From 6ebc4739b9e750822fefba5954a00cae5ad9069e Mon Sep 17 00:00:00 2001
From: mcrcortex <{ID}+{username}@users.noreply.github.com>
Date: Mon, 15 Apr 2024 01:08:32 +1000
Subject: [PATCH] wip hiz

---
 .../voxy/client/core/gl/GlFramebuffer.java    |  6 +-
 .../Gl46MeshletsFarWorldRenderer.java         | 15 ++-
 .../voxy/client/core/rendering/HiZBuffer.java | 98 +++++++++++++++++++
 .../rendering/building/RenderDataFactory.java |  2 +-
 .../assets/voxy/shaders/hiz/blit.fsh          |  9 ++
 .../assets/voxy/shaders/hiz/blit.vsh          |  8 ++
 .../voxy/shaders/lod/gl46mesh/cmdgen.comp     |  2 +-
 .../voxy/shaders/lod/gl46mesh/quads.frag      | 10 ++
 .../voxy/shaders/lod/gl46mesh/quads.vert      |  4 +-
 9 files changed, 149 insertions(+), 5 deletions(-)
 create mode 100644 src/main/java/me/cortex/voxy/client/core/rendering/HiZBuffer.java
 create mode 100644 src/main/resources/assets/voxy/shaders/hiz/blit.fsh
 create mode 100644 src/main/resources/assets/voxy/shaders/hiz/blit.vsh

diff --git a/src/main/java/me/cortex/voxy/client/core/gl/GlFramebuffer.java b/src/main/java/me/cortex/voxy/client/core/gl/GlFramebuffer.java
index b9d428bd..e68c63c0 100644
--- a/src/main/java/me/cortex/voxy/client/core/gl/GlFramebuffer.java
+++ b/src/main/java/me/cortex/voxy/client/core/gl/GlFramebuffer.java
@@ -11,7 +11,11 @@ public class GlFramebuffer extends TrackedObject {
     }
 
     public GlFramebuffer bind(int attachment, GlTexture texture) {
-        glNamedFramebufferTexture(this.id, attachment, texture.id, 0);
+        return this.bind(attachment, texture, 0);
+    }
+
+    public GlFramebuffer bind(int attachment, GlTexture texture, int lvl) {
+        glNamedFramebufferTexture(this.id, attachment, texture.id, lvl);
         return this;
     }
 
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 769bd7b3..9902a2a0 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
@@ -5,11 +5,13 @@ 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.util.UploadStream;
 import me.cortex.voxy.client.mixin.joml.AccessFrustumIntersection;
+import net.minecraft.client.MinecraftClient;
 import net.minecraft.client.render.RenderLayer;
 import org.joml.Matrix4f;
 import org.joml.Vector3f;
 import org.lwjgl.system.MemoryUtil;
 
+import static org.lwjgl.opengl.ARBDirectStateAccess.glGetNamedFramebufferAttachmentParameteriv;
 import static org.lwjgl.opengl.ARBIndirectParameters.GL_PARAMETER_BUFFER_ARB;
 import static org.lwjgl.opengl.GL11.*;
 import static org.lwjgl.opengl.GL14C.glBlendFuncSeparate;
@@ -56,9 +58,10 @@ public class Gl46MeshletsFarWorldRenderer extends AbstractFarWorldRenderer<Gl46M
 
     private final GlBuffer glDrawIndirect;
     private final GlBuffer meshletBuffer;
+    private final HiZBuffer hiZBuffer = new HiZBuffer();
 
     public Gl46MeshletsFarWorldRenderer(int geometrySize, int maxSections) {
-        super(new DefaultGeometryManager(alignUp(geometrySize*8L, 8*64), maxSections, 8*64));
+        super(new DefaultGeometryManager(alignUp(geometrySize*8L, 8*32), maxSections, 8*32));
         this.glDrawIndirect = new GlBuffer(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)
     }
@@ -147,6 +150,14 @@ public class Gl46MeshletsFarWorldRenderer extends AbstractFarWorldRenderer<Gl46M
 
         glMemoryBarrier(GL_SHADER_STORAGE_BARRIER_BIT);
 
+
+
+        var i = new int[1];
+        glGetFramebufferAttachmentParameteriv(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, GL_FRAMEBUFFER_ATTACHMENT_OBJECT_NAME, i);
+        this.hiZBuffer.buildMipChain(i[0], MinecraftClient.getInstance().getFramebuffer().textureWidth, MinecraftClient.getInstance().getFramebuffer().textureHeight);
+
+
+
         glBindVertexArray(0);
         glBindSampler(0, 0);
         glBindTextureUnit(0, 0);
@@ -173,6 +184,8 @@ public class Gl46MeshletsFarWorldRenderer extends AbstractFarWorldRenderer<Gl46M
 
         this.glDrawIndirect.free();
         this.meshletBuffer.free();
+
+        this.hiZBuffer.free();
     }
 
     public static long alignUp(long n, long alignment) {
diff --git a/src/main/java/me/cortex/voxy/client/core/rendering/HiZBuffer.java b/src/main/java/me/cortex/voxy/client/core/rendering/HiZBuffer.java
new file mode 100644
index 00000000..44f2c8c6
--- /dev/null
+++ b/src/main/java/me/cortex/voxy/client/core/rendering/HiZBuffer.java
@@ -0,0 +1,98 @@
+package me.cortex.voxy.client.core.rendering;
+
+import me.cortex.voxy.client.core.gl.GlFramebuffer;
+import me.cortex.voxy.client.core.gl.GlTexture;
+import me.cortex.voxy.client.core.gl.shader.Shader;
+import me.cortex.voxy.client.core.gl.shader.ShaderType;
+import org.lwjgl.opengl.GL11;
+
+import static org.lwjgl.opengl.ARBDirectStateAccess.*;
+import static org.lwjgl.opengl.GL11C.*;
+import static org.lwjgl.opengl.GL30C.*;
+import static org.lwjgl.opengl.GL33.glBindSampler;
+import static org.lwjgl.opengl.GL33.glGenSamplers;
+import static org.lwjgl.opengl.GL33C.glDeleteSamplers;
+import static org.lwjgl.opengl.GL45C.glNamedFramebufferTexture;
+
+public class HiZBuffer {
+    private final Shader hiz = Shader.make()
+            .add(ShaderType.VERTEX, "voxy:hiz/blit.vsh")
+            .add(ShaderType.FRAGMENT, "voxy:hiz/blit.fsh")
+            .compile();
+    private final GlFramebuffer fb = new GlFramebuffer();
+    private final int sampler = glGenSamplers();
+    private GlTexture texture;
+    private int levels;
+    private int width;
+    private int height;
+
+    public HiZBuffer() {
+        glNamedFramebufferDrawBuffer(this.fb.id, GL_NONE);
+    }
+
+    private void alloc(int width, int height) {
+        this.levels = (int)Math.ceil(Math.log(Math.max(width, height))/Math.log(2));
+        //We dont care about e.g. 1x1 size texture since you dont get meshlets that big to cover such a large area
+        this.levels -= 3;//Arbitrary size, shinks the max level by alot and saves a significant amount of processing time
+        // (could probably increase it to be defined by a max meshlet coverage computation thing)
+
+        //We do a hack where we assume (which is probably wrong) that we will on average never
+        // get a meshlet that uses 0 mipping
+        this.levels--;
+        this.texture = new GlTexture().store(GL_DEPTH_COMPONENT32, this.levels, width/2, height/2);
+        glTextureParameteri(this.texture.id, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+        glTextureParameteri(this.texture.id, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+        glTextureParameteri(this.texture.id, GL_TEXTURE_COMPARE_MODE, GL_NONE);
+        glTextureParameteri(this.texture.id, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+        glTextureParameteri(this.texture.id, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+        this.width  = width;
+        this.height = height;
+    }
+
+    public void buildMipChain(int srcDepthTex, int width, int height) {
+        if (this.width != width || this.height != height) {
+            if (this.texture != null) {
+                this.texture.free();
+                this.texture = null;
+            }
+            this.alloc(width, height);
+        }
+        glBindVertexArray(AbstractFarWorldRenderer.STATIC_VAO);
+        int boundFB = GL11.glGetInteger(GL_DRAW_FRAMEBUFFER_BINDING);
+        this.hiz.bind();
+        this.fb.bind(GL_DEPTH_ATTACHMENT, this.texture, 0).verify();
+        glBindFramebuffer(GL_FRAMEBUFFER, this.fb.id);
+
+        glDepthFunc(GL_ALWAYS);
+
+        glBindTextureUnit(0, srcDepthTex);
+        glBindSampler(0, this.sampler);
+        glUniform1i(0, 0);
+        int cw = this.width /2;
+        int ch = this.height/2;
+        for (int i = 0; i < this.levels; i++) {
+            this.fb.bind(GL_DEPTH_ATTACHMENT, this.texture, i);
+            glViewport(0, 0, cw, ch); cw /= 2; ch /= 2;
+            glDrawArrays(GL_TRIANGLE_FAN, 0, 4);
+            if (i == 0) {
+                glBindTextureUnit(0, this.texture.id);
+            }
+            glTextureParameteri(this.texture.id, GL_TEXTURE_BASE_LEVEL, i);
+            glTextureParameteri(this.texture.id, GL_TEXTURE_MAX_LEVEL, i+1);
+        }
+        glTextureParameteri(this.texture.id, GL_TEXTURE_BASE_LEVEL, 0);
+        glTextureParameteri(this.texture.id, GL_TEXTURE_MAX_LEVEL, this.levels-1);//TODO: CHECK IF ITS -1 or -0
+
+        glDepthFunc(GL_LEQUAL);
+        glBindFramebuffer(GL_FRAMEBUFFER, boundFB);
+        glViewport(0, 0, width, height);
+        glBindVertexArray(0);
+    }
+
+    public void free() {
+        this.fb.free();
+        this.texture.free();
+        this.texture = null;
+        glDeleteSamplers(this.sampler);
+    }
+}
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 eec4d928..925d8b98 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
@@ -53,7 +53,7 @@ public class RenderDataFactory {
     // can do funny stuff like double rendering
 
     private static final boolean USE_UINT64 = false;//FIXME: replace with automatic detection of uint64 shader extension support
-    private static final int QUADS_PER_MESHLET = 62;
+    private static final int QUADS_PER_MESHLET = 30;
     private static void writePos(long ptr, long pos) {
         if (USE_UINT64) {
             MemoryUtil.memPutLong(ptr, pos);
diff --git a/src/main/resources/assets/voxy/shaders/hiz/blit.fsh b/src/main/resources/assets/voxy/shaders/hiz/blit.fsh
new file mode 100644
index 00000000..8dc82bc1
--- /dev/null
+++ b/src/main/resources/assets/voxy/shaders/hiz/blit.fsh
@@ -0,0 +1,9 @@
+#version 450
+
+
+layout(location = 0) in vec2 uv;
+layout(binding = 0) uniform sampler2D depthTex;
+void main() {
+    vec4 depths = textureGather(depthTex, uv, 0); // Get depth values from all surrounding texels.
+    gl_FragDepth = max(max(depths.x, depths.y), max(depths.z, depths.w)); // Write conservative depth.
+}
\ No newline at end of file
diff --git a/src/main/resources/assets/voxy/shaders/hiz/blit.vsh b/src/main/resources/assets/voxy/shaders/hiz/blit.vsh
new file mode 100644
index 00000000..29735329
--- /dev/null
+++ b/src/main/resources/assets/voxy/shaders/hiz/blit.vsh
@@ -0,0 +1,8 @@
+#version 450
+
+layout(location = 0) out vec2 uv;
+void main() {
+    vec2 corner = vec2[](vec2(0,0), vec2(1,0), vec2(1,1),vec2(0,1))[gl_VertexID];
+    uv = corner;
+    gl_Position = vec4(corner*2-1, 0, 1);
+}
\ No newline at end of file
diff --git a/src/main/resources/assets/voxy/shaders/lod/gl46mesh/cmdgen.comp b/src/main/resources/assets/voxy/shaders/lod/gl46mesh/cmdgen.comp
index 0094118c..a2b05aa6 100644
--- a/src/main/resources/assets/voxy/shaders/lod/gl46mesh/cmdgen.comp
+++ b/src/main/resources/assets/voxy/shaders/lod/gl46mesh/cmdgen.comp
@@ -7,7 +7,7 @@
 #import <voxy:lod/section.glsl>
 #define extractMeshletStart extractQuadStart
 layout(local_size_x = 64) in;
-#define QUADS_PER_MESHLET 62
+#define QUADS_PER_MESHLET 30
 
 void emitMeshlets(inout uint mli, inout uint meshletPtr, uint mskedCnt, uint cnt) {
     for (;mskedCnt != 0; mskedCnt--,mli++) {
diff --git a/src/main/resources/assets/voxy/shaders/lod/gl46mesh/quads.frag b/src/main/resources/assets/voxy/shaders/lod/gl46mesh/quads.frag
index 7732974e..8b244fa3 100644
--- a/src/main/resources/assets/voxy/shaders/lod/gl46mesh/quads.frag
+++ b/src/main/resources/assets/voxy/shaders/lod/gl46mesh/quads.frag
@@ -10,6 +10,7 @@ layout(location = 2) in flat vec4 tinting;
 layout(location = 3) in flat vec4 addin;
 layout(location = 4) in flat uint flags;
 layout(location = 5) in flat vec4 conditionalTinting;
+layout(location = 6) in flat uint meshlet;
 //layout(location = 6) in flat vec4 solidColour;
 
 layout(location = 0) out vec4 outColour;
@@ -29,4 +30,13 @@ void main() {
     outColour = (colour * tinting) + addin;
 
     //outColour = vec4(uv + baseUV, 0, 1);
+
+    /*
+    uint hash = meshlet*1231421+123141;
+    hash ^= hash>>16;
+    hash = hash*1231421+123141;
+    hash ^= hash>>16;
+
+    outColour = vec4(float(hash&15u)/15, float((hash>>4)&15u)/15, float((hash>>8)&15u)/15, 1);
+    */
 }
\ No newline at end of file
diff --git a/src/main/resources/assets/voxy/shaders/lod/gl46mesh/quads.vert b/src/main/resources/assets/voxy/shaders/lod/gl46mesh/quads.vert
index 8241f8d3..06f23145 100644
--- a/src/main/resources/assets/voxy/shaders/lod/gl46mesh/quads.vert
+++ b/src/main/resources/assets/voxy/shaders/lod/gl46mesh/quads.vert
@@ -1,6 +1,6 @@
 #version 450
 #extension GL_ARB_gpu_shader_int64 : enable
-#define QUADS_PER_MESHLET 62
+#define QUADS_PER_MESHLET 30
 
 #define MESHLET_ACCESS readonly
 //There are 16 bytes of metadata at the start of the meshlet
@@ -36,6 +36,7 @@ uint extractDetail(PosHeader pos) {
 }
 #endif
 
+layout(location = 6) out flat uint meshlet;
 PosHeader meshletPosition;
 Quad quad;
 bool setupMeshlet() {
@@ -49,6 +50,7 @@ bool setupMeshlet() {
         return true;
     }
 
+    meshlet = data;
     uint baseId = (data*MESHLET_SIZE);
     uint quadIndex = baseId + (gl_VertexID>>2) + 2;
     meshletPosition = geometryPool[baseId];