Paging Dr. Instancer

- Goal: avoid needing to re-upload everything when instance count for
  one instancer changes
- Solution: store instances in pages of 32
- Allocate pages in a GPU arena
- Store one uint per page to indicate which model the instances in the
  page belong to, and how many instances are actually stored in the page
- Instancers eagerly allocate and free pages as their instance count
  changes
- Instancers will not necessarily store instances contiguously anymore,
  but that's okay because any given cull workgroup will only reference a
  single page
- Culling threads *will* write instances contiguously however, and so we
  still need to keep track of a base instance per instancer, and the
  target buffer logic does not change

common/src/backend/java/dev/engine_room/flywheel/backend/engine/AbstractArena.java

@@ -1,23 +1,17 @@
 package dev.engine_room.flywheel.backend.engine;
 
-import dev.engine_room.flywheel.lib.memory.MemoryBlock;
 import it.unimi.dsi.fastutil.ints.IntArrayList;
 import it.unimi.dsi.fastutil.ints.IntList;
 
-public class Arena {
-	private final long elementSizeBytes;
-
-	private MemoryBlock memoryBlock;
-
-	// Monotonic index, generally represents the size of the arena.
-	private int top = 0;
+public abstract class AbstractArena {
+	protected final long elementSizeBytes;
 	// List of free indices.
 	private final IntList freeStack = new IntArrayList();
+	// Monotonic index, generally represents the size of the arena.
+	private int top = 0;
 
-	public Arena(long elementSizeBytes, int initialCapacity) {
+	public AbstractArena(long elementSizeBytes) {
 		this.elementSizeBytes = elementSizeBytes;
-
-		memoryBlock = MemoryBlock.malloc(elementSizeBytes * initialCapacity);
 	}
 
 	public int alloc() {
@@ -27,8 +21,8 @@ public class Arena {
 		}
 
 		// Make sure there's room to increment top.
-		if (top * elementSizeBytes >= memoryBlock.size()) {
-			memoryBlock = memoryBlock.realloc(memoryBlock.size() * 2);
+		if (top * elementSizeBytes >= byteCapacity()) {
+			resize();
 		}
 
 		// Return the top index and increment.
@@ -40,19 +34,15 @@ public class Arena {
 		freeStack.add(i);
 	}
 
-	public long indexToPointer(int i) {
-		return memoryBlock.ptr() + i * elementSizeBytes;
-	}
-
-	public void delete() {
-		memoryBlock.free();
+	public long byteOffsetOf(int i) {
+		return i * elementSizeBytes;
 	}
 
 	public int capacity() {
 		return top;
 	}
 
-	public long byteCapacity() {
-		return memoryBlock.size();
-	}
+	public abstract long byteCapacity();
+
+	protected abstract void resize();
 }

common/src/backend/java/dev/engine_room/flywheel/backend/engine/AbstractInstancer.java

@@ -122,12 +122,12 @@ public abstract class AbstractInstancer<I extends Instance> implements Instancer
 
 		if (writePos < newSize) {
 			// Since we'll be shifting everything into this space we can consider it all changed.
-			changed.set(writePos, newSize);
+			setRangeChanged(writePos, newSize);
 		}
 
 		// We definitely shouldn't consider the deleted instances as changed though,
 		// else we might try some out of bounds accesses later.
-		changed.clear(newSize, oldSize);
+		clearChangedRange(newSize, oldSize);
 
 		// Punch out the deleted instances, shifting over surviving instances to fill their place.
 		for (int scanPos = writePos; (scanPos < oldSize) && (writePos < newSize); scanPos++, writePos++) {
@@ -155,6 +155,14 @@ public abstract class AbstractInstancer<I extends Instance> implements Instancer
 				.clear();
 	}
 
+	protected void clearChangedRange(int start, int end) {
+		changed.clear(start, end);
+	}
+
+	protected void setRangeChanged(int start, int end) {
+		changed.set(start, end);
+	}
+
 	/**
 	 * Clear all instances without freeing resources.
 	 */

common/src/backend/java/dev/engine_room/flywheel/backend/engine/CpuArena.java

@@ -0,0 +1,30 @@
+package dev.engine_room.flywheel.backend.engine;
+
+import dev.engine_room.flywheel.lib.memory.MemoryBlock;
+
+public class CpuArena extends AbstractArena {
+
+	private MemoryBlock memoryBlock;
+
+	public CpuArena(long elementSizeBytes, int initialCapacity) {
+		super(elementSizeBytes);
+
+		memoryBlock = MemoryBlock.malloc(elementSizeBytes * initialCapacity);
+	}
+
+	public long indexToPointer(int i) {
+		return memoryBlock.ptr() + i * elementSizeBytes;
+	}
+
+	public void delete() {
+		memoryBlock.free();
+	}
+
+	public long byteCapacity() {
+		return memoryBlock.size();
+	}
+
+	protected void resize() {
+		memoryBlock = memoryBlock.realloc(memoryBlock.size() * 2);
+	}
+}

common/src/backend/java/dev/engine_room/flywheel/backend/engine/LightStorage.java

@@ -46,7 +46,7 @@ public class LightStorage {
 
 	private final LevelAccessor level;
 
-	private final Arena arena;
+	private final CpuArena arena;
 	private final Long2IntMap section2ArenaIndex = new Long2IntOpenHashMap();
 	{
 		section2ArenaIndex.defaultReturnValue(INVALID_SECTION);
@@ -62,7 +62,7 @@ public class LightStorage {
 	public LightStorage(LevelAccessor level) {
 		this.level = level;
 
-		arena = new Arena(SECTION_SIZE_BYTES, DEFAULT_ARENA_CAPACITY_SECTIONS);
+		arena = new CpuArena(SECTION_SIZE_BYTES, DEFAULT_ARENA_CAPACITY_SECTIONS);
 	}
 
 	/**

common/src/backend/java/dev/engine_room/flywheel/backend/engine/embed/EnvironmentStorage.java

@@ -1,6 +1,6 @@
 package dev.engine_room.flywheel.backend.engine.embed;
 
-import dev.engine_room.flywheel.backend.engine.Arena;
+import dev.engine_room.flywheel.backend.engine.CpuArena;
 import it.unimi.dsi.fastutil.objects.ReferenceLinkedOpenHashSet;
 import it.unimi.dsi.fastutil.objects.ReferenceSet;
 
@@ -13,7 +13,7 @@ public class EnvironmentStorage {
 
 	// Note than the arena starts indexing at zero, but we reserve zero for the identity matrix.
 	// Any time an ID from the arena is written we want to add one to it.
-	public final Arena arena = new Arena(MATRIX_SIZE_BYTES, 32);
+	public final CpuArena arena = new CpuArena(MATRIX_SIZE_BYTES, 32);
 
 	{
 		// Reserve the identity matrix. Burns a few bytes but oh well.

common/src/backend/java/dev/engine_room/flywheel/backend/engine/indirect/IndirectBuffers.java

@@ -61,9 +61,9 @@ public class IndirectBuffers {
 	 */
 	private final MemoryBlock multiBindBlock;
 	private final long instanceStride;
-	public final ResizableStorageArray instance;
+
+	public final InstancePager pageFile;
 	public final ResizableStorageArray target;
-	public final ResizableStorageArray modelIndex;
 	public final ResizableStorageArray model;
 	public final ResizableStorageArray draw;
 
@@ -71,30 +71,27 @@ public class IndirectBuffers {
 		this.instanceStride = instanceStride;
 		this.multiBindBlock = MemoryBlock.calloc(BUFFERS_SIZE_BYTES, 1);
 
-		instance = new ResizableStorageArray(instanceStride, INSTANCE_GROWTH_FACTOR);
+		pageFile = new InstancePager(instanceStride);
 		target = new ResizableStorageArray(INT_SIZE, INSTANCE_GROWTH_FACTOR);
-		modelIndex = new ResizableStorageArray(INT_SIZE, INSTANCE_GROWTH_FACTOR);
 		model = new ResizableStorageArray(MODEL_STRIDE, MODEL_GROWTH_FACTOR);
 		draw = new ResizableStorageArray(DRAW_COMMAND_STRIDE, DRAW_GROWTH_FACTOR);
 	}
 
 	void updateCounts(int instanceCount, int modelCount, int drawCount) {
-		instance.ensureCapacity(instanceCount);
 		target.ensureCapacity(instanceCount);
-		modelIndex.ensureCapacity(instanceCount);
 		model.ensureCapacity(modelCount);
 		draw.ensureCapacity(drawCount);
 
 		final long ptr = multiBindBlock.ptr();
-		MemoryUtil.memPutInt(ptr + INSTANCE_HANDLE_OFFSET, instance.handle());
+		MemoryUtil.memPutInt(ptr + INSTANCE_HANDLE_OFFSET, pageFile.storage.handle());
 		MemoryUtil.memPutInt(ptr + TARGET_HANDLE_OFFSET, target.handle());
-		MemoryUtil.memPutInt(ptr + MODEL_INDEX_HANDLE_OFFSET, modelIndex.handle());
+		MemoryUtil.memPutInt(ptr + MODEL_INDEX_HANDLE_OFFSET, pageFile.pageTable.handle());
 		MemoryUtil.memPutInt(ptr + MODEL_HANDLE_OFFSET, model.handle());
 		MemoryUtil.memPutInt(ptr + DRAW_HANDLE_OFFSET, draw.handle());
 
-		MemoryUtil.memPutAddress(ptr + INSTANCE_SIZE_OFFSET, instanceStride * instanceCount);
+		MemoryUtil.memPutAddress(ptr + INSTANCE_SIZE_OFFSET, pageFile.storage.byteCapacity());
 		MemoryUtil.memPutAddress(ptr + TARGET_SIZE_OFFSET, INT_SIZE * instanceCount);
-		MemoryUtil.memPutAddress(ptr + MODEL_INDEX_SIZE_OFFSET, INT_SIZE * instanceCount);
+		MemoryUtil.memPutAddress(ptr + MODEL_INDEX_SIZE_OFFSET, pageFile.pageTable.byteCapacity());
 		MemoryUtil.memPutAddress(ptr + MODEL_SIZE_OFFSET, MODEL_STRIDE * modelCount);
 		MemoryUtil.memPutAddress(ptr + DRAW_SIZE_OFFSET, DRAW_COMMAND_STRIDE * drawCount);
 	}
@@ -124,9 +121,8 @@ public class IndirectBuffers {
 	public void delete() {
 		multiBindBlock.free();
 
-		instance.delete();
+		pageFile.delete();
 		target.delete();
-		modelIndex.delete();
 		model.delete();
 		draw.delete();
 	}

common/src/backend/java/dev/engine_room/flywheel/backend/engine/indirect/IndirectCullingGroup.java

@@ -74,8 +74,8 @@ public class IndirectCullingGroup<I extends Instance> {
 				continue;
 			}
 
-			instancer.modelIndex = modelIndex;
-			instancer.baseInstance = instanceCountThisFrame;
+			instancer.modelIndex(modelIndex);
+			instancer.baseInstance(instanceCountThisFrame);
 			instanceCountThisFrame += instanceCount;
 
 			modelIndex++;
@@ -96,6 +96,8 @@ public class IndirectCullingGroup<I extends Instance> {
 		// Upload only instances that have changed.
 		uploadInstances(stagingBuffer);
 
+		buffers.pageFile.uploadTable(stagingBuffer);
+
 		// We need to upload the models every frame to reset the instance count.
 		uploadModels(stagingBuffer);
 
@@ -118,7 +120,7 @@ public class IndirectCullingGroup<I extends Instance> {
 		cullProgram.bind();
 
 		buffers.bindForCompute();
-		glDispatchCompute(GlCompat.getComputeGroupCount(instanceCountThisFrame), 1, 1);
+		glDispatchCompute(buffers.pageFile.capacity(), 1, 1);
 	}
 
 	public void dispatchApply() {
@@ -171,7 +173,9 @@ public class IndirectCullingGroup<I extends Instance> {
 	}
 
 	public void add(IndirectInstancer<I> instancer, InstancerKey<I> key, MeshPool meshPool) {
-		instancer.modelIndex = instancers.size();
+		instancer.pageFile = buffers.pageFile.createPage();
+		instancer.modelIndex(instancers.size());
+
 		instancers.add(instancer);
 
         List<Model.ConfiguredMesh> meshes = key.model()
@@ -242,12 +246,7 @@ public class IndirectCullingGroup<I extends Instance> {
 
 	private void uploadInstances(StagingBuffer stagingBuffer) {
 		for (var instancer : instancers) {
-			instancer.uploadInstances(stagingBuffer, buffers.instance.handle());
-		}
-
-		for (var instancer : instancers) {
-			instancer.uploadModelIndices(stagingBuffer, buffers.modelIndex.handle());
-			instancer.resetChanged();
+			instancer.uploadInstances(stagingBuffer, buffers.pageFile.storage.handle());
 		}
 	}
 

common/src/backend/java/dev/engine_room/flywheel/backend/engine/indirect/IndirectDraw.java

@@ -72,9 +72,9 @@ public class IndirectDraw {
 		MemoryUtil.memPutInt(ptr + 4, 0); // instanceCount - to be set by the apply shader
 		MemoryUtil.memPutInt(ptr + 8, mesh.firstIndex()); // firstIndex
 		MemoryUtil.memPutInt(ptr + 12, mesh.baseVertex()); // baseVertex
-		MemoryUtil.memPutInt(ptr + 16, instancer.baseInstance); // baseInstance
+		MemoryUtil.memPutInt(ptr + 16, instancer.baseInstance()); // baseInstance
 
-		MemoryUtil.memPutInt(ptr + 20, instancer.modelIndex); // modelIndex
+		MemoryUtil.memPutInt(ptr + 20, instancer.modelIndex()); // modelIndex
 
 		MemoryUtil.memPutInt(ptr + 24, instancer.environment.matrixIndex()); // matrixIndex
 
@@ -89,9 +89,9 @@ public class IndirectDraw {
 		MemoryUtil.memPutInt(ptr + 4, 1); // instanceCount - only drawing one instance
 		MemoryUtil.memPutInt(ptr + 8, mesh.firstIndex()); // firstIndex
 		MemoryUtil.memPutInt(ptr + 12, mesh.baseVertex()); // baseVertex
-		MemoryUtil.memPutInt(ptr + 16, instancer.baseInstance + instanceIndex); // baseInstance
+		MemoryUtil.memPutInt(ptr + 16, instancer.baseInstance() + instanceIndex); // baseInstance
 
-		MemoryUtil.memPutInt(ptr + 20, instancer.modelIndex); // modelIndex
+		MemoryUtil.memPutInt(ptr + 20, instancer.modelIndex()); // modelIndex
 
 		MemoryUtil.memPutInt(ptr + 24, instancer.environment.matrixIndex()); // matrixIndex
 

common/src/backend/java/dev/engine_room/flywheel/backend/engine/indirect/IndirectInstancer.java

@@ -12,6 +12,7 @@ import dev.engine_room.flywheel.api.instance.InstanceWriter;
 import dev.engine_room.flywheel.api.model.Model;
 import dev.engine_room.flywheel.backend.engine.AbstractInstancer;
 import dev.engine_room.flywheel.backend.engine.embed.Environment;
+import dev.engine_room.flywheel.backend.util.AtomicBitSet;
 import dev.engine_room.flywheel.lib.math.MoreMath;
 
 public class IndirectInstancer<I extends Instance> extends AbstractInstancer<I> {
@@ -20,11 +21,12 @@ public class IndirectInstancer<I extends Instance> extends AbstractInstancer<I>
 	private final List<IndirectDraw> associatedDraws = new ArrayList<>();
 	private final Vector4fc boundingSphere;
 
-	public int modelIndex = -1;
-	public int baseInstance = -1;
-	private int lastModelIndex = -1;
-	private int lastBaseInstance = -1;
-	private int lastInstanceCount = -1;
+	private final AtomicBitSet changedPages = new AtomicBitSet();
+
+	public InstancePager.Allocation pageFile;
+
+	private int modelIndex = -1;
+	private int baseInstance = -1;
 
 	public IndirectInstancer(InstanceType<I> type, Environment environment, Model model) {
 		super(type, environment);
@@ -34,6 +36,29 @@ public class IndirectInstancer<I extends Instance> extends AbstractInstancer<I>
 		boundingSphere = model.boundingSphere();
 	}
 
+	@Override
+	public void notifyDirty(int index) {
+		if (index < 0 || index >= instanceCount()) {
+			return;
+		}
+		changed.set(index);
+		changedPages.set(pageFile.object2Page(index));
+	}
+
+	@Override
+	protected void setRangeChanged(int start, int end) {
+		super.setRangeChanged(start, end);
+
+		changedPages.set(pageFile.object2Page(start), pageFile.object2Page(end));
+	}
+
+	@Override
+	protected void clearChangedRange(int start, int end) {
+		super.clearChangedRange(start, end);
+
+		// changedPages.clear(pageFile.object2Page(start), pageFile);
+	}
+
 	public void addDraw(IndirectDraw draw) {
 		associatedDraws.add(draw);
 	}
@@ -44,6 +69,8 @@ public class IndirectInstancer<I extends Instance> extends AbstractInstancer<I>
 
 	public void update() {
 		removeDeletedInstances();
+
+		pageFile.activeCount(instanceCount());
 	}
 
 	public void writeModel(long ptr) {
@@ -57,71 +84,38 @@ public class IndirectInstancer<I extends Instance> extends AbstractInstancer<I>
 	}
 
 	public void uploadInstances(StagingBuffer stagingBuffer, int instanceVbo) {
-		long baseByte = baseInstance * instanceStride;
+		int numPages = pageFile.pageCount();
 
-		if (baseInstance != lastBaseInstance) {
-			uploadAllInstances(stagingBuffer, baseByte, instanceVbo);
-		} else {
-			uploadChangedInstances(stagingBuffer, baseByte, instanceVbo);
-		}
+		var instanceCount = instances.size();
+
+		for (int page = 0; page < numPages; page++) {
+			page = changedPages.nextSetBit(0);
+
+			if (page == -1) {
+				break;
 			}
 
-	public void uploadModelIndices(StagingBuffer stagingBuffer, int modelIndexVbo) {
-		long modelIndexBaseByte = baseInstance * IndirectBuffers.INT_SIZE;
+			int startObject = pageFile.page2Object(page);
 
-		if (baseInstance != lastBaseInstance || modelIndex != lastModelIndex || instances.size() > lastInstanceCount) {
-			uploadAllModelIndices(stagingBuffer, modelIndexBaseByte, modelIndexVbo);
-		}
+			if (startObject >= instanceCount) {
+				break;
+			}
+
+			int endObject = Math.min(instanceCount, pageFile.page2Object(page + 1) - 1);
+
+			long baseByte = pageFile.page2ByteOffset(page);
+			long size = (endObject - startObject) * instanceStride;
+
+			stagingBuffer.enqueueCopy(size, instanceVbo, baseByte, ptr -> {
+				for (int i = startObject; i < endObject; i++) {
+					writer.write(ptr, instances.get(i));
+					ptr += instanceStride;
+				}
+			});
 		}
 
-	public void resetChanged() {
-		lastModelIndex = modelIndex;
-		lastBaseInstance = baseInstance;
-		lastInstanceCount = instances.size();
 		changed.clear();
-	}
-
-	private void uploadChangedInstances(StagingBuffer stagingBuffer, long baseByte, int instanceVbo) {
-		changed.forEachSetSpan((startInclusive, endInclusive) -> {
-			// Generally we're good about ensuring we don't have changed bits set out of bounds, but check just in case
-			if (startInclusive >= instances.size()) {
-				return;
-			}
-			int actualEnd = Math.min(endInclusive, instances.size() - 1);
-
-			int instanceCount = actualEnd - startInclusive + 1;
-			long totalSize = instanceCount * instanceStride;
-
-			stagingBuffer.enqueueCopy(totalSize, instanceVbo, baseByte + startInclusive * instanceStride, ptr -> {
-				for (int i = startInclusive; i <= actualEnd; i++) {
-					var instance = instances.get(i);
-					writer.write(ptr, instance);
-					ptr += instanceStride;
-				}
-			});
-		});
-	}
-
-	private void uploadAllInstances(StagingBuffer stagingBuffer, long baseByte, int instanceVbo) {
-		long totalSize = instances.size() * instanceStride;
-
-		stagingBuffer.enqueueCopy(totalSize, instanceVbo, baseByte, ptr -> {
-			for (I instance : instances) {
-				writer.write(ptr, instance);
-				ptr += instanceStride;
-			}
-		});
-	}
-
-	private void uploadAllModelIndices(StagingBuffer stagingBuffer, long modelIndexBaseByte, int modelIndexVbo) {
-		long modelIndexTotalSize = instances.size() * IndirectBuffers.INT_SIZE;
-
-		stagingBuffer.enqueueCopy(modelIndexTotalSize, modelIndexVbo, modelIndexBaseByte, ptr -> {
-			for (int i = 0; i < instances.size(); i++) {
-				MemoryUtil.memPutInt(ptr, modelIndex);
-				ptr += IndirectBuffers.INT_SIZE;
-			}
-		});
+		changedPages.clear();
 	}
 
 	@Override
@@ -130,4 +124,21 @@ public class IndirectInstancer<I extends Instance> extends AbstractInstancer<I>
 			draw.delete();
 		}
 	}
+
+	public void modelIndex(int modelIndex) {
+		this.modelIndex = modelIndex;
+		pageFile.modelIndex(modelIndex);
+	}
+
+	public int modelIndex() {
+		return modelIndex;
+	}
+
+	public void baseInstance(int baseInstance) {
+		this.baseInstance = baseInstance;
+	}
+
+	public int baseInstance() {
+		return baseInstance;
+	}
 }

common/src/backend/java/dev/engine_room/flywheel/backend/engine/indirect/InstancePager.java

@@ -0,0 +1,158 @@
+package dev.engine_room.flywheel.backend.engine.indirect;
+
+import java.util.ArrayList;
+import java.util.Arrays;
+import java.util.List;
+
+import org.jetbrains.annotations.UnknownNullability;
+import org.lwjgl.system.MemoryUtil;
+
+import dev.engine_room.flywheel.backend.engine.AbstractArena;
+import dev.engine_room.flywheel.lib.memory.MemoryBlock;
+
+public class InstancePager extends AbstractArena {
+	// 32 objects per page. Allows for convenient bitsets on the gpu.
+	public static final int DEFAULT_PAGE_SIZE_OBJECTS = 5;
+	public static final int INITIAL_PAGES_ALLOCATED = 4;
+
+	private final int log2PageSize;
+	/**
+	 * The number of objects in a page.
+	 */
+	private final int pageSize;
+
+	private final long objectSizeBytes;
+
+	@UnknownNullability
+	private MemoryBlock pageData;
+
+	private final int pageMask;
+	public final ResizableStorageArray storage;
+	public final ResizableStorageArray pageTable;
+
+	private final List<Allocation> allocations = new ArrayList<>();
+
+	public InstancePager(long objectSizeBytes) {
+		this(DEFAULT_PAGE_SIZE_OBJECTS, objectSizeBytes);
+	}
+
+	public InstancePager(int log2PageSize, long objectSizeBytes) {
+		super((1L << log2PageSize) * objectSizeBytes);
+		this.log2PageSize = log2PageSize;
+		this.pageSize = 1 << log2PageSize;
+		this.pageMask = pageSize - 1;
+		this.objectSizeBytes = objectSizeBytes;
+
+		this.storage = new ResizableStorageArray(this.elementSizeBytes);
+		this.pageTable = new ResizableStorageArray(Integer.BYTES);
+	}
+
+	public Allocation createPage() {
+		var out = new Allocation();
+		allocations.add(out);
+		return out;
+	}
+
+	@Override
+	public long byteCapacity() {
+		return storage.byteCapacity();
+	}
+
+	@Override
+	protected void resize() {
+		if (pageData == null) {
+			pageData = MemoryBlock.malloc(INITIAL_PAGES_ALLOCATED * Integer.BYTES);
+			storage.ensureCapacity(INITIAL_PAGES_ALLOCATED);
+			pageTable.ensureCapacity(INITIAL_PAGES_ALLOCATED);
+		} else {
+			pageData = pageData.realloc(pageData.size() * 2);
+			storage.ensureCapacity(storage.capacity() * 2);
+			pageTable.ensureCapacity(pageTable.capacity() * 2);
+		}
+	}
+
+	public void uploadTable(StagingBuffer stagingBuffer) {
+		for (Allocation allocation : allocations) {
+			allocation.updatePageTable();
+		}
+		stagingBuffer.enqueueCopy(pageData.ptr(), pageData.size(), pageTable.handle(), 0);
+	}
+
+	public void delete() {
+		storage.delete();
+		pageTable.delete();
+		pageData.free();
+	}
+
+	public class Allocation {
+		public int[] pages = new int[0];
+
+		private int modelIndex = -1;
+
+		public void modelIndex(int modelIndex) {
+			if (this.modelIndex != modelIndex) {
+				this.modelIndex = modelIndex;
+			}
+		}
+
+		private void updatePageTable() {
+			var ptr = pageData.ptr();
+
+			int fullPage = (modelIndex & 0x3FFFFF) | 0x8000000;
+
+			for (int page : pages) {
+				MemoryUtil.memPutInt(ptr + page * Integer.BYTES, fullPage);
+			}
+		}
+
+		public void activeCount(int objectCount) {
+			var neededPages = object2Page((objectCount + pageMask));
+
+			var oldLength = pages.length;
+
+			if (oldLength > neededPages) {
+				shrink(oldLength, neededPages);
+			} else if (oldLength < neededPages) {
+				grow(neededPages, oldLength);
+			}
+		}
+
+		private void grow(int neededPages, int oldLength) {
+			pages = Arrays.copyOf(pages, neededPages);
+
+			for (int i = oldLength; i < neededPages; i++) {
+				pages[i] = InstancePager.this.alloc();
+			}
+		}
+
+		private void shrink(int oldLength, int neededPages) {
+			for (int i = oldLength - 1; i > neededPages; i--) {
+				var page = pages[i];
+				InstancePager.this.free(page);
+				MemoryUtil.memPutInt(pageData.ptr() + page * Integer.BYTES, 0);
+			}
+
+			pages = Arrays.copyOf(pages, neededPages);
+		}
+
+		public int capacity() {
+			return pages.length << log2PageSize;
+		}
+
+		public int pageCount() {
+			return pages.length;
+		}
+
+		public int object2Page(int objectIndex) {
+			return objectIndex >> log2PageSize;
+		}
+
+		public int page2Object(int pageIndex) {
+			return pageIndex << log2PageSize;
+		}
+
+		public long page2ByteOffset(int page) {
+			return InstancePager.this.byteOffsetOf(pages[page]);
+		}
+	}
+}

common/src/backend/resources/assets/flywheel/flywheel/internal/indirect/cull.glsl

@@ -4,14 +4,19 @@
 #include "flywheel:util/matrix.glsl"
 #include "flywheel:internal/indirect/matrices.glsl"
 
-layout(local_size_x = _FLW_SUBGROUP_SIZE) in;
+layout(local_size_x = 32) in;
 
 layout(std430, binding = _FLW_TARGET_BUFFER_BINDING) restrict writeonly buffer TargetBuffer {
     uint _flw_instanceIndices[];
 };
 
+// High 6 bits for the number of instances in the page.
+const uint _FLW_PAGE_COUNT_OFFSET = 25u;
+// Bottom 24 bits for the model index.
+const uint _FLW_MODEL_INDEX_MASK = 0x3FFFFFF;
+
 layout(std430, binding = _FLW_MODEL_INDEX_BUFFER_BINDING) restrict readonly buffer ModelIndexBuffer {
-    uint _flw_modelIndices[];
+    uint _flw_pageTable[];
 };
 
 layout(std430, binding = _FLW_MODEL_BUFFER_BINDING) restrict buffer ModelBuffer {
@@ -55,13 +60,23 @@ bool _flw_isVisible(uint instanceIndex, uint modelIndex) {
 }
 
 void main() {
-    uint instanceIndex = gl_GlobalInvocationID.x;
+    uint pageIndex = gl_WorkGroupID.x;
 
-    if (instanceIndex >= _flw_modelIndices.length()) {
+    if (pageIndex >= _flw_pageTable.length()) {
         return;
     }
 
-    uint modelIndex = _flw_modelIndices[instanceIndex];
+    uint packedModelIndexAndCount = _flw_pageTable[pageIndex];
+
+    uint pageInstanceCount = packedModelIndexAndCount >> _FLW_PAGE_COUNT_OFFSET;
+
+    if (gl_LocalInvocationID.x >= pageInstanceCount) {
+        return;
+    }
+
+    uint instanceIndex = gl_GlobalInvocationID.x;
+
+    uint modelIndex = packedModelIndexAndCount & _FLW_MODEL_INDEX_MASK;
 
     if (_flw_isVisible(instanceIndex, modelIndex)) {
         uint localIndex = atomicAdd(_flw_models[modelIndex].instanceCount, 1);