Implement chunk unloading, frustrum culling and some other minor optimisations

.gitignore

@@ -4,5 +4,6 @@
 /perf.data*
 /chunks/
 
+GPUCache/
 profile.txt
 callgrind.out.*

src/aabb.rs

@@ -1,49 +1,23 @@
-use cgmath::Vector3;
-use std::ops::{Div, Sub};
+use cgmath::Point3;
 
-pub struct Aabb<T> {
-    pub min: Vector3<T>,
-    pub max: Vector3<T>,
+pub struct Aabb {
+    pub min: Point3<f32>,
+    pub max: Point3<f32>,
 }
 
-impl<T: Ord + Copy + Sub<Output = T> + Div<Output = T>> Aabb<T> {
-    pub fn intersects_ray(
-        &self,
-        ray_origin: Vector3<T>,
-        ray_direction: Vector3<T>,
-    ) -> Option<(T, T)> {
-        let mut t_min = (self.min.x - ray_origin.x) / ray_direction.x;
-        let mut t_max = (self.max.x - ray_origin.x) / ray_direction.x;
-        if t_min > t_max {
-            std::mem::swap(&mut t_min, &mut t_max);
-        }
-
-        let mut ty_min = (self.min.y - ray_origin.y) / ray_direction.y;
-        let mut ty_max = (self.max.y - ray_origin.y) / ray_direction.y;
-        if ty_min > ty_max {
-            std::mem::swap(&mut ty_min, &mut ty_max);
-        }
-
-        if t_min > ty_max || ty_min > t_max {
-            return None;
-        }
-
-        t_min = T::max(t_min, ty_min);
-        t_max = T::min(t_min, ty_max);
-
-        let mut tz_min = (self.min.z - ray_origin.z) / ray_direction.z;
-        let mut tz_max = (self.max.z - ray_origin.z) / ray_direction.z;
-        if tz_min > tz_max {
-            std::mem::swap(&mut tz_min, &mut tz_max);
-        }
-
-        if t_min > tz_max || tz_min > t_max {
-            return None;
-        }
-
-        t_min = T::max(t_min, tz_min);
-        t_max = T::max(t_max, tz_max);
-
-        Some((t_min, t_max))
+impl Aabb {
+    pub fn intersects(&self, other: &Self) -> bool {
+        (self.min.x <= other.max.x && self.max.x >= other.min.x)
+            && (self.min.y <= other.max.y && self.max.y >= other.min.y)
+            && (self.min.z <= other.max.z && self.max.z >= other.min.z)
+    }
+}
+
+impl Default for Aabb {
+    fn default() -> Self {
+        Self {
+            min: Point3::new(0.0, 0.0, 0.0),
+            max: Point3::new(0.0, 0.0, 0.0),
+        }
     }
 }

src/camera.rs

@@ -1,7 +1,5 @@
 use cgmath::{Matrix4, Point3, Rad, Vector3};
 
-use crate::aabb::Aabb;
-
 #[rustfmt::skip]
 pub const OPENGL_TO_WGPU_MATRIX: Matrix4<f32> = Matrix4::new(
     1.0, 0.0, 0.0, 0.0,
@@ -10,11 +8,6 @@ pub const OPENGL_TO_WGPU_MATRIX: Matrix4<f32> = Matrix4::new(
     0.0, 0.0, 0.5, 1.0,
 );
 
-const AABB: Aabb<f32> = Aabb {
-    min: Vector3::new(-0.3, 1.62, -0.3),
-    max: Vector3::new(0.3, 1.8 - 1.62, 0.3),
-};
-
 pub struct Camera {
     pub position: Point3<f32>,
     pub yaw: Rad<f32>,

src/chunk.rs

@@ -1,6 +1,12 @@
 use std::{collections::VecDeque, usize};
 
-use crate::{geometry::Geometry, quad::Quad, vertex::BlockVertex};
+use crate::{
+    aabb::Aabb,
+    geometry::{Geometry, GeometryBuffers},
+    quad::Quad,
+    vertex::BlockVertex,
+    view::View,
+};
 use ahash::{AHashMap, AHashSet};
 use cgmath::{Point3, Vector3};
 use noise::utils::{NoiseMapBuilder, PlaneMapBuilder};
@@ -68,9 +74,18 @@ pub const CHUNK_ISIZE: isize = CHUNK_SIZE as isize;
 
 type ChunkBlocks = [[[Option<Block>; CHUNK_SIZE]; CHUNK_SIZE]; CHUNK_SIZE];
 
-#[derive(Clone, Default)]
 pub struct Chunk {
     pub blocks: ChunkBlocks,
+    pub buffers: Option<GeometryBuffers<u16>>,
+}
+
+impl Default for Chunk {
+    fn default() -> Self {
+        Self {
+            blocks: [[[None; CHUNK_SIZE]; CHUNK_SIZE]; CHUNK_SIZE],
+            buffers: None,
+        }
+    }
 }
 
 impl Serialize for Chunk {
@@ -179,7 +194,7 @@ impl Chunk {
                     });
                 }
 
-                if dirt_max >= 0 && dirt_max < CHUNK_ISIZE {
+                if (0..CHUNK_ISIZE).contains(&dirt_max) {
                     self.blocks[dirt_max as usize][z][x] = Some(Block {
                         block_type: BlockType::Grass,
                     });
@@ -410,4 +425,15 @@ impl Chunk {
             Ok(true)
         }
     }
+
+    pub fn is_visible(&self, position: Point3<isize>, view: &View) -> bool {
+        let aabb = Aabb {
+            min: position.cast().unwrap(),
+            max: (position + Vector3::new(CHUNK_ISIZE, CHUNK_ISIZE, CHUNK_ISIZE))
+                .cast()
+                .unwrap(),
+        };
+
+        aabb.intersects(&view.aabb)
+    }
 }

src/geometry.rs

@@ -77,7 +77,7 @@ impl<I: bytemuck::Pod> GeometryBuffers<I> {
 }
 
 impl GeometryBuffers<u16> {
-    pub fn set_buffers<'a>(&'a self, render_pass: &mut RenderPass<'a>) {
+    pub fn apply_buffers<'a>(&'a self, render_pass: &mut RenderPass<'a>) {
         render_pass.set_vertex_buffer(0, self.vertices.slice(..));
         render_pass.set_index_buffer(self.indices.slice(..), wgpu::IndexFormat::Uint16);
     }

src/main.rs

@@ -2,6 +2,7 @@ mod aabb;
 mod camera;
 mod chunk;
 mod geometry;
+mod npc;
 mod quad;
 mod render_context;
 mod state;
@@ -11,7 +12,6 @@ mod time;
 mod vertex;
 mod view;
 mod world;
-mod npc;
 
 use std::time::{Duration, Instant};
 use wgpu::SwapChainError;
@@ -133,9 +133,21 @@ fn main() {
                     let fps_max = 1_000_000 / frametime_min.as_micros();
                     let fps_min = 1_000_000 / frametime_max.as_micros();
 
+                    print!("{:>4} frames | ", frames);
+                    print!(
+                        "frametime avg={:>5.2}ms min={:>5.2}ms max={:>5.2}ms | ",
+                        frametime.as_secs_f32() * 1000.0,
+                        frametime_min.as_secs_f32() * 1000.0,
+                        frametime_max.as_secs_f32() * 1000.0,
+                    );
+                    print!(
+                        "fps avg={:>5} min={:>5} max={:>5} | ",
+                        fps, fps_min, fps_max
+                    );
                     println!(
-                        "{:>4} frames | frametime avg={:>5.2}ms min={:>5.2}ms max={:>5.2}ms | fps avg={:>5} min={:>5} max={:>5} | {:>8} tris",
-                        frames, frametime.as_secs_f32() * 1000.0, frametime_min.as_secs_f32() * 1000.0, frametime_max.as_secs_f32() * 1000.0, fps, fps_min, fps_max, triangle_count,
+                        "{:>8} tris | {:>5} chunks",
+                        triangle_count,
+                        state.world_state.world.chunks.len()
                     );
 
                     elapsed = Duration::from_secs(0);
@@ -150,17 +162,18 @@ fn main() {
                 state.update(dt);
 
                 match state.render() {
-                    Err(root_cause) =>
-                        match root_cause.downcast_ref::<SwapChainError>() {
+                    Err(root_cause) => match root_cause.downcast_ref::<SwapChainError>() {
                         // Recreate the swap_chain if lost
                         Some(wgpu::SwapChainError::Lost) => state.resize(state.window_size),
                         // The system is out of memory, we should probably quit
-                            Some(wgpu::SwapChainError::OutOfMemory) => *control_flow = ControlFlow::Exit,
+                        Some(wgpu::SwapChainError::OutOfMemory) => {
+                            *control_flow = ControlFlow::Exit
+                        }
                         // All other errors (Outdated, Timeout) should be resolved by the next frame
                         Some(_) | None => eprintln!("{:?}", root_cause),
-                        }
+                    },
 
-                    Ok(v) => triangle_count = v
+                    Ok(v) => triangle_count = v,
                 }
             }
             Event::MainEventsCleared => {

src/npc.rs

@@ -55,13 +55,13 @@ impl Npc {
             }
         }
 
-        return Self {
+        Self {
             position,
             scale,
             rotation,
             geometry: Geometry::new(vertices, indices),
             geometry_buffers: None,
-        };
+        }
     }
 
     pub fn load_geometry(&mut self, render_context: &RenderContext) {

src/state/hud_state.rs

@@ -124,19 +124,19 @@ impl HudState {
         render_pass.set_pipeline(&self.render_pipeline);
 
         // Render the HUD elements
-        self.hud_geometry_buffers.set_buffers(&mut render_pass);
+        self.hud_geometry_buffers.apply_buffers(&mut render_pass);
         render_pass.set_bind_group(0, &self.texture_bind_group, &[]);
         self.hud_geometry_buffers.draw_indexed(&mut render_pass);
         render_pass.draw_indexed(0..self.hud_geometry_buffers.index_count as u32, 0, 0..1);
 
         // Render the FPS text
-        self.fps_geometry_buffers.set_buffers(&mut render_pass);
+        self.fps_geometry_buffers.apply_buffers(&mut render_pass);
         render_pass.set_bind_group(0, &self.text_renderer.bind_group, &[]);
         self.fps_geometry_buffers.draw_indexed(&mut render_pass);
 
         // Render the coordinates text
         self.coordinates_geometry_buffers
-            .set_buffers(&mut render_pass);
+            .apply_buffers(&mut render_pass);
         render_pass.set_bind_group(0, &self.text_renderer.bind_group, &[]);
         self.coordinates_geometry_buffers
             .draw_indexed(&mut render_pass);

src/state/mod.rs

@@ -28,7 +28,7 @@ pub const PRIMITIVE_STATE: wgpu::PrimitiveState = wgpu::PrimitiveState {
 pub struct State {
     pub window_size: PhysicalSize<u32>,
     render_context: RenderContext,
-    world_state: WorldState,
+    pub world_state: WorldState,
     hud_state: HudState,
 
     pub mouse_grabbed: bool,

src/state/world_state.rs

@@ -67,7 +67,7 @@ impl WorldState {
             render_context.swap_chain_descriptor.height,
             cgmath::Deg(45.0),
             0.1,
-            5000.0,
+            300.0,
         );
 
         (camera, projection)
@@ -85,7 +85,7 @@ impl WorldState {
             .device
             .create_buffer_init(&BufferInitDescriptor {
                 label: Some("view_buffer"),
-                contents: bytemuck::cast_slice(&[view]),
+                contents: bytemuck::cast_slice(&[view.to_raw()]),
                 usage: wgpu::BufferUsage::UNIFORM | wgpu::BufferUsage::COPY_DST,
             });
 
@@ -324,7 +324,7 @@ impl WorldState {
         render_pass.set_bind_group(1, &self.view_bind_group, &[]);
         render_pass.set_bind_group(2, &self.time_bind_group, &[]);
 
-        triangle_count += self.world.render(&mut render_pass, &self.camera);
+        triangle_count += self.world.render(&mut render_pass, &self.view);
 
         triangle_count
     }
@@ -428,13 +428,15 @@ impl WorldState {
     pub fn update(&mut self, dt: Duration, render_context: &RenderContext) {
         self.update_position(dt);
 
-        self.world.update(render_context, &self.camera);
+        self.world.update(dt, render_context, &self.camera);
 
         self.view
             .update_view_projection(&self.camera, &self.projection);
-        render_context
-            .queue
-            .write_buffer(&self.view_buffer, 0, bytemuck::cast_slice(&[self.view]));
+        render_context.queue.write_buffer(
+            &self.view_buffer,
+            0,
+            bytemuck::cast_slice(&[self.view.to_raw()]),
+        );
 
         self.time.time += dt.as_secs_f32();
         render_context.queue.write_buffer(

src/view.rs

@@ -1,24 +1,73 @@
-use cgmath::SquareMatrix;
+use cgmath::{EuclideanSpace, Matrix4, Point3, SquareMatrix, Vector4, Zero};
 
-use crate::camera::{Camera, Projection};
+use crate::{
+    aabb::Aabb,
+    camera::{Camera, Projection, OPENGL_TO_WGPU_MATRIX},
+};
 
-#[repr(C)]
-#[derive(Debug, Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
 pub struct View {
-    view_position: [f32; 4],
-    view_projection: [[f32; 4]; 4],
+    view_position: Vector4<f32>,
+    view_projection: Matrix4<f32>,
+    pub aabb: Aabb,
 }
 
 impl View {
+    pub fn to_raw(&self) -> ViewRaw {
+        ViewRaw {
+            view_position: self.view_position.into(),
+            view_projection: self.view_projection.into(),
+        }
+    }
+
     pub fn new() -> Self {
         Self {
-            view_position: [0.0; 4],
-            view_projection: cgmath::Matrix4::identity().into(),
+            view_position: Vector4::zero(),
+            view_projection: Matrix4::identity(),
+            aabb: Aabb::default(),
         }
     }
 
     pub fn update_view_projection(&mut self, camera: &Camera, projection: &Projection) {
-        self.view_position = camera.position.to_homogeneous().into();
-        self.view_projection = (projection.calculate_matrix() * camera.calculate_matrix()).into();
+        self.view_position = camera.position.to_homogeneous();
+        self.view_projection = projection.calculate_matrix() * camera.calculate_matrix();
+        self.aabb = self.frustrum_aabb();
+    }
+
+    fn frustrum_aabb(&self) -> Aabb {
+        let projection = OPENGL_TO_WGPU_MATRIX.invert().unwrap() * self.view_projection;
+        let inverse_matrix = projection.invert().unwrap();
+
+        let corners = &[
+            Vector4::new(-1.0, -1.0, 1.0, 1.0),
+            Vector4::new(-1.0, -1.0, -1.0, 1.0),
+            Vector4::new(-1.0, 1.0, 1.0, 1.0),
+            Vector4::new(-1.0, 1.0, -1.0, 1.0),
+            Vector4::new(1.0, -1.0, 1.0, 1.0),
+            Vector4::new(1.0, -1.0, -1.0, 1.0),
+            Vector4::new(1.0, 1.0, 1.0, 1.0),
+            Vector4::new(1.0, 1.0, -1.0, 1.0),
+        ];
+
+        let mut min = Vector4::new(f32::INFINITY, f32::INFINITY, f32::INFINITY, 1.0);
+        let mut max = Vector4::new(0.0, 0.0, 0.0, 1.0);
+        for &corner in corners {
+            let corner = inverse_matrix * corner;
+            let corner = corner / corner.w;
+
+            min = min.zip(corner, f32::min);
+            max = max.zip(corner, f32::max);
+        }
+
+        Aabb {
+            min: Point3::from_vec(min.truncate()),
+            max: Point3::from_vec(max.truncate()),
         }
     }
+}
+
+#[repr(C)]
+#[derive(Debug, Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
+pub struct ViewRaw {
+    view_position: [f32; 4],
+    view_projection: [[f32; 4]; 4],
+}

src/world.rs

@@ -1,4 +1,4 @@
-use std::collections::{HashMap, VecDeque};
+use std::{collections::VecDeque, time::Duration};
 
 use crate::{
     camera::Camera,
@@ -6,30 +6,34 @@ use crate::{
     geometry::GeometryBuffers,
     npc::Npc,
     render_context::RenderContext,
+    view::View,
 };
 use ahash::AHashMap;
-use cgmath::{EuclideanSpace, InnerSpace, Point3, Vector2, Vector3};
+use cgmath::{EuclideanSpace, InnerSpace, Point3, Vector3};
 use wgpu::{BufferUsage, RenderPass};
 
 pub struct World {
-    pub chunks: HashMap<Point3<isize>, Chunk>,
     pub npc: Npc,
 
+    pub chunks: AHashMap<Point3<isize>, Chunk>,
     pub chunk_database: sled::Db,
-    pub chunk_save_queue: VecDeque<Point3<isize>>,
+    pub chunk_save_queue: VecDeque<(Point3<isize>, bool)>,
     pub chunk_load_queue: VecDeque<Point3<isize>>,
     pub chunk_generate_queue: VecDeque<Point3<isize>>,
-    pub chunk_buffers: AHashMap<Point3<isize>, GeometryBuffers<u16>>,
 
     pub highlighted: Option<(Point3<isize>, Vector3<i32>)>,
+
+    pub unload_timer: Duration,
 }
 
 pub const RENDER_DISTANCE: isize = 8;
 pub const WORLD_HEIGHT: isize = 16 * 16 / CHUNK_ISIZE;
 
+const DEBUG_IO: bool = false;
+
 impl World {
     pub fn new() -> Self {
-        let chunks = HashMap::new();
+        let chunks = AHashMap::new();
         let npc = Npc::load();
 
         let chunk_database = sled::Config::new()
@@ -47,13 +51,14 @@ impl World {
             chunk_load_queue: VecDeque::new(),
             chunk_save_queue: VecDeque::new(),
             chunk_generate_queue: VecDeque::new(),
-            chunk_buffers: AHashMap::new(),
 
             highlighted: None,
+
+            unload_timer: Duration::new(0, 0),
         }
     }
 
-    pub fn update(&mut self, render_context: &RenderContext, camera: &Camera) {
+    pub fn update(&mut self, dt: Duration, render_context: &RenderContext, camera: &Camera) {
         if let Some(position) = self.chunk_load_queue.pop_front() {
             let chunk = self.chunks.entry(position).or_default();
             match chunk.load(position, &self.chunk_database) {
@@ -62,20 +67,33 @@ impl World {
                 }
                 Ok(true) => {
                     self.update_chunk_geometry(render_context, position);
-                    self.chunk_save_queue.push_back(position);
-                    // println!("Generated chunk {:?}", position);
+                    self.enqueue_chunk_save(position, false);
+                    if DEBUG_IO {
+                        println!("Generated chunk {:?}", position);
+                    }
                 }
                 Ok(false) => {
                     self.update_chunk_geometry(render_context, position);
-                    // println!("Loaded chunk {:?}", position);
+                    if DEBUG_IO {
+                        println!("Loaded chunk {:?}", position);
                     }
                 }
-        } else if let Some(position) = self.chunk_save_queue.pop_front() {
-            let chunk = self.chunks.get(&position).unwrap();
+            }
+        } else if let Some((position, unload)) = self.chunk_save_queue.pop_front() {
+            if let Some(chunk) = self.chunks.get(&position) {
                 if let Err(err) = chunk.save(position, &self.chunk_database) {
                     eprintln!("Failed to save chunk {:?}: {:?}", position, err);
+                } else if unload {
+                    self.chunks.remove(&position);
+
+                    if DEBUG_IO {
+                        println!("Saved and unloaded chunk {:?}", position);
+                    }
+                } else if DEBUG_IO {
+                    println!("Saved chunk {:?}", position);
+                }
             } else {
-                // println!("Saved chunk {:?}", position);
+                eprintln!("Tried to save unloaded chunk {:?}", position);
             }
         }
 
@@ -102,23 +120,43 @@ impl World {
         });
 
         self.chunk_load_queue.extend(load_queue);
+
+        // Unload chunks that are far away
+        self.unload_timer += dt;
+        if self.unload_timer.as_secs() >= 10 {
+            self.unload_timer = Duration::new(0, 0);
+
+            let camera_pos = camera.position.to_vec();
+            let unload_distance = (RENDER_DISTANCE * CHUNK_ISIZE) as f32 * 1.5;
+
+            let mut unload_chunks = Vec::new();
+            for point in self.chunks.keys() {
+                let pos: Point3<f32> = (point * CHUNK_ISIZE).cast().unwrap();
+                if (pos.x - camera_pos.x).abs() > unload_distance
+                    || (pos.z - camera_pos.z).abs() > unload_distance
+                {
+                    unload_chunks.push(*point);
+                }
+            }
+            for point in unload_chunks {
+                self.enqueue_chunk_save(point, true);
+            }
+        }
     }
 
-    pub fn render<'a>(&'a self, render_pass: &mut RenderPass<'a>, camera: &Camera) -> usize {
-        let camera_pos = camera.position.to_vec();
-        let camera_pos = Vector2::new(camera_pos.x, camera_pos.z);
+    pub fn render<'a>(&'a self, render_pass: &mut RenderPass<'a>, view: &View) -> usize {
         let mut triangle_count = 0;
 
-        for (position, buffers) in &self.chunk_buffers {
-            let pos = (position * CHUNK_ISIZE).cast().unwrap();
-            let pos = Vector2::new(pos.x, pos.z);
-            if (pos - camera_pos).magnitude() > 300.0 {
+        for (position, chunk) in &self.chunks {
+            if !chunk.is_visible(position * CHUNK_ISIZE, &view) {
                 continue;
             }
 
-            buffers.set_buffers(render_pass);
+            if let Some(buffers) = chunk.buffers.as_ref() {
+                buffers.apply_buffers(render_pass);
                 triangle_count += buffers.draw_indexed(render_pass);
             }
+        }
 
         {
             let buffers = self.npc.geometry_buffers.as_ref().unwrap();
@@ -129,21 +167,35 @@ impl World {
         triangle_count
     }
 
+    pub fn enqueue_chunk_save(&mut self, position: Point3<isize>, unload: bool) {
+        if let Some((_, unload_)) = self
+            .chunk_save_queue
+            .iter_mut()
+            .find(|(pos, _)| pos == &position)
+        {
+            *unload_ = *unload_ || unload;
+        } else {
+            self.chunk_save_queue.push_back((position, unload));
+        }
+    }
+
     pub fn update_chunk_geometry(
         &mut self,
         render_context: &RenderContext,
         chunk_position: Point3<isize>,
     ) {
-        let chunk = &mut self.chunks.get(&chunk_position).unwrap();
+        let chunk = self.chunks.get_mut(&chunk_position).unwrap();
         let offset = chunk_position * CHUNK_ISIZE;
         let geometry = chunk.to_geometry(
             offset,
             World::highlighted_for_chunk(self.highlighted, &chunk_position).as_ref(),
         );
 
-        let buffers =
-            GeometryBuffers::from_geometry(render_context, &geometry, BufferUsage::empty());
-        self.chunk_buffers.insert(chunk_position, buffers);
+        chunk.buffers = Some(GeometryBuffers::from_geometry(
+            render_context,
+            &geometry,
+            BufferUsage::empty(),
+        ));
     }
 
     fn update_highlight(&mut self, render_context: &RenderContext, camera: &Camera) {
@@ -247,7 +299,7 @@ impl World {
         }
 
         self.chunk_save_queue
-            .push_back(chunk_position / CHUNK_ISIZE);
+            .push_back((chunk_position / CHUNK_ISIZE, false));
     }
 
     fn calc_scale(vector: Vector3<f32>, scalar: f32) -> f32 {