fix(simd-util): allow different align offsets (#612)

This handles the case where `prev` and `current` from
`copy_and_get_diff` have
different align offsets. On my machine (aarch64-unknown-linux-gnu) the
code
still usually takes the "optimized" path where the align offsets are the
same
(`before_prev.len() == before_current.len()`), so this shouldn't cause
significant
performance issues.

Fixes #608

---------

Co-authored-by: Andrew Gazelka <andrew.gazelka@gmail.com>

Author: TestingPlant

Cargo.lock

@@ -205,6 +205,9 @@ version = '0.10.6'
 [workspace.dependencies.sha2]
 version = '0.10.8'
 
+[workspace.dependencies.simd-utils]
+path = 'crates/simd-utils'
+
 [workspace.dependencies.tokio-util]
 features = ['full']
 version = '0.7.12'

Cargo.toml

@@ -23,6 +23,7 @@ anyhow = {workspace = true}
 base64 = {workspace = true}
 geometry = {workspace = true}
 bitfield-struct = {workspace = true}
+simd-utils = {workspace = true}
 bitvec = {workspace = true}
 heck = {workspace = true}
 papaya = {workspace = true}

crates/hyperion/Cargo.toml

@@ -31,6 +31,7 @@ unsafe extern "C-unwind" fn aligned_malloc(size: ecs_size_t) -> *mut core::ffi::
     };
 
     let ptr = unsafe { alloc(layout) };
+
     if ptr.is_null() {
         return null_mut();
     }
@@ -50,6 +51,7 @@ unsafe extern "C-unwind" fn aligned_calloc(size: ecs_size_t) -> *mut core::ffi::
             std::ptr::write_bytes(ptr, 0, size as usize);
         };
     }
+
     ptr
 }
 

crates/hyperion/src/alloc.rs

@@ -40,27 +40,61 @@ impl Module for EntityStateSyncModule {
             .singleton()
             .multi_threaded()
             .kind::<flecs::pipeline::OnStore>()
-            .each_entity(|entity, (compose, net, prev, current)| {
-                let world = entity.world();
-                if prev == current {
-                    return;
+            .run(|mut table| {
+                while table.next() {
+                    let count = table.count();
+                    let world = table.world();
+
+                    unsafe {
+                        const _: () = assert!(size_of::<Xp>() == size_of::<u16>());
+                        const _: () = assert!(align_of::<Xp>() == align_of::<u16>());
+
+                        /// Number of lanes in the SIMD vector
+                        const LANES: usize = 32; // up to AVX512
+
+                        let compose = table.field_unchecked::<Compose>(0);
+                        let compose = compose.first().unwrap();
+
+                        let net = table.field_unchecked::<NetworkStreamRef>(1);
+                        let net = net.get(..).unwrap();
+
+                        let mut prev_xp = table.field_unchecked::<Xp>(2);
+                        let prev_xp = prev_xp.get_mut(..).unwrap();
+                        let prev_xp: &mut [u16] =
+                            core::slice::from_raw_parts_mut(prev_xp.as_mut_ptr().cast(), count);
+
+                        let mut xp = table.field_unchecked::<Xp>(3);
+                        let xp = xp.get_mut(..).unwrap();
+                        let xp: &mut [u16] =
+                            core::slice::from_raw_parts_mut(xp.as_mut_ptr().cast(), count);
+
+                        simd_utils::copy_and_get_diff::<_, LANES>(
+                            prev_xp,
+                            xp,
+                            |idx, prev, current| {
+                                debug_assert!(prev != current);
+
+                                let net = net.get(idx).unwrap();
+
+                                let current = Xp::from(*current);
+                                let visual = current.get_visual();
+
+                                let packet = play::ExperienceBarUpdateS2c {
+                                    bar: visual.prop,
+                                    level: VarInt(i32::from(visual.level)),
+                                    total_xp: VarInt::default(),
+                                };
+
+                                let entity = table.entity(idx);
+                                entity.modified::<Xp>();
+
+                                compose
+                                    .unicast(&packet, *net, SystemId(100), &world)
+                                    .unwrap();
+                            },
+                        );
+                    }
                 }
-
-                entity.modified::<Xp>();
-
-                let visual = current.get_visual();
-
-                let packet = play::ExperienceBarUpdateS2c {
-                    bar: visual.prop,
-                    level: VarInt(i32::from(visual.level)),
-                    total_xp: VarInt::default(),
-                };
-
-                compose
-                    .unicast(&packet, *net, SystemId(100), &world)
-                    .unwrap();
-
-                *prev = *current;
             });
 
         system!("entity_metadata_sync", world, &Compose($), &mut MetadataChanges)

crates/hyperion/src/egress/sync_entity_state.rs

@@ -23,6 +23,7 @@
 #![feature(array_chunks)]
 #![feature(portable_simd)]
 #![feature(trivial_bounds)]
+#![feature(pointer_is_aligned_to)]
 
 pub const NUM_THREADS: usize = 8;
 pub const CHUNK_HEIGHT_SPAN: u32 = 384; // 512; // usually 384
@@ -39,7 +40,10 @@ use std::{
 use anyhow::{Context, bail};
 use derive_more::{Deref, DerefMut};
 use egress::EgressModule;
-use flecs_ecs::prelude::*;
+use flecs_ecs::{
+    prelude::*,
+    sys::{ecs_os_get_api, ecs_os_set_api, ecs_os_set_api_defaults},
+};
 pub use glam;
 use glam::IVec2;
 use ingress::IngressModule;
@@ -197,19 +201,19 @@ impl Hyperion {
                 .map_err(|_| anyhow::anyhow!("failed to create compression level"))?,
         });
 
-        // unsafe {
-        //     ecs_os_set_api_defaults();
-        //     let mut os_api = ecs_os_get_api();
-        //
-        //     let (malloc, calloc, realloc, free) = alloc::setup_custom_allocators();
-        //
-        //     os_api.malloc_ = malloc;
-        //     os_api.calloc_ = calloc;
-        //     os_api.realloc_ = realloc;
-        //     os_api.free_ = free;
-        //
-        //     ecs_os_set_api(&mut os_api);
-        // }
+        unsafe {
+            ecs_os_set_api_defaults();
+            let mut os_api = ecs_os_get_api();
+
+            let (malloc, calloc, realloc, free) = alloc::setup_custom_allocators();
+
+            os_api.malloc_ = malloc;
+            os_api.calloc_ = calloc;
+            os_api.realloc_ = realloc;
+            os_api.free_ = free;
+
+            ecs_os_set_api(&mut os_api);
+        }
 
         let world = World::new();