Optimize renderer

client/level/client_level.ts

@@ -12,7 +12,7 @@ import { AIR, CHUNK_AREA, CHUNK_HEIGHT, CHUNK_SIZE, Faces, ID_MASK, VOID } from
 import { AssetManager } from "../assets.ts";
 import { ChunkWorkerPool } from "../chunk_workers.ts";
 import { worldgen_mods } from "../mods.ts";
-import { type FromChunkWorker, TERRAIN_VERTEX_FLOATS } from "../workers/chunk_messages.ts";
+import type { FaceGroup, FromChunkWorker } from "../workers/chunk_messages.ts";
 import { create_index_buffer, create_vertex_buffer, destroy_buffer, Texture } from "../renderer/mod.ts";
 import { crosses_planes } from "../workers/translucent_sort.ts";
 import { Camera } from "../camera.ts";
@@ -37,6 +37,9 @@ export interface Chunk {
 	// bumped on every mesh request so late results from older requests get ignored
 	mesh_version: number;
 	meshes: Partial<Record<RenderLayer, ChunkMesh>>;
+	// the lowest and highest y of its meshes, for frustum culling
+	min_y: number;
+	max_y: number;
 	// only for translucent meshes that have to be sorted again as the camera moves
 	translucent_sort?: TranslucentSort;
 	// blocks its generation put in neighboring chunks (leaves), as x, y, z, numeric id. kept so a neighbor that
@@ -49,6 +52,8 @@ export interface ChunkMesh {
 	quad_count: number;
 	// the translucent layer's quads sorted back to front, the others are drawn in order
 	index_buffer?: GPUBuffer;
+	// solid and cutout quads by the way they face, see FACE_GROUPS in chunk_messages.ts
+	groups?: FaceGroup[];
 }
 
 interface TranslucentSort {
@@ -63,8 +68,6 @@ interface TranslucentSort {
 	applied: number;
 }
 
-const FLOATS_PER_QUAD = 4 * TERRAIN_VERTEX_FLOATS;
-
 export { chunk_key };
 
 // the block being looked at and which face of it
@@ -199,6 +202,8 @@ export class ClientLevel {
 			generated: false,
 			mesh_version: 0,
 			meshes: {},
+			min_y: 0,
+			max_y: 0,
 		};
 		this.chunks.set(chunk_key(x, z), chunk);
 		return chunk;
@@ -548,16 +553,15 @@ export class ClientLevel {
 		}
 
 		this.delete_chunk_mesh(chunk);
+		chunk.min_y = message.min_y;
+		chunk.max_y = message.max_y;
 
 		for (const layer of RENDER_LAYERS) {
-			const { vertices, quad_count } = message[layer];
+			const { vertices, quad_count, groups } = message[layer];
 			if (quad_count === 0) {
 				continue;
 			}
-			chunk.meshes[layer] = {
-				vertex_buffer: create_vertex_buffer(vertices.subarray(0, quad_count * FLOATS_PER_QUAD)),
-				quad_count,
-			};
+			chunk.meshes[layer] = { vertex_buffer: create_vertex_buffer(vertices), quad_count, groups };
 		}
 
 		const translucent = message.translucent;

client/renderer/core.ts

@@ -1,7 +1,7 @@
 import type { Camera } from "../camera.ts";
 import { mat4 } from "gl-matrix";
 import type { RenderLayer } from "$/common/everything_registry.ts";
-import { TERRAIN_VERTEX_FLOATS } from "../workers/chunk_messages.ts";
+import { TERRAIN_VERTEX_BYTES } from "../workers/chunk_messages.ts";
 
 export let device: GPUDevice;
 export let canvas: HTMLCanvasElement;
@@ -375,10 +375,16 @@ export function flush_batch() {
 	vert_index = 0;
 }
 
-// draws a chunk mesh made of quads (4 vertices each). without an index buffer the quads are drawn in order
+// what's bound for terrain draws in the current pass, so drawing hundreds of chunks doesn't bind the same things
+// again for each one. cleared whenever the pipeline changes
+let terrain_binds: { slot: number; texture: GPUTexture; index_buffer: GPUBuffer } | undefined;
+
+// draws quads first to first + quad_count of a chunk mesh (4 vertices each). without an index buffer the quads are
+// drawn in order
 export function draw_terrain(
 	layer: RenderLayer,
 	vertex_buffer: GPUBuffer,
+	first: number,
 	quad_count: number,
 	index_buffer?: GPUBuffer,
 ) {
@@ -386,7 +392,7 @@ export function draw_terrain(
 		return;
 	}
 	if (!index_buffer) {
-		ensure_quad_indices(quad_count);
+		ensure_quad_indices(first + quad_count);
 		index_buffer = quad_index_buffer!;
 	}
 
@@ -395,17 +401,33 @@ export function draw_terrain(
 	if (current_pipeline !== pipeline) {
 		render_pass.setPipeline(pipeline);
 		current_pipeline = pipeline;
+		terrain_binds = undefined;
 	}
 
-	render_pass.setBindGroup(0, uniform_bind_group, [uniform_slot * UNIFORM_SLOT_SIZE]);
-	render_pass.setBindGroup(1, get_texture_bind_group(current_texture));
-	render_pass.setBindGroup(2, lightmap_bind_group);
+	if (!terrain_binds) {
+		render_pass.setBindGroup(2, lightmap_bind_group);
+	}
+	if (terrain_binds?.slot !== uniform_slot) {
+		render_pass.setBindGroup(0, uniform_bind_group, [uniform_slot * UNIFORM_SLOT_SIZE]);
+	}
+	if (terrain_binds?.texture !== current_texture) {
+		render_pass.setBindGroup(1, get_texture_bind_group(current_texture));
+	}
+	if (terrain_binds?.index_buffer !== index_buffer) {
+		render_pass.setIndexBuffer(index_buffer, "uint32");
+	}
+	terrain_binds = { slot: uniform_slot, texture: current_texture, index_buffer };
+
 	render_pass.setVertexBuffer(0, vertex_buffer);
-	render_pass.setIndexBuffer(index_buffer, "uint32");
-	render_pass.drawIndexed(quad_count * 6);
+	render_pass.drawIndexed(quad_count * 6, 1, first * 6);
 }
 
-export function create_vertex_buffer(vertices: Float32Array): GPUBuffer {
+// the camera's view and projection, for frustum culling. only meaningful in 3d mode
+export function view_projection(): Readonly<Float32Array> {
+	return mvp as Float32Array;
+}
+
+export function create_vertex_buffer(vertices: Float32Array<ArrayBuffer> | Uint8Array<ArrayBuffer>): GPUBuffer {
 	const buffer = device.createBuffer({
 		size: Math.max(4, vertices.byteLength),
 		usage: GPUBufferUsage.VERTEX | GPUBufferUsage.COPY_DST,
@@ -598,6 +620,7 @@ function ensure_pass(): GPURenderPassEncoder {
 	pending_color_clear = false;
 	pending_depth_clear = false;
 	current_pipeline = undefined;
+	terrain_binds = undefined;
 	apply_scissor(pass);
 
 	return pass;
@@ -762,12 +785,12 @@ function create_pipelines() {
 		module: terrain_module,
 		entryPoint: "vs_terrain",
 		buffers: [{
-			arrayStride: TERRAIN_VERTEX_FLOATS * 4,
+			arrayStride: TERRAIN_VERTEX_BYTES,
 			attributes: [
 				{ shaderLocation: 0, offset: 0, format: "float32x3" },
-				{ shaderLocation: 1, offset: 12, format: "float32x2" },
-				{ shaderLocation: 2, offset: 20, format: "float32x4" },
-				{ shaderLocation: 3, offset: 36, format: "float32x2" },
+				{ shaderLocation: 1, offset: 12, format: "unorm16x2" },
+				{ shaderLocation: 2, offset: 16, format: "unorm8x4" },
+				{ shaderLocation: 3, offset: 20, format: "unorm8x2" },
 			],
 		}],
 	};

client/rendering/frustum.ts

@@ -0,0 +1,41 @@
+// what the camera can see: the six planes around it, taken from its view projection matrix (gl-matrix's column major
+// order, with webgpu's 0 to 1 depth). a box is out of view when it's entirely behind one of them
+export class Frustum {
+	// a, b, c, d per plane, where a * x + b * y + c * z + d >= 0 is the inside
+	#planes = new Float32Array(24);
+
+	update(m: Readonly<Float32Array>) {
+		const planes = this.#planes;
+		const set = (i: number, a: number, b: number, c: number, d: number) => {
+			planes[i * 4] = a;
+			planes[i * 4 + 1] = b;
+			planes[i * 4 + 2] = c;
+			planes[i * 4 + 3] = d;
+		};
+		// row r of the matrix is m[r], m[4 + r], m[8 + r], m[12 + r]
+		set(0, m[3] + m[0], m[7] + m[4], m[11] + m[8], m[15] + m[12]); // left
+		set(1, m[3] - m[0], m[7] - m[4], m[11] - m[8], m[15] - m[12]); // right
+		set(2, m[3] + m[1], m[7] + m[5], m[11] + m[9], m[15] + m[13]); // bottom
+		set(3, m[3] - m[1], m[7] - m[5], m[11] - m[9], m[15] - m[13]); // top
+		set(4, m[2], m[6], m[10], m[14]); // near, depth 0
+		set(5, m[3] - m[2], m[7] - m[6], m[11] - m[10], m[15] - m[14]); // far
+	}
+
+	// whether any of the box can be seen. may say yes for boxes just outside a corner, never no for one inside
+	intersects_box(min_x: number, min_y: number, min_z: number, max_x: number, max_y: number, max_z: number) {
+		const planes = this.#planes;
+		for (let i = 0; i < 24; i += 4) {
+			const a = planes[i];
+			const b = planes[i + 1];
+			const c = planes[i + 2];
+			// the box's corner furthest along the plane's normal
+			const x = a > 0 ? max_x : min_x;
+			const y = b > 0 ? max_y : min_y;
+			const z = c > 0 ? max_z : min_z;
+			if (a * x + b * y + c * z + planes[i + 3] < 0) {
+				return false;
+			}
+		}
+		return true;
+	}
+}

client/rendering/level_renderer.ts

@@ -1,5 +1,5 @@
 import { TEXTURE_SIZE } from "$/common/constants.ts";
-import { CHUNK_SIZE, ClientLevel } from "$/client/level/client_level.ts";
+import { type Chunk, CHUNK_SIZE, type ChunkMesh, ClientLevel } from "$/client/level/client_level.ts";
 import { Camera } from "$/client/camera.ts";
 import { AssetManager } from "$/client/assets.ts";
 import { get_sprite_region } from "$/client/sprites.ts";
@@ -20,8 +20,13 @@ import {
 	push_top_face,
 	set_current_texture,
 	Texture,
+	view_projection,
 	white_tex,
 } from "$/client/renderer/mod.ts";
+import type { RenderLayer } from "$/common/everything_registry.ts";
+import { FACE_GROUPS, UNALIGNED_GROUP } from "$/client/workers/chunk_messages.ts";
+import { FACE_AXIS, FACE_NORMALS } from "$/client/workers/translucent_sort.ts";
+import { Frustum } from "./frustum.ts";
 
 const BREAKING_FACES = [
 	push_back_face,
@@ -34,48 +39,59 @@ const BREAKING_FACES = [
 
 // draws the level, like minecraft's LevelRenderer: terrain in layers, block breaking and entities
 export class LevelRenderer {
+	frustum = new Frustum();
+	// chunks in view this frame, worked out once for all the layers
+	#visible: Chunk[] = [];
+
 	// solid and cutout terrain, drawn before entities
 	render_opaque(level: ClientLevel, camera: Camera) {
 		level.request_meshes(camera);
 		level.update_translucent_sorting(camera);
 
-		set_current_texture(level.image.tex);
-
+		this.frustum.update(view_projection());
+		this.#visible.length = 0;
 		for (const chunk of level.chunks.values()) {
-			const mesh = chunk.meshes.solid;
-			if (mesh) {
-				draw_terrain("solid", mesh.vertex_buffer, mesh.quad_count);
+			if (Object.keys(chunk.meshes).length > 0 && this.#in_view(chunk)) {
+				this.#visible.push(chunk);
 			}
 		}
 
-		for (const chunk of level.chunks.values()) {
-			const mesh = chunk.meshes.cutout;
-			if (mesh) {
-				draw_terrain("cutout", mesh.vertex_buffer, mesh.quad_count);
+		set_current_texture(level.image.tex);
+
+		for (const layer of ["solid", "cutout"] as const) {
+			for (const chunk of this.#visible) {
+				const mesh = chunk.meshes[layer];
+				if (mesh) {
+					draw_facing_camera(layer, mesh, camera);
+				}
 			}
 		}
 	}
 
 	// translucent terrain, drawn after entities so they show through water and glass.
 	// chunks go back to front, and each chunk's quads are already sorted back to front
-	render_translucent(level: ClientLevel, camera: Camera) {
+	render_translucent(_level: ClientLevel, camera: Camera) {
 		const distance_sq = (x: number, z: number) => {
 			const dx = (x + 0.5) * CHUNK_SIZE - camera.x;
 			const dz = (z + 0.5) * CHUNK_SIZE - camera.z;
 			return dx * dx + dz * dz;
 		};
-		const chunks = [...level.chunks.values()]
+		const chunks = this.#visible
 			.filter((chunk) => chunk.meshes.translucent)
 			.map((chunk) => ({ mesh: chunk.meshes.translucent!, distance: distance_sq(chunk.x, chunk.z) }))
 			.sort((a, b) => b.distance - a.distance);
 
-		set_current_texture(level.image.tex);
-
 		for (const { mesh } of chunks) {
-			draw_terrain("translucent", mesh.vertex_buffer, mesh.quad_count, mesh.index_buffer);
+			draw_terrain("translucent", mesh.vertex_buffer, 0, mesh.quad_count, mesh.index_buffer);
 		}
 	}
 
+	#in_view(chunk: Chunk) {
+		const x = chunk.x * CHUNK_SIZE;
+		const z = chunk.z * CHUNK_SIZE;
+		return this.frustum.intersects_box(x, chunk.min_y, z, x + CHUNK_SIZE, chunk.max_y, z + CHUNK_SIZE);
+	}
+
 	// every entity but the one the camera is in
 	render_entities(level: ClientLevel, camera_entity: Entity, partial_tick: number) {
 		flush_batch();
@@ -127,6 +143,44 @@ export class LevelRenderer {
 	}
 }
 
+// only the face groups that can face the camera: quads facing +x can only be seen from beyond the lowest x plane any
+// of them lie on. neighboring groups that are both drawn go in one draw call
+function draw_facing_camera(layer: RenderLayer, mesh: ChunkMesh, camera: Camera) {
+	const groups = mesh.groups;
+	if (!groups) {
+		draw_terrain(layer, mesh.vertex_buffer, 0, mesh.quad_count);
+		return;
+	}
+	let start = 0;
+	let end = 0;
+	for (let g = 0; g < FACE_GROUPS; g++) {
+		const group = groups[g];
+		if (group.count === 0 || !group_faces_camera(g, group.min, group.max, camera)) {
+			continue;
+		}
+		if (group.first !== end) {
+			draw_terrain(layer, mesh.vertex_buffer, start, end - start);
+			start = group.first;
+		}
+		end = group.first + group.count;
+	}
+	draw_terrain(layer, mesh.vertex_buffer, start, end - start);
+}
+
+export function group_faces_camera(
+	group: number,
+	min: number,
+	max: number,
+	camera: { x: number; y: number; z: number },
+) {
+	if (group === UNALIGNED_GROUP) {
+		return true;
+	}
+	const axis = FACE_AXIS[group];
+	const position = axis === 0 ? camera.x : axis === 1 ? camera.y : camera.z;
+	return FACE_NORMALS[group][axis] > 0 ? position > min : position < max;
+}
+
 // a box body and head in the player's color
 function render_player(player: RemotePlayer, partial_tick: number) {
 	const [r, g, b] = player.color;

client/workers/chunk_messages.ts

@@ -6,8 +6,23 @@ import type { SortType } from "./translucent_sort.ts";
 
 // messages between the main thread and the chunk workers
 
-// position 3, uv 2, color 4 (directional shade and ambient occlusion, alpha), lightmap coordinates 2
-export const TERRAIN_VERTEX_FLOATS = 11;
+// a terrain vertex, 24 bytes: position as float32x3, atlas uv as unorm16x2, directional shade times ambient occlusion
+// and alpha as unorm8x4 (shade repeated in rgb), and lightmap coordinates as unorm8x2 plus two unused bytes
+export const TERRAIN_VERTEX_BYTES = 24;
+export const TERRAIN_QUAD_BYTES = 4 * TERRAIN_VERTEX_BYTES;
+
+// solid and cutout quads are grouped by the way they face, so groups facing away from the camera can be skipped, like
+// sodium's block face culling. the groups follow FACE_NORMALS' order, then the quads that aren't axis aligned
+export const FACE_GROUPS = 7;
+export const UNALIGNED_GROUP = 6;
+
+export interface FaceGroup {
+	first: number;
+	count: number;
+	// the lowest and highest plane the group's quads lie on, along its axis
+	min: number;
+	max: number;
+}
 
 export type ToChunkWorker =
 	| {
@@ -44,10 +59,12 @@ export type ToChunkWorker =
 		camera: number[];
 	};
 
-// 4 vertices per quad
+// 4 vertices per quad, TERRAIN_VERTEX_BYTES each
 export interface LayerMesh {
-	vertices: Float32Array;
+	vertices: Uint8Array<ArrayBuffer>;
 	quad_count: number;
+	// solid and cutout only, see FACE_GROUPS. the quads are stored group after group
+	groups?: FaceGroup[];
 }
 
 export type FromChunkWorker =
@@ -64,6 +81,9 @@ export type FromChunkWorker =
 		chunk_x: number;
 		chunk_z: number;
 		version: number;
+		// the lowest and highest y of the chunk's quads, for frustum culling
+		min_y: number;
+		max_y: number;
 		solid: LayerMesh;
 		cutout: LayerMesh;
 		translucent: LayerMesh & {

client/workers/chunk_worker.ts

@@ -16,13 +16,22 @@ import {
 	TEXTURE_SIZE,
 } from "$/common/constants.ts";
 import type { Texture } from "../renderer/types.ts";
-import { type FromChunkWorker, TERRAIN_VERTEX_FLOATS, type ToChunkWorker } from "./chunk_messages.ts";
+import {
+	FACE_GROUPS,
+	type FaceGroup,
+	type FromChunkWorker,
+	TERRAIN_QUAD_BYTES,
+	TERRAIN_VERTEX_BYTES,
+	type ToChunkWorker,
+	UNALIGNED_GROUP,
+} from "./chunk_messages.ts";
 import { generate_raw_chunk, WorldgenSetup } from "$/common/generation.ts";
 import { load_worldgen } from "$/common/worldgen_loader.ts";
 import { default_block_value, get_state_value } from "$/common/utils.ts";
 import { bake_model, block_variant, FACE_CORNERS, find_model, type ModelJson } from "$/common/block_models.ts";
 import {
 	choose_sort_type,
+	FACE_AXIS,
 	FACE_NORMALS,
 	quad_indices,
 	quad_planes,
@@ -42,7 +51,6 @@ import {
 
 type TexturesInfo = Record<string, SpriteRegion>;
 
-const FLOATS_PER_QUAD = 4 * TERRAIN_VERTEX_FLOATS;
 // keeps texture lookups off the sprite's edge
 const UV_PAD = 0.5;
 
@@ -91,6 +99,8 @@ interface MeshQuad {
 	face: number;
 	cull: number;
 	flush: boolean;
+	// the face group it goes in, see FACE_GROUPS
+	group: number;
 	shade: number;
 	sprite: SpriteRegion;
 	// 4 weights per corner
@@ -132,7 +142,7 @@ self.onmessage = async (event: MessageEvent<ToChunkWorker>) => {
 		case "mesh": {
 			region.fill(message.chunks);
 			region.compute(light_tables);
-			const { solid, cutout, translucent } = make_chunk_mesh(
+			const { solid, cutout, translucent, min_y, max_y } = make_chunk_mesh(
 				message.chunk_x,
 				message.chunk_z,
 				message.chunks[4]!,
@@ -144,6 +154,8 @@ self.onmessage = async (event: MessageEvent<ToChunkWorker>) => {
 					chunk_x: message.chunk_x,
 					chunk_z: message.chunk_z,
 					version: message.version,
+					min_y,
+					max_y,
 					solid,
 					cutout,
 					translucent,
@@ -285,28 +297,57 @@ function should_flip() {
 	return light(0) + light(2) > light(1) + light(3);
 }
 
-function push_quad(vertices: Float32Array, i: number, quad: MeshQuad, x: number, y: number, z: number, alpha: number) {
+// quads being built, in the terrain vertex format
+class QuadBuffer {
+	count = 0;
+	bytes = new Uint8Array(TERRAIN_QUAD_BYTES * 64);
+	f32 = new Float32Array(this.bytes.buffer);
+	u16 = new Uint16Array(this.bytes.buffer);
+
+	// room for one more quad
+	reserve() {
+		if ((this.count + 1) * TERRAIN_QUAD_BYTES <= this.bytes.length) return;
+		const bytes = new Uint8Array(this.bytes.length * 2);
+		bytes.set(this.bytes);
+		this.bytes = bytes;
+		this.f32 = new Float32Array(bytes.buffer);
+		this.u16 = new Uint16Array(bytes.buffer);
+	}
+}
+
+// the lowest and highest y of any quad in the chunk being meshed
+let mesh_min_y = Infinity;
+let mesh_max_y = -Infinity;
+
+function push_quad(buffer: QuadBuffer, quad: MeshQuad, x: number, y: number, z: number, alpha: number) {
+	buffer.reserve();
+	const { f32, u16, bytes } = buffer;
 	const sprite = quad.sprite;
 	// starting from the second corner moves the diagonal, the winding stays the same
 	const first = should_flip() ? 1 : 0;
+	const alpha_byte = Math.round(alpha * 255);
 
 	for (let k = 0; k < 4; k++) {
 		const corner = (first + k) & 3;
-		const brightness = quad.shade * corner_ao[corner];
-		vertices[i++] = x + quad.positions[corner * 3];
-		vertices[i++] = y + quad.positions[corner * 3 + 1];
-		vertices[i++] = z + quad.positions[corner * 3 + 2];
-		vertices[i++] = atlas_u(sprite, quad.uvs[corner * 2]);
-		vertices[i++] = atlas_v(sprite, quad.uvs[corner * 2 + 1]);
-		vertices[i++] = brightness;
-		vertices[i++] = brightness;
-		vertices[i++] = brightness;
-		vertices[i++] = alpha;
+		const byte = (buffer.count * 4 + k) * TERRAIN_VERTEX_BYTES;
+		const vy = y + quad.positions[corner * 3 + 1];
+		f32[byte / 4] = x + quad.positions[corner * 3];
+		f32[byte / 4 + 1] = vy;
+		f32[byte / 4 + 2] = z + quad.positions[corner * 3 + 2];
+		u16[byte / 2 + 6] = Math.round(atlas_u(sprite, quad.uvs[corner * 2]) * 65535);
+		u16[byte / 2 + 7] = Math.round(atlas_v(sprite, quad.uvs[corner * 2 + 1]) * 65535);
+		const brightness = Math.round(quad.shade * corner_ao[corner] * 255);
+		bytes[byte + 16] = brightness;
+		bytes[byte + 17] = brightness;
+		bytes[byte + 18] = brightness;
+		bytes[byte + 19] = alpha_byte;
 		// where to read the lightmap, block light across and sky light down
-		vertices[i++] = (corner_block[corner] + 0.5) / 16;
-		vertices[i++] = (corner_sky[corner] + 0.5) / 16;
+		bytes[byte + 20] = Math.round((corner_block[corner] + 0.5) / 16 * 255);
+		bytes[byte + 21] = Math.round((corner_sky[corner] + 0.5) / 16 * 255);
+		if (vy < mesh_min_y) mesh_min_y = vy;
+		if (vy > mesh_max_y) mesh_max_y = vy;
 	}
-	return i;
+	buffer.count += 1;
 }
 
 // pixels of a sprite to atlas coordinates, kept off the sprite's edge
@@ -340,6 +381,7 @@ function block_quads(value: number): MeshQuad[] {
 		face: quad.face,
 		cull: quad.cull,
 		flush: quad.flush,
+		group: quad.aligned ? quad.face : UNALIGNED_GROUP,
 		shade: quad.shade ? FACE_SHADE[quad.face] : 1,
 		sprite: textures_info[quad.texture] ?? textures_info["engine:missing"],
 		light_weights: light_weights(quad.face, quad.positions),
@@ -408,7 +450,13 @@ function light_quad(quad: MeshQuad, index: number, y: number, face_offsets: numb
 // region has to be filled and lit first
 // values is the middle chunk's blocks with their states, for models that change with them
 function make_chunk_mesh(chunk_x: number, chunk_z: number, values: Uint32Array, camera: number[]) {
-	const layers = [SOLID, CUTOUT, TRANSLUCENT].map(() => ({ vertices: new Float32Array(4096), floats: 0 }));
+	// solid and cutout get a buffer per face group, translucent one for everything since it's sorted instead
+	const opaque = [SOLID, CUTOUT].map(() =>
+		Array.from({ length: FACE_GROUPS }, () => ({ buffer: new QuadBuffer(), min: Infinity, max: -Infinity }))
+	);
+	const translucent_quads = new QuadBuffer();
+	mesh_min_y = Infinity;
+	mesh_max_y = -Infinity;
 	// for sorting the translucent quads
 	let centers = new Float32Array(256);
 	let faces = new Uint8Array(256);
@@ -426,7 +474,6 @@ function make_chunk_mesh(chunk_x: number, chunk_z: number, values: Uint32Array,
 
 				const block_info = blocks_registry[block_nid];
 				const layer_id = block_layers[block_nid];
-				const layer = layers[layer_id];
 				const alpha = layer_id === TRANSLUCENT ? block_info.alpha ?? 1 : 1;
 
 				const wx = chunk_x * CHUNK_SIZE + x;
@@ -443,12 +490,20 @@ function make_chunk_mesh(chunk_x: number, chunk_z: number, values: Uint32Array,
 					}
 
 					light_quad(quad, index, y, face_offsets);
-					layer.vertices = ensure_capacity(layer.vertices, layer.floats + FLOATS_PER_QUAD);
-					layer.floats = push_quad(layer.vertices, layer.floats, quad, wx, y, wz, alpha);
 
-					if (layer_id === TRANSLUCENT) {
+					if (layer_id !== TRANSLUCENT) {
+						const group = opaque[layer_id][quad.group];
+						push_quad(group.buffer, quad, wx, y, wz, alpha);
+						if (quad.group !== UNALIGNED_GROUP) {
+							const axis = FACE_AXIS[quad.face];
+							const plane = (axis === 0 ? wx : axis === 1 ? y : wz) + quad.positions[axis];
+							if (plane < group.min) group.min = plane;
+							if (plane > group.max) group.max = plane;
+						}
+					} else {
+						push_quad(translucent_quads, quad, wx, y, wz, alpha);
 						// sorting treats every quad as facing along an axis, rotated ones too
-						const q = layer.floats / FLOATS_PER_QUAD - 1;
+						const q = translucent_quads.count - 1;
 						centers = ensure_capacity(centers, (q + 1) * 3);
 						faces = ensure_capacity(faces, q + 1);
 						const p = quad.positions;
@@ -462,16 +517,31 @@ function make_chunk_mesh(chunk_x: number, chunk_z: number, values: Uint32Array,
 		}
 	}
 
-	const [solid, cutout, translucent] = layers.map((layer) => ({
-		vertices: layer.vertices,
-		quad_count: layer.floats / FLOATS_PER_QUAD,
-	}));
+	// each layer's groups one after another, in one buffer
+	const [solid, cutout] = opaque.map((groups) => {
+		const quad_count = groups.reduce((sum, group) => sum + group.buffer.count, 0);
+		const vertices = new Uint8Array(quad_count * TERRAIN_QUAD_BYTES);
+		const ranges: FaceGroup[] = [];
+		let first = 0;
+		for (const { buffer, min, max } of groups) {
+			vertices.set(buffer.bytes.subarray(0, buffer.count * TERRAIN_QUAD_BYTES), first * TERRAIN_QUAD_BYTES);
+			ranges.push({ first, count: buffer.count, min, max });
+			first += buffer.count;
+		}
+		return { vertices, quad_count, groups: ranges };
+	});
+	const translucent = {
+		vertices: translucent_quads.bytes.slice(0, translucent_quads.count * TERRAIN_QUAD_BYTES),
+		quad_count: translucent_quads.count,
+	};
 
 	const quads = { centers, faces, count: translucent.quad_count };
 	const sort_type = choose_sort_type(quads);
 	const [camera_x, camera_y, camera_z] = camera;
 
 	return {
+		min_y: mesh_min_y === Infinity ? 0 : mesh_min_y,
+		max_y: mesh_max_y === -Infinity ? 0 : mesh_max_y,
 		solid,
 		cutout,
 		translucent: {