Cache deposits

- Make the scatter shader more cache friendly by doing one copy per
  workgroup and making copies up to 64 uints long

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

@@ -12,10 +12,11 @@ public class ScatterList {
 	private long usedBytes;
 
 	public ScatterList() {
-		this(64);
+		// Should be the same as the local_size_x of the scatter shader
+		this(64 * 4);
 	}
 
-	public ScatterList(long maxBytesPerScatter) {
+	private ScatterList(long maxBytesPerScatter) {
 		if ((maxBytesPerScatter & 0b1111111100L) != maxBytesPerScatter) {
 			throw new IllegalArgumentException("Max bytes per scatter must be a multiple of 4 and less than 1024");
 		}

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

@@ -8,7 +8,6 @@ import org.lwjgl.opengl.GL45C;
 import org.lwjgl.system.MemoryUtil;
 
 import com.jozufozu.flywheel.backend.compile.IndirectPrograms;
-import com.jozufozu.flywheel.backend.gl.GlCompat;
 import com.jozufozu.flywheel.backend.gl.GlFence;
 import com.jozufozu.flywheel.backend.gl.buffer.GlBuffer;
 import com.jozufozu.flywheel.backend.gl.shader.GlProgram;
@@ -282,7 +281,7 @@ public class StagingBuffer {
 
 		GL45.glBindBufferBase(GL45C.GL_SHADER_STORAGE_BUFFER, 2, dstVbo);
 
-		GL45.glDispatchCompute(GlCompat.getComputeGroupCount(scatterList.copyCount()), 1, 1);
+		GL45.glDispatchCompute(scatterList.copyCount(), 1, 1);
 
 		scatterList.reset();
 	}

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

@@ -1,4 +1,4 @@
-layout(local_size_x = _FLW_SUBGROUP_SIZE) in;
+layout(local_size_x = 64) in;
 
 const uint SRC_OFFSET_MASK = 0xFFFFFF;
 
@@ -24,19 +24,27 @@ layout(std430, binding = 2) restrict writeonly buffer Dst {
 };
 
 void main() {
-    uint copy = gl_GlobalInvocationID.x;
+    // Each work group is responsible for one of the copies in the buffer.
+    // We dispatch exactly as many work groups as there are copies, so no need to check bounds.
+    uint copy = gl_WorkGroupID.x;
 
-    if (copy >= copies.length()) {
-        return;
-    }
+    // Each invocation in the work group is responsible for one uint in the copy.
+    uint i = gl_LocalInvocationID.x;
 
+    // Unpack the copy.
     uint sizeAndSrcOffset = copies[copy].sizeAndSrcOffset;
+    uint dstOffset = copies[copy].dstOffset;
     uint srcOffset = sizeAndSrcOffset & SRC_OFFSET_MASK;
     uint size = sizeAndSrcOffset >> 24;
 
-    uint dstOffset = copies[copy].dstOffset;
+    // Fetch the uint to copy before exiting to make instruction reordering happy.
+    // With 20mb going through a 24mb staging buffer, this made a 1ms/frame difference.
+    // Should properly test with nsight at some point.
+    uint toCopy = src[srcOffset + i];
 
-    for (uint i = 0; i < size; i++) {
-        dst[dstOffset + i] = src[srcOffset + i];
+    if (i >= size) {
+        return;
     }
+
+    dst[dstOffset + i] = toCopy;
 }