More direct buffers

- Merge object and batch ID buffer.
- ShaderCompiler accepts a Compilation callback.
- Use callback to enable the conservative depth extension only in
  fragment shaders.
- Query subgroup size if available and use callback to set a compile
  definition in compute shaders.

src/main/java/com/jozufozu/flywheel/backend/compile/Compile.java

@@ -6,11 +6,13 @@ import java.util.List;
 import java.util.Map;
 import java.util.function.BiConsumer;
 import java.util.function.BiFunction;
+import java.util.function.Consumer;
 import java.util.function.Function;
 
 import org.jetbrains.annotations.NotNull;
 import org.jetbrains.annotations.Nullable;
 
+import com.jozufozu.flywheel.backend.compile.core.Compilation;
 import com.jozufozu.flywheel.backend.compile.core.ProgramLinker;
 import com.jozufozu.flywheel.backend.compile.core.ShaderCompiler;
 import com.jozufozu.flywheel.gl.shader.GlProgram;
@@ -83,6 +85,8 @@ public class Compile {
 	public static class ShaderCompilerBuilder<K> {
 		private final GLSLVersion glslVersion;
 		private final ShaderType shaderType;
+		private Consumer<Compilation> compilationCallbacks = $ -> {
+		};
 		private final List<BiFunction<K, SourceLoader, SourceComponent>> fetchers = new ArrayList<>();
 
 		public ShaderCompilerBuilder(GLSLVersion glslVersion, ShaderType shaderType) {
@@ -111,6 +115,19 @@ public class Compile {
 			return withResource($ -> resourceLocation);
 		}
 
+		public ShaderCompilerBuilder<K> onCompile(Consumer<Compilation> cb) {
+			compilationCallbacks = compilationCallbacks.andThen(cb);
+			return this;
+		}
+
+		public ShaderCompilerBuilder<K> define(String def, int value) {
+			return onCompile(ctx -> ctx.define(def, String.valueOf(value)));
+		}
+
+		public ShaderCompilerBuilder<K> enableExtension(String extension) {
+			return onCompile(ctx -> ctx.enableExtension(extension));
+		}
+
 		@Nullable
 		private GlShader compile(K key, ShaderCompiler compiler, SourceLoader loader) {
 			var components = new ArrayList<SourceComponent>();
@@ -127,7 +144,7 @@ public class Compile {
 				return null;
 			}
 
-			return compiler.compile(glslVersion, shaderType, components);
+			return compiler.compile(glslVersion, shaderType, compilationCallbacks, components);
 		}
 	}
 }

src/main/java/com/jozufozu/flywheel/backend/compile/IndirectPrograms.java

@@ -12,6 +12,7 @@ import com.jozufozu.flywheel.api.vertex.VertexType;
 import com.jozufozu.flywheel.backend.compile.component.IndirectComponent;
 import com.jozufozu.flywheel.backend.compile.component.MaterialAdapterComponent;
 import com.jozufozu.flywheel.backend.compile.component.UniformComponent;
+import com.jozufozu.flywheel.gl.GlCompat;
 import com.jozufozu.flywheel.gl.shader.GlProgram;
 import com.jozufozu.flywheel.gl.shader.ShaderType;
 import com.jozufozu.flywheel.glsl.GLSLVersion;
@@ -75,6 +76,7 @@ public class IndirectPrograms {
 	private static CompilationHarness<InstanceType<?>> createCullingCompiler(UniformComponent uniformComponent, ShaderSources sources) {
 		return new CompilationHarness<>(sources, createCullingKeys(), Compile.<InstanceType<?>>program()
 				.link(Compile.<InstanceType<?>>shader(GLSLVersion.V460, ShaderType.COMPUTE)
+						.define("FLW_SUBGROUP_SIZE", GlCompat.SUBGROUP_SIZE)
 						.withComponent(uniformComponent)
 						.withComponent(IndirectComponent::create)
 						.withResource(InstanceType::instanceShader)

src/main/java/com/jozufozu/flywheel/backend/compile/PipelineCompiler.java

@@ -23,6 +23,7 @@ public class PipelineCompiler {
 								.vertexShader())
 						.withResource(pipeline.vertexShader()))
 				.link(Compile.<PipelineProgramKey>shader(pipeline.glslVersion(), ShaderType.FRAGMENT)
+						.enableExtension("GL_ARB_conservative_depth")
 						.withComponent(uniformComponent)
 						.withComponent(fragmentMaterialComponent)
 						.withResource(key -> key.contextShader()

src/main/java/com/jozufozu/flywheel/backend/compile/core/Compilation.java

@@ -70,6 +70,14 @@ public class Compilation {
 				.append(" : enable\n");
 	}
 
+	public void define(String key, String value) {
+		fullSource.append("#define ")
+				.append(key)
+				.append(' ')
+				.append(value)
+				.append('\n');
+	}
+
 	public void appendComponent(SourceComponent component) {
 		var source = component.source();
 

src/main/java/com/jozufozu/flywheel/backend/compile/core/ShaderCompiler.java

@@ -24,7 +24,7 @@ public class ShaderCompiler {
 	}
 
 	@Nullable
-	public GlShader compile(GLSLVersion glslVersion, ShaderType shaderType, List<SourceComponent> sourceComponents) {
+	public GlShader compile(GLSLVersion glslVersion, ShaderType shaderType, Consumer<Compilation> callback, List<SourceComponent> sourceComponents) {
 		var key = new ShaderKey(glslVersion, shaderType, sourceComponents);
 		var cached = shaderCache.get(key);
 		if (cached != null) {
@@ -32,7 +32,8 @@ public class ShaderCompiler {
 		}
 
 		Compilation ctx = new Compilation(glslVersion, shaderType);
-		ctx.enableExtension("GL_ARB_conservative_depth");
+
+		callback.accept(ctx);
 
 		expand(sourceComponents, ctx::appendComponent);
 

src/main/java/com/jozufozu/flywheel/backend/engine/indirect/IndirectBuffers.java

@@ -24,7 +24,7 @@ import com.jozufozu.flywheel.lib.memory.FlwMemoryTracker;
 import com.jozufozu.flywheel.lib.memory.MemoryBlock;
 
 public class IndirectBuffers {
-	public static final int BUFFER_COUNT = 4;
+	public static final int BUFFER_COUNT = 3;
 	public static final long INT_SIZE = Integer.BYTES;
 	public static final long PTR_SIZE = Pointer.POINTER_SIZE;
 
@@ -45,18 +45,15 @@ public class IndirectBuffers {
 
 	private static final long OBJECT_SIZE_OFFSET = SIZE_OFFSET;
 	private static final long TARGET_SIZE_OFFSET = OBJECT_SIZE_OFFSET + PTR_SIZE;
-	private static final long BATCH_SIZE_OFFSET = TARGET_SIZE_OFFSET + PTR_SIZE;
+	private static final long DRAW_SIZE_OFFSET = TARGET_SIZE_OFFSET + PTR_SIZE;
-	private static final long DRAW_SIZE_OFFSET = BATCH_SIZE_OFFSET + PTR_SIZE;
 
 	private final MemoryBlock buffers;
 	private final long objectStride;
 	private int object;
 	private int target;
-	private int batch;
 	private int draw;
 
 	long objectPtr;
-	long batchPtr;
 	long drawPtr;
 
 	private int maxObjectCount = 0;
@@ -72,11 +69,10 @@ public class IndirectBuffers {
 
 	void createBuffers() {
 		final long ptr = buffers.ptr();
-		nglCreateBuffers(4, ptr);
+		nglCreateBuffers(BUFFER_COUNT, ptr);
 		object = MemoryUtil.memGetInt(ptr);
 		target = MemoryUtil.memGetInt(ptr + 4);
-		batch = MemoryUtil.memGetInt(ptr + 8);
+		draw = MemoryUtil.memGetInt(ptr + 8);
-		draw = MemoryUtil.memGetInt(ptr + 12);
 	}
 
 	void updateCounts(int objectCount, int drawCount) {
@@ -94,7 +90,6 @@ public class IndirectBuffers {
 		final long ptr = buffers.ptr();
 		MemoryUtil.memPutAddress(ptr + OBJECT_SIZE_OFFSET, objectSize);
 		MemoryUtil.memPutAddress(ptr + TARGET_SIZE_OFFSET, targetSize);
-		MemoryUtil.memPutAddress(ptr + BATCH_SIZE_OFFSET, targetSize);
 		MemoryUtil.memPutAddress(ptr + DRAW_SIZE_OFFSET, drawSize);
 	}
 
@@ -105,38 +100,31 @@ public class IndirectBuffers {
 
 		if (maxObjectCount > 0) {
 			final long ptr = buffers.ptr();
-			nglCreateBuffers(3, ptr);
+			nglCreateBuffers(BUFFER_COUNT - 1, ptr);
 
 			int objectNew = MemoryUtil.memGetInt(ptr);
 			int targetNew = MemoryUtil.memGetInt(ptr + 4);
-			int batchNew = MemoryUtil.memGetInt(ptr + 8);
 
 			glNamedBufferStorage(objectNew, objectSize, PERSISTENT_BITS);
 			glNamedBufferStorage(targetNew, targetSize, GPU_ONLY_BITS);
-			glNamedBufferStorage(batchNew, targetSize, PERSISTENT_BITS);
 
 			glCopyNamedBufferSubData(object, objectNew, 0, 0, objectStride * maxObjectCount);
 			glCopyNamedBufferSubData(target, targetNew, 0, 0, INT_SIZE * maxObjectCount);
-			glCopyNamedBufferSubData(batch, batchNew, 0, 0, INT_SIZE * maxObjectCount);
 
 			glDeleteBuffers(object);
 			glDeleteBuffers(target);
-			glDeleteBuffers(batch);
 
 			object = objectNew;
 			target = targetNew;
-			batch = batchNew;
 		} else {
 			glNamedBufferStorage(object, objectSize, PERSISTENT_BITS);
 			glNamedBufferStorage(target, targetSize, GPU_ONLY_BITS);
-			glNamedBufferStorage(batch, targetSize, PERSISTENT_BITS);
 		}
 
 		objectPtr = nglMapNamedBufferRange(object, 0, objectSize, MAP_BITS);
-		batchPtr = nglMapNamedBufferRange(batch, 0, targetSize, MAP_BITS);
 		maxObjectCount = objectCount;
 
-		FlwMemoryTracker._allocGPUMemory(maxObjectCount * objectStride + maxObjectCount * INT_SIZE);
+		FlwMemoryTracker._allocGPUMemory(maxObjectCount * objectStride);
 	}
 
 	void createDrawStorage(int drawCount) {
@@ -150,7 +138,7 @@ public class IndirectBuffers {
 
 			glDeleteBuffers(draw);
 
-			MemoryUtil.memPutInt(buffers.ptr() + INT_SIZE * 3, drawNew);
+			MemoryUtil.memPutInt(buffers.ptr() + INT_SIZE * 2, drawNew);
 			draw = drawNew;
 			drawPtr = MemoryUtil.nmemRealloc(drawPtr, drawSize);
 		} else {
@@ -163,7 +151,7 @@ public class IndirectBuffers {
 	}
 
 	private void freeObjectStogare() {
-		FlwMemoryTracker._freeGPUMemory(maxObjectCount * objectStride + maxObjectCount * INT_SIZE);
+		FlwMemoryTracker._freeGPUMemory(maxObjectCount * objectStride);
 	}
 
 	private void freeDrawStorage() {
@@ -184,10 +172,6 @@ public class IndirectBuffers {
 		nglBindBuffersRange(GL_SHADER_STORAGE_BUFFER, 0, IndirectBuffers.BUFFER_COUNT, ptr, ptr + OFFSET_OFFSET, ptr + SIZE_OFFSET);
 	}
 
-	void flushBatchIDs(long length) {
-		glFlushMappedNamedBufferRange(batch, 0, length);
-	}
-
 	void flushObjects(long length) {
 		glFlushMappedNamedBufferRange(object, 0, length);
 	}

src/main/java/com/jozufozu/flywheel/backend/engine/indirect/IndirectCullingGroup.java

@@ -21,7 +21,7 @@ public class IndirectCullingGroup<I extends Instance> {
 
 	private final GlProgram compute;
 	private final GlProgram draw;
-	private final long instanceStride;
+	private final long objectStride;
 	private final IndirectBuffers buffers;
 	public final IndirectMeshPool meshPool;
 	public final IndirectDrawSet<I> drawSet = new IndirectDrawSet<>();
@@ -30,9 +30,10 @@ public class IndirectCullingGroup<I extends Instance> {
 	private int instanceCountThisFrame;
 
 	IndirectCullingGroup(InstanceType<I> instanceType, VertexType vertexType) {
-		instanceStride = instanceType.getLayout()
+		objectStride = instanceType.getLayout()
-				.getStride();
+				.getStride() + IndirectBuffers.INT_SIZE;
-		buffers = new IndirectBuffers(instanceStride);
+
+		buffers = new IndirectBuffers(objectStride);
 		buffers.createBuffers();
 		buffers.createObjectStorage(128);
 		buffers.createDrawStorage(2);
@@ -108,20 +109,17 @@ public class IndirectCullingGroup<I extends Instance> {
 
 	private void uploadInstances() {
 		long objectPtr = buffers.objectPtr;
-		long batchIDPtr = buffers.batchPtr;
 
 		for (int i = 0, batchesSize = drawSet.indirectDraws.size(); i < batchesSize; i++) {
 			var batch = drawSet.indirectDraws.get(i);
 			var instanceCount = batch.instancer()
 					.getInstanceCount();
-			batch.writeObjects(objectPtr, batchIDPtr, i);
+			batch.writeObjects(objectPtr, i);
 
-			objectPtr += instanceCount * instanceStride;
+			objectPtr += instanceCount * objectStride;
-			batchIDPtr += instanceCount * IndirectBuffers.INT_SIZE;
 		}
 
 		buffers.flushObjects(objectPtr - buffers.objectPtr);
-		buffers.flushBatchIDs(batchIDPtr - buffers.batchPtr);
 	}
 
 	private void uploadIndirectCommands() {

src/main/java/com/jozufozu/flywheel/backend/engine/indirect/IndirectDraw.java

@@ -55,11 +55,11 @@ public class IndirectDraw<I extends Instance> {
 		needsFullWrite = true;
 	}
 
-	public void writeObjects(long objectPtr, long batchIDPtr, int batchID) {
+	public void writeObjects(long objectPtr, int batchID) {
 		if (needsFullWrite) {
-			instancer.writeFull(objectPtr, batchIDPtr, batchID);
+			instancer.writeFull(objectPtr, batchID);
 		} else {
-			instancer.writeSparse(objectPtr, batchIDPtr, batchID);
+			instancer.writeSparse(objectPtr, batchID);
 		}
 	}
 

src/main/java/com/jozufozu/flywheel/backend/engine/indirect/IndirectInstancer.java

@@ -9,40 +9,42 @@ import com.jozufozu.flywheel.backend.engine.AbstractInstancer;
 
 public class IndirectInstancer<I extends Instance> extends AbstractInstancer<I> {
 	private final long instanceStride;
+	private final long objectStride;
 
 	public IndirectInstancer(InstanceType<I> type) {
 		super(type);
 		this.instanceStride = type.getLayout()
 				.getStride();
+		this.objectStride = instanceStride + IndirectBuffers.INT_SIZE;
 	}
 
 	public void update() {
 		removeDeletedInstances();
 	}
 
-	public void writeSparse(long objectPtr, long batchIDPtr, int batchID) {
+	public void writeSparse(long objectPtr, int batchID) {
 		int count = instances.size();
 		InstanceWriter<I> writer = type.getWriter();
 		for (int i = changed.nextSetBit(0); i >= 0 && i < count; i = changed.nextSetBit(i + 1)) {
-			// write object
+			long ptr = objectPtr + objectStride * i;
-			writer.write(objectPtr + instanceStride * i, instances.get(i));
-
 			// write batchID
-			MemoryUtil.memPutInt(batchIDPtr + IndirectBuffers.INT_SIZE * i, batchID);
+			MemoryUtil.memPutInt(ptr, batchID);
+			// write object
+			writer.write(ptr + IndirectBuffers.INT_SIZE, instances.get(i));
 		}
 		changed.clear();
 	}
 
-	public void writeFull(long objectPtr, long batchIDPtr, int batchID) {
+	public void writeFull(long objectPtr, int batchID) {
 		InstanceWriter<I> writer = type.getWriter();
 		for (I object : instances) {
+			// write batchID
+			MemoryUtil.memPutInt(objectPtr, batchID);
+			objectPtr += IndirectBuffers.INT_SIZE;
+
 			// write object
 			writer.write(objectPtr, object);
 			objectPtr += instanceStride;
-
-			// write batchID
-			MemoryUtil.memPutInt(batchIDPtr, batchID);
-			batchIDPtr += IndirectBuffers.INT_SIZE;
 		}
 		changed.clear();
 	}

src/main/java/com/jozufozu/flywheel/gl/GlCompat.java

@@ -5,7 +5,9 @@ import java.nio.ByteBuffer;
 import org.lwjgl.PointerBuffer;
 import org.lwjgl.opengl.GL;
 import org.lwjgl.opengl.GL20C;
+import org.lwjgl.opengl.GL31C;
 import org.lwjgl.opengl.GLCapabilities;
+import org.lwjgl.opengl.KHRShaderSubgroup;
 import org.lwjgl.system.MemoryStack;
 
 import net.minecraft.Util;
@@ -19,14 +21,16 @@ import net.minecraft.Util;
 public class GlCompat {
 	public static final boolean ALLOW_DSA = true;
 	public static final GLCapabilities CAPABILITIES = GL.createCapabilities();
-	private static final boolean amd = _decideIfWeAreAMDWindows();
+	private static final boolean amd = _decideIfWeAreAMD();
+	private static final boolean windows = _decideIfWeAreWindows();
 	private static final boolean supportsIndirect = _decideIfWeSupportIndirect();
+	public static final int SUBGROUP_SIZE = _subgroupSize();
 
 	private GlCompat() {
 	}
 
 	public static boolean onAMDWindows() {
-		return amd;
+		return amd && windows;
 	}
 
 	public static boolean supportsInstancing() {
@@ -41,6 +45,14 @@ public class GlCompat {
 		return CAPABILITIES.OpenGL46 || (CAPABILITIES.GL_ARB_compute_shader && CAPABILITIES.GL_ARB_shader_draw_parameters && CAPABILITIES.GL_ARB_base_instance && CAPABILITIES.GL_ARB_multi_draw_indirect && CAPABILITIES.GL_ARB_direct_state_access);
 	}
 
+	private static int _subgroupSize() {
+		if (CAPABILITIES.GL_KHR_shader_subgroup) {
+			return GL31C.glGetInteger(KHRShaderSubgroup.GL_SUBGROUP_SIZE_KHR);
+		}
+		// try to guess
+		return amd ? 64 : 32;
+	}
+
 	/**
 	 * Modified from:
 	 * <br> <a href="https://github.com/grondag/canvas/commit/820bf754092ccaf8d0c169620c2ff575722d7d96">canvas</a>
@@ -62,11 +74,11 @@ public class GlCompat {
 		}
 	}
 
-	private static boolean _decideIfWeAreAMDWindows() {
+	private static boolean _decideIfWeAreWindows() {
-		if (Util.getPlatform() != Util.OS.WINDOWS) {
+		return Util.getPlatform() == Util.OS.WINDOWS;
-			return false;
+	}
-		}
 
+	private static boolean _decideIfWeAreAMD() {
 		String vendor = GL20C.glGetString(GL20C.GL_VENDOR);
 
 		if (vendor == null) {

src/main/resources/assets/flywheel/flywheel/internal/indirect_cull.glsl

@@ -1,22 +1,22 @@
-#define FLW_SUBGROUP_SIZE 32
 layout(local_size_x = FLW_SUBGROUP_SIZE) in;
 
 #include "flywheel:internal/indirect_draw_command.glsl"
 
+struct Object {
+    uint batchID;
+    FlwPackedInstance instance;
+};
+
 // populated by instancers
 layout(std430, binding = 0) restrict readonly buffer ObjectBuffer {
-    FlwPackedInstance objects[];
+    Object objects[];
 };
 
 layout(std430, binding = 1) restrict writeonly buffer TargetBuffer {
     uint objectIDs[];
 };
 
-layout(std430, binding = 2) restrict readonly buffer BatchBuffer {
+layout(std430, binding = 2) restrict buffer DrawCommands {
-    uint batchIDs[];
-};
-
-layout(std430, binding = 3) restrict buffer DrawCommands {
     MeshDrawCommand drawCommands[];
 };
 
@@ -38,7 +38,7 @@ bool isVisible() {
     float radius;
     unpackBoundingSphere(sphere, center, radius);
 
-    FlwInstance object = _flw_unpackInstance(objects[flw_objectID]);
+    FlwInstance object = _flw_unpackInstance(objects[flw_objectID].instance);
     flw_transformBoundingSphere(object, center, radius);
 
     return testSphere(center, radius);
@@ -51,7 +51,7 @@ void main() {
         return;
     }
 
-    flw_batchID = batchIDs[flw_objectID];
+    flw_batchID = objects[flw_objectID].batchID;
 
     if (isVisible()) {
         uint batchIndex = atomicAdd(drawCommands[flw_batchID].instanceCount, 1);

src/main/resources/assets/flywheel/flywheel/internal/indirect_draw.vert

@@ -1,26 +1,28 @@
 #include "flywheel:api/vertex.glsl"
 #include "flywheel:internal/indirect_draw_command.glsl"
 
+struct Object {
+    uint batchID;
+    FlwPackedInstance instance;
+};
+
+
 layout(std430, binding = 0) restrict readonly buffer ObjectBuffer {
-    FlwPackedInstance objects[];
+    Object objects[];
 };
 
 layout(std430, binding = 1) restrict readonly buffer TargetBuffer {
     uint objectIDs[];
 };
 
-layout(std430, binding = 2) restrict readonly buffer BatchBuffer {
+layout(std430, binding = 2) restrict readonly buffer DrawCommands {
-    uint batchIDs[];
-};
-
-layout(std430, binding = 3) restrict readonly buffer DrawCommands {
     MeshDrawCommand drawCommands[];
 };
 
 void main() {
     uint instanceIndex = objectIDs[gl_BaseInstance + gl_InstanceID];
-    uint batchID = batchIDs[instanceIndex];
+    uint batchID = objects[instanceIndex].batchID;
-    FlwInstance i = _flw_unpackInstance(objects[instanceIndex]);
+    FlwInstance i = _flw_unpackInstance(objects[instanceIndex].instance);
 
     _flw_materialVertexID = drawCommands[batchID].vertexMaterialID;
     _flw_materialFragmentID = drawCommands[batchID].fragmentMaterialID;