Auto align in hercules box

Cargo.lock

@@ -1065,6 +1065,9 @@ dependencies = [
 [[package]]
 name = "hercules_rt"
 version = "0.1.0"
+dependencies = [
+ "aligned-vec",
+]
 
 [[package]]
 name = "hercules_tests"

hercules_rt/Cargo.toml

@@ -9,4 +9,4 @@ cuda = []
 debug = []
 
 [dependencies]
-
+aligned-vec = "*"
\ No newline at end of file

hercules_rt/src/lib.rs

@@ -4,9 +4,10 @@ use std::alloc::{alloc, dealloc, Layout};
 use std::marker::PhantomData;
 use std::ptr::{copy_nonoverlapping, write_bytes, NonNull};
 use std::slice::{from_raw_parts, from_raw_parts_mut};
-
 use std::sync::OnceLock;
 
+use aligned_vec::AVec;
+
 /*
  * Define supporting types, functions, and macros for Hercules RT functions. For
  * a more in-depth discussion of the design of these utilities, see hercules_cg/
@@ -463,7 +464,7 @@ unsafe impl Sync for __RawPtrSendSync {}
  */
 pub struct HerculesImmBox<'a, T> {
     #[allow(dead_code)]
-    cpu_alloc: OnceLock<Vec<T>>,
+    cpu_alloc: OnceLock<AVec<T>>,
     #[cfg(feature = "cuda")]
     cuda_alloc: OnceLock<CUDABox>,
 
@@ -472,16 +473,32 @@ pub struct HerculesImmBox<'a, T> {
     cuda_ref: OnceLock<HerculesCUDARef<'a>>,
 }
 
-impl<'a, T> From<&'a [T]> for HerculesImmBox<'a, T> {
+impl<'a, T: Clone> From<&'a [T]> for HerculesImmBox<'a, T> {
     fn from(value: &'a [T]) -> Self {
-        HerculesImmBox {
-            cpu_alloc: OnceLock::new(),
-            #[cfg(feature = "cuda")]
-            cuda_alloc: OnceLock::new(),
-
-            cpu_ref: OnceLock::from(HerculesCPURef::from_slice(value)),
-            #[cfg(feature = "cuda")]
-            cuda_ref: OnceLock::new(),
+        if value.as_ptr().is_aligned_to(32) {
+            HerculesImmBox {
+                cpu_alloc: OnceLock::new(),
+                #[cfg(feature = "cuda")]
+                cuda_alloc: OnceLock::new(),
+
+                cpu_ref: OnceLock::from(HerculesCPURef::from_slice(value)),
+                #[cfg(feature = "cuda")]
+                cuda_ref: OnceLock::new(),
+            }
+        } else {
+            let cpu_alloc = AVec::from_slice(32, value);
+            let size = value.len() * size_of::<T>();
+            let cpu_ref =
+                unsafe { HerculesCPURef::__from_parts(cpu_alloc.as_ptr() as *mut u8, size) };
+            HerculesImmBox {
+                cpu_alloc: OnceLock::from(cpu_alloc),
+                #[cfg(feature = "cuda")]
+                cuda_alloc: OnceLock::new(),
+
+                cpu_ref: OnceLock::from(cpu_ref),
+                #[cfg(feature = "cuda")]
+                cuda_ref: OnceLock::new(),
+            }
         }
     }
 }