Implement DeviceCopy for UnitComplex, UnitQuaternion, and Unit<Matrix> instead of using a blanket impl

src/base/unit.rs

@@ -26,10 +26,10 @@ use crate::{Dim, Matrix, OMatrix, RealField, Scalar, SimdComplexField, SimdRealF
 /// in their documentation, read their dedicated pages directly.
 #[repr(transparent)]
 #[derive(Clone, Hash, Copy)]
-#[cfg_attr(
+// #[cfg_attr(
-    all(not(target_os = "cuda"), feature = "cuda"),
+//     all(not(target_os = "cuda"), feature = "cuda"),
-    derive(cust::DeviceCopy)
+//     derive(cust::DeviceCopy)
-)]
+// )]
 pub struct Unit<T> {
     pub(crate) value: T,
 }
@@ -122,6 +122,17 @@ mod rkyv_impl {
     }
 }
 
+#[cfg(all(not(target_os = "cuda"), feature = "cuda"))]
+unsafe impl<T: cust::memory::DeviceCopy, R, C, S> cust::memory::DeviceCopy
+    for Unit<Matrix<T, R, C, S>>
+where
+    T: Scalar,
+    R: Dim,
+    C: Dim,
+    S: RawStorage<T, R, C> + Copy,
+{
+}
+
 impl<T, R, C, S> PartialEq for Unit<Matrix<T, R, C, S>>
 where
     T: Scalar + PartialEq,

src/geometry/quaternion.rs

@@ -1068,6 +1068,9 @@ impl<T: RealField + fmt::Display> fmt::Display for Quaternion<T> {
 /// A unit quaternions. May be used to represent a rotation.
 pub type UnitQuaternion<T> = Unit<Quaternion<T>>;
 
+#[cfg(all(not(target_os = "cuda"), feature = "cuda"))]
+unsafe impl<T: cust::memory::DeviceCopy> cust::memory::DeviceCopy for UnitQuaternion<T> {}
+
 impl<T: Scalar + ClosedNeg + PartialEq> PartialEq for UnitQuaternion<T> {
     #[inline]
     fn eq(&self, rhs: &Self) -> bool {

src/geometry/unit_complex.rs

@@ -31,6 +31,9 @@ use std::cmp::{Eq, PartialEq};
 /// * [Conversion to a matrix <span style="float:right;">`to_rotation_matrix`, `to_homogeneous`…</span>](#conversion-to-a-matrix)
 pub type UnitComplex<T> = Unit<Complex<T>>;
 
+#[cfg(all(not(target_os = "cuda"), feature = "cuda"))]
+unsafe impl<T: cust::memory::DeviceCopy> cust::memory::DeviceCopy for UnitComplex<T> {}
+
 impl<T: Scalar + PartialEq> PartialEq for UnitComplex<T> {
     #[inline]
     fn eq(&self, rhs: &Self) -> bool {