More smarter less upload

- Only write draw calls when they change.
- Update baseInstance in the apply shader.
- Simplify object upload logic.

src/main/java/com/jozufozu/flywheel/backend/engine/AbstractInstancer.java

@@ -15,7 +15,6 @@ public abstract class AbstractInstancer<I extends Instance> implements Instancer
 	protected final ArrayList<I> instances = new ArrayList<>();
 	protected final ArrayList<InstanceHandleImpl> handles = new ArrayList<>();
 
-	// TODO: atomic bitset?
 	protected final BitSet changed = new BitSet();
 	protected final BitSet deleted = new BitSet();
 
@@ -45,6 +44,7 @@ public abstract class AbstractInstancer<I extends Instance> implements Instancer
 		if (index < 0 || index >= getInstanceCount()) {
 			return;
 		}
+		// TODO: Atomic bitset. Synchronizing here blocks the task executor and causes massive overhead.
 		synchronized (lock) {
 			changed.set(index);
 		}

src/main/java/com/jozufozu/flywheel/backend/engine/indirect/IndirectCullingGroup.java

@@ -45,7 +45,7 @@ public class IndirectCullingGroup<I extends Instance> {
 	private final List<IndirectDraw> indirectDraws = new ArrayList<>();
 	private final Map<RenderStage, List<MultiDraw>> multiDraws = new EnumMap<>(RenderStage.class);
 	private boolean needsDrawBarrier;
-	private boolean needsSortDraws;
+	private boolean hasNewDraws;
 	private int instanceCountThisFrame;
 
 	IndirectCullingGroup(InstanceType<I> instanceType) {
@@ -72,15 +72,23 @@ public class IndirectCullingGroup<I extends Instance> {
 
 		buffers.updateCounts(instanceCountThisFrame, indirectModels.size(), indirectDraws.size());
 
-		if (needsSortDraws) {
+		// Must flush the mesh pool first so everything else has the right baseVertex and baseIndex.
-			sortDraws();
-			needsSortDraws = false;
-		}
-
 		meshPool.flush(stagingBuffer);
+
+		// Upload only objects that have changed.
 		uploadObjects(stagingBuffer);
+
+		// We need to upload the models every frame to reset the instance count.
 		uploadModels(stagingBuffer);
-		uploadDraws(stagingBuffer);
+
+		if (hasNewDraws) {
+			sortDraws();
+			// Draws, however, only need to be updated when we get new ones.
+			// The instanceCount and baseInstance will be updated by the applyProgram,
+			// and all other fields are constant to the lifetime of the draw.
+			uploadDraws(stagingBuffer);
+			hasNewDraws = false;
+		}
 	}
 
 	public void dispatchCull() {
@@ -161,7 +169,7 @@ public class IndirectCullingGroup<I extends Instance> {
 			indirectDraws.add(new IndirectDraw(indirectModel, entry.getKey(), bufferedMesh, stage));
 		}
 
-		needsSortDraws = true;
+		hasNewDraws = true;
 	}
 
 	public void submit(RenderStage stage) {
@@ -193,9 +201,9 @@ public class IndirectCullingGroup<I extends Instance> {
 
 	private void uploadObjects(StagingBuffer stagingBuffer) {
 		long pos = 0;
-		for (IndirectModel batch : indirectModels) {
+		for (IndirectModel model : indirectModels) {
-			var instanceCount = batch.instancer.getInstanceCount();
+			var instanceCount = model.instancer.getInstanceCount();
-			batch.writeObjects(stagingBuffer, pos, buffers.object.handle());
+			model.writeObjects(stagingBuffer, pos, buffers.object.handle());
 
 			pos += instanceCount * objectStride;
 		}
@@ -216,15 +224,15 @@ public class IndirectCullingGroup<I extends Instance> {
 	}
 
 	private void writeModels(long writePtr) {
-		for (var batch : indirectModels) {
+		for (var model : indirectModels) {
-			batch.write(writePtr);
+			model.write(writePtr);
 			writePtr += IndirectBuffers.MODEL_STRIDE;
 		}
 	}
 
 	private void writeCommands(long writePtr) {
-		for (var batch : indirectDraws) {
+		for (var draw : indirectDraws) {
-			batch.write(writePtr);
+			draw.write(writePtr);
 			writePtr += IndirectBuffers.DRAW_COMMAND_STRIDE;
 		}
 	}

src/main/java/com/jozufozu/flywheel/backend/engine/indirect/IndirectDraw.java

@@ -47,9 +47,9 @@ 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, model.baseInstance()); // baseInstance
+		MemoryUtil.memPutInt(ptr + 16, 0); // baseInstance - to be set by the apply shader
 
-		MemoryUtil.memPutInt(ptr + 20, model.index); // modelIndex
+		MemoryUtil.memPutInt(ptr + 20, model.index); // modelIndex - never changes
 
 		MemoryUtil.memPutInt(ptr + 24, materialVertexIndex); // materialVertexIndex
 		MemoryUtil.memPutInt(ptr + 28, materialFragmentIndex); // materialFragmentIndex

src/main/java/com/jozufozu/flywheel/backend/engine/indirect/IndirectInstancer.java

@@ -12,6 +12,8 @@ public class IndirectInstancer<I extends Instance> extends AbstractInstancer<I>
 	private final InstanceWriter<I> writer;
 	private int modelIndex;
 
+	private long lastStartPos = -1;
+
 	public IndirectInstancer(InstanceType<I> type) {
 		super(type);
 		long instanceStride = type.getLayout()
@@ -24,21 +26,38 @@ public class IndirectInstancer<I extends Instance> extends AbstractInstancer<I>
 		removeDeletedInstances();
 	}
 
-	public void writeChanged(StagingBuffer stagingBuffer, long start, int dstVbo) {
+	public void write(StagingBuffer stagingBuffer, long startPos, int dstVbo) {
+		if (shouldWriteAll(startPos)) {
+			writeAll(stagingBuffer, startPos, dstVbo);
+		} else {
+			writeChanged(stagingBuffer, startPos, dstVbo);
+		}
+
+		changed.clear();
+		lastStartPos = startPos;
+	}
+
+	private boolean shouldWriteAll(long startPos) {
+		// If enough of the buffer has changed, write the whole thing to avoid the overhead of a bunch of small writes.
+		return startPos != lastStartPos || moreThanTwoThirdsChanged();
+	}
+
+	private boolean moreThanTwoThirdsChanged() {
+		return (changed.cardinality() * 3) > (instances.size() * 2);
+	}
+
+	private void writeChanged(StagingBuffer stagingBuffer, long start, int dstVbo) {
 		int count = instances.size();
 		for (int i = changed.nextSetBit(0); i >= 0 && i < count; i = changed.nextSetBit(i + 1)) {
 			var instance = instances.get(i);
 			stagingBuffer.enqueueCopy(objectStride, dstVbo, start + i * objectStride, ptr -> writeOne(ptr, instance));
 		}
-		changed.clear();
 	}
 
-	public void writeAll(StagingBuffer stagingBuffer, long start, int dstVbo) {
+	private void writeAll(StagingBuffer stagingBuffer, long start, int dstVbo) {
 		long totalSize = objectStride * instances.size();
 
 		stagingBuffer.enqueueCopy(totalSize, dstVbo, start, this::writeAll);
-
-		changed.clear();
 	}
 
 	private void writeAll(long ptr) {
@@ -49,9 +68,7 @@ public class IndirectInstancer<I extends Instance> extends AbstractInstancer<I>
 	}
 
 	private void writeOne(long ptr, I instance) {
-		// write modelID
 		MemoryUtil.memPutInt(ptr, modelIndex);
-		// write object
 		writer.write(ptr + IndirectBuffers.INT_SIZE, instance);
 	}
 

src/main/java/com/jozufozu/flywheel/backend/engine/indirect/IndirectModel.java

@@ -9,7 +9,6 @@ public class IndirectModel {
 	private final Vector4fc boundingSphere;
 
 	private int baseInstance = -1;
-	private boolean needsFullWrite = true;
 
 	public IndirectModel(IndirectInstancer<?> instancer, int index, Vector4fc boundingSphere) {
 		this.instancer = instancer;
@@ -23,20 +22,11 @@ public class IndirectModel {
 
 	public void prepare(int baseInstance) {
 		instancer.update();
-		if (baseInstance == this.baseInstance) {
-			needsFullWrite = false;
-			return;
-		}
 		this.baseInstance = baseInstance;
-		needsFullWrite = true;
 	}
 
 	public void writeObjects(StagingBuffer stagingBuffer, long start, int dstVbo) {
-		if (needsFullWrite) {
+		instancer.write(stagingBuffer, start, dstVbo);
-			instancer.writeAll(stagingBuffer, start, dstVbo);
-		} else {
-			instancer.writeChanged(stagingBuffer, start, dstVbo);
-		}
 	}
 
 	public void write(long ptr) {

src/main/java/com/jozufozu/flywheel/backend/engine/indirect/StagingBuffer.java

@@ -14,7 +14,8 @@ import com.jozufozu.flywheel.lib.memory.MemoryBlock;
 import it.unimi.dsi.fastutil.PriorityQueue;
 import it.unimi.dsi.fastutil.objects.ObjectArrayFIFOQueue;
 
-// https://github.com/CaffeineMC/sodium-fabric/blob/dev/src/main/java/me/jellysquid/mods/sodium/client/gl/arena/staging/MappedStagingBuffer.java
+// Used https://github.com/CaffeineMC/sodium-fabric/blob/dev/src/main/java/me/jellysquid/mods/sodium/client/gl/arena/staging/MappedStagingBuffer.java
+// as a reference for implementation. Modified to be less safe and to allow for writing directly into the staging buffer.
 public class StagingBuffer {
 	private static final long DEFAULT_CAPACITY = 1024 * 1024 * 16;
 	private static final int STORAGE_FLAGS = GL45C.GL_MAP_PERSISTENT_BIT | GL45C.GL_MAP_WRITE_BIT | GL45C.GL_CLIENT_STORAGE_BIT;

src/main/resources/assets/flywheel/flywheel/internal/indirect/apply.glsl

@@ -21,6 +21,6 @@ void main() {
     }
 
     uint modelIndex = drawCommands[drawIndex].modelIndex;
-    uint instanceCount = models[modelIndex].instanceCount;
+    drawCommands[drawIndex].instanceCount = models[modelIndex].instanceCount;
-    drawCommands[drawIndex].instanceCount = instanceCount;
+    drawCommands[drawIndex].baseInstance = models[modelIndex].baseInstance;
 }