texture cache: select color-metadata clears on the GPU with conditional rendering

CMakeLists.txt

@@ -265,6 +265,7 @@ set(gpu_tiler_shader_names
 	promote_d16
 	demote_d16
 	color_transform
+	color_clear
 )
 file(GLOB gpu_tiler_shader_includes CONFIGURE_DEPENDS "${gpu_tiler_shader_dir}/gpu_tiler_*.inc")
 foreach(shader_name IN LISTS gpu_tiler_shader_names)

src/common/debugCounters.h

@@ -31,6 +31,7 @@ enum class Counter {
 	TextureEvictions,
 	ColorMetadataReadbacks, // CPU blocked reading back GPU-written DCC/CMASK metadata.
 	ColorMetadataReadbackNs,
+	ColorMetadataChecks, // GPU clear-key checks of DCC/CMASK metadata.
 	BufferUploadBytes,
 	BufferDownloadBytes,
 	BufferEvictions,

src/graphics/host_gpu/graphicContext.h

@@ -35,6 +35,7 @@ struct GraphicContext {
 	bool                               attachment_feedback_loop_enabled      = false;
 	bool                               provoking_vertex_last_enabled         = false;
 	bool                               maintenance8_enabled                  = false;
+	bool                               conditional_rendering_enabled         = false;
 	bool                               supports_block_texel_view              = false;
 	bool                                      mesh_shader_enabled                   = false;
 	vk::PhysicalDeviceMeshShaderPropertiesEXT mesh_shader_properties                = {};

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

@@ -174,6 +174,7 @@ bool BufferCache::DownloadBufferMemory(Buffer& buffer, uint64_t vaddr, uint64_t
 			    total_size += Common::AlignUp(end - start, 64);
 		    });
 		    m_gpu_modified_ranges.Subtract(address, bytes);
+		    ForgetColorMetadataChecks(address, bytes);
 	    });
 	if (copies.empty()) {
 		return false;
@@ -519,6 +520,7 @@ std::pair<Buffer*, uint64_t> BufferCache::ObtainBuffer(uint64_t vaddr, uint64_t
 	(void)SynchronizeBuffer(buffer, vaddr, size, is_written, is_texel_buffer);
 	if (is_written) {
 		m_gpu_modified_ranges.Add(vaddr, size);
+		ForgetColorMetadataChecks(vaddr, size);
 	}
 	return {&buffer, buffer.Offset(vaddr)};
 }
@@ -623,6 +625,29 @@ bool BufferCache::IsRegionRegistered(uint64_t vaddr, uint64_t size) {
 	return address + m_slot_buffers[id].Size() > vaddr;
 }
 
+void BufferCache::MarkColorMetadataChecked(uint64_t vaddr, uint64_t size, uint32_t keys) {
+	for (uint32_t slot = 0; slot < m_checked_color_metadata.size(); slot++) {
+		if ((keys & (1u << slot)) != 0) {
+			m_checked_color_metadata[slot].Add(vaddr, size);
+		}
+	}
+}
+
+bool BufferCache::IsColorMetadataChecked(uint64_t vaddr, uint64_t size, uint32_t keys) const {
+	for (uint32_t slot = 0; slot < m_checked_color_metadata.size(); slot++) {
+		if ((keys & (1u << slot)) != 0 && !m_checked_color_metadata[slot].Contains(vaddr, size)) {
+			return false;
+		}
+	}
+	return true;
+}
+
+void BufferCache::ForgetColorMetadataChecks(uint64_t vaddr, uint64_t size) {
+	for (auto& checked: m_checked_color_metadata) {
+		checked.Subtract(vaddr, size);
+	}
+}
+
 bool BufferCache::IsRegionGpuModified(uint64_t vaddr, uint64_t size) {
 	return m_memory_tracker.IsRegionGpuModified(vaddr, size);
 }

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

@@ -11,6 +11,7 @@
 #include "graphics/host_gpu/renderer/cache/multiLevelPageTable.h"
 #include "graphics/host_gpu/renderer/cache/streamBuffer.h"
 
+#include <array>
 #include <map>
 #include <span>
 #include <utility>
@@ -79,6 +80,10 @@ public:
 	[[nodiscard]] bool HasGpuDirtyBytes(uint64_t vaddr, uint64_t size);
 	[[nodiscard]] bool IsRegionCpuModified(uint64_t vaddr, uint64_t size);
 	[[nodiscard]] bool IsRegionGpuModified(uint64_t vaddr, uint64_t size);
+	// Color metadata the GPU found free of the clear keys in `keys` (bits of code >> 5) stays so
+	// until a GPU write or download.
+	void               MarkColorMetadataChecked(uint64_t vaddr, uint64_t size, uint32_t keys);
+	[[nodiscard]] bool IsColorMetadataChecked(uint64_t vaddr, uint64_t size, uint32_t keys) const;
 	void               ProcessFaultBuffer();
 	void               SynchronizeBda(const RangeSet& mapped);
 	void               QueueBdaSync(uint64_t vaddr, uint64_t size) {
@@ -124,6 +129,7 @@ private:
 	// Synchronous downloads publish before returning; asynchronous callers wait before reuse.
 	template <bool async>
 	[[nodiscard]] bool DownloadBufferMemory(Buffer& buffer, uint64_t vaddr, uint64_t size);
+	void               ForgetColorMetadataChecks(uint64_t vaddr, uint64_t size);
 
 	GraphicContext&                                   m_graphics;
 	CommandScheduler&                                 m_scheduler;
@@ -135,6 +141,8 @@ private:
 	BufferMap                                         m_buffers;
 	PageTable                                         m_page_table;
 	RangeSet                                          m_gpu_modified_ranges;
+	// One per TileManager::ColorClearSlots key slot.
+	std::array<RangeSet, 8>                           m_checked_color_metadata;
 	MemoryTracker                                     m_memory_tracker;
 	StreamBuffer                                      m_staging_buffer;
 	StreamBuffer                                      m_stream_buffer;

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

@@ -1067,15 +1067,78 @@ void TextureCache::MaterializeColorClear(ImageId id, const ImageDesc& desc,
 	if (first >= layers || count > layers - first) {
 		EXIT("TextureCache: color view exceeds its native metadata slices\n");
 	}
-	// Finish native metadata writes before reading backing bytes. This can submit the scheduler,
+	const auto       slice_size = range.size / layers;
-	// so discovery runs before final draw uploads and never holds the texture lock across it.
+	const GuestRange viewed {range.address + slice_size * first, slice_size * count};
-	if (m_buffer_cache.IsRegionGpuModified(range.address, range.size)) {
+	if (m_buffer_cache.IsRegionGpuModified(viewed.address, viewed.size)) {
+		std::array<vk::ClearAttachment, TileManager::ColorClearSlots> clears {};
+		uint32_t keys = 0;
+		for (uint32_t key = 0; key < clears.size(); key++) {
+			clears[key].aspectMask = vk::ImageAspectFlagBits::eColor;
+			if (DecodeColorClear(desc, static_cast<uint8_t>(key << 5u),
+			                     clears[key].clearValue.color)) {
+				keys |= 1u << key;
+			}
+		}
+		// No selected key to clear, or none left since the last check.
+		if (m_buffer_cache.IsColorMetadataChecked(viewed.address, viewed.size, keys)) {
+			return;
+		}
+		{
+			std::scoped_lock lock {m_lock};
+			auto&            image = m_slot_images[id];
+			// Compressed images cannot upload, so they must already hold native contents.
+			const bool native_current = (image.usage.render_target || image.IsGpuModified()) &&
+			                            !image.IsBufferModified() && !image.IsCpuDirty();
+			// 2D-array attachment views select 2D layers or, of a volume, depth slices.
+			if (m_graphics.conditional_rendering_enabled &&
+			    image.backing.image_type != vk::ImageType::e1D &&
+			    count <= m_graphics.GetPhysicalDeviceProperties().limits.maxFramebufferLayers &&
+			    (image.backing.usage & vk::ImageUsageFlagBits::eColorAttachment) &&
+			    (m_graphics.GetFormatProperties(view.format).optimalTilingFeatures &
+			     vk::FormatFeatureFlagBits::eColorAttachment) &&
+			    (desc.info.metadata.compression == VideoOutCompression::Uncompressed ||
+			     native_current)) {
+				// The GPU selects each slice's clear, so the CPU never waits for the metadata.
+				Common::DebugCounters::Add(Common::DebugCounters::Counter::ColorMetadataChecks);
+				InitializeImage(id);
+				const bool expand = desc.type != BindingType::VideoOut;
+				const auto [buffer, offset] =
+				    m_buffer_cache.ObtainBuffer(viewed.address, viewed.size, expand);
+				const auto predicates = m_tiler.CheckColorClear(
+				    {buffer->Handle(), offset, viewed.size}, count, keys, expand);
+				if (expand) {
+					m_buffer_cache.MarkColorMetadataChecked(viewed.address, viewed.size, keys);
+				}
+				const auto command = m_scheduler.Current().Handle();
+				const auto area    = BeginColorRendering(image, view.format, view.base_level,
+				                                         image_first, count, nullptr, command);
+				vk::ClearRect rect {.rect = area, .layerCount = 1};
+				for (uint32_t slice = 0; slice < count; slice++) {
+					rect.baseArrayLayer = slice;
+					for (uint32_t key = 0; key < clears.size(); key++) {
+						if ((keys & (1u << key)) != 0) {
+							const vk::ConditionalRenderingBeginInfoEXT predicate {
+							    .buffer = predicates,
+							    .offset = (slice * clears.size() + key) * sizeof(uint32_t)};
+							command.beginConditionalRenderingEXT(&predicate);
+							command.clearAttachments(1, &clears[key], 1, &rect);
+							command.endConditionalRenderingEXT();
+						}
+					}
+				}
+				command.endRendering();
+				CommitGpuWrite(image);
+				return;
+			}
+		}
+		// Finish native metadata writes before reading backing bytes. This can submit the
+		// scheduler, so discovery runs before final draw uploads and never holds the texture
+		// lock across it.
 		Common::DebugCounters::ScopedTimer timer(
 		    Common::DebugCounters::Counter::ColorMetadataReadbacks,
 		    Common::DebugCounters::Counter::ColorMetadataReadbackNs);
-		m_buffer_cache.ReadMemory(range.address, range.size, false);
+		m_buffer_cache.ReadMemory(viewed.address, viewed.size, false);
 	}
-	const auto slice_size = range.size / layers;
 	for (uint32_t slice = 0; slice < count; slice++) {
 		const auto address = range.address + slice_size * (first + slice);
 		uint8_t code = 0;
@@ -1768,6 +1831,7 @@ bool BufferCache::SynchronizeBufferFromImage(Buffer& buffer, uint64_t vaddr, uin
 	     copy_size != image.info.data.size)) {
 		return false;
 	}
+	ForgetColorMetadataChecks(image.info.data.address, copy_size);
 	m_texture_cache.DownloadImage(image, buffer, buf_offset, copy_size, levels);
 	return true;
 }

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

@@ -14,6 +14,7 @@
 #include "gpu_tiler_shaders/gpu_tiler_standard64_3d_spv.h"
 #include "gpu_tiler_shaders/gpu_tiler_standard64_spv.h"
 #include "gpu_tiler_shaders/gpu_tiler_color_transform_spv.h"
+#include "gpu_tiler_shaders/gpu_tiler_color_clear_spv.h"
 #include "graphics/host_gpu/graphicContext.h"
 #include "graphics/host_gpu/renderer/cache/streamBuffer.h"
 #include "graphics/host_gpu/renderer/commandScheduler.h"
@@ -71,8 +72,8 @@ TileManager::~TileManager() {
 			m_graphics.device.destroyPipeline(pipeline, nullptr);
 		}
 	}
-	for (auto pipeline:
+	for (auto pipeline: {m_d16_to_d24, m_d16_to_d32, m_d24_to_d16, m_d32_to_d16, m_color_transform,
-	     {m_d16_to_d24, m_d16_to_d32, m_d24_to_d16, m_d32_to_d16, m_color_transform}) {
+	                     m_color_clear}) {
 		if (pipeline != nullptr) {
 			m_graphics.device.destroyPipeline(pipeline, nullptr);
 		}
@@ -627,4 +628,60 @@ TileManager::Result TileManager::TransformColor(Result input, ColorTransform tra
 	return result;
 }
 
+vk::Buffer TileManager::CheckColorClear(Result metadata, uint32_t slices, uint32_t keys,
+                                        bool expand) {
+	struct Check {
+		uint32_t metadata_base;
+		uint32_t words;
+		uint32_t keys;
+		uint32_t expand;
+	};
+	static_assert(sizeof(Check) <= sizeof(Conversion));
+	const uint32_t max_slices =
+	    m_graphics.GetPhysicalDeviceProperties().limits.maxFramebufferLayers;
+	EXIT_IF(slices == 0 || slices > max_slices || metadata.size % (4ull * slices) != 0 ||
+	        metadata.size / 4 > UINT32_MAX);
+	if (m_color_clear == nullptr) {
+		m_color_clear      = CreatePipeline(GPU_TILER_COLOR_CLEAR_SPV, nullptr);
+		m_clear_predicates = std::make_unique<Buffer>(
+		    m_graphics, m_scheduler, MemoryUsage::DeviceLocal, 0,
+		    vk::BufferUsageFlagBits::eStorageBuffer |
+		        vk::BufferUsageFlagBits::eConditionalRenderingEXT,
+		    uint64_t {max_slices} * ColorClearSlots * sizeof(uint32_t));
+	}
+	const auto source = BindStorage(metadata, metadata.size);
+	const vk::DescriptorBufferInfo infos[] {source.info,
+	                                        {m_clear_predicates->Handle(), 0, VK_WHOLE_SIZE}};
+	std::array<vk::WriteDescriptorSet, 2> writes {};
+	for (uint32_t index = 0; index < writes.size(); index++) {
+		writes[index].dstBinding      = index;
+		writes[index].descriptorCount = 1;
+		writes[index].descriptorType  = vk::DescriptorType::eStorageBuffer;
+		writes[index].pBufferInfo     = &infos[index];
+	}
+	// Orders the metadata writers and the previous check's predicate reads before this check.
+	vk::MemoryBarrier barrier {
+	    .srcAccessMask = vk::AccessFlagBits::eMemoryWrite,
+	    .dstAccessMask = vk::AccessFlagBits::eShaderRead | vk::AccessFlagBits::eShaderWrite};
+	m_scheduler.EndRendering();
+	auto command = m_scheduler.Current().Handle();
+	command.pipelineBarrier(vk::PipelineStageFlagBits::eAllCommands,
+	                        vk::PipelineStageFlagBits::eComputeShader, {}, 1, &barrier, 0, nullptr,
+	                        0, nullptr);
+	command.bindPipeline(vk::PipelineBindPoint::eCompute, m_color_clear);
+	command.pushDescriptorSetKHR(vk::PipelineBindPoint::eCompute, m_pipeline_layout, 0,
+	                             static_cast<uint32_t>(writes.size()), writes.data());
+	const Check push {source.base, static_cast<uint32_t>(metadata.size / 4 / slices), keys,
+	                  expand};
+	command.pushConstants(m_pipeline_layout, vk::ShaderStageFlagBits::eCompute, 0, sizeof(push),
+	                      &push);
+	command.dispatch(slices, 1, 1);
+	barrier.srcAccessMask = vk::AccessFlagBits::eShaderWrite;
+	barrier.dstAccessMask = vk::AccessFlagBits::eMemoryRead | vk::AccessFlagBits::eMemoryWrite;
+	command.pipelineBarrier(vk::PipelineStageFlagBits::eComputeShader,
+	                        vk::PipelineStageFlagBits::eAllCommands, {}, 1, &barrier, 0, nullptr,
+	                        0, nullptr);
+	return m_clear_predicates->Handle();
+}
+
 } // namespace Libs::Graphics

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

@@ -54,6 +54,13 @@ public:
 	                                const D16Layout& layout);
 	[[nodiscard]] Result TransformColor(Result input, ColorTransform transform, bool to_host);
 	void TransformColor(Result input, Result output, ColorTransform transform, bool to_host);
+	// Color-clear keys are multiples of 0x20, so key >> 5 selects one of these slots.
+	static constexpr uint32_t ColorClearSlots = 8;
+	// Returns the buffer whose dword slice * ColorClearSlots + (key >> 5) is set when every byte
+	// of that metadata slice equals a key whose bit `keys` selects; a hit expands to 0xff when
+	// `expand` is set.
+	[[nodiscard]] vk::Buffer CheckColorClear(Result metadata, uint32_t slices, uint32_t keys,
+	                                         bool expand);
 
 private:
 	friend struct TileManagerTestAccess;
@@ -132,6 +139,8 @@ private:
 	vk::Pipeline                            m_d24_to_d16      = nullptr;
 	vk::Pipeline                            m_d32_to_d16      = nullptr;
 	vk::Pipeline                            m_color_transform = nullptr;
+	vk::Pipeline                            m_color_clear     = nullptr;
+	std::unique_ptr<Buffer>                 m_clear_predicates;
 };
 
 } // namespace Libs::Graphics

src/graphics/host_gpu/shaders/gpu_tiler_color_clear.comp

@@ -0,0 +1,50 @@
+#version 450
+
+// One workgroup per metadata slice; see TileManager::CheckColorClear.
+layout(local_size_x = 256) in;
+
+layout(set = 0, binding = 0, std430) buffer Metadata { uint data[]; } metadata;
+layout(set = 0, binding = 1, std430) writeonly buffer Predicates { uint data[]; } predicates;
+
+layout(push_constant, std430) uniform Check {
+	uint metadata_base;
+	uint words;
+	uint keys;
+	uint expand;
+} params;
+
+const uint Slots = 8u; // TileManager::ColorClearSlots
+
+shared uint key;
+shared uint mismatch;
+
+void main() {
+	const uint slice = gl_WorkGroupID.x;
+	const uint base  = (params.metadata_base >> 2u) + slice * params.words;
+	// Each invocation reads, and may rewrite, only its own words.
+	if (gl_LocalInvocationIndex == 0u) {
+		key      = metadata.data[base] & 0xffu;
+		mismatch = 0u;
+	}
+	barrier();
+	const uint pattern = key * 0x01010101u;
+	bool       same    = true;
+	for (uint i = gl_LocalInvocationIndex; same && i < params.words; i += gl_WorkGroupSize.x) {
+		same = metadata.data[base + i] == pattern;
+	}
+	if (!same) {
+		atomicOr(mismatch, 1u);
+	}
+	barrier();
+	const uint slot = key >> 5u;
+	const bool hit  = mismatch == 0u && (key & 0x1fu) == 0u && (params.keys & (1u << slot)) != 0u;
+	if (gl_LocalInvocationIndex < Slots) {
+		predicates.data[slice * Slots + gl_LocalInvocationIndex] =
+		    uint(hit && slot == gl_LocalInvocationIndex);
+	}
+	if (hit && params.expand != 0u) {
+		for (uint i = gl_LocalInvocationIndex; i < params.words; i += gl_WorkGroupSize.x) {
+			metadata.data[base + i] = 0xffffffffu;
+		}
+	}
+}

src/graphics/presentation/performanceOverlay.cpp

@@ -313,6 +313,7 @@ void GpuActivityRows(const Snapshot& shown) {
 	Row("DCC/CMASK readbacks",
 	    Fmt("%.0f ms/s  (%.0f/s)", Rate(shown, Counter::ColorMetadataReadbackNs) / 1e6,
 	        Rate(shown, Counter::ColorMetadataReadbacks)));
+	Row("DCC/CMASK GPU checks", Fmt("%.0f/s", Rate(shown, Counter::ColorMetadataChecks)));
 }
 
 void CompilationRows(const Snapshot& shown) {

src/graphics/presentation/window/vulkanWindow.cpp

@@ -499,6 +499,13 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 		image_atomic_int64.pNext = supported_features2.pNext;
 		supported_features2.pNext = &image_atomic_int64;
 	}
+	const bool conditional_rendering_extension =
+	    HasExtension(device_extensions, VK_EXT_CONDITIONAL_RENDERING_EXTENSION_NAME);
+	vk::PhysicalDeviceConditionalRenderingFeaturesEXT conditional_rendering {};
+	if (conditional_rendering_extension) {
+		conditional_rendering.pNext = supported_features2.pNext;
+		supported_features2.pNext   = &conditional_rendering;
+	}
 	const bool barycentric_extension =
 	    HasExtension(device_extensions, VK_KHR_FRAGMENT_SHADER_BARYCENTRIC_EXTENSION_NAME);
 	vk::PhysicalDeviceFragmentShaderBarycentricFeaturesKHR supported_barycentric {};
@@ -572,6 +579,8 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 	    feedback_extensions && feedback_layout.attachmentFeedbackLoopLayout &&
 	    feedback_dynamic.attachmentFeedbackLoopDynamicState;
 	graphics.maintenance8_enabled = maintenance8_extension && maintenance8.maintenance8;
+	graphics.conditional_rendering_enabled =
+	    conditional_rendering_extension && conditional_rendering.conditionalRendering;
 	LOGF("Vulkan maintenance8 support: %s\n",
 	     graphics.maintenance8_enabled ? "true" : "false");
 	LOGF("Vulkan depth feedback support: %s\n",
@@ -670,6 +679,10 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 		maintenance8.pNext = const_cast<void*>(create_info.pNext);
 		create_info.pNext = &maintenance8;
 	}
+	if (graphics.conditional_rendering_enabled) {
+		conditional_rendering.pNext = const_cast<void*>(create_info.pNext);
+		create_info.pNext = &conditional_rendering;
+	}
 	create_info.pQueueCreateInfos       = &queue_create_info;
 	create_info.queueCreateInfoCount    = 1;
 	create_info.enabledExtensionCount   = static_cast<uint32_t>(device_extensions.size());
@@ -1046,7 +1059,8 @@ void WindowContext::CreateVulkan() {
 		                             VK_EXT_MESH_SHADER_EXTENSION_NAME,
 		                             VK_KHR_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_EXTENSION_NAME,
 		                             VK_EXT_DEPTH_RANGE_UNRESTRICTED_EXTENSION_NAME,
-		                             VK_KHR_FRAGMENT_SHADER_BARYCENTRIC_EXTENSION_NAME}) {
+		                             VK_KHR_FRAGMENT_SHADER_BARYCENTRIC_EXTENSION_NAME,
+		                             VK_EXT_CONDITIONAL_RENDERING_EXTENSION_NAME}) {
 			if (HasExtension(available_extensions, extension)) {
 				device_extensions.push_back(extension);
 			}

tests/ShaderRecompilerComputeTests.cpp

@@ -10266,6 +10266,32 @@ public:
                   ReadCachedTexel(name, context, color.image_id, {0, 0, 31}) ==
                       std::vector<u32>{0x5a5a5a5a},
               "clearing one volume slice changed untouched slice contents");
+      // A GPU writer clears slice 9's metadata; its bind selects the clear on the GPU.
+      auto [writer, writer_offset] = resources.GetBufferCache().ObtainBuffer(
+          dcc_address + 9 * metadata_slice_size, metadata_slice_size, true);
+      writer->Fill(writer_offset, metadata_slice_size, 0);
+      Common::DebugCounters::SetEnabled(true);
+      const auto readbacks =
+          Common::DebugCounters::Get(Common::DebugCounters::Counter::ColorMetadataReadbacks);
+      const auto checks =
+          Common::DebugCounters::Get(Common::DebugCounters::Counter::ColorMetadataChecks);
+      RenderExecutorTestAccess::ResetBindings(executor);
+      registers.SetColorView(0, {.base_array_slice_index = 9, .last_array_slice_index = 9});
+      RenderColorInfo gpu_slice{};
+      RenderExecutorTestAccess::ResolveRenderColorTarget(
+          executor, scheduler.Current(), gpu_slice, 0);
+      Common::DebugCounters::SetEnabled(false);
+      Require(name, "GPU-written volume metadata clears on the GPU",
+              gpu_slice.image_id == color.image_id &&
+                  Common::DebugCounters::Get(
+                      Common::DebugCounters::Counter::ColorMetadataReadbacks) == readbacks &&
+                  Common::DebugCounters::Get(
+                      Common::DebugCounters::Counter::ColorMetadataChecks) == checks + 1 &&
+                  ReadCachedTexel(name, context, color.image_id, {0, 0, 9}) ==
+                      std::vector<u32>{0} &&
+                  ReadCachedTexel(name, context, color.image_id, {0, 0, 31}) ==
+                      std::vector<u32>{0x5a5a5a5a},
+              "a volume slice's GPU-written metadata was read back or cleared other slices");
       auto storage_desc = color.desc;
       storage_desc.type = BindingType::Storage;
       storage_desc.info.metadata = {};
@@ -12271,14 +12297,43 @@ public:
         const auto readbacks =
             Common::DebugCounters::Get(Common::DebugCounters::Counter::ColorMetadataReadbacks);
         bind();
-        Common::DebugCounters::SetEnabled(false);
-        Require(name, "metadata readback counter",
-                Common::DebugCounters::Get(
-                    Common::DebugCounters::Counter::ColorMetadataReadbacks) == readbacks + 1,
-                "binding GPU-written metadata did not count exactly one readback");
         Require(name, "same-value raw native re-clear",
                 read_texel() == expected,
                 "a repeated metadata fill did not clear an already drawn target");
+        const auto checks =
+            Common::DebugCounters::Get(Common::DebugCounters::Counter::ColorMetadataChecks);
+        bind();
+        Require(name, "checked metadata skips rebinding",
+                Common::DebugCounters::Get(
+                    Common::DebugCounters::Counter::ColorMetadataChecks) == checks,
+                "rebinding unwritten GPU-checked metadata repeated its clear check");
+        check_expanded_metadata();
+        paint();
+        fill_metadata(metadata_words - 1, true);
+        bind();
+        Require(name, "partial raw native metadata fill", retains_painted_texel(),
+                "a nonuniform GPU-written metadata slice cleared the whole target");
+        fill_metadata(metadata_words, true);
+        bind();
+        Require(name, "complete raw native fill after partial fill", read_texel() == expected,
+                "completing a GPU-written metadata fill did not restore its clear");
+        if (layout == Prospero::ChannelLayout::k16_16_16_16 && !fill_case.cmask &&
+            !fill_case.reuse_unorm) {
+          auto video = color.desc;
+          video.type = BindingType::VideoOut;
+          paint();
+          fill_metadata(metadata_words, true);
+          const bool cleared =
+              texture_cache.FindImage(video) == color.image_id && read_texel() == expected;
+          std::vector<uint32_t> keys(metadata_words);
+          context.GetBufferCache().ReadMemory(dcc_address, metadata_size, false);
+          Require(name, "video-out clear keeps its metadata keys",
+                  cleared &&
+                      LibKernel::Memory::TryReadBacking(dcc_address, keys.data(), metadata_size) &&
+                      std::ranges::all_of(keys,
+                                          [&](uint32_t word) { return word == fill_case.fill; }),
+                  "a video-out bind skipped its clear or expanded the flip's metadata keys");
+        }
         paint();
         WriteMetadata(context, dcc_address, metadata_size, UINT32_MAX);
         fill_metadata(metadata_words - 1);
@@ -12287,15 +12342,15 @@ public:
                 retains_painted_texel(),
                 "a nonuniform metadata allocation cleared the whole target");
         fill_metadata(metadata_words);
-        if (fill_case.goat) {
-          Require(name, "repeated Goat fill stays CPU-owned",
-                  !resources.GetBufferCache().IsRegionGpuModified(dcc_address, metadata_size),
-                  "re-clearing expanded metadata required another GPU readback");
-        }
         bind();
         Require(name, "complete fill after partial fill",
                 read_texel() == expected,
                 "completing the metadata overwrite did not restore its clear");
+        Common::DebugCounters::SetEnabled(false);
+        Require(name, "GPU-written metadata stays on the GPU",
+                Common::DebugCounters::Get(
+                    Common::DebugCounters::Counter::ColorMetadataReadbacks) == readbacks,
+                "binding GPU-written color metadata read it back");
 
         // Consuming a clear key updates metadata backing without invalidating
         // a pooled image whose first texel precedes the metadata range.
@@ -20434,6 +20489,7 @@ private:
     m_runtime_context.queue = m_queue;
     m_runtime_context.shader_image_int64_atomics_enabled = true;
     m_runtime_context.maintenance8_enabled = true;
+    m_runtime_context.conditional_rendering_enabled = true;
     m_runtime_context.attachment_feedback_loop_enabled = m_rasterization_supported;
     m_runtime_context.provoking_vertex_last_enabled = m_rasterization_supported;
     const vk::PhysicalDeviceImageFormatInfo2 block_texel_view_info{
@@ -20717,7 +20773,10 @@ private:
     vk::PhysicalDeviceRobustness2FeaturesEXT robustness2{};
     robustness2.nullDescriptor = true;
     robustness2.pNext = &maintenance8;
-    device_info.pNext = &robustness2;
+    vk::PhysicalDeviceConditionalRenderingFeaturesEXT conditional_rendering{};
+    conditional_rendering.conditionalRendering = true;
+    conditional_rendering.pNext = &robustness2;
+    device_info.pNext = &conditional_rendering;
     vk::PhysicalDeviceFeatures device_features{};
     device_features.shaderStorageImageWriteWithoutFormat = true;
     device_features.shaderImageGatherExtended = true;
@@ -20736,7 +20795,8 @@ private:
         VK_EXT_SHADER_IMAGE_ATOMIC_INT64_EXTENSION_NAME,
         VK_EXT_ROBUSTNESS_2_EXTENSION_NAME,
         VK_KHR_MAINTENANCE_8_EXTENSION_NAME,
-        VK_EXT_IMAGE_VIEW_MIN_LOD_EXTENSION_NAME};
+        VK_EXT_IMAGE_VIEW_MIN_LOD_EXTENSION_NAME,
+        VK_EXT_CONDITIONAL_RENDERING_EXTENSION_NAME};
     if (m_rasterization_supported) {
       device_extensions.push_back(VK_EXT_DEPTH_CLIP_ENABLE_EXTENSION_NAME);
       device_extensions.push_back(VK_EXT_DEPTH_CLIP_CONTROL_EXTENSION_NAME);