Set up cava benchmark

• Set up cava benchmark w/ criterion.
• GPU schedule for cava.
• Convert a bunch of benchmarks from using hercules refs to use hercules boxes.
• Define AlignedAlloc global allocator in hercules_rt that aligns everything to 32 bytes.
• Bump nom version.

hercules_rt/src/lib.rs

-#![feature(once_cell_try, pointer_is_aligned_to)]
+#![feature(once_cell_try)]
 
-use std::alloc::{alloc, dealloc, Layout};
+use std::alloc::{alloc, dealloc, GlobalAlloc, Layout, System};
 use std::marker::PhantomData;
 use std::ptr::{copy_nonoverlapping, write_bytes, NonNull};
 use std::slice::{from_raw_parts, from_raw_parts_mut};
@@ -189,7 +189,6 @@ pub struct CUDABox {
 
 impl<'a> HerculesCPURef<'a> {
     pub fn from_slice<T>(slice: &'a [T]) -> Self {
-        assert!(slice.as_ptr().is_aligned_to(32));
         let ptr = unsafe { NonNull::new_unchecked(slice.as_ptr() as *mut u8) };
         let size = slice.len() * size_of::<T>();
         Self {
@@ -202,6 +201,7 @@ impl<'a> HerculesCPURef<'a> {
     pub fn as_slice<T>(self) -> &'a [T] {
         let ptr = self.ptr.as_ptr() as *const T;
         assert_eq!(self.size % size_of::<T>(), 0);
+        assert!(ptr.is_aligned());
         unsafe { from_raw_parts(ptr, self.size / size_of::<T>()) }
     }
 
@@ -214,7 +214,6 @@ impl<'a> HerculesCPURef<'a> {
     }
 
     pub unsafe fn __from_parts(ptr: *mut u8, size: usize) -> Self {
-        assert!(ptr.is_aligned_to(32));
         Self {
             ptr: NonNull::new(ptr).unwrap(),
             size,
@@ -225,7 +224,6 @@ impl<'a> HerculesCPURef<'a> {
 
 impl<'a> HerculesCPURefMut<'a> {
     pub fn from_slice<T>(slice: &'a mut [T]) -> Self {
-        assert!(slice.as_ptr().is_aligned_to(32));
         let ptr = unsafe { NonNull::new_unchecked(slice.as_ptr() as *mut u8) };
         let size = slice.len() * size_of::<T>();
         Self {
@@ -259,7 +257,6 @@ impl<'a> HerculesCPURefMut<'a> {
     }
 
     pub unsafe fn __from_parts(ptr: *mut u8, size: usize) -> Self {
-        assert!(ptr.is_aligned_to(32));
         Self {
             ptr: NonNull::new(ptr).unwrap(),
             size,
@@ -271,7 +268,6 @@ impl<'a> HerculesCPURefMut<'a> {
 #[cfg(feature = "cuda")]
 impl<'a> HerculesCUDARef<'a> {
     pub fn to_cpu_ref<'b, T>(self, dst: &'b mut [T]) -> HerculesCPURefMut<'b> {
-        assert!(dst.as_ptr().is_aligned_to(32));
         unsafe {
             let size = self.size;
             assert_eq!(size, dst.len() * size_of::<T>());
@@ -313,7 +309,6 @@ impl<'a> HerculesCUDARefMut<'a> {
     }
 
     pub fn to_cpu_ref<'b, T>(self, dst: &mut [T]) -> HerculesCPURefMut<'b> {
-        assert!(dst.as_ptr().is_aligned_to(32));
         unsafe {
             let size = self.size;
             let ptr = NonNull::new(dst.as_ptr() as *mut u8).unwrap();
@@ -872,3 +867,24 @@ impl<'a, T> HerculesRefInto<'a> for Box<[T]> {
         HerculesCPURef::from_slice(self)
     }
 }
+
+/*
+ * We need all allocations to be aligned to LARGEST_ALIGNMENT bytes for
+ * vectorization. This is the easiest way to do that.
+ */
+pub struct AlignedAlloc;
+
+unsafe impl GlobalAlloc for AlignedAlloc {
+    unsafe fn alloc(&self, layout: Layout) -> *mut u8 {
+        let layout = layout.align_to(LARGEST_ALIGNMENT).unwrap();
+        System.alloc(layout)
+    }
+
+    unsafe fn dealloc(&self, ptr: *mut u8, layout: Layout) {
+        let layout = layout.align_to(LARGEST_ALIGNMENT).unwrap();
+        System.dealloc(ptr, layout)
+    }
+}
+
+#[global_allocator]
+static A: AlignedAlloc = AlignedAlloc;