From cce96fc86791f5e6a6c4301b300bdae1ac82d11e Mon Sep 17 00:00:00 2001
From: mcrcortex <{ID}+{username}@users.noreply.github.com>
Date: Sun, 14 Apr 2024 04:05:07 +1000
Subject: [PATCH] made it 62 quad in size

---
 .../Gl46MeshletsFarWorldRenderer.java         |  9 +++---
 .../core/rendering/SharedIndexBuffer.java     | 14 +++++++++
 .../rendering/building/RenderDataFactory.java | 31 ++++++++++---------
 .../voxy/client/core/util/IndexUtil.java      | 19 ++++++++++++
 .../voxy/shaders/lod/gl46mesh/cmdgen.comp     |  2 +-
 .../voxy/shaders/lod/gl46mesh/quads.vert      |  2 +-
 6 files changed, 55 insertions(+), 22 deletions(-)

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 99b79e2a..769bd7b3 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
@@ -58,7 +58,7 @@ public class Gl46MeshletsFarWorldRenderer extends AbstractFarWorldRenderer<Gl46M
     private final GlBuffer meshletBuffer;
 
     public Gl46MeshletsFarWorldRenderer(int geometrySize, int maxSections) {
-        super(new DefaultGeometryManager(alignUp(geometrySize*8L, 8*128), maxSections, 8*128));
+        super(new DefaultGeometryManager(alignUp(geometrySize*8L, 8*64), maxSections, 8*64));
         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)
     }
@@ -74,7 +74,7 @@ public class Gl46MeshletsFarWorldRenderer extends AbstractFarWorldRenderer<Gl46M
         glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 7, this.models.getColourBufferId());
         glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 8, this.lightDataBuffer.id);//Lighting LUT
         glBindBuffer(GL_DRAW_INDIRECT_BUFFER, bindToDrawIndirect?this.glDrawIndirect.id:0);
-        glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, SharedIndexBuffer.INSTANCE.id());
+        glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, SharedIndexBuffer.INSTANCE_BYTE.id());
 
         //Bind the texture atlas
         glBindSampler(0, this.models.getSamplerId());
@@ -132,8 +132,7 @@ public class Gl46MeshletsFarWorldRenderer extends AbstractFarWorldRenderer<Gl46M
         this.lodShader.bind();
         this.bindResources(viewport, true);
         glDisable(GL_CULL_FACE);
-        //glDrawElementsInstanced(GL_TRIANGLES, 6*126, GL_UNSIGNED_SHORT, 0, 40000);
-        glDrawElementsIndirect(GL_TRIANGLES, GL_UNSIGNED_SHORT, 0);
+        glDrawElementsIndirect(GL_TRIANGLES, GL_UNSIGNED_BYTE, 0);
         glEnable(GL_CULL_FACE);
 
         glMemoryBarrier(GL_PIXEL_BUFFER_BARRIER_BIT | GL_FRAMEBUFFER_BARRIER_BIT);
@@ -142,7 +141,7 @@ public class Gl46MeshletsFarWorldRenderer extends AbstractFarWorldRenderer<Gl46M
         this.bindResources(viewport, false);
         glDepthMask(false);
         glColorMask(false, false, false, false);
-        glDrawElementsInstanced(GL_TRIANGLES, 6 * 2 * 3, GL_UNSIGNED_BYTE, (1 << 16) * 6 * 2, this.geometry.getSectionCount());
+        glDrawElementsInstanced(GL_TRIANGLES, 6 * 2 * 3, GL_UNSIGNED_BYTE, (1 << 8) * 6, this.geometry.getSectionCount());
         glColorMask(true, true, true, true);
         glDepthMask(true);
 
diff --git a/src/main/java/me/cortex/voxy/client/core/rendering/SharedIndexBuffer.java b/src/main/java/me/cortex/voxy/client/core/rendering/SharedIndexBuffer.java
index 8dcab04d..c234bd90 100644
--- a/src/main/java/me/cortex/voxy/client/core/rendering/SharedIndexBuffer.java
+++ b/src/main/java/me/cortex/voxy/client/core/rendering/SharedIndexBuffer.java
@@ -10,6 +10,7 @@ import org.lwjgl.system.MemoryUtil;
 //Has a base index buffer of 16380 quads, and also a 1 cube byte index buffer at the end
 public class SharedIndexBuffer {
     public static final SharedIndexBuffer INSTANCE = new SharedIndexBuffer();
+    public static final SharedIndexBuffer INSTANCE_BYTE = new SharedIndexBuffer(true);
 
     private final GlBuffer indexBuffer;
 
@@ -26,6 +27,19 @@ public class SharedIndexBuffer {
         cubeBuff.free();
     }
 
+    private SharedIndexBuffer(boolean type2) {
+        this.indexBuffer = new GlBuffer((1<<8)*6 + 6*2*3);
+        var quadIndexBuff = IndexUtil.generateQuadIndicesByte(63);
+        var cubeBuff = generateCubeIndexBuffer();
+
+        long ptr = UploadStream.INSTANCE.upload(this.indexBuffer, 0, this.indexBuffer.size());
+        quadIndexBuff.cpyTo(ptr);
+        cubeBuff.cpyTo((1<<8)*6 + ptr);
+
+        quadIndexBuff.free();
+        cubeBuff.free();
+    }
+
     private static MemoryBuffer generateCubeIndexBuffer() {
         var buffer = new MemoryBuffer(6*2*3);
         long ptr = buffer.address;
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 724e8d36..eec4d928 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
@@ -53,6 +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 void writePos(long ptr, long pos) {
         if (USE_UINT64) {
             MemoryUtil.memPutLong(ptr, pos);
@@ -88,10 +89,10 @@ public class RenderDataFactory {
 
         int bufferSize;
         if (this.generateMeshlets) {
-            bufferSize = getMeshletHoldingCount(this.doubleSidedQuadCollector.size(), 126, 128) +
-                    getMeshletHoldingCount(this.translucentQuadCollector.size(), 126, 128);
+            bufferSize = getMeshletHoldingCount(this.doubleSidedQuadCollector.size(), QUADS_PER_MESHLET, QUADS_PER_MESHLET+2) +
+                    getMeshletHoldingCount(this.translucentQuadCollector.size(), QUADS_PER_MESHLET, QUADS_PER_MESHLET+2);
             for (var collector : this.directionalQuadCollectors) {
-                bufferSize += getMeshletHoldingCount(collector.size(), 126, 128);
+                bufferSize += getMeshletHoldingCount(collector.size(), QUADS_PER_MESHLET, QUADS_PER_MESHLET+2);
             }
         } else {
             bufferSize = this.doubleSidedQuadCollector.size() + this.translucentQuadCollector.size();
@@ -122,11 +123,11 @@ public class RenderDataFactory {
                     //Write out meshlet header
 
                     //Write out the section position
-                    writePos(ptr + meshlet * 8L * 128L, key);
-                    MemoryUtil.memPutLong(ptr + meshlet * 8L * 128L + 8, 0);
+                    writePos(ptr + meshlet * 8L * (QUADS_PER_MESHLET+2), key);
+                    MemoryUtil.memPutLong(ptr + meshlet * 8L * (QUADS_PER_MESHLET+2) + 8, 0);
                 }
-                MemoryUtil.memPutLong(ptr + meshlet * 8L * 128 + (2 + innerQuadCount++) * 8L, data);
-                if (innerQuadCount == 126) {
+                MemoryUtil.memPutLong(ptr + meshlet * 8L * (QUADS_PER_MESHLET+2) + (2 + innerQuadCount++) * 8L, data);
+                if (innerQuadCount == QUADS_PER_MESHLET) {
                     innerQuadCount = 0;
                     meshlet++;
                 }
@@ -143,11 +144,11 @@ public class RenderDataFactory {
                     //Write out meshlet header
 
                     //Write out the section position
-                    writePos(ptr + meshlet * 8L * 128L, key);
-                    MemoryUtil.memPutLong(ptr + meshlet * 8L * 128L + 8, 0);
+                    writePos(ptr + meshlet * 8L * (QUADS_PER_MESHLET+2), key);
+                    MemoryUtil.memPutLong(ptr + meshlet * 8L * (QUADS_PER_MESHLET+2) + 8, 0);
                 }
-                MemoryUtil.memPutLong(ptr + meshlet * 8L * 128 + (2 + innerQuadCount++) * 8L, data);
-                if (innerQuadCount == 126) {
+                MemoryUtil.memPutLong(ptr + meshlet * 8L * (QUADS_PER_MESHLET+2) + (2 + innerQuadCount++) * 8L, data);
+                if (innerQuadCount == QUADS_PER_MESHLET) {
                     innerQuadCount = 0;
                     meshlet++;
                 }
@@ -165,11 +166,11 @@ public class RenderDataFactory {
                         //Write out meshlet header
 
                         //Write out the section position
-                        writePos(ptr + meshlet * 8L * 128L, key);
-                        MemoryUtil.memPutLong(ptr + meshlet * 8L * 128L + 8, 0);
+                        writePos(ptr + meshlet * 8L * (QUADS_PER_MESHLET+2), key);
+                        MemoryUtil.memPutLong(ptr + meshlet * 8L * (QUADS_PER_MESHLET+2) + 8, 0);
                     }
-                    MemoryUtil.memPutLong(ptr + meshlet * 8L * 128 + (2 + innerQuadCount++) * 8L, data);
-                    if (innerQuadCount == 126) {
+                    MemoryUtil.memPutLong(ptr + meshlet * 8L * (QUADS_PER_MESHLET+2) + (2 + innerQuadCount++) * 8L, data);
+                    if (innerQuadCount == QUADS_PER_MESHLET) {
                         innerQuadCount = 0;
                         meshlet++;
                     }
diff --git a/src/main/java/me/cortex/voxy/client/core/util/IndexUtil.java b/src/main/java/me/cortex/voxy/client/core/util/IndexUtil.java
index 57754cab..91e4924c 100644
--- a/src/main/java/me/cortex/voxy/client/core/util/IndexUtil.java
+++ b/src/main/java/me/cortex/voxy/client/core/util/IndexUtil.java
@@ -4,6 +4,25 @@ import me.cortex.voxy.common.util.MemoryBuffer;
 import org.lwjgl.system.MemoryUtil;
 
 public class IndexUtil {
+    public static MemoryBuffer generateQuadIndicesByte(int quadCount) {
+        if ((quadCount*4) >= 1<<8) {
+            throw new IllegalArgumentException("Quad count to large");
+        }
+        MemoryBuffer buffer = new MemoryBuffer(quadCount * 6L);
+        long ptr = buffer.address;
+        for(int i = 0; i < quadCount*4; i += 4) {
+            MemoryUtil.memPutByte(ptr + (0), (byte) i);
+            MemoryUtil.memPutByte(ptr + (1), (byte) (i + 1));
+            MemoryUtil.memPutByte(ptr + (2), (byte) (i + 2));
+            MemoryUtil.memPutByte(ptr + (3), (byte) (i + 1));
+            MemoryUtil.memPutByte(ptr + (4), (byte) (i + 3));
+            MemoryUtil.memPutByte(ptr + (5), (byte) (i + 2));
+
+            ptr += 6;
+        }
+
+        return buffer;
+    }
     public static MemoryBuffer generateQuadIndicesShort(int quadCount) {
         if ((quadCount*4) >= 1<<16) {
             throw new IllegalArgumentException("Quad count to large");
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 a00d0047..0094118c 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 126
+#define QUADS_PER_MESHLET 62
 
 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.vert b/src/main/resources/assets/voxy/shaders/lod/gl46mesh/quads.vert
index 8db9ceb7..8241f8d3 100644
--- a/src/main/resources/assets/voxy/shaders/lod/gl46mesh/quads.vert
+++ b/src/main/resources/assets/voxy/shaders/lod/gl46mesh/quads.vert
@@ -1,8 +1,8 @@
 #version 450
 #extension GL_ARB_gpu_shader_int64 : enable
+#define QUADS_PER_MESHLET 62
 
 #define MESHLET_ACCESS readonly
-#define QUADS_PER_MESHLET 126
 //There are 16 bytes of metadata at the start of the meshlet
 #define MESHLET_SIZE (QUADS_PER_MESHLET+2)
 #import <voxy:lod/quad_format.glsl>