_U64_U32(inst);
 		case O::V_SAD_U32: return V_SAD_U32(inst);
 		case O::V_ADD3_U32: return V_ADD3_U32(inst);

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

@@ -118,15 +118,14 @@ ResourcePlan& ResourcePlan::operator=(ResourcePlan&& other) noexcept {
 }
 
 Program::~Program() {
-	// Planning expressions can refer to block values but outlive block storage in the base class.
+	// Planning expressions can refer to block values but outlive block storage in the base class,
+	// and values cross block boundaries: detach everything before any storage is destroyed.
 	for (auto& inst: value_storage) {
-		inst.Invalidate();
+		inst.DropArgs();
 	}
-	// Values may cross block boundaries. Detach all arguments before any block starts destroying
-	// its instruction storage so reverse-use links always point to live definitions.
 	for (auto* block: blocks) {
 		for (auto& inst: *block) {
-			inst.Invalidate();
+			inst.DropArgs();
 		}
 	}
 }

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

@@ -155,7 +155,7 @@ KYTY_IO_FIELDS(SampledResourcePair, a, b, c)
 KYTY_IO_FIELDS(StageInput, a, b, c, d, e)
 KYTY_IO_FIELDS(StageOutput, a, b, c, d)
 KYTY_IO_FIELDS(ShaderInfo, a, b, c, d, e, f, g, h, i, j, k, l, m, n, o, p, q, r, s, t, u, v, w, x, y, z,
-               A, B, C, D, E, F, G, H, I, J, K)
+               A, B, C, D, E, F, G, H, I, J, K, L, M)
 KYTY_IO_FIELDS(DescriptorSource::IndirectDescriptor::SelectorRead::CountedRows, a, b)
 KYTY_IO_FIELDS(DescriptorSource::IndirectDescriptor::SelectorRead::MaskRanks, a, b, c)
 KYTY_IO_FIELDS(DescriptorSource::IndirectDescriptor::SelectorRead, a, b, c, d, e, f, g, h, i, j, k,

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

@@ -251,6 +251,8 @@ enum class StageInputKind {
 	LocalInvocationId,
 	LocalInvocationIndex,
 	GlobalInvocationId,
+	BaseVertex,
+	BaseInstance,
 	Parameter,
 };
 
@@ -541,6 +543,8 @@ struct ShaderInfo {
 	bool                            uses_image_heap               = false;
 	bool                            uses_dispatch_threads         = false;
 	bool                            bvh                           = false;
+	bool                            bvh8                          = false;
+	bool                            bvh_stack                     = false;
 	bool                            subgroup_ballot               = false;
 	bool                            subgroup_vote                 = false;
 	bool                            subgroup_barrier              = false;

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

@@ -31,6 +31,7 @@ std::string TypeName(Type type) {
 	    "SamplerResource",
 	    "ImageAddress",
 	    "F64",
+	    "Bvh8Result",
 	};
 	const auto bits = static_cast<uint32_t>(type);
 	if (bits == 0) {

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

@@ -28,6 +28,7 @@ enum class Type : uint32_t {
 	SamplerResource = 1u << 18u,
 	ImageAddress    = 1u << 19u,
 	F64             = 1u << 20u,
+	Bvh8Result      = 1u << 21u,
 };
 
 constexpr Type operator|(Type lhs, Type rhs) {

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

@@ -25,22 +25,6 @@ Value Value::F32(float value) {
 	return Value(Type::F32, std::bit_cast<uint32_t>(value));
 }
 
-bool Value::IsEmpty() const {
-	return type == Type::Void;
-}
-
-bool Value::IsImmediate() const {
-	return type != Type::Opaque;
-}
-
-bool Value::IsIdentity() const {
-	return type == Type::Opaque && inst->GetOpcode() == ValueOpcode::Identity;
-}
-
-bool Value::IsPhi() const {
-	return type == Type::Opaque && inst->GetOpcode() == ValueOpcode::Phi;
-}
-
 Type Value::GetType() const {
 	if (IsPhi()) {
 		return inst->Flags<Type>();
@@ -56,19 +40,6 @@ Inst* Value::Instruction() const {
 	return inst;
 }
 
-Inst* Value::TryInstruction() const {
-	return type == Type::Opaque ? inst : nullptr;
-}
-
-Inst* Value::ResolveInstruction() const {
-	EXIT_IF(type != Type::Opaque);
-	return IsIdentity() ? inst->Arg(0).ResolveInstruction() : inst;
-}
-
-Value Value::Resolve() const {
-	return IsIdentity() ? inst->Arg(0).Resolve() : *this;
-}
-
 ScalarReg Value::ScalarRegister() const {
 	EXIT_IF(type != Type::ScalarReg);
 	return scalar_reg;
@@ -94,11 +65,6 @@ uint16_t Value::U16() const {
 	return imm_u16;
 }
 
-uint32_t Value::U32() const {
-	EXIT_IF(type != Type::U32);
-	return imm_u32;
-}
-
 uint64_t Value::U64() const {
 	EXIT_IF(type != Type::U64);
 	return imm_u64;
@@ -152,10 +118,6 @@ Inst::~Inst() {
 	ClearArgs();
 }
 
-ValueOpcode Inst::GetOpcode() const {
-	return opcode;
-}
-
 Type Inst::GetType() const {
 	if (opcode == ValueOpcode::Phi) {
 		return static_cast<Type>(flags);
@@ -178,51 +140,25 @@ size_t Inst::UseCount() const {
 	return uses.size();
 }
 
-size_t Inst::NumArgs() const {
-	return num_args == PhiArity ? phi_args.size() : num_args;
-}
-
 size_t Inst::NumPhiBlocks() const {
 	return num_args == PhiArity ? phi_args.size() : 0;
 }
 
-Value Inst::Arg(size_t index) const {
-	EXIT_IF(index >= NumArgs());
-	if (num_args <= InlineArity) {
-		return fixed_args[index];
-	}
-	return num_args == PhiArity ? phi_args[index].second : large_args[index];
-}
-
 Block* Inst::PhiBlock(size_t index) const {
 	EXIT_IF(opcode != ValueOpcode::Phi || index >= phi_args.size());
 	return phi_args[index].first;
 }
 
-Block* Inst::Parent() const {
-	return parent;
-}
-
-const std::vector<Use>& Inst::Uses() const {
-	return uses;
-}
-
 void Inst::SetParent(Block* block) {
 	parent = block;
 }
 
 void Inst::SetArg(size_t index, Value value) {
-	const auto old = Arg(index);
-	if (auto* old_inst = old.TryInstruction(); old_inst != nullptr) {
+	auto& arg = ArgSlot(*this, index);
+	if (auto* old_inst = arg.TryInstruction(); old_inst != nullptr) {
 		RemoveUse(old_inst, index);
 	}
-	if (num_args <= InlineArity) {
-		fixed_args[index] = value;
-	} else if (num_args == PhiArity) {
-		phi_args[index].second = value;
-	} else {
-		large_args[index] = value;
-	}
+	arg = value;
 	if (auto* new_inst = value.TryInstruction(); new_inst != nullptr) {
 		AddUse(new_inst, index);
 	}
@@ -238,11 +174,15 @@ void Inst::AddPhiOperand(Block* predecessor, Value value) {
 }
 
 void Inst::ReplaceUsesWith(Value replacement, bool preserve) {
-	const auto old_uses = uses;
-	for (const auto& use: old_uses) {
-		use.user->SetArg(use.operand, replacement);
-	}
+	// Detach the arguments first (a phi can use itself), then move every use over in order.
 	Invalidate();
+	const auto old_uses = std::exchange(uses, {});
+	for (const auto& use: old_uses) {
+		ArgSlot(*use.user, use.operand) = replacement;
+	}
+	if (auto* target = replacement.TryInstruction(); target != nullptr) {
+		target->uses.insert(target->uses.end(), old_uses.begin(), old_uses.end());
+	}
 	if (preserve) {
 		opcode = ValueOpcode::Identity;
 		num_args = 1;
@@ -272,6 +212,10 @@ void Inst::ClearArgs() {
 			RemoveUse(value_inst, index);
 		}
 	}
+	DropArgs();
+}
+
+void Inst::DropArgs() {
 	if (num_args == PhiArity) {
 		std::destroy_at(&phi_args);
 		std::construct_at(&fixed_args);

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

@@ -33,14 +33,14 @@ public:
 	static Value F16(uint16_t bits);
 	static Value F32(float value);
 
-	[[nodiscard]] bool IsEmpty() const;
-	[[nodiscard]] bool IsImmediate() const;
+	[[nodiscard]] bool IsEmpty() const { return type == Type::Void; }
+	[[nodiscard]] bool IsImmediate() const { return type != Type::Opaque; }
 	[[nodiscard]] bool IsIdentity() const;
 	[[nodiscard]] bool IsPhi() const;
 	[[nodiscard]] Type GetType() const;
 
 	[[nodiscard]] Inst*     Instruction() const;
-	[[nodiscard]] Inst*     TryInstruction() const;
+	[[nodiscard]] Inst*     TryInstruction() const { return type == Type::Opaque ? inst : nullptr; }
 	[[nodiscard]] Inst*     ResolveInstruction() const;
 	[[nodiscard]] Value     Resolve() const;
 	[[nodiscard]] ScalarReg ScalarRegister() const;
@@ -112,7 +112,7 @@ public:
 	Inst(Inst&&)                 = delete;
 	Inst& operator=(Inst&&)      = delete;
 
-	[[nodiscard]] ValueOpcode             GetOpcode() const;
+	[[nodiscard]] ValueOpcode             GetOpcode() const { return opcode; }
 	[[nodiscard]] Type                    GetType() const;
 	[[nodiscard]] bool                    MayHaveSideEffects() const;
 	[[nodiscard]] bool                    HasUses() const;
@@ -121,8 +121,8 @@ public:
 	[[nodiscard]] size_t                  NumPhiBlocks() const;
 	[[nodiscard]] Value                   Arg(size_t index) const;
 	[[nodiscard]] Block*                  PhiBlock(size_t index) const;
-	[[nodiscard]] Block*                  Parent() const;
-	[[nodiscard]] const std::vector<Use>& Uses() const;
+	[[nodiscard]] Block*                  Parent() const { return parent; }
+	[[nodiscard]] const std::vector<Use>& Uses() const { return uses; }
 	// SRT walker output index, assigned by the resource plan that owns this instruction.
 	[[nodiscard]] uint32_t WalkOutput() const { return walk_output; }
 	void SetWalkOutput(uint32_t index) const { walk_output = index; }
@@ -138,6 +138,8 @@ public:
 	void AddPhiOperand(Block* predecessor, Value value);
 	void ReplaceUsesWith(Value replacement, bool preserve = true);
 	void Invalidate();
+	// Drops the arguments without updating their use lists, for teardown that destroys both sides.
+	void DropArgs();
 
 	template <typename T>
 	requires(sizeof(T) <= sizeof(uint64_t) && std::is_trivially_copyable_v<T>)
@@ -157,6 +159,8 @@ public:
 private:
 	friend void EliminateDeadCode(const std::vector<Block*>& blocks);
 
+	template <typename Self>
+	static auto& ArgSlot(Self& self, size_t index);
 	void AddUse(Inst* used, size_t operand);
 	void RemoveUse(Inst* used, size_t operand);
 	void ClearArgs();
@@ -181,4 +185,44 @@ private:
 
 static_assert(sizeof(Inst) <= 120, "Inst operand and backend storage unintentionally increased");
 
+// Called per operand by the IR passes and the resource walker; inline, as the build has no LTO.
+inline bool Value::IsIdentity() const {
+	return type == Type::Opaque && inst->GetOpcode() == ValueOpcode::Identity;
+}
+
+inline bool Value::IsPhi() const {
+	return type == Type::Opaque && inst->GetOpcode() == ValueOpcode::Phi;
+}
+
+inline Inst* Value::ResolveInstruction() const {
+	EXIT_IF(type != Type::Opaque);
+	return IsIdentity() ? inst->Arg(0).ResolveInstruction() : inst;
+}
+
+inline Value Value::Resolve() const {
+	return IsIdentity() ? inst->Arg(0).Resolve() : *this;
+}
+
+inline uint32_t Value::U32() const {
+	EXIT_IF(type != Type::U32);
+	return imm_u32;
+}
+
+inline size_t Inst::NumArgs() const {
+	return num_args == PhiArity ? phi_args.size() : num_args;
+}
+
+template <typename Self>
+auto& Inst::ArgSlot(Self& self, size_t index) {
+	EXIT_IF(index >= self.NumArgs());
+	if (self.num_args <= InlineArity) {
+		return self.fixed_args[index];
+	}
+	return self.num_args == PhiArity ? self.phi_args[index].second : self.large_args[index];
+}
+
+inline Value Inst::Arg(size_t index) const {
+	return ArgSlot(*this, index);
+}
+
 } // namespace Libs::Graphics::ShaderRecompiler::IR

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

@@ -44,6 +44,7 @@ constexpr Type AddressResource = Type::AddressResource;
 constexpr Type ImageResource   = Type::ImageResource;
 constexpr Type SamplerResource = Type::SamplerResource;
 constexpr Type ImageAddress    = Type::ImageAddress;
+constexpr Type Bvh8Result      = Type::Bvh8Result;
 
 constexpr std::array<OpcodeMeta, static_cast<size_t>(ValueOpcode::Count)> MetaTable = {{
 #define VALUE_OPCODE(name, ...) MakeMeta(#name __VA_OPT__(, ) __VA_ARGS__),
@@ -76,7 +77,8 @@ bool HasSideEffects(ValueOpcode opcode) {
 		return true;
 	}
 	const auto shared_access = SharedAccessOf(opcode);
-	if (shared_access == SharedAccess::Write || shared_access == SharedAccess::Atomic ||
+	if (shared_access == SharedAccess::Write || shared_access == SharedAccess::ReadWrite ||
+	    shared_access == SharedAccess::Atomic ||
 	    shared_access == SharedAccess::Append || shared_access == SharedAccess::Consume) {
 		return true;
 	}
@@ -156,6 +158,7 @@ uint32_t BufferComponentCount(ValueOpcode opcode) {
 
 SharedAccess SharedAccessOf(ValueOpcode opcode) {
 	switch (opcode) {
+		case ValueOpcode::BvhStack: return SharedAccess::ReadWrite;
 		case ValueOpcode::LoadSharedU8:
 		case ValueOpcode::LoadSharedU16:
 		case ValueOpcode::LoadSharedU32:

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

@@ -17,7 +17,7 @@ enum class ValueOpcode : uint16_t {
 };
 
 enum class BufferAccess { None, Read, Write, Atomic };
-enum class SharedAccess { None, Read, Write, Atomic, Append, Consume };
+enum class SharedAccess { None, Read, Write, ReadWrite, Atomic, Append, Consume };
 enum class AddressAccess { None, Read, Write };
 enum class ImageAccess { None, Read, Write, Atomic };
 enum class ImageResourceClass { None, Sampled, Storage };

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

@@ -318,6 +318,9 @@ VALUE_OPCODE(SwizzleU32, U32, U32, U32, U1)
 
 VALUE_OPCODE(ImageQueryDimensions, U32x4, ImageResource, ImageAddress)
 VALUE_OPCODE(BvhIntersect, U32x4, U32x4, ImageAddress, U1)
+VALUE_OPCODE(Bvh8Intersect, Bvh8Result, U32x4, ImageAddress, U1)
+VALUE_OPCODE(Bvh8Extract, U32, Bvh8Result, U32)
+VALUE_OPCODE(BvhStack, U32x2, U32, U32, U32x4, U32x4, U1)
 VALUE_OPCODE(ImageQueryLod, U32x4, ImageResource, SamplerResource, ImageAddress)
 VALUE_OPCODE(ImageRead, U32x4, ImageResource, ImageAddress, U1)
 VALUE_OPCODE(ImageWrite, Void, ImageResource, ImageAddress, U32x4, U1)

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

@@ -4182,7 +4182,7 @@ private:
 
 	void Collect(Inst& inst) {
 		const auto op           = inst.GetOpcode();
-		if (op == ValueOpcode::BvhIntersect) {
+		if (op == ValueOpcode::BvhIntersect || op == ValueOpcode::Bvh8Intersect) {
 			m_info.uses_dma = true;
 			return;
 		}

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

@@ -175,6 +175,8 @@ void AddBuiltinInput(ShaderInfo& info, StageInputKind kind) {
 		case StageInputKind::GlobalInvocationId:
 			AddInput(info, kind, 0, 3, "gl_GlobalInvocationID");
 			break;
+		case StageInputKind::BaseVertex: AddInput(info, kind, 0, 1, "gl_BaseVertex"); break;
+		case StageInputKind::BaseInstance: AddInput(info, kind, 0, 1, "gl_BaseInstance"); break;
 		case StageInputKind::PackedAncillary:
 		case StageInputKind::Parameter: break;
 	}
@@ -444,6 +446,8 @@ void Visit(Program& program, ShaderStageInputInfo input_info, InputUsage& inputs
 				case StageInputKind::InstanceIndex:
 				case StageInputKind::FrontFacing:
 				case StageInputKind::LocalInvocationIndex:
+				case StageInputKind::BaseVertex:
+				case StageInputKind::BaseInstance:
 					if (component != 0u) {
 						return Fail("typed scalar builtin component is out of range");
 					}
@@ -511,6 +515,8 @@ void Visit(Program& program, ShaderStageInputInfo input_info, InputUsage& inputs
 		case ValueOpcode::StoreCompletion: info.subgroup_barrier = true; break;
 		case ValueOpcode::DataCacheInvalidate: info.coherent_buffers = true; break;
 		case ValueOpcode::BvhIntersect: info.bvh = true; break;
+		case ValueOpcode::Bvh8Intersect: info.bvh8 = true; break;
+		case ValueOpcode::BvhStack: info.bvh_stack = true; break;
 		case ValueOpcode::ConditionRef:
 			info.subgroup_vote |=
 			    inst.Flags<CFG::BranchCondition>() != CFG::BranchCondition::ScalarInstruction;

src/graphics/shader/shader.cpp

@@ -291,10 +291,9 @@ static bool ShaderGetStaticVertexInputInfo(uint64_t shader_addr, const HW::UserS
 	const auto& metadata = data.vertex;
 
 	if (metadata.vertex_attrib_reg >= 0) {
-		info.fetch_external   = false;
-		info.fetch_embedded   = true;
-		info.fetch_attrib_reg = metadata.vertex_attrib_reg;
-		info.fetch_buffer_reg = metadata.vertex_buffer_reg;
+		info.has_dynamic_vertex_fetch = true;
+		info.fetch_attrib_reg         = metadata.vertex_attrib_reg;
+		info.fetch_buffer_reg         = metadata.vertex_buffer_reg;
 
 		const auto* attrib = reinterpret_cast<const uint32_t*>(
 		    static_cast<uint64_t>(user_sgpr.value[metadata.vertex_attrib_reg]) |
@@ -413,7 +412,7 @@ static void ShaderGetStaticInputInfoCS(const HW::ComputeShaderInfo& regs,
 void BuildStageStaticKey(const ShaderVertexInputInfo& info, std::vector<uint32_t>& key) {
 	EXIT_IF(info.resources_num < 0 || info.resources_num > ShaderVertexInputInfo::RES_MAX);
 	key.clear();
-	key.push_back(static_cast<uint32_t>(info.fetch_embedded));
+	key.push_back(static_cast<uint32_t>(info.has_dynamic_vertex_fetch));
 	key.push_back(static_cast<uint32_t>(info.fetch_attrib_reg));
 	key.push_back(static_cast<uint32_t>(info.fetch_buffer_reg));
 	key.push_back(info.resources_num);
@@ -422,6 +421,8 @@ void BuildStageStaticKey(const ShaderVertexInputInfo& info, std::vector<uint32_t
 	key.push_back(info.lds_size_dwords);
 	key.push_back(info.pa_cl_vs_out_cntl);
 	key.push_back(info.user_clip_planes);
+	key.push_back(info.indirect_offsets.vertex | (uint32_t {info.indirect_offsets.instance} << 8u) |
+	              (uint32_t {info.indirect_draw} << 16u));
 	key.push_back(static_cast<uint32_t>(info.clip_space.enabled));
 	if (info.clip_space.enabled) {
 		for (const float value: info.clip_space.scale) {
@@ -692,9 +693,7 @@ void ShaderDbgDumpInputInfo(const ShaderVertexInputInfo& info) {
 
 	LOGF("ShaderDbgDumpInputInfo()\n");
 
-	LOGF("\t fetch_external = %s\n"
-	     "\t fetch_embedded = %s\n",
-	     info.fetch_external ? "true" : "false", info.fetch_embedded ? "true" : "false");
+	LOGF("\t has_dynamic_vertex_fetch = %s\n", info.has_dynamic_vertex_fetch ? "true" : "false");
 
 	for (int i = 0; i < info.resources_num; i++) {
 		LOGF("\t input %d\n", i);

src/graphics/shader/shader.h

@@ -124,6 +124,17 @@ struct ShaderTessellationInputInfo {
 	uint32_t output_topology       = 0;
 };
 
+// GS user-data slots an indirect draw packet writes its record's base vertex and first instance
+// to. A packet can leave both slots unused even when its record contains nonzero offsets.
+struct ShaderIndirectOffsets {
+	static constexpr uint8_t None = UINT8_MAX;
+
+	uint8_t vertex   = None;
+	uint8_t instance = None;
+
+	bool operator==(const ShaderIndirectOffsets&) const = default;
+};
+
 struct ShaderVertexInputInfo {
 	static constexpr int RES_MAX = 32;
 
@@ -146,8 +157,11 @@ struct ShaderVertexInputInfo {
 	ShaderClipSpaceTransform clip_space;
 	ShaderMeshInputInfo      mesh;
 	ShaderTessellationInputInfo tess;
-	bool                    fetch_external      = false;
-	bool                    fetch_embedded      = false;
+	ShaderIndirectOffsets   indirect_offsets;
+	// GPU-read records add Vulkan bases even when the packet patches no user SGPRs.
+	bool                    indirect_draw       = false;
+	// Attribute-table metadata enables dynamic fetch detection; loads are matched later.
+	bool                    has_dynamic_vertex_fetch = false;
 };
 
 struct ShaderComputeInputInfo: ShaderWorkgroupInputInfo {

src/kernel/memory.h

@@ -26,9 +26,10 @@ struct DebugStats {
 	uint64_t flexible_total      = 0;
 	uint64_t flexible_used       = 0;
 	uint64_t pool_committed      = 0;
-	uint64_t cpu_page_entries    = 0; // 2 MiB page-table entries in use.
-	uint64_t gpu_page_entries    = 0;
-	uint64_t page_entries_total  = 0; // Per side.
+	uint64_t cpu_page_entries       = 0; // 2 MiB page-table entries in use.
+	uint64_t gpu_page_entries       = 0;
+	uint64_t cpu_page_entries_total = 0;
+	uint64_t gpu_page_entries_total = 0;
 };
 [[nodiscard]] DebugStats GetDebugStats();
 
@@ -125,7 +126,6 @@ static_assert(sizeof(KernelMemoryPoolBlockStats) == 16,
               "KernelMemoryPoolBlockStats struct size is incorrect");
 
 void                   RegisterCallbacks(callback_func_t alloc_func, callback_func_t free_func);
-void                   SetFlexibleMemorySize(uint64_t size);
 int AllocateDirectMemory(int64_t search_start, int64_t search_end, size_t size, size_t alignment,
                          int memory_type, int64_t* phys_addr_out, bool automatic = false);
 int MapAutomaticMemory(uint64_t vaddr, size_t size, int type, int prot);
@@ -225,6 +225,11 @@ bool     TestPlaceholderRangeIsFree(uint64_t vaddr, uint64_t size);
 bool     TestGuestAddressRangeIsOwned(uint64_t vaddr, uint64_t size);
 bool     TestGuestBackingOutsideAddressSpace();
 uint64_t TestGuestBackingSize();
+uint64_t TestFlexibleBackingHighWater();
+uint64_t TestGuestFreeRangeVisits();
+uint64_t TestDirectMappingCount();
+uint64_t TestFlexibleMappingCount();
+uint64_t TestBackingBytesZeroed();
 bool     TestGuestFreeRangeBounds();
 #endif
 

src/kernel/memoryAddressSpace.inc

@@ -203,6 +203,9 @@ public:
 		}
 		std::lock_guard lock(m_mutex);
 		std::memset(m_backing_base + backing_offset, 0, size);
+#if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
+		g_test_backing_bytes_zeroed += size;
+#endif
 		return true;
 	}
 
@@ -287,6 +290,18 @@ public:
 		return true;
 	}
 
+	bool ProtectViews(uint64_t vaddr, uint64_t size, VirtualMemory::Mode mode) {
+		std::shared_lock lock(m_mutex);
+		bool             ok = true;
+		ForEachSparseRange(
+		    m_maps, vaddr, size,
+		    [&](const auto&, uint64_t start, uint64_t bytes) {
+			    ok = ok && VirtualMemory::Protect(start, bytes, mode);
+		    },
+		    [&](uint64_t, uint64_t) { ok = false; });
+		return ok;
+	}
+
 	bool MapFixed(uint64_t vaddr, uint64_t size, uint64_t backing_offset,
 	              VirtualMemory::Mode mode, FailureReason* failure_reason = nullptr) {
 		if (!IsAvailable() || !RangeOk(backing_offset, size)) {
@@ -324,15 +339,11 @@ public:
 #endif
 	}
 
-	bool Unmap(uint64_t vaddr, uint64_t size, bool* placeholder_preserved = nullptr) {
+	bool Unmap(uint64_t vaddr, uint64_t size) {
 		if (size == 0 || UINT64_MAX - vaddr < size) {
 			return false;
 		}
 
-		if (placeholder_preserved != nullptr) {
-			*placeholder_preserved = false;
-		}
-
 		const auto            end = vaddr + size;
 		std::vector<MapEntry> targets;
 		{
@@ -350,42 +361,36 @@ public:
 			}
 		}
 
-		std::vector<MapEntry> removed;
-		removed.reserve(targets.size());
+		size_t removed  = 0;
+		auto   rollback = [&] {
+			while (removed != 0) {
+				const auto& previous = targets[--removed];
+				EXIT_IF(!MapFixed(previous.vaddr, previous.size, previous.backing_offset,
+				                  previous.mode));
+			}
+		};
 		for (const auto& entry: targets) {
-			bool part_placeholder_preserved = false;
-			if (!UnmapOne(entry.vaddr, entry.size, &part_placeholder_preserved)) {
-				for (auto it = removed.rbegin(); it != removed.rend(); ++it) {
-					EXIT_IF(!MapFixed(it->vaddr, it->size, it->backing_offset, it->mode));
-				}
-				if (placeholder_preserved != nullptr) {
-					*placeholder_preserved = false;
-				}
+			if (!UnmapOne(entry.vaddr, entry.size)) {
+				rollback();
 				return false;
 			}
-			EXIT_IF(!part_placeholder_preserved);
-			removed.push_back(entry);
+			++removed;
 		}
 
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 		// Each unmapped file view becomes a separate Windows placeholder. The address-space owner
 		// records this contiguous guest span as one free range, so normalize fragmented backing
 		// views into one placeholder before publishing it there.
-		if (removed.size() > 1 && !CoalescePlaceholders(vaddr, size)) {
-			for (auto it = removed.rbegin(); it != removed.rend(); ++it) {
-				EXIT_IF(!MapFixed(it->vaddr, it->size, it->backing_offset, it->mode));
-			}
+		if (removed > 1 && !CoalescePlaceholders(vaddr, size)) {
+			rollback();
 			return false;
 		}
 #endif
 
-		if (placeholder_preserved != nullptr) {
-			*placeholder_preserved = true;
-		}
 		return true;
 	}
 
-	bool UnmapOne(uint64_t vaddr, uint64_t size, bool* placeholder_preserved = nullptr) {
+	bool UnmapOne(uint64_t vaddr, uint64_t size) {
 #if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
 		if (g_test_backing_store_unmaps_before_failure == 0) {
 			g_test_backing_store_unmaps_before_failure = UINT32_MAX;
@@ -428,9 +433,7 @@ public:
 				m_maps.emplace(old.vaddr, old);
 				return false;
 			}
-			if (placeholder_preserved != nullptr) {
-				*placeholder_preserved = true;
-			}
+
 			return true;
 		}
 
@@ -451,9 +454,6 @@ public:
 				     ok;
 			}
 			if (ok) {
-				if (placeholder_preserved != nullptr) {
-					*placeholder_preserved = true;
-				}
 				return true;
 			}
 
@@ -481,9 +481,6 @@ public:
 		std::lock_guard lock(m_mutex);
 		RemoveMappingsUnlocked(vaddr, size);
 
-		if (placeholder_preserved != nullptr) {
-			*placeholder_preserved = true;
-		}
 		return true;
 #endif
 	}
@@ -634,8 +631,6 @@ private:
 
 	HANDLE m_handle = nullptr;
 #elif KYTY_PLATFORM == KYTY_PLATFORM_LINUX
-	[[nodiscard]] static uint64_t GetMappableTestSize() { return 0x4000; }
-
 	static int GetProtect(VirtualMemory::Mode mode) {
 		switch (mode) {
 			case VirtualMemory::Mode::Read: return PROT_READ;
@@ -760,7 +755,7 @@ public:
 		if (size == 0 || UINT64_MAX - vaddr < size) {
 			return false;
 		}
-		std::lock_guard lock(m_mutex);
+		std::shared_lock lock(m_mutex);
 		return std::any_of(m_owned.begin(), m_owned.end(), [vaddr, size](const auto& range) {
 			return VirtualRangesOverlap(vaddr, size, range.first, range.second);
 		});
@@ -864,11 +859,7 @@ public:
 		if (!ContainsMappedUnlocked(vaddr, size, MappingType::Backing)) {
 			return false;
 		}
-		bool placeholder_preserved = false;
-		if (!m_backing->Unmap(vaddr, size, &placeholder_preserved)) {
-			return false;
-		}
-		if (!placeholder_preserved) {
+		if (!m_backing->Unmap(vaddr, size)) {
 			return false;
 		}
 		EXIT_IF(!RemoveMappedUnlocked(vaddr, size, MappingType::Backing));
@@ -876,12 +867,13 @@ public:
 		return true;
 	}
 
+	// Reservations live in VirtualRanges; these free-range checks only read the host placeholders.
 	uint64_t ReserveAligned(uint64_t search_addr, uint64_t size, uint64_t alignment) {
 		if (size == 0) {
 			return 0;
 		}
 		alignment = std::max<uint64_t>(alignment != 0 ? alignment : PageSize(), PageSize());
-		std::lock_guard lock(m_mutex);
+		std::shared_lock lock(m_mutex);
 		if (search_addr == 0) {
 			auto addr = FindFreeAlignedUnlocked(HOST_SYSTEM_MANAGED_MIN,
 			                                    HOST_SYSTEM_MANAGED_MAX + 1u, size, alignment);
@@ -898,7 +890,7 @@ public:
 			return 0;
 		}
 		alignment = std::max<uint64_t>(alignment, PageSize());
-		std::lock_guard lock(m_mutex);
+		std::shared_lock lock(m_mutex);
 		return FindFreeAlignedUnlocked(search_start, search_end, size, alignment);
 	}
 
@@ -906,19 +898,19 @@ public:
 		if (vaddr == 0 || size == 0 || !IsPageAligned(vaddr) || !IsPageAligned(size)) {
 			return false;
 		}
-		std::lock_guard lock(m_mutex);
+		std::shared_lock lock(m_mutex);
 		return FindContaining(vaddr, size) != m_free.end();
 	}
 
 #if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
 	bool TestContainsFree(uint64_t vaddr, uint64_t size) {
-		std::lock_guard lock(m_mutex);
+		std::shared_lock lock(m_mutex);
 		return FindContaining(vaddr, size) != m_free.end();
 	}
 #endif
 
 	bool ReleaseFree(uint64_t vaddr, uint64_t size) {
-		std::lock_guard lock(m_mutex);
+		std::shared_lock lock(m_mutex);
 		return FindContaining(vaddr, size) != m_free.end();
 	}
 
@@ -1041,10 +1033,6 @@ private:
 		return (value & (PageSize() - 1u)) == 0;
 	}
 
-	[[nodiscard]] bool IsGranularityAligned(uint64_t value) const {
-		return (value & (m_granularity - 1u)) == 0;
-	}
-
 	bool ProtectMappedUnlocked(uint64_t vaddr, uint64_t size, VirtualMemory::Mode mode) {
 		const auto end = vaddr + size;
 		auto       it  = m_mapped.upper_bound(vaddr);
@@ -1055,8 +1043,20 @@ private:
 			const auto& region      = it->second;
 			const auto  range_start = std::max(vaddr, region.vaddr);
 			const auto  range_end   = std::min(end, region.vaddr + region.size);
-			if (range_start < range_end &&
-			    !VirtualMemory::Protect(range_start, range_end - range_start, mode)) {
+			if (range_start >= range_end) {
+				continue;
+			}
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
+			// Rollback can split a backing view without changing the owner interval.
+			// Windows protection calls must stay within each actual view.
+			if (region.type == MappingType::Backing) {
+				if (!m_backing->ProtectViews(range_start, range_end - range_start, mode)) {
+					return false;
+				}
+				continue;
+			}
+#endif
+			if (!VirtualMemory::Protect(range_start, range_end - range_start, mode)) {
 				return false;
 			}
 		}
@@ -1312,12 +1312,21 @@ private:
 	}
 
 	uint64_t FindFreeAlignedUnlocked(uint64_t search_start, uint64_t search_end, uint64_t size,
-	                                 uint64_t alignment) {
+	                                 uint64_t alignment) const {
 		if (search_start >= search_end || size > search_end - search_start) {
 			return 0;
 		}
-		for (const auto& [base, range_size]: m_free) {
-			const auto candidate = AlignUp(std::max(base, search_start), alignment);
+		// Include the predecessor because search_start can lie inside a free range.
+		auto it = m_free.upper_bound(search_start);
+		if (it != m_free.begin()) {
+			--it;
+		}
+		for (; it != m_free.end() && it->first < search_end; ++it) {
+#if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
+			g_test_free_range_visits.fetch_add(1, std::memory_order_relaxed);
+#endif
+			const auto& [base, range_size] = *it;
+			const auto candidate           = AlignUp(std::max(base, search_start), alignment);
 			if (candidate < search_end && size <= search_end - candidate &&
 			    GuestFreeRangeContains(base, range_size, candidate, size)) {
 				return candidate;
@@ -1331,9 +1340,8 @@ private:
 		if (next == m_free.begin()) {
 			return m_free.end();
 		}
-		auto it  = std::prev(next);
-		auto end = it->first + it->second;
-		return (vaddr >= it->first && vaddr + size <= end ? it : m_free.end());
+		auto it = std::prev(next);
+		return GuestFreeRangeContains(it->first, it->second, vaddr, size) ? it : m_free.end();
 	}
 
 	bool SplitUnlocked(std::map<uint64_t, uint64_t>::iterator it, uint64_t vaddr, uint64_t size) {

src/kernel/pthread.cpp

@@ -2277,7 +2277,10 @@ int KYTY_SYSV_ABI PthreadRwlockDestroy(PthreadRwlock* rwlock) {
 		return KERNEL_ERROR_EINVAL;
 	}
 
-	EXIT_NOT_IMPLEMENTED(*rwlock == nullptr);
+	// SCE_PTHREAD_RWLOCK_INITIALIZER has no allocated resources until first use.
+	if (*rwlock == nullptr) {
+		return OK;
+	}
 
 	{
 		std::lock_guard lock((*rwlock)->m);

src/launcher/CMakeLists.txt

@@ -125,6 +125,7 @@ if(APPLE)
 	install(TARGETS kyty_emulator DESTINATION "KytyPS5.app/Contents/MacOS")
 	install(TARGETS kyty_updater DESTINATION "KytyPS5.app/Contents/MacOS")
 	install(FILES "${KYTY_TROPHY_UNLOCK_SOUND}" DESTINATION "KytyPS5.app/Contents/Resources/assets/sounds")
+	install(FILES ${KYTY_NOTO_SANS_FONTS} DESTINATION "KytyPS5.app/Contents/Resources/assets/fonts")
 	install(DIRECTORY "${CMAKE_SOURCE_DIR}/assets/cursors" DESTINATION "KytyPS5.app/Contents/Resources/assets")
 	# Deploy and ad-hoc sign the Qt frameworks and plugins in the bundle.
 	find_program(KYTY_MACDEPLOYQT NAMES macdeployqt PATHS "${Qt6_DIR}/../../../bin" NO_DEFAULT_PATH)

src/launcher/forms/configuration_edit_dialog.ui

@@ -115,7 +115,17 @@
             </item>
             <item row="0" column="1">
              <widget class="QComboBox" name="comboBox_screen_resolution">
-              <property name="toolTip"><string>Window resolution</string></property>
+              <property name="toolTip"><string>Window resolution as WIDTHxHEIGHT; pick a preset or type a custom size. Non-16:9 sizes such as 21:9 show the game with black bars unless Stretch to window is on.</string></property>
+              <property name="editable"><bool>true</bool></property>
+              <property name="insertPolicy"><enum>QComboBox::NoInsert</enum></property>
+              <item><property name="text"><string notr="true">1280x720</string></property></item>
+              <item><property name="text"><string notr="true">1920x1080</string></property></item>
+              <item><property name="text"><string notr="true">2560x1440</string></property></item>
+              <item><property name="text"><string notr="true">3840x2160</string></property></item>
+              <item><property name="text"><string notr="true">2560x1080</string></property></item>
+              <item><property name="text"><string notr="true">3440x1440</string></property></item>
+              <item><property name="text"><string notr="true">3840x1600</string></property></item>
+              <item><property name="text"><string notr="true">5120x2160</string></property></item>
              </widget>
             </item>
             <item row="1" column="0">
@@ -244,6 +254,12 @@
           <widget class="QGroupBox" name="general_group">
            <property name="title"><string>General</string></property>
            <layout class="QVBoxLayout" name="generalLayout">
+            <item>
+             <widget class="QCheckBox" name="checkBox_pro">
+              <property name="toolTip"><string>Enable experimental Pro mode for supported titles. PSSR uses linear scaling. Disabled by default.</string></property>
+              <property name="text"><string>Pro mode (experimental)</string></property>
+             </widget>
+            </item>
             <item>
              <widget class="QCheckBox" name="checkBox_fullscreen">
               <property name="toolTip"><string>Run in borderless desktop fullscreen</string></property>
@@ -256,6 +272,12 @@
               <property name="text"><string>Auto-hide cursor</string></property>
              </widget>
             </item>
+            <item>
+             <widget class="QCheckBox" name="checkBox_stretch">
+              <property name="toolTip"><string>Stretch the game image to fill the window instead of keeping its aspect ratio with black bars</string></property>
+              <property name="text"><string>Stretch to window</string></property>
+             </widget>
+            </item>
             <item>
              <widget class="QCheckBox" name="checkBox_perf_overlay">
               <property name="toolTip"><string>Show FPS, frame times, CPU/GPU usage, RAM and memory pressure in the game window. Press F2 (Cmd+P on macOS) to toggle it, and add Shift for the detailed view.</string></property>
@@ -390,6 +412,11 @@
          <property name="text"><string>Enable network connection</string></property>
         </widget>
        </item>
+       <item>
+        <widget class="QCheckBox" name="checkBox_psn_emulation">
+         <property name="text"><string>Emulate PSN (compatibility)</string></property>
+        </widget>
+       </item>
        <item>
         <spacer name="networkSpacer">
          <property name="orientation"><enum>Qt::Vertical</enum></property>

src/launcher/include/configuration.h

@@ -21,10 +21,7 @@ inline QStringList EnumToList() {
 	auto        me    = QMetaEnum::fromType<T>();
 	int         count = me.keyCount();
 	for (int i = 0; i < count; i++) {
-		auto key = QString(me.key(i));
-		ret << (key.startsWith('R') && key.size() > 2 && key.at(1).isDigit()
-		            ? key.remove('R').toLower()
-		            : key);
+		ret << QString(me.key(i));
 	}
 	return ret;
 }
@@ -32,16 +29,12 @@ inline QStringList EnumToList() {
 template <class T>
 T TextToEnum(const QString& text) {
 	auto me = QMetaEnum::fromType<T>();
-	return static_cast<T>(me.keyToValue(
-	    ((text.size() > 1 && text.at(0).isDigit()) ? 'R' + text.toUpper() : text).toUtf8().data()));
+	return static_cast<T>(me.keyToValue(text.toUtf8().constData()));
 }
 
 template <class T>
 QString EnumToText(T value) {
-	auto me  = QMetaEnum::fromType<T>();
-	auto key = QString(me.valueToKey(static_cast<int>(value)));
-	return (key.startsWith('R') && key.size() > 2 && key.at(1).isDigit() ? key.remove('R').toLower()
-	                                                                     : key);
+	return QString(QMetaEnum::fromType<T>().valueToKey(static_cast<int>(value)));
 }
 
 struct ControllerSettings {
@@ -80,13 +73,10 @@ public:
 	static constexpr int DEFAULT_CONSOLE_LANGUAGE = 1;
 	static constexpr int MAX_CONSOLE_LANGUAGE     = 29;
 
-	enum class Resolution {
-		R1280X720,
-		R1920X1080,
-		R2560X1440,
-		R3840X2160,
-	};
-	Q_ENUM(Resolution)
+	static constexpr int MIN_SCREEN_SIZE = 320;
+	static constexpr int MAX_SCREEN_SIZE = 16384;
+
+	static QString NormalizeResolution(const QString& text);
 
 	enum class ShaderOptimizationType { None, Size, Performance };
 	Q_ENUM(ShaderOptimizationType)
@@ -117,7 +107,7 @@ public:
 	// Controller preferences always come from the global configuration.
 	ControllerSettings controller;
 
-	Resolution             screen_resolution           = Resolution::R1280X720;
+	QString                screen_resolution           = QStringLiteral("1280x720");
 	QString                user_name                   = "Kyty";
 	int                    user_id                     = Config::DEFAULT_USER_ID;
 	QString                audio_input_device;
@@ -126,6 +116,7 @@ public:
 	bool                   fullscreen_enabled          = false;
 	bool                   hdr_enabled                 = false;
 	bool                   hide_cursor_enabled         = false;
+	bool                   stretch_enabled             = false;
 	bool                   readback_linear_images      = false;
 	bool                   readback_tiled_images       = false;
 	bool                   sync_raw_image_buffers      = false;
@@ -137,6 +128,7 @@ public:
 	bool                   perf_overlay_enabled        = false;
 	bool                   skip_notice_screen          = false;
 	bool                   network_connected           = false;
+	bool                   psn_emulation_enabled       = false;
 	int                    vblank_frequency            = 60;
 	int                    console_language            = DEFAULT_CONSOLE_LANGUAGE;
 	bool                   vulkan_validation_enabled   = false;
@@ -150,6 +142,7 @@ public:
 	QString                printf_output_file          = "_kyty.txt";
 	bool                   profiler_enabled            = false;
 	bool                   renderdoc_enabled           = false;
+	bool                   pro_enabled                 = false;
 	bool                   amd_cpu_enabled             = false;
 #if defined(_WIN32)
 	bool red_zone_protection_enabled = false;
@@ -168,6 +161,7 @@ public:
 		fullscreen_enabled          = other.fullscreen_enabled;
 		hdr_enabled                 = other.hdr_enabled;
 		hide_cursor_enabled         = other.hide_cursor_enabled;
+		stretch_enabled             = other.stretch_enabled;
 		readback_linear_images      = other.readback_linear_images;
 		readback_tiled_images       = other.readback_tiled_images;
 		sync_raw_image_buffers      = other.sync_raw_image_buffers;
@@ -179,6 +173,7 @@ public:
 		perf_overlay_enabled        = other.perf_overlay_enabled;
 		skip_notice_screen          = other.skip_notice_screen;
 		network_connected           = other.network_connected;
+		psn_emulation_enabled       = other.psn_emulation_enabled;
 		vblank_frequency            = other.vblank_frequency;
 		console_language            = other.console_language;
 		vulkan_validation_enabled   = other.vulkan_validation_enabled;
@@ -192,6 +187,7 @@ public:
 		printf_output_file          = other.printf_output_file;
 		profiler_enabled            = other.profiler_enabled;
 		renderdoc_enabled           = other.renderdoc_enabled;
+		pro_enabled                 = other.pro_enabled;
 		amd_cpu_enabled             = other.amd_cpu_enabled;
 #if defined(_WIN32)
 		red_zone_protection_enabled = other.red_zone_protection_enabled;

src/launcher/include/configurationEditDialog.h

@@ -46,11 +46,15 @@ private:
 	QLineEdit*                   m_addons_dir_edit        = nullptr;
 	QToolButton*                 m_remove_game_dir_button = nullptr;
 	bool                         m_global_settings        = false;
+	bool                         m_geometry_restored      = false;
 
 protected:
 	void Init(const Configuration& info);
 	void InitGameDirectories();
 	void AddGameDirectoryItem(const QString& dir);
+	void MakeTabsScrollable();
+	void GrowToFitContents();
+	void ShowTab(QWidget* page);
 
 	void moveEvent(QMoveEvent* event) override;
 	void resizeEvent(QResizeEvent* event) override;

src/launcher/src/configuration.cpp

@@ -1,5 +1,23 @@
 #include "configuration.h"
 
+#include <QRegularExpression>
+
+QString Configuration::NormalizeResolution(const QString& text) {
+	// Older settings stored the enum key, e.g. "R1920X1080".
+	static const QRegularExpression re(QStringLiteral("^\\s*R?(\\d+)\\s*[xX]\\s*(\\d+)\\s*$"));
+	const auto match = re.match(text);
+	if (!match.hasMatch()) {
+		return {};
+	}
+	const int width  = match.captured(1).toInt();
+	const int height = match.captured(2).toInt();
+	if (width < MIN_SCREEN_SIZE || height < MIN_SCREEN_SIZE || width > MAX_SCREEN_SIZE ||
+	    height > MAX_SCREEN_SIZE) {
+		return {};
+	}
+	return QStringLiteral("%1x%2").arg(width).arg(height);
+}
+
 QVariantMap Configuration::GameSettings() const {
 	QVariantMap values;
 #define KYTY_CFG_SET(n) values.insert(QStringLiteral(#n), QVariant::fromValue(n).toString())
@@ -12,6 +30,7 @@ QVariantMap Configuration::GameSettings() const {
 	KYTY_CFG_SET(fullscreen_enabled);
 	KYTY_CFG_SET(hdr_enabled);
 	KYTY_CFG_SET(hide_cursor_enabled);
+	KYTY_CFG_SET(stretch_enabled);
 	KYTY_CFG_SET(readback_linear_images);
 	KYTY_CFG_SET(readback_tiled_images);
 	KYTY_CFG_SET(sync_raw_image_buffers);
@@ -23,6 +42,7 @@ QVariantMap Configuration::GameSettings() const {
 	KYTY_CFG_SET(perf_overlay_enabled);
 	KYTY_CFG_SET(skip_notice_screen);
 	KYTY_CFG_SET(network_connected);
+	KYTY_CFG_SET(psn_emulation_enabled);
 	KYTY_CFG_SET(vblank_frequency);
 	KYTY_CFG_SET(console_language);
 	KYTY_CFG_SET(vulkan_validation_enabled);
@@ -36,6 +56,7 @@ QVariantMap Configuration::GameSettings() const {
 	KYTY_CFG_SET(printf_output_file);
 	KYTY_CFG_SET(profiler_enabled);
 	KYTY_CFG_SET(renderdoc_enabled);
+	KYTY_CFG_SET(pro_enabled);
 	KYTY_CFG_SET(amd_cpu_enabled);
 #if defined(_WIN32)
 	KYTY_CFG_SET(red_zone_protection_enabled);
@@ -48,7 +69,8 @@ QVariantMap Configuration::GameSettings() const {
 template <class Settings>
 void Configuration::ReadGameSettingsValues(const Settings& s) {
 #define KYTY_CFG_GET(n) n = s.value(QStringLiteral(#n)).template value<decltype(n)>()
-	KYTY_CFG_GET(screen_resolution);
+	const auto resolution = NormalizeResolution(s.value("screen_resolution").toString());
+	screen_resolution     = resolution.isEmpty() ? QStringLiteral("1280x720") : resolution;
 	user_name          = s.value("user_name", user_name).toString();
 	bool user_id_ok    = false;
 	auto saved_user_id = s.value("user_id", user_id).toInt(&user_id_ok);
@@ -64,6 +86,7 @@ void Configuration::ReadGameSettingsValues(const Settings& s) {
 	KYTY_CFG_GET(fullscreen_enabled);
 	KYTY_CFG_GET(hdr_enabled);
 	KYTY_CFG_GET(hide_cursor_enabled);
+	KYTY_CFG_GET(stretch_enabled);
 	KYTY_CFG_GET(readback_linear_images);
 	KYTY_CFG_GET(readback_tiled_images);
 	KYTY_CFG_GET(sync_raw_image_buffers);
@@ -75,6 +98,7 @@ void Configuration::ReadGameSettingsValues(const Settings& s) {
 	KYTY_CFG_GET(perf_overlay_enabled);
 	KYTY_CFG_GET(skip_notice_screen);
 	KYTY_CFG_GET(network_connected);
+	KYTY_CFG_GET(psn_emulation_enabled);
 	vblank_frequency = s.value("vblank_frequency", vblank_frequency).toInt();
 	console_language = s.value("console_language", console_language).toInt();
 	if (console_language < 0 || console_language > MAX_CONSOLE_LANGUAGE) {
@@ -91,6 +115,7 @@ void Configuration::ReadGameSettingsValues(const Settings& s) {
 	KYTY_CFG_GET(printf_output_file);
 	KYTY_CFG_GET(profiler_enabled);
 	KYTY_CFG_GET(renderdoc_enabled);
+	pro_enabled     = s.value("pro_enabled", false).toBool();
 	amd_cpu_enabled = s.value("amd_cpu_enabled", false).toBool();
 #if defined(_WIN32)
 	red_zone_protection_enabled =
@@ -134,7 +159,9 @@ bool Configuration::SetGameSettings(const QJsonObject& settings, QString& error)
 		if (!it.value().isString()) {
 			return fail(tr("Game setting %1 must be a string.").arg(it.key()));
 		}
-		values.insert(it.key(), it.value().toString());
+		const auto value = it.value().toString();
+		values.insert(it.key(), it.key() == "screen_resolution" ? NormalizeResolution(value)
+		                                                       : value);
 	}
 
 	ReadGameSettingsValues(values);

src/launcher/src/configurationEditDialog.cpp

@@ -24,6 +24,8 @@
 #include <QMessageBox>
 #include <QPushButton>
 #include <QResizeEvent>
+#include <QScreen>
+#include <QScrollArea>
 #include <QSettings>
 #include <QSlider>
 #include <QSpinBox>
@@ -122,10 +124,10 @@ static QString GameDirectoryKey(const QString& dir) {
 }
 
 ConfigurationEditDialog::ConfigurationEditDialog(Configuration& info, QWidget* parent)
-    : QDialog(parent, Qt::WindowCloseButtonHint), m_ui(new Ui::ConfigurationEditDialog),
-      m_info(info) {
+    : QDialog(parent, Qt::WindowMaximizeButtonHint | Qt::WindowCloseButtonHint),
+      m_ui(new Ui::ConfigurationEditDialog), m_info(info) {
 	m_ui->setupUi(this);
-	setMinimumWidth(width());
+	MakeTabsScrollable();
 	InitGameDirectories();
 	m_ui->controller_group->setVisible(false);
 	m_ui->settings_button->setVisible(false);
@@ -170,9 +172,11 @@ ConfigurationEditDialog::ConfigurationEditDialog(Configuration& info, QWidget* p
 
 	setSizeGripEnabled(true);
 
-	restoreGeometry(g_last_geometry);
-
 	Init(info);
+	m_geometry_restored = restoreGeometry(g_last_geometry);
+	if (!m_geometry_restored) {
+		GrowToFitContents();
+	}
 }
 
 QByteArray ConfigurationEditDialog::g_last_geometry;
@@ -248,7 +252,7 @@ void ConfigurationEditDialog::Init(const Configuration& info) {
 		                    info.audio_input_device);
 	}
 	microphone->setCurrentIndex(microphone->findData(info.audio_input_device));
-	ListInit(m_ui->comboBox_screen_resolution, info.screen_resolution);
+	m_ui->comboBox_screen_resolution->setCurrentText(info.screen_resolution);
 	ListInit(m_ui->comboBox_present_mode, info.present_mode);
 	m_ui->comboBox_gpu->clear();
 	m_ui->comboBox_gpu->addItem(tr("Auto"));
@@ -285,6 +289,7 @@ void ConfigurationEditDialog::Init(const Configuration& info) {
 	m_ui->checkBox_fullscreen->setChecked(info.fullscreen_enabled);
 	m_ui->checkBox_hdr->setChecked(info.hdr_enabled);
 	m_ui->checkBox_hide_cursor->setChecked(info.hide_cursor_enabled);
+	m_ui->checkBox_stretch->setChecked(info.stretch_enabled);
 	m_ui->checkBox_readback->setChecked(info.readback_linear_images);
 	m_ui->checkBox_readback_tiled->setChecked(info.readback_tiled_images);
 	m_ui->checkBox_sync_raw_image_buffers->setChecked(info.sync_raw_image_buffers);
@@ -293,6 +298,7 @@ void ConfigurationEditDialog::Init(const Configuration& info) {
 	m_ui->checkBox_shader_cache->setChecked(info.shader_cache_enabled);
 	m_ui->checkBox_skip_notice_screen->setChecked(info.skip_notice_screen);
 	m_ui->checkBox_network_connected->setChecked(info.network_connected);
+	m_ui->checkBox_psn_emulation->setChecked(info.psn_emulation_enabled);
 	m_ui->checkBox_trophy_notifications->setChecked(info.trophy_enabled);
 	m_ui->spinBox_trophy_duration->setValue(info.trophy_duration_seconds);
 	m_ui->checkBox_perf_overlay->setChecked(info.perf_overlay_enabled);
@@ -306,6 +312,7 @@ void ConfigurationEditDialog::Init(const Configuration& info) {
 	m_ui->checkBox_shader_validation->setChecked(info.shader_validation_enabled);
 	m_ui->checkBox_vulkan_validation->setChecked(info.vulkan_validation_enabled);
 	m_ui->checkBox_renderdoc_capture->setChecked(info.renderdoc_enabled);
+	m_ui->checkBox_pro->setChecked(info.pro_enabled);
 	m_ui->checkBox_amd_cpu->setChecked(info.amd_cpu_enabled);
 #if defined(__APPLE__)
 	m_ui->checkBox_amd_cpu->setVisible(false);
@@ -413,8 +420,46 @@ void ConfigurationEditDialog::SetGlobalSettings(const QStringList& dirs, const Q
 	m_game_dirs_group->setVisible(true);
 	m_ui->controller_group->setVisible(true);
 	update_game_directory_buttons();
+	if (!m_geometry_restored) {
+		GrowToFitContents();
+	}
+}
+
+void ConfigurationEditDialog::MakeTabsScrollable() {
+	auto* tabs = m_ui->settings_tabs;
+	for (int i = 0; i < tabs->count(); i++) {
+		auto* page   = tabs->widget(i);
+		auto  title  = tabs->tabText(i);
+		auto* scroll = new QScrollArea;
+		scroll->setWidgetResizable(true);
+		scroll->setFrameShape(QFrame::NoFrame);
+		tabs->removeTab(i);
+		scroll->setWidget(page);
+		tabs->insertTab(i, scroll, title);
+	}
+	tabs->setCurrentIndex(0);
+}
+
+void ConfigurationEditDialog::GrowToFitContents() {
 	layout()->activate();
-	resize(size().expandedTo(minimumSizeHint()));
+	auto* tabs = m_ui->settings_tabs;
+	QSize content;
+	for (int i = 0; i < tabs->count(); i++) {
+		content = content.expandedTo(
+		    static_cast<QScrollArea*>(tabs->widget(i))->widget()->minimumSizeHint());
+	}
+	const QSize chrome = size() - tabs->currentWidget()->size();
+	resize(size().expandedTo(content + chrome).boundedTo(screen()->availableGeometry().size()));
+}
+
+void ConfigurationEditDialog::ShowTab(QWidget* page) {
+	auto* tabs = m_ui->settings_tabs;
+	for (int i = 0; i < tabs->count(); i++) {
+		if (tabs->widget(i)->isAncestorOf(page)) {
+			tabs->setCurrentIndex(i);
+			return;
+		}
+	}
 }
 
 QString ConfigurationEditDialog::GetAddonsDirectory() const {
@@ -474,13 +519,14 @@ static void UpdateInfo(Configuration& info, Ui::ConfigurationEditDialog& ui, boo
 		info.controller.keyboard_player     = ui.comboBox_keyboard_player->currentData().toInt();
 	}
 	info.screen_resolution =
-	    TextToEnum<Configuration::Resolution>(ui.comboBox_screen_resolution->currentText());
+	    Configuration::NormalizeResolution(ui.comboBox_screen_resolution->currentText());
 	info.present_mode =
 	    TextToEnum<Configuration::PresentMode>(ui.comboBox_present_mode->currentText());
 	info.gpu_index                 = ui.comboBox_gpu->currentIndex() - 1;
 	info.fullscreen_enabled        = ui.checkBox_fullscreen->isChecked();
 	info.hdr_enabled               = ui.checkBox_hdr->isChecked();
 	info.hide_cursor_enabled       = ui.checkBox_hide_cursor->isChecked();
+	info.stretch_enabled           = ui.checkBox_stretch->isChecked();
 	info.readback_linear_images    = ui.checkBox_readback->isChecked();
 	info.readback_tiled_images     = ui.checkBox_readback_tiled->isChecked();
 	info.sync_raw_image_buffers    = ui.checkBox_sync_raw_image_buffers->isChecked();
@@ -489,6 +535,7 @@ static void UpdateInfo(Configuration& info, Ui::ConfigurationEditDialog& ui, boo
 	info.shader_cache_enabled      = ui.checkBox_shader_cache->isChecked();
 	info.skip_notice_screen        = ui.checkBox_skip_notice_screen->isChecked();
 	info.network_connected         = ui.checkBox_network_connected->isChecked();
+	info.psn_emulation_enabled     = ui.checkBox_psn_emulation->isChecked();
 	info.trophy_enabled            = ui.checkBox_trophy_notifications->isChecked();
 	info.trophy_duration_seconds   = ui.spinBox_trophy_duration->value();
 	info.perf_overlay_enabled      = ui.checkBox_perf_overlay->isChecked();
@@ -497,6 +544,7 @@ static void UpdateInfo(Configuration& info, Ui::ConfigurationEditDialog& ui, boo
 	info.vulkan_validation_enabled = ui.checkBox_vulkan_validation->isChecked();
 	info.shader_validation_enabled = ui.checkBox_shader_validation->isChecked();
 	info.renderdoc_enabled         = ui.checkBox_renderdoc_capture->isChecked();
+	info.pro_enabled               = ui.checkBox_pro->isChecked();
 	info.amd_cpu_enabled           = ui.checkBox_amd_cpu->isChecked();
 #if defined(_WIN32)
 	info.red_zone_protection_enabled = ui.checkBox_red_zone_protection->isChecked();
@@ -518,7 +566,7 @@ void ConfigurationEditDialog::save() {
 	for (auto* field: {m_ui->lineEdit_shader_log_folder, m_ui->lineEdit_printf_file,
 	                   m_ui->lineEdit_cmd_dump_folder}) {
 		if (field->isEnabled() && field->text().isEmpty()) {
-			m_ui->settings_tabs->setCurrentWidget(m_ui->debug_tab);
+			ShowTab(m_ui->debug_tab);
 			field->setFocus();
 			break;
 		}
@@ -540,6 +588,16 @@ void ConfigurationEditDialog::save() {
 		                      tr("User ID cannot be 254 (everyone) or 255 (system)"));
 		return;
 	}
+	if (Configuration::NormalizeResolution(m_ui->comboBox_screen_resolution->currentText())
+	        .isEmpty()) {
+		ShowTab(m_ui->settings_tab);
+		m_ui->comboBox_screen_resolution->setFocus();
+		QMessageBox::critical(this, tr("Save failed"),
+		                      tr("Screen resolution must be WIDTHxHEIGHT, each between %1 and %2")
+		                          .arg(Configuration::MIN_SCREEN_SIZE)
+		                          .arg(Configuration::MAX_SCREEN_SIZE));
+		return;
+	}
 
 	UpdateInfo(m_info, *m_ui, m_global_settings);
 

src/launcher/src/mainDialog.cpp

@@ -134,7 +134,8 @@ void MainDialogPrivate::Setup(MainDialog* main_dialog) {
 	m_ui->close_terminal->setVisible(false); // The emulator runs without a terminal.
 #endif
 
-	main_dialog->setWindowFlags(Qt::Dialog /*| Qt::MSWindowsFixedSizeDialogHint*/);
+	main_dialog->setWindowFlags(Qt::Dialog | Qt::WindowMinMaxButtonsHint |
+	                            Qt::WindowCloseButtonHint);
 
 	connect(main_dialog, &MainDialog::Start, this, &MainDialogPrivate::FindInterpreter,
 	        Qt::QueuedConnection);
@@ -240,7 +241,7 @@ static void AddFlag(QStringList& args, bool enabled, const char* name) {
 
 static QStringList CreateEmulatorArgs(const Configuration& info) {
 	QStringList args;
-	auto        r = EnumToText(info.screen_resolution).split('x');
+	auto        r = Configuration::NormalizeResolution(info.screen_resolution).split('x');
 
 	if (r.size() != 2) {
 		return {};
@@ -266,6 +267,7 @@ static QStringList CreateEmulatorArgs(const Configuration& info) {
 	AddFlag(args, info.fullscreen_enabled, "--fullscreen");
 	AddFlag(args, info.hdr_enabled, "--hdr");
 	AddFlag(args, info.hide_cursor_enabled, "--hide-cursor");
+	AddFlag(args, info.stretch_enabled, "--stretch");
 	AddFlag(args, info.readback_linear_images, "--readback-linear-images");
 	AddFlag(args, info.readback_tiled_images, "--readback-tiled-images");
 	AddFlag(args, info.sync_raw_image_buffers, "--sync-raw-image-buffers");
@@ -274,6 +276,7 @@ static QStringList CreateEmulatorArgs(const Configuration& info) {
 	AddFlag(args, info.perf_overlay_enabled, "--perf-overlay");
 	AddFlag(args, info.skip_notice_screen, "--skip-notice-screen");
 	AddFlag(args, info.network_connected, "--network-connected");
+	AddFlag(args, info.psn_emulation_enabled, "--emulate-psn");
 	AddFlag(args, info.tessellation_enabled, "--tessellation");
 	AddFlag(args, info.occlusion_queries_enabled, "--occlusion-queries");
 	AddFlag(args, info.shader_cache_enabled, "--shader-cache");
@@ -289,6 +292,7 @@ static QStringList CreateEmulatorArgs(const Configuration& info) {
 	args << "--printf-direction" << EnumToText(info.printf_direction);
 	args << "--printf-output-file" << info.printf_output_file;
 	AddFlag(args, info.profiler_enabled, "--profile");
+	AddFlag(args, info.pro_enabled, "--trinity");
 	AddFlag(args, info.amd_cpu_enabled, "--amd-cpu");
 #if defined(_WIN32)
 	AddFlag(args, info.red_zone_protection_enabled, "--redzone");

src/libs/agc.cpp

@@ -24,6 +24,7 @@
 #include "kernel/pthread.h"
 #include "libs/errno.h"
 #include "libs/libs.h"
+#include "loader/systemContent.h"
 
 #include <algorithm>
 #include <array>
@@ -839,8 +840,8 @@ static bool is_native_reg_indirect_packet(const uint32_t* cmd, RegIndirectPacket
 	return ((cmd[0] >> 8u) & 0xffu) == reg_indirect_native_op(type);
 }
 
-static constexpr bool is_trinity_mode() {
-	return false;
+static bool is_trinity_mode() {
+	return Loader::SystemContentIsTrinityMode();
 }
 
 static uint32_t reg_indirect_pm4_r(RegIndirectPacket type) {

src/libs/ajm.cpp

@@ -6,6 +6,7 @@
 #include "libs/ajm/decoder.h"
 #include "libs/ajm/mp3_decoder.h"
 #include "libs/ajm/opus_decoder.h"
+#include "libs/ajm/resampler.h"
 #include "libs/audio.h"
 #include "libs/errno.h"
 #include "libs/libs.h"
@@ -63,12 +64,6 @@ struct AjmSidebandMFrame {
 	uint32_t reserved;
 };
 
-struct AjmSidebandResampleInfo {
-	float    ratio;
-	int32_t  num_samples;
-	uint32_t reserved[8];
-};
-
 struct AjmGetStatisticsResult {
 	AjmSidebandResult result;
 	float             engine_usage_batch;
@@ -104,6 +99,8 @@ constexpr uint64_t AJM_FLAG_RUN_GET_CODEC_INFO        = 1ull << 11u;
 constexpr uint64_t AJM_FLAG_RUN_MULTIPLE_FRAMES       = 1ull << 12u;
 constexpr uint64_t AJM_FLAG_CONTROL_RESET             = 1ull << 13u;
 constexpr uint64_t AJM_FLAG_CONTROL_INITIALIZE        = 1ull << 14u;
+constexpr uint64_t AJM_FLAG_CONTROL_RESAMPLE          = 1ull << 15u;
+constexpr uint64_t AJM_FLAG_CONTROL_RESAMPLE_EX       = 1ull << 16u;
 constexpr uint64_t AJM_FLAG_SIDEBAND_RESAMPLE_INFO    = 1ull << 44u;
 constexpr uint64_t AJM_FLAG_SIDEBAND_GAPLESS          = 1ull << 45u;
 constexpr uint64_t AJM_FLAG_SIDEBAND_FORMAT           = 1ull << 46u;
@@ -177,6 +174,7 @@ struct AjmInstanceState {
 	uint64_t                    flags   = 0;
 	AjmGaplessState             gapless;
 	std::unique_ptr<AjmDecoder> decoder;
+	AjmResampler                resampler;
 };
 
 static std::mutex                                     g_ajm_instances_mutex;
@@ -358,7 +356,7 @@ static void AjmWriteSideband(uint64_t flags, void* sideband_output, size_t sideb
 	    offset + AJM_SIDEBAND_RESAMPLE_INFO_SIZE <= sideband_output_size) {
 		auto* info = reinterpret_cast<AjmSidebandResampleInfo*>(
 		    static_cast<uint8_t*>(sideband_output) + offset);
-		info->ratio = 1.0f;
+		*info = decode_result.resample_info;
 		offset += AJM_SIDEBAND_RESAMPLE_INFO_SIZE;
 	}
 
@@ -425,6 +423,12 @@ static AjmDecodeResult AjmDecodeInstance(uint32_t instance, const void* input, s
 		return result;
 	}
 
+	// The resampler requires one complete decoded packet per decoder call.
+	if (state->codec == static_cast<uint32_t>(AjmCodec::DecOpus) &&
+	    (state->flags & AJM_INSTANCE_FLAG_RESAMPLE) != 0) {
+		return state->resampler.Decode(*state->decoder, state->flags, input, input_size, output,
+		                               output_size, multiple_frames, &state->gapless);
+	}
 	return state->decoder->Decode(input, input_size, output, output_size, multiple_frames,
 	                              &state->gapless);
 }
@@ -514,6 +518,32 @@ static AjmDecodeResult AjmControlInstance(uint32_t instance, uint64_t flags,
 		sideband_input = static_cast<const uint8_t*>(sideband_input) + AJM_SIDEBAND_GAPLESS_SIZE;
 		sideband_input_size -= AJM_SIDEBAND_GAPLESS_SIZE;
 	}
+	for (const auto flag : {AJM_FLAG_CONTROL_RESAMPLE, AJM_FLAG_CONTROL_RESAMPLE_EX}) {
+		if ((flags & flag) == 0) {
+			continue;
+		}
+		const bool   extended = flag == AJM_FLAG_CONTROL_RESAMPLE_EX;
+		const size_t size     = extended ? 12 : 8;
+		if (sideband_input == nullptr || sideband_input_size < size ||
+		    (state->flags & AJM_INSTANCE_FLAG_RESAMPLE) == 0) {
+			result.result = AJM_RESULT_INVALID_PARAMETER;
+			return result;
+		}
+		float       ratio, change = 0;
+		uint32_t    resample_flags;
+		const auto* data = static_cast<const uint8_t*>(sideband_input);
+		std::memcpy(&ratio, data, sizeof(ratio));
+		if (extended) {
+			std::memcpy(&change, data + sizeof(ratio), sizeof(change));
+		}
+		std::memcpy(&resample_flags, data + size - sizeof(resample_flags), sizeof(resample_flags));
+		result.result = state->resampler.SetParameters(ratio, change, resample_flags);
+		if (result.result != OK) {
+			return result;
+		}
+		sideband_input = data + size;
+		sideband_input_size -= size;
+	}
 	if ((flags & AJM_FLAG_CONTROL_RESET) != 0) {
 		state->gapless.Reset();
 		decoder->Reset();
@@ -524,6 +554,7 @@ static AjmDecodeResult AjmControlInstance(uint32_t instance, uint64_t flags,
 		result = decoder->Initialize(sideband_input, sideband_input_size);
 	}
 
+	result.resample_info = state->resampler.GetInfo();
 	return result;
 }
 
@@ -572,6 +603,7 @@ static AjmSidebandFormat AjmGetInstanceFormat(uint32_t          instance,
 	if (out_result != nullptr) {
 		*out_result        = state->decoder->MakeResult();
 		out_result->format = state->decoder->GetFormat();
+		out_result->resample_info = state->resampler.GetInfo();
 	}
 	return state->decoder->GetFormat();
 }
@@ -921,7 +953,13 @@ int KYTY_SYSV_ABI AjmBatchJobSetResampleParameters(AjmBatchInfo* info, uint32_t
                                                    float ratio, uint32_t flags, void* result) {
 	PRINT_NAME();
 	EXIT_NOT_IMPLEMENTED(info == nullptr);
-	AjmWriteResult(result, AJM_SIDEBAND_RESULT_SIZE);
+	const struct {
+		float    ratio;
+		uint32_t flags;
+	} params {ratio, flags};
+	const auto decoded = AjmControlInstance(instance, AJM_FLAG_CONTROL_RESAMPLE, &params,
+	                                        sizeof(params), nullptr, nullptr);
+	AjmWriteResult(result, AJM_SIDEBAND_RESULT_SIZE, decoded);
 	LOGF("\t instance = 0x%08" PRIx32 ", ratio = %f, flags = 0x%08" PRIx32 "\n", instance,
 	     static_cast<double>(ratio), flags);
 	return AjmAppendJob(info, AJM_JOB_SET_RESAMPLE_PARAMETERS_SIZE, "set-resample");
@@ -933,7 +971,14 @@ int KYTY_SYSV_ABI AjmBatchJobSetResampleParametersEx(AjmBatchInfo* info, uint32_
                                                      void* result) {
 	PRINT_NAME();
 	EXIT_NOT_IMPLEMENTED(info == nullptr);
-	AjmWriteResult(result, AJM_SIDEBAND_RESULT_SIZE);
+	const struct {
+		float    ratio;
+		float    change;
+		uint32_t flags;
+	} params {ratio_start, ratio_change_per_sample, flags};
+	const auto decoded = AjmControlInstance(instance, AJM_FLAG_CONTROL_RESAMPLE_EX, &params,
+	                                        sizeof(params), nullptr, nullptr);
+	AjmWriteResult(result, AJM_SIDEBAND_RESULT_SIZE, decoded);
 	LOGF("\t instance = 0x%08" PRIx32 ", ratio_start = %f, ratio_change = %f, flags = 0x%08" PRIx32
 	     "\n",
 	     instance, static_cast<double>(ratio_start), static_cast<double>(ratio_change_per_sample),
@@ -952,7 +997,7 @@ int KYTY_SYSV_ABI AjmBatchJobGetResampleInfo(AjmBatchInfo* info, uint32_t instan
 	if (result != nullptr) {
 		auto* resample  = reinterpret_cast<AjmSidebandResampleInfo*>(static_cast<uint8_t*>(result) +
 		                                                             AJM_SIDEBAND_RESULT_SIZE);
-		resample->ratio = 1.0f;
+		*resample = decode_result.resample_info;
 	}
 	LOGF("\t instance = 0x%08" PRIx32 "\n", instance);
 	return AjmAppendJob(info, AJM_JOB_RUN_SIZE, "get-resample-info");

src/libs/ajm/decoder.h

@@ -56,15 +56,22 @@ enum class AjmSampleEncoding : uint32_t {
 	Float = 2,
 };
 
+struct AjmSidebandResampleInfo {
+	float    ratio       = 1.0f;
+	int32_t  num_samples = 0;
+	uint32_t reserved[8] {};
+};
+
 struct AjmDecodeResult {
-	int32_t           result                = OK;
-	int32_t           internal_result       = OK;
-	size_t            input_consumed        = 0;
-	size_t            output_written        = 0;
-	uint64_t          total_decoded_samples = 0;
-	uint32_t          frames                = 0;
-	uint32_t          frames_per_packet     = 0;
-	AjmSidebandFormat format {};
+	int32_t                 result                = OK;
+	int32_t                 internal_result       = OK;
+	size_t                  input_consumed        = 0;
+	size_t                  output_written        = 0;
+	uint64_t                total_decoded_samples = 0;
+	uint32_t                frames                = 0;
+	uint32_t                frames_per_packet     = 0;
+	AjmSidebandFormat       format {};
+	AjmSidebandResampleInfo resample_info {};
 };
 
 constexpr int32_t AJM_RESULT_NOT_INITIALIZED    = 0x00000001;

src/libs/ajm/resampler.h

@@ -0,0 +1,182 @@
+#pragma once
+
+#include "libs/ajm/decoder.h"
+
+#include <algorithm>
+#include <cmath>
+#include <cstring>
+#include <vector>
+
+namespace Libs::Audio::Ajm {
+
+constexpr uint64_t AJM_INSTANCE_FLAG_GAPLESS_LOOP           = 1ull << 10u;
+constexpr uint64_t AJM_INSTANCE_FLAG_RESAMPLE               = 1ull << 12u;
+constexpr uint64_t AJM_INSTANCE_FLAG_SEAMLESS_LOOP          = 1ull << 13u;
+constexpr uint32_t AJM_RESAMPLE_FLAG_RESET                  = 1u;
+constexpr uint32_t AJM_RESAMPLE_FLAG_END                    = 2u;
+constexpr uint32_t AJM_RESAMPLE_FLAG_IGNORE_RATIO_PARAMETER = 8u;
+
+class AjmResampler {
+public:
+	int32_t SetParameters(float ratio, float change, uint32_t flags) {
+		if ((flags & AJM_RESAMPLE_FLAG_IGNORE_RATIO_PARAMETER) == 0 &&
+		    (!std::isfinite(ratio) || ratio < 0 || ratio > 4 || !std::isfinite(change))) {
+			return AJM_RESULT_INVALID_PARAMETER;
+		}
+		if ((flags & AJM_RESAMPLE_FLAG_RESET) != 0) {
+			m_pcm.clear();
+			m_position = 0;
+			m_end      = false;
+		}
+		if ((flags & AJM_RESAMPLE_FLAG_IGNORE_RATIO_PARAMETER) == 0) {
+			m_ratio  = ratio;
+			m_change = std::abs(change) < 0.000001f ? 0 : change;
+		}
+		m_end |= (flags & AJM_RESAMPLE_FLAG_END) != 0;
+		return OK;
+	}
+
+	[[nodiscard]] AjmSidebandResampleInfo GetInfo() const {
+		AjmSidebandResampleInfo info;
+		info.ratio       = static_cast<float>(m_ratio);
+		info.num_samples = m_end && m_position >= Samples()
+		                       ? -1
+		                       : static_cast<int32_t>(std::ceil(Samples() - m_position));
+		return info;
+	}
+
+	AjmDecodeResult Decode(AjmDecoder& decoder, uint64_t flags, const void* input,
+	                       size_t input_size, void* output, size_t output_size,
+	                       bool multiple_frames, AjmGaplessState* gapless) {
+		auto result = decoder.MakeResult();
+		if (result.format.channel_num == 0) {
+			return decoder.Decode(input, input_size, output, output_size, multiple_frames, gapless);
+		}
+		bool stop_decoding = false;
+		for (;;) {
+			const auto format       = decoder.GetFormat();
+			const auto encoding     = static_cast<AjmSampleEncoding>(format.sample_encoding);
+			const auto sample_bytes = format.channel_num * AjmBytesPerSample(encoding);
+			if (!m_pcm.empty() && sample_bytes != m_sample_bytes) {
+				result.result = AJM_RESULT_INVALID_PARAMETER;
+				break;
+			}
+			if (output == nullptr || output_size - result.output_written < sample_bytes) {
+				if (m_position < Samples() || result.input_consumed < input_size) {
+					result.result |= AJM_RESULT_NOT_ENOUGH_ROOM;
+				}
+				break;
+			}
+
+			const auto index     = static_cast<size_t>(m_position);
+			const auto fraction  = m_position - static_cast<double>(index);
+			const bool need_next = fraction != 0 && index + 1 >= Samples();
+			if (index >= Samples() || need_next) {
+				Compact();
+				if (!stop_decoding && input != nullptr && result.input_consumed < input_size &&
+				    (gapless == nullptr || !gapless->IsEnd())) {
+					// A decode job can produce up to six 20 ms Opus frames in one packet.
+					const auto channels = std::max<uint32_t>(format.channel_num, flags & 0x7fu);
+					m_frame.resize(5760 * channels * AjmBytesPerSample(encoding));
+					auto decoded =
+					    decoder.Decode(static_cast<const uint8_t*>(input) + result.input_consumed,
+						               input_size - result.input_consumed, m_frame.data(),
+						               m_frame.size(), false, gapless);
+					result.input_consumed += decoded.input_consumed;
+					result.frames += decoded.frames;
+					result.frames_per_packet     = decoded.frames_per_packet;
+					result.total_decoded_samples = decoded.total_decoded_samples;
+					result.format                = decoded.format;
+					result.result |= decoded.result;
+					result.internal_result = decoded.internal_result;
+					m_sample_bytes = decoded.format.channel_num * AjmBytesPerSample(encoding);
+					m_pcm.insert(m_pcm.end(), m_frame.begin(),
+					             m_frame.begin() + decoded.output_written);
+					stop_decoding =
+					    !multiple_frames || decoded.result != OK || decoded.input_consumed == 0;
+					if (gapless != nullptr && gapless->IsEnd() &&
+					    (flags & AJM_INSTANCE_FLAG_GAPLESS_LOOP) != 0) {
+						gapless->Reset();
+						stop_decoding |= (flags & AJM_INSTANCE_FLAG_SEAMLESS_LOOP) == 0;
+					}
+					continue;
+				}
+				if (!m_end || m_position >= Samples()) {
+					if (!m_end && !stop_decoding && (gapless == nullptr || !gapless->IsEnd())) {
+						result.result |= AJM_RESULT_PARTIAL_INPUT;
+					}
+					break;
+				}
+			}
+
+			const auto  current = static_cast<size_t>(m_position);
+			const auto  next    = std::min(current + 1, Samples() - 1);
+			auto*       dst     = static_cast<uint8_t*>(output) + result.output_written;
+			const auto* a       = m_pcm.data() + current * sample_bytes;
+			const auto* b       = m_pcm.data() + next * sample_bytes;
+			const auto  weight  = m_position - static_cast<double>(current);
+			if (weight == 0) {
+				std::memcpy(dst, a, sample_bytes);
+			} else {
+				switch (encoding) {
+					case AjmSampleEncoding::S16:
+						Interpolate<int16_t>(dst, a, b, format.channel_num, weight);
+						break;
+					case AjmSampleEncoding::S32:
+						Interpolate<int32_t>(dst, a, b, format.channel_num, weight);
+						break;
+					case AjmSampleEncoding::Float:
+						Interpolate<float>(dst, a, b, format.channel_num, weight);
+						break;
+				}
+			}
+			result.output_written += sample_bytes;
+			m_position += m_ratio;
+			m_ratio = std::clamp(m_ratio + m_change, 0.0, 4.0);
+			if (m_ratio == 0 || m_ratio == 4) {
+				m_change = 0;
+			}
+		}
+		Compact();
+		result.resample_info = GetInfo();
+		return result;
+	}
+
+private:
+	template <typename T>
+	static void Interpolate(uint8_t* dst, const uint8_t* a, const uint8_t* b, uint32_t channels,
+	                        double weight) {
+		for (uint32_t channel = 0; channel < channels; ++channel) {
+			T first, second;
+			std::memcpy(&first, a + channel * sizeof(T), sizeof(T));
+			std::memcpy(&second, b + channel * sizeof(T), sizeof(T));
+			const auto value =
+			    static_cast<T>(first + (static_cast<double>(second) - first) * weight);
+			std::memcpy(dst + channel * sizeof(T), &value, sizeof(T));
+		}
+	}
+
+	[[nodiscard]] size_t Samples() const {
+		return m_sample_bytes == 0 ? 0 : m_pcm.size() / m_sample_bytes;
+	}
+
+	void Compact() {
+		const auto consumed = std::min(static_cast<size_t>(m_position), Samples());
+		const auto bytes    = consumed * m_sample_bytes;
+		if (bytes != 0) {
+			std::memmove(m_pcm.data(), m_pcm.data() + bytes, m_pcm.size() - bytes);
+			m_pcm.resize(m_pcm.size() - bytes);
+			m_position -= static_cast<double>(consumed);
+		}
+	}
+
+	std::vector<uint8_t> m_pcm;
+	std::vector<uint8_t> m_frame;
+	size_t               m_sample_bytes = 0;
+	double               m_position     = 0;
+	double               m_ratio        = 1;
+	double               m_change       = 0;
+	bool                 m_end          = false;
+};
+
+} // namespace Libs::Audio::Ajm

src/libs/controller.cpp

@@ -519,7 +519,20 @@ void GameController::SetDevices(int active_id) {
 	m_trigger_state      = {};
 }
 
+// Input that changes no button or touch contact updates the newest unread sample, so a burst of
+// stick or sensor events between two reads cannot push unread button or touch changes out.
 void GameController::AddState() {
+	if (m_states_num != 0) {
+		const auto newest = (m_first_state + m_states_num - 1) % STATES_MAX;
+		auto&      last   = m_states[newest];
+		if (!m_obtained[newest] && last.buttons == m_state.buttons &&
+		    last.touch[0].down == m_state.touch[0].down &&
+		    last.touch[1].down == m_state.touch[1].down) {
+			last = m_state;
+			return;
+		}
+	}
+
 	if (m_states_num >= STATES_MAX) {
 		m_states_num  = STATES_MAX - 1;
 		m_first_state = (m_first_state + 1) % STATES_MAX;

src/libs/libAppContent.cpp

@@ -32,6 +32,7 @@ namespace AppContent {
 
 constexpr uint32_t APP_CONTENT_APPPARAM_SKU_FLAG_FULL = 3;
 constexpr int      APP_CONTENT_ERROR_PARAMETER        = -2133262334; /* 0x80D90002 */
+constexpr int      APP_CONTENT_ERROR_BUSY             = -2133262333; /* 0x80D90003 */
 constexpr int      APP_CONTENT_ERROR_NOT_MOUNTED      = -2133262332; /* 0x80D90004 */
 constexpr int      APP_CONTENT_ERROR_NOT_FOUND        = -2133262331; /* 0x80D90005 */
 constexpr int      APP_CONTENT_ERROR_MOUNT_FULL       = -2133262330; /* 0x80D90006 */
@@ -210,10 +211,8 @@ int KYTY_SYSV_ABI AppContentAddcontMount(uint32_t                         servic
 
 	for (const auto& m: g_addcont_mounts) {
 		if (m.label == entry->entitlement_label) {
-			const auto guest_point = m.point.substr(0, m.point.size() - 1);
-			std::strncpy(mount_point->data, guest_point.c_str(), sizeof(mount_point->data) - 1);
-			LOGF("\t already mounted %s\n", guest_point.c_str());
-			return OK;
+			LOGF("\t already mounted %s\n", m.point.c_str());
+			return APP_CONTENT_ERROR_BUSY;
 		}
 	}
 

src/libs/libAudio.cpp

@@ -1162,6 +1162,8 @@ LIB_DEFINE(InitAudio_1_Ngs2) {
 	LIB_FUNC("AbYvTOZ8Pts", Ngs2::Ngs2VoiceRunCommands);
 	LIB_FUNC("-TOuuAQ-buE", Ngs2::Ngs2VoiceGetState);
 	LIB_FUNC("rEh728kXk3w", Ngs2::Ngs2VoiceGetStateFlags);
+	LIB_FUNC("9eic4AmjGVI", Ngs2::Ngs2VoiceQueryInfo);
+	LIB_FUNC("WCayTgob7-o", Ngs2::Ngs2VoiceGetPortInfo);
 	LIB_FUNC("i0VnXM-C9fc", Ngs2::Ngs2SystemRender);
 	LIB_FUNC("l4Q2dWEH6UM", Ngs2::Ngs2SystemSetGrainSamples);
 	LIB_FUNC("hyVLT2VlOYk", Ngs2::Ngs2ParseWaveformData);

src/libs/libAudio2.cpp

@@ -13,6 +13,7 @@
 #include <atomic>
 #include <cmath>
 #include <cstring>
+#include <span>
 #include <vector>
 
 namespace Libs::Audio {
@@ -168,17 +169,12 @@ struct AudioOut2MasteringStatesV2 {
 
 static std::atomic_uint64_t g_audioout2_next_context {1};
 static std::atomic_uint64_t g_audioout2_next_port {1};
-static AudioOut2UserHandle g_audioout2_next_user = 1;
+static AudioOut2UserHandle  g_audioout2_next_user = 1;
 
-struct AudioOut2ContextState {
-	bool                   used                = false;
-	AudioOut2ContextHandle handle              = 0;
-	uint32_t               queue_depth         = 4;
-	uint32_t               queued              = 0;
-	uint32_t               num_grains          = 512;
-	uint64_t               last_update         = 0;
-	int                    object_audio_handle = 0;
-};
+static constexpr uint32_t AUDIOOUT2_CHANNEL_PORTS_MAX = 32;
+static constexpr uint32_t AUDIOOUT2_OBJECT_PORTS_MAX  = 512;
+static constexpr uint32_t AUDIOOUT2_PORT_SLOT_BITS    = 10;
+static constexpr uint32_t AUDIOOUT2_CONTEXT_SLOT_BITS = 4;
 
 struct AudioOut2PortStateEntry {
 	bool                   used          = false;
@@ -196,6 +192,19 @@ struct AudioOut2PortStateEntry {
 	bool                   attributes_locked = false;
 	std::vector<uint8_t>   pcm_data;
 };
+
+struct AudioOut2ContextState {
+	bool                   used                = false;
+	AudioOut2ContextHandle handle              = 0;
+	uint32_t               queue_depth         = 4;
+	uint32_t               queued              = 0;
+	uint32_t               num_grains          = 512;
+	uint64_t               last_update         = 0;
+	int                    object_audio_handle = 0;
+	uint32_t               max_ports           = 0;
+	// Channel slots precede object slots; each partition has its own creation limit.
+	std::vector<AudioOut2PortStateEntry> ports;
+};
 struct AudioOut2SpeakerArrayState {
 	bool                        used          = false;
 	uint32_t                    num_speakers  = 0;
@@ -223,13 +232,16 @@ struct AudioOut2Submission {
 static Common::Mutex                              g_audioout2_state_mutex;
 static std::array<AudioOut2ContextState, 16>      g_audioout2_contexts;
 static std::array<AudioOut2Submission, 16>        g_audioout2_submissions;
-static std::array<AudioOut2PortStateEntry, 256>   g_audioout2_ports;
-static std::vector<AudioOut2UserHandle>          g_audioout2_users;
+static std::vector<AudioOut2UserHandle>           g_audioout2_users;
 static Common::Mutex                              g_audioout2_speaker_array_mutex;
 static std::array<AudioOut2SpeakerArrayState, 32> g_audioout2_speaker_arrays;
 static Common::Mutex                              g_audioout2_latency_mutex;
 static std::array<AudioOut2LatencyState, 16>      g_audioout2_latencies;
 
+static_assert(AUDIOOUT2_CHANNEL_PORTS_MAX + AUDIOOUT2_OBJECT_PORTS_MAX <=
+              (1u << AUDIOOUT2_PORT_SLOT_BITS));
+static_assert(g_audioout2_contexts.size() == (1u << AUDIOOUT2_CONTEXT_SLOT_BITS));
+
 static constexpr int AUDIO_OUT2_ERROR_NOT_READY                   = -2144960504; /* 0x80268008 */
 static constexpr int AUDIO_OUT2_ERROR_PORT_FULL                   = -2144960494; /* 0x80268012 */
 static constexpr int AUDIO_OUT2_ERROR_INVALID_PARAM               = -2144960511; /* 0x80268001 */
@@ -262,6 +274,11 @@ static AudioOut2Submission& audioout2_submission(const AudioOut2ContextState& co
 	return g_audioout2_submissions[&context - g_audioout2_contexts.data()];
 }
 
+// Ticks of a sixth of a microsecond: one 48 kHz sample lasts exactly 125 of them.
+static uint64_t audioout2_clock() {
+	return LibKernel::KernelGetProcessTime() * 6;
+}
+
 static uint32_t audioout2_grain_micros(uint32_t grains) {
 	const auto sample_count = (grains == 0 ? 512u : grains);
 	return std::max<uint32_t>((sample_count * 1000000u) / 48000u, 1000u);
@@ -331,49 +348,51 @@ static void audioout2_update_context_locked(AudioOut2ContextState* state) {
 		return;
 	}
 
-	const auto now = LibKernel::KernelGetProcessTime();
+	const auto now = audioout2_clock();
 	if (state->last_update == 0 || state->queued == 0) {
 		state->last_update = now;
 		return;
 	}
-
-	const auto grain_micros = static_cast<uint64_t>(audioout2_grain_micros(state->num_grains));
-	if (now <= state->last_update || grain_micros == 0) {
+	if (now <= state->last_update) {
 		return;
 	}
 
-	const auto elapsed = now - state->last_update;
-	const auto drained = std::min<uint64_t>(state->queued, elapsed / grain_micros);
+	const auto grain   = uint64_t {state->num_grains} * 125;
+	const auto drained = std::min<uint64_t>(state->queued, (now - state->last_update) / grain);
 	if (drained == 0) {
 		return;
 	}
 
 	state->queued -= static_cast<uint32_t>(drained);
-	state->last_update += drained * grain_micros;
+	state->last_update += drained * grain;
 	if (state->queued == 0) {
 		state->last_update = now;
 	}
 }
 
 static AudioOut2PortStateEntry* audioout2_find_port_locked(AudioOut2PortHandle port) {
-	for (auto& state: g_audioout2_ports) {
-		if (state.used && state.handle == port) {
-			return &state;
-		}
+	const auto context_slot =
+	    (port >> AUDIOOUT2_PORT_SLOT_BITS) & ((1u << AUDIOOUT2_CONTEXT_SLOT_BITS) - 1u);
+	const auto port_slot = port & ((1u << AUDIOOUT2_PORT_SLOT_BITS) - 1u);
+	auto&      context   = g_audioout2_contexts[context_slot];
+	if (context.used && port_slot < context.ports.size()) {
+		auto& state = context.ports[port_slot];
+		if (state.used && state.handle == port) return &state;
 	}
 
 	return nullptr;
 }
 
 static size_t audioout2_pcm_size(const AudioOut2PortStateEntry& state) {
-	const auto bytes_per_sample = (state.data_format & 0x7fu) == 1 ? sizeof(int16_t) : sizeof(float);
+	const auto bytes_per_sample =
+	    (state.data_format & 0x7fu) == 1 ? sizeof(int16_t) : sizeof(float);
 	return static_cast<size_t>(state.samples_num) *
 	       audioout2_data_format_channels(state.data_format) * bytes_per_sample;
 }
 
 static bool audioout2_context_has_queueable_device(const AudioOut2ContextState& context) {
-	for (const auto& state: g_audioout2_ports) {
-		if (state.used && state.context == context.handle && !state.pcm_data.empty()) {
+	for (const auto& state: context.ports) {
+		if (state.used && !state.pcm_data.empty()) {
 			const int handle = audioout2_port_type_is_object(state.port_type)
 			                       ? context.object_audio_handle
 			                       : state.audio_handle;
@@ -427,18 +446,21 @@ static void audioout2_queue_context_audio(AudioOut2ContextHandle ctx, AudioOut2S
 	out.object_mix.clear();
 	std::vector<AudioInternal::OutputParam> params;
 	params.reserve(AudioInternal::OUT_PORTS_MAX);
-	for (const auto& state: g_audioout2_ports) {
-		if (!state.used || state.context != ctx || state.pcm_data.empty()) {
+	// A port's PCM covers one grain: the push consumes it, so a port the game stops feeding
+	// falls silent.
+	for (auto& state: context->ports) {
+		if (!state.used || state.pcm_data.empty()) {
 			continue;
 		}
 		if (audioout2_port_type_is_object(state.port_type)) {
 			audioout2_mix_object(out.object_mix, context->num_grains, state);
 		} else if (state.audio_handle > 0 && params.size() < AudioInternal::OUT_PORTS_MAX) {
 			const auto i = params.size();
-			out.pcm[i]   = state.pcm_data;
+			out.pcm[i].swap(state.pcm_data);
 			out.gains[i] = state.gains;
 			params.push_back({state.audio_handle, out.pcm[i].data(), out.gains[i].data()});
 		}
+		state.pcm_data.clear();
 	}
 	if (!out.object_mix.empty() && params.size() < AudioInternal::OUT_PORTS_MAX) {
 		params.push_back({context->object_audio_handle, out.object_mix.data()});
@@ -467,22 +489,28 @@ int KYTY_SYSV_ABI AudioOut2ContextResetParam(AudioOut2ContextParam* params) {
 	EXIT_NOT_IMPLEMENTED(params == nullptr);
 
 	std::memset(params, 0, sizeof(AudioOut2ContextParam));
-	params->max_ports              = 256;
-	params->max_object_ports       = 256;
-	params->guarantee_object_ports = 0;
-	params->queue_depth            = 4;
-	params->num_grains             = 512;
-	params->flags                  = 1;
+	params->max_ports   = 8;
+	params->queue_depth = 1;
+	params->num_grains  = 256;
 
 	return OK;
 }
 
+static bool audioout2_valid_context_capacity(const AudioOut2ContextParam& params) {
+	return params.max_ports <= AUDIOOUT2_CHANNEL_PORTS_MAX &&
+	       params.max_object_ports <= AUDIOOUT2_OBJECT_PORTS_MAX &&
+	       params.guarantee_object_ports <= params.max_object_ports;
+}
+
 int KYTY_SYSV_ABI AudioOut2ContextQueryMemory(const AudioOut2ContextParam* params,
                                               size_t*                      memory_size) {
 	PRINT_NAME();
 
 	EXIT_NOT_IMPLEMENTED(params == nullptr);
 	EXIT_NOT_IMPLEMENTED(memory_size == nullptr);
+	if (!audioout2_valid_context_capacity(*params)) {
+		return AUDIO_OUT2_ERROR_INVALID_PARAM;
+	}
 
 	const auto queue_depth = (params->queue_depth == 0 ? 4u : params->queue_depth);
 	*memory_size           = 0x10000u + static_cast<size_t>(queue_depth) * 0x590u;
@@ -498,6 +526,9 @@ int KYTY_SYSV_ABI AudioOut2ContextCreate(const AudioOut2ContextParam* params, vo
 
 	EXIT_NOT_IMPLEMENTED(params == nullptr);
 	EXIT_NOT_IMPLEMENTED(ctx == nullptr);
+	if (!audioout2_valid_context_capacity(*params)) {
+		return AUDIO_OUT2_ERROR_INVALID_PARAM;
+	}
 
 	const auto num_grains = params->num_grains == 0 ? 512u : params->num_grains;
 	const int  object_audio_handle =
@@ -524,8 +555,10 @@ int KYTY_SYSV_ABI AudioOut2ContextCreate(const AudioOut2ContextParam* params, vo
 	state->handle              = *ctx;
 	state->queue_depth         = params->queue_depth == 0 ? 4u : params->queue_depth;
 	state->num_grains          = num_grains;
-	state->last_update         = LibKernel::KernelGetProcessTime();
+	state->last_update         = audioout2_clock();
 	state->object_audio_handle = object_audio_handle;
+	state->max_ports           = params->max_ports;
+	state->ports.resize(params->max_ports + params->max_object_ports);
 	g_audioout2_state_mutex.Unlock();
 
 	LOGF("\t buffer      = 0x%016" PRIx64 "\n"
@@ -541,8 +574,8 @@ int KYTY_SYSV_ABI AudioOut2ContextDestroy(AudioOut2ContextHandle ctx) {
 	PRINT_NAME();
 	LOGF("\t ctx = 0x%016" PRIx64 "\n", ctx);
 
-	std::array<int, 257> audio_handles {};
-	size_t audio_handles_num = 0;
+	std::array<int, AUDIOOUT2_CHANNEL_PORTS_MAX + 1> audio_handles {};
+	size_t                                           audio_handles_num = 0;
 
 	g_audioout2_state_mutex.Lock();
 	auto* state = audioout2_find_context_locked(ctx);
@@ -550,15 +583,12 @@ int KYTY_SYSV_ABI AudioOut2ContextDestroy(AudioOut2ContextHandle ctx) {
 		if (state->object_audio_handle > 0) {
 			audio_handles[audio_handles_num++] = state->object_audio_handle;
 		}
-		*state = AudioOut2ContextState {};
-	}
-	for (auto& port_state: g_audioout2_ports) {
-		if (port_state.used && port_state.context == ctx) {
-			if (port_state.audio_handle > 0 && audio_handles_num < audio_handles.size()) {
+		for (const auto& port_state: state->ports) {
+			if (port_state.used && port_state.audio_handle > 0) {
 				audio_handles[audio_handles_num++] = port_state.audio_handle;
 			}
-			port_state = AudioOut2PortStateEntry {};
 		}
+		*state = AudioOut2ContextState {};
 	}
 	g_audioout2_state_mutex.Unlock();
 
@@ -586,9 +616,7 @@ int KYTY_SYSV_ABI AudioOut2ContextAdvance(AudioOut2ContextHandle ctx) {
 	g_audioout2_state_mutex.Lock();
 	if (auto* state = audioout2_find_context_locked(ctx); state != nullptr) {
 		audioout2_update_context_locked(state);
-	}
-	for (auto& port: g_audioout2_ports) {
-		if (port.used && port.context == ctx) {
+		for (auto& port: state->ports) {
 			port.attributes_locked = false;
 		}
 	}
@@ -612,7 +640,7 @@ int KYTY_SYSV_ABI AudioOut2ContextPush(AudioOut2ContextHandle ctx, uint32_t bloc
 			sleep_micros = audioout2_grain_micros(state->num_grains);
 			if (state->queued < state->queue_depth || use_device_clock) {
 				if (state->queued == 0) {
-					state->last_update = LibKernel::KernelGetProcessTime();
+					state->last_update = audioout2_clock();
 				}
 				if (state->queued < state->queue_depth) {
 					state->queued++;
@@ -663,21 +691,19 @@ int KYTY_SYSV_ABI AudioOut2PortCreate(AudioOut2ContextHandle ctx, const AudioOut
 	EXIT_NOT_IMPLEMENTED(params == nullptr);
 	EXIT_NOT_IMPLEMENTED(port == nullptr);
 
-	const auto next_port    = g_audioout2_next_port.fetch_add(1, std::memory_order_relaxed);
 	const auto audio_format = audioout2_data_format_to_audio_format(params->data_format);
 	const auto audio_type   = audioout2_port_type_to_audio_out_type(params->port_type);
+	const bool is_object    = audioout2_port_type_is_object(params->port_type);
 
 	g_audioout2_state_mutex.Lock();
-	const auto* context_state = audioout2_find_context_locked(ctx);
+	auto* context_state = audioout2_find_context_locked(ctx);
 	if (context_state == nullptr) {
 		g_audioout2_state_mutex.Unlock();
 		return AUDIO_OUT2_ERROR_INVALID_PARAM;
 	}
-	if (audioout2_port_type_is_object(params->port_type) &&
-	    (params->port_type != 0x100 || context_state->object_audio_handle == 0 ||
-	     params->sampling_freq != 48000 ||
-	     (audio_format != AudioInternal::Format::FloatMono &&
-	      audio_format != AudioInternal::Format::Signed16bitMono))) {
+	if (is_object && (params->port_type != 0x100 || params->sampling_freq != 48000 ||
+	                  (audio_format != AudioInternal::Format::FloatMono &&
+	                   audio_format != AudioInternal::Format::Signed16bitMono))) {
 		g_audioout2_state_mutex.Unlock();
 		return AUDIO_OUT2_ERROR_INVALID_PARAM;
 	}
@@ -688,42 +714,45 @@ int KYTY_SYSV_ABI AudioOut2PortCreate(AudioOut2ContextHandle ctx, const AudioOut
 		g_audioout2_state_mutex.Unlock();
 		return AUDIO_OUT2_ERROR_INVALID_PARAM;
 	}
-	AudioOut2PortStateEntry* port_state = nullptr;
-	for (auto& candidate: g_audioout2_ports) {
-		if (!candidate.used) {
-			port_state = &candidate;
-			break;
-		}
-	}
-	if (port_state != nullptr) {
-		*port_state               = AudioOut2PortStateEntry {};
-		port_state->used          = true;
-		port_state->handle        = next_port;
-		port_state->context       = ctx;
-		port_state->user          = params->user_handle;
-		port_state->port_type     = params->port_type;
-		port_state->data_format   = params->data_format;
-		port_state->sampling_freq = params->sampling_freq;
-		port_state->samples_num   = samples_num;
-		port_state->audio_format  = audio_format;
-		port_state->gains.fill(1.0f);
-	}
-	g_audioout2_state_mutex.Unlock();
-
-	if (port_state == nullptr) {
+	auto slots = std::span(context_state->ports);
+	slots =
+	    is_object ? slots.subspan(context_state->max_ports) : slots.first(context_state->max_ports);
+	const auto available = std::find_if(slots.begin(), slots.end(),
+	                                    [](const auto& candidate) { return !candidate.used; });
+	if (available == slots.end()) {
+		g_audioout2_state_mutex.Unlock();
 		return AUDIO_OUT2_ERROR_PORT_FULL;
 	}
+	const auto serial       = g_audioout2_next_port.fetch_add(1, std::memory_order_relaxed);
+	const auto context_slot = static_cast<uint64_t>(context_state - g_audioout2_contexts.data());
+	const auto port_slot    = static_cast<uint64_t>(&*available - context_state->ports.data());
+	const auto next_port    = (serial << (AUDIOOUT2_PORT_SLOT_BITS + AUDIOOUT2_CONTEXT_SLOT_BITS)) |
+	                          (context_slot << AUDIOOUT2_PORT_SLOT_BITS) | port_slot;
+	auto*      port_state   = &*available;
+	*port_state             = AudioOut2PortStateEntry {};
+	port_state->used        = true;
+	port_state->handle      = next_port;
+	port_state->context     = ctx;
+	port_state->user        = params->user_handle;
+	port_state->port_type   = params->port_type;
+	port_state->data_format = params->data_format;
+	port_state->sampling_freq = params->sampling_freq;
+	port_state->samples_num   = samples_num;
+	port_state->audio_format  = audio_format;
+	port_state->gains.fill(1.0f);
+	g_audioout2_state_mutex.Unlock();
 
 	int audio_handle = 0;
 
-	if (audio_format != AudioInternal::Format::Unknown &&
-	    !audioout2_port_type_is_object(params->port_type)) {
+	if (audio_format != AudioInternal::Format::Unknown && !is_object) {
 		audio_handle = AudioInternal::AudioOutOpen(audio_type, samples_num, params->sampling_freq,
 		                                           audio_format);
 	}
 
 	g_audioout2_state_mutex.Lock();
-	const bool reserved = port_state->used && port_state->handle == next_port;
+	// Context destruction may have freed or replaced the slots while the backend opened.
+	port_state          = audioout2_find_port_locked(next_port);
+	const bool reserved = port_state != nullptr;
 	if (reserved) {
 		port_state->audio_handle = audio_handle;
 	}
@@ -735,7 +764,7 @@ int KYTY_SYSV_ABI AudioOut2PortCreate(AudioOut2ContextHandle ctx, const AudioOut
 
 	*port = next_port;
 
-	if (next_port <= 16 || (next_port % 600) == 0) {
+	if (serial <= 16 || (serial % 600) == 0) {
 		PRINT_NAME();
 		LOGF("\t ctx           = 0x%016" PRIx64 "\n"
 		     "\t port          = 0x%016" PRIx64 "\n"
@@ -901,21 +930,23 @@ int KYTY_SYSV_ABI AudioOut2UserDestroy(AudioOut2UserHandle handle) {
 	PRINT_NAME();
 	LOGF("\t handle = 0x%016" PRIx64 "\n", static_cast<uint64_t>(handle));
 	Common::LockGuard lock(g_audioout2_state_mutex);
-	const auto it = std::find(g_audioout2_users.begin(), g_audioout2_users.end(), handle);
+	const auto        it = std::find(g_audioout2_users.begin(), g_audioout2_users.end(), handle);
 	if (it == g_audioout2_users.end()) {
 		return AUDIO_OUT2_ERROR_INVALID_PARAM;
 	}
-	if (std::any_of(g_audioout2_ports.begin(), g_audioout2_ports.end(),
-	                [handle](const auto& port) { return port.used && port.user == handle; })) {
-		return AUDIO_OUT2_ERROR_BUSY;
+	for (const auto& context: g_audioout2_contexts) {
+		if (std::any_of(context.ports.begin(), context.ports.end(),
+		                [handle](const auto& port) { return port.used && port.user == handle; })) {
+			return AUDIO_OUT2_ERROR_BUSY;
+		}
 	}
 	g_audioout2_users.erase(it);
 	return OK;
 }
 
 int KYTY_SYSV_ABI AudioOut2UserGetSupportedAttributes(AudioOut2UserHandle handle,
-                                                       uint32_t* context_attributes,
-                                                       uint32_t* port_attributes) {
+                                                      uint32_t*           context_attributes,
+                                                      uint32_t*           port_attributes) {
 	if (context_attributes == nullptr || port_attributes == nullptr) {
 		return AUDIO_OUT2_ERROR_INVALID_PARAM;
 	}

src/libs/libDbgAsan.cpp

@@ -38,8 +38,8 @@ struct Global {
 };
 
 struct Shadow {
-	uintptr_t start = 0;
-	int       count = 0;
+	uintptr_t start         = 0;
+	uint64_t  covered_bytes = 0;
 };
 
 struct AsanContext {
@@ -50,6 +50,14 @@ struct AsanContext {
 static uint32_t     g_fake_stack = 0;
 static AsanContext* g_ctx        = nullptr;
 
+// Allocation and release callbacks can partition the same guest range differently.
+static uint64_t ShadowCoverage(uintptr_t page, uintptr_t min_addr, uintptr_t max_addr) {
+	const auto guest_start = (page - 0x10000000000ull) << 3u;
+	const auto guest_end   = guest_start + 0x80000ull - 1;
+	return std::min<uintptr_t>(max_addr, guest_end) - std::max<uintptr_t>(min_addr, guest_start) +
+	       1;
+}
+
 static void AllocShadow(uintptr_t min_addr, uintptr_t max_addr) {
 	EXIT_IF(g_ctx == nullptr);
 
@@ -67,10 +75,11 @@ static void AllocShadow(uintptr_t min_addr, uintptr_t max_addr) {
 	     static_cast<uint64_t>(min_shadow), static_cast<uint64_t>(size));
 
 	for (uintptr_t page = min_shadow; page <= max_shadow; page += page_size) {
-		const auto it = std::find_if(g_ctx->shadows.begin(), g_ctx->shadows.end(),
-		                             [page](const Shadow& s) { return s.start == page; });
+		const auto bytes = ShadowCoverage(page, min_addr, max_addr);
+		const auto it    = std::find_if(g_ctx->shadows.begin(), g_ctx->shadows.end(),
+		                                [page](const Shadow& s) { return s.start == page; });
 		if (it != g_ctx->shadows.end()) {
-			it->count++;
+			it->covered_bytes += bytes;
 		} else {
 			if (!Common::VirtualMemory::AllocFixed(page, page_size,
 			                                       Common::VirtualMemory::Mode::ReadWrite)) {
@@ -78,8 +87,8 @@ static void AllocShadow(uintptr_t min_addr, uintptr_t max_addr) {
 			}
 
 			Shadow s;
-			s.start = page;
-			s.count = 1;
+			s.start         = page;
+			s.covered_bytes = bytes;
 			g_ctx->shadows.push_back(s);
 		}
 	}
@@ -105,14 +114,15 @@ static void FreeShadow(uintptr_t min_addr, uintptr_t max_addr) {
 		const auto it = std::find_if(g_ctx->shadows.begin(), g_ctx->shadows.end(),
 		                             [page](const Shadow& s) { return s.start == page; });
 		if (it != g_ctx->shadows.end()) {
-			it->count--;
-
-			if (it->count <= 0) {
+			const auto bytes = ShadowCoverage(page, min_addr, max_addr);
+			if (bytes >= it->covered_bytes) {
 				if (!Common::VirtualMemory::Free(page)) {
 					EXIT("can't free shadow memory\n");
 				}
 
 				g_ctx->shadows.erase(it);
+			} else {
+				it->covered_bytes -= bytes;
 			}
 		}
 	}
@@ -126,6 +136,14 @@ static void KernelFree(uintptr_t addr, size_t size) {
 	FreeShadow(addr, addr + size - 1);
 }
 
+#if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
+void TestRegisterMemoryCallbacks() {
+	EXIT_IF(g_ctx != nullptr);
+	g_ctx = new AsanContext;
+	LibKernel::Memory::RegisterCallbacks(KernelAlloc, KernelFree);
+}
+#endif
+
 static void KYTY_SYSV_ABI asan_init() {
 	PRINT_NAME();
 

src/libs/libFont.cpp

@@ -2,10 +2,12 @@
 #include "common/assert.h"
 #include "common/common.h"
 #include "common/logging/log.h"
+#include "common/stringUtils.h"
 #include "libs/errno.h"
 #include "libs/libs.h"
 #include "loader/symbolDatabase.h"
 
+#include <SDL3/SDL.h>
 #include <algorithm>
 #include <array>
 #include <cinttypes>
@@ -675,19 +677,6 @@ static bool font_is_italic(const FontState* font) {
 	       font->render_effect_slant != 0.0f;
 }
 
-static const char* fallback_font_file_name(bool bold, bool italic) {
-	if (bold && italic) {
-		return "Roboto-BoldItalic.ttf";
-	}
-	if (bold) {
-		return "Roboto-Bold.ttf";
-	}
-	if (italic) {
-		return "Roboto-Italic.ttf";
-	}
-	return "Roboto-Regular.ttf";
-}
-
 static std::vector<uint8_t> read_file_bytes(const std::filesystem::path& path) {
 	std::ifstream file(path, std::ios::binary | std::ios::ate);
 	if (!file) {
@@ -709,49 +698,34 @@ static std::vector<uint8_t> read_file_bytes(const std::filesystem::path& path) {
 	return data;
 }
 
-static std::vector<std::filesystem::path> fallback_font_candidates(const char* file_name) {
-	const auto rel =
-	    std::filesystem::path("3rdparty") / "tracy" / "profiler" / "src" / "font" / file_name;
-	std::vector<std::filesystem::path> paths;
-
-	auto source_root = std::filesystem::path(__FILE__).parent_path().parent_path().parent_path();
-	paths.push_back(source_root / rel);
-
-	std::error_code ec;
-	auto            cwd = std::filesystem::current_path(ec);
-	if (!ec) {
-		paths.push_back(cwd / rel);
-		paths.push_back(cwd / ".." / rel);
-		paths.push_back(cwd / ".." / ".." / rel);
+static std::vector<uint8_t> load_fallback_font_data(const char* file_name) {
+	const char* base_path = SDL_GetBasePath();
+	if (base_path == nullptr) {
+		LOGF("Font: SDL_GetBasePath failed: %s\n", SDL_GetError());
+		return {};
 	}
-
-	return paths;
-}
-
-static std::vector<uint8_t> load_fallback_font_data(bool bold, bool italic) {
-	for (const auto& path: fallback_font_candidates(fallback_font_file_name(bold, italic))) {
-		auto data = read_file_bytes(path);
-		if (!data.empty()) {
-			return data;
-		}
+	// SDL returns UTF-8; path(const char*) would read it in the Windows ANSI code page.
+	const auto path = Common::PathFromUtf8(base_path) / "assets" / "fonts" / file_name;
+	auto       data = read_file_bytes(path);
+	if (data.empty()) {
+		LOGF("Font: fallback font not found: %s\n", Common::PathToString(path).c_str());
 	}
-
-	return {};
+	return data;
 }
 
 static const std::vector<uint8_t>& fallback_font_data(bool bold, bool italic) {
-	static const auto regular     = load_fallback_font_data(false, false);
-	static const auto bold_data   = load_fallback_font_data(true, false);
-	static const auto italic_data = load_fallback_font_data(false, true);
-	static const auto bold_italic = load_fallback_font_data(true, true);
+	static const auto regular     = load_fallback_font_data("NotoSans-Regular.ttf");
+	static const auto bold_data   = load_fallback_font_data("NotoSans-Bold.ttf");
+	static const auto italic_data = load_fallback_font_data("NotoSans-Italic.ttf");
+	static const auto bold_italic = load_fallback_font_data("NotoSans-BoldItalic.ttf");
 
-	if (bold && italic && !bold_italic.empty()) {
+	if (bold && italic) {
 		return bold_italic;
 	}
-	if (bold && !bold_data.empty()) {
+	if (bold) {
 		return bold_data;
 	}
-	if (italic && !italic_data.empty()) {
+	if (italic) {
 		return italic_data;
 	}
 	return regular;

src/libs/libKernel.cpp

@@ -20,6 +20,7 @@
 #include "libs/network.h"
 #include "loader/elf.h"
 #include "loader/runtimeLinker.h"
+#include "loader/systemContent.h"
 #include "loader/symbolDatabase.h"
 
 #include <algorithm>
@@ -1321,23 +1322,12 @@ static void* KYTY_SYSV_ABI KernelGetProcParam() {
 }
 
 static int KYTY_SYSV_ABI KernelIsTrinityMode() {
-	static std::atomic_uint32_t call_count {0};
-
-	auto count = call_count.fetch_add(1, std::memory_order_relaxed);
-	if (count < 8 || (count % 1024) == 0) {
-		PRINT_NAME();
-		LOGF("\tcall = %u, return = 0 (base PS5 mode)\n", count + 1);
-	}
-
-	return 0;
+	return Loader::SystemContentIsTrinityMode() ? 1 : 0;
 }
 
 static int KYTY_SYSV_ABI KernelGetOperationMode(int* mode, int* submode) {
-	PRINT_NAME();
-
-	*mode    = 2; // PS5 Base
+	*mode    = Loader::SystemContentIsTrinityMode() ? 3 : 2;
 	*submode = 0; // None
-	LOGF("\t mode = %d, submode = %d\n", *mode, *submode);
 	return OK;
 }