diff --git a/.github/workflows/nalgebra-ci-build.yml b/.github/workflows/nalgebra-ci-build.yml
index dd8d23bb..ee62301d 100644
--- a/.github/workflows/nalgebra-ci-build.yml
+++ b/.github/workflows/nalgebra-ci-build.yml
@@ -36,14 +36,20 @@ jobs:
       run: cargo build;
     - name: Build --features serde-serialize
       run: cargo build --features serde-serialize
-    - name: Build --all-features
-      run: cargo build --all-features;
-    - name: Build nalgebra-glm
-      run: cargo build -p nalgebra-glm --all-features;
     - name: Build nalgebra-lapack
       run: cd nalgebra-lapack; cargo build;
     - name: Build nalgebra-sparse
       run: cd nalgebra-sparse; cargo build;
+  # Run this on it’s own job because it alone takes a lot of time.
+  # So it’s best to let it run in parallel to the other jobs.
+  build-nalgebra-all-features:
+    runs-on: ubuntu-latest
+    steps:
+      # Needed because the --all-features build which enables cuda support.
+      - uses: Jimver/cuda-toolkit@v0.2.4
+      - uses: actions/checkout@v2
+      - run: cargo build --all-features;
+      - run: cargo build -p nalgebra-glm --all-features;
   test-nalgebra:
     runs-on: ubuntu-latest
 #    env:
@@ -110,3 +116,16 @@ jobs:
         run: xargo build --verbose --no-default-features -p nalgebra-glm --target=x86_64-unknown-linux-gnu;
       - name: build thumbv7em-none-eabihf nalgebra-glm
         run: xargo build --verbose --no-default-features -p nalgebra-glm --target=thumbv7em-none-eabihf;
+  build-cuda:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: Jimver/cuda-toolkit@v0.2.4
+      - name: Install nightly-2021-10-17
+        uses: actions-rs/toolchain@v1
+        with:
+          toolchain: nightly-2021-10-17
+          override: true
+      - uses: actions/checkout@v2
+      - run: rustup target add nvptx64-nvidia-cuda
+      - run: cargo build --no-default-features --features cuda
+      - run: cargo build --no-default-features --features cuda --target=nvptx64-nvidia-cuda
\ No newline at end of file
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 281c8f8e..70b4958a 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,17 +4,31 @@ documented here.
 
 This project adheres to [Semantic Versioning](https://semver.org/).
 
-## [0.29.1] - WIP
+## [0.30.0]
+
+### Breaking changes
+- The `Dim` trait is now marked as unsafe.
+
+### Modified
+- Use more concise debug impls for matrices and geometric transformation types.
 
 ### Added
-
-- The conversion trait `From<Vec<T>>` and method `from_vec_storage` for `RowDVector`. See [#975](https://github.com/dimforge/nalgebra/issues/975)
+- Added the conversion trait `From<Vec<T>>` and method `from_vec_storage` for `RowDVector`.
 - Added implementation of `From` and `Into` for converting between `nalgebra` types and types from
   `glam 0.18`. These can be enabled by enabling the `convert-glam018` cargo features.
 - Added the methods `Matrix::product`, `::row_product`, `::row_product_tr`, and `::column_product` to compute the
   product of the components, rows, or columns, of a single matrix or vector.
 - The `Default` trait is now implemented for most geometric types: `Point`, `Isometry`, `Rotation`, `Similarity`,
   `Transform`, `UnitComplex`, and `UnitQuaternion`.
+- Added the `Scale` geometric type for representing non-uniform scaling.
+- `nalgebra-sparse`: provide constructors for unsorted but otherwise valid data using the CSR format.
+- Added `Cholesky::new_with_substitute` that will replace diagonal elements by a given constant whenever `Cholesky`
+  meets a non-definite-positiveness.
+
+### Fixed
+- Fixed a potential unsoundness with `matrix.get(i)` and `matrix.get_mut(i)` where `i`  is an `usize`, and `matrix`
+  is a matrix slice with non-default strides.
+- Fixed potential unsoundness with `vector.perp` where `vector` isn’t actually a 2D vector as expected.
 
 ## [0.29.0]
 ### Breaking changes
diff --git a/Cargo.toml b/Cargo.toml
index f1e50ab9..61f0fd25 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -32,6 +32,7 @@ compare = [ "matrixcompare-core" ]
 libm    = [ "simba/libm" ]
 libm-force = [ "simba/libm_force" ]
 macros = [ "nalgebra-macros" ]
+cuda   = [ "cust", "simba/cuda" ]
 
 # Conversion
 convert-mint = [ "mint" ]
@@ -86,7 +87,7 @@ pest           = { version = "2", optional = true }
 pest_derive    = { version = "2", optional = true }
 bytemuck       = { version = "1.5", optional = true }
 matrixcompare-core = { version = "0.1", optional = true }
-proptest           = { version = "1", optional = true, default-features = false, features = ["std"] }
+proptest       = { version = "1", optional = true, default-features = false, features = ["std"] }
 glam013        = { package = "glam", version = "0.13", optional = true }
 glam014        = { package = "glam", version = "0.14", optional = true }
 glam015        = { package = "glam", version = "0.15", optional = true }
@@ -94,6 +95,9 @@ glam016        = { package = "glam", version = "0.16", optional = true }
 glam017        = { package = "glam", version = "0.17", optional = true }
 glam018        = { package = "glam", version = "0.18", optional = true }
 
+[target.'cfg(not(target_os = "cuda"))'.dependencies]
+cust           = { version = "0.1", optional = true }
+
 
 [dev-dependencies]
 serde_json = "1.0"
@@ -129,3 +133,6 @@ lto = true
 [package.metadata.docs.rs]
 # Enable certain features when building docs for docs.rs
 features = [ "proptest-support", "compare", "macros", "rand" ]
+
+[patch.crates-io]
+simba = { git = "https://github.com/dimforge/simba"}
diff --git a/nalgebra-glm/Cargo.toml b/nalgebra-glm/Cargo.toml
index edad31d7..cfe83bdc 100644
--- a/nalgebra-glm/Cargo.toml
+++ b/nalgebra-glm/Cargo.toml
@@ -22,6 +22,7 @@ std                   = [ "nalgebra/std", "simba/std" ]
 arbitrary             = [ "nalgebra/arbitrary" ]
 serde-serialize       = [ "nalgebra/serde-serialize-no-std" ]
 abomonation-serialize = [ "nalgebra/abomonation-serialize" ]
+cuda                  = [ "nalgebra/cuda" ]
 
 # Conversion
 convert-mint = [ "nalgebra/mint" ]
diff --git a/src/base/array_storage.rs b/src/base/array_storage.rs
index 3fc88ade..b46d442f 100644
--- a/src/base/array_storage.rs
+++ b/src/base/array_storage.rs
@@ -32,6 +32,10 @@ use std::mem;
 /// A array-based statically sized matrix data storage.
 #[repr(transparent)]
 #[derive(Copy, Clone, PartialEq, Eq, Hash)]
+#[cfg_attr(
+    all(not(target_os = "cuda"), feature = "cuda"),
+    derive(cust::DeviceCopy)
+)]
 pub struct ArrayStorage<T, const R: usize, const C: usize>(pub [[T; R]; C]);
 
 impl<T, const R: usize, const C: usize> ArrayStorage<T, R, C> {
diff --git a/src/base/dimension.rs b/src/base/dimension.rs
index 9a51536c..86006f3d 100644
--- a/src/base/dimension.rs
+++ b/src/base/dimension.rs
@@ -13,6 +13,10 @@ use serde::{Deserialize, Deserializer, Serialize, Serializer};
 
 /// Dim of dynamically-sized algebraic entities.
 #[derive(Clone, Copy, Eq, PartialEq, Debug)]
+#[cfg_attr(
+    all(not(target_os = "cuda"), feature = "cuda"),
+    derive(cust::DeviceCopy)
+)]
 pub struct Dynamic {
     value: usize,
 }
@@ -197,6 +201,10 @@ dim_ops!(
 );
 
 #[derive(Debug, Copy, Clone, PartialEq, Eq, Hash)]
+#[cfg_attr(
+    all(not(target_os = "cuda"), feature = "cuda"),
+    derive(cust::DeviceCopy)
+)]
 pub struct Const<const R: usize>;
 
 /// Trait implemented exclusively by type-level integers.
diff --git a/src/base/indexing.rs b/src/base/indexing.rs
index c2bb48ff..2c691bd1 100644
--- a/src/base/indexing.rs
+++ b/src/base/indexing.rs
@@ -589,7 +589,10 @@ where
     where
         S: RawStorageMut<T, R, C>,
     {
-        matrix.data.get_unchecked_linear_mut(self)
+        let nrows = matrix.shape().0;
+        let row = self % nrows;
+        let col = self / nrows;
+        matrix.data.get_unchecked_mut(row, col)
     }
 }
 
diff --git a/src/base/matrix.rs b/src/base/matrix.rs
index 983f9a86..5df2f78e 100644
--- a/src/base/matrix.rs
+++ b/src/base/matrix.rs
@@ -154,6 +154,10 @@ pub type MatrixCross<T, R1, C1, R2, C2> =
 /// some concrete types for `T` and a compatible data storage type `S`).
 #[repr(C)]
 #[derive(Clone, Copy)]
+#[cfg_attr(
+    all(not(target_os = "cuda"), feature = "cuda"),
+    derive(cust::DeviceCopy)
+)]
 pub struct Matrix<T, R, C, S> {
     /// The data storage that contains all the matrix components. Disappointed?
     ///
diff --git a/src/base/unit.rs b/src/base/unit.rs
index 8d548687..22af401a 100644
--- a/src/base/unit.rs
+++ b/src/base/unit.rs
@@ -26,6 +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(
+    all(not(target_os = "cuda"), feature = "cuda"),
+    derive(cust::DeviceCopy)
+)]
 pub struct Unit<T> {
     pub(crate) value: T,
 }
diff --git a/src/geometry/dual_quaternion.rs b/src/geometry/dual_quaternion.rs
index e597cbf5..4280668a 100644
--- a/src/geometry/dual_quaternion.rs
+++ b/src/geometry/dual_quaternion.rs
@@ -39,6 +39,10 @@ use simba::scalar::{ClosedNeg, RealField};
 ///  See <https://github.com/dimforge/nalgebra/issues/487>
 #[repr(C)]
 #[derive(Debug, Copy, Clone)]
+#[cfg_attr(
+    all(not(target_os = "cuda"), feature = "cuda"),
+    derive(cust::DeviceCopy)
+)]
 pub struct DualQuaternion<T> {
     /// The real component of the quaternion
     pub real: Quaternion<T>,
diff --git a/src/geometry/isometry.rs b/src/geometry/isometry.rs
index bdd2de60..f020c0e9 100755
--- a/src/geometry/isometry.rs
+++ b/src/geometry/isometry.rs
@@ -54,7 +54,11 @@ use crate::geometry::{AbstractRotation, Point, Translation};
 /// * [Conversion to a matrix <span style="float:right;">`to_matrix`…</span>](#conversion-to-a-matrix)
 ///
 #[repr(C)]
-#[derive(Debug)]
+#[derive(Debug, Copy, Clone)]
+#[cfg_attr(
+    all(not(target_os = "cuda"), feature = "cuda"),
+    derive(cust::DeviceCopy)
+)]
 #[cfg_attr(feature = "serde-serialize-no-std", derive(Serialize, Deserialize))]
 #[cfg_attr(
     feature = "serde-serialize-no-std",
@@ -170,20 +174,6 @@ where
     }
 }
 
-impl<T: Scalar + Copy, R: Copy, const D: usize> Copy for Isometry<T, R, D> where
-    Owned<T, Const<D>>: Copy
-{
-}
-
-impl<T: Scalar, R: Clone, const D: usize> Clone for Isometry<T, R, D> {
-    #[inline]
-    fn clone(&self) -> Self {
-        Self {
-            rotation: self.rotation.clone(),
-            translation: self.translation.clone(),
-        }
-    }
-}
 /// # From the translation and rotation parts
 impl<T: Scalar, R: AbstractRotation<T, D>, const D: usize> Isometry<T, R, D> {
     /// Creates a new isometry from its rotational and translational parts.
diff --git a/src/geometry/orthographic.rs b/src/geometry/orthographic.rs
index d4008173..18a7852d 100644
--- a/src/geometry/orthographic.rs
+++ b/src/geometry/orthographic.rs
@@ -19,19 +19,15 @@ use crate::geometry::{Point3, Projective3};
 
 /// A 3D orthographic projection stored as a homogeneous 4x4 matrix.
 #[repr(C)]
+#[cfg_attr(
+    all(not(target_os = "cuda"), feature = "cuda"),
+    derive(cust::DeviceCopy)
+)]
+#[derive(Copy, Clone)]
 pub struct Orthographic3<T> {
     matrix: Matrix4<T>,
 }
 
-impl<T: RealField + Copy> Copy for Orthographic3<T> {}
-
-impl<T: RealField> Clone for Orthographic3<T> {
-    #[inline]
-    fn clone(&self) -> Self {
-        Self::from_matrix_unchecked(self.matrix.clone())
-    }
-}
-
 impl<T: RealField> fmt::Debug for Orthographic3<T> {
     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> Result<(), fmt::Error> {
         self.matrix.fmt(f)
diff --git a/src/geometry/perspective.rs b/src/geometry/perspective.rs
index 34af6f0b..59b7f9f2 100644
--- a/src/geometry/perspective.rs
+++ b/src/geometry/perspective.rs
@@ -20,19 +20,15 @@ use crate::geometry::{Point3, Projective3};
 
 /// A 3D perspective projection stored as a homogeneous 4x4 matrix.
 #[repr(C)]
+#[cfg_attr(
+    all(not(target_os = "cuda"), feature = "cuda"),
+    derive(cust::DeviceCopy)
+)]
+#[derive(Copy, Clone)]
 pub struct Perspective3<T> {
     matrix: Matrix4<T>,
 }
 
-impl<T: RealField + Copy> Copy for Perspective3<T> {}
-
-impl<T: RealField> Clone for Perspective3<T> {
-    #[inline]
-    fn clone(&self) -> Self {
-        Self::from_matrix_unchecked(self.matrix.clone())
-    }
-}
-
 impl<T: RealField> fmt::Debug for Perspective3<T> {
     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> Result<(), fmt::Error> {
         self.matrix.fmt(f)
diff --git a/src/geometry/point.rs b/src/geometry/point.rs
index 7930e8fb..0953a9ce 100644
--- a/src/geometry/point.rs
+++ b/src/geometry/point.rs
@@ -74,6 +74,15 @@ where
 {
 }
 
+#[cfg(all(not(target_os = "cuda"), feature = "cuda"))]
+unsafe impl<T: Scalar + cust::memory::DeviceCopy, D: DimName> cust::memory::DeviceCopy
+    for OPoint<T, D>
+where
+    DefaultAllocator: Allocator<T, D>,
+    OVector<T, D>: cust::memory::DeviceCopy,
+{
+}
+
 #[cfg(feature = "bytemuck")]
 unsafe impl<T: Scalar, D: DimName> bytemuck::Zeroable for OPoint<T, D>
 where
diff --git a/src/geometry/quaternion.rs b/src/geometry/quaternion.rs
index 968a9e18..05b73334 100755
--- a/src/geometry/quaternion.rs
+++ b/src/geometry/quaternion.rs
@@ -28,6 +28,10 @@ use crate::geometry::{Point3, Rotation};
 /// that may be used as a rotation.
 #[repr(C)]
 #[derive(Copy, Clone)]
+#[cfg_attr(
+    all(not(target_os = "cuda"), feature = "cuda"),
+    derive(cust::DeviceCopy)
+)]
 pub struct Quaternion<T> {
     /// This quaternion as a 4D vector of coordinates in the `[ x, y, z, w ]` storage order.
     pub coords: Vector4<T>,
diff --git a/src/geometry/rotation.rs b/src/geometry/rotation.rs
index e1da8d67..2d5e5d24 100755
--- a/src/geometry/rotation.rs
+++ b/src/geometry/rotation.rs
@@ -54,6 +54,11 @@ use crate::geometry::Point;
 /// * [Conversion to a matrix <span style="float:right;">`matrix`, `to_homogeneous`…</span>](#conversion-to-a-matrix)
 ///
 #[repr(C)]
+#[cfg_attr(
+    all(not(target_os = "cuda"), feature = "cuda"),
+    derive(cust::DeviceCopy)
+)]
+#[derive(Copy, Clone)]
 pub struct Rotation<T, const D: usize> {
     matrix: SMatrix<T, D, D>,
 }
@@ -73,21 +78,6 @@ where
     }
 }
 
-impl<T: Scalar + Copy, const D: usize> Copy for Rotation<T, D> where
-    <DefaultAllocator as Allocator<T, Const<D>, Const<D>>>::Buffer: Copy
-{
-}
-
-impl<T: Scalar, const D: usize> Clone for Rotation<T, D>
-where
-    <DefaultAllocator as Allocator<T, Const<D>, Const<D>>>::Buffer: Clone,
-{
-    #[inline]
-    fn clone(&self) -> Self {
-        Self::from_matrix_unchecked(self.matrix.clone())
-    }
-}
-
 #[cfg(feature = "bytemuck")]
 unsafe impl<T, const D: usize> bytemuck::Zeroable for Rotation<T, D>
 where
diff --git a/src/geometry/scale.rs b/src/geometry/scale.rs
index 9f3e778c..b1d278d6 100755
--- a/src/geometry/scale.rs
+++ b/src/geometry/scale.rs
@@ -22,6 +22,11 @@ use crate::geometry::Point;
 
 /// A scale which supports non-uniform scaling.
 #[repr(C)]
+#[cfg_attr(
+    all(not(target_os = "cuda"), feature = "cuda"),
+    derive(cust::DeviceCopy)
+)]
+#[derive(Copy, Clone)]
 pub struct Scale<T, const D: usize> {
     /// The scale coordinates, i.e., how much is multiplied to a point's coordinates when it is
     /// scaled.
@@ -43,18 +48,6 @@ where
     }
 }
 
-impl<T: Scalar + Copy, const D: usize> Copy for Scale<T, D> {}
-
-impl<T: Scalar, const D: usize> Clone for Scale<T, D>
-where
-    Owned<T, Const<D>>: Clone,
-{
-    #[inline]
-    fn clone(&self) -> Self {
-        Scale::from(self.vector.clone())
-    }
-}
-
 #[cfg(feature = "bytemuck")]
 unsafe impl<T, const D: usize> bytemuck::Zeroable for Scale<T, D>
 where
diff --git a/src/geometry/similarity.rs b/src/geometry/similarity.rs
index 1b685252..fd6b1d36 100755
--- a/src/geometry/similarity.rs
+++ b/src/geometry/similarity.rs
@@ -23,7 +23,11 @@ use crate::geometry::{AbstractRotation, Isometry, Point, Translation};
 
 /// A similarity, i.e., an uniform scaling, followed by a rotation, followed by a translation.
 #[repr(C)]
-#[derive(Debug)]
+#[derive(Debug, Copy, Clone)]
+#[cfg_attr(
+    all(not(target_os = "cuda"), feature = "cuda"),
+    derive(cust::DeviceCopy)
+)]
 #[cfg_attr(feature = "serde-serialize-no-std", derive(Serialize, Deserialize))]
 #[cfg_attr(
     feature = "serde-serialize-no-std",
@@ -73,22 +77,6 @@ where
     }
 }
 
-impl<T: Scalar + Copy + Zero, R: AbstractRotation<T, D> + Copy, const D: usize> Copy
-    for Similarity<T, R, D>
-where
-    Owned<T, Const<D>>: Copy,
-{
-}
-
-impl<T: Scalar + Zero, R: AbstractRotation<T, D> + Clone, const D: usize> Clone
-    for Similarity<T, R, D>
-{
-    #[inline]
-    fn clone(&self) -> Self {
-        Similarity::from_isometry(self.isometry.clone(), self.scaling.clone())
-    }
-}
-
 impl<T: Scalar + Zero, R, const D: usize> Similarity<T, R, D>
 where
     R: AbstractRotation<T, D>,
diff --git a/src/geometry/transform.rs b/src/geometry/transform.rs
index d87a815d..b0b5cced 100755
--- a/src/geometry/transform.rs
+++ b/src/geometry/transform.rs
@@ -60,14 +60,26 @@ where
 
 /// Tag representing the most general (not necessarily inversible) `Transform` type.
 #[derive(Debug, Copy, Clone, Hash, PartialEq, Eq)]
+#[cfg_attr(
+    all(not(target_os = "cuda"), feature = "cuda"),
+    derive(cust::DeviceCopy)
+)]
 pub enum TGeneral {}
 
 /// Tag representing the most general inversible `Transform` type.
 #[derive(Debug, Copy, Clone, Hash, PartialEq, Eq)]
+#[cfg_attr(
+    all(not(target_os = "cuda"), feature = "cuda"),
+    derive(cust::DeviceCopy)
+)]
 pub enum TProjective {}
 
 /// Tag representing an affine `Transform`. Its bottom-row is equal to `(0, 0 ... 0, 1)`.
 #[derive(Debug, Copy, Clone, Hash, PartialEq, Eq)]
+#[cfg_attr(
+    all(not(target_os = "cuda"), feature = "cuda"),
+    derive(cust::DeviceCopy)
+)]
 pub enum TAffine {}
 
 impl TCategory for TGeneral {
@@ -195,6 +207,16 @@ where
 {
 }
 
+#[cfg(all(not(target_os = "cuda"), feature = "cuda"))]
+unsafe impl<T: RealField + cust::memory::DeviceCopy, C: TCategory, const D: usize>
+    cust::memory::DeviceCopy for Transform<T, C, D>
+where
+    Const<D>: DimNameAdd<U1>,
+    DefaultAllocator: Allocator<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>,
+    Owned<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>: cust::memory::DeviceCopy,
+{
+}
+
 impl<T: RealField, C: TCategory, const D: usize> Clone for Transform<T, C, D>
 where
     Const<D>: DimNameAdd<U1>,
diff --git a/src/geometry/translation.rs b/src/geometry/translation.rs
index 48656f9c..1ce8cba3 100755
--- a/src/geometry/translation.rs
+++ b/src/geometry/translation.rs
@@ -22,6 +22,11 @@ use crate::geometry::Point;
 
 /// A translation.
 #[repr(C)]
+#[cfg_attr(
+    all(not(target_os = "cuda"), feature = "cuda"),
+    derive(cust::DeviceCopy)
+)]
+#[derive(Copy, Clone)]
 pub struct Translation<T, const D: usize> {
     /// The translation coordinates, i.e., how much is added to a point's coordinates when it is
     /// translated.
@@ -43,18 +48,6 @@ where
     }
 }
 
-impl<T: Scalar + Copy, const D: usize> Copy for Translation<T, D> {}
-
-impl<T: Scalar, const D: usize> Clone for Translation<T, D>
-where
-    Owned<T, Const<D>>: Clone,
-{
-    #[inline]
-    fn clone(&self) -> Self {
-        Translation::from(self.vector.clone())
-    }
-}
-
 #[cfg(feature = "bytemuck")]
 unsafe impl<T, const D: usize> bytemuck::Zeroable for Translation<T, D>
 where