diff --git a/shared/src/shapes/mesh.rs b/shared/src/shapes/mesh.rs
new file mode 100644
index 0000000..cb8aae4
--- /dev/null
+++ b/shared/src/shapes/mesh.rs
@@ -0,0 +1,160 @@
+use crate::core::geometry::{Normal3f, Point2f, Point3f, Vector3f};
+use crate::utils::sampling::DevicePiecewiseConstant2D;
+use crate::utils::Transform;
+use crate::{Float, Gvec};
+
+#[repr(C)]
+#[derive(Debug, Copy, Clone)]
+pub struct TriangleMesh {
+    pub p: GVec<Point3f>,
+    pub n: GVec<Normal3f>,
+    pub s: GVec<Vector3f>,
+    pub uv: GVec<Point2f>,
+    pub vertex_indices: GVec<i32>,
+    pub face_indices: GVec<i32>,
+    pub n_triangles: u32,
+    pub n_vertices: u32,
+    pub reverse_orientation: bool,
+    pub transform_swaps_handedness: bool,
+}
+
+#[repr(C)]
+#[derive(Debug, Clone, Copy)]
+pub struct BilinearPatchMesh {
+    pub p: GVec<Point3f>,
+    pub n: GVec<Normal3f>,
+    pub uv: GVec<Point2f>,
+    pub vertex_indices: GVec<i32>,
+    pub n_patches: u32,
+    pub n_vertices: u32,
+    pub reverse_orientation: bool,
+    pub transform_swaps_handedness: bool,
+    pub image_distribution: Ptr<PiecewiseConstant2D>,
+}
+
+unsafe impl Send for TriangleMesh {}
+unsafe impl Sync for TriangleMesh {}
+unsafe impl Send for BilinearPatchMesh {}
+unsafe impl Sync for BilinearPatchMesh {}
+
+impl TriangleMesh {
+    pub fn new(
+        render_from_object: &Transform,
+        reverse_orientation: bool,
+        vertex_indices: &[i32],
+        p: &[Point3f],
+        n: &[Normal3f],
+        s: &[Vector3f],
+        uv: &[Point2f],
+        face_indices: &[i32],
+    ) -> Self {
+        let n_triangles = (vertex_indices.len() / 3) as u32;
+        let n_vertices = p.len() as u32;
+
+        let mut p_gvec = gvec_with_capacity(p.len());
+        for pt in p {
+            p_gvec.push(render_from_object.apply_to_point(*pt));
+        }
+
+        let mut n_gvec = gvec_with_capacity(n.len());
+        if !n.is_empty() {
+            assert_eq!(n_vertices as usize, n.len(), "Normal count mismatch");
+            for nn in n {
+                let mut transformed = render_from_object.apply_to_normal(*nn);
+                if reverse_orientation {
+                    transformed = -transformed;
+                }
+                n_gvec.push(transformed);
+            }
+        }
+
+        let mut s_gvec = gvec_with_capacity(s.len());
+        if !s.is_empty() {
+            assert_eq!(n_vertices as usize, s.len(), "Tangent count mismatch");
+            for ss in s {
+                s_gvec.push(render_from_object.apply_to_vector(*ss));
+            }
+        }
+
+        assert!(
+            uv.is_empty() || uv.len() == n_vertices as usize,
+            "UV count mismatch"
+        );
+        assert!(
+            face_indices.is_empty() || face_indices.len() == n_triangles as usize,
+            "Face index count mismatch"
+        );
+
+        Self {
+            vertex_indices: gvec_from_slice(vertex_indices),
+            p: p_gvec,
+            n: n_gvec,
+            s: s_gvec,
+            uv: gvec_from_slice(uv),
+            face_indices: gvec_from_slice(face_indices),
+            n_triangles,
+            n_vertices,
+            reverse_orientation,
+            transform_swaps_handedness: render_from_object.swaps_handedness(),
+        }
+    }
+
+    pub fn positions(&self) -> &[Point3f] {
+        &self.p
+    }
+    pub fn indices(&self) -> &[i32] {
+        &self.vertex_indices
+    }
+    pub fn normals(&self) -> &[Normal3f] {
+        &self.n
+    }
+    pub fn uvs(&self) -> &[Point2f] {
+        &self.uv
+    }
+}
+
+impl BilinearPatchMesh {
+    pub fn new(
+        render_from_object: &Transform,
+        reverse_orientation: bool,
+        vertex_indices: &[i32],
+        p: &[Point3f],
+        n: &[Normal3f],
+        uv: &[Point2f],
+        image_distribution: Option<PiecewiseConstant2D>,
+    ) -> Self {
+        let n_patches = (vertex_indices.len() / 4) as u32;
+        let n_vertices = p.len() as u32;
+
+        let mut p_gvec = gvec_with_capacity(p.len());
+        for pt in p {
+            p_gvec.push(render_from_object.apply_to_point(*pt));
+        }
+
+        let mut n_gvec = gvec_with_capacity(n.len());
+        if !n.is_empty() {
+            assert_eq!(n_vertices as usize, n.len());
+            for nn in n {
+                let mut transformed = render_from_object.apply_to_normal(*nn);
+                if reverse_orientation {
+                    transformed = -transformed;
+                }
+                n_gvec.push(transformed);
+            }
+        }
+
+        assert!(uv.is_empty() || uv.len() == n_vertices as usize);
+
+        Self {
+            vertex_indices: gvec_from_slice(vertex_indices),
+            p: p_gvec,
+            n: n_gvec,
+            uv: gvec_from_slice(uv),
+            image_distribution: Ptr::from(image_distribution),
+            n_patches,
+            n_vertices,
+            reverse_orientation,
+            transform_swaps_handedness: render_from_object.swaps_handedness(),
+        }
+    }
+}
diff --git a/shared/src/utils/alloc.rs b/shared/src/utils/alloc.rs
new file mode 100644
index 0000000..c6db9a1
--- /dev/null
+++ b/shared/src/utils/alloc.rs
@@ -0,0 +1,238 @@
+#![feature(allocator_api)]
+extern crate alloc;
+
+use alloc::alloc::Global;
+use core::alloc::{AllocError, Allocator, Layout};
+use core::ptr::NonNull;
+
+// CPU fallback, delegates to Global
+#[derive(Debug, Clone, Copy, Default)]
+pub struct SystemAlloc;
+
+unsafe impl Allocator for SystemAlloc {
+    fn allocate(&self, layout: Layout) -> Result<NonNull<[u8]>, AllocError> {
+        Global.allocate(layout)
+    }
+
+    unsafe fn deallocate(&self, ptr: NonNull<u8>, layout: Layout) {
+        Global.deallocate(ptr, layout)
+    }
+
+    unsafe fn grow(
+        &self,
+        ptr: NonNull<u8>,
+        old_layout: Layout,
+        new_layout: Layout,
+    ) -> Result<NonNull<[u8]>, AllocError> {
+        Global.grow(ptr, old_layout, new_layout)
+    }
+
+    unsafe fn shrink(
+        &self,
+        ptr: NonNull<u8>,
+        old_layout: Layout,
+        new_layout: Layout,
+    ) -> Result<NonNull<[u8]>, AllocError> {
+        Global.shrink(ptr, old_layout, new_layout)
+    }
+}
+
+// Unified memory via cudaMallocManaged
+
+#[cfg(feature = "cuda")]
+pub mod cuda {
+    use super::*;
+    use cust::memory::{cuda_free_unified, cuda_malloc_unified, UnifiedPointer};
+
+    #[derive(Debug, Clone, Copy, Default)]
+    pub struct CudaAlloc;
+
+    unsafe impl Allocator for CudaAlloc {
+        fn allocate(&self, layout: Layout) -> Result<NonNull<[u8]>, AllocError> {
+            if layout.size() == 0 {
+                // Zero-sized allocations: return a dangling aligned pointer
+                // with zero length, which is valid for ZSTs.
+                let ptr = NonNull::new(layout.align() as *mut u8).ok_or(AllocError)?;
+                return Ok(NonNull::slice_from_raw_parts(ptr, 0));
+            }
+
+            let ptr = cuda_malloc_unified::<u8>(layout.size()).map_err(|_| AllocError)?;
+
+            let raw = ptr.as_raw_mut();
+            core::mem::forget(ptr); // Arena owns the raw pointer, not the RAII wrapper
+
+            let nn = NonNull::new(raw).ok_or(AllocError)?;
+            Ok(NonNull::slice_from_raw_parts(nn, layout.size()))
+        }
+
+        unsafe fn deallocate(&self, ptr: NonNull<u8>, layout: Layout) {
+            if layout.size() == 0 {
+                return;
+            }
+            let _ = cuda_free_unified(UnifiedPointer::wrap(ptr.as_ptr()));
+        }
+    }
+}
+
+// Host-visible GPU memory via gpu-allocator
+#[cfg(feature = "cuda")]
+pub mod cuda {
+    use super::*;
+    use cust::memory::{cuda_free_unified, cuda_malloc_unified, UnifiedPointer};
+
+    #[derive(Debug, Clone, Copy, Default)]
+    pub struct CudaAlloc;
+
+    unsafe impl Allocator for CudaAlloc {
+        fn allocate(&self, layout: Layout) -> Result<NonNull<[u8]>, AllocError> {
+            if layout.size() == 0 {
+                // Zero-sized allocations: return a dangling aligned pointer
+                // with zero length, which is valid for ZSTs.
+                let ptr = NonNull::new(layout.align() as *mut u8).ok_or(AllocError)?;
+                return Ok(NonNull::slice_from_raw_parts(ptr, 0));
+            }
+
+            let ptr = cuda_malloc_unified::<u8>(layout.size()).map_err(|_| AllocError)?;
+
+            let raw = ptr.as_raw_mut();
+            core::mem::forget(ptr); // Arena owns the raw pointer, not the RAII wrapper
+
+            let nn = NonNull::new(raw).ok_or(AllocError)?;
+            Ok(NonNull::slice_from_raw_parts(nn, layout.size()))
+        }
+
+        unsafe fn deallocate(&self, ptr: NonNull<u8>, layout: Layout) {
+            if layout.size() == 0 {
+                return;
+            }
+            let _ = cuda_free_unified(UnifiedPointer::wrap(ptr.as_ptr()));
+        }
+    }
+}
+
+// Host-visible GPU memory via gpu-allocator
+
+#[cfg(feature = "vulkan")]
+pub mod vulkan {
+    use super::*;
+    use ash::vk;
+    use gpu_allocator::vulkan::{
+        Allocation, AllocationCreateDesc, AllocationScheme, Allocator as VkAllocator,
+        AllocatorCreateDesc,
+    };
+    use gpu_allocator::MemoryLocation;
+    use parking_lot::Mutex;
+    use std::collections::HashMap;
+    use std::sync::Arc;
+
+    /// Wraps a gpu-allocator instance. Clone is cheap
+    #[derive(Clone)]
+    pub struct VulkanAlloc {
+        inner: Arc<Mutex<VulkanInner>>,
+    }
+
+    struct VulkanInner {
+        allocator: VkAllocator,
+        allocations: HashMap<usize, Allocation>,
+    }
+
+    impl VulkanAlloc {
+        pub fn new(
+            instance: &ash::Instance,
+            device: ash::Device,
+            physical_device: vk::PhysicalDevice,
+        ) -> Self {
+            let allocator = VkAllocator::new(&AllocatorCreateDesc {
+                instance: instance.clone(),
+                device: device.clone(),
+                physical_device,
+                debug_settings: Default::default(),
+                buffer_device_address: false,
+                allocation_sizes: Default::default(),
+            })
+            .expect("Failed to create Vulkan allocator");
+
+            Self {
+                inner: Arc::new(Mutex::new(VulkanInner {
+                    allocator,
+                    allocations: HashMap::new(),
+                })),
+            }
+        }
+    }
+
+    unsafe impl Allocator for VulkanAlloc {
+        fn allocate(&self, layout: Layout) -> Result<NonNull<[u8]>, AllocError> {
+            if layout.size() == 0 {
+                let ptr = NonNull::new(layout.align() as *mut u8).ok_or(AllocError)?;
+                return Ok(NonNull::slice_from_raw_parts(ptr, 0));
+            }
+
+            let mut inner = self.inner.lock();
+
+            let allocation = inner
+                .allocator
+                .allocate(&AllocationCreateDesc {
+                    name: "gvec_alloc",
+                    requirements: vk::MemoryRequirements {
+                        size: layout.size() as u64,
+                        alignment: layout.align() as u64,
+                        memory_type_bits: u32::MAX,
+                    },
+                    location: MemoryLocation::CpuToGpu,
+                    linear: true,
+                    allocation_scheme: AllocationScheme::GpuAllocatorManaged,
+                })
+                .map_err(|_| AllocError)?;
+
+            let ptr = allocation.mapped_ptr().ok_or(AllocError)?.as_ptr() as *mut u8;
+
+            let nn = NonNull::new(ptr).ok_or(AllocError)?;
+            inner.allocations.insert(ptr as usize, allocation);
+            Ok(NonNull::slice_from_raw_parts(nn, layout.size()))
+        }
+
+        unsafe fn deallocate(&self, ptr: NonNull<u8>, layout: Layout) {
+            if layout.size() == 0 {
+                return;
+            }
+            let mut inner = self.inner.lock();
+            if let Some(allocation) = inner.allocations.remove(&(ptr.as_ptr() as usize)) {
+                inner
+                    .allocator
+                    .free(allocation)
+                    .expect("Vulkan free failed");
+            }
+        }
+    }
+}
+
+#[cfg(feature = "cuda")]
+pub type GpuAlloc = cuda::CudaAlloc;
+
+#[cfg(all(feature = "vulkan", not(feature = "cuda")))]
+pub type GpuAlloc = vulkan::VulkanAlloc;
+
+#[cfg(not(any(feature = "cuda", feature = "vulkan")))]
+pub type GpuAlloc = SystemAlloc;
+
+pub type GVec<T> = alloc::vec::Vec<T, GpuAlloc>;
+pub type GBox<T> = alloc::boxed::Box<T, GpuAlloc>;
+
+pub fn gvec<T>() -> GVec<T> {
+    Vec::new_in(GpuAlloc::default())
+}
+
+pub fn gvec_with_capacity<T>(cap: usize) -> GVec<T> {
+    Vec::with_capacity_in(cap, GpuAlloc::default())
+}
+
+pub fn gvec_from_slice<T: Clone>(slice: &[T]) -> GVec<T> {
+    let mut v = gvec_with_capacity(slice.len());
+    v.extend_from_slice(slice);
+    v
+}
+
+pub fn gbox<T>(value: T) -> GBox<T> {
+    Box::new_in(value, GpuAlloc::default())
+}
diff --git a/src/utils/upload.rs b/src/utils/upload.rs
new file mode 100644
index 0000000..40b3555
--- /dev/null
+++ b/src/utils/upload.rs
@@ -0,0 +1,167 @@
+use crate::core::image::Image;
+use crate::spectra::DenselySampledSpectrumBuffer;
+use crate::Arena;
+use shared::core::color::RGBToSpectrumTable;
+use shared::core::image::DeviceImage;
+use shared::core::light::Light;
+use shared::core::material::Material;
+use shared::core::shape::Shape;
+use shared::core::spectrum::Spectrum;
+use shared::spectra::{DenselySampledSpectrum, DeviceStandardColorSpaces, RGBColorSpace};
+use shared::Ptr;
+use std::slice::from_raw_parts;
+
+pub trait DeviceRepr {
+    /// The `#[repr(C)] Copy` device-side struct.
+    type Target: Copy;
+
+    /// Upload into the arena and return the device struct by value.
+    /// Use this when embedding the result inline in another device struct.
+    fn upload_value(&self, arena: &Arena) -> Self::Target;
+
+    /// Upload into the arena and return a Ptr to the device struct.
+    /// This is the common entry point — allocates the Target in the arena.
+    fn upload(&self, arena: &Arena) -> Ptr<Self::Target> {
+        let value = self.upload_value(arena);
+        arena.alloc(value)
+    }
+}
+
+impl<T: DeviceRepr> DeviceRepr for Option<T> {
+    type Target = T::Target;
+
+    fn upload_value(&self, arena: &Arena) -> Self::Target {
+        match self {
+            Some(val) => val.upload_value(arena),
+            None => panic!("Cannot upload_value on None — use upload() which returns Ptr::null()"),
+        }
+    }
+
+    fn upload(&self, arena: &Arena) -> Ptr<Self::Target> {
+        match self {
+            Some(val) => val.upload(arena),
+            None => Ptr::null(),
+        }
+    }
+}
+
+impl<T: DeviceRepr> DeviceRepr for std::sync::Arc<T> {
+    type Target = T::Target;
+
+    fn upload_value(&self, arena: &Arena) -> Self::Target {
+        (**self).upload_value(arena)
+    }
+
+    fn upload(&self, arena: &Arena) -> Ptr<Self::Target> {
+        (**self).upload(arena)
+    }
+}
+
+impl<T: DeviceRepr> DeviceRepr for Box<T> {
+    type Target = T::Target;
+
+    fn upload_value(&self, arena: &Arena) -> Self::Target {
+        (**self).upload_value(arena)
+    }
+
+    fn upload(&self, arena: &Arena) -> Ptr<Self::Target> {
+        (**self).upload(arena)
+    }
+}
+
+impl DeviceRepr for Shape {
+    type Target = Shape;
+    fn upload_value(&self, _arena: &Arena) -> Shape {
+        self.clone()
+    }
+}
+
+impl DeviceRepr for Light {
+    type Target = Light;
+    fn upload_value(&self, _arena: &Arena) -> Light {
+        self.clone()
+    }
+}
+
+impl DeviceRepr for Spectrum {
+    type Target = Spectrum;
+    fn upload_value(&self, _arena: &Arena) -> Spectrum {
+        self.clone()
+    }
+}
+
+impl DeviceRepr for Material {
+    type Target = Material;
+    fn upload_value(&self, _arena: &Arena) -> Material {
+        self.clone()
+    }
+}
+
+impl DeviceRepr for Image {
+    type Target = DeviceImage;
+    fn upload_value(&self, _arena: &Arena) -> DeviceImage {
+        *self.device()
+    }
+}
+
+impl DeviceRepr for DenselySampledSpectrumBuffer {
+    type Target = DenselySampledSpectrum;
+    fn upload_value(&self, _arena: &Arena) -> DenselySampledSpectrum {
+        self.device()
+    }
+}
+
+impl DeviceRepr for RGBToSpectrumTable {
+    type Target = RGBToSpectrumTable;
+
+    fn upload_value(&self, arena: &Arena) -> RGBToSpectrumTable {
+        let n_nodes = self.n_nodes as usize;
+
+        // Safety: these Ptrs point into static or previously-uploaded data;
+        // we're copying the contents into the arena for a new lifetime.
+        let z_slice = unsafe { from_raw_parts(self.z_nodes.as_raw(), n_nodes) };
+        let (z_ptr, _) = arena.alloc_slice(z_slice);
+
+        let n_coeffs = 3 * n_nodes.pow(3);
+        let coeffs_slice = unsafe { from_raw_parts(self.coeffs.as_raw(), n_coeffs) };
+        let (c_ptr, _) = arena.alloc_slice(coeffs_slice);
+
+        RGBToSpectrumTable {
+            z_nodes: z_ptr,
+            coeffs: c_ptr,
+            n_nodes: self.n_nodes,
+        }
+    }
+}
+
+impl DeviceRepr for RGBColorSpace {
+    type Target = RGBColorSpace;
+
+    fn upload_value(&self, arena: &Arena) -> RGBColorSpace {
+        let table_ptr = self.rgb_to_spectrum_table.upload(arena);
+
+        RGBColorSpace {
+            r: self.r,
+            g: self.g,
+            b: self.b,
+            w: self.w,
+            illuminant: self.illuminant.clone(),
+            rgb_to_spectrum_table: table_ptr,
+            xyz_from_rgb: self.xyz_from_rgb,
+            rgb_from_xyz: self.rgb_from_xyz,
+        }
+    }
+}
+
+impl DeviceRepr for DeviceStandardColorSpaces {
+    type Target = DeviceStandardColorSpaces;
+
+    fn upload_value(&self, arena: &Arena) -> DeviceStandardColorSpaces {
+        DeviceStandardColorSpaces {
+            srgb: self.srgb.upload(arena),
+            dci_p3: self.dci_p3.upload(arena),
+            rec2020: self.rec2020.upload(arena),
+            aces2065_1: self.aces2065_1.upload(arena),
+        }
+    }
+}