Joining the occult

- Implement hi-z occlusion culling
- Generate depth pyramid just before issuing cull dispatches
- Currently use raw texel fetches but this may be causing loss
- Add _flw_cullData to frame uniforms

common/src/backend/java/dev/engine_room/flywheel/backend/compile/IndirectPrograms.java

@@ -30,6 +30,7 @@ public class IndirectPrograms extends AtomicReferenceCounted {
 	private static final ResourceLocation CULL_SHADER_MAIN = Flywheel.rl("internal/indirect/cull.glsl");
 	private static final ResourceLocation APPLY_SHADER_MAIN = Flywheel.rl("internal/indirect/apply.glsl");
 	private static final ResourceLocation SCATTER_SHADER_MAIN = Flywheel.rl("internal/indirect/scatter.glsl");
+	private static final ResourceLocation DEPTH_REDUCE_SHADER_MAIN = Flywheel.rl("internal/indirect/depth_reduce.glsl");
 
 	private static final Compile<InstanceType<?>> CULL = new Compile<>();
 	private static final Compile<ResourceLocation> UTIL = new Compile<>();
@@ -44,12 +45,14 @@ public class IndirectPrograms extends AtomicReferenceCounted {
 	private final Map<InstanceType<?>, GlProgram> culling;
 	private final GlProgram apply;
 	private final GlProgram scatter;
+	private final GlProgram depthReduce;
 
-	private IndirectPrograms(Map<PipelineProgramKey, GlProgram> pipeline, Map<InstanceType<?>, GlProgram> culling, GlProgram apply, GlProgram scatter) {
+	private IndirectPrograms(Map<PipelineProgramKey, GlProgram> pipeline, Map<InstanceType<?>, GlProgram> culling, GlProgram apply, GlProgram scatter, GlProgram depthReduce) {
 		this.pipeline = pipeline;
 		this.culling = culling;
 		this.apply = apply;
 		this.scatter = scatter;
+		this.depthReduce = depthReduce;
 	}
 
 	private static List<String> getExtensions(GlslVersion glslVersion) {
@@ -94,10 +97,10 @@ public class IndirectPrograms extends AtomicReferenceCounted {
 		try {
 			var pipelineResult = pipelineCompiler.compileAndReportErrors(pipelineKeys);
 			var cullingResult = cullingCompiler.compileAndReportErrors(createCullingKeys());
-			var utils = utilCompiler.compileAndReportErrors(List.of(APPLY_SHADER_MAIN, SCATTER_SHADER_MAIN));
+			var utils = utilCompiler.compileAndReportErrors(List.of(APPLY_SHADER_MAIN, SCATTER_SHADER_MAIN, DEPTH_REDUCE_SHADER_MAIN));
 
 			if (pipelineResult != null && cullingResult != null && utils != null) {
-				newInstance = new IndirectPrograms(pipelineResult, cullingResult, utils.get(APPLY_SHADER_MAIN), utils.get(SCATTER_SHADER_MAIN));
+				newInstance = new IndirectPrograms(pipelineResult, cullingResult, utils.get(APPLY_SHADER_MAIN), utils.get(SCATTER_SHADER_MAIN), utils.get(DEPTH_REDUCE_SHADER_MAIN));
 			}
 		} catch (Throwable t) {
 			FlwPrograms.LOGGER.error("Failed to compile indirect programs", t);
@@ -184,6 +187,10 @@ public class IndirectPrograms extends AtomicReferenceCounted {
 		return scatter;
 	}
 
+	public GlProgram getDepthReduceProgram() {
+		return depthReduce;
+	}
+
 	@Override
 	protected void _delete() {
 		pipeline.values()

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

@@ -0,0 +1,106 @@
+package dev.engine_room.flywheel.backend.engine.indirect;
+
+import org.lwjgl.opengl.GL32;
+import org.lwjgl.opengl.GL46;
+
+import com.mojang.blaze3d.platform.GlStateManager;
+
+import dev.engine_room.flywheel.backend.gl.shader.GlProgram;
+import dev.engine_room.flywheel.lib.math.MoreMath;
+import net.minecraft.client.Minecraft;
+
+public class DepthPyramid {
+	private final GlProgram depthReduceProgram;
+
+	public final int pyramidTextureId;
+
+	private int lastWidth = -1;
+	private int lastHeight = -1;
+
+	public DepthPyramid(GlProgram depthReduceProgram) {
+		this.depthReduceProgram = depthReduceProgram;
+
+		pyramidTextureId = GL32.glGenTextures();
+
+		GlStateManager._bindTexture(pyramidTextureId);
+		GlStateManager._texParameter(GL32.GL_TEXTURE_2D, GL32.GL_TEXTURE_MIN_FILTER, GL32.GL_NEAREST);
+		GlStateManager._texParameter(GL32.GL_TEXTURE_2D, GL32.GL_TEXTURE_MAG_FILTER, GL32.GL_NEAREST);
+		GlStateManager._texParameter(GL32.GL_TEXTURE_2D, GL32.GL_TEXTURE_COMPARE_MODE, GL32.GL_NONE);
+		GlStateManager._texParameter(GL32.GL_TEXTURE_2D, GL32.GL_TEXTURE_WRAP_S, GL32.GL_CLAMP_TO_EDGE);
+		GlStateManager._texParameter(GL32.GL_TEXTURE_2D, GL32.GL_TEXTURE_WRAP_T, GL32.GL_CLAMP_TO_EDGE);
+
+	}
+
+	public void generate() {
+		var mainRenderTarget = Minecraft.getInstance()
+				.getMainRenderTarget();
+
+		int width = mainRenderTarget.width;
+		int height = mainRenderTarget.height;
+
+		int mipLevels = getImageMipLevels(width, height);
+
+		createPyramidMips(mipLevels, width, height);
+
+		int depthBufferId = mainRenderTarget.getDepthTextureId();
+
+		GlStateManager._bindTexture(depthBufferId);
+
+		GL46.glMemoryBarrier(GL46.GL_FRAMEBUFFER_BARRIER_BIT);
+
+		GL46.glActiveTexture(GL32.GL_TEXTURE1);
+
+		depthReduceProgram.bind();
+
+		for (int i = 0; i < mipLevels; i++) {
+			int mipWidth = Math.max(1, width >> i);
+			int mipHeight = Math.max(1, height >> i);
+
+			int srcTexture = (i == 0) ? depthBufferId : pyramidTextureId;
+			GL46.glBindTexture(GL32.GL_TEXTURE_2D, srcTexture);
+
+			GL46.glBindImageTexture(0, pyramidTextureId, i, false, 0, GL32.GL_WRITE_ONLY, GL32.GL_R32F);
+
+			depthReduceProgram.setUVec2("imageSize", mipWidth, mipHeight);
+			depthReduceProgram.setInt("lod", Math.max(0, i - 1));
+
+			GL46.glDispatchCompute(MoreMath.ceilingDiv(mipWidth, 8), MoreMath.ceilingDiv(mipHeight, 8), 1);
+
+			GL46.glMemoryBarrier(GL46.GL_TEXTURE_FETCH_BARRIER_BIT);
+		}
+	}
+
+	public void delete() {
+		GL32.glDeleteTextures(pyramidTextureId);
+	}
+
+	private void createPyramidMips(int mipLevels, int width, int height) {
+		if (lastWidth == width && lastHeight == height) {
+			return;
+		}
+
+		lastWidth = width;
+		lastHeight = height;
+
+		GL32.glBindTexture(GL32.GL_TEXTURE_2D, pyramidTextureId);
+
+		for (int i = 0; i < mipLevels; i++) {
+			int mipWidth = Math.max(1, width >> (i + 1));
+			int mipHeight = Math.max(1, height >> (i + 1));
+
+			GL32.glTexImage2D(GL32.GL_TEXTURE_2D, i, GL32.GL_R32F, mipWidth, mipHeight, 0, GL32.GL_RED, GL32.GL_FLOAT, 0);
+		}
+	}
+
+	private static int getImageMipLevels(int width, int height) {
+		int result = 1;
+
+		while (width > 2 && height > 2) {
+			result++;
+			width /= 2;
+			height /= 2;
+		}
+
+		return result;
+	}
+}

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

@@ -12,6 +12,8 @@ import java.util.HashMap;
 import java.util.List;
 import java.util.Map;
 
+import org.lwjgl.opengl.GL46;
+
 import dev.engine_room.flywheel.api.backend.Engine;
 import dev.engine_room.flywheel.api.instance.Instance;
 import dev.engine_room.flywheel.api.instance.InstanceType;
@@ -46,6 +48,8 @@ public class IndirectDrawManager extends DrawManager<IndirectInstancer<?>> {
 	private final LightBuffers lightBuffers;
 	private final MatrixBuffer matrixBuffer;
 
+	private final DepthPyramid depthPyramid;
+
 	private boolean needsBarrier = false;
 
 	public IndirectDrawManager(IndirectPrograms programs) {
@@ -58,6 +62,8 @@ public class IndirectDrawManager extends DrawManager<IndirectInstancer<?>> {
 		meshPool.bind(vertexArray);
 		lightBuffers = new LightBuffers();
 		matrixBuffer = new MatrixBuffer();
+
+		depthPyramid = new DepthPyramid(programs.getDepthReduceProgram());
 	}
 
 	@Override
@@ -136,6 +142,8 @@ public class IndirectDrawManager extends DrawManager<IndirectInstancer<?>> {
 
 		stagingBuffer.flush();
 
+		depthPyramid.generate();
+
 		// We could probably save some driver calls here when there are
 		// actually zero instances, but that feels like a very rare case
 
@@ -143,6 +151,9 @@ public class IndirectDrawManager extends DrawManager<IndirectInstancer<?>> {
 
 		matrixBuffer.bind();
 
+		GL46.glActiveTexture(GL46.GL_TEXTURE0);
+		GL46.glBindTexture(GL46.GL_TEXTURE_2D, depthPyramid.pyramidTextureId);
+
 		for (var group : cullingGroups.values()) {
 			group.dispatchCull();
 		}

common/src/backend/java/dev/engine_room/flywheel/backend/engine/uniform/FrameUniforms.java

@@ -17,7 +17,7 @@ import net.minecraft.world.level.Level;
 import net.minecraft.world.phys.Vec3;
 
 public final class FrameUniforms extends UniformWriter {
-	private static final int SIZE = 96 + 64 * 9 + 16 * 5 + 8 * 2 + 8 + 4 * 10;
+	private static final int SIZE = 96 + 64 * 9 + 16 * 5 + 8 * 2 + 8 + 4 * 16;
 	static final UniformBuffer BUFFER = new UniformBuffer(Uniforms.FRAME_INDEX, SIZE);
 
 	private static final Matrix4f VIEW = new Matrix4f();
@@ -112,6 +112,8 @@ public final class FrameUniforms extends UniformWriter {
 
 		ptr = writeInt(ptr, debugMode);
 
+		ptr = writeCullData(ptr);
+
 		firstWrite = false;
 		BUFFER.markDirty();
 	}
@@ -179,6 +181,18 @@ public final class FrameUniforms extends UniformWriter {
 		return writeInFluidAndBlock(ptr, level, blockPos, cameraPos);
 	}
 
+	private static long writeCullData(long ptr) {
+		ptr = writeFloat(ptr, 0.05F); // zNear
+		ptr = writeFloat(ptr, Minecraft.getInstance().gameRenderer.getDepthFar()); // zFar
+		ptr = writeFloat(ptr, PROJECTION.m00()); // P00
+		ptr = writeFloat(ptr, PROJECTION.m11()); // P11
+		ptr = writeFloat(ptr, Minecraft.getInstance().getMainRenderTarget().width >> 1); // pyramidWidth
+		ptr = writeFloat(ptr, Minecraft.getInstance().getMainRenderTarget().height >> 1); // pyramidHeight
+		ptr = writeInt(ptr, 0); // useMin
+
+		return ptr;
+	}
+
 	/**
 	 * Writes the frustum planes of the given projection matrix to the given buffer.<p>
 	 * Uses a different format that is friendly towards an optimized instruction-parallel

common/src/backend/java/dev/engine_room/flywheel/backend/gl/shader/GlProgram.java

@@ -11,6 +11,7 @@ import static org.lwjgl.opengl.GL20.glUniform4f;
 import static org.lwjgl.opengl.GL20.glUniformMatrix3fv;
 import static org.lwjgl.opengl.GL20.glUniformMatrix4fv;
 import static org.lwjgl.opengl.GL30.glUniform1ui;
+import static org.lwjgl.opengl.GL30.glUniform2ui;
 import static org.lwjgl.opengl.GL31.GL_INVALID_INDEX;
 import static org.lwjgl.opengl.GL31.glGetUniformBlockIndex;
 import static org.lwjgl.opengl.GL31.glUniformBlockBinding;
@@ -118,6 +119,16 @@ public class GlProgram extends GlObject {
 		glUniform1ui(uniform, value);
 	}
 
+	public void setUVec2(String name, int x, int y) {
+		int uniform = getUniformLocation(name);
+
+		if (uniform < 0) {
+			return;
+		}
+
+		glUniform2ui(uniform, x, y);
+	}
+
 	public void setInt(String glslName, int value) {
 		int uniform = getUniformLocation(glslName);
 

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

@@ -23,10 +23,12 @@ layout(std430, binding = _FLW_MODEL_BUFFER_BINDING) restrict buffer ModelBuffer
     ModelDescriptor _flw_models[];
 };
 
-layout(std430, binding = _FLW_MATRIX_BUFFER_BINDING) restrict buffer MatrixBuffer {
+layout(std430, binding = _FLW_MATRIX_BUFFER_BINDING) restrict readonly buffer MatrixBuffer {
     Matrices _flw_matrices[];
 };
 
+layout(binding = 0) uniform sampler2D _flw_depthPyramid;
+
 // Disgustingly vectorized sphere frustum intersection taking advantage of ahead of time packing.
 // Only uses 6 fmas and some boolean ops.
 // See also:
@@ -40,6 +42,28 @@ bool _flw_testSphere(vec3 center, float radius) {
     return all(xyInside) && all(zInside);
 }
 
+bool projectSphere(vec3 c, float r, float znear, float P00, float P11, out vec4 aabb) {
+    if (c.z > r + znear) {
+        return false;
+    }
+
+    vec3 cr = c * r;
+    float czr2 = c.z * c.z - r * r;
+
+    float vx = sqrt(c.x * c.x + czr2);
+    float minx = (vx * c.x - cr.z) / (vx * c.z + cr.x);
+    float maxx = (vx * c.x + cr.z) / (vx * c.z - cr.x);
+
+    float vy = sqrt(c.y * c.y + czr2);
+    float miny = (vy * c.y - cr.z) / (vy * c.z + cr.y);
+    float maxy = (vy * c.y + cr.z) / (vy * c.z - cr.y);
+
+    aabb = vec4(minx * P00, miny * P11, maxx * P00, maxy * P11);
+    aabb = aabb.xwzy * vec4(-0.5f, -0.5f, -0.5f, -0.5f) + vec4(0.5f); // clip space -> uv space
+
+    return true;
+}
+
 bool _flw_isVisible(uint instanceIndex, uint modelIndex) {
     uint matrixIndex = _flw_models[modelIndex].matrixIndex;
     BoundingSphere sphere = _flw_models[modelIndex].boundingSphere;
@@ -56,7 +80,38 @@ bool _flw_isVisible(uint instanceIndex, uint modelIndex) {
         transformBoundingSphere(_flw_matrices[matrixIndex].pose, center, radius);
     }
 
-    return _flw_testSphere(center, radius);
+    bool isVisible = _flw_testSphere(center, radius);
+
+    if (isVisible) {
+        transformBoundingSphere(flw_view, center, radius);
+
+        vec4 aabb;
+        if (projectSphere(center, radius, _flw_cullData.znear, _flw_cullData.P00, _flw_cullData.P11, aabb))
+        {
+            float width = (aabb.z - aabb.x) * _flw_cullData.pyramidWidth;
+            float height = (aabb.w - aabb.y) * _flw_cullData.pyramidHeight;
+
+            float level = floor(log2(max(width, height)));
+
+            float depth01 = textureLod(_flw_depthPyramid, aabb.xw, level).r;
+            float depth11 = textureLod(_flw_depthPyramid, aabb.zw, level).r;
+            float depth10 = textureLod(_flw_depthPyramid, aabb.zy, level).r;
+            float depth00 = textureLod(_flw_depthPyramid, aabb.xy, level).r;
+
+            float depth;
+            if (_flw_cullData.useMin == 0) {
+                depth = max(max(depth00, depth01), max(depth10, depth11));
+            } else {
+                depth = min(min(depth00, depth01), min(depth10, depth11));
+            }
+
+            float depthSphere = 1. + _flw_cullData.znear / (center.z + radius);
+
+            isVisible = isVisible && depthSphere <= depth;
+        }
+    }
+
+    return isVisible;
 }
 
 void main() {

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

@@ -0,0 +1,29 @@
+layout(local_size_x = 8, local_size_y = 8) in;
+
+layout(binding = 0, r32f) uniform writeonly image2D outImage;
+layout(binding = 1) uniform sampler2D inImage;
+
+uniform uvec2 imageSize;
+uniform int lod;
+
+uniform int useMin = 0;
+
+void main() {
+    uvec2 pos = gl_GlobalInvocationID.xy;
+
+    ivec2 samplePos = ivec2(pos) * 2;
+
+    float depth01 = texelFetchOffset(inImage, samplePos, lod, ivec2(0, 1)).r;
+    float depth11 = texelFetchOffset(inImage, samplePos, lod, ivec2(1, 1)).r;
+    float depth10 = texelFetchOffset(inImage, samplePos, lod, ivec2(1, 0)).r;
+    float depth00 = texelFetchOffset(inImage, samplePos, lod, ivec2(0, 0)).r;
+
+    float depth;
+    if (useMin == 0) {
+        depth = max(max(depth00, depth01), max(depth10, depth11));
+    } else {
+        depth = min(min(depth00, depth01), min(depth10, depth11));
+    }
+
+    imageStore(outImage, ivec2(pos), vec4(depth));
+}

common/src/backend/resources/assets/flywheel/flywheel/internal/uniforms/frame.glsl

@@ -9,6 +9,16 @@ struct FrustumPlanes {
     vec2 zW; // <nz.w, pz.w>
 };
 
+struct _FlwCullData {
+    float znear;
+    float zfar;
+    float P00;
+    float P11;
+    float pyramidWidth;
+    float pyramidHeight;
+    uint useMin;
+};
+
 layout(std140) uniform _FlwFrameUniforms {
     FrustumPlanes flw_frustumPlanes;
 
@@ -47,6 +57,8 @@ layout(std140) uniform _FlwFrameUniforms {
     uint flw_cameraInBlock;
 
     uint _flw_debugMode;
+
+    _FlwCullData _flw_cullData;
 };
 
 #define flw_renderOrigin (_flw_renderOrigin.xyz)