Seeing blue

- Optimize read visibility by having each invocation read a 2x2 area and
  coalescing atomicOrs when all 4 texels are equal
- Also use the fancy remap function for better texture cache locality

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

@@ -16,7 +16,7 @@ import it.unimi.dsi.fastutil.ints.IntSet;
 import net.minecraft.client.Minecraft;
 
 public class VisibilityBuffer {
-	private static final int READ_GROUP_SIZE = 16;
+	private static final int READ_GROUP_SIZE = 32;
 	private static final int ATTACHMENT = GL30.GL_COLOR_ATTACHMENT1;
 
 	private final GlProgram readVisibilityProgram;

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

@@ -1,6 +1,6 @@
 #include "flywheel:internal/indirect/buffer_bindings.glsl"
 
-layout(local_size_x = 16, local_size_y = 16) in;
+layout(local_size_x = 256) in;
 
 layout(binding = 0) uniform usampler2D visBuffer;
 
@@ -8,9 +8,24 @@ layout(std430, binding = _FLW_LAST_FRAME_VISIBILITY_BUFFER_BINDING) restrict buf
     uint _flw_lastFrameVisibility[];
 };
 
-void main() {
+uint extractBits(uint e, uint offset, uint count) {
-    uint instanceID = texelFetch(visBuffer, ivec2(gl_GlobalInvocationID.xy), 0).r;
+    return (e >> offset) & ((1u << count) - 1u);
+}
 
+uint insertBits(uint e, uint newbits, uint offset, uint count) {
+    uint countMask = ((1u << count) - 1u);
+    // zero out the bits we're going to replace first
+    return (e & ~(countMask << offset)) | ((newbits & countMask) << offset);
+}
+
+uvec2 remap_for_wave_reduction(uint a) {
+    return uvec2(
+    insertBits(extractBits(a, 2u, 3u), a, 0u, 1u),
+    insertBits(extractBits(a, 3u, 3u), extractBits(a, 1u, 2u), 0u, 2u)
+    );
+}
+
+void emit(uint instanceID) {
     // Null instance id.
     if (instanceID == 0) {
         return;
@@ -25,3 +40,25 @@ void main() {
 
     atomicOr(_flw_lastFrameVisibility[index], mask);
 }
+
+void main() {
+    uvec2 sub_xy = remap_for_wave_reduction(gl_LocalInvocationIndex % 64u);
+    uint x = sub_xy.x + 8u * ((gl_LocalInvocationIndex >> 6u) % 2u);
+    uint y = sub_xy.y + 8u * (gl_LocalInvocationIndex >> 7u);
+
+    ivec2 tex = ivec2(gl_WorkGroupID.xy) * 32 + ivec2(x, y) * 2;
+
+    uint instanceID01 = texelFetchOffset(visBuffer, tex, 0, ivec2(0, 1)).r;
+    uint instanceID11 = texelFetchOffset(visBuffer, tex, 0, ivec2(1, 1)).r;
+    uint instanceID10 = texelFetchOffset(visBuffer, tex, 0, ivec2(1, 0)).r;
+    uint instanceID00 = texelFetch(visBuffer, tex, 0).r;
+
+    if (instanceID00 == instanceID01 && instanceID01 == instanceID10 && instanceID10 == instanceID11) {
+        emit(instanceID00);
+    } else {
+        emit(instanceID00);
+        emit(instanceID01);
+        emit(instanceID10);
+        emit(instanceID11);
+    }
+}