Merge branch 'zesterer/perf2' into 'master'

Various terrain-related performance improvements

See merge request veloren/veloren!3655

CHANGELOG.md

@@ -24,6 +24,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - HQX upscaling shader for people playing on low internal resolutions
 - Pets can now be traded with.
 - Crafting recipe for black lantern
+- Added redwood and dead trees
 
 ### Changed
 - Use fluent for translations

Cargo.lock

@@ -6729,6 +6729,7 @@ dependencies = [
  "ordered-float 2.10.0",
  "petgraph 0.6.2",
  "rand 0.8.5",
+ "rand_chacha 0.3.1",
  "rayon",
  "ron 0.7.1",
  "roots",

common/Cargo.toml

@@ -82,6 +82,7 @@ specs = { version = "0.18", features = ["serde", "storage-event-control", "night
 [dev-dependencies]
 #bench
 criterion = "0.3"
+rand_chacha = "0.3"
 
 #test
 tracing-subscriber = { version = "0.3.7", default-features = false, features = ["fmt", "time", "ansi", "smallvec", "env-filter"] }

common/benches/chonk_benchmark.rs

@@ -124,6 +124,27 @@ fn criterion_benchmark(c: &mut Criterion) {
             }
         })
     });
+
+    c.bench_function("for_each_in", |b| {
+        use rand::prelude::*;
+        let mut rng = rand_chacha::ChaChaRng::seed_from_u64(thread_rng().gen());
+        b.iter(|| {
+            let pos = Vec3::new(
+                rng.gen_range(0..TerrainChunk::RECT_SIZE.x as i32 - 3),
+                rng.gen_range(0..TerrainChunk::RECT_SIZE.x as i32 - 3),
+                rng.gen_range(MIN_Z..MAX_Z - 6),
+            );
+            chunk.for_each_in(
+                Aabb {
+                    min: pos,
+                    max: pos + Vec3::new(3, 3, 6),
+                },
+                |pos, vox| {
+                    black_box((pos, vox));
+                },
+            );
+        })
+    });
     black_box(chunk);
 }
 

common/src/terrain/chonk.rs

@@ -181,6 +181,49 @@ impl<V, S: RectVolSize, M: Clone> ReadVol for Chonk<V, S, M> {
                 .map_err(Self::Error::SubChunkError)
         }
     }
+
+    #[inline(always)]
+    fn get_unchecked(&self, pos: Vec3<i32>) -> &V {
+        if pos.z < self.get_min_z() {
+            // Below the terrain
+            &self.below
+        } else if pos.z >= self.get_max_z() {
+            // Above the terrain
+            &self.above
+        } else {
+            // Within the terrain
+            let sub_chunk_idx = self.sub_chunk_idx(pos.z);
+            let rpos = pos
+                - Vec3::unit_z()
+                    * (self.z_offset + sub_chunk_idx * SubChunkSize::<S>::SIZE.z as i32);
+            self.sub_chunks[sub_chunk_idx as usize]
+                .get_unchecked(rpos)
+        }
+    }
+
+    fn for_each_in(&self, aabb: Aabb<i32>, mut f: impl FnMut(Vec3<i32>, Self::Vox))
+    where
+        Self::Vox: Copy,
+    {
+        let idx = self.sub_chunk_idx(aabb.min.z);
+        // Special-case for the AABB being entirely within a single sub-chunk as this is very common.
+        if idx == self.sub_chunk_idx(aabb.max.z) && idx >= 0 && idx < self.sub_chunks.len() as i32 {
+            let sub_chunk = &self.sub_chunks[idx as usize];
+            let z_off = self.z_offset + idx * SubChunkSize::<S>::SIZE.z as i32;
+            sub_chunk.for_each_in(
+                Aabb { min: aabb.min.with_z(aabb.min.z - z_off), max: aabb.max.with_z(aabb.max.z - z_off) },
+                |pos, vox| f(pos.with_z(pos.z + z_off), vox),
+            );
+        } else {
+            for z in aabb.min.z..aabb.max.z + 1 {
+                for y in aabb.min.y..aabb.max.y + 1 {
+                    for x in aabb.min.x..aabb.max.x + 1 {
+                        f(Vec3::new(x, y, z), *self.get_unchecked(Vec3::new(x, y, z)));
+                    }
+                }
+            }
+        }
+    }
 }
 
 impl<V: Clone + PartialEq, S: RectVolSize, M: Clone> WriteVol for Chonk<V, S, M> {

common/src/util/grid_hasher.rs

@@ -0,0 +1,20 @@
+use std::hash::{Hasher, BuildHasher};
+
+#[derive(Copy, Clone, Default)]
+pub struct GridHasher(u64);
+
+// It's extremely unlikely that the spatial grid can be used to viably DOS the server given that clients only have
+// control over their player and a handful of entities in their near vicinity. For this reason, we just use an xor
+// hash, which should keep collisions relatively low since the spatial coherence of the grid is distributed fairly
+// evenly with the output of the hash function.
+impl Hasher for GridHasher {
+    fn finish(&self) -> u64 { self.0 }
+    fn write(&mut self, _: &[u8]) { panic!("Hashing arbitrary bytes is unimplemented"); }
+    fn write_i32(&mut self, x: i32) { self.0 = self.0.wrapping_mul(113989) ^ self.0 ^ x as u64; }
+}
+
+impl BuildHasher for GridHasher {
+    type Hasher = Self;
+
+    fn build_hasher(&self) -> Self::Hasher { *self }
+}

common/src/util/mod.rs

@@ -7,6 +7,7 @@ pub mod projection;
 /// Contains [`SpatialGrid`] which is useful for accelerating queries of nearby
 /// entities
 mod spatial_grid;
+mod grid_hasher;
 
 pub const GIT_VERSION_BUILD: &str = include_str!(concat!(env!("OUT_DIR"), "/githash"));
 pub const GIT_TAG_BUILD: &str = include_str!(concat!(env!("OUT_DIR"), "/gittag"));
@@ -39,3 +40,4 @@ pub use option::either_with;
 pub use plane::Plane;
 pub use projection::Projection;
 pub use spatial_grid::SpatialGrid;
+pub use grid_hasher::GridHasher;

common/src/util/spatial_grid.rs

@@ -1,11 +1,12 @@
 use vek::*;
+use super::GridHasher;
 
 #[derive(Debug)]
 pub struct SpatialGrid {
     // Uses two scales of grids so that we can have a hard limit on how far to search in the
     // smaller grid
-    grid: hashbrown::HashMap<Vec2<i32>, Vec<specs::Entity>>,
-    large_grid: hashbrown::HashMap<Vec2<i32>, Vec<specs::Entity>>,
+    grid: hashbrown::HashMap<Vec2<i32>, Vec<specs::Entity>, GridHasher>,
+    large_grid: hashbrown::HashMap<Vec2<i32>, Vec<specs::Entity>, GridHasher>,
     // Log base 2 of the cell size of the spatial grid
     lg2_cell_size: usize,
     // Log base 2 of the cell size of the large spatial grid
@@ -53,7 +54,7 @@ impl SpatialGrid {
     /// NOTE: for best optimization of the iterator use
     /// `for_each` rather than a for loop.
     pub fn in_aabr<'a>(&'a self, aabr: Aabr<i32>) -> impl Iterator<Item = specs::Entity> + 'a {
-        let iter = |max_entity_radius, grid: &'a hashbrown::HashMap<_, _>, lg2_cell_size| {
+        let iter = |max_entity_radius, grid: &'a hashbrown::HashMap<_, _, _>, lg2_cell_size| {
             // Add buffer for other entity radius
             let min = aabr.min - max_entity_radius as i32;
             let max = aabr.max + max_entity_radius as i32;

common/src/vol.rs

@@ -98,6 +98,10 @@ pub trait ReadVol: BaseVol {
     /// Get a reference to the voxel at the provided position in the volume.
     fn get(&self, pos: Vec3<i32>) -> Result<&Self::Vox, Self::Error>;
 
+    /// Get a reference to the voxel at the provided position in the volume. Many volumes provide a fast path,
+    /// provided the position is always in-bounds. Note that this function is still safe.
+    fn get_unchecked(&self, pos: Vec3<i32>) -> &Self::Vox { self.get(pos).unwrap() }
+
     /// NOTE: By default, this ray will simply run from `from` to `to` without
     /// stopping.  To make something interesting happen, call `until` or
     /// `for_each`.

common/src/volumes/chunk.rs

@@ -307,6 +307,28 @@ impl<V, S: VolSize, M> ReadVol for Chunk<V, S, M> {
             Ok(self.get_unchecked(pos))
         }
     }
+
+    #[inline(always)]
+    fn get_unchecked(&self, pos: Vec3<i32>) -> &Self::Vox {
+        self.get_unchecked(pos)
+    }
+
+    fn for_each_in(&self, mut aabb: Aabb<i32>, mut f: impl FnMut(Vec3<i32>, Self::Vox))
+    where
+        Self::Vox: Copy,
+    {
+        aabb.intersect(Aabb {
+            min: Vec3::zero(),
+            max: S::SIZE.map(|e| e as i32) - 1,
+        });
+        for z in aabb.min.z..aabb.max.z + 1 {
+            for y in aabb.min.y..aabb.max.y + 1 {
+                for x in aabb.min.x..aabb.max.x + 1 {
+                    f(Vec3::new(x, y, z), *self.get_unchecked(Vec3::new(x, y, z)));
+                }
+            }
+        }
+    }
 }
 
 impl<V: Clone + PartialEq, S: VolSize, M> WriteVol for Chunk<V, S, M> {

common/src/volumes/vol_grid_2d.rs

@@ -2,6 +2,7 @@ use crate::{
     terrain::MapSizeLg,
     vol::{BaseVol, ReadVol, RectRasterableVol, SampleVol, WriteVol},
     volumes::dyna::DynaError,
+    util::GridHasher,
 };
 use hashbrown::{hash_map, HashMap};
 use std::{fmt::Debug, ops::Deref, sync::Arc};
@@ -24,7 +25,7 @@ pub struct VolGrid2d<V: RectRasterableVol> {
     map_size_lg: MapSizeLg,
     /// Default voxel for use outside of max map bounds.
     default: Arc<V>,
-    chunks: HashMap<Vec2<i32>, Arc<V>>,
+    chunks: HashMap<Vec2<i32>, Arc<V>, GridHasher>,
 }
 
 impl<V: RectRasterableVol> VolGrid2d<V> {
@@ -64,9 +65,10 @@ impl<V: RectRasterableVol + ReadVol + Debug> ReadVol for VolGrid2d<V> {
         let ck = Self::chunk_key(pos);
         self.get_key(ck)
             .ok_or(VolGrid2dError::NoSuchChunk)
-            .and_then(|chunk| {
+            .map(|chunk| {
                 let co = Self::chunk_offs(pos);
-                chunk.get(co).map_err(VolGrid2dError::ChunkError)
+                // Always within bounds of the chunk, so we can use the get_unchecked form
+                chunk.get_unchecked(co)
             })
     }
 
@@ -271,7 +273,7 @@ impl<'a, V: RectRasterableVol + ReadVol> CachedVolGrid2d<'a, V> {
             chunk
         };
         let co = VolGrid2d::<V>::chunk_offs(pos);
-        chunk.get(co).map_err(VolGrid2dError::ChunkError)
+        Ok(chunk.get_unchecked(co))
     }
 }
 

common/systems/src/phys.rs

@@ -1460,11 +1460,9 @@ fn box_voxel_collision<'a, T: BaseVol<Vox = Block> + ReadVol>(
             // Calculate the world space near aabb
             let near_aabb = move_aabb(near_aabb, pos.0);
             let player_overlap = |block_aabb: Aabb<f32>| {
-                ordered_float::OrderedFloat(
-                    (block_aabb.center() - player_aabb.center() - Vec3::unit_z() * 0.5)
-                        .map(f32::abs)
-                        .sum(),
-                )
+                (block_aabb.center() - player_aabb.center() - Vec3::unit_z() * 0.5)
+                    .map(f32::abs)
+                    .sum()
             };
 
             terrain.for_each_in(near_aabb, |block_pos, block| {
@@ -1479,19 +1477,16 @@ fn box_voxel_collision<'a, T: BaseVol<Vox = Block> + ReadVol>(
 
                     // Determine whether the block's AABB collides with the player's AABB
                     if block_aabb.collides_with_aabb(player_aabb) {
-                        most_colliding = match most_colliding {
+                        match &most_colliding {
                             // Select the minimum of the value from `player_overlap`
-                            other @ Some((_, other_block_aabb, _))
+                            Some((_, other_block_aabb, _))
                                 if {
                                     // TODO: comment below is outdated (as of ~1 year ago)
                                     // Find the maximum of the minimum collision axes (this bit
                                     // is weird, trust me that it works)
-                                    player_overlap(block_aabb) >= player_overlap(other_block_aabb)
-                                } =>
-                            {
-                                other
-                            },
-                            _ => Some((block_pos, block_aabb, block)),
+                                    player_overlap(block_aabb) >= player_overlap(*other_block_aabb)
+                                } => {},
+                            _ => most_colliding = Some((block_pos, block_aabb, block)),
                         }
                     }
                 }