Fix HITMAN 3 boot, layered presentation, and socket imports (#851)

* videoout: synchronize layered flips with premultiplied composition

* shader: support unsigned FFBH subword sources

* shader: apply cube-face result scaling

* shader: prune buffer reads proven out of bounds

* net: implement send and receive exports with shared socket handling

---------

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

CMakeLists.txt

@@ -301,6 +301,7 @@ set(gpu_blit_generated_dir "${PROJECT_BINARY_DIR}/gpu_blit_shaders")
 set(gpu_blit_shader_sources
 	"${gpu_tiler_shader_dir}/gpu_blit_fs_triangle.vert"
 	"${gpu_tiler_shader_dir}/gpu_blit_color_to_ms_depth.frag"
+	"${gpu_tiler_shader_dir}/gpu_video_out_overlay.frag"
 )
 foreach(shader_source IN LISTS gpu_blit_shader_sources)
 	get_filename_component(shader_name "${shader_source}" NAME_WE)

src/graphics/guest_gpu/graphicsRun.cpp

@@ -1502,8 +1502,9 @@ void CommandProcessor::PrepareCpuFlip(uint64_t request_id) {
 	ProcessorScope processor_scope(*this);
 
 	m_renderer.GetVideoOut().PrepareFlip(request_id, command);
+	GetScheduler().DeferPriorityOperation(
+	    [this, request_id] { m_renderer.GetVideoOut().CompleteFlip(request_id); });
 	GetScheduler().Flush();
-	m_renderer.GetVideoOut().CompleteFlip(request_id);
 }
 
 void CommandProcessor::SynchronizeGpu() {

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

@@ -798,9 +798,14 @@ void RenderExecutor::FindBuffers(PreparedBindings& prepared) {
 	auto&       cache    = m_context.GetBufferCache();
 
 	prepared.buffer_sources.clear();
-	prepared.buffer_sources.reserve(program.info.buffers.size());
-	for (uint32_t i = 0; i < program.info.buffers.size(); i++) {
-		auto descriptor = DecodeNativeDescriptor<ShaderBufferResource>(snapshot.buffers[i]);
+	const auto& layout = program.bindings;
+	if (layout.memory_offset_count == 0) {
+		return;
+	}
+	const auto& resources = layout.descriptors.front().resources;
+	prepared.buffer_sources.reserve(resources.size());
+	for (const auto resource: resources) {
+		auto descriptor = DecodeNativeDescriptor<ShaderBufferResource>(snapshot.buffers[resource]);
 		const auto address = descriptor.Base48();
 		const auto requested_size = descriptor.GetSize();
 		if (address == 0 || requested_size == 0) {
@@ -818,10 +823,10 @@ void RenderExecutor::RebindBuffers(PreparedBindings& prepared) {
 	const auto& program   = *prepared.runtime->program;
 	const auto& snapshot  = *prepared.runtime->resources;
 	const auto& layout    = program.bindings;
-	EXIT_IF(prepared.buffer_sources.size() != program.info.buffers.size());
+	EXIT_IF(prepared.buffer_sources.size() != layout.memory_offset_count);
 
 	prepared.buffers.clear();
-	prepared.buffers.reserve(program.info.buffers.size());
+	prepared.buffers.reserve(layout.memory_offset_count);
 	EXIT_IF(prepared.shader_data.size() != layout.ShaderDataDwords());
 	std::fill(prepared.shader_data.begin() + layout.memory_offset_dword,
 	          prepared.shader_data.end(), 0);
@@ -830,11 +835,12 @@ void RenderExecutor::RebindBuffers(PreparedBindings& prepared) {
 		const auto shift = (index % 4u) * 8u;
 		prepared.shader_data[dword] |= offset << shift;
 	};
-	for (uint32_t i = 0; i < program.info.buffers.size(); i++) {
+	for (uint32_t i = 0; i < layout.memory_offset_count; i++) {
+		const auto resource = layout.descriptors.front().resources[i];
 		uint32_t buffer_offset = 0;
 		prepared.buffers.push_back(NativeStorageBuffer(m_context, prepared.buffer_sources[i],
-		                                               program.info.buffers[i], program.stage, i,
-		                                               buffer_offset));
+		                                               program.info.buffers[resource],
+		                                               program.stage, resource, buffer_offset));
 		pack_memory_offset(i, buffer_offset);
 	}
 	if (ShaderRecompiler::IR::FindBinding(
@@ -1045,8 +1051,8 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 			} else {
 				switch (binding.kind) {
 					case BindingKind::Buffers:
-						for (const auto resource: binding.resources) {
-							const auto& view = descriptors.buffers.at(resource);
+						EXIT_IF(descriptors.buffers.size() != binding.resources.size());
+						for (const auto& view: descriptors.buffers) {
 							EXIT_IF(view.buffer == nullptr);
 							m_descriptor_buffers.push_back(view);
 						}

src/graphics/host_gpu/renderer/sync.cpp

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

src/graphics/host_gpu/shaders/gpu_video_out_overlay.frag

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

src/graphics/presentation/presenter.h

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

src/graphics/presentation/videoOut.cpp

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

src/graphics/presentation/videoOut.h

@@ -37,7 +37,7 @@ public:
 	void PrepareFlip(uint64_t request_id, Graphics::CommandBuffer& buffer);
 	void CompleteFlip(uint64_t request_id);
 	void SubmitFlipPreparation(uint64_t request_id);
-	void WaitForSubmitSlot();
+	void WaitForSubmitSlot(int handle);
 	void WaitFlipDone(int handle, int index);
 
 	[[nodiscard]] Impl& State() noexcept;
@@ -52,6 +52,9 @@ void                          VideoOutShutdown();
 
 KYTY_SYSV_ABI int  VideoOutOpen(int user_id, int bus_type, int index, const void* param);
 KYTY_SYSV_ABI int  VideoOutClose(int handle);
+KYTY_SYSV_ABI int  VideoOutSetFlipMaster(int slave_handle, int master_handle);
+KYTY_SYSV_ABI void VideoOutAddBufferAttributeOption(VideoOutBufferAttribute2* attribute,
+                                                   uint64_t option);
 KYTY_SYSV_ABI void VideoOutSetBufferAttribute2(VideoOutBufferAttribute2* attribute,
                                                uint64_t pixel_format, uint32_t tiling_mode,
                                                uint32_t width, uint32_t height, uint64_t option,

src/graphics/presentation/window/swapchain.cpp

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

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

@@ -3,6 +3,7 @@
 #include "common/assert.h"
 #include "graphics/shader/recompiler/backend/spirv/spirvEmitterInternal.h"
 #include "graphics/shader/recompiler/ir/ShaderIR.h"
+#include "graphics/shader/recompiler/ir/passes/BindingLayout.h"
 
 #include <algorithm>
 #include <array>
@@ -34,9 +35,6 @@ void ValidateNativeProgram(const IR::Program& program) {
 		present[index]   = true;
 		expected[index]  = std::move(resources);
 	};
-	if (!program.info.buffers.empty()) {
-		Expect(Kind::Buffers, Dense(program.info.buffers.size()));
-	}
 	for (uint32_t i = 0; i < program.info.images.size(); i++) {
 		const auto kind = IR::DescriptorBindingForImage(program.info.images[i]);
 		if (!kind.has_value()) {
@@ -54,23 +52,9 @@ void ValidateNativeProgram(const IR::Program& program) {
 	if (!program.info.samplers.empty()) {
 		Expect(Kind::Samplers, Dense(program.info.samplers.size()));
 	}
-	bool uses_gds = false;
-	for (const auto* block: program.blocks) {
-		for (const auto& inst: *block) {
-			if (IR::SharedAccessOf(inst.GetOpcode()) == IR::SharedAccess::None) {
-				continue;
-			}
-			const auto index = inst.Flags<IR::MemoryFlags>().index;
-			if (index >= program.memory_info.size()) {
-				Fail(program, "shared operation has invalid memory metadata");
-			}
-			const auto kind = program.memory_info[index].kind;
-			if (kind != IR::ResourceKind::Lds && kind != IR::ResourceKind::Gds) {
-				Fail(program, "shared operation has invalid resource kind");
-			}
-			uses_gds |= kind == IR::ResourceKind::Gds;
-		}
-	}
+	auto& buffers = expected[static_cast<size_t>(Kind::Buffers)];
+	const bool uses_gds = IR::CollectMemoryResources(program, buffers);
+	present[static_cast<size_t>(Kind::Buffers)] = !buffers.empty();
 	if (uses_gds) {
 		Expect(Kind::Gds);
 	}
@@ -109,7 +93,7 @@ void ValidateNativeProgram(const IR::Program& program) {
 	if ((program.bindings.UsesPushData() &&
 	     !IR::PushData::CanFit(program.bindings.push_data_start_dword, shader_data_dwords)) ||
 	    program.bindings.memory_offset_dword != program.bindings.user_data_registers.size() ||
-	    program.bindings.memory_offset_count != program.info.buffers.size() ||
+	    program.bindings.memory_offset_count != buffers.size() ||
 	    has_shader_data_storage != (shader_data_dwords != 0 && !program.bindings.UsesPushData()) ||
 	    !std::is_sorted(program.bindings.user_data_registers.begin(),
 	                    program.bindings.user_data_registers.end()) ||

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

@@ -537,6 +537,7 @@ constexpr Vop1SdwaRule VOP1_SDWA_RULES[] = {
     {Opcode::V_NOT_B32, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), SdwaSelBytes() | SdwaSelWords(),
      SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), false},
     {Opcode::V_FFBL_B32, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), 0, 0, false},
+    {Opcode::V_FFBH_U32, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), 0, 0, false},
     {Opcode::V_CVT_F32_F16, SdwaSelWords() | SdwaSelFull(), 0, 0, true},
     {Opcode::V_CVT_F16_F32, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), SdwaSelWords(),
      SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), true},
@@ -1379,6 +1380,7 @@ bool SupportsNativeVop3ResultModifiers(Opcode opcode) {
 		case Opcode::V_MAD_F32:
 		case Opcode::V_FMA_F32:
 		case Opcode::V_FMA_F16:
+		case Opcode::V_CUBEID_F32:
 		case Opcode::V_CVT_PKRTZ_F16_F32:
 		case Opcode::V_LDEXP_F32:
 		case Opcode::V_MIN3_F32:

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

@@ -46,29 +46,45 @@ void AddBinding(BindingLayout& layout, DescriptorBindingKind kind,
 	layout.descriptors.push_back({kind, std::move(resources)});
 }
 
-bool UsesGds(const Program& program) {
+} // namespace
+
+bool CollectMemoryResources(const Program& program, std::vector<uint32_t>& buffers) {
+	std::array<bool, ShaderInfo::MaxBuffers> live_buffers {};
 	bool uses_gds = false;
 	for (const auto* block: program.blocks) {
 		for (const auto& inst: *block) {
-			if (SharedAccessOf(inst.GetOpcode()) == SharedAccess::None) {
+			const auto op = inst.GetOpcode();
+			if (BufferAccessOf(op) == BufferAccess::None &&
+			    SharedAccessOf(op) == SharedAccess::None) {
 				continue;
 			}
 			const auto index = inst.Flags<MemoryFlags>().index;
 			if (index >= program.memory_info.size()) {
-				BindingFail("typed shader contains invalid shared-memory metadata");
+				BindingFail("typed shader contains invalid memory metadata");
 			}
-			const auto kind = program.memory_info[index].kind;
-			if (kind != ResourceKind::Lds && kind != ResourceKind::Gds) {
-				BindingFail("typed shader contains invalid shared-memory metadata");
+			const auto& memory = program.memory_info[index];
+			if (memory.planning_only) {
+				continue;
 			}
-			uses_gds |= kind == ResourceKind::Gds;
+			if (SharedAccessOf(op) != SharedAccess::None) {
+				if (memory.kind != ResourceKind::Lds && memory.kind != ResourceKind::Gds) {
+					BindingFail("typed shader contains invalid shared-memory metadata");
+				}
+				uses_gds |= memory.kind == ResourceKind::Gds;
+			} else if (memory.kind == ResourceKind::Buffer || memory.kind == ResourceKind::ScalarBuffer) {
+				EXIT_IF(memory.resource >= program.info.buffers.size());
+				live_buffers.at(memory.resource) = true;
+			}
+		}
+	}
+	for (uint32_t i = 0; i < program.info.buffers.size(); i++) {
+		if (live_buffers.at(i)) {
+			buffers.push_back(i);
 		}
 	}
 	return uses_gds;
 }
 
-} // namespace
-
 void AllocateBindings(Program& program, uint32_t push_data_start_dword) {
 	if (!program.shader_info_complete || program.binding_layout_complete) {
 		EXIT("shader binding layout failed: %s", !program.shader_info_complete
@@ -76,18 +92,17 @@ void AllocateBindings(Program& program, uint32_t push_data_start_dword) {
 		                                             : "binding layout already allocated");
 	}
 	BindingLayout next;
+	std::vector<uint32_t> buffers;
+	const bool            uses_gds = CollectMemoryResources(program, buffers);
 	next.user_data_registers = CollectUserData(program);
 	next.memory_offset_dword = static_cast<uint32_t>(next.user_data_registers.size());
-	next.memory_offset_count = static_cast<uint32_t>(program.info.buffers.size());
+	next.memory_offset_count       = static_cast<uint32_t>(buffers.size());
 	next.push_data_start_dword =
 	    PushData::StartFor(push_data_start_dword, next.ShaderDataDwords());
 
-	if (!program.info.buffers.empty()) {
-		std::vector<uint32_t> resources(program.info.buffers.size());
-		for (uint32_t i = 0; i < resources.size(); i++) {
-			resources[i] = i;
-		}
-		AddBinding(next, DescriptorBindingKind::Buffers, std::move(resources));
+	if (!buffers.empty()) {
+		// Draw binding accesses this first group directly when memory_offset_count is nonzero.
+		AddBinding(next, DescriptorBindingKind::Buffers, std::move(buffers));
 	}
 
 	std::array<std::vector<uint32_t>, ImageBindingCount> image_groups;
@@ -123,7 +138,7 @@ void AllocateBindings(Program& program, uint32_t push_data_start_dword) {
 		}
 		AddBinding(next, DescriptorBindingKind::Samplers, std::move(resources));
 	}
-	if (UsesGds(program)) {
+	if (uses_gds) {
 		AddBinding(next, DescriptorBindingKind::Gds);
 	}
 	if (program.info.uses_dma) {

src/graphics/shader/recompiler/ir/passes/BindingLayout.h

@@ -7,6 +7,9 @@ namespace Libs::Graphics::ShaderRecompiler::IR {
 
 void AllocateBindings(Program& program, uint32_t push_data_start_dword = 0);
 
+// Collect live direct buffer indices and report whether the shader accesses GDS.
+bool CollectMemoryResources(const Program& program, std::vector<uint32_t>& buffers);
+
 const DescriptorBinding* FindBinding(const BindingLayout& layout, DescriptorBindingKind kind);
 
 } // namespace Libs::Graphics::ShaderRecompiler::IR

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

@@ -2,6 +2,7 @@
 
 #include "common/assert.h"
 #include "graphics/guest_gpu/gpu_format.h"
+#include "graphics/shader/recompiler/BufferFormat.h"
 #include "graphics/shader/recompiler/ir/ShaderIR.h"
 #include "graphics/shader/shaderBindings.h"
 
@@ -446,6 +447,7 @@ static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSna
 		                             : Prospero::BufferFormat::kInvalid,
 		    .descriptor_swizzle =
 		        program.info.buffers[i].formatted ? descriptor.DstSelXYZW() : DstSel(4, 5, 6, 7),
+		    .zero_stride_oob = descriptor.OutOfBounds() == 0u && stride == 0u,
 		});
 	}
 	for (uint32_t i = 0; i < specialization.images.size(); i++) {
@@ -1127,6 +1129,49 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
 	for (auto* block: program.blocks) {
 		for (auto it = block->begin(); it != block->end(); ++it) {
 			auto& inst = *it;
+			if (BufferAccessOf(inst.GetOpcode()) == BufferAccess::Read) {
+				const auto& memory = memory_info[inst.Flags<MemoryFlags>().index];
+				if (memory.kind == ResourceKind::Buffer &&
+				    specialization.buffers[memory.resource].zero_stride_oob) {
+					// Bounds mode 0 checks offset >= stride, so zero stride
+					// makes every vector read out of bounds regardless of its address.
+					const auto count = BufferComponentCount(inst.GetOpcode());
+					std::array<Value, 4> values {Value(0u), Value(0u), Value(0u), Value(0u)};
+					if (memory.formatted && !memory.typed) {
+						const auto& buffer = buffers[memory.resource];
+						const auto format = Format::GetFormatInfo(buffer.descriptor_format);
+						for (uint32_t component = 0; component < count; component++) {
+							if (format.type == Format::ComponentType::Unknown ||
+							    GetDstSel(buffer.descriptor_swizzle, component) != 1u) continue;
+							const auto one = Format::FormattedConstantBits(
+							    format, Format::FormattedSourceKind::One);
+							values[component] = Value(&*block->PrependNewInst(
+							    it, ValueOpcode::SelectU32,
+							    {inst.Arg(inst.NumArgs() - 1), Value(one), Value(0u)}));
+						}
+					}
+					Value result = values[0];
+					switch (inst.GetType()) {
+						case Type::U8: result = Value(uint8_t {0}); break;
+						case Type::U16: result = Value(uint16_t {0}); break;
+						case Type::U32x2:
+							result = Value(&*block->PrependNewInst(it, ValueOpcode::CompositeConstructU32x2,
+							                                      {values[0], values[1]}));
+							break;
+						case Type::U32x3:
+							result = Value(&*block->PrependNewInst(it, ValueOpcode::CompositeConstructU32x3,
+							                                      {values[0], values[1], values[2]}));
+							break;
+						case Type::U32x4:
+							result = Value(&*block->PrependNewInst(it, ValueOpcode::CompositeConstructU32x4,
+							                                      {values[0], values[1], values[2], values[3]}));
+							break;
+						default: break;
+					}
+					inst.ReplaceUsesWith(result);
+				}
+				continue;
+			}
 			const auto image_opcode = ImageOpcodeInfoOf(inst.GetOpcode());
 			if (image_opcode.access == ImageAccess::None) {
 				continue;

src/graphics/shader/recompiler/ir/passes/ResourceMaterialization.h

@@ -12,6 +12,7 @@ struct ResourceSpecialization {
 		uint32_t               packed_stride                   = 0;
 		Prospero::BufferFormat descriptor_format               = Prospero::BufferFormat::kInvalid;
 		uint32_t               descriptor_swizzle              = DstSel(4, 5, 6, 7);
+		bool                   zero_stride_oob                 = false;
 		bool                   operator==(const Buffer&) const = default;
 	};
 

src/libs/libNet.cpp

@@ -7,9 +7,11 @@
 #include "libs/network.h"
 #include "loader/symbolDatabase.h"
 
+#include <algorithm>
 #include <cctype>
 #include <cstddef>
 #include <cstring>
+#include <limits>
 #include <map>
 #include <mutex>
 #include <string>
@@ -163,6 +165,16 @@ int KYTY_SYSV_ABI NetSetsockopt(int s, int level, int optname, const void* optva
 	return FinishSocketCall(Net::Setsockopt(s, level, optname, optval, optlen));
 }
 
+int KYTY_SYSV_ABI NetSend(int s, const void* buf, size_t len, int flags) {
+	const auto size = std::min<size_t>(len, std::numeric_limits<int>::max());
+	return FinishSocketCall(static_cast<int>(Net::Send(s, buf, size, flags | 0x20000)));
+}
+
+int KYTY_SYSV_ABI NetRecv(int s, void* buf, size_t len, int flags) {
+	const auto size = std::min<size_t>(len, std::numeric_limits<int>::max());
+	return FinishSocketCall(static_cast<int>(Net::Recv(s, buf, size, flags)));
+}
+
 uint32_t KYTY_SYSV_ABI NetHtonl(uint32_t host32) {
 	return ((host32 & 0x000000ffu) << 24u) | ((host32 & 0x0000ff00u) << 8u) |
 	       ((host32 & 0x00ff0000u) >> 8u) | ((host32 & 0xff000000u) >> 24u);
@@ -207,6 +219,8 @@ LIB_DEFINE(InitNet_1_Net) {
 	LIB_FUNC("Q4qBuN-c0ZM", LibNet::NetSocket);
 	LIB_FUNC("45ggEzakPJQ", LibNet::NetSocketClose);
 	LIB_FUNC("2mKX2Spso7I", LibNet::NetSetsockopt);
+	LIB_FUNC("beRjXBn-z+o", LibNet::NetSend);
+	LIB_FUNC("9wO9XrMsNhc", LibNet::NetRecv);
 	LIB_FUNC("9T2pDF2Ryqg", LibNet::NetHtonl);
 	LIB_FUNC("iWQWrwiSt8A", LibNet::NetHtons);
 	LIB_FUNC("pQGpHYopAIY", LibNet::NetNtohl);

src/libs/libVideoOut.cpp

@@ -15,6 +15,8 @@ LIB_DEFINE(InitVideoOut_1) {
 
 	LIB_FUNC("Up36PTk687E", VideoOut::VideoOutOpen);
 	LIB_FUNC("uquVH4-Du78", VideoOut::VideoOutClose);
+	LIB_FUNC("WkYtyOg30do", VideoOut::VideoOutSetFlipMaster);
+	LIB_FUNC("MBlv8YWoyOk", VideoOut::VideoOutAddBufferAttributeOption);
 	LIB_FUNC("PjS5uASwcV8", VideoOut::VideoOutSetBufferAttribute2);
 	LIB_FUNC("rKBUtgRrtbk", VideoOut::VideoOutRegisterBuffers2);
 	LIB_FUNC("HuViW4HnrOw", VideoOut::VideoOutSubmitChangeBufferAttribute2);

src/libs/network.cpp

@@ -1003,6 +1003,7 @@ static int ConvertHostSocketError(int error) {
 		case ENOBUFS: posix_error = Posix::POSIX_ENOBUFS; break;
 		case ENOMEM: posix_error = Posix::POSIX_ENOMEM; break;
 		case ENOTSOCK: posix_error = Posix::POSIX_ENOTSOCK; break;
+		case EPIPE: posix_error = Posix::POSIX_EPIPE; break;
 		case EPROTONOSUPPORT: posix_error = Posix::POSIX_EPROTONOSUPPORT; break;
 		default: break;
 	}

tests/KernelFileSystemTests.cpp

@@ -17,6 +17,7 @@
 #include <array>
 
 #include <chrono>
+#include <csignal>
 #include <cstdio>
 #include <cstdlib>
 #include <cstring>
@@ -31,6 +32,10 @@ namespace Libs::LibKernelApr {
 void InitLibKernel_1_Apr(Loader::SymbolDatabase *symbols);
 }
 
+namespace Libs::LibNet {
+void InitNet_1_Net(Loader::SymbolDatabase *symbols);
+}
+
 namespace {
 
 namespace FileSystem = Libs::LibKernel::FileSystem;
@@ -407,6 +412,22 @@ void CheckAprPaths(const std::filesystem::path &root) {
 
 void CheckSocketWakeup() {
   namespace Net = Libs::Network::Net;
+  Loader::SymbolDatabase symbols;
+  Libs::LibNet::InitNet_1_Net(&symbols);
+  const auto *send_symbol = symbols.Find(
+      {"beRjXBn-z+o", "Net", 1, "Net", 1, 1, Loader::SymbolType::Func});
+  const auto *recv_symbol = symbols.Find(
+      {"9wO9XrMsNhc", "Net", 1, "Net", 1, 1, Loader::SymbolType::Func});
+  const auto *errno_symbol = symbols.Find(
+      {"HQOwnfMGipQ", "Net", 1, "Net", 1, 1, Loader::SymbolType::Func});
+  Check(send_symbol && recv_symbol && errno_symbol,
+        "Net send, receive and errno exports resolve with the guest ABI versions");
+  using Send = int (KYTY_SYSV_ABI *)(int, const void *, size_t, int);
+  using Recv = int (KYTY_SYSV_ABI *)(int, void *, size_t, int);
+  using Errno = int *(KYTY_SYSV_ABI *)();
+  const auto net_send = reinterpret_cast<Send>(send_symbol->vaddr);
+  const auto net_recv = reinterpret_cast<Recv>(recv_symbol->vaddr);
+  auto *net_errno = reinterpret_cast<Errno>(errno_symbol->vaddr)();
   // Guest sockaddr_in: length, family, network-order port/address, padding.
   std::array<uint8_t, 16> address {16, 2, 0, 0, 127, 0, 0, 1};
   const int listener = Net::Socket(2, 1, 0);
@@ -441,26 +462,61 @@ void CheckSocketWakeup() {
                     immediate.data()) == 0 && readable[reader / 64] == 0,
         "empty socket is not readable");
   const char payload[] = "wake";
-  Check(Net::Send(writer, payload, sizeof(payload), 0x20000) == sizeof(payload),
-        "send wake bytes with guest MSG_NOSIGNAL");
+  *net_errno = Libs::Posix::POSIX_EINVAL;
+  Check(net_send(writer, payload, sizeof(payload), 0) == sizeof(payload) &&
+            *net_errno == Libs::Posix::POSIX_EINVAL,
+        "Net send forwards bytes and preserves errno on success");
   readable[reader / 64] = bit;
   const std::array<int64_t, 2> deadline {1, 0};
   Check(Net::Select(reader + 1, readable.data(), nullptr, nullptr,
                     deadline.data()) == 1 && readable[reader / 64] == bit,
         "select reports the guest descriptor after wake");
   std::array<char, sizeof(payload)> received {};
-  Check(Net::Recv(reader, received.data(), received.size(), 0x42) == sizeof(payload) &&
+  Check(net_recv(reader, received.data(), received.size(), 0x42) == sizeof(payload) &&
+            std::memcmp(received.data(), payload, sizeof(payload)) == 0 &&
+            *net_errno == Libs::Posix::POSIX_EINVAL,
+        "Net receive forwards PEEK and WAITALL without consuming bytes");
+  received.fill(0);
+  Check(net_recv(reader, received.data(), received.size(), 0x40) == sizeof(payload) &&
             std::memcmp(received.data(), payload, sizeof(payload)) == 0,
-        "guest PEEK and WAITALL preserve the wake bytes");
-  Check(Net::Recv(reader, received.data(), received.size(), 0x40) == sizeof(payload),
-        "consume wake bytes with guest WAITALL");
+        "Net receive consumes the same bytes after peeking");
 #if !defined(_WIN32)
   Check(Net::Recv(reader, received.data(), received.size(), 0x80) == -1 &&
             *Libs::Posix::GetErrorAddr() == Libs::Posix::POSIX_EWOULDBLOCK,
         "empty nonblocking receive translates guest errno");
+  Check(net_recv(reader, received.data(), received.size(), 0x80) ==
+            Libs::Network::NET_ERROR_EWOULDBLOCK &&
+            *net_errno == Libs::Posix::POSIX_EWOULDBLOCK,
+        "Net nonblocking receive translates POSIX failure and Net errno");
 #endif
-  Check(Net::SocketClose(reader) == 0 && Net::SocketClose(writer) == 0,
-        "close wake sockets");
+  Check(net_send(-1, payload, sizeof(payload), 0) == Libs::Network::NET_ERROR_EBADF &&
+            *net_errno == Libs::Posix::POSIX_EBADF,
+        "Net send translates an invalid socket instead of returning POSIX minus one");
+  Check(net_recv(reader, nullptr, received.size(), 0) == Libs::Network::NET_ERROR_EFAULT &&
+            *net_errno == Libs::Posix::POSIX_EFAULT,
+        "Net receive translates an invalid output buffer");
+  Check(net_send(writer, payload, sizeof(payload), 0x100000) ==
+            Libs::Network::NET_ERROR_EOPNOTSUPP &&
+            *net_errno == Libs::Posix::POSIX_EOPNOTSUPP,
+        "Net send preserves the backend's unsupported crypto flag error");
+#if defined(__linux__)
+  const int disconnected = Net::Socket(2, 1, 0);
+  Check(disconnected >= 0, "create unconnected socket for broken pipe check");
+  const auto previous_sigpipe = std::signal(SIGPIPE, SIG_DFL);
+  Check(previous_sigpipe != SIG_ERR, "set default SIGPIPE disposition for Net send");
+  const auto broken_send = net_send(disconnected, payload, sizeof(payload), 0);
+  std::signal(SIGPIPE, previous_sigpipe);
+  Check(broken_send == Libs::Network::NET_ERROR_EPIPE &&
+            *net_errno == Libs::Posix::POSIX_EPIPE,
+        "Net send reports a broken pipe without raising host SIGPIPE");
+  Check(Net::SocketClose(disconnected) == 0, "close unconnected socket");
+#endif
+  Check(Net::SocketClose(writer) == 0, "close wake writer");
+  *net_errno = Libs::Posix::POSIX_EINVAL;
+  Check(net_recv(reader, received.data(), received.size(), 0) == 0 &&
+            *net_errno == Libs::Posix::POSIX_EINVAL,
+        "Net receive returns EOF without replacing errno");
+  Check(Net::SocketClose(reader) == 0, "close wake reader");
   readable[reader / 64] = bit;
   Check(Net::Select(reader + 1, readable.data(), nullptr, nullptr,
                     immediate.data()) == -1 &&

tests/ShaderRecompilerComputeTests.cpp

@@ -1208,6 +1208,7 @@ struct TestCase {
   std::vector<std::pair<std::string, size_t>> decoded_counts;
   std::vector<std::pair<std::string, size_t>> ir_counts;
   u32 expected_storage_mip_descriptors = 0;
+  std::optional<std::vector<u32>> expected_buffer_resources;
 };
 
 struct GraphicsCase {
@@ -1241,7 +1242,9 @@ struct CompiledShader {
 
 std::array<u32, 64> MakeNativeUserData(const std::array<u32, 64> *source) {
   std::array<u32, 64> data{};
+  data[3] = 3u << 28u;
   data[50] = 1u << 20u;
+  data[51] = 3u << 28u;
   if (source != nullptr) {
     data = *source;
   }
@@ -1564,18 +1567,26 @@ CompiledShader CompileCase(const TestCase &test, u32 host_subgroup_size = 64) {
           "recompiler returned empty SPIR-V");
   ValidateSpirv(test.name, result.spirv);
   CheckSpirvText(test, result.spirv);
+  const auto *buffer_binding = ShaderRecompiler::IR::FindBinding(
+      result.program.bindings, ShaderRecompiler::IR::DescriptorBindingKind::Buffers);
+  if (test.expected_buffer_resources) {
+    Require(test.name, "live buffer bindings",
+            buffer_binding != nullptr &&
+                buffer_binding->resources == *test.expected_buffer_resources,
+            "dead buffer changed the surviving descriptor order");
+  }
   std::vector<u32> packed_user_data;
   for (const auto reg : result.program.bindings.user_data_registers) {
     packed_user_data.push_back(
         resources.user_data[reg - result.program.user_data_base]);
   }
   packed_user_data.resize(result.program.bindings.ShaderDataDwords());
-  const auto buffer_count =
-      static_cast<u32>(result.program.info.buffers.size());
+  const auto buffer_count = result.program.bindings.memory_offset_count;
   for (u32 i = 0; i < buffer_count; i++) {
     u32 offset = 0;
-    if (i < test.storage_buffer_offsets.size()) {
-      offset = test.storage_buffer_offsets[i];
+    const auto resource = buffer_binding->resources[i];
+    if (resource < test.storage_buffer_offsets.size()) {
+      offset = test.storage_buffer_offsets[resource];
     }
     Require(test.name, "shader data", offset % sizeof(u32) == 0 && offset < 256,
             "storage buffer offset is not representable");
@@ -1593,7 +1604,8 @@ std::array<u32, 64> MakeStructuredStorageBufferData(u32 stride_bytes,
   std::array<u32, 64> data{};
   data[1] = (stride_bytes & 0x3fffu) << 16u;
   data[2] = num_records;
-  data[3] = DstSel(4, 5, 6, 7) | (1u << 24u);
+  data[3] = DstSel(4, 5, 6, 7) | (1u << 24u) |
+            (stride_bytes == 0 ? 3u << 28u : 0u);
   if (add_tid) {
     data[3] |= 1u << 23u;
   }
@@ -13233,8 +13245,9 @@ public:
         info.offset = 0;
         info.range = buffer.size;
         if (test.storage_buffer_range_dwords != 0) {
-          const auto offset = i < test.storage_buffer_offsets.size()
-                                  ? test.storage_buffer_offsets[i]
+          const auto resource = buffers->resources[i];
+          const auto offset = resource < test.storage_buffer_offsets.size()
+                                  ? test.storage_buffer_offsets[resource]
                                   : 0u;
           info.range = static_cast<vk::DeviceSize>(
               test.storage_buffer_range_dwords * sizeof(u32) + offset);
@@ -18876,6 +18889,31 @@ TestCase VectorFfbhI32NativeAndVop3OnGpu() {
   return test;
 }
 
+TestCase Vop1SdwaFfbhCapturedScalarLowWordSource() {
+  using O = ShaderOpcode;
+
+  constexpr std::array inputs = {0xdead0000u, 0xffff0001u, 0x00008000u,
+                                 0x00010008u, 0xffff4000u, 0x0000ffffu};
+  std::vector<u32> code;
+  for (u32 i = 0; i < inputs.size(); i++) {
+    AppendSmemLoadOpcode(&code, 0x08, 39, i * 4u);
+    code.insert(code.end(), {0x7e0072f9u, 0x00840627u}); // v_ffbh_u32 v0, s39.word0
+    AppendStoreVgpr(&code, 0, i);
+  }
+  AppendEnd(&code);
+
+  TestCase test;
+  test.name = "Vop1SdwaFfbhCapturedScalarLowWordSource";
+  test.code = std::move(code);
+  test.initial.assign(inputs.begin(), inputs.end());
+  test.expected = {0xffffffffu, 31u, 16u, 28u, 17u, 16u};
+  test.opcodes = {O::S_BUFFER_LOAD_DWORD, O::V_FFBH_U32, O::V_MOV_B32,
+                  O::BUFFER_STORE_DWORD, O::S_ENDPGM};
+  test.decoded_counts = {{"V_FFBH_U32 v0, s39.sdwa(sel=4,sext=0)", inputs.size()}};
+  test.required_spirv = {"OpBitFieldUExtract", "FindUMsb"};
+  return test;
+}
+
 TestCase Vop1SdwaFfblCapturedHighWordSource() {
   using O = ShaderOpcode;
 
@@ -21192,6 +21230,37 @@ TestCase Vop1MoveRelDestination() {
   return test;
 }
 
+TestCase CubeIdCapturedNegationAndOutputScale() {
+  using O = ShaderOpcode;
+
+  constexpr std::array directions = {
+      std::array{4.0f, 2.0f, 1.0f}, std::array{-4.0f, 2.0f, 1.0f},
+      std::array{1.0f, -4.0f, 2.0f}, std::array{1.0f, 4.0f, 2.0f},
+      std::array{1.0f, 2.0f, -4.0f}, std::array{1.0f, 2.0f, 4.0f}};
+  constexpr std::array registers = {47u, 49u, 48u};
+  TestCase test;
+  test.name = "CubeIdCapturedNegationAndOutputScale";
+  for (u32 i = 0; i < directions.size(); i++) {
+    for (u32 component = 0; component < registers.size(); component++) {
+      test.initial.push_back(std::bit_cast<u32>(directions[i][component]));
+      AppendVMovU32(&test.code, 30, (i * 3u + component) * 4u);
+      AppendBufferLoadDword(&test.code, registers[component], 30);
+    }
+    // The captured instruction negates Y/Z and divides the floating face ID by two.
+    test.code.insert(test.code.end(), {0xd544000bu, 0xdcc2632fu});
+    AppendStoreVgpr(&test.code, 11, i);
+  }
+  AppendEnd(&test.code);
+  test.expected = {0x00000000u, 0x3f000000u, 0x3f800000u,
+                   0x3fc00000u, 0x40000000u, 0x40200000u};
+  test.opcodes = {O::V_MOV_B32, O::BUFFER_LOAD_DWORD, O::V_CUBEID_F32,
+                  O::BUFFER_STORE_DWORD, O::S_ENDPGM};
+  test.decoded_counts = {
+      {"V_CUBEID_F32 v11.omod(3), v47, v49.neg, v48.neg", directions.size()}};
+  test.required_spirv = {"OpFNegate", "OpFMul"};
+  return test;
+}
+
 TestCase VectorFloatSpecialOps() {
   using O = ShaderOpcode;
 
@@ -24503,6 +24572,113 @@ TestCase BufferStoreFormatXDropsOutOfRangeRecord() {
   return test;
 }
 
+TestCase TBufferCapturedZeroStrideOob(bool raw_bounds = false, bool scalar = false) {
+  using O = ShaderOpcode;
+  TestCase test;
+  test.name = scalar ? "TBufferZeroStrideOobPreservesScalarRead"
+                    : raw_bounds ? "TBufferZeroStrideRawBoundsStillLoads"
+                                 : "TBufferCapturedZeroStrideOob";
+  test.initial = {0x11223344u, 0x55667788u, 0x99aabbccu, 0xddeeff00u,
+                  0x12345678u, 0xdeadbeefu, 0xdeadbeefu};
+  test.expected = test.initial;
+  for (u32 i = 0; i < 4; i++) {
+    test.expected[1 + i] = raw_bounds ? test.initial[i] : 0u;
+  }
+  test.expected[5] = test.initial[4];
+  if (scalar) test.expected[6] = test.initial[0];
+  test.user_data = MakeStructuredStorageBufferData(4, test.initial.size());
+  std::copy_n(test.user_data.begin(), 4, test.user_data.begin() + 4);
+  test.user_data[4] = 0x2000u;
+  // Captured HITMAN descriptor in s[32:35]: mode 0, zero stride, arbitrary base.
+  test.user_data[32] = 0x4a398620u;
+  test.user_data[33] = 0;
+  test.user_data[34] = 0x2b5u;
+  test.user_data[35] = 0x0004d000u | (raw_bounds ? 3u << 28u : 0u);
+  test.has_user_data = true;
+  test.storage_buffer_offsets = {0, 0, 16};
+  test.expected_buffer_resources = raw_bounds || scalar
+                                       ? std::vector<u32>{0, 1, 2}
+                                       : std::vector<u32>{0, 2};
+  auto &code = test.code;
+  AppendVMovLiteral(&code, 0, test.initial[0]);
+  AppendStoreVgpr(&code, 0, 0); // Keep resource zero before the removed middle slot.
+  AppendVMovU32(&code, 30, 0);
+  code.push_back(0xea6b2000u);
+  code.push_back(0x8008191eu); // Exact pc 0x47c: typed float4, s[32:35], v30.
+  AppendVMovU32(&code, 20, 0);
+  code.push_back(EncodeMubuf0(0x0cu));
+  code.push_back(EncodeMubuf1(33, 1, 20));
+  if (scalar) {
+    code.push_back(EncodeSmem0(0x08u, 40, 16));
+    code.push_back(EncodeSmem1(0));
+    code.push_back(EncodeVop1(0x01u, 32, 40));
+  }
+  for (u32 i = 0; i < 4; i++) AppendStoreVgpr(&code, 25 + i, 1 + i);
+  AppendStoreVgpr(&code, 33, 5);
+  if (scalar) AppendStoreVgpr(&code, 32, 6);
+  AppendEnd(&code);
+  test.opcodes = {O::V_MOV_B32, O::TBUFFER_LOAD_FORMAT_XYZW,
+                  O::BUFFER_LOAD_DWORD, O::BUFFER_STORE_DWORD, O::S_ENDPGM};
+  if (scalar) test.opcodes.push_back(O::S_BUFFER_LOAD_DWORD);
+  test.ir_counts = {{"LoadBufferU32x4", raw_bounds ? 1u : 0u}};
+  return test;
+}
+
+TestCase BufferZeroStrideOobFormatsAndWidths() {
+  using O = ShaderOpcode;
+  TestCase test;
+  test.name = "BufferZeroStrideOobFormatsAndWidths";
+  test.initial = std::vector<u32>(20, 0xdeadbeefu);
+  test.expected = std::vector<u32>(20, 0);
+  test.expected[10] = 0x3f800000u;
+  test.expected[12] = 0x3f800000u;
+  test.expected[14] = 1;
+  test.expected[16] = 1;
+  test.expected[18] = 0xabcdef01u;
+  test.expected[19] = 0xdeadbeefu;
+  test.user_data = MakeStructuredStorageBufferData(4, 20);
+  for (u32 base : {4u, 8u}) {
+    test.user_data[base] = 0x4a398620u;
+    test.user_data[base + 2] = 0x2b5u;
+    const u32 format = base == 4 ? 77u : 75u; // Four 32-bit float or unsigned integer components.
+    test.user_data[base + 3] = (format << 12u) | DstSel(1, 4, 1, 0);
+  }
+  test.has_user_data = true;
+  test.expected_buffer_resources = std::vector<u32>{0};
+  auto &code = test.code;
+  AppendVMovU32(&code, 0, 0);
+  AppendStoreVgpr(&code, 0, 0);
+  AppendVMovU32(&code, 20, 0);
+  u32 destination = 0;
+  for (const auto [opcode, width] :
+       {std::pair{0x08u, 1u}, {0x0au, 1u}, {0x0cu, 1u}, {0x0du, 2u}, {0x0fu, 3u}}) {
+    code.push_back(EncodeMubuf0(opcode));
+    code.push_back(EncodeMubuf1(destination, 1, 20));
+    destination += width;
+  }
+  AppendVMovU32(&code, 8, 0);
+  AppendVMovU32(&code, 9, 0);
+  for (u32 base : {1u, 2u}) {
+    code.push_back(EncodeMubuf0(0x03u));
+    code.push_back(EncodeMubuf1(base == 1 ? 10 : 14, base, 20));
+  }
+  AppendVMovLiteral(&code, 18, 0xabcdef01u);
+  code.push_back(EncodeSop1(0x04u, 60, 126u)); // Preserve the partial wave's active lanes.
+  code.push_back(EncodeSop1(0x04u, 126, InlineU32(0)));
+  code.push_back(EncodeMubuf0(0x00u));
+  code.push_back(EncodeMubuf1(18, 1, 20));
+  code.push_back(EncodeSop1(0x04u, 126, 60u));
+  for (u32 i = 0; i < 19; i++) AppendStoreVgpr(&code, i, i);
+  AppendEnd(&code);
+  test.opcodes = {O::V_MOV_B32, O::S_MOV_B64, O::BUFFER_LOAD_UBYTE,
+                  O::BUFFER_LOAD_USHORT, O::BUFFER_LOAD_DWORD,
+                  O::BUFFER_LOAD_DWORDX2, O::BUFFER_LOAD_DWORDX3,
+                  O::BUFFER_LOAD_FORMAT_X, O::BUFFER_LOAD_FORMAT_XYZW,
+                  O::BUFFER_STORE_DWORD, O::S_ENDPGM};
+  test.ir_counts = {{"LoadBuffer", 0}};
+  return test;
+}
+
 TestCase TBufferLoadVariants() {
   using O = ShaderOpcode;
 
@@ -29001,6 +29177,7 @@ std::vector<TestCase> MakeCases() {
   AddCase(VectorVop3MoveAppliesFloatSourceModifiers);
   AddCase(VectorIntegerOps);
   AddCase(VectorFfbhI32NativeAndVop3OnGpu);
+  AddCase(Vop1SdwaFfbhCapturedScalarLowWordSource);
   AddCase(Vop1SdwaFfblCapturedHighWordSource);
   AddCase(Vop1SdwaNotCapturedByte0Source);
   AddCase(Vop1SdwaNotPreservesHighWordDestination);
@@ -29072,6 +29249,7 @@ std::vector<TestCase> MakeCases() {
   AddCase(Vop1MoveRelSource);
   AddCase(Vop1MoveRelDestination);
   AddCase(VectorFloatSpecialOps);
+  AddCase(CubeIdCapturedNegationAndOutputScale);
   AddCase(MadMixF16LiteralHalfSourceUsesOpsel);
   AddCase(MadMixF16NegHiIsAbsAndNegIsIndependent);
   AddCase(VectorVop3FmaF16UsesRdna2Opcode34b);
@@ -29173,6 +29351,10 @@ std::vector<TestCase> MakeCases() {
   AddCase(BufferStoreFormatXAddTidUsesLaneIndex);
   AddCase(BufferStoreFormatXDropsOutOfRangeRecord);
   AddCase(TBufferLoadVariants);
+  AddCase([] { return TBufferCapturedZeroStrideOob(); });
+  AddCase([] { return TBufferCapturedZeroStrideOob(true); });
+  AddCase([] { return TBufferCapturedZeroStrideOob(false, true); });
+  AddCase(BufferZeroStrideOobFormatsAndWidths);
   AddCase(TBufferLoadFormatXyzwSnapshotsOverlappingAddress);
   AddCase(TBufferLoadFormatXyzwPackedSnapshotsOverlappingAddress);
   AddCase(TBufferLoadFormatX8UintZeroExtendsByte);
@@ -34160,6 +34342,28 @@ int main(int argc, char **argv) {
     RunCase(&vulkan, Vop1SdwaMovByteDestinations());
     return 0;
   }
+  if (argc == 2 && std::strcmp(argv[1], "--sdwa-ffbh-only") == 0) {
+    VulkanHarness vulkan;
+    RunCase(&vulkan, Vop1SdwaFfbhCapturedScalarLowWordSource());
+    RunCase(&vulkan, VectorFfbhI32NativeAndVop3OnGpu());
+    RunCase(&vulkan, Vop1SdwaFfblCapturedHighWordSource());
+    return 0;
+  }
+  if (argc == 2 && std::strcmp(argv[1], "--cubeid-omod-only") == 0) {
+    VulkanHarness vulkan;
+    RunCase(&vulkan, CubeIdCapturedNegationAndOutputScale());
+    RunCase(&vulkan, VectorFloatSpecialOps());
+    return 0;
+  }
+  if (argc == 2 && std::strcmp(argv[1], "--zero-stride-oob-only") == 0) {
+    VulkanHarness vulkan;
+    RunCase(&vulkan, TBufferCapturedZeroStrideOob());
+    RunCase(&vulkan, TBufferCapturedZeroStrideOob(true));
+    RunCase(&vulkan, TBufferCapturedZeroStrideOob(false, true));
+    RunCase(&vulkan, BufferZeroStrideOobFormatsAndWidths());
+    RunCase(&vulkan, TBufferLoadFormatXIdxenUsesDescriptorStride());
+    return 0;
+  }
   if (argc == 2 && std::strcmp(argv[1], "--sdwa-ashr-only") == 0) {
     VulkanHarness vulkan;
     RunCase(&vulkan, Vop2SdwaAshrrevCapturedWord0SignExtends());

tests/shaderCfgTests.cpp

@@ -86,7 +86,11 @@ ShaderRecompiler::CompileOptions MakeCompileOptions(ShaderType stage) {
   static const ShaderVertexInputInfo vertex{};
   static const ShaderPixelInputInfo pixel{};
   static const ShaderComputeInputInfo compute{};
-  static const std::array<uint32_t, 64> user_data{};
+  static const auto user_data = [] {
+    std::array<uint32_t, 64> data{};
+    data[3] = 3u << 28u; // Default fixture buffer uses raw offset bounds.
+    return data;
+  }();
 
   ShaderRecompiler::CompileOptions options;
   options.stage = stage;
@@ -6512,6 +6516,7 @@ void TestNewShaderRecompilerNativeWideScalarMemoryIr() {
 
 void TestNewShaderRecompilerNativeWideBufferIr() {
   const uint32_t shader[] = {
+      EncodeSMovB32(83, 255), 3u << 28u, // Raw bounds for the s[80:83] fixture.
       EncodeMubuf0(0x0d, 0),
       EncodeMubuf1(0, 20, 1), // buffer_load_dwordx2 v[0:1]
       EncodeMubuf0(0x1d, 16),