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

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
@@ -522,6 +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"
 )
@@ -624,6 +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)
 
@@ -668,6 +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>)
@@ -684,6 +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
@@ -692,6 +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
@@ -700,8 +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
@@ -764,6 +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

src/common/emulatorConfig.cpp

@@ -113,6 +113,10 @@ bool VrEnabled() {
 	return g_config->vr_enabled;
 }
 
+bool Ps5ProEnabled() {
+	return g_config->pro_enabled;
+}
+
 bool AmdCpuEnabled() {
 	return g_config->amd_cpu_enabled;
 }
@@ -133,6 +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

@@ -67,8 +67,10 @@ struct ConfigOptions {
 	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;
@@ -128,8 +130,10 @@ 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/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);
-					}
 					for (auto& queue: gpu->m_queues) {
 						if (!queue.empty()) {
 							queue.front().blocked = false;
@@ -866,9 +856,8 @@ 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);
-	// The GPU reads the records if the shaders allow; the VS prolog loads the patched SGPRs from
-	// Vulkan's bases. With no SGPR patched, Vulkan applies the offsets, which matches while they
-	// are zero.
+	// 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(),
@@ -940,6 +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) {
@@ -960,11 +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,

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/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;
 		}
 	}
@@ -289,6 +297,7 @@ ClassifyVideoOutCompression(bool compressed, uint64_t metadata_address, uint32_t
                             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_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) {
+	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);
 	};
@@ -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/pipelineCache.cpp

@@ -689,6 +689,7 @@ void PipelineCache::GetGraphicsPrograms(const CommandBuffer& command, DrawRender
 	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.

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

@@ -139,6 +139,7 @@ struct DrawRenderState {
 	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/render.h

@@ -150,15 +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);
@@ -219,6 +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/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};
 	}
 
@@ -1251,16 +1251,26 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 }
 
 // Vulkan offsets for a prepared draw, or nullopt when its indirect records need the CPU: mesh
-// and quad-list draws for counts, tessellation draws for patched SGPRs. Indirect records carry
-// their own offsets.
+// 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 ||
+		if (!vs.indirect_draw ||
+		    vs.stage.module->info.stage == ShaderType::Mesh ||
 		    ucfg.GetPrimType() == Prospero::PrimitiveType::kQuadListLegacy ||
 		    vs.indirect_offsets != indirect.offsets) {
 			return std::nullopt;
@@ -1269,7 +1279,7 @@ static std::optional<DrawEmitInfo> ResolveDrawEmit(const HW::UserConfig&
 	}
 	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};
@@ -1326,7 +1336,8 @@ bool RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
 
 	hw_check(buffer);
 
-	DrawRenderState state {.indirect_offsets = args.indirect.offsets};
+	DrawRenderState state {.indirect_offsets = args.indirect.offsets,
+	                       .indirect_draw    = args.indirect.address != 0};
 	if (!GetDrawTopology(ucfg, state.topology)) {
 		return true;
 	}
@@ -1431,7 +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 {.indirect_offsets = args.indirect.offsets};
+	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/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/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 has_request = [this, &cfg, index] {
 		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;
+		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/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/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

@@ -1026,10 +1026,8 @@ 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);
@@ -1072,7 +1070,6 @@ uint32_t LoadIndirectFormattedX(ValueEmitContext& ctx, const IR::Inst& inst) {
 	                          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/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

@@ -30,11 +30,13 @@ uint32_t TessellationPointer(ValueEmitContext& ctx, const IR::Inst& inst) {
 	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 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)));
 	};
+	IR::U32 quads;
+	if (is64) {
 		const auto source = ReadU32Pair(inst.src0);
-	const auto quads  = ir.BitwiseOr(
+		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})})));
+		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)));
@@ -1380,21 +1388,17 @@ 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.
+			// 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 (offsets == ShaderIndirectOffsets {}) {
+				if (!options.input_info.vertex->indirect_draw) {
 					return builtin(kind);
 				}
 				const auto base = builtin(base_kind);

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/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

@@ -543,6 +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/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

@@ -515,6 +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

@@ -291,8 +291,7 @@ 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.has_dynamic_vertex_fetch = true;
 		info.fetch_attrib_reg         = metadata.vertex_attrib_reg;
 		info.fetch_buffer_reg         = metadata.vertex_buffer_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);
@@ -422,7 +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));
+	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) {
@@ -693,9 +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

@@ -125,8 +125,7 @@ struct ShaderTessellationInputInfo {
 };
 
 // GS user-data slots an indirect draw packet writes its record's base vertex and first instance
-// to. Like the hardware index VGPRs, the vertex shader's indices exclude both bases once either
-// slot is set.
+// to. A packet can leave both slots unused even when its record contains nonzero offsets.
 struct ShaderIndirectOffsets {
 	static constexpr uint8_t None = UINT8_MAX;
 
@@ -159,8 +158,10 @@ struct ShaderVertexInputInfo {
 	ShaderMeshInputInfo      mesh;
 	ShaderTessellationInputInfo tess;
 	ShaderIndirectOffsets   indirect_offsets;
-	bool                    fetch_external      = false;
-	bool                    fetch_embedded      = false;
+	// 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.cpp

@@ -1006,8 +1006,10 @@ void Initialize() {
 	const auto table_overhead = g_cpu_table_memory + g_gpu_table_memory +
 	                            std::max(params.amm_table_memory, INCLUDED_TABLE_MEMORY) -
 	                            3 * INCLUDED_TABLE_MEMORY;
-	EXIT_IF(g_flexible_memory_size + table_overhead >= GUEST_MEMORY_BUDGET - RESERVED_CACHE_SIZE);
-	g_guest_memory_size = GUEST_MEMORY_BUDGET - RESERVED_CACHE_SIZE - table_overhead;
+	const auto memory_budget = GUEST_MEMORY_BUDGET +
+	                           (Loader::SystemContentIsTrinityMode() ? 1216ull * 1024ull * 1024ull : 0);
+	EXIT_IF(g_flexible_memory_size + table_overhead >= memory_budget - RESERVED_CACHE_SIZE);
+	g_guest_memory_size = memory_budget - RESERVED_CACHE_SIZE - table_overhead;
 	LOGF("\t flexible memory size = %" PRIu64 " MiB, direct memory size = %" PRIu64 " MiB\n",
 	     g_flexible_memory_size / (1024ull * 1024ull),
 	     (g_guest_memory_size - g_flexible_memory_size) / (1024ull * 1024ull));

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/forms/configuration_edit_dialog.ui

@@ -254,6 +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>
@@ -406,6 +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

@@ -128,6 +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;
@@ -141,6 +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;
@@ -171,6 +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;
@@ -184,6 +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/src/configuration.cpp

@@ -42,6 +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);
@@ -55,6 +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);
@@ -96,6 +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) {
@@ -112,6 +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 =

src/launcher/src/configurationEditDialog.cpp

@@ -298,6 +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);
@@ -311,6 +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);
@@ -533,6 +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();
@@ -541,6 +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();

src/launcher/src/mainDialog.cpp

@@ -276,6 +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");
@@ -291,6 +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

@@ -56,6 +56,12 @@ 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;
@@ -65,6 +71,7 @@ struct AjmDecodeResult {
 	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/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/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;
 }
 

src/libs/libNet.cpp

@@ -1685,12 +1685,34 @@ LIB_VERSION("NpAuth", 1, "NpAuth", 1, 1);
 constexpr int NP_AUTH_MAX_REQUEST_NUM         = 16;
 constexpr int NP_AUTH_REQUEST_ID_OFFSET       = 0x10000000;
 constexpr int NP_AUTH_ERROR_INVALID_ARGUMENT  = -2141912319; /* 0x80550301 */
+constexpr int NP_AUTH_ERROR_INVALID_SIZE      = -2141912318; /* 0x80550302 */
 constexpr int NP_AUTH_ERROR_ABORTED           = -2141912316; /* 0x80550304 */
 constexpr int NP_AUTH_ERROR_REQUEST_MAX       = -2141912315; /* 0x80550305 */
 constexpr int NP_AUTH_ERROR_REQUEST_NOT_FOUND = -2141912314; /* 0x80550306 */
 constexpr int NP_AUTH_ERROR_INVALID_ID        = -2141912313; /* 0x80550307 */
+constexpr int NP_AUTH_ERROR_SERVICE_DOWN      = -2141912063; /* 0x80550401 */
 constexpr int NP_ERROR_SIGNED_OUT             = -2141913082; /* 0x80550006 */
 
+struct NpAuthGetAuthorizationCodeParameterV3 {
+	size_t      size;
+	int         user_id;
+	uint8_t     padding[4];
+	const void* client_id;
+	const char* scope;
+};
+
+struct NpAuthGetIdTokenParameterV3 {
+	size_t      size;
+	int         user_id;
+	uint8_t     padding[4];
+	const void* client_id;
+	const void* client_secret;
+	const char* scope;
+};
+
+static_assert(sizeof(NpAuthGetAuthorizationCodeParameterV3) == 32);
+static_assert(sizeof(NpAuthGetIdTokenParameterV3) == 40);
+
 enum class NpAuthRequestState {
 	None = 0,
 	Ready,
@@ -1822,7 +1844,7 @@ static int np_auth_wait_or_poll_async(int req_id, int* result) {
 	return 0;
 }
 
-static int np_auth_complete_signed_out(NpAuthRequest* request) {
+static int np_auth_complete_unavailable(NpAuthRequest* request, int user_id) {
 	if (request->state == NpAuthRequestState::Complete) {
 		request->result = NP_AUTH_ERROR_INVALID_ARGUMENT;
 		return NP_AUTH_ERROR_INVALID_ARGUMENT;
@@ -1833,9 +1855,11 @@ static int np_auth_complete_signed_out(NpAuthRequest* request) {
 	}
 
 	request->state  = NpAuthRequestState::Complete;
-	request->result = NP_ERROR_SIGNED_OUT;
+	// Emulating sign-in cannot provide a real authorization code or ID token.
+	request->result = Network::NpManager::IsSignedIn(user_id) ? NP_AUTH_ERROR_SERVICE_DOWN
+	                                                        : NP_ERROR_SIGNED_OUT;
 
-	return request->async ? 0 : NP_ERROR_SIGNED_OUT;
+	return request->async ? 0 : request->result;
 }
 
 static int KYTY_SYSV_ABI NpAuthWaitAsync(int req_id, int* result) {
@@ -1856,8 +1880,9 @@ static int KYTY_SYSV_ABI NpAuthPollAsync(int req_id, int* result) {
 	return np_auth_wait_or_poll_async(req_id, result);
 }
 
-static int KYTY_SYSV_ABI NpAuthGetAuthorizationCodeV3(int req_id, const void* param,
-                                                      void* auth_code, int* issuer_id) {
+static int KYTY_SYSV_ABI NpAuthGetAuthorizationCodeV3(
+    int req_id, const NpAuthGetAuthorizationCodeParameterV3* param, void* auth_code,
+    int* issuer_id) {
 	PRINT_NAME();
 
 	LOGF("\t req_id    = %d\n", req_id);
@@ -1865,9 +1890,17 @@ static int KYTY_SYSV_ABI NpAuthGetAuthorizationCodeV3(int req_id, const void* pa
 	LOGF("\t auth_code = 0x%016" PRIx64 "\n", reinterpret_cast<uint64_t>(auth_code));
 	LOGF("\t issuer_id = 0x%016" PRIx64 "\n", reinterpret_cast<uint64_t>(issuer_id));
 
-	if (auth_code != nullptr) {
-		memset(auth_code, 0, 136);
+	if (req_id <= 0 || param == nullptr || auth_code == nullptr) {
+		return NP_AUTH_ERROR_INVALID_ARGUMENT;
 	}
+	if (param->size < sizeof(*param)) {
+		return NP_AUTH_ERROR_INVALID_SIZE;
+	}
+	if (param->client_id == nullptr || param->scope == nullptr) {
+		return NP_AUTH_ERROR_INVALID_ARGUMENT;
+	}
+
+	memset(auth_code, 0, 136);
 	if (issuer_id != nullptr) {
 		*issuer_id = 0;
 	}
@@ -1877,46 +1910,35 @@ static int KYTY_SYSV_ABI NpAuthGetAuthorizationCodeV3(int req_id, const void* pa
 	if (request == nullptr) {
 		return NP_AUTH_ERROR_REQUEST_NOT_FOUND;
 	}
-	if (request->state == NpAuthRequestState::Aborted) {
-		// request->result = 0;
-		// return 0;
-		request->result = NP_AUTH_ERROR_ABORTED;
-		return NP_AUTH_ERROR_ABORTED;
-	}
-	// request->state  = NpAuthRequestState::Complete;
-	// request->result = 0;
-
-	// return 0;
-	return np_auth_complete_signed_out(request);
+	return np_auth_complete_unavailable(request, param->user_id);
 }
 
-static int KYTY_SYSV_ABI NpAuthGetIdTokenV3(int req_id, const void* param, void* id_token) {
+static int KYTY_SYSV_ABI NpAuthGetIdTokenV3(int req_id, const NpAuthGetIdTokenParameterV3* param,
+                                        void* id_token) {
 	PRINT_NAME();
 
 	LOGF("\t req_id   = %d\n", req_id);
 	LOGF("\t param    = 0x%016" PRIx64 "\n", reinterpret_cast<uint64_t>(param));
 	LOGF("\t id_token = 0x%016" PRIx64 "\n", reinterpret_cast<uint64_t>(id_token));
 
-	if (id_token != nullptr) {
-		memset(id_token, 0, 4104);
+	if (req_id <= 0 || param == nullptr || id_token == nullptr) {
+		return NP_AUTH_ERROR_INVALID_ARGUMENT;
 	}
+	if (param->size < sizeof(*param)) {
+		return NP_AUTH_ERROR_INVALID_SIZE;
+	}
+	if (param->client_id == nullptr || param->client_secret == nullptr || param->scope == nullptr) {
+		return NP_AUTH_ERROR_INVALID_ARGUMENT;
+	}
+
+	memset(id_token, 0, 4104);
 
 	std::scoped_lock lock(g_np_auth_request_mutex);
 	auto*            request = np_auth_find_request(req_id);
 	if (request == nullptr) {
 		return NP_AUTH_ERROR_REQUEST_NOT_FOUND;
 	}
-	if (request->state == NpAuthRequestState::Aborted) {
-		// request->result = 0;
-		// return 0;
-		request->result = NP_AUTH_ERROR_ABORTED;
-		return NP_AUTH_ERROR_ABORTED;
-	}
-	// request->state  = NpAuthRequestState::Complete;
-	// request->result = 0;
-
-	// return 0;
-	return np_auth_complete_signed_out(request);
+	return np_auth_complete_unavailable(request, param->user_id);
 }
 
 LIB_DEFINE(InitNet_1_NpAuth) {

src/libs/libPsml.cpp

@@ -1,349 +1,376 @@
+#include "graphics/guest_gpu/pm4.h"
+#include "graphics/guest_gpu/pssr.h"
+#include "graphics/shader/shaderBindings.h"
 #include "libs/errno.h"
 #include "libs/libs.h"
 #include "loader/symbolDatabase.h"
+#include "loader/systemContent.h"
 
-#include <atomic>
+#include <algorithm>
+#include <cmath>
+#include <cstddef>
 #include <cstdint>
+#include <cstring>
+#include <memory>
+#include <mutex>
+#include <unordered_map>
 
-namespace Libs {
-
-namespace LibPsml {
+namespace Libs::LibPsml {
 
 LIB_VERSION("Psml", 1, "Psml", 1, 1);
 
-namespace Psml {
+namespace {
 
-constexpr int PSML_ERROR_NOT_INITIALIZED = static_cast<int>(0x8a810001u);
-constexpr int PSML_ERROR_INVALID_OBJECT  = static_cast<int>(0x8a810005u);
-constexpr int PSML_ERROR_INVALID_POINTER = static_cast<int>(0x8a810009u);
-constexpr int PSML_ERROR_INVALID_VALUE   = static_cast<int>(0x8a81000du);
-constexpr int PSML_ERROR_NULL_OBJECT     = static_cast<int>(0x8a810014u);
+constexpr int      NOT_INITIALIZED           = static_cast<int>(0x8a810001u);
+constexpr int      ALREADY_INITIALIZED       = static_cast<int>(0x8a810002u);
+constexpr int      TOO_MANY_CONTEXTS         = static_cast<int>(0x8a810004u);
+constexpr int      INVALID_CONTEXT           = static_cast<int>(0x8a810005u);
+constexpr int      COMMAND_BUFFER_FULL       = static_cast<int>(0x8a810006u);
+constexpr int      COMMAND_BUFFER_BAD        = static_cast<int>(0x8a810007u);
+constexpr int      INSUFFICIENT_BUFFER       = static_cast<int>(0x8a810008u);
+constexpr int      ARGUMENT_NULL             = static_cast<int>(0x8a810009u);
+constexpr int      TEXTURE_NULL              = static_cast<int>(0x8a81000au);
+constexpr int      INVALID_TEXTURE           = static_cast<int>(0x8a81000bu);
+constexpr int      INVALID_PARAMETER         = static_cast<int>(0x8a81000du);
+constexpr int      UNSUPPORTED_RENDER_SIZE   = static_cast<int>(0x8a81000eu);
+constexpr int      INVALID_OUTPUT_TEXTURE    = static_cast<int>(0x8a81000fu);
+constexpr int      NOT_TRINITY_MODE          = static_cast<int>(0x8a810012u);
+constexpr int      UNSUPPORTED_OUTPUT_SIZE   = static_cast<int>(0x8a810013u);
+constexpr int      INVALID_SHARED_RESOURCES  = static_cast<int>(0x8a810014u);
+constexpr int      TOO_MANY_SHARED_RESOURCES = static_cast<int>(0x8a810015u);
+constexpr uint64_t MEMORY_BLOCK_SIZE         = 0x200000;
+constexpr uint32_t PACKET_DWORDS             = 1 + Graphics::Pssr::RequestDwords;
 
-constexpr uint32_t SHARED_RESOURCES_MAGIC = 0xa9c4;
-constexpr uint32_t CONTEXT_MAGIC          = 0x9231;
-
-constexpr uint64_t MAIN_MEMORY_BLOCK_SIZE = 0x200000;
-constexpr uint64_t MAIN_MEMORY_ALIGNMENT  = 0x200000;
-constexpr uint64_t EXTRA_VA_BYTES         = 0x600000;
-
-struct MainMemoryRequirements {
+struct MemoryRequirements {
 	uint64_t block_size;
 	uint64_t alignment;
 	uint64_t block_count;
 };
 
-struct MainMemoryParameters {
-	uint32_t type;
-	uint32_t reserved;
-};
-
 struct DirectMemoryBlock {
-	uint64_t address;
+	uint64_t offset;
 	uint64_t size;
 };
 
-struct SharedResourcesInitParameters {
-	uint32_t                 type;
-	uint32_t                 reserved;
+struct MemoryParameters {
+	uint32_t                 profile;
+	uint32_t                 flags;
 	const DirectMemoryBlock* blocks;
 	uint64_t                 block_count;
-	uint64_t                 virtual_address_start;
 };
 
-static_assert(sizeof(MainMemoryRequirements) == 24);
-static_assert(sizeof(DirectMemoryBlock) == 16);
+struct SharedParameters {
+	MemoryParameters memory;
+	uint64_t         virtual_address;
+};
 
-static bool                 g_initialized      = false;
-static uint32_t             g_supported_modes  = 0x3;
-static std::atomic<int32_t> g_shared_ref_count = 0;
+struct ContextParameters {
+	MemoryParameters memory;
+	uint16_t         width;
+	uint16_t         height;
+	uint32_t         reserved;
+	const void*      depth_copy;
+	const void*      motion_copy;
+	const char*      name;
+	void*            shared;
+};
+static_assert(offsetof(ContextParameters, width) == 0x18);
+static_assert(offsetof(ContextParameters, shared) == 0x38);
 
-static uint64_t BaseVaBytes(uint32_t type) {
-	// The native implementation indexes this table before adding one 6 MiB range.
-	switch (type) {
-		case 0: return 0x06200000;
-		case 1: return 0x18200000;
-		case 2: return 0x12200000;
-		default: return 0;
-	}
+struct SharedResources {
+	uint32_t contexts = 0;
+};
+
+struct Context {
+	uint32_t         width;
+	uint32_t         height;
+	SharedResources* shared;
+};
+
+struct CommandSpace {
+	uint32_t* cmd;
+	uint32_t  dwords;
+};
+
+std::mutex                                                g_mutex;
+bool                                                      g_initialized = false;
+std::unique_ptr<SharedResources>                          g_shared;
+std::unordered_map<const void*, std::unique_ptr<Context>> g_contexts;
+
+uint64_t SharedBlockCount(uint32_t profile) {
+	constexpr uint64_t bytes[] = {0x06200000, 0x18200000, 0x12200000};
+	return profile < 3 ? (bytes[profile] + 0x600000) / MEMORY_BLOCK_SIZE : 0;
 }
 
-static bool IsSupportedType(uint32_t type) {
-	switch (type) {
-		case 0: return true;
-		case 1: return (g_supported_modes & 0x1u) != 0;
-		case 2: return (g_supported_modes & 0x2u) != 0;
-		default: return false;
+int CheckContext(const void* context) {
+	if (!g_initialized) {
+		return NOT_INITIALIZED;
 	}
+	return g_contexts.contains(context) ? OK : INVALID_CONTEXT;
 }
 
-static uint64_t RequiredBlockCount(uint32_t type) {
-	const auto base = BaseVaBytes(type);
-	if (base == 0) {
-		return 0;
+int KYTY_SYSV_ABI Initialize() {
+	std::lock_guard lock(g_mutex);
+	if (!Loader::SystemContentIsTrinityMode()) {
+		return NOT_TRINITY_MODE;
 	}
-	return (base + EXTRA_VA_BYTES) / MAIN_MEMORY_BLOCK_SIZE;
+	if (g_initialized) {
+		return ALREADY_INITIALIZED;
 	}
-
-static int CheckInitialized() {
-	return (g_initialized ? OK : PSML_ERROR_NOT_INITIALIZED);
-}
-
-static void MarkSharedResources(void* resources, const SharedResourcesInitParameters& params) {
-	auto* bytes = static_cast<uint8_t*>(resources);
-
-	*reinterpret_cast<uint32_t*>(bytes)                        = SHARED_RESOURCES_MAGIC;
-	*reinterpret_cast<const DirectMemoryBlock**>(bytes + 0x08) = params.blocks;
-	*reinterpret_cast<uint64_t*>(bytes + 0x18)                 = params.block_count;
-	*reinterpret_cast<uint32_t*>(bytes + 0x20)                 = params.type;
-	*reinterpret_cast<uint64_t*>(bytes + 0x28)                 = params.virtual_address_start;
-}
-
-static bool HasKnownObjectMagic(const void* object) {
-	if (object == nullptr) {
-		return false;
-	}
-	const auto magic = *static_cast<const uint32_t*>(object);
-	return magic == SHARED_RESOURCES_MAGIC || magic == CONTEXT_MAGIC;
-}
-
-static int KYTY_SYSV_ABI PsmlInitialize() {
-	PRINT_NAME();
-
 	g_initialized = true;
-	g_supported_modes  = 0x3;
-	g_shared_ref_count = 0;
-
 	return OK;
 }
 
-static int KYTY_SYSV_ABI PsmlGetMainMemoryRequirements(MainMemoryRequirements*     out,
-                                                       const MainMemoryParameters* params) {
-	PRINT_NAME();
-
-	const auto init_error = CheckInitialized();
-	if (init_error != OK) {
-		return init_error;
+int KYTY_SYSV_ABI GetSharedRequirements(MemoryRequirements* out, const MemoryParameters* params) {
+	std::lock_guard lock(g_mutex);
+	if (!g_initialized) {
+		return NOT_INITIALIZED;
 	}
 	if (out == nullptr || params == nullptr) {
-		return PSML_ERROR_INVALID_POINTER;
+		return ARGUMENT_NULL;
 	}
-	if (!IsSupportedType(params->type)) {
-		return PSML_ERROR_INVALID_VALUE;
+	const auto blocks = SharedBlockCount(params->profile);
+	if (blocks == 0) {
+		return INVALID_PARAMETER;
 	}
-
-	out->block_size  = MAIN_MEMORY_BLOCK_SIZE;
-	out->alignment   = MAIN_MEMORY_ALIGNMENT;
-	out->block_count = RequiredBlockCount(params->type);
-
+	*out = {MEMORY_BLOCK_SIZE, MEMORY_BLOCK_SIZE, blocks};
 	return OK;
 }
 
-static int KYTY_SYSV_ABI
-PsmlSharedResourcesInitialize(void* resources, const SharedResourcesInitParameters* params) {
-	PRINT_NAME();
-
-	const auto init_error = CheckInitialized();
-	if (init_error != OK) {
-		return init_error;
+int KYTY_SYSV_ABI CreateSharedResources(void** out, const SharedParameters* params) {
+	std::lock_guard lock(g_mutex);
+	if (!g_initialized) {
+		return NOT_INITIALIZED;
 	}
-	if (resources == nullptr || params == nullptr || params->blocks == nullptr) {
-		return PSML_ERROR_INVALID_POINTER;
+	if (out == nullptr || params == nullptr || params->memory.blocks == nullptr) {
+		return ARGUMENT_NULL;
 	}
-	if (params->reserved != 0 || !IsSupportedType(params->type) ||
-	    params->block_count < RequiredBlockCount(params->type)) {
-		return PSML_ERROR_INVALID_VALUE;
+	const auto required = SharedBlockCount(params->memory.profile);
+	if (required == 0 || params->memory.flags != 0) {
+		return INVALID_PARAMETER;
 	}
-
-	MarkSharedResources(resources, *params);
-	g_shared_ref_count.fetch_add(1, std::memory_order_relaxed);
-
+	if (params->memory.block_count < required) {
+		return INSUFFICIENT_BUFFER;
+	}
+	if (g_shared) {
+		return TOO_MANY_SHARED_RESOURCES;
+	}
+	g_shared = std::make_unique<SharedResources>();
+	*out     = g_shared.get();
 	return OK;
 }
 
-static int KYTY_SYSV_ABI PsmlSharedResourcesFinalize(void* resources) {
-	PRINT_NAME();
-
-	const auto init_error = CheckInitialized();
-	if (init_error != OK) {
-		return init_error;
+int KYTY_SYSV_ABI ReleaseSharedResources(void* resources) {
+	std::lock_guard lock(g_mutex);
+	if (!g_initialized) {
+		return NOT_INITIALIZED;
 	}
-	if (resources == nullptr) {
-		return PSML_ERROR_NULL_OBJECT;
+	if (!g_shared || resources != g_shared.get()) {
+		return INVALID_SHARED_RESOURCES;
 	}
-
-	g_shared_ref_count.fetch_sub(1, std::memory_order_relaxed);
-
+	if (g_shared->contexts != 0) {
+		return INVALID_CONTEXT;
+	}
+	g_shared.reset();
 	return OK;
 }
 
-static int KYTY_SYSV_ABI PsmlGetContextMemoryRequirements(MainMemoryRequirements* out,
-                                                          uint64_t /*params*/) {
-	PRINT_NAME();
-
-	const auto init_error = CheckInitialized();
-	if (init_error != OK) {
-		return init_error;
+int KYTY_SYSV_ABI GetContextRequirements(MemoryRequirements* out, const void* params) {
+	std::lock_guard lock(g_mutex);
+	if (!g_initialized) {
+		return NOT_INITIALIZED;
 	}
-	if (out == nullptr) {
-		return PSML_ERROR_INVALID_POINTER;
+	if (out == nullptr || params == nullptr) {
+		return ARGUMENT_NULL;
 	}
-
-	out->block_size  = MAIN_MEMORY_BLOCK_SIZE;
-	out->alignment   = MAIN_MEMORY_ALIGNMENT;
-	out->block_count = 1;
-
+	*out = {MEMORY_BLOCK_SIZE, MEMORY_BLOCK_SIZE, 1};
 	return OK;
 }
 
-static int KYTY_SYSV_ABI PsmlContextInitialize(void* context, const void* params) {
-	PRINT_NAME();
-
-	const auto init_error = CheckInitialized();
-	if (init_error != OK) {
-		return init_error;
+int CreateContext(void** out, const MemoryParameters* memory, uint32_t width, uint32_t height,
+                  void* shared, bool use_shared) {
+	std::lock_guard lock(g_mutex);
+	if (!g_initialized) {
+		return NOT_INITIALIZED;
 	}
-	if (context == nullptr || params == nullptr) {
-		return PSML_ERROR_INVALID_POINTER;
+	if (out == nullptr || memory == nullptr || memory->blocks == nullptr) {
+		return ARGUMENT_NULL;
 	}
-
-	auto* bytes            = static_cast<uint8_t*>(context);
-	auto* params_bytes     = static_cast<const uint8_t*>(params);
-	auto* shared_resources = *reinterpret_cast<void* const*>(params_bytes + 0x08);
-	if (shared_resources == nullptr) {
-		return PSML_ERROR_INVALID_POINTER;
+	if (memory->profile > 2) {
+		return INVALID_PARAMETER;
 	}
-
-	*reinterpret_cast<uint32_t*>(bytes)        = CONTEXT_MAGIC;
-	*reinterpret_cast<void**>(bytes + 0x360)   = shared_resources;
-	*reinterpret_cast<uint8_t*>(bytes + 0x368) = 0;
-
+	if (memory->block_count == 0) {
+		return INSUFFICIENT_BUFFER;
+	}
+	if (width == 0 || height == 0 || width > 16384 || height > 16384) {
+		return UNSUPPORTED_OUTPUT_SIZE;
+	}
+	if (use_shared && (!g_shared || shared != g_shared.get())) {
+		return INVALID_SHARED_RESOURCES;
+	}
+	if (g_contexts.size() == 4) {
+		return TOO_MANY_CONTEXTS;
+	}
+	auto record =
+	    std::make_unique<Context>(Context {width, height, use_shared ? g_shared.get() : nullptr});
+	auto* handle = record.get();
+	g_contexts.emplace(handle, std::move(record));
+	if (handle->shared != nullptr) {
+		++handle->shared->contexts;
+	}
+	*out = handle;
 	return OK;
 }
 
-static int KYTY_SYSV_ABI PsmlContextFinalize(void* context) {
-	PRINT_NAME();
-
-	const auto init_error = CheckInitialized();
-	if (init_error != OK) {
-		return init_error;
-	}
-	if (!HasKnownObjectMagic(context)) {
-		return PSML_ERROR_INVALID_OBJECT;
+int KYTY_SYSV_ABI CreateFixedContext(void** out, const MemoryParameters* params) {
+	const bool large = params != nullptr && params->profile == 1;
+	return CreateContext(out, params, large ? 7680 : 3840, large ? 4320 : 2160, nullptr, false);
 }
 
-	*static_cast<uint32_t*>(context) = 0;
+int KYTY_SYSV_ABI CreatePrivateContext(void** out, const ContextParameters* params) {
+	return CreateContext(out, params == nullptr ? nullptr : &params->memory,
+	                     params == nullptr ? 0 : params->width,
+	                     params == nullptr ? 0 : params->height, nullptr, false);
+}
 
+int KYTY_SYSV_ABI CreateSharedContext(void** out, const ContextParameters* params) {
+	return CreateContext(
+	    out, params == nullptr ? nullptr : &params->memory, params == nullptr ? 0 : params->width,
+	    params == nullptr ? 0 : params->height, params == nullptr ? nullptr : params->shared, true);
+}
+
+int KYTY_SYSV_ABI ReleaseContext(void* context) {
+	std::lock_guard lock(g_mutex);
+	if (const int error = CheckContext(context); error != OK) {
+		return error;
+	}
+	auto& record = g_contexts.at(context);
+	if (record->shared != nullptr) {
+		--record->shared->contexts;
+	}
+	g_contexts.erase(context);
 	return OK;
 }
 
-static int KYTY_SYSV_ABI PsmlGetWorkAreaSize(const void* object, uint32_t* out_size) {
-	PRINT_NAME();
-
-	const auto init_error = CheckInitialized();
-	if (init_error != OK) {
-		return init_error;
-	}
-	if (!HasKnownObjectMagic(object)) {
-		return PSML_ERROR_INVALID_OBJECT;
+int KYTY_SYSV_ABI GetPacketSize(const void* context, uint32_t* out_size) {
+	std::lock_guard lock(g_mutex);
+	if (const int error = CheckContext(context); error != OK) {
+		return error;
 	}
 	if (out_size == nullptr) {
-		return PSML_ERROR_INVALID_POINTER;
+		return ARGUMENT_NULL;
 	}
-
-	*out_size = 0x600;
-
+	*out_size = PACKET_DWORDS;
 	return OK;
 }
 
-static int KYTY_SYSV_ABI PsmlDispatch(void* context, uint64_t command, uint64_t params) {
-	PRINT_NAME();
-
-	const auto init_error = CheckInitialized();
-	if (init_error != OK) {
-		return init_error;
+int KYTY_SYSV_ABI Dispatch(const void* context, const CommandSpace* space, const void* params) {
+	std::lock_guard lock(g_mutex);
+	if (const int error = CheckContext(context); error != OK) {
+		return error;
 	}
-	if (!HasKnownObjectMagic(context)) {
-		return PSML_ERROR_INVALID_OBJECT;
+	if (space == nullptr || params == nullptr || space->cmd == nullptr) {
+		return ARGUMENT_NULL;
 	}
-	if (command == 0 || params == 0) {
-		return PSML_ERROR_INVALID_POINTER;
+	if (space->dwords < PACKET_DWORDS) {
+		return COMMAND_BUFFER_FULL;
 	}
-
+	if (space->dwords > 0x4001) {
+		return COMMAND_BUFFER_BAD;
+	}
+	const uint32_t* descriptors[2];
+	std::memcpy(descriptors, params, sizeof(descriptors));
+	if (descriptors[0] == nullptr || descriptors[1] == nullptr) {
+		return TEXTURE_NULL;
+	}
+	Graphics::Pssr::Request request {};
+	std::memcpy(request.output, descriptors[0], sizeof(request.output));
+	std::memcpy(request.input, descriptors[1], sizeof(request.input));
+	std::memcpy(&request.width, static_cast<const uint8_t*>(params) + 0xf0, sizeof(uint32_t) * 2);
+	Graphics::ShaderTextureResource input, output;
+	std::memcpy(input.fields, request.input, sizeof(input.fields));
+	std::memcpy(output.fields, request.output, sizeof(output.fields));
+	namespace Gpu    = Graphics::Prospero;
+	const auto is_2d = [](const Graphics::ShaderTextureResource& descriptor) {
+		return !descriptor.IsNull() &&
+		       (descriptor.Type() == Gpu::ImageType::kColor2D ||
+		        descriptor.Type() == Gpu::ImageType::kColor2DArray) &&
+		       descriptor.TileMode() != Gpu::TileMode::kDepth &&
+		       descriptor.BaseLevel() <= descriptor.LastLevel() &&
+		       descriptor.LastLevel() <= descriptor.MaxMip() &&
+		       descriptor.BaseArray5() <= descriptor.Depth();
+	};
+	if (!is_2d(input) || (input.Format() != Gpu::BufferFormat::k11_11_10Float &&
+	                      input.Format() != Gpu::BufferFormat::k10_11_11Float &&
+	                      input.Format() != Gpu::BufferFormat::k16_16_16_16Float)) {
+		return INVALID_TEXTURE;
+	}
+	if (!is_2d(output) || output.Format() != Gpu::BufferFormat::k11_11_10Float ||
+	    output.TileMode() != Gpu::TileMode::kStandard256B || output.BaseLevel() != 0 ||
+	    output.LastLevel() != 0 || output.MaxMip() != 0 || output.BaseArray5() != 0 ||
+	    output.Depth() != 0 || input.Base40() == output.Base40()) {
+		return INVALID_OUTPUT_TEXTURE;
+	}
+	const uint32_t input_width  = std::max(1u, (input.Width5() + 1u) >> input.BaseLevel());
+	const uint32_t input_height = std::max(1u, (input.Height5() + 1u) >> input.BaseLevel());
+	if (request.width == 0 || request.height == 0 || request.width > input_width ||
+	    request.height > input_height) {
+		return UNSUPPORTED_RENDER_SIZE;
+	}
+	const auto& record = *g_contexts.at(context);
+	if (uint32_t(output.Width5()) + 1 != record.width ||
+	    uint32_t(output.Height5()) + 1 != record.height) {
+		return INVALID_OUTPUT_TEXTURE;
+	}
+	namespace Pm4 = Graphics::Pm4;
+	space->cmd[0] = KYTY_PM4(space->dwords, Pm4::IT_NOP, Pm4::R_PSSR);
+	std::memcpy(space->cmd + 1, &request, sizeof(request));
+	std::memset(space->cmd + PACKET_DWORDS, 0, (space->dwords - PACKET_DWORDS) * sizeof(uint32_t));
 	return OK;
 }
 
-static int KYTY_SYSV_ABI PsmlGetProgress(const void* object, float* out_progress,
-                                         uint32_t /*frame_index*/) {
-	PRINT_NAME();
-
-	const auto init_error = CheckInitialized();
-	if (init_error != OK) {
-		return init_error;
+int KYTY_SYSV_ABI GetMipmapBias(const void* context, float* bias, uint32_t width, uint32_t height) {
+	std::lock_guard lock(g_mutex);
+	if (const int error = CheckContext(context); error != OK) {
+		return error;
 	}
-	if (!HasKnownObjectMagic(object)) {
-		return PSML_ERROR_INVALID_OBJECT;
+	if (bias == nullptr) {
+		return ARGUMENT_NULL;
 	}
-	if (out_progress == nullptr) {
-		return PSML_ERROR_INVALID_POINTER;
+	if (width == 0 || height == 0) {
+		return UNSUPPORTED_RENDER_SIZE;
 	}
-
-	*out_progress = 0.0f;
-
+	*bias = std::log2(float(width) / float(g_contexts.at(context)->width)) - 1.0f;
 	return OK;
 }
 
-static int KYTY_SYSV_ABI PsmlRequestCapture(const void* object) {
-	PRINT_NAME();
-
-	const auto init_error = CheckInitialized();
-	if (init_error != OK) {
-		return init_error;
-	}
-	if (!HasKnownObjectMagic(object)) {
-		return PSML_ERROR_INVALID_OBJECT;
+int KYTY_SYSV_ABI ValidateContext(const void* context) {
+	std::lock_guard lock(g_mutex);
+	return CheckContext(context);
 }
 
-	return OK;
-}
-
-static int KYTY_SYSV_ABI PsmlValidateObject(const void* object) {
-	PRINT_NAME();
-
-	const auto init_error = CheckInitialized();
-	if (init_error != OK) {
-		return init_error;
-	}
-	if (!HasKnownObjectMagic(object)) {
-		return PSML_ERROR_INVALID_OBJECT;
-	}
-
-	return OK;
-}
-
-} // namespace Psml
+} // namespace
 
 LIB_DEFINE(InitPsml_1) {
-	LIB_FUNC("3WVD91e12ZQ", Psml::PsmlInitialize);
-	LIB_FUNC("+2KpvixvL6E", Psml::PsmlGetMainMemoryRequirements);
-	LIB_FUNC("eWoKNeB6V-k", Psml::PsmlSharedResourcesInitialize);
-	LIB_FUNC("jEevBXmagOQ", Psml::PsmlSharedResourcesFinalize);
-	LIB_FUNC("EO9YQXmJEN8", Psml::PsmlGetContextMemoryRequirements);
-	LIB_FUNC("IMlj247LdTo", Psml::PsmlGetContextMemoryRequirements);
-	LIB_FUNC("sjTyYheKTrU", Psml::PsmlGetContextMemoryRequirements);
-	LIB_FUNC("ArakEpzsZo0", Psml::PsmlGetContextMemoryRequirements);
-	LIB_FUNC("VGjrQa-WqdU", Psml::PsmlGetContextMemoryRequirements);
-	LIB_FUNC("2ecEbQaf9VU", Psml::PsmlContextInitialize);
-	LIB_FUNC("vUk2pWMx3KQ", Psml::PsmlContextInitialize);
-	LIB_FUNC("gxv3i+MTEzU", Psml::PsmlContextInitialize);
-	LIB_FUNC("JaLBe0P3jSU", Psml::PsmlContextFinalize);
-	LIB_FUNC("AHalTX9wFZY", Psml::PsmlGetWorkAreaSize);
-	LIB_FUNC("RUNLFro+qok", Psml::PsmlDispatch);
-	LIB_FUNC("GHna9-DvnUk", Psml::PsmlGetProgress);
-	LIB_FUNC("GJY0MvuTcs8", Psml::PsmlRequestCapture);
-	LIB_FUNC("LXq+6mIxpCw", Psml::PsmlValidateObject);
-	LIB_FUNC("FSGaTQze0UY", Psml::PsmlValidateObject);
+	LIB_FUNC("3WVD91e12ZQ", Initialize);
+	LIB_FUNC("+2KpvixvL6E", GetSharedRequirements);
+	LIB_FUNC("eWoKNeB6V-k", CreateSharedResources);
+	LIB_FUNC("jEevBXmagOQ", ReleaseSharedResources);
+	LIB_FUNC("EO9YQXmJEN8", GetContextRequirements);
+	LIB_FUNC("IMlj247LdTo", GetContextRequirements);
+	LIB_FUNC("sjTyYheKTrU", GetContextRequirements);
+	LIB_FUNC("ArakEpzsZo0", GetContextRequirements);
+	LIB_FUNC("VGjrQa-WqdU", GetContextRequirements);
+	LIB_FUNC("2ecEbQaf9VU", CreateFixedContext);
+	LIB_FUNC("vUk2pWMx3KQ", CreatePrivateContext);
+	LIB_FUNC("gxv3i+MTEzU", CreateSharedContext);
+	LIB_FUNC("JaLBe0P3jSU", ReleaseContext);
+	LIB_FUNC("AHalTX9wFZY", GetPacketSize);
+	LIB_FUNC("RUNLFro+qok", Dispatch);
+	LIB_FUNC("GHna9-DvnUk", GetMipmapBias);
+	LIB_FUNC("GJY0MvuTcs8", ValidateContext);
+	LIB_FUNC("LXq+6mIxpCw", ValidateContext);
+	LIB_FUNC("FSGaTQze0UY", ValidateContext);
 }
 
-} // namespace LibPsml
-
-} // namespace Libs
+} // namespace Libs::LibPsml

src/libs/network.cpp

@@ -4002,6 +4002,11 @@ struct NpCountryCode {
 	char padding[1];
 };
 
+struct NpLanguageCode {
+	char code[36];
+	char padding[12];
+};
+
 struct NpAgeRestriction {
 	NpCountryCode country_code;
 	int8_t        age;
@@ -4056,8 +4061,23 @@ constexpr int      np_error_aborted                     = -2141913070; /* 0x8055
 constexpr int      np_error_request_max                 = -2141913069; /* 0x80550013 */
 constexpr int      np_error_request_not_found           = -2141913068; /* 0x80550014 */
 constexpr int      np_error_invalid_id                  = -2141913067; /* 0x80550015 */
+constexpr int      np_error_network_not_available       = -2141872086; /* 0x8055a02a */
 constexpr int      np_request_max                       = 128;
 constexpr uint32_t np_state_signed_out                  = 1;
+constexpr uint32_t np_state_signed_in                   = 2;
+
+// A simulated account belongs only to the primary user and survives network loss.
+bool IsSignedIn(int user_id) {
+	return Config::PsnEmulationEnabled() && user_id == Config::GetUserId();
+}
+
+static int np_reachability_result(int user_id) {
+	if (!IsSignedIn(user_id)) {
+		return np_error_signed_out;
+	}
+	// This is an explicit compatibility simulation, not a connection to PSN.
+	return NetCtl::NetCtlConnected() ? OK : np_error_network_not_available;
+}
 
 struct NpStateCallback {
 	NpStateCallbackA callback = nullptr;
@@ -4117,7 +4137,7 @@ static NpRequest* np_get_request_locked(int req_id) {
 	return &g_np_requests[index];
 }
 
-static int np_complete_signed_out_locked(NpRequest* request) {
+static int np_complete_locked(NpRequest* request, int result) {
 	if (request->state == NpRequestState::Complete) {
 		request->result = np_error_invalid_argument;
 		return np_error_invalid_argument;
@@ -4128,9 +4148,9 @@ static int np_complete_signed_out_locked(NpRequest* request) {
 	}
 
 	request->state  = NpRequestState::Complete;
-	request->result = np_error_signed_out;
+	request->result = result;
 
-	return (request->async ? OK : np_error_signed_out);
+	return (request->async ? OK : result);
 }
 
 int KYTY_SYSV_ABI NpCheckCallback() {
@@ -4144,7 +4164,9 @@ int KYTY_SYSV_ABI NpCheckCallback() {
 		}
 	}
 	for (const auto& entry: pending) {
-		entry.callback(Config::GetUserId(), np_state_signed_out, entry.userdata);
+		entry.callback(Config::GetUserId(),
+		               IsSignedIn(Config::GetUserId()) ? np_state_signed_in : np_state_signed_out,
+		               entry.userdata);
 	}
 
 	return OK;
@@ -4239,12 +4261,13 @@ int KYTY_SYSV_ABI NpGetNpId(int user_id, NpId* np_id) {
 
 	EXIT_NOT_IMPLEMENTED(np_id == nullptr);
 
-	// int s = snprintf(np_id->handle.data, 16, "Kyty");
-	// EXIT_NOT_IMPLEMENTED(s >= 16);
-	// np_id->handle.term = 0;
 	std::memset(np_id, 0, sizeof(*np_id));
 
-	// return OK;
+	if (IsSignedIn(user_id)) {
+		std::memcpy(np_id->handle.data, "Kyty", 4);
+		return OK;
+	}
+
 	return np_error_signed_out;
 }
 
@@ -4255,12 +4278,13 @@ int KYTY_SYSV_ABI NpGetOnlineId(int user_id, NpOnlineId* online_id) {
 
 	EXIT_NOT_IMPLEMENTED(online_id == nullptr);
 
-	// int s = snprintf(online_id->data, 16, "Kyty");
-	// EXIT_NOT_IMPLEMENTED(s >= 16);
-	// online_id->term = 0;
 	std::memset(online_id, 0, sizeof(*online_id));
 
-	// return OK;
+	if (IsSignedIn(user_id)) {
+		std::memcpy(online_id->data, "Kyty", 4);
+		return OK;
+	}
+
 	return np_error_signed_out;
 }
 
@@ -4271,10 +4295,13 @@ int KYTY_SYSV_ABI NpGetAccountIdA(int user_id, uint64_t* account_id) {
 
 	EXIT_NOT_IMPLEMENTED(account_id == nullptr);
 
-	// *account_id = 0x00000000feedfaceull;
 	*account_id = 0;
 
-	// return OK;
+	if (IsSignedIn(user_id)) {
+		*account_id = 0x00000000feedfaceull;
+		return OK;
+	}
+
 	return np_error_signed_out;
 }
 
@@ -4287,10 +4314,12 @@ int KYTY_SYSV_ABI NpGetAccountCountryA(int user_id, void* country_code) {
 
 	auto* code = static_cast<NpCountryCode*>(country_code);
 	std::memset(code, 0, sizeof(*code));
-	// code->data[0] = '';
-	// code->data[1] = 'S';
 
-	// return OK;
+	if (IsSignedIn(user_id)) {
+		std::memcpy(code->data, "US", 2);
+		return OK;
+	}
+
 	return np_error_signed_out;
 }
 
@@ -4305,11 +4334,14 @@ int KYTY_SYSV_ABI NpGetAccountAge(int req_id, int user_id, uint8_t* age) {
 		return np_error_invalid_argument;
 	}
 
-	// *age = 21;
-	*age = 0;
+	std::lock_guard lock(g_np_request_mutex);
+	auto* request = np_get_request_locked(req_id);
+	if (request == nullptr) {
+		return np_error_request_not_found;
+	}
 
-	// return OK;
-	return np_error_signed_out;
+	*age = IsSignedIn(user_id) ? 21 : 0;
+	return np_complete_locked(request, IsSignedIn(user_id) ? OK : np_error_signed_out);
 }
 
 int KYTY_SYSV_ABI NpGetAccountLanguage2(int req_id, int user_id, void* language_code) {
@@ -4325,11 +4357,17 @@ int KYTY_SYSV_ABI NpGetAccountLanguage2(int req_id, int user_id, void* language_
 
 	std::lock_guard lock(g_np_request_mutex);
 
-	if (np_get_request_locked(req_id) == nullptr) {
+	auto* request = np_get_request_locked(req_id);
+	if (request == nullptr) {
 		return np_error_request_not_found;
 	}
 
-	return np_error_signed_out;
+	auto* code = static_cast<NpLanguageCode*>(language_code);
+	std::memset(code, 0, sizeof(*code));
+	if (IsSignedIn(user_id)) {
+		std::memcpy(code->code, "en-US", 5);
+	}
+	return np_complete_locked(request, IsSignedIn(user_id) ? OK : np_error_signed_out);
 }
 
 int KYTY_SYSV_ABI NpCreateRequest() {
@@ -4428,8 +4466,9 @@ int KYTY_SYSV_ABI NpCheckNpAvailability(int req_id, const char* user, void* resu
 		return np_error_request_not_found;
 	}
 
-	// return OK;
-	return np_complete_signed_out_locked(request);
+	return np_complete_locked(request, std::strcmp(user, "Kyty") == 0
+	                                       ? np_reachability_result(Config::GetUserId())
+	                                       : np_error_signed_out);
 }
 
 int KYTY_SYSV_ABI NpCheckNpReachability(int req_id, int user_id) {
@@ -4451,8 +4490,7 @@ int KYTY_SYSV_ABI NpCheckNpReachability(int req_id, int user_id) {
 		return np_error_request_not_found;
 	}
 
-	// return OK;
-	return np_complete_signed_out_locked(request);
+	return np_complete_locked(request, np_reachability_result(user_id));
 }
 
 int KYTY_SYSV_ABI NpPollAsync(int req_id, int* result) {
@@ -4509,23 +4547,10 @@ int KYTY_SYSV_ABI NpCheckPremium(int req_id, const NpCheckPremiumParameter* para
 		return np_error_request_not_found;
 	}
 
-	if (request->state == NpRequestState::Complete) {
-		request->result = np_error_invalid_argument;
-		return np_error_invalid_argument;
-	}
-
-	if (request->state == NpRequestState::Aborted) {
-		return np_error_aborted;
-	}
-
 	std::memset(result, 0, sizeof(*result));
-	// result->authorized = true;
 
-	// request->state  = NpRequestState::Complete;
-	// request->result = OK;
-
-	// return OK;
-	return np_complete_signed_out_locked(request);
+	// The simulated account has no premium eligibility.
+	return np_complete_locked(request, np_reachability_result(param->user_id));
 }
 
 int KYTY_SYSV_ABI NpGetState(int user_id, uint32_t* state) {
@@ -4535,7 +4560,7 @@ int KYTY_SYSV_ABI NpGetState(int user_id, uint32_t* state) {
 
 	LOGF("\t user_id = %d\n", user_id);
 
-	*state = np_state_signed_out;
+	*state = IsSignedIn(user_id) ? np_state_signed_in : np_state_signed_out;
 
 	return OK;
 }
@@ -4549,8 +4574,7 @@ int KYTY_SYSV_ABI NpGetNpReachabilityState(int user_id, uint32_t* state) {
 
 	LOGF("\t user_id = %d\n", user_id);
 
-	// *state = 2; // SCE_NP_REACHABILITY_STATE_REACHABLE
-	*state = 0; // SCE_NP_REACHABILITY_STATE_UNAVAILABLE
+	*state = np_reachability_result(user_id) == OK ? 2 : 0; // REACHABLE / UNAVAILABLE
 
 	return OK;
 }
@@ -4564,7 +4588,7 @@ int KYTY_SYSV_ABI NpHasSignedUp(int user_id, bool* has_signed_up) {
 
 	LOGF("\t user_id = %d\n", user_id);
 
-	*has_signed_up = false;
+	*has_signed_up = IsSignedIn(user_id);
 
 	return OK;
 }

src/libs/network.h

@@ -182,6 +182,8 @@ struct NpCheckPremiumResult;
 
 using NpStateCallbackA = KYTY_SYSV_ABI void (*)(int user_id, uint32_t state, void* userdata);
 
+bool IsSignedIn(int user_id);
+
 int KYTY_SYSV_ABI  NpCheckCallback();
 int KYTY_SYSV_ABI  NpSetNpTitleId(const NpTitleId* title_id, const NpTitleSecret* title_secret);
 int KYTY_SYSV_ABI  NpSetContentRestriction(const NpContentRestriction* restriction);

src/libs/ngs2.cpp

@@ -546,6 +546,7 @@ struct Ngs2VoiceInternal {
 	};
 	struct Module {
 		std::vector<uint8_t>     param, work, state;
+		std::vector<float>       channel_matrix;
 		void                     KYTY_SYSV_ABI (*handler)(Ngs2UserFxProcessContext*) = nullptr;
 		std::array<uintptr_t, 3> user_data {};
 		uint32_t                 flags          = 1;
@@ -563,9 +564,13 @@ struct Ngs2VoiceInternal {
 	std::vector<std::vector<float>> matrices;
 	std::vector<Module>             modules;
 	std::vector<float>              samples;
+	std::vector<float>              mix_samples;
 	std::vector<float>              output_matrix;
 	uint32_t                        channels        = 0;
 	uint32_t                        output_id       = 0;
+	float                           fbw_level       = 1.0f;
+	float                           lfe_level       = 1.0f;
+	float                           dry_level       = 0.0f;
 	bool                            rendering       = false;
 	bool                            rendered        = false;
 	bool                            has_samples     = false;
@@ -614,6 +619,27 @@ struct Ngs2VoiceInternal {
 		return samples.data() + size_t(id) * grain * channels;
 	}
 
+	void MixChannels(const std::vector<float>& matrix, uint32_t buffer, uint32_t grain) {
+		if (matrix.empty()) {
+			return;
+		}
+		EXIT_NOT_IMPLEMENTED(channels > 37);
+		auto* input = Buffer(buffer, grain);
+		mix_samples.assign(size_t(channels) * grain, 0.0f);
+		for (uint32_t dst = 0; dst < channels; ++dst) {
+			for (uint32_t src = 0; src < channels; ++src) {
+				const float level = matrix[dst * 37 + src];
+				if (level == 0.0f) {
+					continue;
+				}
+				for (uint32_t i = 0; i < grain; ++i) {
+					mix_samples[dst * grain + i] += input[src * grain + i] * level;
+				}
+			}
+		}
+		std::copy(mix_samples.begin(), mix_samples.end(), input);
+	}
+
 	void ResetSetupState() {
 		SetEvent(4);
 		std::ranges::fill(ports, Port {});
@@ -687,10 +713,11 @@ struct Ngs2VoiceInternal {
 		if (count > capacity) {
 			return NGS2_ERROR_INVALID_NUM_WAVEFORM_BLOCKS;
 		}
-		EXIT_NOT_IMPLEMENTED(!accepts_blocks || (flags & ~0x17u) != 0);
 		const bool clear     = (flags & 4u) != 0;
 		const bool only_data = (flags & 2u) != 0;
 		const bool join      = only_data && (flags & 0x10u) == 0;
+		// CLEAR reopens a list that the previous addition closed.
+		EXIT_NOT_IMPLEMENTED((!accepts_blocks && !clear) || (flags & ~0x17u) != 0);
 		const auto retained = clear ? 0 : blocks.size();
 		Block      tail     = retained == 0 ? Block {} : blocks.back();
 		std::vector<Block> pending;
@@ -1777,6 +1804,18 @@ static void Ngs2ConsumeSamples(Ngs2VoiceInternal& voice, uint32_t grain) {
 	voice.has_samples = output != 0;
 }
 
+// Mono and stereo convert at -3 dB, mono feeds the surround front center, others map 1:1.
+static float Ngs2DefaultMatrixLevel(uint32_t src_channels, uint32_t src, uint32_t dst_channels,
+                                    size_t dst) {
+	if ((src_channels == 1 && dst_channels == 2) || (src_channels == 2 && dst_channels == 1)) {
+		return std::numbers::sqrt2_v<float> / 2;
+	}
+	if (src_channels == 1 && dst_channels > 2) {
+		return dst == 2 ? 1.0f : 0.0f;
+	}
+	return src == dst ? 1.0f : 0.0f;
+}
+
 static void Ngs2RenderVoice(Ngs2VoiceInternal& voice, const std::vector<Ngs2VoiceInternal*>& voices,
                             uint32_t grain) {
 	if (voice.rendered) {
@@ -1825,7 +1864,11 @@ static void Ngs2RenderVoice(Ngs2VoiceInternal& voice, const std::vector<Ngs2Voic
 				EXIT_NOT_IMPLEMENTED(port.input >= voice.rack->NumBuffers());
 				const auto* input  = source->Buffer(source->rack->PortBufferId(p), grain);
 				auto*       output = voice.Buffer(port.input, grain);
+				// A port without a matrix, or whose matrix has no levels, uses the default mix.
 				const auto* matrix = port.matrix < 0 ? nullptr : &source->matrices.at(port.matrix);
+				if (matrix != nullptr && matrix->empty()) {
+					matrix = nullptr;
+				}
 				const size_t output_channels =
 				    matrix == nullptr
 				        ? voice.channels
@@ -1834,7 +1877,8 @@ static void Ngs2RenderVoice(Ngs2VoiceInternal& voice, const std::vector<Ngs2Voic
 					for (uint32_t src = 0; src < source->channels; ++src) {
 						const float level =
 						    port.volume * (matrix == nullptr
-						                       ? (src == dst ? 1.0f : 0.0f)
+						                       ? Ngs2DefaultMatrixLevel(source->channels, src,
+						                                                voice.channels, dst)
 						                       : (*matrix)[dst * source->channels + src]);
 						for (uint32_t i = 0; i < grain; ++i) {
 							output[dst * grain + i] += input[src * grain + i] * level;
@@ -1844,12 +1888,21 @@ static void Ngs2RenderVoice(Ngs2VoiceInternal& voice, const std::vector<Ngs2Voic
 				voice.has_samples = true;
 			}
 		}
+		if (voice.has_samples && voice.rack->type == Ngs2RackType::Reverb) {
+			for (auto& sample: voice.samples) {
+				sample *= voice.dry_level;
+			}
+		}
 		if (voice.has_samples) {
 			const auto& custom = voice.rack->option.custom_sampler.custom_rack_option;
 			const auto  rate   = voice.rack->ngs->option.sample_rate;
 			for (size_t m = 0; m < voice.modules.size(); ++m) {
 				const auto& info   = custom.module[m];
 				auto&       module = voice.modules[m];
+				if (info.module_id == 0x17) {
+					voice.MixChannels(module.channel_matrix, info.source_buffer_id, grain);
+					continue;
+				}
 				if (info.module_id != 0x19 && info.module_id != 0x1f) {
 					if (!module.render_warned) {
 						Log::WriteToConsoleAndLog(fmt::sprintf(
@@ -1934,7 +1987,8 @@ int KYTY_SYSV_ABI Ngs2SystemRender(uintptr_t system_handle, const Ngs2RenderBuff
 		EXIT_NOT_IMPLEMENTED(voice->output_id >= num_buffer_info);
 		const auto& output = buffer_info[voice->output_id];
 		EXIT_NOT_IMPLEMENTED(output.buffer == nullptr);
-		EXIT_NOT_IMPLEMENTED((output.waveform_type != 0x12 && output.waveform_type != 0x18) ||
+		EXIT_NOT_IMPLEMENTED(output.waveform_type != 0x12 && output.waveform_type != 0x18);
+		EXIT_NOT_IMPLEMENTED(voice->rack->type == Ngs2RackType::Mastering &&
 		                     output.num_channels != voice->channels);
 		EXIT_NOT_IMPLEMENTED(output.buffer_size < grain * output.num_channels *
 		                                               (output.waveform_type == 0x12 ? sizeof(int16_t)
@@ -1950,10 +2004,16 @@ int KYTY_SYSV_ABI Ngs2SystemRender(uintptr_t system_handle, const Ngs2RenderBuff
 			pcm = static_cast<float*>(output.buffer);
 		}
 		for (uint32_t dst = 0; dst < output.num_channels; ++dst) {
+			// Output channel 3 is LFE; every other channel is full band.
+			const float gain = dst == 3 ? voice->lfe_level : voice->fbw_level;
 			for (uint32_t src = 0; src < voice->channels; ++src) {
-				const float level = voice->output_matrix.empty()
-				                        ? (src == dst ? 1.0f : 0.0f)
-				                        : voice->output_matrix.at(src * output.num_channels + dst);
+				const float level =
+				    gain * (voice->output_matrix.empty()
+				                ? static_cast<float>(src == dst)
+				                : voice->output_matrix.at(src * output.num_channels + dst));
+				if (level == 0.0f) {
+					continue;
+				}
 				for (uint32_t i = 0; i < grain; ++i) {
 					pcm[i * output.num_channels + dst] += voice->samples[src * grain + i] * level;
 				}
@@ -2308,6 +2368,17 @@ int KYTY_SYSV_ABI Ngs2RackGetVoiceHandle(uintptr_t rack_handle, uint32_t voice_i
 	return OK;
 }
 
+// A patch to a rack handle connects to voice #0 of that rack.
+static Ngs2VoiceInternal* Ngs2FindPatchDestination(uintptr_t handle) {
+	Common::LockGuard racks_lock(g_racks_mutex);
+	for (auto* rack = g_racks_list; rack != nullptr; rack = rack->next) {
+		if (reinterpret_cast<uintptr_t>(rack) == handle) {
+			return reinterpret_cast<Ngs2VoiceInternal*>(rack + 1);
+		}
+	}
+	return reinterpret_cast<Ngs2VoiceInternal*>(handle);
+}
+
 // NOLINTNEXTLINE(readability-function-cognitive-complexity)
 int KYTY_SYSV_ABI Ngs2VoiceControl(uintptr_t voice_handle, const Ngs2VoiceParamHeader* param_list) {
 	PRINT_NAME();
@@ -2377,7 +2448,7 @@ int KYTY_SYSV_ABI Ngs2VoiceControl(uintptr_t voice_handle, const Ngs2VoiceParamH
 						const auto* patch = reinterpret_cast<const Ngs2VoicePatchParam*>(param);
 						EXIT_NOT_IMPLEMENTED(patch->port >= voice->ports.size());
 						voice->ports[patch->port].dest =
-						    reinterpret_cast<Ngs2VoiceInternal*>(patch->dest_handle);
+						    Ngs2FindPatchDestination(patch->dest_handle);
 						voice->ports[patch->port].input = patch->dest_input_id;
 						LOGF("\t connect->port          = %u\n"
 						     "\t connect->dest_input_id = %u\n"
@@ -2425,7 +2496,31 @@ int KYTY_SYSV_ABI Ngs2VoiceControl(uintptr_t voice_handle, const Ngs2VoiceParamH
 				}
 				break;
 			}
-			case 0x2001: EXIT_NOT_IMPLEMENTED(voice->rack->type != Ngs2RackType::Reverb); break;
+			case 0x2001: {
+				EXIT_NOT_IMPLEMENTED(voice->rack->type != Ngs2RackType::Reverb);
+				const auto* values = reinterpret_cast<const uint32_t*>(param + 1);
+				switch (param->id & 0xffffu) {
+					case 0:
+						// Input and output channels, flags; channel conversion stays silent.
+						EXIT_NOT_IMPLEMENTED(values[0] == 0 || values[2] != 0);
+						voice->ResetSetupState();
+						voice->channels = values[1] == values[0] ? values[0] : 0;
+						break;
+					case 1: {
+						// The reverb parameters start with the wet and dry levels.
+						const auto* levels = reinterpret_cast<const float*>(values);
+						voice->dry_level   = std::clamp(levels[1], 0.0f, 1.0f);
+
+						static std::atomic_bool wet_warned = false;
+						if (levels[0] != 0.0f && !wet_warned.exchange(true)) {
+							Log::WriteToConsoleAndLog("warning: NGS2 reverb wet signal is muted\n");
+						}
+						break;
+					}
+					default: EXIT("unsupported reverb control: 0x%08" PRIx32 "\n", param->id);
+				}
+				break;
+			}
 			case 0x3000: {
 				EXIT_NOT_IMPLEMENTED(voice->rack->type != Ngs2RackType::Mastering);
 				const auto* values = reinterpret_cast<const uint32_t*>(param + 1);
@@ -2439,6 +2534,12 @@ int KYTY_SYSV_ABI Ngs2VoiceControl(uintptr_t voice_handle, const Ngs2VoiceParamH
 						                            matrix->levels + matrix->num_levels);
 						break;
 					}
+					case 4: {
+						const auto* levels = reinterpret_cast<const float*>(values);
+						voice->fbw_level   = std::clamp(levels[0], 0.0f, 4.0f);
+						voice->lfe_level   = std::clamp(levels[1], 0.0f, 4.0f);
+						break;
+					}
 					case 5: voice->output_id = values[0]; break;
 					default: EXIT("unsupported mastering control: 0x%08" PRIx32 "\n", param->id);
 				}
@@ -2452,7 +2553,7 @@ int KYTY_SYSV_ABI Ngs2VoiceControl(uintptr_t voice_handle, const Ngs2VoiceParamH
 				    voice->rack->option.custom_sampler.custom_rack_option.module[index].module_id;
 				auto& module = voice->modules[index];
 				if ((param->id & 0xffffffe0u) != (0x40000000u | module_id << 8u) ||
-				    (module_id != 0x19 && module_id != 0x1f)) {
+				    (module_id != 0x17 && module_id != 0x19 && module_id != 0x1f)) {
 					if (!module.control_warned) {
 						Log::WriteToConsoleAndLog(fmt::sprintf(
 						    "warning: NGS2 custom control 0x%08" PRIx32 " is ignored\n",
@@ -2461,6 +2562,16 @@ int KYTY_SYSV_ABI Ngs2VoiceControl(uintptr_t voice_handle, const Ngs2VoiceParamH
 					}
 					break;
 				}
+				if (module_id == 0x17) {
+					EXIT_NOT_IMPLEMENTED(param->size != sizeof(Ngs2VoiceParamHeader) +
+					                                       37 * 37 * sizeof(float));
+					const auto* levels = reinterpret_cast<const float*>(param + 1);
+					module.channel_matrix.resize(37 * 37);
+					std::transform(levels, levels + module.channel_matrix.size(),
+					               module.channel_matrix.begin(),
+					               [](float level) { return std::clamp(level, -32.0f, 32.0f); });
+					break;
+				}
 				if (module_id == 0x19) {
 					struct FxParam {
 						Ngs2VoiceParamHeader     header;
@@ -2660,24 +2771,22 @@ int KYTY_SYSV_ABI Ngs2VoiceGetState(uintptr_t voice_handle, Ngs2VoiceState* stat
 			submixer->voice_state.state_flags = Ngs2GetStateFlags(voice);
 			break;
 		}
-		case Ngs2RackType::CustomMastering: {
-			const auto configured_size =
-			    voice->rack->option.custom_mastering.custom_rack_option.state_size;
-			EXIT_NOT_IMPLEMENTED(configured_size < sizeof(Ngs2CustomMasteringVoiceState));
-			EXIT_NOT_IMPLEMENTED(state_size != configured_size);
-			std::memset(state, 0, state_size);
-			auto* mastering = reinterpret_cast<Ngs2CustomMasteringVoiceState*>(state);
-			mastering->voice_state.state_flags = Ngs2GetStateFlags(voice);
-			break;
-		}
+		case Ngs2RackType::CustomMastering:
 		case Ngs2RackType::CustomSampler: {
+			const bool sampler = voice->rack->type == Ngs2RackType::CustomSampler;
 			const auto configured_size =
 			    voice->rack->option.custom_sampler.custom_rack_option.state_size;
-			EXIT_NOT_IMPLEMENTED(configured_size < sizeof(Ngs2CustomSamplerVoiceState));
-			EXIT_NOT_IMPLEMENTED(state_size != configured_size);
-			std::memset(state, 0, state_size);
-			auto* sampler = reinterpret_cast<Ngs2CustomSamplerVoiceState*>(state);
-			Ngs2FillSamplerState(*voice, *sampler);
+			EXIT_NOT_IMPLEMENTED(configured_size < (sampler ? sizeof(Ngs2CustomSamplerVoiceState)
+			                                                : sizeof(Ngs2CustomMasteringVoiceState)));
+			if (state_size < configured_size) {
+				return static_cast<int32_t>(0x804a8011u);
+			}
+			std::memset(state, 0, configured_size);
+			if (sampler) {
+				Ngs2FillSamplerState(*voice, *reinterpret_cast<Ngs2CustomSamplerVoiceState*>(state));
+			} else {
+				state->state_flags = Ngs2GetStateFlags(voice);
+			}
 			break;
 		}
 		case Ngs2RackType::Sampler: {
@@ -2722,6 +2831,50 @@ int KYTY_SYSV_ABI Ngs2VoiceGetStateFlags(uintptr_t voice_handle, uint32_t* state
 	return OK;
 }
 
+int KYTY_SYSV_ABI Ngs2VoiceQueryInfo(uintptr_t voice_handle, uint32_t query_id, void* info,
+                                     size_t info_size) {
+	PRINT_NAME();
+
+	struct SetupInfo {
+		uint32_t channels, flags;
+	};
+	EXIT_NOT_IMPLEMENTED(voice_handle == 0 || info == nullptr);
+	EXIT_NOT_IMPLEMENTED(query_id != 0x4001 || info_size != sizeof(SetupInfo));
+
+	auto* voice = reinterpret_cast<Ngs2VoiceInternal*>(voice_handle);
+
+	Common::LockGuard lock(voice->rack->ngs->mutex);
+
+	*static_cast<SetupInfo*>(info) = {voice->channels, 0};
+
+	return OK;
+}
+
+int KYTY_SYSV_ABI Ngs2VoiceGetPortInfo(uintptr_t voice_handle, uint32_t port, void* info,
+                                       size_t info_size) {
+	PRINT_NAME();
+
+	struct PortInfo {
+		int32_t   matrix;
+		float     volume;
+		uint32_t  delay_samples, input;
+		uintptr_t dest;
+	};
+	EXIT_NOT_IMPLEMENTED(voice_handle == 0 || info == nullptr || info_size != sizeof(PortInfo));
+
+	auto* voice = reinterpret_cast<Ngs2VoiceInternal*>(voice_handle);
+
+	Common::LockGuard lock(voice->rack->ngs->mutex);
+
+	EXIT_NOT_IMPLEMENTED(port >= voice->ports.size());
+	const auto& p = voice->ports[port];
+
+	*static_cast<PortInfo*>(info) = {p.matrix, p.volume, 0, p.input,
+	                                 reinterpret_cast<uintptr_t>(p.dest)};
+
+	return OK;
+}
+
 int KYTY_SYSV_ABI Ngs2SystemLock(uintptr_t system_handle) {
 	PRINT_NAME();
 

src/libs/ngs2.h

@@ -60,6 +60,10 @@ int KYTY_SYSV_ABI Ngs2VoiceRunCommands(uintptr_t voice_handle, const void* comma
 int KYTY_SYSV_ABI Ngs2VoiceGetState(uintptr_t voice_handle, Ngs2VoiceState* state,
                                     size_t state_size);
 int KYTY_SYSV_ABI Ngs2VoiceGetStateFlags(uintptr_t voice_handle, uint32_t* state_flags);
+int KYTY_SYSV_ABI Ngs2VoiceQueryInfo(uintptr_t voice_handle, uint32_t query_id, void* info,
+                                     size_t info_size);
+int KYTY_SYSV_ABI Ngs2VoiceGetPortInfo(uintptr_t voice_handle, uint32_t port, void* info,
+                                       size_t info_size);
 int KYTY_SYSV_ABI Ngs2SystemRender(uintptr_t system_handle, const Ngs2RenderBufferInfo* buffer_info,
                                    uint32_t num_buffer_info);
 int KYTY_SYSV_ABI Ngs2ParseWaveformData(const void* data, size_t data_size, Ngs2WaveformInfo* info);

src/loader/systemContent.cpp

@@ -2,6 +2,7 @@
 
 #include "common/assert.h"
 #include "common/common.h"
+#include "common/emulatorConfig.h"
 #include "common/file.h"
 #include "common/logging/log.h"
 #include "common/singleton.h"
@@ -90,6 +91,7 @@ private:
 };
 
 struct SystemContent {
+	bool                  trinity_mode = false;
 	std::filesystem::path psf_path;
 	Psf                   psf;
 	std::filesystem::path playgo_path;
@@ -603,6 +605,14 @@ void SystemContentLoadParamSfo(const std::filesystem::path& file_name) {
 
 	sc->psf_path = file_name;
 	sc->psf.DbgPrint();
+	int32_t attribute3 = 0;
+	const bool title_supports_trinity = sc->psf.GetParamInt("ATTRIBUTE3", &attribute3) &&
+	                                    (static_cast<uint32_t>(attribute3) & (1u << 22u)) != 0;
+	sc->trinity_mode = Config::Ps5ProEnabled() && title_supports_trinity;
+	if (Config::Ps5ProEnabled()) {
+		Log::WriteToConsoleAndLog(fmt::format("[PS5] title Trinity support={}\n",
+		                                    title_supports_trinity ? "yes" : "no"));
+	}
 
 	sc->icon_path = file_name.parent_path() / "icon0.png";
 
@@ -626,6 +636,10 @@ void SystemContentLoadParamSfo(const std::filesystem::path& file_name) {
 	}
 }
 
+bool SystemContentIsTrinityMode() {
+	return Common::Singleton<SystemContent>::Instance()->trinity_mode;
+}
+
 bool SystemContentParamSfoGetInt(const char* name, int32_t* value) {
 	if (name == nullptr || value == nullptr) {
 		return false;

src/loader/systemContent.h

@@ -11,6 +11,7 @@
 namespace Loader {
 
 void     SystemContentLoadParamSfo(const std::filesystem::path& file_name);
+bool     SystemContentIsTrinityMode();
 bool     SystemContentParamSfoGetInt(const char* name, int32_t* value);
 bool     SystemContentParamSfoGetString(const char* name, std::string* value);
 bool     SystemContentParamSfoGetString(const char* name, char* value, size_t value_size);

src/main.cpp

@@ -74,11 +74,14 @@ static void PrintUsage() {
 	    "  --stretch                            Stretch the image to the window. Default: off.\n");
 	::printf("  --no-splash                          Disable the startup splash screen.\n");
 	::printf("  --vr                                 Enable the virtual VR headset.\n");
+	::printf("  --trinity                            Enable experimental PS5 Pro mode for supported titles\n"
+	         "                                       (PSSR uses linear scaling). Default: off.\n");
 	::printf("  --amd-cpu                            Apply AMD CPU instruction patches.\n");
 	::printf("  --vblank-frequency <num>             Virtual vblank frequency. Default: 60.\n");
 	::printf("  --console-language <0-29>            Console language. Default: 1 (English US).\n");
 	::printf("  --vulkan-validation                  Enable Vulkan validation.\n");
 	::printf("  --network-connected                  Use the host local network.\n");
+	::printf("  --emulate-psn                        Simulate PSN sign-in for compatibility.\n");
 	::printf("  --gpu-assisted-validation            Bounds-check shader accesses on the GPU.\n"
 	         "                                       Implies --vulkan-validation; very slow.\n");
 	::printf("  --shader-validation                  Enable shader validation.\n");
@@ -204,8 +207,10 @@ static constexpr Flag FLAGS[] = {
     {"--stretch", &ConfigOptions::stretch_enabled},
     {"--no-splash", &ConfigOptions::splash_enabled, false},
     {"--vr", &ConfigOptions::vr_enabled},
+    {"--trinity", &ConfigOptions::pro_enabled},
     {"--amd-cpu", &ConfigOptions::amd_cpu_enabled},
     {"--network-connected", &ConfigOptions::network_connected},
+    {"--emulate-psn", &ConfigOptions::psn_emulation_enabled},
     {"--vulkan-validation", &ConfigOptions::vulkan_validation_enabled},
     {"--gpu-assisted-validation", &ConfigOptions::gpu_assisted_validation_enabled},
     {"--shader-validation", &ConfigOptions::shader_validation_enabled},

tests/AjmResampleTests.cpp

@@ -0,0 +1,360 @@
+#include "libs/ajm.cpp"
+
+#include <array>
+#include <cmath>
+#include <cstdlib>
+#include <fstream>
+#include <opus.h>
+
+using namespace Libs::Audio::Ajm;
+
+namespace {
+
+void Check(bool condition, const char* message) {
+	if (!condition) {
+		std::fprintf(stderr, "AjmResampleTests: %s\n", message);
+		std::abort();
+	}
+}
+
+struct DecodeResult {
+	AjmSidebandResult result;
+	AjmSidebandStream stream;
+	AjmSidebandMFrame frames;
+};
+
+struct Fixture {
+	uint32_t                  instance = 0;
+	uint32_t                  channels;
+	std::array<uint8_t, 4096> batch_data {};
+	AjmBatchInfo              batch {};
+
+	explicit Fixture(uint32_t channels = 1, AjmSampleEncoding encoding = AjmSampleEncoding::S16,
+	                 uint64_t flags = 0x1000)
+	    : channels(channels) {
+		Check(AjmInstanceCreate(1, 24, flags | channels | (uint64_t(encoding) << 7), &instance) ==
+		          OK,
+		      "create Opus instance");
+		Initialize();
+	}
+	~Fixture() { AjmInstanceDestroy(1, instance); }
+
+	AjmBatchInfo* Batch() {
+		batch = {batch_data.data(), 0, batch_data.size(), nullptr, nullptr};
+		return &batch;
+	}
+
+	void Initialize() {
+		AjmDecOpusInitializeParameters params {channels, 48000, 0};
+		AjmSidebandResult              result {};
+		AjmBatchJobInitialize(Batch(), instance, &params, sizeof(params), &result);
+		Check(result.result == OK, "initialize Opus instance");
+	}
+
+	void Parameters(float ratio, uint32_t flags, float change = 0, bool extended = false) {
+		AjmSidebandResult result {};
+		if (extended) {
+			AjmBatchJobSetResampleParametersEx(Batch(), instance, ratio, change, flags, &result);
+		} else {
+			AjmBatchJobSetResampleParameters(Batch(), instance, ratio, flags, &result);
+		}
+		Check(result.result == OK, "set resampling parameters");
+	}
+
+	AjmSidebandResampleInfo Info() {
+		struct {
+			AjmSidebandResult       result;
+			AjmSidebandResampleInfo info;
+		} result {};
+		AjmBatchJobGetResampleInfo(Batch(), instance, &result);
+		Check(result.result.result == OK, "get resampling info");
+		return result.info;
+	}
+
+	void Gapless(uint32_t samples, uint16_t skip) {
+		AjmSidebandGaplessDecode params {samples, skip, 0};
+		AjmSidebandResult        result {};
+		AjmBatchJobSetGaplessDecode(Batch(), instance, &params, 1, &result);
+		Check(result.result == OK, "set gapless parameters");
+	}
+
+	DecodeResult Decode(const void* input, size_t size, void* output, size_t capacity) {
+		DecodeResult result {};
+		AjmBatchJobDecode(Batch(), instance, input, size, output, capacity, &result);
+		Check((result.result.result & ~(AJM_RESULT_NOT_ENOUGH_ROOM | AJM_RESULT_PARTIAL_INPUT)) ==
+		          0,
+		      "decode returned unexpected error");
+		return result;
+	}
+};
+
+std::vector<uint8_t> Encode(uint32_t channels, int packets) {
+	int   error   = 0;
+	auto* encoder = opus_encoder_create(48000, channels, OPUS_APPLICATION_AUDIO, &error);
+	Check(encoder != nullptr && error == OPUS_OK, "create test encoder");
+	std::vector<uint8_t>      stream;
+	std::vector<int16_t>      pcm(960 * channels);
+	std::array<uint8_t, 4096> packet;
+	for (int p = 0; p < packets; ++p) {
+		for (size_t i = 0; i < pcm.size(); ++i) {
+			pcm[i] = static_cast<int16_t>(
+			    12000 * std::sin((i / channels + p * 960) * 0.041 + (i % channels)));
+		}
+		const int size = opus_encode(encoder, pcm.data(), 960, packet.data(), packet.size());
+		Check(size > 0, "encode test packet");
+		stream.push_back(static_cast<uint8_t>(size));
+		stream.push_back(static_cast<uint8_t>(size >> 8));
+		stream.insert(stream.end(), packet.begin(), packet.begin() + size);
+	}
+	opus_encoder_destroy(encoder);
+	return stream;
+}
+
+void SmallOutputs(const std::vector<uint8_t>& packet, uint16_t skip = 0) {
+	Fixture f;
+	f.Gapless(960 - skip, skip);
+	f.Parameters(1, 3);
+	std::array<uint8_t, 1030> output;
+	output.fill(0x5a);
+	auto first = f.Decode(packet.data(), packet.size(), output.data(), 1024);
+	Check(first.stream.size_consumed == static_cast<int>(packet.size()),
+	      "small output must consume packet");
+	Check(first.stream.size_produced == 1024, "small output must produce 1024 bytes");
+	Check(first.frames.num_frames == 1 && first.stream.total_decoded_samples == 960 - skip,
+	      "decoded frame counted once before resampling");
+	Check(f.Info().num_samples == 448 - skip, "report retained source samples");
+	Check(std::all_of(output.begin() + 1024, output.end(), [](auto v) { return v == 0x5a; }),
+	      "do not write past output capacity");
+	const auto last = f.Decode(nullptr, 0, output.data(), 1024);
+	Check(last.result.result == OK && last.stream.size_consumed == 0 &&
+	          last.stream.size_produced == 896 - skip * 2 && last.frames.num_frames == 0 &&
+	          last.stream.total_decoded_samples == first.stream.total_decoded_samples,
+	      "END drains retained PCM without decoding or reporting partial input");
+	Check(f.Decode(nullptr, 0, output.data(), 1024).stream.size_produced == 0,
+	      "END drains only once");
+}
+
+void UnityMatchesWholeDecode(uint32_t channels, AjmSampleEncoding encoding) {
+	const auto stream       = Encode(channels, 6);
+	const auto sample_bytes = channels * AjmBytesPerSample(encoding);
+	Fixture    reference(channels, encoding, 0);
+	reference.Gapless(5328, 312);
+	std::vector<uint8_t> expected(5760 * sample_bytes);
+	const auto           full =
+	    reference.Decode(stream.data(), stream.size(), expected.data(), expected.size());
+	expected.resize(full.stream.size_produced);
+	Check(expected.size() == 5328 * sample_bytes, "reference gapless length");
+
+	Fixture f(channels, encoding);
+	f.Gapless(5328, 312);
+	f.Parameters(1, 3);
+	std::vector<uint8_t>      actual;
+	std::array<uint8_t, 1025> output;
+	size_t                    offset = 0;
+	uint32_t                  frames = 0;
+	for (int i = 0; i < 100; ++i) {
+		const auto decoded = f.Decode(offset < stream.size() ? stream.data() + offset : nullptr,
+		                              stream.size() - offset, output.data(), output.size());
+		offset += decoded.stream.size_consumed;
+		frames += decoded.frames.num_frames;
+		actual.insert(actual.end(), output.begin(), output.begin() + decoded.stream.size_produced);
+		if (decoded.stream.size_produced == 0) {
+			break;
+		}
+	}
+	Check(actual == expected && frames == 6 && offset == stream.size(),
+	      "chunked PCM must exactly match whole decode, without loss or duplication");
+}
+
+void ChunkedRates(const std::vector<uint8_t>& stream, uint32_t samples, uint16_t skip) {
+	for (float ratio: {0.125f, 0.749153f, 1.334840f}) {
+		Fixture reference;
+		reference.Gapless(samples, skip);
+		reference.Parameters(ratio, 3);
+		std::vector<uint8_t> expected(
+		    (static_cast<size_t>(std::ceil(samples / double(ratio))) + 1) * 2);
+		const auto full =
+		    reference.Decode(stream.data(), stream.size(), expected.data(), expected.size());
+		expected.resize(full.stream.size_produced);
+		Check(expected.size() == static_cast<size_t>(std::ceil(samples / double(ratio))) * 2,
+		      "fractional resampling rounds up final output sample count");
+
+		Fixture f;
+		f.Gapless(samples, skip);
+		f.Parameters(ratio, 1);
+		std::vector<uint8_t>      actual;
+		std::array<uint8_t, 1024> output;
+		size_t                    offset = 0;
+		for (int i = 0; i < 1000; ++i) {
+			if (offset == stream.size()) {
+				// The caller can signal END only after submitting the last packet.
+				f.Parameters(0, 2 | 8);
+			}
+			const auto decoded = f.Decode(offset < stream.size() ? stream.data() + offset : nullptr,
+			                              stream.size() - offset, output.data(), output.size());
+			offset += decoded.stream.size_consumed;
+			actual.insert(actual.end(), output.begin(),
+			              output.begin() + decoded.stream.size_produced);
+			if (decoded.stream.size_produced == 0) {
+				break;
+			}
+		}
+		Check(actual == expected && offset == stream.size(),
+		      "non-unity PCM is independent of output chunk boundaries and deferred END");
+	}
+}
+
+void LoopBoundaries() {
+	const auto loop   = Encode(1, 2);
+	auto       stream = loop;
+	stream.insert(stream.end(), loop.begin(), loop.end());
+	for (uint64_t flags: {0x1400u, 0x3400u}) {
+		Fixture f(1, AjmSampleEncoding::S16, flags);
+		f.Gapless(1500, 100);
+		f.Parameters(1, 1);
+		std::array<uint8_t, 8192> pcm;
+		const auto                first = f.Decode(stream.data(), stream.size(), pcm.data(), 1024);
+		Check(first.stream.size_produced == 1024, "loop first partial output");
+		const auto second =
+		    f.Decode(stream.data() + first.stream.size_consumed,
+			         stream.size() - first.stream.size_consumed, pcm.data(), pcm.size());
+		const bool seamless = flags == 0x3400;
+		Check(second.stream.size_produced == (seamless ? 6000 - 1024 : 3000 - 1024),
+		      "non-seamless loop stops job at stream boundary");
+		const auto consumed = first.stream.size_consumed + second.stream.size_consumed;
+		Check(consumed == static_cast<int>(seamless ? stream.size() : loop.size()),
+		      "loop boundary input consumption");
+		if (!seamless) {
+			Check(
+			    f.Decode(stream.data() + consumed, stream.size() - consumed, pcm.data(), pcm.size())
+			            .stream.size_produced == 3000,
+			    "next job resumes gapless loop");
+		}
+	}
+}
+
+void RatesAndControls(const std::vector<uint8_t>& packet) {
+	for (float ratio: {0.125f, 0.5f, 1.0f, 1.25f, 2.0f, 4.0f}) {
+		Fixture f;
+		f.Parameters(ratio, 3);
+		std::vector<int16_t> pcm(8000);
+		const auto result = f.Decode(packet.data(), packet.size(), pcm.data(), pcm.size() * 2);
+		Check(result.stream.size_produced == static_cast<int>(std::ceil(960.0 / ratio)) * 2,
+		      "resampling ratio controls output duration");
+		Check(result.stream.total_decoded_samples == 960,
+		      "ratio does not change decoded sample count");
+	}
+	Fixture f;
+	f.Parameters(1, 1);
+	std::array<int16_t, 512> pcm;
+	f.Decode(packet.data(), packet.size(), pcm.data(), sizeof(pcm));
+	f.Initialize();
+	Check(f.Info().num_samples == 448, "codec initialize must preserve resampler PCM");
+	f.Parameters(0, 0);
+	const auto frozen = f.Decode(nullptr, 0, pcm.data(), sizeof(pcm));
+	Check(
+	    frozen.stream.size_produced == sizeof(pcm) && f.Info().num_samples == 448 &&
+	        std::all_of(pcm.begin(), pcm.end(), [&](auto sample) { return sample == pcm.front(); }),
+	    "zero ratio repeats one sample without advancing input");
+	f.Parameters(1, 0, 0.25f, true);
+	f.Parameters(std::numeric_limits<float>::quiet_NaN(), 8);
+	f.Decode(nullptr, 0, pcm.data(), 4 * sizeof(int16_t));
+	Check(f.Info().ratio == 2.0f, "ignore preserves ratio ramp");
+	f.Parameters(1, 1);
+	Check(f.Info().num_samples == 0, "resample RESET discards retained PCM");
+	Check(f.Decode(nullptr, 0, pcm.data(), sizeof(pcm)).stream.size_produced == 0,
+	      "RESET leaves no stale audio");
+}
+
+void UnbufferedAndSplit(const std::vector<uint8_t>& packet) {
+	Fixture                   raw(1, AjmSampleEncoding::S16, 0);
+	std::array<uint8_t, 1024> pcm;
+	const auto too_small = raw.Decode(packet.data(), packet.size(), pcm.data(), pcm.size());
+	Check(too_small.result.result == AJM_RESULT_NOT_ENOUGH_ROOM &&
+	          too_small.stream.size_consumed == 0 && too_small.stream.size_produced == 0,
+	      "ordinary decoding still requires room for a whole packet");
+	Fixture f;
+	f.Parameters(1, 3);
+	AjmBuffer    input[] {{const_cast<uint8_t*>(packet.data()), 1},
+	                      {const_cast<uint8_t*>(packet.data()) + 1, packet.size() - 1}};
+	AjmBuffer    output[] {{pcm.data(), 257}, {pcm.data() + 257, 767}};
+	DecodeResult result {};
+	AjmBatchJobDecodeSplit(f.Batch(), f.instance, input, 2, output, 2, &result);
+	Check(result.stream.size_produced == 1024 &&
+	          result.stream.size_consumed == static_cast<int>(packet.size()),
+	      "split buffers retain PCM across jobs");
+}
+
+void FormatChangeRequiresReset() {
+	Fixture f(2);
+	f.Parameters(1, 1);
+	const auto                stereo = Encode(2, 1);
+	std::array<uint8_t, 4096> pcm;
+	f.Decode(stereo.data(), stereo.size(), pcm.data(), 1024);
+	f.channels = 1;
+	f.Initialize();
+	DecodeResult result {};
+	AjmBatchJobDecode(f.Batch(), f.instance, nullptr, 0, pcm.data(), pcm.size(), &result);
+	Check(result.result.result == AJM_RESULT_INVALID_PARAMETER && result.stream.size_produced == 0,
+	      "format change cannot reinterpret retained PCM with a different channel stride");
+	f.Parameters(1, 3);
+	const auto mono = Encode(1, 1);
+	Check(f.Decode(mono.data(), mono.size(), pcm.data(), pcm.size()).stream.size_produced == 1920,
+	      "resampler RESET permits a new stream format");
+}
+
+void EndPersistsUntilReset(const std::vector<uint8_t>& packet) {
+	Fixture f;
+	f.Parameters(0.5f, 1);
+	std::array<int16_t, 64> pcm;
+	auto   decoded  = f.Decode(packet.data(), packet.size(), pcm.data(), sizeof(pcm));
+	size_t produced = decoded.stream.size_produced;
+	f.Parameters(0.5f, 2);
+	for (int i = 0; i < 40; ++i) {
+		f.Parameters(0.5f, 0);
+		decoded = f.Decode(nullptr, 0, pcm.data(), sizeof(pcm));
+		produced += decoded.stream.size_produced;
+		Check((decoded.result.result & AJM_RESULT_PARTIAL_INPUT) == 0,
+		      "END remains active across subsequent ratio controls");
+		if (f.Info().num_samples < 0) {
+			break;
+		}
+	}
+	Check(produced == 3840 && f.Info().num_samples < 0,
+	      "fully drained END reports a negative retained sample count");
+	f.Parameters(1, 1);
+	Check(f.Info().num_samples == 0, "RESET clears end-of-stream state");
+}
+
+} // namespace
+
+int main(int argc, char** argv) {
+	const auto packet = Encode(1, 1);
+	SmallOutputs(packet);
+	SmallOutputs(packet, 312);
+	for (uint32_t channels: {1u, 2u}) {
+		for (auto encoding:
+		     {AjmSampleEncoding::S16, AjmSampleEncoding::S32, AjmSampleEncoding::Float}) {
+			UnityMatchesWholeDecode(channels, encoding);
+		}
+	}
+	RatesAndControls(packet);
+	UnbufferedAndSplit(packet);
+	ChunkedRates(Encode(1, 6), 5328, 312);
+	LoopBoundaries();
+	FormatChangeRequiresReset();
+	EndPersistsUntilReset(packet);
+	if (argc >= 2) {
+		std::ifstream        file(argv[1], std::ios::binary);
+		std::vector<uint8_t> real_packet((std::istreambuf_iterator<char>(file)), {});
+		Check(real_packet.size() == 209, "expected real 209-byte packet fixture");
+		SmallOutputs(real_packet, 312);
+	}
+	if (argc >= 3) {
+		std::ifstream        file(argv[2], std::ios::binary);
+		std::vector<uint8_t> real_stream((std::istreambuf_iterator<char>(file)), {});
+		Check(real_stream.size() == 4048, "expected complete real stream fixture");
+		ChunkedRates(real_stream, 25920, 312);
+	}
+	std::puts("AjmResampleTests: passed");
+}

tests/KernelFileSystemTests.cpp

@@ -73,6 +73,10 @@ namespace Libs::LibNpCommerce {
 void InitNet_1_NpCommerce(Loader::SymbolDatabase *symbols);
 }
 
+namespace Libs::LibNpAuth {
+void InitNet_1_NpAuth(Loader::SymbolDatabase *symbols);
+}
+
 namespace Libs::LibNpEntitlementAccess {
 void InitNet_1_NpEntitlementAccess(Loader::SymbolDatabase *symbols);
 }
@@ -97,6 +101,17 @@ void WaitForMutexTest(Predicate ready) {
   }
 }
 
+void TestStaticRwlockDestroy() {
+  using namespace Libs::LibKernel;
+  PthreadRwlock rwlock = nullptr;
+  Check(PthreadRwlockDestroy(nullptr) == KERNEL_ERROR_EINVAL &&
+            PthreadRwlockDestroy(&rwlock) == OK && rwlock == nullptr,
+        "destroy an untouched static rwlock without allocating it");
+  Check(PthreadRwlockRdlock(&rwlock) == OK && PthreadRwlockUnlock(&rwlock) == OK &&
+            PthreadRwlockDestroy(&rwlock) == OK,
+        "static rwlock remains usable after destroying its untouched initializer");
+}
+
 struct MutexWaiter {
   Libs::LibKernel::PthreadMutex *mutex;
   Libs::LibKernel::KernelUseconds timeout = 0;
@@ -573,17 +588,16 @@ void TestAddContent(const std::filesystem::path& root) {
   Check(mount(0, &info[3].label, &point) == OK &&
             !std::string_view(point.data).ends_with('/'),
         "mount owned add-on without trailing slash");
+  MountPoint repeated {};
+  Check(mount(0, &info[3].label, &repeated) == static_cast<int>(0x80d90003u),
+        "duplicate mount reports BUSY");
   const std::string mounted_file = std::string(point.data) + "/payload.txt";
   const int fd = FileSystem::KernelOpen(mounted_file.c_str(), 0, 0);
   std::array<char, 11> payload {};
   Check(fd >= 3 && FileSystem::KernelRead(fd, payload.data(), payload.size()) == payload.size() &&
             std::string_view(payload.data(), payload.size()) == "DLC payload" &&
             FileSystem::KernelClose(fd) == OK,
-        "guest reads files from the selected title's add-on");
-  MountPoint repeated {};
-  Check(mount(0, &info[3].label, &repeated) == OK &&
-            std::strcmp(repeated.data, point.data) == 0,
-        "duplicate mount returns the same mount point");
+        "guest reads the original mount after a rejected duplicate mount");
   Check(unmount(&point) == OK && FileSystem::GetRealFilename(mounted_file).empty() &&
             unmount(&point) == static_cast<int>(0x80d90004u),
         "unmount removes mapping and repeated unmount reports NOT_MOUNTED");
@@ -2082,6 +2096,256 @@ void TestNetCtlHostIpv4() {
 }
 #endif
 
+template <int Mode>
+void KYTY_SYSV_ABI RecordNpState(int user_id, uint32_t state, void* arg) {
+  auto& event = *static_cast<std::array<int, 3>*>(arg);
+  ++event[0];
+  event[1] = user_id;
+  event[2] = static_cast<int>(state);
+}
+
+void TestNpEmulation() {
+  using namespace Libs::Network::NpManager;
+  using namespace Libs::Network::NetCtl;
+  constexpr int SignedOut = static_cast<int>(0x80550006u);
+  constexpr int NetworkUnavailable = static_cast<int>(0x8055a02au);
+  constexpr int Aborted = static_cast<int>(0x80550012u);
+  struct AsyncParameters {
+    size_t size = sizeof(AsyncParameters);
+    uint64_t affinity = 0;
+    int priority = 0;
+    uint8_t padding[4] {};
+  } async_parameters;
+  struct PremiumParameters {
+    size_t size = sizeof(PremiumParameters);
+    int user_id = 0;
+    uint8_t padding[4] {};
+    uint64_t features = 1;
+    uint8_t reserved[32] {};
+  } premium_parameters;
+  struct PremiumResult {
+    bool authorized = true;
+    uint8_t reserved[32] {};
+  } premium_result;
+
+  const auto create_request = [&] (bool async) {
+    const int request = async
+        ? NpCreateAsyncRequest(reinterpret_cast<const NpCreateAsyncRequestParameter*>(
+              &async_parameters))
+        : NpCreateRequest();
+    Check(request > 0, "create NP request");
+    return request;
+  };
+  const auto check_request = [&] (bool async, int expected, auto operation) {
+    const int request = create_request(async);
+    Check(operation(request) == (async ? OK : expected),
+          "NP call separates asynchronous submission from its execution result");
+    if (async) {
+      int result = 123;
+      Check(NpPollAsync(request, &result) == OK && result == expected,
+            "NP polling reports the completed operation result");
+    }
+    Check(NpDeleteRequest(request) == OK, "release completed NP request");
+  };
+
+  Config::ConfigOptions options;
+  options.printf_direction = Config::LogDirection::Silent;
+  Check(!options.network_connected && !options.psn_emulation_enabled,
+        "host networking and PSN emulation both require explicit opt-in");
+  static std::array<std::array<int, 3>, 4> events {};
+  const std::array<NpStateCallbackA, 4> callbacks {
+      RecordNpState<0>, RecordNpState<1>, RecordNpState<2>, RecordNpState<3>};
+  int mode = 0;
+  for (bool network : {false, true}) {
+    int previous_network_state = -1;
+    for (bool emulate : {false, true}) {
+      options.network_connected = network;
+      options.psn_emulation_enabled = emulate;
+      Config::Load(options);
+      Check(Config::IsNetworkConnected() == network && Config::PsnEmulationEnabled() == emulate,
+            "loading PSN emulation does not change the independent network option");
+      Check(NetCtlInit() == OK, "initialize network state for NP mode");
+      int network_state = -1;
+      Check(NetCtlGetState(&network_state) == OK && (network || network_state == 0),
+            "PSN emulation cannot connect disabled host networking");
+      Check(previous_network_state == -1 || previous_network_state == network_state,
+            "PSN emulation does not change host network connectivity");
+      previous_network_state = network_state;
+      const bool connected = network_state == 3;
+
+      Check(NpRegisterStateCallbackA(callbacks[mode], &events[mode]) > 0 &&
+                NpCheckCallback() == OK && events[mode][0] == 1 &&
+                events[mode][1] == Config::GetUserId() &&
+                events[mode][2] == (emulate ? 2 : 1),
+            "initial NP callback matches the selected emulated sign-in state");
+      ++mode;
+      for (int player = 0; player < Config::MAX_PLAYERS; ++player) {
+        const int user_id = Config::GetPlayerUserId(player);
+        const bool signed_in = emulate && player == 0;
+        const int identity_result = signed_in ? OK : SignedOut;
+        const int online_result = !signed_in ? SignedOut : connected ? OK : NetworkUnavailable;
+        uint32_t state = 99;
+        bool signed_up = false;
+        Check(NpGetState(user_id, &state) == OK && state == (signed_in ? 2u : 1u) &&
+                  NpHasSignedUp(user_id, &signed_up) == OK && signed_up == signed_in,
+              "only the primary emulated user has a persistent signed-in account");
+        Check(NpGetNpReachabilityState(user_id, &state) == OK &&
+                  state == (signed_in && connected ? 2u : 0u),
+              "NP reachability requires both the emulated account and host connectivity");
+
+        uint64_t account_id = 0;
+        std::array<char, 20> online_id {};
+        std::array<char, 36> np_id {};
+        std::array<char, 4> country {};
+        Check(NpGetAccountIdA(user_id, &account_id) == identity_result &&
+                  NpGetOnlineId(user_id, reinterpret_cast<NpOnlineId*>(online_id.data())) ==
+                      identity_result &&
+                  NpGetNpId(user_id, reinterpret_cast<NpId*>(np_id.data())) == identity_result &&
+                  NpGetAccountCountryA(user_id, country.data()) == identity_result,
+              "NP identity and account-country getters agree on primary and guest sign-in");
+        if (signed_in) {
+          Check(account_id != 0 && online_id[0] != 0 && online_id[16] == 0 &&
+                    std::strcmp(online_id.data(), np_id.data()) == 0 &&
+                    country[0] != 0 && country[1] != 0 && country[2] == 0,
+                "emulated account identity and country remain usable while disconnected");
+        }
+
+        for (bool async : {false, true}) {
+          check_request(async, online_result, [&] (int request) {
+            return NpCheckNpReachability(request, user_id);
+          });
+          uint8_t age = 0;
+          check_request(async, identity_result, [&] (int request) {
+            return NpGetAccountAge(request, user_id, &age);
+          });
+          std::array<char, 48> language {};
+          check_request(async, identity_result, [&] (int request) {
+            return NpGetAccountLanguage2(request, user_id, language.data());
+          });
+          Check(!signed_in || (age > 0 && language[0] != 0 && language[35] == 0),
+                "signed-in account metadata is available independently of host networking");
+          premium_parameters.user_id = user_id;
+          premium_result.authorized = true;
+          check_request(async, online_result, [&] (int request) {
+            return NpCheckPremium(request,
+                reinterpret_cast<const NpCheckPremiumParameter*>(&premium_parameters),
+                reinterpret_cast<NpCheckPremiumResult*>(&premium_result));
+          });
+          Check(!premium_result.authorized, "PSN emulation does not grant premium eligibility");
+          if (player == 0) {
+            check_request(async, online_result, [&] (int request) {
+              return NpCheckNpAvailability(request, "Kyty", nullptr);
+            });
+          }
+        }
+      }
+      const int aborted = create_request(true);
+      int result = 123;
+      Check(NpAbortRequest(aborted) == OK &&
+                NpCheckNpReachability(aborted, Config::GetUserId()) == Aborted &&
+                NpPollAsync(aborted, &result) == OK && result == Aborted &&
+                NpDeleteRequest(aborted) == OK,
+            "NP emulation does not turn an aborted request into a successful one");
+      NetCtlTerm();
+    }
+  }
+  Config::Load(Config::ConfigOptions {});
+}
+
+void TestNpAuthEmulation() {
+  Loader::SymbolDatabase symbols;
+  Libs::LibNpAuth::InitNet_1_NpAuth(&symbols);
+  const auto resolve = [&] (const char* nid) {
+    const auto* symbol = symbols.FindByNid(nid, Loader::SymbolType::Func);
+    Check(symbol != nullptr, "resolve NP authentication exports");
+    return symbol->vaddr;
+  };
+  const auto create = reinterpret_cast<int(KYTY_SYSV_ABI*)()>(resolve("6bwFkosYRQg"));
+  const auto create_async = reinterpret_cast<int(KYTY_SYSV_ABI*)(const void*)>(
+      resolve("N+mr7GjTvr8"));
+  const auto destroy = reinterpret_cast<int(KYTY_SYSV_ABI*)(int)>(resolve("H8wG9Bk-nPc"));
+  const auto poll = reinterpret_cast<int(KYTY_SYSV_ABI*)(int, int*)>(resolve("gjSyfzSsDcE"));
+  const auto get_code = reinterpret_cast<int(KYTY_SYSV_ABI*)(int, const void*, void*, int*)>(
+      resolve("KI4dHLlTNl0"));
+  const auto get_token = reinterpret_cast<int(KYTY_SYSV_ABI*)(int, const void*, void*)>(
+      resolve("RdsFVsgSpZY"));
+  constexpr int InvalidArgument = static_cast<int>(0x80550301u);
+  constexpr int InvalidSize = static_cast<int>(0x80550302u);
+  constexpr int ServiceUnavailable = static_cast<int>(0x80550401u);
+  constexpr int SignedOut = static_cast<int>(0x80550006u);
+  struct AuthorizationCodeParameters {
+    size_t size = sizeof(AuthorizationCodeParameters);
+    int user_id = 0;
+    uint8_t padding[4] {};
+    const void* client_id = "test-client";
+    const char* scope = "test-scope";
+  } code_parameters;
+  struct IdTokenParameters {
+    size_t size = sizeof(IdTokenParameters);
+    int user_id = 0;
+    uint8_t padding[4] {};
+    const void* client_id = "test-client";
+    const void* client_secret = "test-secret";
+    const char* scope = "test-scope";
+  } token_parameters;
+  std::array<uint8_t, 136> code {};
+  std::array<uint8_t, 4104> token {};
+  const std::array<uint64_t, 3> async_parameters {24, 0, 0};
+  int issuer = 0;
+  const int invalid_request = create();
+  code_parameters.size = 0;
+  token_parameters.size = 0;
+  Check(invalid_request > 0 &&
+            get_code(invalid_request, nullptr, code.data(), &issuer) == InvalidArgument &&
+            get_token(invalid_request, nullptr, token.data()) == InvalidArgument &&
+            get_code(invalid_request, &code_parameters, code.data(), &issuer) == InvalidSize &&
+            get_token(invalid_request, &token_parameters, token.data()) == InvalidSize &&
+            destroy(invalid_request) == OK,
+        "NP authentication rejects missing and undersized user parameters");
+  code_parameters.size = sizeof(code_parameters);
+  token_parameters.size = sizeof(token_parameters);
+
+  Config::ConfigOptions options;
+  options.printf_direction = Config::LogDirection::Silent;
+  for (bool network : {false, true}) {
+    for (bool emulate : {false, true}) {
+      options.network_connected = network;
+      options.psn_emulation_enabled = emulate;
+      Config::Load(options);
+      for (int player = 0; player < Config::MAX_PLAYERS; ++player) {
+        code_parameters.user_id = Config::GetPlayerUserId(player);
+        token_parameters.user_id = code_parameters.user_id;
+        const int expected = emulate && player == 0 ? ServiceUnavailable : SignedOut;
+        for (bool async : {false, true}) {
+          for (bool id_token : {false, true}) {
+            const int request = async ? create_async(async_parameters.data()) : create();
+            Check(request > 0, "create NP authentication request");
+            code.fill(0xff);
+            token.fill(0xff);
+            issuer = -1;
+            const int result = id_token
+                ? get_token(request, &token_parameters, token.data())
+                : get_code(request, &code_parameters, code.data(), &issuer);
+            Check(result == (async ? OK : expected),
+                  "NP authentication distinguishes an unavailable service from sign-out");
+            if (async) {
+              int completion = 123;
+              Check(poll(request, &completion) == OK && completion == expected,
+                    "asynchronous NP authentication delivers the actual failure");
+            }
+            const auto zero = [] (uint8_t value) { return value == 0; };
+            Check(id_token ? std::all_of(token.begin(), token.end(), zero)
+                           : std::all_of(code.begin(), code.end(), zero) && issuer == 0,
+                  "PSN emulation does not fabricate authorization codes or ID tokens");
+            Check(destroy(request) == OK, "release NP authentication request");
+          }
+        }
+      }
+    }
+  }
+  Config::Load(Config::ConfigOptions {});
+}
+
 void CheckEtherAddressFormatting() {
   Loader::SymbolDatabase symbols;
   Libs::LibNet::InitNet_1_Net(&symbols);
@@ -2458,7 +2722,19 @@ int main(int argc, char** argv) {
   Config::Load(options);
   subsystems.Initialize<Log::Lifecycle>();
 
+  if (argc == 2 && std::string_view(argv[1]) == "--network-only") {
+#if defined(__linux__)
+    TestNetCtlHostIpv4();
+#endif
+    TestNpEmulation();
+    TestNpAuthEmulation();
+    subsystems.Destroy();
+    std::printf("KernelFileSystemTests: network cases passed\n");
+    return 0;
+  }
+
   TestMutexHandoff();
+  TestStaticRwlockDestroy();
   if (argc == 2 && std::string_view(argv[1]) == "--mutex-only") {
     subsystems.Destroy();
     std::printf("KernelFileSystemTests: mutex cases passed\n");
@@ -2503,6 +2779,8 @@ int main(int argc, char** argv) {
 #if defined(__linux__)
   TestNetCtlHostIpv4();
 #endif
+  TestNpEmulation();
+  TestNpAuthEmulation();
   TestNpWebApi2Memory();
   TestNpCommerceDialog();
   graphics.reset();

tests/Ngs2SamplerTests.cpp

@@ -751,6 +751,26 @@ void TestCustomSideChainUserFx() {
 	      "user FX did not receive the side-chain buffer after the source channels");
 }
 
+struct RenderRack {
+	std::unique_ptr<std::byte[]> storage;
+	uintptr_t handle = 0;
+
+	RenderRack(Ngs2Internal& system, uint32_t id, const Ngs2RackOption& option) {
+		Ngs2ContextBufferInfo buffer {};
+		Check(Ngs2RackQueryBufferSize(id, &option, &buffer) == OK, "rack size query failed");
+		storage = std::make_unique<std::byte[]>(buffer.host_buffer_size);
+		buffer.host_buffer = storage.get();
+		Check(Ngs2RackCreate(reinterpret_cast<uintptr_t>(&system), id, &option, &buffer,
+		                     &handle) == OK, "render rack creation failed");
+	}
+
+	~RenderRack() { Check(Ngs2RackDestroy(handle, nullptr) == OK, "render rack cleanup failed"); }
+
+	Ngs2VoiceInternal& Voice() {
+		return *reinterpret_cast<Ngs2VoiceInternal*>(reinterpret_cast<Ngs2RackInternal*>(handle) + 1);
+	}
+};
+
 void TestPcm16RenderOutput() {
 	Ngs2Internal system;
 	system.option                   = Ngs2DefaultSystemOption();
@@ -760,23 +780,11 @@ void TestPcm16RenderOutput() {
 	Ngs2FillDefaultRackOption(0x3000, &mastering_option);
 	sampler_option.common.max_voices = 2;
 	sampler_option.common.max_ports  = 1;
-	alignas(Ngs2RackInternal) alignas(Ngs2VoiceInternal)
-	    std::byte sampler_storage[sizeof(Ngs2RackInternal) + 2 * sizeof(Ngs2VoiceInternal)];
-	alignas(Ngs2RackInternal) alignas(Ngs2VoiceInternal)
-	    std::byte mastering_storage[sizeof(Ngs2RackInternal) + sizeof(Ngs2VoiceInternal)];
-	Ngs2ContextBufferInfo sampler_buffer {sampler_storage, sizeof(sampler_storage)};
-	Ngs2ContextBufferInfo mastering_buffer {mastering_storage, sizeof(mastering_storage)};
-	uintptr_t sampler_handle = 0, mastering_handle = 0;
+	RenderRack sampler(system, 0x1000, sampler_option.common);
+	RenderRack master(system, 0x3000, mastering_option.common);
 	const auto system_handle = reinterpret_cast<uintptr_t>(&system);
-	Check(Ngs2RackCreate(system_handle, 0x1000, &sampler_option.common, &sampler_buffer,
-	                     &sampler_handle) == OK &&
-	          Ngs2RackCreate(system_handle, 0x3000, &mastering_option.common,
-	                         &mastering_buffer, &mastering_handle) == OK,
-	      "PCM16 render rack creation failed");
-	auto* samplers = reinterpret_cast<Ngs2VoiceInternal*>(
-	    reinterpret_cast<Ngs2RackInternal*>(sampler_handle) + 1);
-	auto& mastering = *reinterpret_cast<Ngs2VoiceInternal*>(
-	    reinterpret_cast<Ngs2RackInternal*>(mastering_handle) + 1);
+	auto* samplers = &sampler.Voice();
+	auto& mastering = master.Voice();
 	mastering.channels = 1;
 	mastering.state    = Ngs2VoicePlayState::Playing;
 	const int16_t source[2][4] = {{16384, -16384, 32767, -32768},
@@ -797,9 +805,174 @@ void TestPcm16RenderOutput() {
 	Check(Ngs2SystemRender(system_handle, &output, 1) == OK &&
 	          std::ranges::all_of(output_samples, [](int16_t sample) { return sample == 0; }),
 	      "PCM16 output retained samples from the previous render");
-	Check(Ngs2RackDestroy(mastering_handle, nullptr) == OK &&
-	          Ngs2RackDestroy(sampler_handle, nullptr) == OK,
-	      "PCM16 render rack cleanup failed");
+}
+
+void TestPatchToRackHandle() {
+	Ngs2Internal system;
+	Ngs2RackOptionUnion sampler_option {}, mastering_option {};
+	Ngs2FillDefaultRackOption(0x1000, &sampler_option);
+	Ngs2FillDefaultRackOption(0x3000, &mastering_option);
+	sampler_option.common.max_voices = 1;
+	RenderRack sampler(system, 0x1000, sampler_option.common);
+	RenderRack master(system, 0x3000, mastering_option.common);
+	auto& source = sampler.Voice();
+	Ngs2VoicePatchParam patch {{24, 0, 5}, 0, 0, master.handle};
+	Check(Ngs2VoiceControl(reinterpret_cast<uintptr_t>(&source), &patch.header) == OK &&
+	          source.ports[0].dest == &master.Voice(),
+	      "patch to a rack handle did not connect to its voice #0");
+}
+
+void TestMasteringGain() {
+	Ngs2Internal system;
+	system.option                   = Ngs2DefaultSystemOption();
+	system.option.num_grain_samples = 1;
+	Ngs2RackOptionUnion sampler_option {}, mastering_option {};
+	Ngs2FillDefaultRackOption(0x1000, &sampler_option);
+	Ngs2FillDefaultRackOption(0x3000, &mastering_option);
+	sampler_option.common.max_voices = 1;
+	RenderRack sampler(system, 0x1000, sampler_option.common);
+	RenderRack master(system, 0x3000, mastering_option.common);
+	auto& source = sampler.Voice();
+	auto& mastering = master.Voice();
+	mastering.channels = 8;
+	mastering.state = Ngs2VoicePlayState::Playing;
+	const int16_t frame[8] = {8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192};
+	source.SetupSampler({0x12, 8, 48000});
+	source.blocks.push_back({reinterpret_cast<const uint8_t*>(frame), {0, sizeof(frame), 0, 0, 1}});
+	source.ports[0].dest = &mastering;
+	source.SetEvent(1);
+	struct {
+		Ngs2VoiceParamHeader header {16, 0, 0x30000004};
+		float fbw = 5.0f, lfe = 0.5f;
+	} gain;
+	float pcm[8] = {};
+	Ngs2RenderBufferInfo output {pcm, sizeof(pcm), 0x18, 8};
+	const float expected[8] = {1.0f, 1.0f, 1.0f, 0.125f, 1.0f, 1.0f, 1.0f, 1.0f};
+	Check(Ngs2VoiceControl(reinterpret_cast<uintptr_t>(&mastering), &gain.header) == OK &&
+	          Ngs2SystemRender(reinterpret_cast<uintptr_t>(&system), &output, 1) == OK &&
+	          std::ranges::equal(pcm, expected),
+	      "mastering gain was not clamped or applied per full band and LFE channel");
+}
+
+void TestReverbDryLevel() {
+	Fixture          f(48000);
+	Ngs2RackInternal reverb_rack {};
+	reverb_rack.ngs  = &f.system;
+	reverb_rack.type = Ngs2RackType::Reverb;
+	Ngs2VoiceInternal reverb;
+	reverb.rack = &reverb_rack;
+	f.voice.ports.push_back({&reverb, 0, 1.0f, -1});
+	const int16_t pcm = 8192;
+	f.Queue(&pcm, 1, 0);
+	struct {
+		Ngs2VoiceParamHeader header {24, 24, 0x20010000};
+		uint32_t             channels[4] = {1, 1, 0, 0};
+		Ngs2VoiceParamHeader levels_header {96, 0, 0x20010001};
+		float                levels[22] = {0.6f, 0.4f};
+	} setup;
+	Ngs2VoiceControl(reinterpret_cast<uintptr_t>(&reverb), &setup.header);
+	reverb.SetEvent(1);
+	Ngs2RenderVoice(reverb, {&f.voice, &reverb}, 1);
+	Check(reverb.has_samples && reverb.samples[0] == 0.25f * 0.4f,
+	      "reverb did not pass its dry signal");
+}
+
+void TestCustomMasteringChannelMixer() {
+	Ngs2Internal system;
+	system.option = Ngs2DefaultSystemOption();
+	system.option.num_grain_samples = 6;
+	Ngs2RackOptionUnion sampler_option {};
+	Ngs2FillDefaultRackOption(0x1000, &sampler_option);
+	sampler_option.common.max_voices = 1;
+	sampler_option.common.max_ports = 1;
+	Ngs2CustomMasteringRackOption master_option {};
+	auto& custom = master_option.custom_rack_option;
+	custom.rack_option.size = sizeof(master_option);
+	custom.rack_option.max_voices = 1;
+	custom.rack_option.max_voice_channels = 6;
+	custom.num_buffers = 1;
+	custom.num_modules = 1;
+	custom.module[0].module_id = 0x17;
+	RenderRack sampler(system, 0x1000, sampler_option.common);
+	RenderRack master(system, 0x4003, custom.rack_option);
+	auto& source = sampler.Voice();
+	auto& output = master.Voice();
+	output.channels = 6;
+	output.SetEvent(1);
+	struct {
+		Ngs2VoiceParamHeader header {5484, 0, 0x40001700};
+		float levels[37][37] {};
+	} matrix;
+	constexpr float center = 0.70710677f;
+	matrix.levels[0][0] = matrix.levels[1][1] = 1.0f;
+	matrix.levels[0][2] = matrix.levels[1][2] = center;
+	matrix.levels[0][4] = matrix.levels[1][5] = center;
+	int16_t impulses[6][6] {};
+	for (uint32_t c = 0; c < 6; ++c) impulses[c][c] = 8192;
+	for (unsigned pass = 0; pass < 2; ++pass) {
+		if (pass == 1) {
+			std::memset(matrix.levels, 0, sizeof(matrix.levels));
+			matrix.levels[0][1] = 64.0f;
+			matrix.levels[1][0] = -64.0f;
+		}
+		Check(Ngs2VoiceControl(reinterpret_cast<uintptr_t>(&output), &matrix.header) == OK,
+		      "custom channel mixer control failed");
+		source.SetupSampler({0x12, 6, 48000});
+		source.blocks.push_back({reinterpret_cast<const uint8_t*>(impulses),
+		                         {0, sizeof(impulses), 0, 0, 6}});
+		source.ports[0].dest = &output;
+		source.SetEvent(1);
+		std::array<float, 14> pcm;
+		pcm.fill(99.0f);
+		Ngs2RenderBufferInfo buffer {pcm.data() + 1, 12 * sizeof(float), 0x18, 2};
+		Check(Ngs2SystemRender(reinterpret_cast<uintptr_t>(&system), &buffer, 1) == OK,
+		      "custom mastering rejected stereo output");
+		const std::array<float, 12> expected = pass == 0
+		    ? std::array<float, 12> {0.25f, 0, 0, 0.25f, center / 4, center / 4,
+		                             0, 0, center / 4, 0, 0, center / 4}
+		    : std::array<float, 12> {0, -8, 8, 0};
+		Check(pcm.front() == 99 && pcm.back() == 99 &&
+		          std::equal(expected.begin(), expected.end(), pcm.begin() + 1),
+		      "custom mixer lost channel routing, in-place input, clamping or output bounds");
+	}
+}
+
+void TestCustomVoiceStateCapacity() {
+	for (auto type: {Ngs2RackType::CustomSampler, Ngs2RackType::CustomMastering}) {
+		Fixture f(48000, 1, Ngs2RackType::CustomSampler);
+		f.rack.type = type;
+		f.rack.option.custom_sampler.custom_rack_option.state_size = 212;
+		f.voice.state_flags = 3;
+		f.voice.decoded_samples = 37;
+		f.voice.decoded_bytes = 128;
+		alignas(Ngs2CustomSamplerVoiceState) std::array<uint8_t, 384> storage;
+		auto* state = reinterpret_cast<Ngs2VoiceState*>(storage.data() + 8);
+		const auto handle = reinterpret_cast<uintptr_t>(&f.voice);
+		for (size_t capacity: {212u, 368u}) {
+			storage.fill(0xa5);
+			Check(Ngs2VoiceGetState(handle, state, capacity) == OK && state->state_flags == 3 &&
+			          state->error_code == 0 &&
+			          std::all_of(storage.begin(), storage.begin() + 8,
+			                      [](uint8_t v) { return v == 0xa5; }) &&
+			          std::all_of(storage.begin() + 220, storage.end(),
+			                      [](uint8_t v) { return v == 0xa5; }),
+			      "oversized custom state buffer was rejected or overwritten past configured state");
+			if (type == Ngs2RackType::CustomSampler) {
+				const auto* sampler = reinterpret_cast<const Ngs2CustomSamplerVoiceState*>(state);
+				Check(sampler->num_decoded_samples == 37 && sampler->decoded_data_size == 128,
+				      "custom sampler state omitted playback progress");
+			}
+			const size_t base_size = type == Ngs2RackType::CustomSampler
+			    ? sizeof(Ngs2CustomSamplerVoiceState) : sizeof(Ngs2CustomMasteringVoiceState);
+			Check(std::all_of(storage.begin() + 8 + base_size, storage.begin() + 220,
+			                  [](uint8_t v) { return v == 0; }),
+			      "configured custom state extension was not initialized");
+		}
+		storage.fill(0xa5);
+		Check(Ngs2VoiceGetState(handle, state, 211) == static_cast<int32_t>(0x804a8011u) &&
+		          std::ranges::all_of(storage, [](uint8_t v) { return v == 0xa5; }),
+		      "undersized custom state buffer was accepted or modified");
+	}
 }
 
 void TestPitchLoopAndSkip() {
@@ -923,6 +1096,35 @@ void TestMonoRateAndRouting() {
 	      "22.05 kHz effect duration was incorrect");
 }
 
+void TestDefaultMatrix() {
+	struct Case {
+		uint32_t           src_channels;
+		int32_t            matrix;
+		uint32_t           dst_channels;
+		std::vector<float> expected;
+	};
+	constexpr float minus_3db = std::numbers::sqrt2_v<float> / 2;
+	const int16_t   pcm[]     = {8192, -16384};
+	for (const auto& c: {Case {1, -1, 2, {0.25f * minus_3db, 0.25f * minus_3db}},
+	                     Case {1, 0, 2, {0.25f * minus_3db, 0.25f * minus_3db}},
+	                     Case {2, -1, 1, {0.25f * minus_3db + -0.5f * minus_3db}},
+	                     Case {1, -1, 6, {0, 0, 0.25f, 0, 0, 0}}}) {
+		Fixture          f(48000, c.src_channels);
+		Ngs2RackInternal rack {};
+		rack.ngs  = &f.system;
+		rack.type = Ngs2RackType::Mastering;
+		Ngs2VoiceInternal dest;
+		dest.rack     = &rack;
+		dest.channels = c.dst_channels;
+		dest.state    = Ngs2VoicePlayState::Playing;
+		f.voice.matrices.resize(1);
+		f.voice.ports.push_back({&dest, 0, 1.0f, c.matrix});
+		f.Queue(pcm, 1, 0);
+		Ngs2RenderVoice(dest, {&f.voice, &dest}, 1);
+		Check(dest.samples == c.expected, "default port mix produced the wrong channel levels");
+	}
+}
+
 void TestMatrixOutputChannels() {
 	Fixture f(48000, 2);
 	Ngs2RackInternal mastering_rack {};
@@ -1008,6 +1210,38 @@ void TestDeferredSamplerSetup() {
 	      "sampler did not play after deferred stream setup");
 }
 
+void TestClearReopensClosedList() {
+	Fixture       f(48000);
+	const int16_t first = 8192, second = -16384;
+	f.Queue(&first, 1, 0);
+	f.Queue(&second, 1, 4);
+	f.Render(1);
+	Check(f.voice.samples[0] == -0.5f, "CLEAR did not reopen a closed waveform list");
+}
+
+void TestSetupQuery() {
+	Fixture  f(48000, 2);
+	uint32_t info[2] = {99, 99};
+	Ngs2VoiceQueryInfo(reinterpret_cast<uintptr_t>(&f.voice), 0x4001, info, sizeof(info));
+	Check(info[0] == 2 && info[1] == 0, "setup query did not report the voice channels");
+}
+
+void TestPortInfo() {
+	Fixture           f(48000);
+	Ngs2VoiceInternal dest;
+	f.voice.ports.push_back({&dest, 1, 0.5f, 0});
+	struct {
+		int32_t   matrix;
+		float     volume;
+		uint32_t  delay_samples, input;
+		uintptr_t dest;
+	} info {99, 99, 99, 99, 99};
+	Ngs2VoiceGetPortInfo(reinterpret_cast<uintptr_t>(&f.voice), 0, &info, sizeof(info));
+	Check(info.matrix == 0 && info.volume == 0.5f && info.delay_samples == 0 && info.input == 1 &&
+	          info.dest == reinterpret_cast<uintptr_t>(&dest),
+	      "port info did not report the port routing and volume");
+}
+
 void TestPlayStateBeforeRender() {
 	Fixture       f(48000);
 	const int16_t pcm[] = {8192, -16384};
@@ -1700,6 +1934,11 @@ int main() {
 	TestAtrac9Vibration();
 	TestAtrac9Discrete();
 	TestPcm16RenderOutput();
+	TestPatchToRackHandle();
+	TestMasteringGain();
+	TestReverbDryLevel();
+	TestCustomMasteringChannelMixer();
+	TestCustomVoiceStateCapacity();
 	TestSubmixerSetupAndRouting();
 	TestCustomSideChainUserFx();
 	TestWaveformBlockGeometry();
@@ -1732,8 +1971,12 @@ int main() {
 	TestPcmSkipAcrossBlocks();
 	TestPcmBlocksEndWithTheirData();
 	TestMonoRateAndRouting();
+	TestDefaultMatrix();
 	TestMatrixOutputChannels();
 	TestCustomPcmStillPlays();
+	TestClearReopensClosedList();
+	TestSetupQuery();
+	TestPortInfo();
 	TestPlayStateBeforeRender();
 	TestStatePublication();
 	TestStopThenPlayBeforeRender();

tests/PsmlTests.cpp

@@ -0,0 +1,346 @@
+#include "common/abi.h"
+#include "graphics/guest_gpu/pm4.h"
+#include "graphics/guest_gpu/pssr.h"
+#include "loader/symbolDatabase.h"
+
+#include <gtest/gtest.h>
+
+#include <algorithm>
+#include <array>
+#include <cstddef>
+#include <cstdint>
+#include <cstring>
+#include <vector>
+
+namespace Libs::LibPsml {
+void InitPsml_1(Loader::SymbolDatabase *symbols);
+}
+
+namespace {
+bool g_trinity = true;
+}
+
+namespace Loader {
+bool SystemContentIsTrinityMode() { return g_trinity; }
+} // namespace Loader
+
+namespace {
+
+namespace Pm4 = Libs::Graphics::Pm4;
+namespace Pssr = Libs::Graphics::Pssr;
+
+constexpr int ERROR_ALREADY_INITIALIZED = static_cast<int>(0x8a810002u);
+constexpr int ERROR_TOO_MANY_CONTEXTS = static_cast<int>(0x8a810004u);
+constexpr int ERROR_INVALID_CONTEXT = static_cast<int>(0x8a810005u);
+constexpr int ERROR_NO_SPACE = static_cast<int>(0x8a810006u);
+constexpr int ERROR_BAD_COMMAND_BUFFER = static_cast<int>(0x8a810007u);
+constexpr int ERROR_NULL_ARGUMENT = static_cast<int>(0x8a810009u);
+constexpr int ERROR_NULL_TEXTURE = static_cast<int>(0x8a81000au);
+constexpr int ERROR_INVALID_TEXTURE = static_cast<int>(0x8a81000bu);
+constexpr int ERROR_RENDER_SIZE = static_cast<int>(0x8a81000eu);
+constexpr int ERROR_OUTPUT_TEXTURE = static_cast<int>(0x8a81000fu);
+constexpr int ERROR_NOT_TRINITY = static_cast<int>(0x8a810012u);
+constexpr int ERROR_INVALID_SHARED = static_cast<int>(0x8a810014u);
+constexpr uint64_t CANARY = 0xfedcba9876543210;
+
+struct MemoryBlock {
+  uint64_t offset = 0;
+  uint64_t size = 0x200000;
+};
+
+struct SharedParameters {
+  uint32_t profile = 0;
+  uint32_t reserved = 0;
+  const MemoryBlock *blocks = nullptr;
+  uint64_t block_count = 0;
+  uint64_t virtual_address = 0;
+};
+
+struct ContextParameters {
+  uint32_t profile = 0;
+  uint32_t reserved = 0;
+  const MemoryBlock *blocks = nullptr;
+  uint64_t block_count = 0;
+  uint16_t width = 960;
+  uint16_t height = 540;
+  uint32_t reserved1 = 0;
+  const void *depth_copy = nullptr;
+  const void *motion_copy = nullptr;
+  const char *name = nullptr;
+  void *shared = nullptr;
+};
+static_assert(offsetof(ContextParameters, shared) == 0x38);
+
+struct CommandSpace {
+  uint32_t *cmd;
+  uint32_t dwords;
+};
+
+struct DispatchParameters {
+  const uint32_t *output;
+  const uint32_t *input;
+  std::array<uint8_t, 0xe0> unused{};
+  uint32_t width;
+  uint32_t height;
+};
+static_assert(offsetof(DispatchParameters, width) == 0xf0);
+
+struct GuardedHandle {
+  uint64_t before = CANARY;
+  void *value = nullptr;
+  std::array<uint64_t, 128> after;
+
+  GuardedHandle() { after.fill(CANARY); }
+
+  bool Intact() const {
+    return before == CANARY &&
+           std::all_of(after.begin(), after.end(),
+                       [](uint64_t value) { return value == CANARY; });
+  }
+};
+
+using Initialize = int(KYTY_SYSV_ABI *)();
+using CreateShared = int(KYTY_SYSV_ABI *)(void **, const SharedParameters *);
+using CreateContext = int(KYTY_SYSV_ABI *)(void **, const ContextParameters *);
+using Release = int(KYTY_SYSV_ABI *)(void *);
+using GetSize = int(KYTY_SYSV_ABI *)(const void *, uint32_t *);
+using Dispatch = int(KYTY_SYSV_ABI *)(const void *, const CommandSpace *,
+                                      const void *);
+using GetBias = int(KYTY_SYSV_ABI *)(const void *, float *, uint32_t, uint32_t);
+using Predicate = int(KYTY_SYSV_ABI *)(const void *);
+
+Loader::SymbolDatabase g_symbols;
+
+template <typename T> T Function(const char *nid) {
+  const auto *symbol = g_symbols.FindByNid(nid, Loader::SymbolType::Func);
+  EXPECT_NE(symbol, nullptr);
+  return symbol == nullptr ? nullptr : reinterpret_cast<T>(symbol->vaddr);
+}
+
+std::array<uint32_t, 8> Texture(uint32_t width, uint32_t height,
+                                uint32_t address) {
+  std::array<uint32_t, 8> result{};
+  result[0] = address;
+  result[1] = (((width - 1) & 3) << 30) | (36u << 20);
+  result[2] = ((width - 1) >> 2) | ((height - 1) << 14);
+  result[3] = (9u << 28) | (1u << 20) | 0xfacu;
+  return result;
+}
+
+class PsmlTest : public testing::Test {
+protected:
+  static void SetUpTestSuite() {
+    Libs::LibPsml::InitPsml_1(&g_symbols);
+    g_trinity = false;
+    ASSERT_EQ(Function<Initialize>("3WVD91e12ZQ")(), ERROR_NOT_TRINITY);
+    g_trinity = true;
+    ASSERT_EQ(Function<Initialize>("3WVD91e12ZQ")(), 0);
+  }
+
+  void SetUp() override {
+    SharedParameters params;
+    params.blocks = blocks.data();
+    params.block_count = blocks.size();
+    GuardedHandle result;
+    ASSERT_EQ(Function<CreateShared>("eWoKNeB6V-k")(&result.value, &params), 0);
+    ASSERT_TRUE(result.Intact());
+    ASSERT_GT(reinterpret_cast<uintptr_t>(result.value), 0x10000u);
+    shared = result.value;
+
+    context_params.blocks = blocks.data();
+    context_params.block_count = 1;
+    context_params.shared = shared;
+    result.value = nullptr;
+    ASSERT_EQ(
+        Function<CreateContext>("gxv3i+MTEzU")(&result.value, &context_params),
+        0);
+    ASSERT_TRUE(result.Intact());
+    ASSERT_GT(reinterpret_cast<uintptr_t>(result.value), 0x10000u);
+    context = result.value;
+  }
+
+  void TearDown() override {
+    for (auto *extra : extra_contexts) {
+      EXPECT_EQ(Function<Release>("JaLBe0P3jSU")(extra), 0);
+    }
+    if (context != nullptr) {
+      EXPECT_EQ(Function<Release>("JaLBe0P3jSU")(context), 0);
+    }
+    if (shared != nullptr) {
+      EXPECT_EQ(Function<Release>("jEevBXmagOQ")(shared), 0);
+    }
+  }
+
+  std::array<MemoryBlock, 52> blocks;
+  ContextParameters context_params;
+  void *shared = nullptr;
+  void *context = nullptr;
+  std::vector<void *> extra_contexts;
+};
+
+TEST_F(PsmlTest, HandlesPreserveAdjacentMemoryAndRejectReleasedObjects) {
+  EXPECT_EQ(Function<Initialize>("3WVD91e12ZQ")(), ERROR_ALREADY_INITIALIZED);
+  uint32_t size = 0;
+  EXPECT_EQ(Function<GetSize>("AHalTX9wFZY")(context, &size), 0);
+  EXPECT_EQ(size, Pssr::RequestDwords + 1);
+  EXPECT_EQ(Function<Release>("jEevBXmagOQ")(shared), ERROR_INVALID_CONTEXT);
+  EXPECT_EQ(Function<Predicate>("LXq+6mIxpCw")(context), 0);
+  EXPECT_EQ(Function<Release>("JaLBe0P3jSU")(context), 0);
+  EXPECT_EQ(Function<Release>("JaLBe0P3jSU")(context), ERROR_INVALID_CONTEXT);
+  EXPECT_EQ(Function<GetSize>("AHalTX9wFZY")(context, &size),
+            ERROR_INVALID_CONTEXT);
+  EXPECT_EQ(Function<Predicate>("LXq+6mIxpCw")(context), ERROR_INVALID_CONTEXT);
+  context = nullptr;
+  EXPECT_EQ(Function<Release>("jEevBXmagOQ")(shared), 0);
+  EXPECT_EQ(Function<Release>("jEevBXmagOQ")(shared), ERROR_INVALID_SHARED);
+  shared = nullptr;
+  EXPECT_EQ(
+      Function<GetSize>("AHalTX9wFZY")(reinterpret_cast<void *>(1), &size),
+      ERROR_INVALID_CONTEXT);
+}
+
+TEST_F(PsmlTest, LegacyContextVariantsUseTheirOwnLayoutsAndOutputResolution) {
+  context_params.shared = reinterpret_cast<void *>(1);
+  GuardedHandle result;
+  ASSERT_EQ(
+      Function<CreateContext>("gxv3i+MTEzU")(&result.value, &context_params),
+      ERROR_INVALID_SHARED);
+  ASSERT_TRUE(result.Intact());
+  ASSERT_EQ(result.value, nullptr);
+  ASSERT_EQ(
+      Function<CreateContext>("vUk2pWMx3KQ")(&result.value, &context_params),
+      0);
+  ASSERT_TRUE(result.Intact());
+  extra_contexts.push_back(result.value);
+  float bias = 0;
+  EXPECT_EQ(Function<GetBias>("GHna9-DvnUk")(result.value, &bias, 480, 270), 0);
+  EXPECT_FLOAT_EQ(bias, -2.0f);
+  ASSERT_EQ(
+      Function<CreateContext>("2ecEbQaf9VU")(&result.value, &context_params),
+      0);
+  ASSERT_TRUE(result.Intact());
+  extra_contexts.push_back(result.value);
+  EXPECT_EQ(Function<GetBias>("GHna9-DvnUk")(result.value, &bias, 1920, 1080),
+            0);
+  EXPECT_FLOAT_EQ(bias, -2.0f);
+}
+
+TEST_F(PsmlTest, CapsLiveContextsAndReleasesCapacity) {
+  for (int i = 0; i < 3; ++i) {
+    void *extra = nullptr;
+    ASSERT_EQ(Function<CreateContext>("gxv3i+MTEzU")(&extra, &context_params),
+              0);
+    extra_contexts.push_back(extra);
+  }
+  void *fifth = nullptr;
+  EXPECT_EQ(Function<CreateContext>("gxv3i+MTEzU")(&fifth, &context_params),
+            ERROR_TOO_MANY_CONTEXTS);
+  EXPECT_EQ(fifth, nullptr);
+  EXPECT_EQ(Function<Release>("JaLBe0P3jSU")(extra_contexts.back()), 0);
+  extra_contexts.pop_back();
+  ASSERT_EQ(Function<CreateContext>("gxv3i+MTEzU")(&fifth, &context_params), 0);
+  extra_contexts.push_back(fifth);
+}
+
+TEST_F(PsmlTest, DispatchSnapshotsFixedSlotsAndActiveExtent) {
+  auto output = Texture(960, 540, 0x10000);
+  auto input = Texture(1920, 1080, 0x20000);
+  DispatchParameters params{output.data(), input.data(), {}, 640, 360};
+  std::array<uint32_t, 32> command;
+  command.fill(0xdeadbeef);
+  const CommandSpace space{command.data() + 1, 28};
+  ASSERT_EQ(Function<Dispatch>("RUNLFro+qok")(context, &space, &params), 0);
+  EXPECT_EQ(KYTY_PM4_LEN(command[1]), space.dwords);
+  EXPECT_EQ(KYTY_PM4_R(command[1]), Pm4::R_PSSR);
+  EXPECT_EQ((command[1] >> 8) & 0xffu, Pm4::IT_NOP);
+  Pssr::Request request;
+  std::memcpy(&request, command.data() + 2, sizeof(request));
+  EXPECT_EQ(std::memcmp(request.output, output.data(), sizeof(request.output)),
+            0);
+  EXPECT_EQ(std::memcmp(request.input, input.data(), sizeof(request.input)), 0);
+  EXPECT_EQ(request.width, 640u);
+  EXPECT_EQ(request.height, 360u);
+  EXPECT_EQ(command.front(), 0xdeadbeefu);
+  EXPECT_EQ(command[29], 0xdeadbeefu);
+  for (size_t i = 2 + Pssr::RequestDwords; i < 29; ++i) {
+    EXPECT_EQ(command[i], 0u);
+  }
+  output.fill(0);
+  input.fill(0);
+  params = {};
+  Pssr::Request snapshot;
+  std::memcpy(&snapshot, command.data() + 2, sizeof(snapshot));
+  EXPECT_EQ(std::memcmp(&snapshot, &request, sizeof(request)), 0);
+}
+
+TEST_F(PsmlTest, DispatchAcceptsEqualSizedBuffersAndValidatesWithoutWriting) {
+  auto output = Texture(960, 540, 0x10000);
+  auto input = Texture(960, 540, 0x20000);
+  DispatchParameters params{output.data(), input.data(), {}, 960, 540};
+  uint32_t dwords = 0;
+  ASSERT_EQ(Function<GetSize>("AHalTX9wFZY")(context, &dwords), 0);
+  std::vector<uint32_t> command(dwords, 0xdeadbeef);
+  CommandSpace space{command.data(), dwords};
+  ASSERT_EQ(Function<Dispatch>("RUNLFro+qok")(context, &space, &params), 0);
+  std::fill(command.begin(), command.end(), 0xdeadbeef);
+  space.dwords = dwords - 1;
+  EXPECT_EQ(Function<Dispatch>("RUNLFro+qok")(context, &space, &params),
+            ERROR_NO_SPACE);
+  space.dwords = 0x4002;
+  EXPECT_EQ(Function<Dispatch>("RUNLFro+qok")(context, &space, &params),
+            ERROR_BAD_COMMAND_BUFFER);
+  space.dwords = dwords;
+  params.width = 961;
+  EXPECT_EQ(Function<Dispatch>("RUNLFro+qok")(context, &space, &params),
+            ERROR_RENDER_SIZE);
+  params.width = 960;
+  output = Texture(960, 539, 0x10000);
+  EXPECT_EQ(Function<Dispatch>("RUNLFro+qok")(context, &space, &params),
+            ERROR_OUTPUT_TEXTURE);
+  params.input = nullptr;
+  EXPECT_EQ(Function<Dispatch>("RUNLFro+qok")(context, &space, &params),
+            ERROR_NULL_TEXTURE);
+  EXPECT_EQ(Function<Dispatch>("RUNLFro+qok")(context, &space, nullptr),
+            ERROR_NULL_ARGUMENT);
+  EXPECT_TRUE(std::all_of(command.begin(), command.end(),
+                          [](uint32_t value) { return value == 0xdeadbeef; }));
+}
+
+TEST_F(PsmlTest, DispatchChecksDescriptorContractAndSelectedMip) {
+  auto output = Texture(960, 540, 0x10000);
+  auto input = Texture(1920, 1080, 0x20000);
+  input[3] |= (1u << 12) | (1u << 16);
+  input[5] |= 1u << 4;
+  DispatchParameters params{output.data(), input.data(), {}, 960, 540};
+  std::array<uint32_t, Pssr::RequestDwords + 1> command{};
+  const CommandSpace space{command.data(), command.size()};
+  ASSERT_EQ(Function<Dispatch>("RUNLFro+qok")(context, &space, &params), 0);
+  command.fill(0xdeadbeef);
+  params.width = 961;
+  EXPECT_EQ(Function<Dispatch>("RUNLFro+qok")(context, &space, &params),
+            ERROR_RENDER_SIZE);
+  params.width = 960;
+  input[1] &= ~(0x1ffu << 20);
+  EXPECT_EQ(Function<Dispatch>("RUNLFro+qok")(context, &space, &params),
+            ERROR_INVALID_TEXTURE);
+  input = Texture(960, 540, 0x20000);
+  input[3] = (input[3] & ~(0xfu << 28)) | (14u << 28);
+  EXPECT_EQ(Function<Dispatch>("RUNLFro+qok")(context, &space, &params),
+            ERROR_INVALID_TEXTURE);
+  input = Texture(960, 540, 0x20000);
+  output[3] &= ~(0x1fu << 20);
+  EXPECT_EQ(Function<Dispatch>("RUNLFro+qok")(context, &space, &params),
+            ERROR_OUTPUT_TEXTURE);
+  output = Texture(960, 540, 0x20000);
+  EXPECT_EQ(Function<Dispatch>("RUNLFro+qok")(context, &space, &params),
+            ERROR_OUTPUT_TEXTURE);
+  EXPECT_TRUE(std::all_of(command.begin(), command.end(),
+                          [](uint32_t value) { return value == 0xdeadbeef; }));
+}
+
+} // namespace
+
+int main(int argc, char **argv) {
+  testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}

tests/PssrGpuTests.inc

@@ -0,0 +1,228 @@
+  void CheckPssr() {
+    constexpr const char *name = "PssrLinear";
+    constexpr uintptr_t base = 0x0000000206a00000ull;
+    constexpr uint64_t size = 0x100000;
+    constexpr auto packed = Prospero::BufferFormat::k11_11_10Float;
+    constexpr auto linear = Prospero::TileMode::kLinear;
+    constexpr auto tiled = Prospero::TileMode::kStandard256B;
+    constexpr u32 red = 0x000003c0, green = 0x001e0000, blue = 0x78000000;
+    const auto descriptor = [](uint64_t address, u32 width, u32 height,
+                               Prospero::BufferFormat format, Prospero::TileMode tile,
+                               u32 swizzle = DstSel(4, 5, 6, 7)) {
+      ShaderTextureResource result{};
+      result.fields[0] = static_cast<u32>(address >> 8);
+      result.fields[1] = static_cast<u32>(address >> 40) |
+                         (static_cast<u32>(format) << 20) | (((width - 1) & 3) << 30);
+      result.fields[2] = ((width - 1) >> 2) | ((height - 1) << 14);
+      result.fields[3] = swizzle | (static_cast<u32>(tile) << 20) |
+                         (static_cast<u32>(Prospero::ImageType::kColor2D) << 28);
+      result.fields[5] = 0x00700000;
+      return result;
+    };
+    EnsureRuntimeContext();
+    GraphicsInitJmpTables();
+    int64_t physical = -1;
+    Require(name, "allocate", LibKernel::Memory::KernelAllocateDirectMemory(
+        0, LibKernel::Memory::KernelGetDirectMemorySize(), size, 0x10000, 0, &physical) == 0,
+        "upscaling allocation failed");
+    void *mapped = reinterpret_cast<void *>(base);
+    Require(name, "map", LibKernel::Memory::KernelMapDirectMemory(
+        &mapped, size, 0x3, 0x10, physical, 0x10000) == 0 &&
+            mapped == reinterpret_cast<void *>(base), "upscaling mapping failed");
+    std::memset(mapped, 0, size);
+    RenderContext context(m_runtime_context);
+    context.InitializeGpu(nullptr);
+    LibKernel::Memory::InstallGpuResources(&context);
+    context.GetGpu().SendCommandSync([&] {
+      CommandProcessor processor(context, 0);
+      processor.BufferInit();
+      context.MapMemory(base, size);
+      auto &executor = context.GetRenderExecutor();
+      auto &cache = context.GetTextureCache();
+      auto &scheduler = context.GetCommandScheduler();
+      const auto resolve = [&](const ShaderTextureResource &texture) {
+        ShaderRecompiler::IR::ImageResource resource{};
+        resource.resource_class = ShaderRecompiler::IR::ImageResourceClass::Sampled;
+        resource.numeric_class = Prospero::TextureNumericClass::Float;
+        resource.dimension = ShaderRecompiler::Decoder::ImageDimension::Dim2D;
+        resource.read = true;
+        ShaderRecompiler::IR::DescriptorValue value{};
+        value.dword_count = 8;
+        std::copy_n(texture.fields, 8, value.dwords.begin());
+        return executor.ResolveTexture(resource, value);
+      };
+      const auto write_linear = [&](const ShaderTextureResource &texture,
+                                    std::span<const u32> texels) {
+        auto info = MakeTilingImage(texture.Format(), texture.Width5() + 1u,
+            texture.Height5() + 1u, 1, 1, linear, 0, false);
+        const u32 words_per_texel = info.bytes_per_block / sizeof(u32);
+        Require(name, "source size", texels.size() ==
+            info.extent.width * info.extent.height * words_per_texel,
+            "source texel fixture size is wrong");
+        std::vector<u32> words(info.data.size / sizeof(u32));
+        for (u32 y = 0; y < info.extent.height; ++y) {
+          std::copy_n(texels.begin() + y * info.extent.width * words_per_texel,
+              info.extent.width * words_per_texel,
+              words.begin() + y * info.mip_layout[0].pitch * words_per_texel);
+        }
+        Require(name, "invalidate source", context.InvalidateMemory(
+            texture.Base40(), info.data.size), "source fixture is outside mapped memory");
+        LibKernel::Memory::WriteBacking(texture.Base40(), words.data(), info.data.size);
+      };
+      const auto run = [&](const ShaderTextureResource &input,
+                            const ShaderTextureResource &output, u32 width, u32 height) {
+        Pssr::Request request{};
+        std::copy_n(input.fields, 8, request.input);
+        std::copy_n(output.fields, 8, request.output);
+        request.width = width;
+        request.height = height;
+        // The reserved packet includes trailing words that the parser must consume.
+        constexpr u32 reserved_words = Pssr::RequestDwords + 9;
+        std::vector<u32> packet(reserved_words, 0xdeadbeef);
+        packet[0] = KYTY_PM4(reserved_words, Pm4::IT_NOP, Pm4::R_PSSR);
+        std::memcpy(packet.data() + 1, &request, sizeof(request));
+        u32 sentinel = 0;
+        const auto address = reinterpret_cast<uint64_t>(&sentinel);
+        packet.insert(packet.end(), {KYTY_PM4(5, Pm4::IT_WRITE_DATA, 0), 0,
+            static_cast<u32>(address), static_cast<u32>(address >> 32), 33});
+        Pm4Execution execution;
+        Require(name, "packet and suffix", processor.Process(execution, packet) ==
+            Pm4ProcessResult::Complete && sentinel == 33,
+            "upscaling packet did not consume its reserved payload before the suffix");
+        const auto binding = resolve(output);
+        return ReadCachedTexel(name, context, binding.image_id, {},
+            {output.Width5() + 1u, output.Height5() + 1u, 1},
+            binding.desc.view_info.base_layer, binding.desc.view_info.base_level);
+      };
+
+      auto input = descriptor(base, 2, 1, packed, linear);
+      auto output = descriptor(base + 0x10000, 4, 1, packed, tiled);
+      write_linear(input, std::array<u32, 2>{red, green});
+      Require(name, "CPU-first bilinear source", run(input, output, 2, 1) ==
+          std::vector<u32>{red, 0x3a0 | (0x340 << 11), 0x340 | (0x3a0 << 11), green},
+          "first-use source upload or linear interpolation is wrong");
+      write_linear(input, std::array<u32, 2>{blue, blue});
+      Require(name, "CPU source refresh", run(input, output, 2, 1) ==
+          std::vector<u32>(4, blue), "reused source retained stale GPU texels");
+
+      input = descriptor(base + 0x20000, 2, 1, packed, linear);
+      output = descriptor(base + 0x30000, 2, 1, packed, tiled);
+      write_linear(input, std::array<u32, 2>{red, green});
+      Require(name, "same-size copy", run(input, output, 2, 1) ==
+          std::vector<u32>{red, green}, "same-size upscaling lost source pixels");
+
+      input = descriptor(base + 0x40000, 8, 4, packed, linear);
+      output = descriptor(base + 0x50000, 8, 4, packed, tiled);
+      std::vector<u32> cropped(32, blue);
+      cropped[0] = cropped[1] = cropped[8] = cropped[9] = red;
+      write_linear(input, cropped);
+      Require(name, "active rectangle", run(input, output, 2, 2) ==
+          std::vector<u32>(32, red),
+          "equal allocation extents bypassed scaling or sampled outside the active rectangle");
+
+      input = descriptor(base + 0x60000, 2, 1, packed, linear, DstSel(6, 5, 4, 7));
+      output = descriptor(base + 0x70000, 2, 1, packed, tiled);
+      write_linear(input, std::array<u32, 2>{red, green});
+      Require(name, "sampled view swizzle", run(input, output, 2, 1) ==
+          std::vector<u32>{blue, green}, "sampled descriptor swizzle was ignored");
+      output.fields[3] = (output.fields[3] & ~0xfffu) | DstSel(6, 5, 4, 7);
+      Require(name, "storage swizzle", run(input, output, 2, 1) ==
+          std::vector<u32>{red, green}, "output channel mapping was ignored");
+
+      input = descriptor(base + 0x80000, 2, 1,
+          Prospero::BufferFormat::k16_16_16_16Float, linear);
+      output = descriptor(base + 0x90000, 2, 1, packed, tiled);
+      write_linear(input, std::array<u32, 4>{0x00003c00, 0x3c000000,
+                                            0x3c000000, 0x3c000000});
+      Require(name, "float16 input", run(input, output, 2, 1) ==
+          std::vector<u32>{red, green}, "float16 input conversion failed");
+
+      ImageDesc parent{};
+      parent.type = BindingType::Texture;
+      parent.info = MakeTilingImage(packed, 4, 2, 1, 2, tiled, 0, false);
+      parent.info.data.address = base + 0xa0000;
+      parent.info.pixel_format = vk::Format::eB10G11R11UfloatPack32;
+      parent.view_info.format = parent.info.pixel_format;
+      parent.view_info.type = vk::ImageViewType::e2DArray;
+      parent.view_info.layer_count = 2;
+      const auto parent_id = cache.FindImage(parent);
+      (void)cache.FindTexture(parent_id, parent);
+      vk::ClearValue clear{};
+      clear.color.float32 = std::array<float, 4>{1, 0, 0, 1};
+      TextureCacheTestAccess::ClearImage(cache, scheduler.Current(), parent_id,
+          {vk::ImageAspectFlagBits::eColor, 0, 1, 0, 1}, clear);
+      clear.color.float32 = std::array<float, 4>{0, 0, 1, 1};
+      TextureCacheTestAccess::ClearImage(cache, scheduler.Current(), parent_id,
+          {vk::ImageAspectFlagBits::eColor, 0, 1, 1, 1}, clear);
+      input = descriptor(parent.info.data.address, 4, 2, packed, tiled);
+      output = descriptor(parent.info.data.address + parent.info.tiled_slice_stride,
+          4, 2, packed, tiled);
+      const auto resolved_output = resolve(output);
+      Require(name, "parent array resolution", resolved_output.image_id == parent_id &&
+          resolved_output.desc.view_info.base_layer == 1,
+          "standalone output did not resolve into the existing array's second layer");
+      Require(name, "resolved output layer", run(input, output, 4, 2) ==
+          std::vector<u32>(8, red), "upscaling wrote layer zero instead of the resolved layer");
+      Require(name, "source layer preserved", ReadCachedTexel(name, context, parent_id,
+          {}, {4, 2, 1}) == std::vector<u32>(8, red),
+          "upscaling modified the source layer of its shared native image");
+      clear.color.float32 = std::array<float, 4>{0, 1, 0, 1};
+      TextureCacheTestAccess::ClearImage(cache, scheduler.Current(), parent_id,
+          {vk::ImageAspectFlagBits::eColor, 0, 1, 1, 1}, clear);
+      Require(name, "resolved source layer", run(output, input, 4, 2) ==
+          std::vector<u32>(8, green), "upscaling read layer zero instead of the resolved layer");
+
+      input = descriptor(base + 0xc0000, 2, 1, packed, linear);
+      output = descriptor(base + 0xd0000, 2, 1, packed, tiled);
+      write_linear(input, std::array<u32, 2>{red, green});
+      ImageDesc raw{};
+      raw.type = BindingType::Texture;
+      raw.info = MakeTilingImage(Prospero::BufferFormat::k32UInt, 2, 1, 1, 1,
+                                 linear, 0, false);
+      raw.info.data.address = input.Base40();
+      raw.info.pixel_format = vk::Format::eR32Uint;
+      raw.view_info.format = raw.info.pixel_format;
+      const auto raw_id = cache.FindImage(raw);
+      (void)cache.FindTexture(raw_id, raw);
+      const auto float_binding = resolve(input);
+      Require(name, "reinterpreted sampled view", float_binding.image_id == raw_id &&
+          cache.GetImage(raw_id).backing.format == vk::Format::eR32Uint &&
+          float_binding.desc.view_info.format == vk::Format::eB10G11R11UfloatPack32,
+          "float sampled descriptor did not reuse the integer native allocation");
+      Require(name, "sampled view format", run(input, output, 2, 1) ==
+          std::vector<u32>{red, green}, "upscaling used the allocation format instead of its view");
+
+      ImageDesc mip_parent{};
+      mip_parent.type = BindingType::Texture;
+      mip_parent.info = MakeTilingImage(packed, 128, 64, 2, 1, tiled, 0, false);
+      mip_parent.info.data.address = base + 0xe0000;
+      mip_parent.info.pixel_format = vk::Format::eB10G11R11UfloatPack32;
+      mip_parent.view_info.format = mip_parent.info.pixel_format;
+      mip_parent.view_info.level_count = 2;
+      const auto mip_parent_id = cache.FindImage(mip_parent);
+      (void)cache.FindTexture(mip_parent_id, mip_parent);
+      clear.color.float32 = std::array<float, 4>{0, 0, 1, 1};
+      TextureCacheTestAccess::ClearImage(cache, scheduler.Current(), mip_parent_id,
+          {vk::ImageAspectFlagBits::eColor, 0, 2, 0, 1}, clear);
+      clear.color.float32 = std::array<float, 4>{0, 1, 0, 1};
+      TextureCacheTestAccess::ClearImage(cache, scheduler.Current(), mip_parent_id,
+          {vk::ImageAspectFlagBits::eColor, 1, 1, 0, 1}, clear);
+      input = descriptor(mip_parent.info.data.address + mip_parent.info.mip_layout[1].offset,
+          64, 32, packed, tiled);
+      output = descriptor(base + 0xf0000, 64, 32, packed, tiled);
+      const auto resolved_mip = resolve(input);
+      Require(name, "parent mip resolution", resolved_mip.image_id == mip_parent_id &&
+          resolved_mip.desc.view_info.base_level == 1,
+          "standalone source did not resolve into the existing image's second mip");
+      Require(name, "resolved source mip", run(input, output, 64, 32) ==
+          std::vector<u32>(64 * 32, green),
+          "upscaling read mip zero instead of the resolved mip");
+      context.UnmapMemory(base, size);
+      scheduler.Finish();
+    });
+    LibKernel::Memory::InstallGpuResources(nullptr);
+    context.ShutdownGpu();
+    Require(name, "release", LibKernel::Memory::KernelCheckedReleaseDirectMemory(
+        physical, size) == 0, "upscaling allocation release failed");
+    std::printf("[gpu]     %-32s ok\n", name);
+  }

tests/ShaderRayTracingGpuTests.inc

@@ -1,3 +1,435 @@
+TestCase TrinityBvhIntersections(bool compressed, bool bary, bool sorted,
+                                 bool wide_sort, u32 heuristic = 0, bool triangles_first = false,
+                                 u32 wave_size = 32, bool edges = false) {
+  TestCase test;
+  test.name = compressed ? (sorted ? "TrinityCompressedSorted" : "TrinityCompressedUnsorted")
+                         : (sorted ? "TrinityFp32Sorted" : "TrinityFp32Unsorted");
+  if (!bary) test.name = "TrinityTriangleIds";
+  if (heuristic == 1) test.name = "TrinityLargestInterval";
+  if (heuristic == 2) test.name = "TrinityClosestMidpoint";
+  if (triangles_first) test.name = "TrinityTrianglesFirst";
+  if (wave_size == 64) test.name = "TrinityBvhWave64";
+  if (edges) test.name = compressed ? "TrinityCompressedEdges" : "TrinityFp32Edges";
+  constexpr u32 NodeStart = 4096, OutputStart = 2048, GuestBase = 0x10000;
+  constexpr u32 Invalid = ~0u, Infinity = 0x7f800000, Flag = 0x8484, Shape = 0x12345678;
+  const auto bits = [](float x) { return std::bit_cast<u32>(x); };
+  test.initial.resize(NodeStart + 256 * 16);
+  auto& data = test.initial;
+  const std::array<float, 12> vertices {-1,-1,2, 1,-1,2, 0,1,2, 2,1,2};
+  for (u32 i = 0; i < vertices.size(); ++i) data[NodeStart + i] = bits(vertices[i]);
+  data[NodeStart + 14] = Shape;
+  data[NodeStart + 15] = Flag;
+  const std::array<float, 12> matrix {2,0,0,1, 0,3,0,-2, 0,0,4,3};
+  for (u32 i = 0; i < matrix.size(); ++i) data[NodeStart + 16 + i] = bits(matrix[i]);
+  data[NodeStart + 28] = 0x1000;
+  data[NodeStart + 29] = 9u << 22; // Force opaque and flip facing.
+  data[NodeStart + 30] = 0x98;
+  data[NodeStart + 31] = 0x04123456;
+  const std::array<float, 8> nearDist {5,1,3,7,4,2,6,8};
+  const std::array<float, 8> length {1,8,3,5,7,2,4,6};
+  const std::array<u32, 8> children {128,138,171,190,256,266,299,318};
+  const auto set_bits = [&](u32 base, u32 offset, u32 count, u32 value) {
+    for (u32 bit = 0; bit < count; ++bit)
+      data[base + (offset + bit) / 32] |= ((value >> bit) & 1) << ((offset + bit) % 32);
+  };
+  for (u32 slot = 0; slot < 2; ++slot) {
+    const u32 base = NodeStart + (3 + slot * (compressed ? 1 : 2)) * 16;
+    if (compressed) {
+      set_bits(base, 0, 29, slot ? 32 : 16);
+      for (u32 i = 0; i < 4; ++i) set_bits(base, 29 + i * 3, 3, children[slot * 4 + i] & 7);
+      set_bits(base, 56, 8, 127); set_bits(base, 64, 8, 127); set_bits(base, 72, 8, 130);
+    }
+    for (u32 i = 0; i < 4; ++i) {
+      const u32 index = slot * 4 + i;
+      if (compressed) {
+        // Exact integer minima; positive maxima round outward within one quantization step.
+        const std::array<u32, 6> planes {0x30000, 0x30000, u32(nearDist[index]) * 8192,
+            0x10000, 0x10000, u32(nearDist[index] + length[index]) * 8192};
+        for (u32 axis = 0; axis < 6; ++axis) set_bits(base, 80 + i * 108 + axis * 18, 18, planes[axis]);
+      } else {
+        data[base + i] = children[index];
+        const std::array<float, 6> bounds {-1,-1,nearDist[index], 1,1,nearDist[index] + length[index]};
+        for (u32 axis = 0; axis < 6; ++axis) data[base + 4 + i * 6 + axis] = bits(bounds[axis]);
+      }
+    }
+  }
+  if (compressed) {
+    std::copy_n(data.begin() + NodeStart + 3 * 16, 16, data.begin() + NodeStart + 10 * 16);
+  } else {
+    constexpr std::array<u32, 4> lo {0x4500, 0x3c00, 0x4200, 0x4700};
+    constexpr std::array<u32, 4> hi {0x4600, 0x4880, 0x4600, 0x4a00};
+    for (u32 i = 0; i < 4; ++i) {
+      data[NodeStart + 160 + i] = children[i];
+      data[NodeStart + 164 + i * 3] = 0xbc00bc00;
+      data[NodeStart + 165 + i * 3] = 0x3c000000 | lo[i];
+      data[NodeStart + 166 + i * 3] = (hi[i] << 16) | 0x3c00;
+    }
+  }
+  if (edges) {
+    // Slot zero ends exactly at the descriptor limit; slot one lies beyond it.
+    std::copy_n(data.begin() + NodeStart + 3 * 16, compressed ? 16 : 32,
+                data.begin() + NodeStart + (compressed ? 253 : 252) * 16);
+    for (u32 index : {12u, 13u})
+      std::copy_n(data.begin() + NodeStart, 16, data.begin() + NodeStart + index * 16);
+    data[NodeStart + 12 * 16 + 15] = 0xc4c4; // Opaque procedural pair.
+    data[NodeStart + 13 * 16 + 15] = 0x9494; // Opaque, double-sided pair.
+    if (compressed) {
+      std::copy_n(data.begin() + NodeStart + 3 * 16, 16, data.begin() + NodeStart + 11 * 16);
+      set_bits(NodeStart + 11 * 16, 41, 4, 1);
+      set_bits(NodeStart + 11 * 16, 45, 4, 2);
+      set_bits(NodeStart + 11 * 16, 49, 4, 3);
+    }
+  }
+  std::vector<std::array<u32, 16>> expected(wave_size);
+  for (u32 lane = 0; lane < wave_size; ++lane) {
+    const u32 which = lane % 16, input = lane * 16;
+    auto& out = expected[lane];
+    out.fill(Invalid);
+    data[input + 2] = bits(20); data[input + 3] = 4; data[input + 9] = bits(1);
+    const auto miss = [&](u32 slot, bool invalid = false) {
+      out[slot * 4] = Infinity; out[slot * 4 + 1] = bits(1);
+      out[slot * 4 + 2] = !invalid && bary ? 0 : Invalid; out[slot * 4 + 3] = 0;
+    };
+    const auto hit = [&](u32 slot, float u, float v) {
+      out[slot * 4] = bits(-8); out[slot * 4 + 1] = bits(-4);
+      out[slot * 4 + 2] = bary ? bits(u) : Flag + slot;
+      out[slot * 4 + 3] = bary ? bits(v) : 1;
+    };
+    if (which <= 2 || which >= 9 && which <= 14) {
+      data[input + 10] = which == 1 || which == 2 ? 1 : 0;
+      miss(0); miss(1);
+      out[8] = Shape;
+      if (which == 1) {
+        data[input + 4] = bits(1); hit(1, 2, 1); out[9] = Shape;
+      } else {
+        hit(0, 1, 1);
+        if (which == 2) { miss(1); out[9] = Shape; }
+      }
+      if (which == 9 || which == 10 || which == 11) {
+        if (which == 9) data[input + 4] = 0x7fc00000;
+        if (which == 10) data[input + 10] = 0xfff8;
+        if (which == 11) data[input + 1] = 0x41u << 22;
+        miss(0, true); miss(1, true); out[8] = out[9] = Invalid;
+      } else if (which == 12) {
+        data[input + 1] = 0x100u << 22; miss(0);
+      } else if (which == 13) {
+        data[input + 1] = 8u << 22; out[0] = bits(8); out[1] = bits(4);
+      } else if (which == 14) {
+        data[input + 1] = 2u << 22;
+        if (bary) { out[2] = bits(-1); out[3] = bits(-1); }
+      }
+    } else if (which >= 3 && which <= 5) {
+      data[input + 10] = which == 3 ? (3 * 8 + (compressed ? 3 : 2))
+                           : which == 4 ? (3 * 8 + (compressed ? 4 : 5)) : (10 * 8 + 4);
+      std::array<u32, 8> order {0,1,2,3,4,5,6,7};
+      const auto compare = [&](u32 x, u32 y) {
+        const bool xt = (children[x] & 7) < 2, yt = (children[y] & 7) < 2;
+        if (triangles_first && xt != yt) return xt;
+        if (heuristic == 1) return length[x] > length[y];
+        if (heuristic == 2) return 2 * nearDist[x] + length[x] < 2 * nearDist[y] + length[y];
+        return nearDist[x] < nearDist[y];
+      };
+      if (sorted) {
+        if (wide_sort && which == 3) std::stable_sort(order.begin(), order.end(), compare);
+        else {
+          std::stable_sort(order.begin(), order.begin() + 4, compare);
+          std::stable_sort(order.begin() + 4, order.end(), compare);
+        }
+      }
+      for (u32 i = 0; i < (which == 3 ? 8u : 4u); ++i) out[i] = children[order[i]];
+    } else if (which == 6 || which == 7 || which == 15) {
+      data[input + 10] = 14;
+      if (which == 15) { data[input] = 8; data[input + 10] = 6; }
+      data[input + 4] = bits(1); data[input + 5] = bits(2); data[input + 6] = bits(3);
+      if (which == 7) data[input + 3] = 2;
+      for (u32 i = 0; i < 10; ++i) out[i] = 0xabc00000 + i;
+      out[2] = 0x1000; out[3] = 9u << 22; out[6] = 0x123456; out[7] = which == 7 ? Invalid : 0x98;
+    } else data[input + 10] = 7;
+    if (edges) {
+      switch (which) {
+        case 0:
+        case 1:
+        case 2:
+        case 5:
+          if (which == 0) data[input] = 1; // Misaligned encoded base must not read across a block.
+          if (which == 1) data[input + 4] = Infinity;
+          if (which == 2) data[input + 2] = 0x7fc00000;
+          if (which == 5) { data[input + 1] = 0x2000; data[input + 10] = 0; }
+          miss(0, true); miss(1, true); out[8] = out[9] = Invalid;
+          break;
+        case 3:
+          data[input + 10] = compressed ? 253 * 8 + 3 : 252 * 8 + 2;
+          for (u32 i = 4; i < 8; ++i) out[i] = Invalid;
+          break;
+        case 4:
+          if (compressed) {
+            data[input + 10] = 11 * 8 + 4; data[input + 1] = 0x40u << 22;
+            out[0] = out[2] = Invalid;
+          } else {
+            data[input + 2] = bits(1);
+            for (u32 i = 0; i < 4; ++i) out[i] = Invalid;
+          }
+          break;
+        case 6:
+          data[input + 1] = 8u << 22; out[3] = 1u << 22;
+          break;
+        case 7:
+          data[input + 1] = 2u << 22; out[3] = 10u << 22;
+          break;
+        case 9:
+        case 10:
+          data[input + 10] = 12 * 8;
+          data[input + 4] = bits(100);
+          out[8] = Shape; miss(0); miss(1);
+          if (which == 9) { out[0] = bits(0); out[1] = bits(1); out[2] = bits(-1); out[3] = bits(1); }
+          else data[input + 1] = 0x200u << 22;
+          break;
+        case 11:
+          data[input + 1] = 0x10u << 22; miss(0); miss(1); out[8] = Shape;
+          break;
+        case 12:
+        case 13:
+          data[input + 1] = (which == 12 ? 0x14u : 0x10u) << 22;
+          if (which == 13) data[input + 10] = 13 * 8;
+          hit(0, 1, 1);
+          break;
+        case 14:
+          // The compressed +X bound rounds outward beyond exactly one.
+          if (compressed) {
+            data[input + 1] = 0; data[input + 4] = bits(1.0f + 0x1p-17f);
+            data[input + 10] = 3 * 8 + 4;
+            out.fill(Invalid);
+            for (u32 i = 0; i < 4; ++i) out[i] = children[i];
+          }
+          break;
+      }
+    }
+    for (u32 i = 0; i < 6; ++i) out[10 + i] = data[input + 4 + i];
+    if (which == 6 || which == 7 || which == 15) {
+      out[10] = bits(3); out[11] = bits(4); out[12] = bits(15); out[15] = bits(4);
+    }
+    if (lane % 32 >= 16) {
+      for (u32 i = 0; i < 10; ++i) out[i] = 0xabc00000 + i;
+      for (u32 i = 0; i < 6; ++i) out[10 + i] = data[input + 4 + i];
+    }
+  }
+  test.expected = data;
+  for (u32 lane = 0; lane < wave_size; ++lane)
+    std::copy(expected[lane].begin(), expected[lane].end(), test.expected.begin() + OutputStart + lane * 16);
+  auto& code = test.code;
+  AppendSMovLiteral(&code, 8, GuestBase >> 8);
+  AppendSMovLiteral(&code, 9, (sorted ? 1u << 31 : 0) | (triangles_first ? 1u << 20 : 0) | (heuristic << 21));
+  AppendSMovLiteral(&code, 10, edges ? 253 : 255);
+  AppendSMovLiteral(&code, 11, 0x80c00000 | (bary ? 1u << 24 : 0) |
+      (compressed ? 1u << 20 : 0) | (wide_sort ? 1u << 21 : 0));
+  code.push_back(EncodeVop2(0x1a, 1, InlineU32(6), 0));
+  for (u32 offset : {0u,4u,8u})
+    code.insert(code.end(), {EncodeMubuf0(offset == 8 ? 0x0f : 0x0e, offset * 4), EncodeMubuf1(4 + offset, 0, 1)});
+  for (u32 i = 0; i < 10; ++i) AppendVMovLiteral(&code, 32 + i, 0xabc00000 + i);
+  AppendSMovLiteral(&code, 126, 0x0000ffff);
+  if (wave_size == 64) AppendSMovLiteral(&code, 127, 0x0000ffff);
+  code.insert(code.end(), {EncodeMimg0(0xe5, 15, 2, false, 0, true), EncodeMimg1(32, 4, 2),
+      0x08070605, 0x00000e0b});
+  code.push_back(EncodeSop1(0x04, 126, 193));
+  code.push_back(EncodeVop2(0x1a, 2, InlineU32(4), 0));
+  for (u32 i = 0; i < 10; ++i) AppendStoreVgprAtLaneDwordOffset(&code, 32 + i, 2, OutputStart + i);
+  for (u32 i = 0; i < 6; ++i) AppendStoreVgprAtLaneDwordOffset(&code, 8 + i, 2, OutputStart + 10 + i);
+  AppendEnd(&code);
+  test.bda_mappings = {{GuestBase, NodeStart * sizeof(u32)}};
+  test.compute_info.wave_size = wave_size;
+  test.compute_info.threads_num[0] = wave_size;
+  test.compute_info.threads_num[1] = test.compute_info.threads_num[2] = 1;
+  test.compute_info.thread_ids_num = 1;
+  test.has_compute_info = true;
+  test.decoded_counts = {{"IMAGE_BVH8_INTERSECT_RAY", 1}};
+  test.required_spirv = {"trinity_bvh_intersect"};
+  return test;
+}
+
+TestCase TrinityBvhAliases() {
+  auto test = TrinityBvhIntersections(false, true, false, false);
+  test.name = "TrinityBvhRegisterAliases";
+  constexpr u32 OutputStart = 2048;
+  for (u32 lane = 0; lane < 32; ++lane) {
+    auto* out = test.expected.data() + OutputStart + lane * 16;
+    const auto* input = test.initial.data() + lane * 16;
+    const bool active = lane < 16, instance = active && (lane == 6 || lane == 7 || lane == 15);
+    if (!active) {
+      for (u32 i = 0; i < 6; ++i) out[4 + i] = input[i];
+    } else if (instance) {
+      out[4] = input[0]; out[5] = input[1];
+      out[8] = out[10]; out[9] = out[11];
+    } else {
+      // A non-instance writes its shape IDs into registers formerly holding ray XY.
+      out[10] = out[8]; out[11] = out[9];
+    }
+  }
+  auto& code = test.code;
+  code.clear();
+  AppendSMovLiteral(&code, 8, 0x100);
+  AppendSMovLiteral(&code, 9, 0);
+  AppendSMovLiteral(&code, 10, 255);
+  AppendSMovLiteral(&code, 11, 0x81c00000);
+  code.push_back(EncodeVop2(0x1a, 50, InlineU32(6), 0));
+  code.push_back(EncodeVop2(0x1a, 51, InlineU32(4), 0));
+  for (u32 i = 0; i < 10; ++i) AppendVMovLiteral(&code, i, 0xabc00000 + i);
+  for (u32 offset : {0u, 4u, 8u})
+    code.insert(code.end(), {EncodeMubuf0(offset == 8 ? 0x0f : 0x0e, offset * 4),
+        EncodeMubuf1(4 + offset, 0, 50)});
+  AppendSMovLiteral(&code, 126, 0x0000ffff);
+  code.insert(code.end(), {EncodeMimg0(0xe5, 15, 2, false, 0, true), EncodeMimg1(0, 4, 2),
+      0x08070605, 0x00000e0b});
+  code.push_back(EncodeSop1(0x04, 126, 193));
+  for (u32 i = 0; i < 10; ++i) AppendStoreVgprAtLaneDwordOffset(&code, i, 51, OutputStart + i);
+  for (u32 i = 0; i < 6; ++i) AppendStoreVgprAtLaneDwordOffset(&code, 8 + i, 51, OutputStart + 10 + i);
+  AppendEnd(&code);
+  return test;
+}
+
+std::vector<TestCase> TrinityBvhCases() {
+  return {TrinityBvhIntersections(false, true, false, false),
+          TrinityBvhIntersections(false, true, true, true),
+          TrinityBvhIntersections(true, true, false, false),
+          TrinityBvhIntersections(true, true, true, false),
+          TrinityBvhIntersections(false, false, true, true),
+          TrinityBvhIntersections(false, true, true, true, 1),
+          TrinityBvhIntersections(false, true, true, true, 2),
+          TrinityBvhIntersections(true, true, true, true, 0, true),
+          TrinityBvhIntersections(true, true, true, true, 0, false, 64),
+          TrinityBvhIntersections(false, true, false, false, 0, false, 32, true),
+          TrinityBvhIntersections(true, true, false, false, 0, false, 32, true),
+          TrinityBvhAliases()};
+}
+
+TestCase TrinityStack(u32 wave_size, u32 capacity, bool storage = false) {
+  TestCase test;
+  static constexpr const char* names[2][3] = {
+      {"TrinityStack32Capacity8", "TrinityStack32Capacity16", "TrinityStack32Capacity31"},
+      {"TrinityStack64Capacity8", "TrinityStack64Capacity16", "TrinityStack64Capacity31"}};
+  test.name = storage ? (wave_size == 32 ? "TrinityStackStorage32" : "TrinityStackStorage64")
+                      : names[wave_size == 64][capacity == 8 ? 0 : capacity == 16 ? 1 : 2];
+  constexpr u32 Invalid = 0xffffffffu, Terminal = 0xfffffffeu;
+  constexpr u32 Parent = 1u << 29, Overflow = 1u << 30, Return = 1u << 31;
+  const auto state = [](u32 count, u32 parent, u32 ring, u32 flags = 0) {
+    return count | (parent << 5) | (ring << 10) | flags;
+  };
+  test.initial.resize(wave_size * 32);
+  struct Output { u32 control, node; std::array<u32, 8> lds; };
+  std::vector<Output> outputs(wave_size);
+  for (u32 lane = 0; lane < wave_size; ++lane) {
+    const u32 which = lane % 16;
+    const u32 base = which == 15 ? 0x3fff : (lane / 32) * capacity * 32 + lane % 32;
+    auto* input = test.initial.data() + lane * 32;
+    auto& out = outputs[lane];
+    input[0] = base << 15;
+    input[1] = Invalid;
+    std::fill(input + 2, input + 10, Invalid);
+    input[10] = base * 4;
+    for (u32 i = 0; i < 8; ++i) out.lds[i] = 100 + i * 8;
+    if (capacity == 8) out.lds[7] = 88;
+    out.node = Terminal;
+    out.control = 0;
+    const auto children = [&] {
+      for (u32 i = 0; i < 8; ++i) input[2 + i] = (i + 1) * 8;
+    };
+    switch (which) {
+      case 0: break;
+      case 1: input[0] |= Overflow; out.control = Overflow; out.node = Invalid; break;
+      case 2:
+        children(); out.control = state(7, 7, 7); out.node = 8;
+        out.lds = {64, 56, 48, 40, 32, 24, 16, 8}; break;
+      case 3:
+        input[2] = 8; input[4] = 24; input[7] = 48; input[9] = 64;
+        out.control = state(3, 3, 3); out.node = 8;
+        out.lds[0] = 64; out.lds[1] = 48; out.lds[2] = 24; out.lds[3] = 8; break;
+      case 4:
+      case 14:
+        input[0] |= state(capacity, capacity, 0, which == 14 ? Parent : 0);
+        input[2] = 8; out.control = state(capacity - 1, capacity - 1, 0, Overflow);
+        out.node = out.lds[0] = 8; break;
+      case 5:
+        input[0] |= state(1, 1, 0); out.control = state(0, 0, capacity - 1); out.node = 88; break;
+      case 6:
+        input[0] |= state(1, 0, 1, Parent);
+        out.control = state(0, 31, 0, Return); out.node = 100; break;
+      case 7:
+        input[2] = 14; out.control = Parent; out.node = out.lds[0] = 14; break;
+      case 8:
+        input[0] |= state(3, 1, 3, Parent);
+        out.lds[2] = 14;
+        out.control = state(0, 0, 2, Parent | Overflow); out.node = 14; break;
+      case 9:
+        children(); input[1] = 24; out.control = state(4, 4, 4); out.node = 32;
+        out.lds[0] = 64; out.lds[1] = 56; out.lds[2] = 48; out.lds[3] = 40; out.lds[4] = 32; break;
+      case 10:
+        children(); input[1] = 96; input[0] |= state(1, 1, 1);
+        out.node = 100; break;
+      case 11:
+        children(); input[9] = 0xfffffffbu; out.control = state(3, 3, 3); out.node = 8;
+        out.lds[0] = 32; out.lds[1] = 24; out.lds[2] = 16; out.lds[3] = 8; break;
+      case 12: children(); input[9] = 0xfffffff9u; break;
+      case 13:
+        children(); input[5] = 0xfffffffdu; out.control = state(3, 3, 3); out.node = 40;
+        out.lds[0] = 64; out.lds[1] = 56; out.lds[2] = 48; out.lds[3] = 40; break;
+      case 15:
+        input[0] |= state(1, 1, 1); out.node = 0; out.lds.fill(0); break;
+    }
+    out.control |= base << 15;
+    if ((lane % 32) >= 16) {
+      out.control = input[0]; out.node = input[2];
+      for (u32 i = 0; i < 8; ++i) out.lds[i] = 100 + i * 8;
+      if (capacity == 8) out.lds[7] = 88;
+      if (which == 8) out.lds[2] = 14;
+      if (which == 15) out.lds.fill(0);
+    }
+  }
+  test.expected = test.initial;
+  for (u32 lane = 0; lane < wave_size; ++lane) {
+    test.expected[lane * 32 + 20] = outputs[lane].control;
+    test.expected[lane * 32 + 21] = outputs[lane].node;
+    std::copy(outputs[lane].lds.begin(), outputs[lane].lds.end(), test.expected.begin() + lane * 32 + 24);
+  }
+  auto& code = test.code;
+  code.push_back(EncodeVop2(0x1a, 1, InlineU32(7), 0));
+  code.insert(code.end(), {EncodeMubuf0(0x0d), EncodeMubuf1(8, 0, 1),
+                           EncodeMubuf0(0x0e, 8), EncodeMubuf1(16, 0, 1),
+                           EncodeMubuf0(0x0e, 24), EncodeMubuf1(20, 0, 1),
+                           EncodeMubuf0(0x0c, 40), EncodeMubuf1(2, 0, 1)});
+  for (u32 i = 0; i < 8; ++i) {
+    AppendVMovU32(&code, 3, 100 + i * 8);
+    code.insert(code.end(), {EncodeDs0(0x0d, i * 128), EncodeDs1(0, 3, 2)});
+  }
+  AppendVMovU32(&code, 3, 88);
+  code.insert(code.end(), {EncodeDs0(0x0d, (capacity - 1) * 128), EncodeDs1(0, 3, 2)});
+  // Lane eight starts with a nested instance on top of its stack.
+  AppendSMovLiteral(&code, 126, 0x01000100u);
+  if (wave_size == 64) AppendSMovLiteral(&code, 127, 0x01000100u);
+  AppendVMovU32(&code, 3, 14);
+  code.insert(code.end(), {EncodeDs0(0x0d, 256), EncodeDs1(0, 3, 2)});
+  AppendSMovLiteral(&code, 126, 0x0000ffffu);
+  if (wave_size == 64) AppendSMovLiteral(&code, 127, 0x0000ffffu);
+  code.insert(code.end(), {EncodeDs0(0xe1, capacity), EncodeDs1Ex(16, 16, 9, 8)});
+  code.push_back(EncodeSop1(0x04, 126, 193));
+  code.push_back(EncodeVop2(0x1a, 3, InlineU32(5), 0));
+  AppendStoreVgprAtLaneDwordOffset(&code, 8, 3, 20);
+  AppendStoreVgprAtLaneDwordOffset(&code, 16, 3, 21);
+  for (u32 i = 0; i < 8; ++i) {
+    code.insert(code.end(), {EncodeDs0(0x36, i * 128), EncodeDs1(24, 0, 2)});
+    AppendStoreVgprAtLaneDwordOffset(&code, 24, 3, 24 + i);
+  }
+  AppendEnd(&code);
+  test.compute_info.wave_size = wave_size;
+  test.compute_info.threads_num[0] = wave_size;
+  test.compute_info.threads_num[1] = test.compute_info.threads_num[2] = 1;
+  test.compute_info.thread_ids_num = 1;
+  test.compute_info.lds_size_dwords = wave_size * capacity;
+  test.compute_info.lds_storage = storage;
+  test.has_compute_info = true;
+  test.decoded_counts = {{"DS_BVH_STACK_PUSH8_POP1_RTN_B32", 1}};
+  test.forbidden_spirv = {"OpControlBarrier", "OpMemoryBarrier", "get_bda_pointer"};
+  return test;
+}
+
 TestCase FlatStackApertures(u32 wave_size) {
   TestCase test;
   test.name = wave_size == 32 ? "FlatStackApertures32" : "FlatStackApertures64";

tests/ShaderRayTracingTests.inc

@@ -1,4 +1,86 @@
+void TestTrinityStackInstruction() {
+  namespace D = ShaderRecompiler::Decoder;
+  const std::array code{0xdb840008u, 0x01010a09u,
+                        0xdb840010u, 0x02010a09u, EncodeSopp(0x01)};
+  D::Program decoded;
+  D::DecodeProgram(code, decoded);
+  const auto& first = decoded.instructions.front();
+  Check(first.opcode == D::Opcode::DS_BVH_STACK_PUSH8_POP1_RTN_B32 &&
+            first.src_count == 3 && first.offset == 8 && first.dst.reg == 1 &&
+            first.src0.reg == 9 && first.src1.reg == 10 && first.src2.reg == 1,
+        "Trinity stack operands or capacity were not decoded");
+  ShaderComputeInputInfo compute{};
+  compute.lds_size_dwords = 1024;
+  auto options = MakeCompileOptions(ShaderType::Compute);
+  options.input_info.compute = &compute;
+  const auto result = RecompileForTest(code, options);
+  Check(CountSourceOccurrences(result.ir_dump, " = BvhStack ") == 2 &&
+            result.program.info.uses_lds && result.program.info.bvh_stack &&
+            !result.program.info.shared_int64_atomics,
+        "unused stack results discarded LDS writes or selected 64-bit atomics");
+  Check(SpirvInstructionOpcodeCount(result.spirv, 54u) == 2 &&
+            SpirvInstructionOpcodeCount(result.spirv, 246u) == 1,
+        "Trinity stack calls duplicated the function or unrolled its eight pushes");
+  CheckSpirvBinaryValidates(result.spirv);
+  compute.lds_storage = true;
+  CheckSpirvBinaryValidates(RecompileForTest(code, options).spirv);
+  options = MakeCompileOptions(ShaderType::Pixel);
+  CheckSpirvBinaryValidates(RecompileForTest(code, options).spirv);
+  for (const auto words : {std::array{0xdb860008u, 0x01010a09u},
+                           std::array{0xdb840007u, 0x01010a09u},
+                           std::array{0xdb840020u, 0x01010a09u},
+                           std::array{0xdb840008u, 0x01f90a09u},
+                           std::array{0xdb840008u, 0x09010a09u}}) {
+    const std::array invalid{words[0], words[1], EncodeSopp(0x01)};
+    D::DecodeProgram(invalid, decoded);
+    Check(decoded.instructions[0].opcode == D::Opcode::UNSUPPORTED,
+          "unsupported stack encoding was silently accepted");
+  }
+}
+
+void TestTrinityBvhInstruction() {
+  namespace D = ShaderRecompiler::Decoder;
+  const std::array code{0xf1949f05u, 0x00010000u, 0x0a030201u, 0x0000040du,
+      EncodeMubuf0(0x1e, 0, false), EncodeMubuf1(0, 0, 0),
+      EncodeMubuf0(0x1e, 16, false), EncodeMubuf1(4, 0, 0),
+      EncodeMubuf0(0x1d, 32, false), EncodeMubuf1(8, 0, 0),
+      EncodeMubuf0(0x1e, 40, false), EncodeMubuf1(10, 0, 0),
+      EncodeMubuf0(0x1d, 56, false), EncodeMubuf1(14, 0, 0), EncodeSopp(0x01)};
+  D::Program decoded;
+  D::DecodeProgram(code, decoded);
+  const auto& inst = decoded.instructions.front();
+  Check(inst.opcode == D::Opcode::IMAGE_BVH8_INTERSECT_RAY && inst.word_count == 4 &&
+            inst.data_dwords == 10 && inst.image_address_components == 11 &&
+            inst.image_nsa_addr[3] == 10 && inst.image_nsa_addr[5] == 12 &&
+            inst.image_nsa_addr[6] == 13 && inst.image_nsa_addr[8] == 15 &&
+            inst.image_nsa_addr[9] == 4 && inst.src1.reg == 4,
+        "Trinity BVH8 grouped ray registers were not decoded");
+  const auto result = RecompileForTest(code, MakeCompileOptions(ShaderType::Compute));
+  Check(result.program.info.bvh8 && !result.program.info.bvh &&
+            result.program.info.uses_dma && result.program.info.images.empty(),
+        "Trinity BVH8 did not use its separate raw-memory helper");
+  Check(CountSourceOccurrences(result.ir_dump, " = Bvh8Intersect ") == 1,
+        "Trinity BVH8 result and ray writes duplicated the intersection");
+  CheckSpirvBinaryValidates(result.spirv);
+  Check(SpirvInstructionOpcodeCount(result.spirv, 54u) == 6 &&
+            SpirvInstructionOpcodeCount(result.spirv, 246u) == 1,
+        "Trinity BVH8 duplicated helpers or unrolled its four-child loop");
+  CheckSpirvBinaryValidates(RecompileForTest(code, MakeCompileOptions(ShaderType::Pixel)).spirv);
+  for (const auto words : {std::array{0xf1949f05u, 0x40010000u, 0x0a030201u, 0x0000040du},
+                           std::array{0xf1949f05u, 0x0001f700u, 0x0a030201u, 0x0000040du},
+                           std::array{0xf1949f05u, 0x00010000u, 0xfe030201u, 0x0000040du},
+                           std::array{0xf1949f05u, 0x00010700u, 0x0a030201u, 0x0000040du},
+                           std::array{0xf1949f05u, 0x00010000u, 0x0a030201u, 0x0000040bu}}) {
+    const std::array invalid{words[0], words[1], words[2], words[3], EncodeSopp(0x01)};
+    D::DecodeProgram(invalid, decoded);
+    Check(decoded.instructions[0].opcode == D::Opcode::UNSUPPORTED,
+          "Trinity BVH8 accepted truncated ray or destination registers");
+  }
+}
+
 void TestRayTracingInstructions() {
+  TestTrinityBvhInstruction();
+  TestTrinityStackInstruction();
   namespace D = ShaderRecompiler::Decoder;
   const std::array code{
       0xf1989f01u, 0x00081014u,

tests/TrinityTests.cpp

@@ -0,0 +1,129 @@
+#include "common/emulatorConfig.h"
+#include "common/logging/log.h"
+#include "common/subsystems.h"
+#include "common/threads.h"
+#include "common/virtualMemory.h"
+#include "graphics/guest_gpu/pm4.h"
+#include "kernel/memory.h"
+#include "libs/agc.h"
+#include "loader/symbolDatabase.h"
+#include "loader/systemContent.h"
+
+#include <gtest/gtest.h>
+
+#include <array>
+#include <filesystem>
+#include <fstream>
+
+namespace Libs {
+void InitLibKernel_1(Loader::SymbolDatabase*);
+}
+
+namespace {
+
+namespace Pm4 = Libs::Graphics::Pm4;
+
+TEST(Trinity, ModeQueriesPacketsAndMemoryAgree) {
+	static Common::Subsystems subsystems;
+	Common::VirtualMemory::Init();
+	Common::InitializeThreads();
+	subsystems.Initialize<Config::Lifecycle>();
+	Config::ConfigOptions options;
+	options.printf_direction = Config::LogDirection::Silent;
+	Config::Load(options);
+	subsystems.Initialize<Log::Lifecycle>();
+	Loader::SymbolDatabase symbols;
+	Libs::InitLibKernel_1(&symbols);
+	const auto* query = symbols.FindByNid("tU5e3f9gSiU", Loader::SymbolType::Func);
+	const auto* operation = symbols.FindByNid("NH6xARDOVv8", Loader::SymbolType::Func);
+	ASSERT_NE(query, nullptr);
+	ASSERT_NE(operation, nullptr);
+	const auto is_trinity = reinterpret_cast<int(KYTY_SYSV_ABI*)()>(query->vaddr);
+	const auto operation_mode = reinterpret_cast<int(KYTY_SYSV_ABI*)(int*, int*)>(operation->vaddr);
+
+	struct CommandBuffer {
+		uint32_t* bottom;
+		uint32_t* top;
+		uint32_t* cursor_up;
+		uint32_t* cursor_down;
+		void* callback;
+		void* user_data;
+		uint32_t reserved_dw;
+	};
+	const auto path = std::filesystem::temp_directory_path() /
+	                  ("kyty_trinity_" + std::to_string(reinterpret_cast<uintptr_t>(&symbols)) + ".json");
+	struct Case { bool enabled; uint32_t attribute; bool expected; };
+	for (const auto test: {Case {false, 0, false}, Case {false, 1u << 22, false}, Case {true, 0, false},
+	                       Case {true, 1u << 21, false}, Case {true, 1u << 22, true}}) {
+		options.pro_enabled = test.enabled;
+		Config::Load(options);
+		{
+			std::ofstream file(path);
+			file << "{\"attribute3\":" << test.attribute
+			     << R"(,"kernel":{"flexibleMemorySize":3221225472,"cpuPageTableSize":134217728,"gpuPageTableSize":0},"amm":{"pagetableMemorySizeInMib":96}})";
+			ASSERT_TRUE(file.good());
+		}
+		testing::internal::CaptureStdout();
+		Loader::SystemContentLoadParamSfo(path);
+		std::filesystem::remove(path);
+		EXPECT_EQ(Loader::SystemContentIsTrinityMode(), test.expected);
+		EXPECT_EQ(is_trinity(), test.expected ? 1 : 0);
+		int mode = -1, submode = -1;
+		EXPECT_EQ(operation_mode(&mode, &submode), 0);
+		EXPECT_EQ(mode, test.expected ? 3 : 2);
+		EXPECT_EQ(submode, 0);
+		uint8_t graphics_mode = 0xff;
+		Libs::Graphics::Gen5::AgcGetIsTrinityMode(&graphics_mode);
+		EXPECT_EQ(graphics_mode, test.expected ? 1 : 0);
+		std::array<uint32_t, 5> packet {};
+		CommandBuffer buffer {packet.data(), packet.data() + packet.size(), packet.data(),
+		                      packet.data() + packet.size(), nullptr, nullptr, 0};
+		Libs::Graphics::Gen5::AgcDcbSetUcRegistersIndirect(
+		    reinterpret_cast<Libs::Graphics::Gen5::CommandBuffer*>(&buffer), nullptr, 0);
+		EXPECT_EQ(packet[0], KYTY_PM4(5, Pm4::IT_SET_UCONFIG_REG_INDIRECT,
+		                              test.expected ? 1 : 0));
+		const auto console = testing::internal::GetCapturedStdout();
+		if (test.enabled) {
+			EXPECT_EQ(console, test.expected ? "[PS5] title Trinity support=yes\n"
+			                                 : "[PS5] title Trinity support=no\n");
+		} else {
+			EXPECT_TRUE(console.empty()) << console;
+		}
+	}
+
+	using namespace Libs::LibKernel::Memory;
+	subsystems.Initialize<Lifecycle>();
+	constexpr uint64_t mib = 1024 * 1024;
+	constexpr uint64_t backing = (12800 + 1216 - 64 - 96) * mib;
+	constexpr uint64_t flexible = 3072 * mib;
+	EXPECT_EQ(TestGuestBackingSize(), backing);
+	EXPECT_EQ(KernelGetDirectMemorySize(), backing - flexible);
+	size_t configured = 0;
+	ASSERT_EQ(KernelConfiguredFlexibleMemorySize(&configured), 0);
+	EXPECT_EQ(configured, flexible);
+	int64_t available_start = -1;
+	size_t available_size = 0;
+	ASSERT_EQ(KernelAvailableDirectMemorySize(0, backing - flexible, 0x4000,
+	                                         &available_start, &available_size), 0);
+	EXPECT_EQ(available_start, 0);
+	EXPECT_EQ(available_size, backing - flexible);
+	int64_t physical = -1;
+	ASSERT_EQ(KernelAllocateDirectMemory(backing - flexible - 0x4000, backing - flexible,
+	                                     0x4000, 0x4000, 0, &physical), 0);
+	void* mapped = nullptr;
+	ASSERT_EQ(KernelMapDirectMemory(&mapped, 0x4000, 2, 0, physical, 0), 0);
+	*static_cast<volatile uint64_t*>(mapped) = 0x123456789abcdef0ull;
+	EXPECT_EQ(*static_cast<volatile uint64_t*>(mapped), 0x123456789abcdef0ull);
+	EXPECT_EQ(KernelMunmap(reinterpret_cast<uint64_t>(mapped), 0x4000), 0);
+	EXPECT_EQ(KernelCheckedReleaseDirectMemory(physical, 0x4000), 0);
+}
+
+} // namespace
+
+int RunTrinityTests() {
+	int argc = 1;
+	char name[] = "trinity_tests";
+	char* argv[] = {name, nullptr};
+	testing::InitGoogleTest(&argc, argv);
+	return RUN_ALL_TESTS();
+}

tests/VirtualMemoryAllocationTests.cpp

@@ -5708,7 +5708,12 @@ void TestTlsStubsPreserveFlags() {
 }
 #endif
 
+int RunTrinityTests();
+
 int main(int argc, char** argv) {
+	if (argc == 2 && std::strcmp(argv[1], "--trinity-only") == 0) {
+		return RunTrinityTests();
+	}
 	InitSubsystems();
 	if (argc == 2 && std::strcmp(argv[1], "--memory-bottlenecks-only") == 0) {
 		RunTest(TestFlexibleZeroingTracksBackingLifetime);

tests/shaderCfgTests.cpp

@@ -1938,7 +1938,7 @@ void TestVertexBufferGrouping() {
   Check(input.fetch_attrib_reg == 2, "a draw reread the vertex metadata of a mapped shader");
   ShaderMapUserData(regs.es_regs.data_addr, mapped);
   (void)PrepareProgram(regs, HW::Context{}, HW::UserConfig{}, input);
-  Check(!input.fetch_embedded, "remapping a shader kept its previous vertex metadata");
+  Check(!input.has_dynamic_vertex_fetch, "remapping a shader kept its previous vertex metadata");
 }
 
 void TestNggVertexEntryState() {
@@ -2397,6 +2397,50 @@ void TestNewShaderRecompilerVop3LaneReadDestinationEncoding() {
   CheckSpirvBinaryValidates(result.spirv);
 }
 
+void TestVop3FloatConversionDecode() {
+  namespace Decoder = ShaderRecompiler::Decoder;
+  const std::pair<uint32_t, Decoder::Opcode> conversions[] = {
+      {0x18b, Decoder::Opcode::V_CVT_F32_F16},
+      {0x191, Decoder::Opcode::V_CVT_F32_UBYTE0},
+      {0x192, Decoder::Opcode::V_CVT_F32_UBYTE1},
+      {0x193, Decoder::Opcode::V_CVT_F32_UBYTE2},
+      {0x194, Decoder::Opcode::V_CVT_F32_UBYTE3},
+  };
+  for (const auto &[encoding, opcode] : conversions) {
+    for (const uint32_t source : {261u, 12u, 128u, 255u}) {
+      const uint32_t code[] = {
+          EncodeVop3Word0(encoding, 20, 0, 1, true),
+          EncodeVop3Word1(source, 0, 0) | (2u << 27u) | (1u << 29u),
+          0xa5a53e00u,
+      };
+      Decoder::Instruction inst;
+      Decoder::DecodeInstruction(code, 0, inst);
+      Check(inst.opcode == opcode && inst.family == Decoder::Family::VOP3 &&
+                inst.src_count == 1 && inst.word_count == (source == 255u ? 3u : 2u),
+            "VOP3 conversion lost its opcode, unary source, or literal length");
+      Check(inst.dst.kind == Decoder::OperandKind::Vgpr && inst.dst.reg == 20 &&
+                inst.dst.sdwa_sel == 6 && inst.dst.clamp && inst.dst.omod == 2,
+            "VOP3 conversion lost its full-width destination or output modifiers");
+      const bool half = opcode == Decoder::Opcode::V_CVT_F32_F16;
+      Check(inst.src0.absolute == half && inst.src0.negate == half,
+            "VOP3 conversion must apply half-float modifiers and ignore byte modifiers");
+      Decoder::Operand expected_source;
+      Decoder::DecodeScalarSource(source, 0, expected_source);
+      Check(inst.src0.kind == expected_source.kind && inst.src0.reg == expected_source.reg &&
+                (source != 255u || inst.src0.value == code[2]),
+            "VOP3 conversion changed its source register or literal value");
+    }
+    for (const uint32_t op_sel : {1u, 2u, 4u, 8u}) {
+      const uint32_t code[] = {EncodeVop3Word0(encoding, 20, op_sel),
+                               EncodeVop3Word1(261, 0, 0)};
+      Decoder::Instruction inst;
+      Decoder::DecodeInstruction(code, 0, inst);
+      Check(inst.opcode == Decoder::Opcode::UNSUPPORTED,
+            "VOP3 conversion accepted unsupported OP_SEL bits");
+    }
+  }
+}
+
 void TestNewShaderRecompilerMoreAluFamilies() {
   const uint32_t shader[] = {
       EncodeSopk(0x00, 9, 7), // s_movk_i32 s9, 7
@@ -10550,6 +10594,43 @@ void TestNewShaderRecompilerNativeU64Translation() {
         "captured VOP3 V_LSHLREV_B64 was not present in the decoded dump");
   CheckSpirvBinaryValidates(result.spirv);
 
+  const uint32_t ashrrev_i64_shader[] = {
+      0xd7010001u, 0x0000d50cu,
+      EncodeSopp(0x01),
+  };
+  Decoder::Instruction decoded_ashrrev_i64;
+  ShaderRecompiler::Decoder::DecodeInstruction(ashrrev_i64_shader, 0u,
+                                               decoded_ashrrev_i64);
+  Check(decoded_ashrrev_i64.opcode == Decoder::Opcode::V_ASHRREV_I64 &&
+            decoded_ashrrev_i64.dst.kind == Decoder::OperandKind::Vgpr &&
+            decoded_ashrrev_i64.dst.reg == 1u &&
+            decoded_ashrrev_i64.src_count == 2u &&
+            decoded_ashrrev_i64.src0.kind == Decoder::OperandKind::Vgpr &&
+            decoded_ashrrev_i64.src0.reg == 12u &&
+            decoded_ashrrev_i64.src1.kind == Decoder::OperandKind::VccLo,
+        "decoder rejected VOP3 V_ASHRREV_I64 fields");
+  result = RecompileForTest(ashrrev_i64_shader, options);
+  Check((result.decoded_dump.find("V_ASHRREV_I64 v1, v12, vcc_lo") != std::string::npos),
+        "VOP3 V_ASHRREV_I64 was not present in the decoded dump");
+  CheckSpirvBinaryValidates(result.spirv);
+
+  const uint32_t ashrrev_captured[] = {0xd7010007u, 0x00020498u, EncodeSopp(0x01)};
+  Decoder::Instruction decoded_ashrrev_captured;
+  ShaderRecompiler::Decoder::DecodeInstruction(ashrrev_captured, 0u,
+                                               decoded_ashrrev_captured);
+  Check(decoded_ashrrev_captured.opcode == Decoder::Opcode::V_ASHRREV_I64 &&
+            decoded_ashrrev_captured.family == Decoder::Family::VOP3 &&
+            decoded_ashrrev_captured.src_count == 2u &&
+            decoded_ashrrev_captured.dst.kind == Decoder::OperandKind::Vgpr &&
+            decoded_ashrrev_captured.dst.reg == 7u &&
+            decoded_ashrrev_captured.src0.kind == Decoder::OperandKind::IntegerInlineConstant &&
+            decoded_ashrrev_captured.src0.value == 24u &&
+            decoded_ashrrev_captured.src1.kind == Decoder::OperandKind::Vgpr &&
+            decoded_ashrrev_captured.src1.reg == 2u,
+        "captured VOP3 V_ASHRREV_I64 opcode 0x301 was not decoded");
+  result = RecompileForTest(ashrrev_captured, options);
+  CheckSpirvBinaryValidates(result.spirv);
+
   const uint32_t cmpx_i64_shader[] = {
       0xd4b5007eu, 0x00020e80u, // v_cmpx_ne_i64 exec, 0, v[7:8]
       EncodeSopp(0x01),
@@ -11575,7 +11656,7 @@ void TestEmbeddedFetchPreservesSharedScalarLoad() {
                              reinterpret_cast<uint64_t>(attributes.data())};
   std::memcpy(user_data.data() + 12, tables, sizeof(tables));
   ShaderVertexInputInfo input{};
-  input.fetch_embedded = true;
+  input.has_dynamic_vertex_fetch = true;
   input.fetch_buffer_reg = 12;
   input.fetch_attrib_reg = 14;
   input.resources_num = 1;
@@ -17121,6 +17202,7 @@ int main() {
   using namespace Libs::Graphics;
 
   EnsureConfigInitialized();
+  TestVop3FloatConversionDecode();
   TestRayTracingInstructions();
   TestResourceDescriptorClassification();
   TestShaderBufferResourceSize();