From b885d7c1ec1763d0884113d005dc33aa41b8e492 Mon Sep 17 00:00:00 2001
From: mcrcortex <{ID}+{username}@users.noreply.github.com>
Date: Thu, 9 May 2024 21:46:42 +1000
Subject: [PATCH] Size 62 meshlets + major fix for hiz culling

---
 .../voxy/client/core/gl/shader/Shader.java    | 28 ++++++++++++++++++-
 .../Gl46MeshletsFarWorldRenderer.java         | 16 ++++++++---
 .../rendering/building/RenderDataFactory.java |  2 +-
 .../voxy/shaders/lod/gl46mesh/bindings.glsl   |  2 +-
 .../voxy/shaders/lod/gl46mesh/meshlet.glsl    |  2 --
 .../shaders/lod/gl46mesh/meshletculler.comp   | 11 +++++---
 6 files changed, 48 insertions(+), 13 deletions(-)

diff --git a/src/main/java/me/cortex/voxy/client/core/gl/shader/Shader.java b/src/main/java/me/cortex/voxy/client/core/gl/shader/Shader.java
index a3f3814d..63d215b5 100644
--- a/src/main/java/me/cortex/voxy/client/core/gl/shader/Shader.java
+++ b/src/main/java/me/cortex/voxy/client/core/gl/shader/Shader.java
@@ -5,6 +5,7 @@ import org.lwjgl.opengl.GL20C;
 
 import java.util.HashMap;
 import java.util.Map;
+import java.util.stream.Collectors;
 
 import static org.lwjgl.opengl.GL20.glDeleteProgram;
 import static org.lwjgl.opengl.GL20.glUseProgram;
@@ -37,12 +38,23 @@ public class Shader extends TrackedObject {
     }
 
     public static class Builder {
+        private final Map<String, String> defines = new HashMap<>();
         private final Map<ShaderType, String> sources = new HashMap<>();
         private final IShaderProcessor processor;
         private Builder(IShaderProcessor processor) {
             this.processor = processor;
         }
 
+        public Builder define(String name) {
+            this.defines.put(name, "");
+            return this;
+        }
+
+        public Builder define(String name, int value) {
+            this.defines.put(name, Integer.toString(value));
+            return this;
+        }
+
         public Builder add(ShaderType type, String id) {
             this.addSource(type, ShaderLoader.parse(id));
             return this;
@@ -55,7 +67,21 @@ public class Shader extends TrackedObject {
 
         public Shader compile() {
             int program = GL20C.glCreateProgram();
-            int[] shaders = this.sources.entrySet().stream().mapToInt(a->createShader(a.getKey(), a.getValue())).toArray();
+            int[] shaders = new int[this.sources.size()];
+            {
+                String defs = this.defines.entrySet().stream().map(a->"#define " + a.getKey() + " " + a.getValue() + "\n").collect(Collectors.joining());
+                int i = 0;
+                for (var entry : this.sources.entrySet()) {
+                    String src = entry.getValue();
+
+                    //Inject defines
+                    src = src.substring(0, src.indexOf('\n')+1) +
+                            defs
+                            + src.substring(src.indexOf('\n')+1);
+
+                    shaders[i++] = createShader(entry.getKey(), src);
+                }
+            }
 
             for (int i : shaders) {
                 GL20C.glAttachShader(program, i);
diff --git a/src/main/java/me/cortex/voxy/client/core/rendering/Gl46MeshletsFarWorldRenderer.java b/src/main/java/me/cortex/voxy/client/core/rendering/Gl46MeshletsFarWorldRenderer.java
index 35f0ccc9..03efe9dc 100644
--- a/src/main/java/me/cortex/voxy/client/core/rendering/Gl46MeshletsFarWorldRenderer.java
+++ b/src/main/java/me/cortex/voxy/client/core/rendering/Gl46MeshletsFarWorldRenderer.java
@@ -3,6 +3,7 @@ package me.cortex.voxy.client.core.rendering;
 import me.cortex.voxy.client.core.gl.GlBuffer;
 import me.cortex.voxy.client.core.gl.shader.Shader;
 import me.cortex.voxy.client.core.gl.shader.ShaderType;
+import me.cortex.voxy.client.core.rendering.building.RenderDataFactory;
 import me.cortex.voxy.client.core.rendering.util.UploadStream;
 import me.cortex.voxy.client.mixin.joml.AccessFrustumIntersection;
 import net.minecraft.client.MinecraftClient;
@@ -44,20 +45,24 @@ import static org.lwjgl.opengl.NVRepresentativeFragmentTest.GL_REPRESENTATIVE_FR
 // this could potentially result in a fair bit of memory savings (especially if used in normal mc terrain rendering)
 public class Gl46MeshletsFarWorldRenderer extends AbstractFarWorldRenderer<Gl46MeshletViewport, DefaultGeometryManager> {
     private final Shader lodShader = Shader.make()
+            .define("QUADS_PER_MESHLET", RenderDataFactory.QUADS_PER_MESHLET)
             .add(ShaderType.VERTEX, "voxy:lod/gl46mesh/quads.vert")
             .add(ShaderType.FRAGMENT, "voxy:lod/gl46mesh/quads.frag")
             .compile();
 
     private final Shader cullShader = Shader.make()
+            .define("QUADS_PER_MESHLET", RenderDataFactory.QUADS_PER_MESHLET)
             .add(ShaderType.VERTEX, "voxy:lod/gl46mesh/cull.vert")
             .add(ShaderType.FRAGMENT, "voxy:lod/gl46mesh/cull.frag")
             .compile();
 
     private final Shader meshletGenerator = Shader.make()
+            .define("QUADS_PER_MESHLET", RenderDataFactory.QUADS_PER_MESHLET)
             .add(ShaderType.COMPUTE, "voxy:lod/gl46mesh/cmdgen.comp")
             .compile();
 
     private final Shader meshletCuller = Shader.make()
+            .define("QUADS_PER_MESHLET", RenderDataFactory.QUADS_PER_MESHLET)
             .add(ShaderType.COMPUTE, "voxy:lod/gl46mesh/meshletculler.comp")
             .compile();
 
@@ -67,15 +72,19 @@ public class Gl46MeshletsFarWorldRenderer extends AbstractFarWorldRenderer<Gl46M
     private final int hizSampler = glGenSamplers();
 
     public Gl46MeshletsFarWorldRenderer(int geometrySize, int maxSections) {
-        super(new DefaultGeometryManager(alignUp(geometrySize*8L, 8*32), maxSections, 8*32));
+        super(new DefaultGeometryManager(alignUp(geometrySize*8L, 8* (RenderDataFactory.QUADS_PER_MESHLET+2)), maxSections, 8*(RenderDataFactory.QUADS_PER_MESHLET+2)));
         this.glDrawIndirect = new GlBuffer(4*(4+5));
         this.meshletBuffer = new GlBuffer(4*1000000);//TODO: Make max meshlet count configurable, not just 1 million (even tho thats a max of 126 million quads per frame)
 
-        glSamplerParameteri(this.hizSampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST_MIPMAP_NEAREST);
+        glSamplerParameteri(this.hizSampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);//This is so that using the shadow sampler works correctly
         glTextureParameteri(this.hizSampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
-        glTextureParameteri(this.hizSampler, GL_TEXTURE_COMPARE_MODE, GL_NONE);
+        glTextureParameteri(this.hizSampler, GL_TEXTURE_COMPARE_MODE, GL_COMPARE_REF_TO_TEXTURE);
+        glTextureParameteri(this.hizSampler, GL_TEXTURE_COMPARE_FUNC, GL_LEQUAL);
         glTextureParameteri(this.hizSampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
         glTextureParameteri(this.hizSampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+
+        nglClearNamedBufferData(this.meshletBuffer.id, GL_R32UI, GL_RED_INTEGER, GL_UNSIGNED_INT, 0);
+        nglClearNamedBufferData(this.glDrawIndirect.id, GL_R32UI, GL_RED_INTEGER, GL_UNSIGNED_INT, 0);
     }
 
     protected void bindResources(Gl46MeshletViewport viewport, boolean bindToDrawIndirect, boolean bindToDispatchIndirect, boolean bindHiz) {
@@ -122,7 +131,6 @@ public class Gl46MeshletsFarWorldRenderer extends AbstractFarWorldRenderer<Gl46M
         }
         innerTranslation.getToAddress(ptr); ptr += 4*3;
         MemoryUtil.memPutInt(ptr, viewport.frameId++); ptr += 4;
-        //Divided by 2 cause hiz is half the size of the viewport
         MemoryUtil.memPutInt(ptr, viewport.width); ptr += 4;
         MemoryUtil.memPutInt(ptr, viewport.height); ptr += 4;
     }
diff --git a/src/main/java/me/cortex/voxy/client/core/rendering/building/RenderDataFactory.java b/src/main/java/me/cortex/voxy/client/core/rendering/building/RenderDataFactory.java
index e5b68d99..9e0f8de3 100644
--- a/src/main/java/me/cortex/voxy/client/core/rendering/building/RenderDataFactory.java
+++ b/src/main/java/me/cortex/voxy/client/core/rendering/building/RenderDataFactory.java
@@ -55,7 +55,7 @@ public class RenderDataFactory {
     // can do funny stuff like double rendering
 
     private static final boolean USE_UINT64 = Capabilities.INSTANCE.INT64_t;
-    private static final int QUADS_PER_MESHLET = 30;
+    public static final int QUADS_PER_MESHLET = 62;
     private static void writePos(long ptr, long pos) {
         if (USE_UINT64) {
             MemoryUtil.memPutLong(ptr, pos);
diff --git a/src/main/resources/assets/voxy/shaders/lod/gl46mesh/bindings.glsl b/src/main/resources/assets/voxy/shaders/lod/gl46mesh/bindings.glsl
index 7061e2e9..b0c59efa 100644
--- a/src/main/resources/assets/voxy/shaders/lod/gl46mesh/bindings.glsl
+++ b/src/main/resources/assets/voxy/shaders/lod/gl46mesh/bindings.glsl
@@ -46,7 +46,7 @@ struct DispatchIndirect {
 };
 
 #ifdef BIND_SAMPLER_AS_HIZ
-layout(binding = 0) uniform sampler2D hizSampler;
+layout(binding = 0) uniform sampler2DShadow hizSampler;
 #else
 layout(binding = 0) uniform sampler2D blockModelAtlas;
 #endif
diff --git a/src/main/resources/assets/voxy/shaders/lod/gl46mesh/meshlet.glsl b/src/main/resources/assets/voxy/shaders/lod/gl46mesh/meshlet.glsl
index 5b4d7b40..e628edb7 100644
--- a/src/main/resources/assets/voxy/shaders/lod/gl46mesh/meshlet.glsl
+++ b/src/main/resources/assets/voxy/shaders/lod/gl46mesh/meshlet.glsl
@@ -1,5 +1,3 @@
-#define QUADS_PER_MESHLET 30
-
 #define extractMeshletStart extractQuadStart
 #define PosHeader Quad
 #define AABBHeader Quad
diff --git a/src/main/resources/assets/voxy/shaders/lod/gl46mesh/meshletculler.comp b/src/main/resources/assets/voxy/shaders/lod/gl46mesh/meshletculler.comp
index 5892d391..cd968931 100644
--- a/src/main/resources/assets/voxy/shaders/lod/gl46mesh/meshletculler.comp
+++ b/src/main/resources/assets/voxy/shaders/lod/gl46mesh/meshletculler.comp
@@ -32,19 +32,22 @@ bool testHiZ(PosHeader secPos, AABBHeader aabb) {
         maxBB = max(maxBB, point);
     }
 
-    minBB = minBB*0.5+0.5;
-    maxBB = maxBB*0.5+0.5;
+    minBB = clamp(minBB*0.5+0.5, vec3(0), vec3(1));
+    maxBB = clamp(maxBB*0.5+0.5, vec3(0), vec3(1));
 
     vec2 size = (maxBB.xy - minBB.xy) * vec2(ivec2(screensize));
     float miplevel = ceil(log2(max(max(size.x, size.y),1)));
-
+    /*
     float a = textureLod(hizSampler,minBB.xy,miplevel).r;
     float b = textureLod(hizSampler,vec2(minBB.x,maxBB.y),miplevel).r;
     float c = textureLod(hizSampler,maxBB.xy,miplevel).r;
     float d = textureLod(hizSampler,vec2(maxBB.x,minBB.y),miplevel).r;
     float depth = max(max(a,b),max(c,d));
-
     return minBB.z <= depth;
+    */
+
+    vec2 midpoint = (maxBB.xy + minBB.xy)*0.5;
+    return textureLod(hizSampler, vec3(midpoint, minBB.z - 0.000000001), miplevel) > 0.0001;
 }