gpu: compile RDNA2 tessellation through native LS HS and TES stages

2026-10-24 21:06:55 +02:00 · 2026-09-13 07:45:08 +04:00
parent 6d1ba5870a
commit 7516068662
26 changed files with 1035 additions and 161 deletions
@@ -93,22 +93,15 @@ static const char* ShaderStageResourceName(ShaderType stage) {
 	switch (stage) {
 		case ShaderType::Vertex: return "Vertex";
 		case ShaderType::Mesh: return "Mesh";
+		case ShaderType::Local: return "Local";
+		case ShaderType::TessellationControl: return "Hull";
+		case ShaderType::TessellationEvaluation: return "Domain";
 		case ShaderType::Pixel: return "Pixel";
 		case ShaderType::Compute: return "Compute";
 		default: return "Unknown";
 	}
 }

-static vk::ShaderStageFlags NativeShaderStage(ShaderType stage) {
-	switch (stage) {
-		case ShaderType::Vertex: return vk::ShaderStageFlagBits::eVertex;
-		case ShaderType::Mesh: return vk::ShaderStageFlagBits::eMeshEXT;
-		case ShaderType::Pixel: return vk::ShaderStageFlagBits::eFragment;
-		case ShaderType::Compute: return vk::ShaderStageFlagBits::eCompute;
-		default: EXIT("unknown native shader stage\n");
-	}
-}
-
 static Prospero::ImageType TextureType(const ShaderTextureResource& descriptor) {
 	const auto type = descriptor.Type();
 	return type == Prospero::ImageType::kCube ? Prospero::ImageType::kColor2DArray : type;
@@ -884,9 +877,10 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 	size_t write_count      = 0;
 	ShaderRecompiler::IR::PushData push_data;
 	bool                           has_push_data = false;
-	constexpr auto                 GraphicsStages = vk::ShaderStageFlagBits::eVertex |
-	                                                vk::ShaderStageFlagBits::eMeshEXT |
-	                                                vk::ShaderStageFlagBits::eFragment;
+	constexpr auto                 GraphicsStages =
+	    vk::ShaderStageFlagBits::eVertex | vk::ShaderStageFlagBits::eMeshEXT |
+	    vk::ShaderStageFlagBits::eTessellationControl |
+	    vk::ShaderStageFlagBits::eTessellationEvaluation | vk::ShaderStageFlagBits::eFragment;
 	vk::ShaderStageFlags push_stages = pipeline_bind_point == vk::PipelineBindPoint::eGraphics
 	                                       ? vk::ShaderStageFlagBits::eFragment
 	                                       : vk::ShaderStageFlags {};
@@ -236,6 +236,9 @@ struct PipelineCache::ProgramCache {
 		switch (options.stage) {
 			case ShaderType::Vertex: stage_name = "vs"; break;
 			case ShaderType::Mesh: stage_name = "ms"; break;
+			case ShaderType::Local: stage_name = "ls"; break;
+			case ShaderType::TessellationControl: stage_name = "hs"; break;
+			case ShaderType::TessellationEvaluation: stage_name = "ds"; break;
 			case ShaderType::Pixel: stage_name = "ps"; break;
 			case ShaderType::Compute: stage_name = "cs"; break;
 			default: EXIT("invalid pipeline shader stage\n");
@@ -273,7 +276,7 @@ struct PipelineCache::ProgramCache {
 	                  uint32_t& push_data_cursor) {
 		ShaderType stage;
 		if constexpr (std::is_same_v<InputInfo, ShaderVertexInputInfo>) {
-			stage = input_info.mesh.threads_num[0] != 0 ? ShaderType::Mesh : ShaderType::Vertex;
+			stage = input_info.logical_stage;
 		} else if constexpr (std::is_same_v<InputInfo, ShaderPixelInputInfo>) {
 			stage = ShaderType::Pixel;
 		} else {
@@ -325,6 +328,9 @@ struct PipelineCache::ProgramCache {
 		switch (stage) {
 			case ShaderType::Vertex: label = "ShaderRecompiler VS"; break;
 			case ShaderType::Mesh: label = "ShaderRecompiler MS"; break;
+			case ShaderType::Local: label = "ShaderRecompiler LS"; break;
+			case ShaderType::TessellationControl: label = "ShaderRecompiler HS"; break;
+			case ShaderType::TessellationEvaluation: label = "ShaderRecompiler DS"; break;
 			case ShaderType::Pixel: label = "ShaderRecompiler PS"; break;
 			case ShaderType::Compute: label = "ShaderRecompiler CS"; break;
 			default: EXIT("invalid pipeline shader stage\n");
@@ -341,9 +347,9 @@ struct PipelineCache::ProgramCache {

 		if constexpr (std::is_same_v<InputInfo, ShaderVertexInputInfo>) {
 			options.user_data_base = 8;
-			if (stage == ShaderType::Mesh) {
+			if (stage == ShaderType::Mesh || stage == ShaderType::TessellationControl) {
 				options.user_data_base = 0;
-				options.wave_size      = input_info.mesh.wave_size;
+				options.wave_size = stage == ShaderType::Mesh ? input_info.mesh.wave_size : 64u;
 			}
 		} else if constexpr (std::is_same_v<InputInfo, ShaderComputeInputInfo>) {
 			options.wave_size = input_info.wave_size;
@@ -361,16 +367,19 @@ struct PipelineCache::ProgramCache {
 		input_info.stage = {.program = &permutation.program, .resources = std::move(resources)};
 		permutation.program.bindings.AdvancePushData(push_data_cursor);

-		std::array<size_t, static_cast<size_t>(ShaderType::Mesh) + 1> counts {};
+		std::array<size_t, static_cast<size_t>(ShaderType::TessellationEvaluation) + 1> counts {};
 		for (const auto& [key, source]: programs) {
 			counts[static_cast<size_t>(key.stage)] += source.permutations.size();
 		}
 		// Guest geometry shaders are compiled through the host mesh stage.
-		std::printf("Shaders: VS %zu | PS %zu | CS %zu | GS %zu\n",
+		std::printf("Shaders: VS %zu | PS %zu | CS %zu | GS %zu | LS %zu | HS %zu | TES %zu\n",
 		            counts[static_cast<size_t>(ShaderType::Vertex)],
 		            counts[static_cast<size_t>(ShaderType::Pixel)],
 		            counts[static_cast<size_t>(ShaderType::Compute)],
-		            counts[static_cast<size_t>(ShaderType::Mesh)]);
+		            counts[static_cast<size_t>(ShaderType::Mesh)],
+		            counts[static_cast<size_t>(ShaderType::Local)],
+		            counts[static_cast<size_t>(ShaderType::TessellationControl)],
+		            counts[static_cast<size_t>(ShaderType::TessellationEvaluation)]);
 		return permutation.handle;
 	}

@@ -564,12 +573,18 @@ PipelineCache::GraphicsPrograms PipelineCache::GetGraphicsPrograms(
    const HW::VertexShaderInfo& vertex_regs, const HW::PixelShaderInfo& pixel_regs,
    const HW::ShaderRegisters& sh, const HW::Context& context, const HW::UserConfig& user_config,
    std::span<const Prospero::ColorComponentMapping, 8> target_export_mapping, bool pixel_active,
-    ShaderVertexInputInfo& vertex_info, ShaderPixelInputInfo& pixel_info) {
-	const auto vertex_params = PrepareProgram(vertex_regs, context, user_config, vertex_info);
-	const bool mesh_active   = vertex_info.mesh.threads_num[0] != 0;
+    std::array<ShaderVertexInputInfo, 3>& vertex_info, ShaderPixelInputInfo& pixel_info) {
+	const bool tess_active = user_config.GetPrimType() == Prospero::PrimitiveType::kPatch;
+	std::array<ShaderParams, 3> vertex_params;
+	if (tess_active) {
+		vertex_params = PrepareTessellationPrograms(vertex_regs, context, vertex_info);
+	} else {
+		vertex_params[0] = PrepareProgram(vertex_regs, context, user_config, vertex_info[0]);
+	}
+	const bool mesh_active = vertex_info[0].logical_stage == ShaderType::Mesh;
 	if (mesh_active) {
 		EXIT_NOT_IMPLEMENTED(!m_graphics.mesh_shader_enabled);
-		auto& mesh              = vertex_info.mesh;
+		auto& mesh              = vertex_info[0].mesh;
 		mesh.host_subgroup_size = m_graphics.subgroup_size;
 		const auto& limits      = m_graphics.mesh_shader_properties;
 		const auto  logical_threads =
@@ -592,7 +607,7 @@ PipelineCache::GraphicsPrograms PipelineCache::GetGraphicsPrograms(
 	if (context.GetClipControl().clip_disable) {
 		const auto& viewport = context.GetScreenViewport().viewports[0];
 		const auto& limits   = m_graphics.GetPhysicalDeviceProperties().limits;
-		auto&       clip     = vertex_info.clip_space;
+		auto&       clip     = vertex_info[tess_active ? 2u : 0u].clip_space;
 		clip.scale[0]        = viewport.xscale;
 		clip.scale[1]        = viewport.yscale;
 		clip.offset[0]       = viewport.xoffset;
@@ -610,7 +625,9 @@ PipelineCache::GraphicsPrograms PipelineCache::GetGraphicsPrograms(
 	if (pixel_active) {
 		result.pixel = m_program_cache->Get(pixel_params, pixel_info, push_data_cursor);
 	}
-	result.vertex = m_program_cache->Get(vertex_params, vertex_info, push_data_cursor);
+	for (uint32_t i = 0; i < (tess_active ? 3u : 1u); i++) {
+		result.vertex[i] = m_program_cache->Get(vertex_params[i], vertex_info[i], push_data_cursor);
+	}
 	return result;
 }

@@ -630,10 +647,12 @@ bool PipelineStaticParameters::operator==(const PipelineStaticParameters& other)

 PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(
    std::span<const RenderColorInfo> colors, const RenderDepthInfo& depth,
-    const ShaderVertexInputInfo& vs_input_info, CommandBuffer& command,
+    std::span<const ShaderVertexInputInfo> vertex_info, CommandBuffer& command,
    const ShaderPixelInputInfo* ps_input_info, vk::PrimitiveTopology topology,
-    bool primitive_restart_enable, const ShaderProgram& vertex_program,
-    const ShaderProgram& pixel_program) {
+    bool primitive_restart_enable, const GraphicsPrograms& programs) {
+	const auto& vs_input_info  = vertex_info.front();
+	const auto& vertex_program = programs.vertex[0];
+	const auto& pixel_program  = programs.pixel;
 	KYTY_PROFILER_BLOCK("PipelineCache::CreatePipeline(Gfx)", profiler::colors::DeepOrangeA200);

 	EXIT_IF(colors.size() > RENDER_COLOR_ATTACHMENTS_MAX);
@@ -651,7 +670,9 @@ PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(
 	const auto ps_id = ps_active ? pixel_program.id : 0;

 	GraphicsPipelineKey key {};
-	key.vs_shader_id            = vs_id;
+	for (uint32_t i = 0; i < programs.vertex.size(); i++) {
+		key.vertex_shader_ids[i] = programs.vertex[i].id;
+	}
 	key.ps_shader_id            = ps_id;
 	auto& static_params         = key.static_params;
 	auto& rendering             = key.rendering;
@@ -729,7 +750,8 @@ PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(
 	static_params.stencil_test_enable      = depth.stencil_test_enable;
 	static_params.stencil_front            = depth.stencil_static_front;
 	static_params.stencil_back             = depth.stencil_static_back;
-	const bool rect_list     = topology == vk::PrimitiveTopology::ePatchList;
+	const bool rect_list =
+	    command.GetUserConfig().GetPrimType() == Prospero::PrimitiveType::kRectList;
 	static_params.cull_back  = !rect_list && mc.cull_back;
 	static_params.cull_front = !rect_list && mc.cull_front;
 	static_params.face       = mc.face;
@@ -780,9 +802,8 @@ PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(

 	auto cached = std::make_unique<Pipeline>();
 	LogPipelineTrace("CreatePipelineInternal begin", vs_id, ps_id);
-	CreatePipelineInternal(m_graphics, *cached, rendering, key.vertex_input, vs_input_info,
-	                       vertex_program, ps_input_info, pixel_program, static_params,
-	                       m_driver_cache);
+	CreatePipelineInternal(m_graphics, *cached, rendering, key.vertex_input, vertex_info,
+	                       ps_input_info, programs, static_params, m_driver_cache);
 	LogPipelineTrace("CreatePipelineInternal done", vs_id, ps_id);

 	EXIT_NOT_IMPLEMENTED(cached->pipeline == nullptr);
@@ -121,8 +121,10 @@ public:
 	};

 	struct GraphicsPrograms {
-		ShaderProgram vertex;
+		std::array<ShaderProgram, 3> vertex;
 		ShaderProgram pixel;
+
+		[[nodiscard]] uint32_t VertexStageCount() const { return vertex[1] ? 3u : 1u; }
 	};

 	GraphicsPrograms
@@ -130,18 +132,18 @@ public:
 	                    const HW::PixelShaderInfo& pixel_regs, const HW::ShaderRegisters& sh,
 	                    const HW::Context& context, const HW::UserConfig& user_config,
 	                    std::span<const Prospero::ColorComponentMapping, 8> target_export_mapping,
-	                    bool pixel_active, ShaderVertexInputInfo& vertex_info,
+	                    bool pixel_active, std::array<ShaderVertexInputInfo, 3>& vertex_info,
 	                    ShaderPixelInputInfo& pixel_info);
 	ShaderProgram GetComputeProgram(const HW::ComputeShaderInfo& regs,
 	                                const HW::ShaderRegisters&   sh,
 	                                ShaderComputeInputInfo&      input_info);

-	Pipeline&
-	GetGraphicsPipeline(std::span<const RenderColorInfo> colors, const RenderDepthInfo& depth,
-	                    const ShaderVertexInputInfo& vs_input_info, CommandBuffer& command,
-	                    const ShaderPixelInputInfo* ps_input_info,
-	                    vk::PrimitiveTopology topology, bool primitive_restart_enable,
-	                    const ShaderProgram& vertex_program, const ShaderProgram& pixel_program);
+	Pipeline& GetGraphicsPipeline(std::span<const RenderColorInfo>       colors,
+	                              const RenderDepthInfo&                 depth,
+	                              std::span<const ShaderVertexInputInfo> vertex_info,
+	                              CommandBuffer& command, const ShaderPixelInputInfo* ps_input_info,
+	                              vk::PrimitiveTopology topology, bool primitive_restart_enable,
+	                              const GraphicsPrograms& programs);
 	Pipeline& GetComputePipeline(const ShaderComputeInputInfo& input_info,
 	                             const ShaderProgram&          compute_program);

@@ -150,13 +152,13 @@ private:

 	struct GraphicsPipelineKey {
 		PipelineRenderingState   rendering;
-		uint64_t                 vs_shader_id = 0;
+		std::array<uint64_t, 3>  vertex_shader_ids {};
 		uint64_t                 ps_shader_id = 0;
 		PipelineVertexInputState vertex_input;
 		PipelineStaticParameters static_params;

 		bool operator==(const GraphicsPipelineKey& other) const {
-			return rendering == other.rendering && vs_shader_id == other.vs_shader_id &&
+			return rendering == other.rendering && vertex_shader_ids == other.vertex_shader_ids &&
 			       ps_shader_id == other.ps_shader_id && vertex_input == other.vertex_input &&
 			       static_params == other.static_params;
 		}
@@ -189,7 +191,9 @@ private:
 		std::size_t operator()(const GraphicsPipelineKey& key) const {
 			std::size_t hash = 0;
 			PipelineKeyHash::MixRendering(hash, key.rendering);
-			PipelineKeyHash::Mix(hash, key.vs_shader_id);
+			for (const auto id: key.vertex_shader_ids) {
+				PipelineKeyHash::Mix(hash, id);
+			}
 			PipelineKeyHash::Mix(hash, key.ps_shader_id);
 			PipelineKeyHash::Mix(hash, key.vertex_input.binding_count);
 			for (uint32_t i = 0; i < key.vertex_input.binding_count; i++) {
@@ -219,12 +223,14 @@ private:
 };

 void LogPipelineTrace(const char* phase, uint64_t vertex_program_id, uint64_t pixel_program_id);
-void CreatePipelineInternal(
-    GraphicContext& graphics, PipelineCache::Pipeline& pipeline,
-    const PipelineRenderingState& rendering, const PipelineVertexInputState& vertex_input,
-    const ShaderVertexInputInfo& vs_input_info, const ShaderProgram& vertex_program,
-    const ShaderPixelInputInfo* ps_input_info, const ShaderProgram& pixel_program,
-    const PipelineStaticParameters& static_params, vk::PipelineCache driver_cache);
+void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& pipeline,
+                            const PipelineRenderingState&          rendering,
+                            const PipelineVertexInputState&        vertex_input,
+                            std::span<const ShaderVertexInputInfo> vertex_info,
+                            const ShaderPixelInputInfo*            ps_input_info,
+                            const PipelineCache::GraphicsPrograms& programs,
+                            const PipelineStaticParameters&        static_params,
+                            vk::PipelineCache                      driver_cache);
 void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& pipeline,
                            const ShaderComputeInputInfo& input_info,
                            vk::ShaderModule compute_module, vk::PipelineCache driver_cache);
@@ -8,6 +8,20 @@

 namespace Libs::Graphics {

+vk::ShaderStageFlagBits NativeShaderStage(ShaderType stage) {
+	switch (stage) {
+		case ShaderType::Local:
+		case ShaderType::Vertex: return vk::ShaderStageFlagBits::eVertex;
+		case ShaderType::Mesh: return vk::ShaderStageFlagBits::eMeshEXT;
+		case ShaderType::TessellationControl: return vk::ShaderStageFlagBits::eTessellationControl;
+		case ShaderType::TessellationEvaluation:
+			return vk::ShaderStageFlagBits::eTessellationEvaluation;
+		case ShaderType::Pixel: return vk::ShaderStageFlagBits::eFragment;
+		case ShaderType::Compute: return vk::ShaderStageFlagBits::eCompute;
+		default: EXIT("unknown native shader stage\n");
+	}
+}
+
 vk::PipelineStageFlags ShaderPipelineStages(vk::ShaderStageFlags stages) {
 	vk::PipelineStageFlags result = {};
 	if (stages & vk::ShaderStageFlagBits::eVertex) {
@@ -16,6 +30,12 @@ vk::PipelineStageFlags ShaderPipelineStages(vk::ShaderStageFlags stages) {
 	if (stages & vk::ShaderStageFlagBits::eMeshEXT) {
 		result |= vk::PipelineStageFlagBits::eMeshShaderEXT;
 	}
+	if (stages & vk::ShaderStageFlagBits::eTessellationControl) {
+		result |= vk::PipelineStageFlagBits::eTessellationControlShader;
+	}
+	if (stages & vk::ShaderStageFlagBits::eTessellationEvaluation) {
+		result |= vk::PipelineStageFlagBits::eTessellationEvaluationShader;
+	}
 	if (stages & vk::ShaderStageFlagBits::eFragment) {
 		result |= vk::PipelineStageFlagBits::eFragmentShader;
 	}
@@ -8,6 +8,7 @@ namespace Libs::Graphics {

 struct ShaderStageRuntime;

+vk::ShaderStageFlagBits NativeShaderStage(ShaderType stage);
 vk::PipelineStageFlags  ShaderPipelineStages(vk::ShaderStageFlags stages);
 VulkanMemoryBarrier     MakeShaderAccessDependency();
 VulkanMemoryBarrier     MakeShaderWriteHazardDependency();
@@ -8,6 +8,7 @@
 #include "graphics/host_gpu/renderer/debug.h"
 #include "graphics/host_gpu/renderer/pipeline/descriptors.h"
 #include "graphics/host_gpu/renderer/pipeline/pipelineCache.h"
+#include "graphics/host_gpu/renderer/pipeline/shaderResourceBarrier.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
 #include "graphics/host_gpu/renderer/renderTarget.h"
@@ -205,12 +206,18 @@ static void CreateDescriptorLayout(GraphicContext& graphics, PipelineCache::Pipe
 }

 // NOLINTNEXTLINE(readability-function-cognitive-complexity)
-void CreatePipelineInternal(
-    GraphicContext& graphics, PipelineCache::Pipeline& pipeline,
-    const PipelineRenderingState& rendering, const PipelineVertexInputState& vertex_input,
-    const ShaderVertexInputInfo& vs_input_info, const ShaderProgram& vertex_program,
-    const ShaderPixelInputInfo* ps_input_info, const ShaderProgram& pixel_program,
-    const PipelineStaticParameters& static_params, vk::PipelineCache driver_cache) {
+void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& pipeline,
+                            const PipelineRenderingState&          rendering,
+                            const PipelineVertexInputState&        vertex_input,
+                            std::span<const ShaderVertexInputInfo> vertex_info,
+                            const ShaderPixelInputInfo*            ps_input_info,
+                            const PipelineCache::GraphicsPrograms& programs,
+                            const PipelineStaticParameters&        static_params,
+                            vk::PipelineCache                      driver_cache) {
+	const auto& vs_input_info  = vertex_info.front();
+	const auto& vertex_program = programs.vertex[0];
+	const auto& pixel_program  = programs.pixel;
+	const bool  tessellation   = vertex_info.size() == 3;
 	const bool ps_active = ps_input_info != nullptr;
 	EXIT_IF(!vertex_program || (ps_active && !pixel_program));
 	const bool with_depth = rendering.depth_format != vk::Format::eUndefined ||
@@ -218,10 +225,8 @@ void CreatePipelineInternal(
 	EXIT_IF(!vs_input_info.stage);
 	const bool mesh = vs_input_info.stage.program->stage == ShaderType::Mesh;
 	EXIT_NOT_IMPLEMENTED(mesh && !graphics.mesh_shader_enabled);
-	const auto vertex_stage =
-	    mesh ? vk::ShaderStageFlagBits::eMeshEXT : vk::ShaderStageFlagBits::eVertex;
-
-	const bool rect_list = !mesh && static_params.topology == vk::PrimitiveTopology::ePatchList;
+	const bool rect_list =
+	    !mesh && !tessellation && static_params.topology == vk::PrimitiveTopology::ePatchList;

 	vk::ShaderModule tess_control_shader_module = nullptr;
 	vk::ShaderModule tess_eval_shader_module    = nullptr;
@@ -255,9 +260,14 @@ void CreatePipelineInternal(
 	EXIT_NOT_IMPLEMENTED(
 	    rect_list && (tess_control_shader_module == nullptr || tess_eval_shader_module == nullptr));

-	vk::PipelineShaderStageCreateInfo shader_stages[4] = {
-	    {.stage = vertex_stage, .module = vertex_program.module, .pName = "main"}};
-	uint32_t shader_stage_count = 1;
+	vk::PipelineShaderStageCreateInfo shader_stages[4] {};
+	uint32_t                          shader_stage_count = 0;
+	for (uint32_t i = 0; i < vertex_info.size(); i++) {
+		shader_stages[shader_stage_count++] = {.stage =
+		                                           NativeShaderStage(vertex_info[i].logical_stage),
+		                                       .module = programs.vertex[i].module,
+		                                       .pName  = "main"};
+	}
 	if (rect_list) {
 		shader_stages[shader_stage_count++] = {.stage =
 		                                           vk::ShaderStageFlagBits::eTessellationControl,
@@ -432,15 +442,19 @@ void CreatePipelineInternal(
 	color_blending.pAttachments    = color_blend_attachment;

 	std::vector<vk::DescriptorSetLayoutBinding> descriptor_bindings;
-	AddLayoutBindings(descriptor_bindings, *vs_input_info.stage.program, vertex_stage);
+	vk::ShaderStageFlags graphics_stages = vk::ShaderStageFlagBits::eFragment;
+	for (const auto& stage: vertex_info) {
+		const auto native_stage = NativeShaderStage(stage.logical_stage);
+		AddLayoutBindings(descriptor_bindings, *stage.stage.program, native_stage);
+		graphics_stages |= native_stage;
+	}
 	if (ps_active) {
 		EXIT_IF(!ps_input_info->stage);
 		AddLayoutBindings(descriptor_bindings, *ps_input_info->stage.program,
 		                  vk::ShaderStageFlagBits::eFragment);
 	}
 	CreateDescriptorLayout(graphics, pipeline, descriptor_bindings);
-	const auto                  GraphicsStages = vertex_stage | vk::ShaderStageFlagBits::eFragment;
-	const vk::PushConstantRange push_constants {GraphicsStages, 0,
+	const vk::PushConstantRange push_constants {graphics_stages, 0,
 	                                            ShaderRecompiler::IR::NativePushConstantSize};

 	vk::PipelineLayoutCreateInfo pipeline_layout_info {};
@@ -524,8 +538,9 @@ void CreatePipelineInternal(
 	pipeline_info.pVertexInputState        = mesh ? nullptr : &vertex_input_info;
 	pipeline_info.pInputAssemblyState      = mesh ? nullptr : &input_assembly;
 	vk::PipelineTessellationStateCreateInfo tessellation_state {};
-	tessellation_state.patchControlPoints = 3;
-	pipeline_info.pTessellationState      = (rect_list ? &tessellation_state : nullptr);
+	tessellation_state.patchControlPoints =
+	    tessellation ? vs_input_info.tess.input_control_points : 3u;
+	pipeline_info.pTessellationState = (rect_list || tessellation) ? &tessellation_state : nullptr;
 	pipeline_info.pViewportState          = &viewport_state;
 	pipeline_info.pRasterizationState     = &rasterizer;
 	pipeline_info.pMultisampleState       = &multisampling;
@@ -170,7 +170,7 @@ private:
 	void DrawAuto(uint64_t submit_id, CommandBuffer& buffer, const DrawAutoArgs& args);

 	struct GraphicsBindings {
-		PreparedBindings                vertex;
+		std::array<PreparedBindings, 3> vertex;
 		std::optional<PreparedBindings> pixel;
 	};

@@ -453,7 +453,7 @@ struct DrawRenderState {
 	RenderColorInfo       color_info[RENDER_COLOR_ATTACHMENTS_MAX] = {};
 	uint32_t              color_count                              = 0;
 	bool                  ps_active                                = true;
-	ShaderVertexInputInfo vs_input_info;
+	std::array<ShaderVertexInputInfo, 3> vertex_info;
 	ShaderPixelInputInfo  ps_input_info;
 	PipelineCache::GraphicsPrograms programs;
 };
@@ -831,6 +831,7 @@ static bool GetDrawTopology(const HW::UserConfig& ucfg, bool auto_draw,
 		case Prospero::PrimitiveType::kTriStrip:
 			topology = vk::PrimitiveTopology::eTriangleStrip;
 			break;
+		case Prospero::PrimitiveType::kPatch:
 		case Prospero::PrimitiveType::kRectList:
 			topology = vk::PrimitiveTopology::ePatchList;
 			break;
@@ -953,7 +954,7 @@ static void RefreshShaders(CommandBuffer& buffer, const DrawCallInfo& draw,
 	}
 	state.programs = pipeline_cache.GetGraphicsPrograms(
 	    vertex_shader_info, pixel_shader_info, shader_regs, ctx, buffer.GetUserConfig(),
-	    target_export_mapping, state.ps_active, state.vs_input_info, state.ps_input_info);
+	    target_export_mapping, state.ps_active, state.vertex_info, state.ps_input_info);
 }

 static PreparedIndexBuffer PrepareIndexBuffer(CommandBuffer&               buffer,
@@ -1016,7 +1017,7 @@ static void LogDrawStateIfNeeded(const CommandBuffer& buffer, const DrawCallInfo
 		LogDrawTargetState(draw.Name(), state.color_info[0], state.depth_info, buffer,
 		                   state.ps_input_info, draw.index_count, 0);
 	}
-	LogDrawInputState(buffer, state.color_info[0], state.vs_input_info, index_type_and_size,
+	LogDrawInputState(buffer, state.color_info[0], state.vertex_info[0], index_type_and_size,
 	                  draw.index_count, index_addr);
 }

@@ -1031,6 +1032,7 @@ static void EmitDrawPrimitives(const HW::UserConfig& ucfg, vk::CommandBuffer vk_
 		case Prospero::PrimitiveType::kTriFan:
 		case Prospero::PrimitiveType::kTriStrip:
 		case Prospero::PrimitiveType::kRectList:
+		case Prospero::PrimitiveType::kPatch:
 			if (draw.IsIndexed()) {
 				vk_buffer.drawIndexed(draw.index_count, draw.instance_count, 0, emit.vertex_offset,
 				                      emit.first_instance);
@@ -1069,10 +1071,12 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
                                         const DrawIndexBufferSource& index_source,
 	                                     bool primitive_restart_enable) {
 	auto& ucfg = buffer.GetUserConfig();
-	const bool mesh_active = state.vs_input_info.stage.program->stage == ShaderType::Mesh;
+	const auto vertex_stages =
+	    std::span {state.vertex_info.data(), state.programs.VertexStageCount()};
+	const bool mesh_active = state.vertex_info[0].stage.program->stage == ShaderType::Mesh;
 	uint32_t   mesh_groups = 0;
 	if (mesh_active) {
-		const auto& mesh = state.vs_input_info.mesh;
+		const auto& mesh = state.vertex_info[0].mesh;
 		if (primitive_restart_enable || mesh.primitives_per_group == 0) {
 			EXIT("unsupported mesh draw: primitive=%u indexed=%u restart=%u\n",
 			     static_cast<uint32_t>(ucfg.GetPrimType()), draw.IsIndexed(), primitive_restart_enable);
@@ -1100,19 +1104,24 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 		                              index_source.guest_element_size);
 	}
 	LogDrawPhase(draw.Name(), "PrepareBindings");
-	GraphicsBindings bindings {.vertex = PrepareBindings(state.vs_input_info.stage)};
-	std::array<PreparedBindings*, 2> descriptor_stages {&bindings.vertex, nullptr};
+	GraphicsBindings                 bindings;
+	std::array<PreparedBindings*, 4> descriptor_stages {};
+	uint32_t                         stage_count = 0;
+	for (uint32_t i = 0; i < vertex_stages.size(); i++) {
+		bindings.vertex[i]               = PrepareBindings(state.vertex_info[i].stage);
+		descriptor_stages[stage_count++] = &bindings.vertex[i];
+	}
 	if (state.ps_active) {
 		bindings.pixel.emplace(PrepareBindings(state.ps_input_info.stage));
-		descriptor_stages[1] = &*bindings.pixel;
+		descriptor_stages[stage_count++] = &*bindings.pixel;
 	}
-	const auto stages = std::span {descriptor_stages.data(), state.ps_active ? 2u : 1u};
+	const auto stages = std::span {descriptor_stages.data(), stage_count};
 	PrepareGraphicsBindings(stages);
 	PreparedVertexBuffers vertex_bindings;
 	PreparedIndexBuffer   index_binding;
 	if (!mesh_active) {
 		LogDrawPhase(draw.Name(), "PrepareVertexBuffers");
-		vertex_bindings = AcquireVertexBuffers(buffer, state.vs_input_info);
+		vertex_bindings = AcquireVertexBuffers(buffer, state.vertex_info[0]);
 		index_binding   = PrepareIndexBuffer(buffer, index_source);
 	}
 	const auto rendering =
@@ -1123,9 +1132,9 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 		LogDrawPhase(draw.Name(), "CreatePipeline");
 	}
 	auto& pipeline = m_context.GetPipelineCache().GetGraphicsPipeline(
-	    std::span {state.color_info, state.color_count}, state.depth_info, state.vs_input_info, buffer,
+	    std::span {state.color_info, state.color_count}, state.depth_info, vertex_stages, buffer,
 	    state.ps_active ? &state.ps_input_info : nullptr, topology, primitive_restart_enable,
-	    state.programs.vertex, state.programs.pixel);
+	    state.programs);

 	// Resource preparation above may synchronously finish and restart the scheduler. From this
 	// point onward, every operation targets the current command buffer and cannot touch guest
@@ -1155,7 +1164,7 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 		CommitIndexBuffer(vk_buffer, index_binding);
 	}

-	SetGraphicsDynamicParams(buffer, vk_buffer, state.vs_input_info, state.color_info,
+	SetGraphicsDynamicParams(buffer, vk_buffer, vertex_stages.back(), state.color_info,
 	                         state.color_count, state.depth_info);
 	if (m_context.GetGraphics().attachment_feedback_loop_enabled) {
 		vk_buffer.setAttachmentFeedbackLoopEnableEXT(
@@ -1177,16 +1186,17 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 	if (mesh_active) {
 		vk_buffer.drawMeshTasksEXT(mesh_groups, draw.instance_count, 1);
 	} else {
-		EmitDrawPrimitives(ucfg, vk_buffer, state.vs_input_info, draw, emit);
+		EmitDrawPrimitives(ucfg, vk_buffer, state.vertex_info[0], draw, emit);
 	}

 	if (!draw.IsIndexed()) {
 		SetDrawDebugPhase(buffer, submit_id, draw, 0x600u);
 	}
 	vk::PipelineStageFlags shader_write_stages = {};
-	if (HasShaderBufferWrites(state.vs_input_info.stage)) {
-		shader_write_stages |= mesh_active ? vk::PipelineStageFlagBits::eMeshShaderEXT
-		                                   : vk::PipelineStageFlagBits::eVertexShader;
+	for (const auto& stage: vertex_stages) {
+		if (HasShaderBufferWrites(stage.stage)) {
+			shader_write_stages |= ShaderPipelineStages(NativeShaderStage(stage.logical_stage));
+		}
 	}
 	if (state.ps_active && HasShaderBufferWrites(state.ps_input_info.stage)) {
 		shader_write_stages |= vk::PipelineStageFlagBits::eFragmentShader;
@@ -1302,7 +1312,7 @@ void RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
 	const bool indirect = args.offset_source == DrawOffsetSource::IndirectArgs;
 	const auto [vertex_offset, instance_offset] =
 	    indirect ? std::pair<int32_t, uint32_t> {0, args.first_instance}
-	             : ResolveDrawOffsets(ucfg.GetIndexOffset(), state.vs_input_info);
+	             : ResolveDrawOffsets(ucfg.GetIndexOffset(), state.vertex_info[0]);

 	DrawEmitInfo emit {};
 	emit.vertex_offset  = vertex_offset + args.base_vertex;
@@ -1374,9 +1384,9 @@ void RenderExecutor::DrawAuto(uint64_t submit_id, CommandBuffer& buffer, const D
 	}
 	RefreshShaders(buffer, draw, state);

-	const bool rect_list = topology == vk::PrimitiveTopology::ePatchList;
-	if (rect_list && state.vs_input_info.buffers_num == 0 &&
-	    state.vs_input_info.stage.program->param_export_mask == 0 &&
+	const bool rect_list = ucfg.GetPrimType() == Prospero::PrimitiveType::kRectList;
+	if (rect_list && state.vertex_info[0].buffers_num == 0 &&
+	    state.vertex_info[0].stage.program->param_export_mask == 0 &&
 	    state.ps_input_info.input_num != 0) {
 		if (graphics_debug_dump_enabled()) {
 			LOGF("DrawIndexAuto: skipping rect-list draw with no VS param exports and PS inputs: "
@@ -1393,7 +1403,7 @@ void RenderExecutor::DrawAuto(uint64_t submit_id, CommandBuffer& buffer, const D
 	const bool indirect = args.offset_source == DrawOffsetSource::IndirectArgs;
 	const auto [vertex_offset, instance_offset] =
 	    indirect ? std::pair<int32_t, uint32_t> {0, args.first_instance}
-	             : ResolveDrawOffsets(ucfg.GetIndexOffset(), state.vs_input_info);
+	             : ResolveDrawOffsets(ucfg.GetIndexOffset(), state.vertex_info[0]);
 	DrawEmitInfo emit {};
 	emit.first_vertex = static_cast<uint32_t>(vertex_offset + static_cast<int32_t>(args.first_vertex));
 	emit.first_instance = instance_offset;
@@ -47,6 +47,9 @@ const char* StageName(ShaderType stage) {
 	switch (stage) {
 		case ShaderType::Compute: return "CS";
 		case ShaderType::Vertex: return "VS";
+		case ShaderType::Local: return "LS";
+		case ShaderType::TessellationControl: return "HS";
+		case ShaderType::TessellationEvaluation: return "TES";
 		case ShaderType::Mesh: return "MS";
 		case ShaderType::Pixel: return "PS";
 		default: return "unknown";
@@ -211,6 +214,8 @@ EmbeddedFetchData DetectEmbeddedVertexFetch(const Decoder::Program&      decoded
                                            const ShaderVertexInputInfo* input_info,
                                            uint32_t user_data_base, uint32_t user_data_count,
                                            uint32_t wave_size) {
+	const uint32_t    vertex_index_reg   = input_info->logical_stage == ShaderType::Local ? 2u : 5u;
+	const uint32_t    instance_index_reg = input_info->logical_stage == ShaderType::Local ? 5u : 8u;
 	EmbeddedFetchData data;
 	data.loads.reserve(input_info->resources_num);
 	int32_t vertex_offset_candidate   = -1;
@@ -251,16 +256,17 @@ EmbeddedFetchData DetectEmbeddedVertexFetch(const Decoder::Program&      decoded
 		// corresponding direct-draw offsets before fetching.
 		const bool vertex_index_accumulator =
 		    IsDecodedVgpr(inst.dst) &&
-		    (inst.dst.reg == 0 || (user_data_base == 8 && inst.dst.reg == 5));
+		    (inst.dst.reg == 0 || (user_data_base == 8 && inst.dst.reg == vertex_index_reg));
 		const bool instance_index_accumulator =
-		    IsDecodedVgpr(inst.dst) && (inst.dst.reg == (user_data_base == 8 ? 8u : 3u));
+		    IsDecodedVgpr(inst.dst) &&
+		    (inst.dst.reg == (user_data_base == 8 ? instance_index_reg : 3u));
 		uint32_t   sad_zero = 0;
 		const bool index_offset_add =
-		    (vertex_index_accumulator || instance_index_accumulator) &&
-		    IsDecodedSgpr(inst.src0) &&
+		    (vertex_index_accumulator || instance_index_accumulator) && IsDecodedSgpr(inst.src0) &&
 		    ((inst.opcode == Decoder::Opcode::V_ADD_I32 && IsDecodedVgpr(inst.src1) &&
 		      inst.src1.reg == inst.dst.reg) ||
-		     (user_data_base == 8 && (inst.dst.reg == 5 || inst.dst.reg == 8) &&
+		     (user_data_base == 8 &&
+		      (inst.dst.reg == vertex_index_reg || inst.dst.reg == instance_index_reg) &&
 		      inst.opcode == Decoder::Opcode::V_SAD_U32 && IsDecodedVgpr(inst.src2) &&
 		      inst.src2.reg == inst.dst.reg &&
 		      TryDecodedOperandConstant(sgprs, inst.src1, sad_zero) && sad_zero == 0));
@@ -391,8 +397,8 @@ EmbeddedFetchData DetectEmbeddedVertexFetch(const Decoder::Program&      decoded
 						ClearEmbeddedFetchVectorLanes(&vector_lanes, inst.dst.reg);
 					}
 					if (IsDecodedVgpr(inst.dst) && inst.dst.reg < vgpr_is_index.size() &&
-					    IsDecodedVgpr(inst.src0) && inst.src0.reg == 8 &&
-					    IsDecodedVgpr(inst.src1) && inst.src1.reg == 5) {
+					    IsDecodedVgpr(inst.src0) && inst.src0.reg == instance_index_reg &&
+					    IsDecodedVgpr(inst.src1) && inst.src1.reg == vertex_index_reg) {
 						vgpr_is_index[inst.dst.reg] = true;
 					}
 				} else if (IsEmbeddedFetchAttribPropagationAlu(inst)) {
@@ -459,9 +465,7 @@ EmbeddedFetchData DetectEmbeddedVertexFetch(const Decoder::Program&      decoded
 	return data;
 }

-Decoder::Program DecodeFusedProgram(std::span<const uint32_t> front, std::span<const uint32_t> back,
-                                    std::vector<uint32_t>& joined_code) {
-	EXIT_IF(back.empty());
+Decoder::Program DecodeFrontProgram(std::span<const uint32_t> front) {
 	Decoder::Program result;
 	uint32_t         front_words = 0;
 	while (front_words < front.size()) {
@@ -477,7 +481,16 @@ Decoder::Program DecodeFusedProgram(std::span<const uint32_t> front, std::span<c
 	}
 	EXIT_IF(result.instructions.empty() ||
 	        result.instructions.back().opcode != Decoder::Opcode::S_SETPC_B64);
-	joined_code.assign(front.begin(), front.begin() + front_words);
+	result.code = front.first(front_words);
+	return result;
+}
+
+Decoder::Program DecodeFusedProgram(std::span<const uint32_t> front, std::span<const uint32_t> back,
+                                    std::vector<uint32_t>& joined_code) {
+	EXIT_IF(back.empty());
+	auto       result      = DecodeFrontProgram(front);
+	const auto front_words = static_cast<uint32_t>(result.code.size());
+	joined_code.assign(result.code.begin(), result.code.end());
 	joined_code.insert(joined_code.end(), back.begin(), back.end());
 	// The merged-stage ABI passes the back shader in s[6:7]. Give that handoff an
 	// ordinary CFG edge, retaining both bodies in one register and LDS lifetime.
@@ -500,12 +513,224 @@ Decoder::Program DecodeFusedProgram(std::span<const uint32_t> front, std::span<c

 } // namespace

+void AnalyzeTessellationPrograms(std::span<const uint32_t> local, std::span<const uint32_t> control,
+                                 ShaderTessellationInputInfo& info) {
+	using namespace Decoder;
+	const auto constant = [](const Operand& operand, uint32_t value) {
+		return (operand.kind == OperandKind::IntegerInlineConstant ||
+		        operand.kind == OperandKind::LiteralConstant) &&
+		       operand.value == value;
+	};
+	auto     decoded       = DecodeFrontProgram(local);
+	uint32_t local_address = UINT32_MAX;
+	for (const auto& inst: decoded.instructions) {
+		if (inst.opcode == Opcode::V_MUL_U32_U24 && inst.src1.kind == OperandKind::Vgpr &&
+		    inst.src1.reg == 3u &&
+		    (inst.src0.kind == OperandKind::IntegerInlineConstant ||
+		     inst.src0.kind == OperandKind::LiteralConstant)) {
+			EXIT_IF(info.ls_stride != 0u && info.ls_stride != inst.src0.value);
+			info.ls_stride = inst.src0.value;
+			local_address  = inst.dst.reg;
+		}
+		if (inst.family == Family::DS) {
+			EXIT_NOT_IMPLEMENTED(inst.src0.kind != OperandKind::Vgpr ||
+			                     inst.src0.reg != local_address);
+		}
+	}
+	decoded = {};
+	DecodeProgram(control, decoded);
+	uint32_t control_point   = UINT32_MAX;
+	uint32_t control_address = UINT32_MAX;
+	for (const auto& inst: decoded.instructions) {
+		if (inst.opcode == Opcode::V_BFE_U32 && inst.src0.kind == OperandKind::Vgpr &&
+		    inst.src0.reg == 1u && constant(inst.src1, 8u) && constant(inst.src2, 5u)) {
+			control_point = inst.dst.reg;
+		}
+		if (inst.opcode == Opcode::V_LSHL_ADD_U32 && inst.src0.kind == OperandKind::Vgpr &&
+		    inst.src0.reg == control_point &&
+		    (inst.src1.kind == OperandKind::IntegerInlineConstant ||
+		     inst.src1.kind == OperandKind::LiteralConstant) &&
+		    inst.src1.value < 32u) {
+			info.hs_stride  = 1u << inst.src1.value;
+			control_address = inst.dst.reg;
+		}
+	}
+	EXIT_NOT_IMPLEMENTED(local_address == UINT32_MAX || control_address == UINT32_MAX ||
+	                     info.ls_stride == 0u || info.hs_stride == 0u ||
+	                     (info.ls_stride & 3u) != 0u || (info.hs_stride & 3u) != 0u);
+	LOGF("Tessellation interface: input_cp=%u output_cp=%u ls_stride=%u hs_stride=%u\n",
+	     info.input_control_points, info.output_control_points, info.ls_stride, info.hs_stride);
+}
+
+namespace {
+
+const IR::Inst* TessellationBufferBase(const IR::Inst& inst) {
+	if (IR::BufferAccessOf(inst.GetOpcode()) == IR::BufferAccess::None || inst.NumArgs() <= 3u)
+		return nullptr;
+	const auto* base = inst.Arg(3).Resolve().TryInstruction();
+	return base != nullptr && base->GetOpcode() == IR::ValueOpcode::TessellationBase ? base
+	                                                                                 : nullptr;
+}
+
+IR::Value ActiveAddress(IR::Block& block, IR::Block::iterator before, IR::Value value,
+                        IR::Value predicate, std::unordered_map<IR::Inst*, IR::Value>& resolved) {
+	using namespace IR;
+	value        = value.Resolve();
+	auto* source = value.TryInstruction();
+	if (source == nullptr) return value;
+	if (const auto found = resolved.find(source); found != resolved.end()) return found->second;
+	if (source->GetOpcode() == ValueOpcode::SelectU32) {
+		return source->Arg(0).Resolve() == predicate.Resolve()
+		           ? ActiveAddress(block, before, source->Arg(1), predicate, resolved)
+		           : value;
+	}
+	if (source->GetOpcode() != ValueOpcode::IAdd32 && source->GetOpcode() != ValueOpcode::ISub32)
+		return value;
+	const auto lhs = ActiveAddress(block, before, source->Arg(0), predicate, resolved);
+	const auto rhs = ActiveAddress(block, before, source->Arg(1), predicate, resolved);
+	if (lhs != source->Arg(0).Resolve() || rhs != source->Arg(1).Resolve()) {
+		auto copy = block.PrependNewInst(before, source->GetOpcode(), {lhs, rhs},
+		                                 source->Flags<uint64_t>());
+		value     = Value(&*copy);
+	}
+	resolved.emplace(source, value);
+	return value;
+}
+
+void LowerTessellationMemory(IR::Program& program, const CompileOptions& options) {
+	using namespace IR;
+	if (options.stage != ShaderType::Local && options.stage != ShaderType::TessellationControl &&
+	    options.stage != ShaderType::TessellationEvaluation) {
+		return;
+	}
+	const auto& tess = options.input_info.vertex->tess;
+	// Ring addresses can reuse data VGPRs. Their inactive values are irrelevant to
+	// a store guarded by the same EXEC predicate, but must remain intact elsewhere.
+	for (auto* block: program.blocks) {
+		for (auto it = block->begin(); it != block->end(); ++it) {
+			if (TessellationBufferBase(*it) == nullptr) continue;
+			std::unordered_map<Inst*, Value> resolved;
+			it->SetArg(
+			    2, ActiveAddress(*block, it, it->Arg(2), it->Arg(it->NumArgs() - 1u), resolved));
+		}
+	}
+	ConstantPropagationPass(program.blocks);
+	uint32_t reads = 0, writes = 0, factors = 0;
+	for (auto* block: program.blocks) {
+		for (auto it = block->begin(); it != block->end(); ++it) {
+			auto&      inst   = *it;
+			const auto shared = SharedAccessOf(inst.GetOpcode());
+			const auto buffer = BufferAccessOf(inst.GetOpcode());
+			if (shared == SharedAccess::None && buffer == BufferAccess::None) {
+				continue;
+			}
+			const auto&           memory = program.memory_info.at(inst.Flags<MemoryFlags>().index);
+			bool                  write  = false;
+			uint32_t              components = 0;
+			TessellationAttribute kind;
+			Value                 address, predicate;
+			if (shared != SharedAccess::None && memory.kind == ResourceKind::Lds) {
+				write = shared == SharedAccess::Write;
+				EXIT_NOT_IMPLEMENTED(memory.data_bits != 32u ||
+				                     (options.stage == ShaderType::Local
+				                          ? !write
+				                          : options.stage != ShaderType::TessellationControl ||
+				                                shared != SharedAccess::Read));
+				kind       = write ? TessellationAttribute::LocalOutput
+				                   : TessellationAttribute::ControlInput;
+				components = SharedComponentCount(inst.GetOpcode());
+				address    = inst.Arg(0);
+				predicate  = inst.Arg(inst.NumArgs() - 1u);
+			} else if (buffer != BufferAccess::None && memory.kind == ResourceKind::Buffer) {
+				const auto* base = TessellationBufferBase(inst);
+				if (base == nullptr) {
+					continue;
+				}
+				EXIT_NOT_IMPLEMENTED(memory.data_bits != 32u || memory.formatted || memory.idxen ||
+				                     !memory.offen || buffer == BufferAccess::Atomic);
+				write      = buffer == BufferAccess::Write;
+				components = BufferComponentCount(inst.GetOpcode());
+				address    = inst.Arg(2).Resolve();
+				predicate  = inst.Arg(inst.NumArgs() - 1u);
+				if (base->Arg(0).U32() == 1u) {
+					EXIT_NOT_IMPLEMENTED(!write ||
+					                     options.stage != ShaderType::TessellationControl);
+					kind = TessellationAttribute::Factor;
+					factors += components;
+				} else {
+					kind = write ? TessellationAttribute::ControlOutput
+					             : TessellationAttribute::EvaluationInput;
+					EXIT_NOT_IMPLEMENTED(write
+					                         ? options.stage != ShaderType::TessellationControl
+					                         : options.stage != ShaderType::TessellationEvaluation);
+					if (address.IsImmediate() &&
+					    address.U32() >= tess.hs_stride * tess.output_control_points) {
+						EXIT_NOT_IMPLEMENTED(!write);
+						kind = TessellationAttribute::PatchOutput;
+					}
+				}
+			} else {
+				continue;
+			}
+			const auto emit = [&](ValueOpcode opcode, std::initializer_list<Value> args) {
+				return Value(&*block->PrependNewInst(it, opcode, args));
+			};
+			std::array<Value, 4> values;
+			for (uint32_t component = 0; component < components; component++) {
+				const auto offset = memory.offset + 4u * component;
+				const auto byte_address =
+				    offset == 0u ? address : emit(ValueOpcode::IAdd32, {address, Value(offset)});
+				if (write) {
+					Value data = shared != SharedAccess::None ? inst.Arg(component + 1u)
+					             : components == 1u
+					                 ? inst.Arg(4)
+					                 : emit(components == 2u   ? ValueOpcode::CompositeExtractU32x2
+					                        : components == 3u ? ValueOpcode::CompositeExtractU32x3
+					                                           : ValueOpcode::CompositeExtractU32x4,
+					                        {inst.Arg(4), Value(component)});
+					emit(ValueOpcode::SetTessellationAttribute,
+					     {Value(static_cast<uint32_t>(kind)), byte_address, data, predicate});
+					writes++;
+				} else {
+					values[component] =
+					    emit(ValueOpcode::GetTessellationAttribute,
+					         {Value(static_cast<uint32_t>(kind)), byte_address, predicate});
+					reads++;
+				}
+			}
+			if (!write) {
+				Value replacement = values[0];
+				if (components == 2u)
+					replacement =
+					    emit(ValueOpcode::CompositeConstructU32x2, {values[0], values[1]});
+				if (components == 3u)
+					replacement = emit(ValueOpcode::CompositeConstructU32x3,
+					                   {values[0], values[1], values[2]});
+				if (components == 4u)
+					replacement = emit(ValueOpcode::CompositeConstructU32x4,
+					                   {values[0], values[1], values[2], values[3]});
+				inst.ReplaceUsesWith(replacement);
+			}
+			inst.Invalidate();
+		}
+	}
+	ConstantPropagationPass(program.blocks);
+	RemoveIdentities(program.blocks);
+	EliminateDeadCode(program.blocks);
+	LOGF("%s tessellation lowering: reads=%u writes=%u factors=%u\n", StageName(options.stage),
+	     reads, writes, factors);
+}
+
+} // namespace
+
 TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOptions& options) {
 	if (code.empty()) {
 		EXIT("shader recompiler input is empty\n");
 	}
 	if (options.stage != ShaderType::Compute && options.stage != ShaderType::Vertex &&
-	    options.stage != ShaderType::Pixel && options.stage != ShaderType::Mesh) {
+	    options.stage != ShaderType::Pixel && options.stage != ShaderType::Mesh &&
+	    options.stage != ShaderType::Local && options.stage != ShaderType::TessellationControl &&
+	    options.stage != ShaderType::TessellationEvaluation) {
 		EXIT("shader recompiler received unsupported stage %u\n",
 		     static_cast<unsigned>(options.stage));
 	}
@@ -525,6 +750,12 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 	std::vector<uint32_t> joined_code;
 	if (!options.back_code.empty()) {
 		decoded = DecodeFusedProgram(code, options.back_code, joined_code);
+	} else if (options.stage == ShaderType::Local) {
+		decoded = DecodeFrontProgram(code);
+		// The separately compiled hull half runs in the next Vulkan stage.
+		auto& handoff     = decoded.instructions.back();
+		handoff.opcode    = Decoder::Opcode::S_ENDPGM;
+		handoff.src_count = 0;
 	} else {
 		Decoder::DecodeProgram(code, decoded);
 	}
@@ -568,8 +799,8 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 	}

 	EmbeddedFetchData embedded_fetch;
-	if (options.stage == ShaderType::Vertex && options.input_info.vertex != nullptr &&
-	    options.input_info.vertex->fetch_embedded) {
+	if ((options.stage == ShaderType::Vertex || options.stage == ShaderType::Local) &&
+	    options.input_info.vertex != nullptr && options.input_info.vertex->fetch_embedded) {
 		embedded_fetch = DetectEmbeddedVertexFetch(
 		    decoded, options.input_info.vertex, options.user_data_base,
 		    static_cast<uint32_t>(options.user_data.size()), options.wave_size);
@@ -608,6 +839,7 @@ TranslateResult TranslateProgram(std::span<const uint32_t> code, const CompileOp
 		IR::RemoveIdentities(ir.blocks);
 		IR::EliminateDeadCode(ir.blocks);
 	}
+	LowerTessellationMemory(ir, options);
 	IR::BuildSrtPlan(ir);
 	IR::EliminateDeadCode(ir.blocks);
 	IR::TrackResources(ir);
@@ -37,6 +37,9 @@ struct CompileResult {
 	IR::Program            program;
 };

+void AnalyzeTessellationPrograms(std::span<const uint32_t> local, std::span<const uint32_t> control,
+                                 ShaderTessellationInputInfo& info);
+
 [[nodiscard]] TranslateResult TranslateProgram(std::span<const uint32_t> code,
                                               const CompileOptions& options);
 [[nodiscard]] CompileResult CompileProgram(TranslateResult translated,
@@ -277,7 +277,8 @@ Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program
 					break;
 				}
 				case IR::ValueOpcode::LaneId:
-					requirements.subgroup_local_invocation_id = true;
+					requirements.subgroup_local_invocation_id |=
+					    program.stage != ShaderType::TessellationControl;
 					break;
 				case IR::ValueOpcode::ImageQueryLod: requirements.compute_derivatives = true; break;
 				case IR::ValueOpcode::ImageGatherRaw:
@@ -306,8 +307,10 @@ std::vector<uint32_t> EmitProgram(const IR::Program& program,
 	using namespace Emitter;

 	if (program.stage != ShaderType::Compute && program.stage != ShaderType::Vertex &&
-	    program.stage != ShaderType::Pixel && program.stage != ShaderType::Mesh) {
-		Fail(program, "binary SPIR-V emitter supports compute, vertex, and pixel shaders");
+	    program.stage != ShaderType::Pixel && program.stage != ShaderType::Mesh &&
+	    program.stage != ShaderType::Local && program.stage != ShaderType::TessellationControl &&
+	    program.stage != ShaderType::TessellationEvaluation) {
+		Fail(program, "binary SPIR-V emitter received an unsupported shader stage");
 	}
 	if (!program.srt_plan_complete || !program.resource_tracking_complete ||
 	    !program.shader_info_complete || !program.binding_layout_complete) {
@@ -250,7 +250,10 @@ void EmitStorageImageWrite(EmitterState& state, uint32_t resource, uint32_t mip_

 spv::ExecutionModel ExecutionModelForStage(ShaderType stage) {
 	switch (stage) {
+		case ShaderType::Local:
 		case ShaderType::Vertex: return spv::ExecutionModelVertex;
+		case ShaderType::TessellationControl: return spv::ExecutionModelTessellationControl;
+		case ShaderType::TessellationEvaluation: return spv::ExecutionModelTessellationEvaluation;
 		case ShaderType::Mesh: return spv::ExecutionModelMeshEXT; // MeshEXT
 		case ShaderType::Pixel: return spv::ExecutionModelFragment;
 		default: return spv::ExecutionModelGLCompute;
@@ -51,6 +51,7 @@ uint32_t EmitBuiltinU32(EmitterState& state, IR::StageInputKind kind, uint32_t c
 		return bits;
 	}
 	if (kind == IR::StageInputKind::VertexIndex || kind == IR::StageInputKind::InstanceIndex ||
+	    kind == IR::StageInputKind::InvocationId || kind == IR::StageInputKind::PrimitiveId ||
 	    kind == IR::StageInputKind::Layer || kind == IR::StageInputKind::SampleId) {
 		const auto value = state.builder.AllocateId();
 		const auto bits  = state.builder.AllocateId();
@@ -58,7 +59,7 @@ uint32_t EmitBuiltinU32(EmitterState& state, IR::StageInputKind kind, uint32_t c
 		state.builder.AddFunction(spv::OpBitcast, TypeU32(state), bits, value);
 		return bits;
 	}
-	if (kind == IR::StageInputKind::FragCoord) {
+	if (kind == IR::StageInputKind::FragCoord || kind == IR::StageInputKind::TessCoord) {
 		const auto pointer = state.builder.AllocateId();
 		const auto value   = state.builder.AllocateId();
 		const auto bits    = state.builder.AllocateId();
@@ -138,7 +139,7 @@ uint32_t EmitAttribute(EmitterState& state, uint32_t attr, uint32_t chan) {
 	if (input == nullptr || input->variable_id == 0) {
 		return ConstantU32(state, 0);
 	}
-	if (state.program.stage == ShaderType::Vertex) {
+	if (state.program.stage == ShaderType::Vertex || state.program.stage == ShaderType::Local) {
 		return EmitVertexParameterComponentU32(state, *input, chan & 3u);
 	}
 	const auto load_per_vertex = [&](uint32_t vertex) {
@@ -517,11 +518,95 @@ uint32_t EmitIdentity(ValueEmitContext&, uint32_t value) {
 void EmitVoid(ValueEmitContext&) {}

 void EmitBarrier(EmitterState& state) {
-	const auto semantics =
-	    spv::MemorySemanticsAcquireReleaseMask | spv::MemorySemanticsWorkgroupMemoryMask;
+	const auto tessellation = state.program.stage == ShaderType::TessellationControl;
+	const auto memory_scope = tessellation ? spv::ScopeInvocation : spv::ScopeWorkgroup;
+	const auto semantics    = tessellation ? spv::MemorySemanticsMaskNone
+	                                       : spv::MemorySemanticsAcquireReleaseMask |
+	                                             spv::MemorySemanticsWorkgroupMemoryMask;
 	state.builder.AddFunction(spv::OpControlBarrier, ConstantU32(state, spv::ScopeWorkgroup),
-	                          ConstantU32(state, spv::ScopeWorkgroup),
-	                          ConstantU32(state, semantics));
+	                          ConstantU32(state, memory_scope), ConstantU32(state, semantics));
+}
+
+uint32_t EmitLaneId(EmitterState& state) {
+	return state.program.stage == ShaderType::TessellationControl
+	           ? EmitBuiltinU32(state, IR::StageInputKind::InvocationId, 0)
+	           : EmitSubgroupLocalInvocationId(state);
+}
+
+namespace {
+
+uint32_t TessellationPointer(ValueEmitContext& ctx, const IR::Inst& inst) {
+	auto& state          = ctx.state;
+	using Attribute      = IR::TessellationAttribute;
+	const auto  kind     = static_cast<Attribute>(inst.Arg(0).U32());
+	const auto& tess     = state.input_info.vertex->tess;
+	const auto  variable = state.tess_variables.at(static_cast<uint32_t>(kind));
+	EXIT_IF(variable == 0);
+	const auto input   = kind == Attribute::ControlInput || kind == Attribute::EvaluationInput;
+	const auto storage = input ? spv::StorageClassInput : spv::StorageClassOutput;
+	const auto pointer = state.builder.AllocateId();
+	if (kind == Attribute::Factor) {
+		EXIT_NOT_IMPLEMENTED(!inst.Arg(1).IsImmediate());
+		const auto index = inst.Arg(1).U32() / 4u;
+		const auto outer = index < 3u;
+		EXIT_NOT_IMPLEMENTED(index >= 4u);
+		state.builder.AddFunction(spv::OpAccessChain, TypePointer(state, storage, TypeF32(state)),
+		                          pointer, outer ? variable : state.tess_inner_variable,
+		                          ConstantU32(state, outer ? index : index - 3u));
+		return pointer;
+	}
+	auto address = ctx.Arg(inst, 1);
+	if (kind == Attribute::PatchOutput) {
+		address =
+		    EmitBinaryU32(state, spv::OpISub, address, ConstantU32(state, state.tess_patch_base));
+	}
+	const bool local  = kind == Attribute::LocalOutput || kind == Attribute::ControlInput;
+	const auto stride = local ? tess.ls_stride : tess.hs_stride;
+	const auto offset = kind == Attribute::PatchOutput ? address
+	                                                   : EmitBinaryU32(state, spv::OpUMod, address,
+	                                                                   ConstantU32(state, stride));
+	const auto attribute =
+	    EmitBinaryU32(state, spv::OpShiftRightLogical, offset, ConstantU32(state, 4));
+	const auto component =
+	    EmitBinaryU32(state, spv::OpBitwiseAnd,
+	                  EmitBinaryU32(state, spv::OpShiftRightLogical, offset, ConstantU32(state, 2)),
+	                  ConstantU32(state, 3));
+	const auto type = TypePointer(state, storage, TypeU32(state));
+	if (kind == Attribute::LocalOutput || kind == Attribute::PatchOutput) {
+		state.builder.AddFunction(spv::OpAccessChain, type, pointer, variable, attribute,
+		                          component);
+	} else {
+		const auto vertex =
+		    kind == Attribute::ControlOutput
+		        ? EmitBuiltinU32(state, IR::StageInputKind::InvocationId, 0)
+		        : EmitBinaryU32(state, spv::OpUDiv, address, ConstantU32(state, stride));
+		state.builder.AddFunction(spv::OpAccessChain, type, pointer, variable, vertex, attribute,
+		                          component);
+	}
+	return pointer;
+}
+
+} // namespace
+
+uint32_t EmitGetTessellationAttribute(ValueEmitContext& ctx, const IR::Inst& inst) {
+	return EmitValueOrZeroIfCondition(ctx.state, ctx.Arg(inst, 2), [&] {
+		const auto value = ctx.state.builder.AllocateId();
+		ctx.state.builder.AddFunction(spv::OpLoad, TypeU32(ctx.state), value,
+		                              TessellationPointer(ctx, inst));
+		return value;
+	});
+}
+
+void EmitSetTessellationAttribute(ValueEmitContext& ctx, const IR::Inst& inst) {
+	EmitIfCondition(ctx.state, ctx.Arg(inst, 3), [&] {
+		auto value = ctx.Arg(inst, 2);
+		if (inst.Arg(0).U32() == static_cast<uint32_t>(IR::TessellationAttribute::Factor)) {
+			const auto floating = ctx.state.builder.AllocateId();
+			ctx.state.builder.AddFunction(spv::OpBitcast, TypeF32(ctx.state), floating, value);
+			value = floating;
+		}
+		ctx.state.builder.AddFunction(spv::OpStore, TessellationPointer(ctx, inst), value);
+	});
 }

 uint32_t EmitMeshDrawParameter(ValueEmitContext& ctx, const IR::Inst& inst) {
@@ -155,6 +155,8 @@ inline constexpr auto EmitInstPrefetch = EmitVoid;
 void                  EmitBarrier(EmitterState& state);
 void                  EmitMeshAllocate(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitMeshDrawParameter(ValueEmitContext& ctx, const IR::Inst& inst);
+uint32_t              EmitGetTessellationAttribute(ValueEmitContext& ctx, const IR::Inst& inst);
+void                  EmitSetTessellationAttribute(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitGetUserData(EmitterState& state, IR::ScalarReg reg);
 uint32_t              EmitGetBuiltin(ValueEmitContext& ctx, IR::Value kind, IR::Value index);
 uint32_t              EmitUndefU1(EmitterState& state, const IR::Inst& inst);
@@ -165,7 +167,7 @@ inline constexpr auto EmitUndefU64 = EmitUndefU1;
 uint32_t              EmitDppMoveU32(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitDppUpdateU32(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitWqmU64(EmitterState& state, uint32_t value);
-inline constexpr auto EmitLaneId = EmitSubgroupLocalInvocationId;
+uint32_t              EmitLaneId(EmitterState& state);
 uint32_t              EmitBallot(ValueEmitContext& ctx, IR::Value predicate);
 uint32_t              EmitReadFirstLane(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitReadLane(ValueEmitContext& ctx, const IR::Inst& inst);
@@ -271,6 +273,7 @@ inline constexpr auto EmitImageAtomicOr32      = EmitImage;
 inline constexpr auto EmitImageAtomicXor32     = EmitImage;
 void                  EmitUnreachable(ValueEmitContext& ctx, const IR::Inst& inst);
 inline constexpr auto EmitPhi                        = EmitUnreachable;
+inline constexpr auto EmitTessellationBase           = EmitUnreachable;
 inline constexpr auto EmitGetThreadBitScalarRegister = EmitUnreachable;
 inline constexpr auto EmitSetThreadBitScalarRegister = EmitUnreachable;
 inline constexpr auto EmitGetScalarMaskTag           = EmitUnreachable;
@@ -80,6 +80,10 @@ struct EmitterState {
 	Builder                                          builder;
 	const IR::Program&                               program;
 	ShaderStageInputInfo                             input_info;
+	std::array<uint32_t, 6>                          tess_variables {};
+	uint32_t                                         tess_inner_variable = 0;
+	uint32_t                                         tess_patch_base     = 0;
+
 	const SpirvRequirements                          requirements;
 	uint32_t                                         lane_count              = 1;
 	uint32_t                                         lane_half               = 0;
@@ -290,7 +290,8 @@ uint32_t GlslStd450(EmitterState& state) {
 }

 VertexInputScalarKind VertexParameterScalarKind(const EmitterState& state, uint32_t location) {
-	if (state.program.stage != ShaderType::Vertex || location >= ShaderVertexInputInfo::RES_MAX ||
+	if ((state.program.stage != ShaderType::Vertex && state.program.stage != ShaderType::Local) ||
+	    location >= ShaderVertexInputInfo::RES_MAX ||
 	    location >= static_cast<uint32_t>(state.input_info.vertex->resources_num)) {
 		return VertexInputScalarKind::Float;
 	}
@@ -347,6 +348,9 @@ uint32_t DefineInterfaceVariable(EmitterState& state, uint32_t type, spv::Storag
 uint32_t BuiltInForInput(IR::StageInputKind kind) {
 	switch (kind) {
 		case IR::StageInputKind::VertexIndex: return spv::BuiltInVertexIndex;
+		case IR::StageInputKind::InvocationId: return spv::BuiltInInvocationId;
+		case IR::StageInputKind::PrimitiveId: return spv::BuiltInPrimitiveId;
+		case IR::StageInputKind::TessCoord: return spv::BuiltInTessCoord;
 		case IR::StageInputKind::InstanceIndex: return spv::BuiltInInstanceIndex;
 		case IR::StageInputKind::FragCoord: return spv::BuiltInFragCoord;
 		case IR::StageInputKind::FrontFacing: return spv::BuiltInFrontFacing;
@@ -387,6 +391,8 @@ void DefineInputs(EmitterState& state) {
 		uint32_t type = TypeU32(state);
 		switch (input.kind) {
 			case IR::StageInputKind::VertexIndex:
+			case IR::StageInputKind::InvocationId:
+			case IR::StageInputKind::PrimitiveId:
 			case IR::StageInputKind::InstanceIndex:
 			case IR::StageInputKind::Layer:
 			case IR::StageInputKind::SampleId: type = TypeI32(state); break;
@@ -394,11 +400,13 @@ void DefineInputs(EmitterState& state) {
 			case IR::StageInputKind::LocalInvocationId:
 			case IR::StageInputKind::GlobalInvocationId: type = TypeU32Vector(state, 3); break;
 			case IR::StageInputKind::FragCoord: type = TypeF32Vector(state, 4); break;
+			case IR::StageInputKind::TessCoord:
 			case IR::StageInputKind::BaryCoordSmooth:
 			case IR::StageInputKind::BaryCoordNoPerspective: type = TypeF32Vector(state, 3); break;
 			case IR::StageInputKind::FrontFacing: type = TypeBool(state); break;
 			case IR::StageInputKind::Parameter:
-				if (state.program.stage == ShaderType::Vertex) {
+				if (state.program.stage == ShaderType::Vertex ||
+				    state.program.stage == ShaderType::Local) {
 					type = VertexParameterScalarType(
 					    state, VertexParameterScalarKind(state, input.location));
 					const auto components = VertexParameterComponentCount(input);
@@ -536,6 +544,70 @@ void DefineOutputs(EmitterState& state) {
 	}
 }

+void DefineTessellationInterfaces(EmitterState& state) {
+	using Attribute = IR::TessellationAttribute;
+	std::array<bool, 6> used {};
+	uint32_t            patch_begin = UINT32_MAX, patch_end = 0;
+	for (const auto* block: state.program.blocks) {
+		for (const auto& inst: *block) {
+			if (inst.GetOpcode() != IR::ValueOpcode::GetTessellationAttribute &&
+			    inst.GetOpcode() != IR::ValueOpcode::SetTessellationAttribute)
+				continue;
+			const auto kind = inst.Arg(0).U32();
+			used.at(kind)   = true;
+			if (kind == static_cast<uint32_t>(Attribute::PatchOutput)) {
+				EXIT_NOT_IMPLEMENTED(!inst.Arg(1).IsImmediate());
+				patch_begin = std::min(patch_begin, inst.Arg(1).U32());
+				patch_end   = std::max(patch_end, inst.Arg(1).U32() + 4u);
+			}
+		}
+	}
+	if (std::ranges::none_of(used, [](bool value) { return value; })) return;
+	const auto& tess  = state.input_info.vertex->tess;
+	const auto  array = [&](uint32_t type, uint32_t count) {
+		return state.builder.Type(spv::OpTypeArray, type, ConstantU32(state, count));
+	};
+	for (uint32_t index = 0; index < used.size(); index++) {
+		if (!used[index]) continue;
+		const auto kind    = static_cast<Attribute>(index);
+		const bool input   = kind == Attribute::ControlInput || kind == Attribute::EvaluationInput;
+		const auto storage = input ? spv::StorageClassInput : spv::StorageClassOutput;
+		uint32_t   type;
+		if (kind == Attribute::Factor) {
+			type = array(TypeF32(state), 4u);
+		} else if (kind == Attribute::PatchOutput) {
+			state.tess_patch_base = patch_begin;
+			type = array(TypeU32Vector(state, 4), (patch_end - patch_begin + 15u) / 16u);
+		} else {
+			const bool local  = kind == Attribute::LocalOutput || kind == Attribute::ControlInput;
+			const auto stride = local ? tess.ls_stride : tess.hs_stride;
+			type              = array(TypeU32Vector(state, 4), (stride + 15u) / 16u);
+			if (kind != Attribute::LocalOutput) {
+				type = array(type, local ? tess.input_control_points : tess.output_control_points);
+			}
+		}
+		auto& variable = state.tess_variables[index];
+		variable       = DefineInterfaceVariable(state, type, storage, "tess_attributes");
+		if (kind == Attribute::Factor) {
+			state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationBuiltIn,
+			                            spv::BuiltInTessLevelOuter);
+			state.tess_inner_variable =
+			    DefineInterfaceVariable(state, array(TypeF32(state), 2u), storage, "tess_inner");
+			state.builder.AddAnnotation(spv::OpDecorate, state.tess_inner_variable,
+			                            spv::DecorationBuiltIn, spv::BuiltInTessLevelInner);
+			state.builder.AddAnnotation(spv::OpDecorate, state.tess_inner_variable,
+			                            spv::DecorationPatch);
+		} else {
+			state.builder.AddAnnotation(
+			    spv::OpDecorate, variable, spv::DecorationLocation,
+			    kind == Attribute::PatchOutput ? (tess.hs_stride + 15u) / 16u : 0u);
+		}
+		if (kind == Attribute::Factor || kind == Attribute::PatchOutput) {
+			state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationPatch);
+		}
+	}
+}
+
 } // namespace

 void DefineModule(EmitterState& state) {
@@ -543,6 +615,7 @@ void DefineModule(EmitterState& state) {
 	                                  state.program.info.outputs.size());
 	DefineInputs(state);
 	DefineOutputs(state);
+	DefineTessellationInterfaces(state);
 	DefineDescriptors(state);
 	if (state.requirements.function_lds) {
 		state.lds_variable = state.builder.AllocateId();
@@ -564,6 +637,21 @@ void DefineModule(EmitterState& state) {
 		state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeOutputPrimitivesEXT,
 		                               state.input_info.vertex->mesh.max_primitives);
 	}
+	if (state.program.stage == ShaderType::TessellationControl ||
+	    state.program.stage == ShaderType::TessellationEvaluation) {
+		const auto& tess = state.input_info.vertex->tess;
+		EXIT_NOT_IMPLEMENTED(tess.domain != 1u || tess.partitioning != 2u ||
+		                     tess.output_topology != 2u);
+		state.builder.RequireCapability(spv::CapabilityTessellation);
+		if (state.program.stage == ShaderType::TessellationControl) {
+			state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeOutputVertices,
+			                               tess.output_control_points);
+		} else {
+			state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeTriangles);
+			state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeSpacingFractionalOdd);
+			state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeVertexOrderCw);
+		}
+	}
 	state.entry_label = state.builder.AllocateId();

 	state.builder.RequireCapability(spv::CapabilityShader);
@@ -898,11 +898,15 @@ void ValidateTranslateOptions(const TranslateOptions& options) {
 	if (options.wave_size != 32u && options.wave_size != 64u) {
 		EXIT("shader translation requires wave32 or wave64, got %u", options.wave_size);
 	}
-	if (options.embedded_fetch != nullptr && options.stage != ShaderType::Vertex) {
-		EXIT("embedded fetch requires the vertex shader stage");
+	if (options.embedded_fetch != nullptr && options.stage != ShaderType::Vertex &&
+	    options.stage != ShaderType::Local) {
+		EXIT("embedded fetch requires a vertex or local shader");
 	}
 	switch (options.stage) {
 		case ShaderType::Vertex:
+		case ShaderType::Local:
+		case ShaderType::TessellationControl:
+		case ShaderType::TessellationEvaluation:
 		case ShaderType::Mesh:
 			if (options.input_info.vertex == nullptr) {
 				EXIT("vertex shader translation has no vertex input metadata");
@@ -941,6 +945,9 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 	result.user_data_count     = options.user_data_count;
 	switch (options.stage) {
 		case ShaderType::Vertex:
+		case ShaderType::Local:
+		case ShaderType::TessellationControl:
+		case ShaderType::TessellationEvaluation:
 			result.scratch_dwords = options.input_info.vertex->scratch_size_dwords;
 			break;
 		case ShaderType::Mesh:
@@ -1164,6 +1171,42 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 			entry_ir.SetVectorReg(
 			    static_cast<IR::VectorReg>(8),
 			    entry_ir.IAdd(draw(2), builtin(IR::StageInputKind::WorkgroupId, 1)));
+		} else if (options.stage == ShaderType::Local) {
+			entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(3), IR::U32(IR::Value(64u)));
+			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(2),
+			                      builtin(IR::StageInputKind::VertexIndex));
+			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(3), IR::U32(IR::Value(0u)));
+			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(5),
+			                      builtin(IR::StageInputKind::InstanceIndex));
+		} else if (options.stage == ShaderType::TessellationControl) {
+			const auto& tess = options.input_info.vertex->tess;
+			entry_ir.SetScalarReg(
+			    static_cast<IR::ScalarReg>(2),
+			    IR::U32(entry_ir.Emit(IR::ValueOpcode::TessellationBase, {IR::Value(0u)})));
+			entry_ir.SetScalarReg(
+			    static_cast<IR::ScalarReg>(4),
+			    IR::U32(entry_ir.Emit(IR::ValueOpcode::TessellationBase, {IR::Value(1u)})));
+			entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(3),
+			                      IR::U32(IR::Value(0x81010000u | tess.input_control_points |
+			                                        (tess.output_control_points << 8u))));
+			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(0),
+			                      builtin(IR::StageInputKind::PrimitiveId));
+			entry_ir.SetVectorReg(
+			    static_cast<IR::VectorReg>(1),
+			    entry_ir.ShiftLeftLogical(builtin(IR::StageInputKind::InvocationId),
+			                              IR::U32(IR::Value(8u))));
+		} else if (options.stage == ShaderType::TessellationEvaluation) {
+			entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(3), IR::U32(IR::Value(64u)));
+			entry_ir.SetScalarReg(
+			    static_cast<IR::ScalarReg>(4),
+			    IR::U32(entry_ir.Emit(IR::ValueOpcode::TessellationBase, {IR::Value(0u)})));
+			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(5),
+			                      builtin(IR::StageInputKind::TessCoord, 0));
+			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(6),
+			                      builtin(IR::StageInputKind::TessCoord, 1));
+			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(7), IR::U32(IR::Value(0u)));
+			entry_ir.SetVectorReg(static_cast<IR::VectorReg>(8),
+			                      builtin(IR::StageInputKind::PrimitiveId));
 		} else if (options.stage == ShaderType::Pixel) {
 			const auto* ps = options.input_info.pixel;
 			if (ps->ps_perspective_center_vgpr != UINT32_MAX) {
@@ -150,8 +150,20 @@ struct SampledResourcePair {
 	bool operator==(const SampledResourcePair& other) const = default;
 };

+enum class TessellationAttribute {
+	LocalOutput,
+	ControlInput,
+	ControlOutput,
+	EvaluationInput,
+	PatchOutput,
+	Factor
+};
+
 enum class StageInputKind {
 	VertexIndex,
+	InvocationId,
+	PrimitiveId,
+	TessCoord,
 	InstanceIndex,
 	FragCoord,
 	FrontFacing,
@@ -287,8 +299,12 @@ static_assert(sizeof(PushData) == 128);
 constexpr uint32_t NativePushConstantSize = sizeof(PushData);

 [[nodiscard]] constexpr uint32_t NativeBinding(ShaderType stage, DescriptorBindingKind kind) {
+	const uint32_t group = stage == ShaderType::Pixel                    ? 1u
+	                       : stage == ShaderType::TessellationControl    ? 2u
+	                       : stage == ShaderType::TessellationEvaluation ? 3u
+	                                                                     : 0u;
 	return static_cast<uint32_t>(kind) +
-	       (stage == ShaderType::Pixel ? static_cast<uint32_t>(DescriptorBindingKind::Count) : 0u);
+	       group * static_cast<uint32_t>(DescriptorBindingKind::Count);
 }

 [[nodiscard]] constexpr ImageResourceClass ImageBindingResourceClass(DescriptorBindingKind kind) {
@@ -90,6 +90,7 @@ bool HasSideEffects(ValueOpcode opcode) {
 		case ValueOpcode::Reference:
 		case ValueOpcode::ReferenceU32:
 		case ValueOpcode::SetAttribute:
+		case ValueOpcode::SetTessellationAttribute:
 		case ValueOpcode::MeshAllocate:
 		case ValueOpcode::Barrier: return true;
 		default: return false;
@@ -10,6 +10,9 @@ VALUE_OPCODE(GetUserData, U32, ScalarReg)
 VALUE_OPCODE(GetShaderBase, U64)
 VALUE_OPCODE(MeshDrawParameter, U32, U32)
 VALUE_OPCODE(MeshAllocate, Void, U32)
+VALUE_OPCODE(TessellationBase, U32, U32)
+VALUE_OPCODE(GetTessellationAttribute, U32, U32, U32, U1)
+VALUE_OPCODE(SetTessellationAttribute, Void, U32, U32, U32, U1)
 VALUE_OPCODE(GetBuiltin, U32, U32, U32)
 VALUE_OPCODE(GetThreadBitScalarRegister, U1, ScalarReg)
 VALUE_OPCODE(SetThreadBitScalarRegister, Void, ScalarReg, U1)
@@ -228,6 +228,15 @@ void FoldInstruction(Block& block, Block::iterator instruction,
 			const auto offset = Arg(inst, 1);
 			const auto count  = Arg(inst, 2);
 			auto* source = value.TryInstruction();
+			if (source != nullptr && source->GetOpcode() == ValueOpcode::ShiftLeftLogical32 &&
+			    IsImmediate(offset, Type::U32) && IsImmediate(count, Type::U32)) {
+				const auto shift = Arg(*source, 1);
+				if (IsImmediate(shift, Type::U32) && shift.U32() < 32u &&
+				    offset.U32() <= shift.U32() && count.U32() <= shift.U32() - offset.U32()) {
+					Replace(inst, Value(0u));
+					return;
+				}
+			}
 			if (source != nullptr && source->GetOpcode() == ValueOpcode::GetBuiltin &&
 			    source->Arg(0) == Value(static_cast<uint32_t>(StageInputKind::PackedAncillary)) &&
 			    IsImmediate(offset, Type::U32) && IsImmediate(count, Type::U32) && count.U32() != 0u) {
@@ -43,6 +43,9 @@ void AddOutput(ShaderInfo& info, StageOutputKind kind, uint32_t index, uint32_t
 void ValidateOptions(const Program& program, ShaderStageInputInfo input_info) {
 	switch (program.stage) {
 		case ShaderType::Vertex:
+		case ShaderType::Local:
+		case ShaderType::TessellationControl:
+		case ShaderType::TessellationEvaluation:
 		case ShaderType::Mesh:
 			if (input_info.vertex == nullptr) {
 				return Fail("vertex shader has no input metadata");
@@ -81,7 +84,8 @@ void ValidateValueReferences(const Program& program, ShaderStageInputInfo input_
 					    !inst.Arg(1).IsImmediate() || inst.Arg(1).GetType() != Type::U32) {
 						return Fail("typed attribute reference is not constant");
 					}
-					if (program.stage == ShaderType::Vertex &&
+					if ((program.stage == ShaderType::Vertex ||
+					     program.stage == ShaderType::Local) &&
 					    (inst.Arg(1).U32() >= 4u ||
 					     inst.Arg(0).U32() >=
 					         static_cast<uint32_t>(input_info.vertex->resources_num))) {
@@ -124,6 +128,8 @@ void ValidateValueReferences(const Program& program, ShaderStageInputInfo input_
 							}
 							break;
 						case StageInputKind::VertexIndex:
+						case StageInputKind::InvocationId:
+						case StageInputKind::PrimitiveId:
 						case StageInputKind::InstanceIndex:
 						case StageInputKind::FrontFacing:
 						case StageInputKind::LocalInvocationIndex:
@@ -142,6 +148,7 @@ void ValidateValueReferences(const Program& program, ShaderStageInputInfo input_
 								return Fail("typed barycentric component is out of range");
 							}
 							break;
+						case StageInputKind::TessCoord:
 						case StageInputKind::WorkgroupId:
 						case StageInputKind::LocalInvocationId:
 						case StageInputKind::GlobalInvocationId:
@@ -161,8 +168,10 @@ void ValidateValueReferences(const Program& program, ShaderStageInputInfo input_
 					}
 					const auto& exp = program.export_info[index];
 					if (exp.kind == ExportTargetKind::Position && exp.index != 0 && exp.en != 0 &&
-					    program.stage != ShaderType::Vertex && program.stage != ShaderType::Mesh) {
-						return Fail("auxiliary position export requires the vertex or mesh stage");
+					    program.stage != ShaderType::Vertex && program.stage != ShaderType::Mesh &&
+					    program.stage != ShaderType::TessellationEvaluation) {
+						return Fail("auxiliary position export requires a vertex, mesh, or "
+						            "tessellation evaluation shader");
 					}
 					break;
 				}
@@ -249,6 +258,10 @@ void CollectComputeInputs(const ShaderComputeInputInfo* compute, ShaderInfo& inf
 void CollectBuiltinInputs(const Program& program, ShaderInfo& info) {
 	for (const auto* block: program.blocks) {
 		for (const auto& inst: *block) {
+			if (program.stage == ShaderType::TessellationControl &&
+			    inst.GetOpcode() == ValueOpcode::LaneId) {
+				AddInput(info, StageInputKind::InvocationId, 0, 1, "gl_InvocationID");
+			}
 			if (inst.GetOpcode() != ValueOpcode::GetBuiltin) {
 				continue;
 			}
@@ -260,6 +273,13 @@ void CollectBuiltinInputs(const Program& program, ShaderInfo& info) {
 				case StageInputKind::InstanceIndex:
 					AddInput(info, kind, 0, 1, "gl_InstanceIndex");
 					break;
+				case StageInputKind::InvocationId:
+					AddInput(info, kind, 0, 1, "gl_InvocationID");
+					break;
+				case StageInputKind::PrimitiveId:
+					AddInput(info, kind, 0, 1, "gl_PrimitiveID");
+					break;
+				case StageInputKind::TessCoord: AddInput(info, kind, 0, 3, "gl_TessCoord"); break;
 				case StageInputKind::FragCoord: AddInput(info, kind, 0, 4, "gl_FragCoord"); break;
 				case StageInputKind::FrontFacing:
 					AddInput(info, kind, 0, 1, "gl_FrontFacing");
@@ -380,7 +400,10 @@ void CollectShaderInfo(Program& program, ShaderStageInputInfo input_info) {
 		    });
 	    });
 	switch (program.stage) {
-		case ShaderType::Vertex: CollectVertexInputs(program, input_info.vertex, next); break;
+		case ShaderType::Vertex:
+		case ShaderType::Local: CollectVertexInputs(program, input_info.vertex, next); break;
+		case ShaderType::TessellationControl:
+		case ShaderType::TessellationEvaluation:
 		case ShaderType::Mesh: break;
 		case ShaderType::Pixel: CollectPixelInputs(program, input_info.pixel, next); break;
 		case ShaderType::Compute: CollectComputeInputs(input_info.compute, next); break;
@@ -11,6 +11,7 @@
 #include "graphics/guest_gpu/graphicsRun.h"
 #include "graphics/guest_gpu/hardwareContext.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
+#include "graphics/shader/recompiler/ShaderRecompiler.h"
 #include "graphics/shader/recompiler/frontend/decode/ShaderDecoder.h"
 #include "graphics/shader/shaderCompiler.h"
 #include "graphics/shader/shaderVertexMetadata.h"
@@ -511,33 +512,28 @@ static uint32_t ShaderCalcPsSystemInputBase(const HW::ShaderRegisters& regs) {
 	return reg;
 }

-static bool ShaderGetStaticInputInfoVS(const HW::VertexShaderInfo& regs,
-	                                   const HW::ShaderRegisters& sh,
-	                                   const ShaderMappedData& data,
-	                                   ShaderVertexInputInfo& info) {
+static bool ShaderGetStaticVertexInputInfo(uint64_t shader_addr, const HW::UserSgprInfo& user_sgpr,
+                                           uint32_t user_sgpr_num, const HW::ShaderRegisters& sh,
+                                           const ShaderMappedData& data,
+                                           ShaderVertexInputInfo&  info) {
 	KYTY_PROFILER_FUNCTION();

 	info = {};

 	info.pa_cl_vs_out_cntl = sh.m_paClVsOutCntl;

-	EXIT_NOT_IMPLEMENTED(regs.es_regs.data_addr == 0);
-
-	uint64_t                shader_addr   = regs.es_regs.data_addr;
-	const HW::UserSgprInfo& user_sgpr     = regs.gs_user_sgpr;
-	auto                    user_sgpr_num = regs.gs_regs.rsrc2.user_sgpr;
+	EXIT_NOT_IMPLEMENTED(shader_addr == 0);
 	info.scratch_size_dwords = data.scratch_size_dwords;

 	if (data.user_data == nullptr) {
-		LOGF("ShaderGetInputInfoVS(): no AGC user data for shader=0x%016" PRIx64 " es=0x%016" PRIx64
-		     " gs=0x%016" PRIx64 " user_sgpr_num=%u\n",
-		     shader_addr, regs.es_regs.data_addr, regs.gs_regs.data_addr,
-		     static_cast<uint32_t>(user_sgpr_num));
+		LOGF("ShaderGetVertexInputInfo(): no AGC user data for shader=0x%016" PRIx64
+		     " user_sgpr_num=%u\n",
+		     shader_addr, user_sgpr_num);
 	}
 	ShaderVertexMetadata metadata;
 	std::string          metadata_error;
 	if (!ShaderReadVertexMetadata(data, HW::UserSgprInfo::SGPRS_MAX, metadata, &metadata_error)) {
-		LOGF("ShaderGetInputInfoVS(): invalid AGC metadata shader=0x%016" PRIx64 ": %s\n",
+		LOGF("ShaderGetVertexInputInfo(): invalid AGC metadata shader=0x%016" PRIx64 ": %s\n",
 		     shader_addr, metadata_error.c_str());
 		return false;
 	}
@@ -556,7 +552,7 @@ static bool ShaderGetStaticInputInfoVS(const HW::VertexShaderInfo& regs,
 		    (static_cast<uint64_t>(user_sgpr.value[metadata.vertex_buffer_reg + 1]) << 32u));

 		if (attrib == nullptr || buffer == nullptr) {
-			LOGF("ShaderGetInputInfoVS(): null vertex table pointer shader=0x%016" PRIx64 "\n",
+			LOGF("ShaderGetVertexInputInfo(): null vertex table pointer shader=0x%016" PRIx64 "\n",
 			     shader_addr);
 			return false;
 		}
@@ -673,6 +669,12 @@ void BuildStageStaticKey(const ShaderVertexInputInfo& info, std::vector<uint32_t
 		                       mesh.primitives_per_group, mesh.vertices_per_group,
 		                       mesh.max_vertices, mesh.max_primitives, mesh.provoking_vertex});
 	}
+	key.push_back(info.tess.input_control_points);
+	if (info.tess.input_control_points != 0) {
+		const auto& tess = info.tess;
+		key.insert(key.end(), {tess.output_control_points, tess.ls_stride, tess.hs_stride,
+		                       tess.domain, tess.partitioning, tess.output_topology});
+	}

 	for (int i = 0; i < info.resources_num; i++) {
 		const auto& resource    = info.resources[i];
@@ -750,7 +752,8 @@ ShaderParams PrepareProgram(const HW::VertexShaderInfo& regs, const HW::Context&
 	    GetDeclaredShaderHash(regs.es_regs.data_addr),
 	    std::span<const uint32_t>(regs.gs_user_sgpr.value, regs.gs_regs.rsrc2.user_sgpr), data);
 	if ((context.GetShaderStages() & 0x20u) == 0) {
-		if (!ShaderGetStaticInputInfoVS(regs, sh, data, info)) {
+		if (!ShaderGetStaticVertexInputInfo(regs.es_regs.data_addr, regs.gs_user_sgpr,
+		                                    regs.gs_regs.rsrc2.user_sgpr, sh, data, info)) {
 			EXIT("failed to prepare vertex shader program\n");
 		}
 		return params;
@@ -759,6 +762,7 @@ ShaderParams PrepareProgram(const HW::VertexShaderInfo& regs, const HW::Context&
 	// its user-data pointer in s0:s1.
 	params.user_data.insert(params.user_data.begin(), 8u, 0u);
 	info                     = {};
+	info.logical_stage       = ShaderType::Mesh;
 	info.pa_cl_vs_out_cntl   = sh.m_paClVsOutCntl;
 	auto& mesh               = info.mesh;
 	mesh.input_primitive     = static_cast<uint32_t>(user_config.GetPrimType());
@@ -806,6 +810,65 @@ ShaderParams PrepareProgram(const HW::VertexShaderInfo& regs, const HW::Context&
 	return params;
 }

+std::array<ShaderParams, 3>
+PrepareTessellationPrograms(const HW::VertexShaderInfo& regs, const HW::Context& context,
+                            std::array<ShaderVertexInputInfo, 3>& input_info) {
+	const auto& sh        = context.GetShaderRegisters();
+	const auto  local     = ShaderGetMappedData(regs.ls_regs.data_addr, "ShaderGetInputInfoLS():");
+	const auto  control   = ShaderGetMappedData(regs.hs_regs.data_addr, "ShaderGetInputInfoHS():");
+	const auto evaluation = ShaderGetMappedData(regs.es_regs.data_addr, "ShaderGetInputInfoTES():");
+	EXIT_NOT_IMPLEMENTED(local.type != Prospero::ShaderBinaryType::kHsFront ||
+	                     control.type != Prospero::ShaderBinaryType::kHsBack ||
+	                     evaluation.type != Prospero::ShaderBinaryType::kGs);
+	EXIT_NOT_IMPLEMENTED((context.GetShaderStages() & 0x4u) == 0 ||
+	                     (context.GetShaderStages() & 0x00600020u) != 0);
+	EXIT_IF(regs.hs_regs.user_data_addr == 0);
+
+	const auto local_users      = std::span(regs.hs_user_sgpr.value, regs.hs_regs.rsrc2.user_sgpr);
+	const auto evaluation_users = std::span(regs.gs_user_sgpr.value, regs.gs_regs.rsrc2.user_sgpr);
+	std::array<ShaderParams, 3> params {
+	    GetShaderParams(regs.ls_regs.data_addr, "ShaderRecompiler LS",
+	                    GetDeclaredShaderHash(regs.ls_regs.data_addr), local_users, local),
+	    GetShaderParams(regs.hs_regs.data_addr, "ShaderRecompiler HS",
+	                    GetDeclaredShaderHash(regs.hs_regs.data_addr), local_users, control),
+	    GetShaderParams(regs.es_regs.data_addr, "ShaderRecompiler TES",
+	                    GetDeclaredShaderHash(regs.es_regs.data_addr), evaluation_users,
+	                    evaluation),
+	};
+	// The fused HS back half receives its separate user-data address in s0:s1.
+	// RDNA2 reserves s0:s7 before the native HS user SGPRs.
+	params[1].user_data.insert(params[1].user_data.begin(), 8u, 0u);
+	params[1].user_data[0] = static_cast<uint32_t>(regs.hs_regs.user_data_addr);
+	params[1].user_data[1] = static_cast<uint32_t>(regs.hs_regs.user_data_addr >> 32u);
+
+	input_info = {};
+	if (!ShaderGetStaticVertexInputInfo(regs.ls_regs.data_addr, regs.hs_user_sgpr,
+	                                    regs.hs_regs.rsrc2.user_sgpr, sh, local, input_info[0])) {
+		EXIT("failed to prepare local shader program\n");
+	}
+	input_info[0].logical_stage       = ShaderType::Local;
+	input_info[1].logical_stage       = ShaderType::TessellationControl;
+	input_info[1].scratch_size_dwords = control.scratch_size_dwords;
+	input_info[2].logical_stage       = ShaderType::TessellationEvaluation;
+	input_info[2].scratch_size_dwords = evaluation.scratch_size_dwords;
+	input_info[2].pa_cl_vs_out_cntl   = sh.m_paClVsOutCntl;
+
+	ShaderTessellationInputInfo tess {
+	    .input_control_points  = (sh.m_vgtLsHsConfig >> 8u) & 0x3fu,
+	    .output_control_points = (sh.m_vgtLsHsConfig >> 14u) & 0x3fu,
+	    .domain                = sh.m_vgtTfParam & 0x3u,
+	    .partitioning          = (sh.m_vgtTfParam >> 2u) & 0x3u,
+	    .output_topology       = (sh.m_vgtTfParam >> 5u) & 0x3u,
+	};
+	EXIT_IF(tess.input_control_points == 0 || tess.input_control_points > 32 ||
+	        tess.output_control_points == 0 || tess.output_control_points > 32);
+	ShaderRecompiler::AnalyzeTessellationPrograms(params[0].code, params[1].code, tess);
+	for (auto& stage: input_info) {
+		stage.tess = tess;
+	}
+	return params;
+}
+
 ShaderParams PrepareProgram(
    const HW::PixelShaderInfo& regs, const HW::ShaderRegisters& sh,
    std::span<const Prospero::ColorComponentMapping, 8> target_export_mapping,
@@ -33,7 +33,17 @@ struct ComputeShaderInfo;
 struct ShaderRegisters;
 } // namespace HW

-enum class ShaderType { Unknown, Vertex, Pixel, Fetch, Compute, Mesh };
+enum class ShaderType {
+	Unknown,
+	Vertex,
+	Pixel,
+	Fetch,
+	Compute,
+	Mesh,
+	Local,
+	TessellationControl,
+	TessellationEvaluation
+};

 namespace ShaderRecompiler::IR {
 struct CompiledShaderInfo;
@@ -99,6 +109,16 @@ struct ShaderMeshInputInfo: ShaderWorkgroupInputInfo {
 	}
 };

+struct ShaderTessellationInputInfo {
+	uint32_t input_control_points  = 0;
+	uint32_t output_control_points = 0;
+	uint32_t ls_stride             = 0;
+	uint32_t hs_stride             = 0;
+	uint32_t domain                = 0;
+	uint32_t partitioning          = 0;
+	uint32_t output_topology       = 0;
+};
+
 struct ShaderVertexInputInfo {
 	static constexpr int RES_MAX = 32;

@@ -106,6 +126,7 @@ struct ShaderVertexInputInfo {
 	ShaderVertexDestination resources_dst[RES_MAX];
 	ShaderVertexInputBuffer buffers[RES_MAX];
 	ShaderStageRuntime      stage;
+	ShaderType                  logical_stage        = ShaderType::Vertex;
 	int                     resources_num       = 0;
 	int                     fetch_attrib_reg    = 0;
 	int                     fetch_buffer_reg    = 0;
@@ -114,6 +135,7 @@ struct ShaderVertexInputInfo {
 	uint32_t                pa_cl_vs_out_cntl    = 0;
 	ShaderClipSpaceTransform clip_space;
 	ShaderMeshInputInfo      mesh;
+	ShaderTessellationInputInfo tess;
 	bool                    fetch_external      = false;
 	bool                    fetch_embedded      = false;
 };
@@ -30,6 +30,9 @@ void BuildStageStaticKey(const ShaderComputeInputInfo& input_info, std::vector<u

 ShaderParams PrepareProgram(const HW::VertexShaderInfo& regs, const HW::Context& context,
                            const HW::UserConfig& user_config, ShaderVertexInputInfo& input_info);
+std::array<ShaderParams, 3>
+PrepareTessellationPrograms(const HW::VertexShaderInfo& regs, const HW::Context& context,
+                            std::array<ShaderVertexInputInfo, 3>& input_info);
 ShaderParams PrepareProgram(
    const HW::PixelShaderInfo& regs, const HW::ShaderRegisters& sh,
    std::span<const Prospero::ColorComponentMapping, 8> target_export_mapping,
@@ -401,7 +401,16 @@ struct RenderExecutorTestAccess {
                                      const ShaderStageRuntime &vertex,
                                      const ShaderStageRuntime &pixel,
                                      bool pixel_active) {
-    return executor.PrepareGraphicsBindings(vertex, pixel, pixel_active);
+    RenderExecutor::GraphicsBindings result;
+    result.vertex[0] = executor.PrepareBindings(vertex);
+    std::array<PreparedBindings *, 2> stages{&result.vertex[0], nullptr};
+    if (pixel_active) {
+      result.pixel.emplace(executor.PrepareBindings(pixel));
+      stages[1] = &*result.pixel;
+    }
+    executor.PrepareGraphicsBindings(
+        std::span{stages.data(), pixel_active ? 2u : 1u});
+    return result;
  }

  static PipelineCache::Pipeline
@@ -9741,7 +9750,7 @@ public:
          ordered_bindings.pixel->images[0].image_id;
      Require(
          name, "VS-before-PS retained-owner order",
-          ordered_bindings.vertex.images[0].image_id == storage_id &&
+          ordered_bindings.vertex[0].images[0].image_id == storage_id &&
              ordered_sampled_id != storage_id &&
              RenderExecutorTestAccess::BoundImagesInOrder(executor, storage_id,
                                                           ordered_sampled_id),
@@ -9753,7 +9762,7 @@ public:
          RenderExecutorTestAccess::PrepareGraphicsBindings(
              executor, storage_runtime, sampled_runtime, true);
      const auto &storage_binding =
-          graphics_bindings.vertex.images[0];
+          graphics_bindings.vertex[0].images[0];
      const auto &sampled_binding =
          graphics_bindings.pixel->images[0];
      Require(name, "storage final acquisition",
@@ -9775,7 +9784,7 @@ public:
              "the production graphics binding path did not complete vertex "
              "storage acquisition before pixel sampling");
      descriptor_pipelines.push_back(RenderExecutorTestAccess::CommitBindings(
-          executor, scheduler.Current(), graphics_bindings.vertex,
+          executor, scheduler.Current(), graphics_bindings.vertex[0],
          *graphics_bindings.pixel));
      Require(name, "early writable alias retention",
              texture_cache.GetImage(storage_id).backing.state.access_mask ==
@@ -9815,15 +9824,15 @@ public:
              "an already sampled image was not promoted when a later "
              "storage alias bound the same backing");
      descriptor_pipelines.push_back(RenderExecutorTestAccess::CommitBindings(
-          executor, scheduler.Current(), writable_alias_bindings.vertex,
+          executor, scheduler.Current(), writable_alias_bindings.vertex[0],
          *writable_alias_bindings.pixel));
      Require(
          name, "forced-general descriptor capture",
-          writable_alias_bindings.vertex.images[0].layout ==
+          writable_alias_bindings.vertex[0].images[0].layout ==
                  vk::ImageLayout::eGeneral &&
              writable_alias_bindings.pixel->images[0].layout ==
                  vk::ImageLayout::eGeneral &&
-              MakeImageInfo(writable_alias_bindings.vertex.images[0])
+              MakeImageInfo(writable_alias_bindings.vertex[0].images[0])
                      .imageLayout == vk::ImageLayout::eGeneral &&
              MakeImageInfo(writable_alias_bindings.pixel->images[0])
                      .imageLayout == vk::ImageLayout::eGeneral &&
@@ -10240,8 +10249,8 @@ public:
        const auto bounds_rendering = RenderExecutorTestAccess::AcquireRenderTargets(
            executor, scheduler.Current(), &no_color, 0, bounds_depth, bounds_bindings.pixel);
        descriptor_pipelines.push_back(RenderExecutorTestAccess::CommitBindings(
-            executor, scheduler.Current(), bounds_bindings.vertex, *bounds_bindings.pixel));
-        const auto &vertex_depth = bounds_bindings.vertex.images[0];
+            executor, scheduler.Current(), bounds_bindings.vertex[0], *bounds_bindings.pixel));
+        const auto &vertex_depth = bounds_bindings.vertex[0].images[0];
        const auto &pixel_depth = bounds_bindings.pixel->images[0];
        constexpr auto readonly_layout = vk::ImageLayout::eDepthReadOnlyOptimal;
        Require(name, "deferred clear with sampled read-only depth bounds",
@@ -10581,7 +10590,7 @@ public:
            RenderExecutorTestAccess::AcquireRenderTargets(
                executor, scheduler.Current(), &no_color, 0, shared_depth, shared_bindings.pixel);
        descriptor_pipelines.push_back(RenderExecutorTestAccess::CommitBindings(
-            executor, scheduler.Current(), shared_bindings.vertex,
+            executor, scheduler.Current(), shared_bindings.vertex[0],
            *shared_bindings.pixel));
        const auto expected_layout =
            stencil_write
@@ -10592,7 +10601,7 @@ public:
            vk::AccessFlagBits2::eDepthStencilAttachmentRead |
            vk::AccessFlagBits2::eDepthStencilAttachmentWrite;
        const auto &shared_image = texture_cache.GetImage(shared_depth.image_id);
-        const auto &vertex_image = shared_bindings.vertex.images[0];
+        const auto &vertex_image = shared_bindings.vertex[0].images[0];
        const auto &pixel_image = shared_bindings.pixel->images[0];
        Require(name, "sampled depth and stencil attachment layout",
                shared_depth.image_id == phased_depth.image_id &&
@@ -12173,8 +12182,9 @@ public:
      mode.provoking_vtx_last = provoking_last;
      registers.SetModeControl(mode);
      return context.GetPipelineCache().GetGraphicsPipeline(
-          std::span{&color, 1u}, depth, vertex, scheduler.Current(), &pixel,
-          vk::PrimitiveTopology::eTriangleList, false, vertex_shader, pixel_shader);
+          std::span{&color, 1u}, depth, std::span{&vertex, 1u}, scheduler.Current(), &pixel,
+          vk::PrimitiveTopology::eTriangleList, false,
+          PipelineCache::GraphicsPrograms{{vertex_shader}, pixel_shader});
    };
    auto &filled = pipeline(true, 2, 2);
    const auto draw = [&](const PipelineCache::Pipeline &selected) {
@@ -12196,7 +12206,7 @@ public:
              feedback_enabled == (depth_feedback && depth.depth_write_enable),
              "feedback was not selected only for an overlapping fragment depth read/write");
      RenderExecutorTestAccess::CommitBindings(
-          executor, command, selected, bindings.vertex, *bindings.pixel);
+          executor, command, selected, bindings.vertex[0], *bindings.pixel);
      rendering.color_attachments[0].is_clear = true;
      command.BeginRendering(rendering);
      auto cmd = command.Handle();
@@ -25504,10 +25514,160 @@ void CheckPs5GameExampleImageClearRuntimeShape() {
  std::printf("[host]    %-32s ok\n", "Ps5GameExampleImageClear");
 }

+void CheckTessellationPrograms() {
+  using namespace ShaderRecompiler;
+  const char *name = "TessellationPrograms";
+  std::array<std::vector<u32>, 3> code;
+  auto &local = code[0];
+  // The fused LS addresses one 124-byte control point using its v3 ordinal.
+  local.push_back(EncodeVop2(0x0b, 21, 255u, 3));
+  local.push_back(124);
+  local.push_back(EncodeDs0(0x0e, (3u << 8u) | 1u));
+  local.push_back(EncodeDs1Ex(0, 5, 2, 21));
+  local.push_back(EncodeSop1(0x20, 0, 6));
+  local.push_back(0xffffffffu); // Shader footer is not executable after LS handoff.
+
+  auto &control = code[1];
+  AppendVop3(&control, 0x365, 0, 193u, InlineU32(0));
+  AppendVop3(&control, 0x366, 0, 193u, Vgpr(0));
+  control.push_back(EncodeSopp(0x0a));
+  control.push_back(EncodeVopc(0xd1, 249u, 3));
+  control.push_back(EncodeVopcSdwa(0, 0, 0, 6, 1, 0, 0, 0, 0, 0, 0, 0, 1));
+  AppendVop3(&control, 0x148, 32, Vgpr(1), InlineU32(0), InlineU32(8));
+  AppendVop3(&control, 0x148, 34, Vgpr(1), InlineU32(8), InlineU32(5));
+  AppendVop3(&control, 0x346, 35, Vgpr(34), InlineU32(7), InlineU32(0));
+  control.push_back(EncodeVop2(0x0b, 29, 255u, 34));
+  control.push_back(124);
+  control.push_back(EncodeDs0(0x37, (3u << 8u) | 1u));
+  control.push_back(EncodeDs1Ex(10, 0, 0, 29));
+  control.push_back(EncodeMubuf0(0x1d, 28));
+  control.push_back(EncodeMubuf1(10, 2, 35, 2));
+  // The HS prologue's wave lane computation must refer to its control point.
+  control.push_back(EncodeMubuf0(0x1c, 120));
+  control.push_back(EncodeMubuf1(0, 2, 35, 2));
+  // The game reuses previously written data VGPRs for CP0-only ring addresses.
+  AppendVMovLiteral(&control, 4, 0x42280000u);
+  AppendVMovLiteral(&control, 5, 0x422c0000u);
+  control.push_back(EncodeVopc(0xd4, InlineU32(1), 34));
+  control.push_back(EncodeVop2(0x1a, 4, InlineU32(4), 32));
+  control.push_back(EncodeVop2(0x1a, 6, InlineU32(6), 32));
+  control.push_back(EncodeVop2(0x26, 5, 255u, 6));
+  control.push_back(0x7fc0);
+  for (u32 i = 0; i < 4; i++) {
+    AppendVMovLiteral(&control, i, std::bit_cast<u32>(static_cast<float>(i + 1)));
+  }
+  control.push_back(EncodeMubuf0(0x1e));
+  control.push_back(EncodeMubuf1(0, 2, 4, 4));
+  for (u32 i = 0; i < 4; i++) {
+    control.push_back(EncodeMubuf0(0x1c, i * 16));
+    control.push_back(EncodeMubuf1(i, 2, 5, 2));
+  }
+  AppendEnd(&control);
+
+  auto &evaluation = code[2];
+  evaluation.push_back(EncodeMubuf0(0x0d, 28));
+  evaluation.push_back(EncodeMubuf1(10, 2, 7, 4));
+  evaluation.push_back(EncodeMubuf0(0x0d, 256 + 28));
+  evaluation.push_back(EncodeMubuf1(12, 2, 7, 4));
+  evaluation.push_back(EncodeExp0(0x0c, 0xf));
+  evaluation.push_back(EncodeExp1(5, 6, 10, 12));
+  AppendEnd(&evaluation);
+
+  ShaderTessellationInputInfo tess{.input_control_points = 3,
+                                  .output_control_points = 3,
+                                  .domain = 1,
+                                  .partitioning = 2,
+                                  .output_topology = 2};
+  AnalyzeTessellationPrograms(local, control, tess);
+  Require(name, "decoded interface", tess.ls_stride == 124 && tess.hs_stride == 128,
+          "captured LS and HS address arithmetic must produce distinct strides");
+
+  constexpr std::array stages{ShaderType::Local, ShaderType::TessellationControl,
+                              ShaderType::TessellationEvaluation};
+  for (u32 stage = 0; stage < stages.size(); stage++) {
+    ShaderVertexInputInfo vertex;
+    vertex.logical_stage = stages[stage];
+    vertex.tess = tess;
+    CompileOptions options;
+    options.stage = stages[stage];
+    options.input_info.vertex = &vertex;
+    auto translated = TranslateProgram(code[stage], options);
+    std::set<u32> local_offsets, evaluation_offsets, factor_offsets, patch_offsets;
+    for (const auto *block : translated.program.blocks) {
+      for (const auto &inst : *block) {
+        if (inst.GetOpcode() != IR::ValueOpcode::GetTessellationAttribute &&
+            inst.GetOpcode() != IR::ValueOpcode::SetTessellationAttribute) {
+          continue;
+        }
+        if (!inst.Arg(1).IsImmediate()) {
+          continue;
+        }
+        switch (static_cast<IR::TessellationAttribute>(inst.Arg(0).U32())) {
+        case IR::TessellationAttribute::LocalOutput:
+          local_offsets.insert(inst.Arg(1).U32());
+          break;
+        case IR::TessellationAttribute::EvaluationInput:
+          evaluation_offsets.insert(inst.Arg(1).U32());
+          break;
+        case IR::TessellationAttribute::Factor:
+          factor_offsets.insert(inst.Arg(1).U32());
+          break;
+        case IR::TessellationAttribute::PatchOutput:
+          patch_offsets.insert(inst.Arg(1).U32());
+          break;
+        default:
+          break;
+        }
+      }
+    }
+    Require(name, "separate DS offsets", stage != 0 || local_offsets == std::set<u32>{4, 12},
+            "DS_WRITE2 must preserve both independently encoded offsets");
+    Require(name, "offchip component addressing",
+            stage != 2 || evaluation_offsets == std::set<u32>{28, 284},
+            "TES control point and component offsets must survive lowering");
+    Require(name, "triangle factor layout",
+            stage != 1 || factor_offsets == std::set<u32>{0, 4, 8, 12},
+            "the four packed guest factors must retain their order");
+    Require(name, "predicated patch addresses",
+            stage != 1 || patch_offsets == std::set<u32>{0x7fc0, 0x7fd0, 0x7fe0, 0x7ff0},
+            "CP0 patch stores must use the address assigned on their active predicate");
+    auto result = CompileProgram(std::move(translated), options, {});
+    ValidateSpirv(name, result.spirv);
+    spvtools::SpirvTools tools(SPV_ENV_VULKAN_1_2);
+    std::string assembly;
+    Require(name, "disassembly", tools.Disassemble(result.spirv, &assembly),
+            "failed to disassemble tessellation shader");
+    Require(name, "native local inputs",
+            stage != 0 || (assembly.find("VertexIndex") != std::string::npos &&
+                           assembly.find("InstanceIndex") != std::string::npos),
+            "LS v2/v5 must read native vertex/instance indices");
+    Require(name, "control point lane",
+            stage != 1 || (assembly.find("InvocationId") != std::string::npos &&
+                           assembly.find("SubgroupLocalInvocationId") == std::string::npos),
+            "HS lane masks must use control point invocation IDs");
+    Require(name, "control point barrier",
+            stage != 1 || (assembly.find("OpMemoryModel Logical GLSL450") != std::string::npos &&
+                           assembly.find("OpControlBarrier %uint_2 %uint_4 %uint_0") != std::string::npos),
+            "HS barriers must match GLSL450 control point synchronization semantics");
+    Require(name, "factor builtins",
+            stage != 1 || (assembly.find("TessLevelOuter") != std::string::npos &&
+                           assembly.find("TessLevelInner") != std::string::npos),
+            "triangle factors require outer[4] and inner[2] Vulkan builtins");
+    Require(name, "native domain and winding",
+            stage != 2 || (assembly.find("TessCoord") != std::string::npos &&
+                           assembly.find("Triangles") != std::string::npos &&
+                           assembly.find("SpacingFractionalOdd") != std::string::npos &&
+                           assembly.find("VertexOrderCw") != std::string::npos),
+            "TES must preserve guest triangle domain, partitioning, and winding");
+  }
+  std::printf("[host]    %-32s ok\n", name);
+}
+
 void CheckEmbeddedFetchVertexOffset() {
  const auto MakeFetch = [](std::initializer_list<std::pair<u32, u32>> adds,
                            std::optional<std::pair<u32, u32>> late_add = {},
-                            u32 accumulator_vgpr = 0, bool ngg_sad = false) {
+                            u32 accumulator_vgpr = 0, bool ngg_sad = false,
+                            bool local = false) {
    std::vector<u32> code;
    code.push_back(EncodeSMovB32(0, InlineU32(0)));
    code.push_back(EncodeSmem0(0x02u, 20, 4));
@@ -25526,7 +25686,7 @@ void CheckEmbeddedFetchVertexOffset() {
    if (ngg_sad) {
      AppendOffsets();
    }
-    code.push_back(EncodeVop2(0x01u, 0, Vgpr(8), 5));
+    code.push_back(EncodeVop2(0x01u, 0, Vgpr(local ? 5 : 8), local ? 2 : 5));
    if (!ngg_sad) {
      AppendOffsets();
    }
@@ -25535,26 +25695,41 @@ void CheckEmbeddedFetchVertexOffset() {
    if (late_add.has_value()) {
      AppendVop3B(&code, 0x30fu, 0, 0, late_add->first, Vgpr(late_add->second));
    }
-    AppendEnd(&code);
+    if (local) {
+      AppendVMovU32(&code, 20, 0);
+      code.push_back(EncodeDs0(0x0e, 1u << 8u));
+      code.push_back(EncodeDs1Ex(0, 10, 9, 20));
+      code.push_back(EncodeDs0(0x0e, (3u << 8u) | 2u));
+      code.push_back(EncodeDs1Ex(0, 12, 11, 20));
+      code.push_back(EncodeSop1(0x20, 0, 6));
+      code.push_back(0xffffffffu);
+    } else {
+      AppendEnd(&code);
+    }
    return code;
  };

  const auto Compile = [&](const char *name, const std::vector<u32> &code,
-                           u32 slot10) {
+                           u32 slot10, ShaderType stage = ShaderType::Vertex,
+                           Prospero::BufferFormat format = Prospero::BufferFormat::k32_32_32_32Float) {
    std::array<u32, 11> user_data{};
    user_data[10] = slot10;
    ShaderVertexInputInfo vertex;
+    vertex.logical_stage = stage;
+    vertex.tess = {.input_control_points = 3, .output_control_points = 3,
+                   .ls_stride = 124, .hs_stride = 128, .domain = 1,
+                   .partitioning = 2, .output_topology = 2};
    vertex.fetch_embedded = true;
    vertex.fetch_buffer_reg = 0;
    vertex.fetch_attrib_reg = 2;
    vertex.resources_num = 1;
    vertex.resources[0].fields[3] =
-        BufferFormat(Prospero::BufferFormat::k32_32_32_32Float) << 12u | DstSel(4, 5, 6, 7);
+        BufferFormat(format) << 12u | DstSel(4, 5, 6, 7);
    vertex.resources_dst[0].attr_id = 0;
    vertex.resources_dst[0].registers_num = 4;

    ShaderRecompiler::CompileOptions options;
-    options.stage = ShaderType::Vertex;
+    options.stage = stage;
    options.user_data_base = 8;
    options.user_data = user_data;
    options.input_info.vertex = &vertex;
@@ -25605,6 +25780,28 @@ void CheckEmbeddedFetchVertexOffset() {
    return ResolveDrawOffsets(0, vertex).second;
  };

+  for (const auto format : {Prospero::BufferFormat::k32_32_32_32Float,
+                            Prospero::BufferFormat::k32_32_32_32UInt}) {
+    const auto local = Compile("LocalEmbeddedFetch", MakeFetch({}, {}, 0, false, true),
+                               0, ShaderType::Local, format);
+    ValidateSpirv("LocalEmbeddedFetch", local.spirv);
+    Require("LocalEmbeddedFetch", "live vertex attributes",
+            std::ranges::any_of(local.program.info.inputs, [](const auto &input) {
+              return input.kind == ShaderRecompiler::IR::StageInputKind::Parameter &&
+                     input.location == 0;
+            }), "LS indexed fetch must consume native vertex attributes");
+    if (format == Prospero::BufferFormat::k32_32_32_32UInt) {
+      spvtools::SpirvTools tools(SPV_ENV_VULKAN_1_2);
+      std::string assembly;
+      Require("LocalEmbeddedFetch", "disassembly", tools.Disassemble(local.spirv, &assembly),
+              "failed to disassemble unsigned LS fetch");
+      Require("LocalEmbeddedFetch", "unsigned input type",
+              assembly.find("OpTypePointer Input %v4uint") != std::string::npos &&
+                  assembly.find("OpConvertUToF") == std::string::npos,
+              "unsigned LS attributes must preserve their native integer bits");
+    }
+  }
+
  const auto valid =
      Compile("EmbeddedFetchVertexOffset", MakeFetch({{18, 0}}), 7);
  Require(
@@ -29355,6 +29552,11 @@ int main(int argc, char **argv) {
  std::setvbuf(stdout, nullptr, _IONBF, 0);
  EnsureConfigInitialized();
  CheckLeastRecentlyUsedCacheOrdering();
+  if (argc == 2 && std::strcmp(argv[1], "--tessellation-only") == 0) {
+    CheckTessellationPrograms();
+    CheckEmbeddedFetchVertexOffset();
+    return 0;
+  }
  if (argc == 2 && std::strcmp(argv[1], "--cmpx-lt-u16-only") == 0) {
    VulkanHarness vulkan;
    RunCase(&vulkan, VectorVopcCmpxLtU16CapturedSdwaExecMask());
@@ -29777,6 +29979,7 @@ int main(int argc, char **argv) {
  CheckPm4CeCompletion(vulkan.RuntimeRenderer());
  CheckEmbeddedFetchVertexOffset();
  CheckEmbeddedFetchLaneSpill();
+  CheckTessellationPrograms();
  CheckRectListShaders();
  CheckIndirectImageKeySwitch();
  CheckPs5GameExampleImageClearRuntimeShape();