Poking and prodding

- Invert image size on CPU to avoid divisions on GPU
- Increase depth reduce group size to 16x16
- Early-out in uploadInstances based on changed cardinality
  - Much faster to calculate cardinality than it is to clear an
    AtomicBitSet, so the check is worth it
- Upload scatter list directly in the staging buffer if there's room

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

@@ -35,7 +35,6 @@ public static void setup(Material material) {
 		setupBackfaceCulling(material.backfaceCulling());
 		setupPolygonOffset(material.polygonOffset());
 		setupDepthTest(material.depthTest());
-//		setupDepthTest(DepthTest.OFF);
 		setupTransparency(material.transparency());
 		setupWriteMask(material.writeMask());
 	}

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

@@ -11,6 +11,8 @@
 import net.minecraft.client.Minecraft;
 
 public class DepthPyramid {
+	private static final int GROUP_SIZE = 16;
+
 	private final GlProgram depthReduceProgram;
 
 	public int pyramidTextureId = -1;
@@ -50,10 +52,10 @@ public void generate() {
 
 			GL46.glBindImageTexture(0, pyramidTextureId, i, false, 0, GL32.GL_WRITE_ONLY, GL32.GL_R32F);
 
-			depthReduceProgram.setVec2("imageSize", mipWidth, mipHeight);
+			depthReduceProgram.setVec2("oneOverImageSize", 1f / (float) mipWidth, 1f / (float) mipHeight);
 			depthReduceProgram.setInt("lod", Math.max(0, i - 1));
 
-			GL46.glDispatchCompute(MoreMath.ceilingDiv(mipWidth, 8), MoreMath.ceilingDiv(mipHeight, 8), 1);
+			GL46.glDispatchCompute(MoreMath.ceilingDiv(mipWidth, GROUP_SIZE), MoreMath.ceilingDiv(mipHeight, GROUP_SIZE), 1);
 
 			GL46.glMemoryBarrier(GL46.GL_TEXTURE_FETCH_BARRIER_BIT);
 		}

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

@@ -82,6 +82,11 @@ public void writeModel(long ptr) {
 	}
 
 	public void uploadInstances(StagingBuffer stagingBuffer, int instanceVbo) {
+		if (changedPages.cardinality() == 0) {
+			// Early return because checking the cardinality is faster than clearing.
+			return;
+		}
+
 		int numPages = mapping.pageCount();
 
 		var instanceCount = instances.size();

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

@@ -23,6 +23,8 @@ public class StagingBuffer {
 	private static final int STORAGE_FLAGS = GL45C.GL_MAP_PERSISTENT_BIT | GL45C.GL_MAP_WRITE_BIT | GL45C.GL_CLIENT_STORAGE_BIT;
 	private static final int MAP_FLAGS = GL45C.GL_MAP_PERSISTENT_BIT | GL45C.GL_MAP_WRITE_BIT | GL45C.GL_MAP_FLUSH_EXPLICIT_BIT | GL45C.GL_MAP_INVALIDATE_BUFFER_BIT;
 
+	private static final int SSBO_ALIGNMENT = GL45.glGetInteger(GL45.GL_SHADER_STORAGE_BUFFER_OFFSET_ALIGNMENT);
+
 	private final int vbo;
 	private final long map;
 	private final long capacity;
@@ -254,7 +256,6 @@ private void dispatchComputeCopies() {
 		scatterProgram.bind();
 
 		// These bindings don't change between dstVbos.
-		GL45.glBindBufferBase(GL45C.GL_SHADER_STORAGE_BUFFER, 0, scatterBuffer.handle());
 		GL45.glBindBufferBase(GL45C.GL_SHADER_STORAGE_BUFFER, 1, vbo);
 
 		int dstVbo;
@@ -276,7 +277,24 @@ private void dispatchComputeCopies() {
 	}
 
 	private void dispatchScatter(int dstVbo) {
-		scatterBuffer.upload(scatterList.ptr(), scatterList.usedBytes());
+		var scatterSize = scatterList.usedBytes();
+
+		long alignedPos = pos + SSBO_ALIGNMENT - 1 - (pos + SSBO_ALIGNMENT - 1) % SSBO_ALIGNMENT;
+
+		long remaining = capacity - alignedPos;
+		if (scatterSize <= remaining && scatterSize <= totalAvailable) {
+			MemoryUtil.memCopy(scatterList.ptr(), map + alignedPos, scatterSize);
+			GL45.glBindBufferRange(GL45C.GL_SHADER_STORAGE_BUFFER, 0, vbo, alignedPos, scatterSize);
+
+			long alignmentCost = alignedPos - pos;
+
+			usedCapacity += scatterSize + alignmentCost;
+			totalAvailable -= scatterSize + alignmentCost;
+			pos += scatterSize + alignmentCost;
+		} else {
+			scatterBuffer.upload(scatterList.ptr(), scatterSize);
+			GL45.glBindBufferBase(GL45C.GL_SHADER_STORAGE_BUFFER, 0, scatterBuffer.handle());
+		}
 
 		GL45.glBindBufferBase(GL45C.GL_SHADER_STORAGE_BUFFER, 2, dstVbo);
 

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

@@ -1,9 +1,9 @@
-layout(local_size_x = 8, local_size_y = 8) in;
+layout(local_size_x = 16, local_size_y = 16) in;
 
 layout(binding = 0, r32f) uniform writeonly image2D outImage;
 layout(binding = 1) uniform sampler2D inImage;
 
-uniform vec2 imageSize;
+uniform vec2 oneOverImageSize;
 uniform int lod;
 
 uniform int useMin = 0;
@@ -13,7 +13,7 @@ void main() {
 
     // Map the output texel to an input texel. Properly do the division because generating mip0 maps from the actual
     // full resolution depth buffer and the aspect ratio may be different from our Po2 pyramid.
-    ivec2 samplePos = ivec2(floor(vec2(pos) * vec2(textureSize(inImage, lod)) / imageSize));
+    ivec2 samplePos = ivec2(floor(vec2(pos) * vec2(textureSize(inImage, lod)) * oneOverImageSize));
 
     float depth01 = texelFetchOffset(inImage, samplePos, lod, ivec2(0, 1)).r;
     float depth11 = texelFetchOffset(inImage, samplePos, lod, ivec2(1, 1)).r;