Hitting myself, and going for a trait for uploading objs to GPU. Keep forgetting to actually upload them

shared/src/core/filter.rs

@@ -13,13 +13,13 @@ pub struct FilterSample {
 
 #[repr(C)]
 #[derive(Clone, Debug, Copy)]
-pub struct FilterSampler {
+pub struct DeviceFilterSampler {
     pub domain: Bounds2f,
     pub distrib: DevicePiecewiseConstant2D,
     pub f: DeviceArray2D<Float>,
 }
 
-impl FilterSampler {
+impl DeviceFilterSampler {
     pub fn sample(&self, u: Point2f) -> FilterSample {
         let (p, pdf, pi) = self.distrib.sample(u);
 
@@ -38,7 +38,7 @@ pub trait FilterTrait {
     fn radius(&self) -> Vector2f;
     fn evaluate(&self, p: Point2f) -> Float;
     fn integral(&self) -> Float;
-    fn sample(&self, u: Point2f) -> FilterSample;
+    fn sample(&self, u: Point2f) -> DeviceFilterSample;
 }
 
 #[repr(C)]

shared/src/core/primitive.rs

@@ -205,6 +205,7 @@ impl PrimitiveTrait for KdTreeAggregate {
     }
 }
 
+#[repr(C)]
 #[derive(Clone, Debug, Copy)]
 #[enum_dispatch(PrimitiveTrait)]
 pub enum Primitive {

shared/src/filters/boxf.rs

@@ -1,5 +1,5 @@
 use crate::Float;
-use crate::core::filter::{FilterSample, FilterTrait};
+use crate::core::filter::{DeviceFilterSample, FilterTrait};
 use crate::core::geometry::{Point2f, Vector2f};
 use crate::utils::math::lerp;
 
@@ -31,7 +31,7 @@ impl FilterTrait for BoxFilter {
         (2.0 * self.radius.x()) * (2.0 * self.radius.y())
     }
 
-    fn sample(&self, u: Point2f) -> FilterSample {
+    fn sample(&self, u: Point2f) -> DeviceFilterSample {
         let p = Point2f::new(
             lerp(u[0], -self.radius.x(), self.radius.x()),
             lerp(u[1], -self.radius.y(), self.radius.y()),

shared/src/filters/gaussian.rs

@@ -1,5 +1,5 @@
 use crate::Float;
-use crate::core::filter::{FilterSample, FilterSampler, FilterTrait};
+use crate::core::filter::{DeviceFilterSample, FilterSampler, FilterTrait};
 use crate::core::geometry::{Point2f, Vector2f};
 use crate::utils::math::{gaussian, gaussian_integral};
 
@@ -30,7 +30,7 @@ impl FilterTrait for GaussianFilter {
                 - 2.0 * self.radius.y() * self.exp_y)
     }
 
-    fn sample(&self, u: Point2f) -> FilterSample {
+    fn sample(&self, u: Point2f) -> DeviceFilterSample {
         self.sampler.sample(u)
     }
 }

shared/src/filters/lanczos.rs

@@ -1,5 +1,5 @@
 use crate::Float;
-use crate::core::filter::{FilterSample, FilterSampler, FilterTrait};
+use crate::core::filter::{DeviceFilterSample, FilterSampler, FilterTrait};
 use crate::core::geometry::{Point2f, Vector2f};
 use crate::utils::math::{lerp, windowed_sinc};
 
@@ -26,7 +26,7 @@ impl FilterTrait for LanczosSincFilter {
         self.integral
     }
 
-    fn sample(&self, u: Point2f) -> FilterSample {
+    fn sample(&self, u: Point2f) -> DeviceFilterSample {
         self.sampler.sample(u)
     }
 }

shared/src/filters/mitchell.rs

@@ -1,5 +1,5 @@
 use crate::Float;
-use crate::core::filter::{FilterSample, FilterSampler, FilterTrait};
+use crate::core::filter::{DeviceFilterSample, FilterSampler, FilterTrait};
 use crate::core::geometry::{Point2f, Vector2f};
 use num_traits::Float as NumFloat;
 
@@ -9,7 +9,7 @@ pub struct MitchellFilter {
     pub radius: Vector2f,
     pub b: Float,
     pub c: Float,
-    pub sampler: FilterSampler,
+    pub sampler: DeviceFilterSampler,
 }
 
 impl MitchellFilter {
@@ -50,7 +50,7 @@ impl FilterTrait for MitchellFilter {
         self.radius.x() * self.radius.y() / 4.0
     }
 
-    fn sample(&self, u: Point2f) -> FilterSample {
+    fn sample(&self, u: Point2f) -> DeviceFilterSample {
         self.sampler.sample(u)
     }
 }

shared/src/filters/triangle.rs

@@ -1,5 +1,5 @@
 use crate::Float;
-use crate::core::filter::{FilterSample, FilterTrait};
+use crate::core::filter::{DeviceFilterSample, FilterTrait};
 use crate::core::geometry::{Point2f, Vector2f};
 use crate::utils::math::sample_tent;
 use num_traits::Float as NumFloat;
@@ -29,11 +29,11 @@ impl FilterTrait for TriangleFilter {
         self.radius.x().powi(2) * self.radius.y().powi(2)
     }
 
-    fn sample(&self, u: Point2f) -> FilterSample {
+    fn sample(&self, u: Point2f) -> DeviceFilterSample {
         let p = Point2f::new(
             sample_tent(u[0], self.radius.x()),
             sample_tent(u[1], self.radius.y()),
         );
-        FilterSample { p, weight: 1.0 }
+        DeviceFilterSample { p, weight: 1.0 }
     }
 }

shared/src/utils/mod.rs

@@ -18,6 +18,15 @@ pub use options::PBRTOptions;
 pub use ptr::Ptr;
 pub use transform::{AnimatedTransform, Transform, TransformGeneric};
 
+use proc_macro::TokenStream;
+use proc_macro2::TokenStream as TokenStream2;
+use quote::{format_ident, quote};
+use syn::{
+    parse_macro_input, Attribute, Data, DeriveInput, Expr, Fields, GenericArgument, Ident, Lit,
+    PathArguments, Type, Variant,
+};
+
+
 use crate::Float;
 use core::sync::atomic::{AtomicU32, Ordering};
 
@@ -128,3 +137,515 @@ pub fn gpu_array_from_fn<T, const N: usize>(mut f: impl FnMut(usize) -> T) -> [T
         arr.assume_init()
     }
 }
+
+/// # Enum variant attributes
+///
+/// | Attribute | Effect |
+/// |-----------|--------|
+/// | *(none)* | Inner type has `DeviceRepr`; auto-call `upload_value` |
+/// | `#[device(clone)]` | Same type on both sides, just clone |
+/// | `#[device(custom = "method")]` | You provide `fn method(inner: &T, arena) -> DeviceT` |
+/// | `#[device(variant_type = "T")]` | Override the device-side variant's inner type |
+///
+/// # Container attribute
+///
+/// `#[device(name = "DeviceFoo")]` — override the generated type name (default: `Device{Name}`).
+#[proc_macro_derive(Device, attributes(device))]
+pub fn derive_device(input: TokenStream) -> TokenStream {
+    let input = parse_macro_input!(input as DeriveInput);
+    match derive_impl(input) {
+        Ok(tokens) => tokens.into(),
+        Err(e) => e.to_compile_error().into(),
+    }
+}
+
+fn derive_impl(input: DeriveInput) -> syn::Result<TokenStream2> {
+    match &input.data {
+        Data::Struct(_) => derive_struct(input),
+        Data::Enum(_) => derive_enum(input),
+        Data::Union(_) => Err(syn::Error::new_spanned(
+            &input.ident,
+            "Device derive does not support unions",
+        )),
+    }
+}
+
+// Struct derivation
+
+fn derive_struct(input: DeriveInput) -> syn::Result<TokenStream2> {
+    let host_name = &input.ident;
+    let vis = &input.vis;
+    let device_name = get_device_name(&input.attrs, host_name)?;
+
+    let fields = match &input.data {
+        Data::Struct(s) => match &s.fields {
+            Fields::Named(named) => &named.named,
+            _ => {
+                return Err(syn::Error::new_spanned(
+                    host_name,
+                    "Device derive only supports structs with named fields",
+                ))
+            }
+        },
+        _ => unreachable!(),
+    };
+
+    let mut device_fields = Vec::new();
+    let mut upload_stmts = Vec::new();
+    let mut device_field_inits = Vec::new();
+    let mut spread_expr: Option<Expr> = None;
+
+    for field in fields {
+        let field_name = field.ident.as_ref().unwrap();
+        let attrs = parse_field_attrs(&field.attrs)?;
+
+        if attrs.skip {
+            continue;
+        }
+
+        if let Some(ref expr_str) = attrs.spread {
+            spread_expr = Some(syn::parse_str(expr_str).map_err(|e| {
+                syn::Error::new_spanned(field, format!("invalid device(spread): {}", e))
+            })?);
+            continue;
+        }
+
+        if let Some(expr_str) = &attrs.expr {
+            let expr: Expr = syn::parse_str(expr_str).map_err(|e| {
+                syn::Error::new_spanned(field, format!("invalid device(expr): {}", e))
+            })?;
+            let ty = &field.ty;
+            device_fields.push(quote! { pub #field_name: #ty });
+            upload_stmts.push(quote! { let #field_name = #expr; });
+            device_field_inits.push(quote! { #field_name });
+            continue;
+        }
+
+        match classify_type(&field.ty) {
+            FieldClass::VecCopy(inner_ty) => {
+                let len_name = format_ident!("{}_len", field_name);
+                device_fields.push(quote! { pub #field_name: Ptr<#inner_ty> });
+                device_fields.push(quote! { pub #len_name: usize });
+                upload_stmts.push(quote! {
+                    let (#field_name, #len_name) = arena.alloc_slice(&self.#field_name);
+                });
+                device_field_inits.push(quote! { #field_name });
+                device_field_inits.push(quote! { #len_name });
+            }
+            FieldClass::VecUploadable(inner_ty) => {
+                let len_name = format_ident!("{}_len", field_name);
+                device_fields.push(quote! {
+                    pub #field_name: Ptr<<#inner_ty as DeviceRepr>::Target>
+                });
+                device_fields.push(quote! { pub #len_name: usize });
+                upload_stmts.push(quote! {
+                    let __up: Vec<<#inner_ty as DeviceRepr>::Target> = self.#field_name
+                        .iter()
+                        .map(|item| DeviceRepr::upload_value(item, arena))
+                        .collect();
+                    let (#field_name, #len_name) = arena.alloc_slice(&__up);
+                });
+                device_field_inits.push(quote! { #field_name });
+                device_field_inits.push(quote! { #len_name });
+            }
+            FieldClass::Option(inner_ty) => {
+                device_fields.push(quote! {
+                    pub #field_name: Ptr<<#inner_ty as DeviceRepr>::Target>
+                });
+                upload_stmts.push(quote! {
+                    let #field_name = match &self.#field_name {
+                        Some(val) => DeviceRepr::upload(val, arena),
+                        None => Ptr::null(),
+                    };
+                });
+                device_field_inits.push(quote! { #field_name });
+            }
+            FieldClass::Arc(inner_ty) => {
+                if attrs.flatten {
+                    device_fields.push(quote! {
+                        pub #field_name: <#inner_ty as DeviceRepr>::Target
+                    });
+                    upload_stmts.push(quote! {
+                        let #field_name = DeviceRepr::upload_value(&*self.#field_name, arena);
+                    });
+                } else {
+                    device_fields.push(quote! {
+                        pub #field_name: Ptr<<#inner_ty as DeviceRepr>::Target>
+                    });
+                    upload_stmts.push(quote! {
+                        let #field_name = DeviceRepr::upload(&*self.#field_name, arena);
+                    });
+                }
+                device_field_inits.push(quote! { #field_name });
+            }
+            FieldClass::Plain => {
+                let ty = &field.ty;
+                if attrs.copy_upload {
+                    device_fields.push(quote! { pub #field_name: #ty });
+                    upload_stmts.push(quote! {
+                        let #field_name = self.#field_name.clone();
+                    });
+                } else if attrs.flatten {
+                    device_fields.push(quote! {
+                        pub #field_name: <#ty as DeviceRepr>::Target
+                    });
+                    upload_stmts.push(quote! {
+                        let #field_name = DeviceRepr::upload_value(&self.#field_name, arena);
+                    });
+                } else if attrs.upload {
+                    device_fields.push(quote! {
+                        pub #field_name: Ptr<<#ty as DeviceRepr>::Target>
+                    });
+                    upload_stmts.push(quote! {
+                        let #field_name = DeviceRepr::upload(&self.#field_name, arena);
+                    });
+                } else {
+                    device_fields.push(quote! { pub #field_name: #ty });
+                    upload_stmts.push(quote! {
+                        let #field_name = self.#field_name;
+                    });
+                }
+                device_field_inits.push(quote! { #field_name });
+            }
+        }
+    }
+
+    let constructor = if let Some(spread) = spread_expr {
+        quote! {
+            #device_name {
+                #(#device_field_inits,)*
+                ..#spread
+            }
+        }
+    } else {
+        quote! {
+            #device_name {
+                #(#device_field_inits,)*
+            }
+        }
+    };
+
+    Ok(quote! {
+        #[repr(C)]
+        #[derive(Debug, Copy, Clone)]
+        #vis struct #device_name {
+            #(#device_fields,)*
+        }
+
+        unsafe impl Send for #device_name {}
+        unsafe impl Sync for #device_name {}
+
+        impl DeviceRepr for #host_name {
+            type Target = #device_name;
+
+            fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> Self::Target {
+                #(#upload_stmts)*
+                #constructor
+            }
+        }
+    })
+}
+
+// Enum derivation
+fn derive_enum(input: DeriveInput) -> syn::Result<TokenStream2> {
+    let host_name = &input.ident;
+    let vis = &input.vis;
+    let device_name = get_device_name(&input.attrs, host_name)?;
+
+    let variants = match &input.data {
+        Data::Enum(e) => &e.variants,
+        _ => unreachable!(),
+    };
+
+    let mut device_variants = Vec::new();
+    let mut match_arms = Vec::new();
+
+    for variant in variants {
+        let var_name = &variant.ident;
+        let var_attrs = parse_variant_attrs(&variant.attrs)?;
+        let inner_ty = get_variant_inner_type(variant)?;
+
+        // Determine the device-side inner type for this variant
+        let device_inner: Type = if let Some(ref ty_str) = var_attrs.variant_type {
+            syn::parse_str(ty_str).map_err(|e| {
+                syn::Error::new_spanned(variant, format!("invalid variant_type: {}", e))
+            })?
+        } else if var_attrs.clone_variant {
+            // clone: same type on both sides
+            inner_ty.clone()
+        } else {
+            // auto-upload: use DeviceRepr::Target
+            syn::parse_str(&format!("<{} as DeviceRepr>::Target", quote!(#inner_ty))).map_err(
+                |e| {
+                    syn::Error::new_spanned(variant, format!("cannot construct Target type: {}", e))
+                },
+            )?
+        };
+
+        device_variants.push(quote! { #var_name(#device_inner) });
+
+        if var_attrs.clone_variant {
+            match_arms.push(quote! {
+                #host_name::#var_name(inner) => #device_name::#var_name(inner.clone())
+            });
+        } else if let Some(ref method) = var_attrs.custom {
+            let method_ident = format_ident!("{}", method);
+            match_arms.push(quote! {
+                #host_name::#var_name(inner) => {
+                    #device_name::#var_name(Self::#method_ident(inner, arena))
+                }
+            });
+        } else {
+            // Default: inner implements DeviceRepr
+            match_arms.push(quote! {
+                #host_name::#var_name(inner) => {
+                    #device_name::#var_name(DeviceRepr::upload_value(inner, arena))
+                }
+            });
+        }
+    }
+
+    Ok(quote! {
+        #[repr(C)]
+        #[derive(Debug, Copy, Clone)]
+        #vis enum #device_name {
+            #(#device_variants,)*
+        }
+
+        unsafe impl Send for #device_name {}
+        unsafe impl Sync for #device_name {}
+
+        impl DeviceRepr for #host_name {
+            type Target = #device_name;
+
+            fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> Self::Target {
+                match self {
+                    #(#match_arms,)*
+                }
+            }
+        }
+    })
+}
+
+fn get_variant_inner_type(variant: &Variant) -> syn::Result<Type> {
+    match &variant.fields {
+        Fields::Unnamed(fields) if fields.unnamed.len() == 1 => {
+            Ok(fields.unnamed.first().unwrap().ty.clone())
+        }
+        Fields::Unit => Err(syn::Error::new_spanned(
+            variant,
+            "Device derive: enum variants must have exactly one field, e.g. Variant(Type)",
+        )),
+        _ => Err(syn::Error::new_spanned(
+            variant,
+            "Device derive: only single-field tuple variants supported, e.g. Variant(Type)",
+        )),
+    }
+}
+
+// Attribute parsing for variants
+struct VariantAttrs {
+    clone_variant: bool,
+    custom: Option<String>,
+    variant_type: Option<String>,
+}
+
+fn parse_variant_attrs(attrs: &[Attribute]) -> syn::Result<VariantAttrs> {
+    let mut result = VariantAttrs {
+        clone_variant: false,
+        custom: None,
+        variant_type: None,
+    };
+
+    for attr in attrs {
+        if !attr.path().is_ident("device") {
+            continue;
+        }
+        attr.parse_nested_meta(|meta| {
+            if meta.path.is_ident("clone") {
+                result.clone_variant = true;
+                Ok(())
+            } else if meta.path.is_ident("custom") {
+                let value = meta.value()?;
+                let lit: Lit = value.parse()?;
+                if let Lit::Str(s) = lit {
+                    result.custom = Some(s.value());
+                    Ok(())
+                } else {
+                    Err(meta.error("expected string literal"))
+                }
+            } else if meta.path.is_ident("variant_type") {
+                let value = meta.value()?;
+                let lit: Lit = value.parse()?;
+                if let Lit::Str(s) = lit {
+                    result.variant_type = Some(s.value());
+                    Ok(())
+                } else {
+                    Err(meta.error("expected string literal"))
+                }
+            } else {
+                Err(meta.error("unknown device variant attribute"))
+            }
+        })?;
+    }
+
+    Ok(result)
+}
+
+// Attribute parsing for fields
+struct FieldAttrs {
+    skip: bool,
+    expr: Option<String>,
+    copy_upload: bool,
+    flatten: bool,
+    upload: bool,
+    spread: Option<String>,
+}
+
+fn parse_field_attrs(attrs: &[Attribute]) -> syn::Result<FieldAttrs> {
+    let mut result = FieldAttrs {
+        skip: false,
+        expr: None,
+        copy_upload: false,
+        flatten: false,
+        upload: false,
+        spread: None,
+    };
+
+    for attr in attrs {
+        if !attr.path().is_ident("device") {
+            continue;
+        }
+        attr.parse_nested_meta(|meta| {
+            if meta.path.is_ident("skip") {
+                result.skip = true;
+                Ok(())
+            } else if meta.path.is_ident("expr") {
+                let value = meta.value()?;
+                let lit: Lit = value.parse()?;
+                if let Lit::Str(s) = lit {
+                    result.expr = Some(s.value());
+                    Ok(())
+                } else {
+                    Err(meta.error("expected string literal"))
+                }
+            } else if meta.path.is_ident("copy_upload") {
+                result.copy_upload = true;
+                Ok(())
+            } else if meta.path.is_ident("flatten") {
+                result.flatten = true;
+                Ok(())
+            } else if meta.path.is_ident("upload") {
+                result.upload = true;
+                Ok(())
+            } else if meta.path.is_ident("spread") {
+                let value = meta.value()?;
+                let lit: Lit = value.parse()?;
+                if let Lit::Str(s) = lit {
+                    result.spread = Some(s.value());
+                    Ok(())
+                } else {
+                    Err(meta.error("expected string literal"))
+                }
+            } else {
+                Err(meta.error("unknown device attribute"))
+            }
+        })?;
+    }
+
+    Ok(result)
+}
+
+// Container-level name attribute
+fn get_device_name(attrs: &[Attribute], host_name: &Ident) -> syn::Result<Ident> {
+    for attr in attrs {
+        if !attr.path().is_ident("device") {
+            continue;
+        }
+        let mut name = None;
+        attr.parse_nested_meta(|meta| {
+            if meta.path.is_ident("name") {
+                let value = meta.value()?;
+                let lit: Lit = value.parse()?;
+                if let Lit::Str(s) = lit {
+                    name = Some(format_ident!("{}", s.value()));
+                    Ok(())
+                } else {
+                    Err(meta.error("expected string literal"))
+                }
+            } else {
+                Ok(())
+            }
+        })?;
+        if let Some(n) = name {
+            return Ok(n);
+        }
+    }
+    Ok(format_ident!("Device{}", host_name))
+}
+
+// Type classification
+enum FieldClass {
+    VecCopy(Type),
+    VecUploadable(Type),
+    Option(Type),
+    Arc(Type),
+    Plain,
+}
+
+fn classify_type(ty: &Type) -> FieldClass {
+    if let Some(inner) = extract_generic_arg(ty, "Vec") {
+        if is_copy_primitive(&inner) {
+            FieldClass::VecCopy(inner)
+        } else {
+            FieldClass::VecUploadable(inner)
+        }
+    } else if let Some(inner) = extract_generic_arg(ty, "Option") {
+        FieldClass::Option(inner)
+    } else if let Some(inner) = extract_generic_arg(ty, "Arc") {
+        FieldClass::Arc(inner)
+    } else {
+        FieldClass::Plain
+    }
+}
+
+fn extract_generic_arg(ty: &Type, wrapper: &str) -> Option<Type> {
+    if let Type::Path(type_path) = ty {
+        let seg = type_path.path.segments.last()?;
+        if seg.ident != wrapper {
+            return None;
+        }
+        if let PathArguments::AngleBracketed(args) = &seg.arguments {
+            if let Some(GenericArgument::Type(inner)) = args.args.first() {
+                return Some(inner.clone());
+            }
+        }
+    }
+    None
+}
+
+fn is_copy_primitive(ty: &Type) -> bool {
+    if let Type::Path(type_path) = ty {
+        if let Some(seg) = type_path.path.segments.last() {
+            let name = seg.ident.to_string();
+            return matches!(
+                name.as_str(),
+                "f32"
+                    | "f64"
+                    | "u8"
+                    | "u16"
+                    | "u32"
+                    | "u64"
+                    | "i8"
+                    | "i16"
+                    | "i32"
+                    | "i64"
+                    | "usize"
+                    | "isize"
+                    | "bool"
+                    | "Float"
+            );
+        }
+    }
+    false
+}

shared/src/utils/sampling.rs

@@ -695,8 +695,8 @@ impl VarianceEstimator {
     }
 }
 
-#[derive(Debug, Copy, Clone, Default)]
 #[repr(C)]
+#[derive(Debug, Copy, Clone, Default)]
 pub struct PLSample {
     pub p: Point2f,
     pub pdf: Float,

src/core/aggregates.rs

@@ -1,5 +1,5 @@
 use rayon::prelude::*;
-use shared::core::aggregates::DeviceBVHAggregate;
+use shared::core::aggregates::{{DeviceBVHAggregate, LinearBVHNode};
 use shared::core::geometry::{Bounds3f, Point3f, Ray, Vector3f};
 use shared::core::primitive::{Primitive, PrimitiveTrait};
 use shared::core::shape::ShapeIntersection;
@@ -26,14 +26,6 @@ struct BVHSplitBucket {
     pub bounds: Bounds3f,
 }
 
-#[derive(Debug, Clone, Default)]
-pub struct LinearBVHNode {
-    pub bounds: Bounds3f,
-    pub primitives_offset: usize,
-    pub n_primitives: u16,
-    pub axis: u8,
-    pub pad: u8,
-}
 
 #[derive(Debug, Clone, Copy, Default)]
 struct MortonPrimitive {
@@ -48,7 +40,7 @@ struct LBVHTreelet {
 
 #[derive(Debug, Clone)]
 pub struct BVHPrimitiveInfo {
-    primitive_number: usize, // Index into the original primitives vector
+    primitive_number: usize,
     bounds: Bounds3f,
     centroid: Point3f,
 }

src/core/filter.rs

@@ -1,12 +1,13 @@
 use crate::filters::*;
 use crate::utils::containers::Array2D;
 use crate::utils::sampling::PiecewiseConstant2D;
+use crate::utils::DeviceRepr;
 use crate::utils::{FileLoc, ParameterDictionary};
-use anyhow::{Result, anyhow};
-use shared::Float;
-use shared::core::filter::{Filter, FilterSampler};
+use anyhow::{anyhow, Result};
+use shared::core::filter::{DeviceFilterSampler, Filter};
 use shared::core::geometry::{Bounds2f, Point2f, Vector2f};
 use shared::filters::*;
+use shared::Float;
 
 pub trait FilterFactory {
     fn create(name: &str, params: &ParameterDictionary, loc: &FileLoc) -> Result<Filter>;
@@ -54,14 +55,16 @@ impl FilterFactory for Filter {
     }
 }
 
-pub trait CreateFilterSampler {
-    fn new<F>(radius: Vector2f, func: F) -> Self
-    where
-        F: Fn(Point2f) -> Float;
+#[repr(C)]
+#[derive(Clone, Debug, Copy)]
+pub struct FilterSampler {
+    pub domain: Bounds2f,
+    pub distrib: PiecewiseConstant2D,
+    pub f: Array2D<Float>,
 }
 
-impl CreateFilterSampler for FilterSampler {
-    fn new<F>(radius: Vector2f, func: F) -> Self
+impl FilterSampler {
+    pub fn new<F>(radius: Vector2f, func: F) -> Self
     where
         F: Fn(Point2f) -> Float,
     {
@@ -72,7 +75,6 @@ impl CreateFilterSampler for FilterSampler {
 
         let nx = (32.0 * radius.x()) as i32;
         let ny = (32.0 * radius.y()) as i32;
-
         let mut f = Array2D::new_dims(nx, ny);
         for y in 0..f.y_size() {
             for x in 0..f.x_size() {
@@ -83,11 +85,21 @@ impl CreateFilterSampler for FilterSampler {
                 f[(x as i32, y as i32)] = func(p);
             }
         }
+
         let distrib = PiecewiseConstant2D::new_with_bounds(&f, domain);
-        Self {
-            domain,
-            f: *f.device(),
-            distrib: distrib.device,
+
+        Self { domain, distrib, f }
+    }
+}
+
+impl DeviceRepr for FilterSampler {
+    type Target = DeviceFilterSampler;
+
+    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> DeviceFilterSampler {
+        DeviceFilterSampler {
+            domain: self.domain,
+            distrib: self.distrib.upload_value(arena),
+            f: self.f.upload_value(arena),
         }
     }
 }

src/filters/gaussian.rs

@@ -5,27 +5,26 @@ use shared::core::geometry::{Point2f, Vector2f};
 use shared::filters::GaussianFilter;
 use shared::utils::math::gaussian;
 
-pub trait GaussianFilterCreator {
-    fn new(radius: Vector2f, sigma: Float) -> Self;
+#[derive(Clone, Debug, Device)]
+#[device(name = "GaussianFilter")]
+pub struct GaussianFilterHost {
+    pub radius: Vector2f,
+    pub sigma: Float,
+    pub exp_x: Float,
+    pub exp_y: Float,
+    #[device(flatten)]
+    pub sampler: FilterSampler,
 }
 
-impl GaussianFilterCreator for GaussianFilter {
-    fn new(radius: Vector2f, sigma: Float) -> Self {
-        let exp_x = gaussian(radius.x(), 0., sigma);
-        let exp_y = gaussian(radius.y(), 0., sigma);
-
+impl GaussianFilterHost {
+    pub fn new(radius: Vector2f, sigma: Float) -> Self {
+        let exp_x = gaussian(radius.x(), 0.0, sigma);
+        let exp_y = gaussian(radius.y(), 0.0, sigma);
         let sampler = FilterSampler::new(radius, move |p: Point2f| {
-            let gx = (gaussian(p.x(), 0., sigma) - exp_x).max(0.0);
-            let gy = (gaussian(p.y(), 0., sigma) - exp_y).max(0.0);
+            let gx = (gaussian(p.x(), 0.0, sigma) - exp_x).max(0.0);
+            let gy = (gaussian(p.y(), 0.0, sigma) - exp_y).max(0.0);
             gx * gy
         });
-
-        Self {
-            radius,
-            sigma,
-            exp_x: gaussian(radius.x(), 0., sigma),
-            exp_y: gaussian(radius.y(), 0., sigma),
-            sampler,
-        }
+        Self { radius, sigma, exp_x, exp_y, sampler }
     }
 }

src/filters/lanczos.rs

@@ -6,43 +6,35 @@ use shared::core::geometry::{Point2f, Vector2f};
 use shared::filters::LanczosSincFilter;
 use shared::utils::math::{lerp, windowed_sinc};
 
-pub trait LanczosFilterCreator {
-    fn new(radius: Vector2f, tau: Float) -> Self;
+#[derive(Clone, Debug, Device)]
+#[device(name = "LanczosSincFilter")]
+pub struct LanczosSincFilterHost {
+    pub radius: Vector2f,
+    pub tau: Float,
+    #[device(flatten)]
+    pub sampler: FilterSampler,
 }
 
-impl LanczosFilterCreator for LanczosSincFilter {
-    fn new(radius: Vector2f, tau: Float) -> Self {
-        let evaluate = |p: Point2f| -> Float {
+impl LanczosSincFilterHost {
+    pub fn new(radius: Vector2f, tau: Float) -> Self {
+        let sampler = FilterSampler::new(radius, move |p: Point2f| {
             windowed_sinc(p.x(), radius.x(), tau) * windowed_sinc(p.y(), radius.y(), tau)
-        };
-
-        let sampler = FilterSampler::new(radius, evaluate);
-        let sqrt_samples = 64;
-        let n_samples = sqrt_samples * sqrt_samples;
-        let area = (2.0 * radius.x()) * (2.0 * radius.y());
-        let mut sum = 0.0;
-        let mut rng = rand::rng();
-
-        for y in 0..sqrt_samples {
-            for x in 0..sqrt_samples {
-                let u = Point2f::new(
-                    (x as Float + rng.random::<Float>()) / sqrt_samples as Float,
-                    (y as Float + rng.random::<Float>()) / sqrt_samples as Float,
-                );
-                let p = Point2f::new(
-                    lerp(u.x(), -radius.x(), radius.x()),
-                    lerp(u.y(), -radius.y(), radius.y()),
-                );
-                sum += evaluate(p);
-            }
-        }
-        let integral = sum / n_samples as Float * area;
-
-        Self {
-            radius,
-            tau,
-            sampler,
-            integral,
-        }
+        });
+        Self { radius, tau, sampler }
+    }
+}
+
+fn windowed_sinc(x: Float, radius: Float, tau: Float) -> Float {
+    use std::f32::consts::PI;
+    let x = x.abs();
+    if x > radius {
+        return 0.0;
+    }
+    if x < 1e-5 {
+        1.0
+    } else {
+        let xpi = x * PI;
+        let xpit = xpi * tau;
+        (xpi.sin() / xpi) * (xpit.sin() / xpit)
     }
 }

src/filters/mitchell.rs

@@ -4,23 +4,39 @@ use shared::core::filter::FilterSampler;
 use shared::core::geometry::{Point2f, Vector2f};
 use shared::filters::MitchellFilter;
 
-pub trait MitchellFilterCreator {
-    fn new(radius: Vector2f, b: Float, c: Float) -> Self;
+#[derive(Clone, Debug, Device)]
+#[device(name = "MitchellFilter")]
+pub struct MitchellFilterHost {
+    pub radius: Vector2f,
+    pub b: Float,
+    pub c: Float,
+    #[device(flatten)]
+    pub sampler: FilterSampler,
 }
 
-impl MitchellFilterCreator for MitchellFilter {
-    fn new(radius: Vector2f, b: Float, c: Float) -> Self {
+impl MitchellFilterHost {
+    pub fn new(radius: Vector2f, b: Float, c: Float) -> Self {
         let sampler = FilterSampler::new(radius, move |p: Point2f| {
-            let nx = 2.0 * p.x() / radius.x();
-            let ny = 2.0 * p.y() / radius.y();
-            Self::mitchell_1d_eval(b, c, nx) * Self::mitchell_1d_eval(b, c, ny)
+            mitchell_1d(p.x() / radius.x(), b, c) * mitchell_1d(p.y() / radius.y(), b, c)
         });
-
-        Self {
-            radius,
-            b,
-            c,
-            sampler,
-        }
+        Self { radius, b, c, sampler }
+    }
+}
+
+fn mitchell_1d(x: Float, b: Float, c: Float) -> Float {
+    let x = (2.0 * x).abs();
+    if x <= 1.0 {
+        ((12.0 - 9.0 * b - 6.0 * c) * x * x * x
+            + (-18.0 + 12.0 * b + 6.0 * c) * x * x
+            + (6.0 - 2.0 * b))
+            / 6.0
+    } else if x <= 2.0 {
+        ((-b - 6.0 * c) * x * x * x
+            + (6.0 * b + 30.0 * c) * x * x
+            + (-12.0 * b - 48.0 * c) * x
+            + (8.0 * b + 24.0 * c))
+            / 6.0
+    } else {
+        0.0
     }
 }

src/shapes/mesh.rs

@@ -21,16 +21,18 @@ pub struct TriQuadMesh {
     pub quad_indices: Vec<i32>,
 }
 
-#[derive(Debug)]
-pub(crate) struct TriangleMeshStorage {
-    pub p: Vec<Point3f>,
-    pub n: Vec<Normal3f>,
-    pub s: Vec<Vector3f>,
-    pub uv: Vec<Point2f>,
-    pub vertex_indices: Vec<i32>,
-    pub face_indices: Vec<i32>,
+#[derive(DeviceRepr)]
+#[device(name = "DeviceTriangleMesh")]
+pub struct TriangleMeshStorage {
+    pub vertex_indices: Vec<i32>,    // → Ptr<i32> + len (always present)
+    pub p: Vec<Point3f>,             // → Ptr<Point3f> + len (always present)
+    pub n: Vec<Normal3f>,            // → Ptr<Normal3f> + len (empty → null Ptr, len 0)
+    pub s: Vec<Vector3f>,            // → Ptr<Vector3f> + len
+    pub uv: Vec<Point2f>,            // → Ptr<Point2f> + len
+    pub face_indices: Vec<i32>,      // → Ptr<i32> + len
 }
 
+
 #[derive(Debug)]
 pub(crate) struct BilinearMeshStorage {
     pub vertex_indices: Vec<i32>,

src/utils/arena.rs

@@ -1,147 +1,148 @@
-use crate::core::image::Image;
-use crate::core::texture::{FloatTexture, SpectrumTexture};
-use crate::shapes::{BilinearPatchMesh, TriangleMesh};
-use crate::spectra::DenselySampledSpectrumBuffer;
 use crate::utils::backend::GpuAllocator;
-use crate::utils::mipmap::MIPMap;
-use crate::utils::sampling::{PiecewiseConstant2D, WindowedPiecewiseConstant2D};
-use parking_lot::Mutex;
-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::material::Material;
 use shared::core::spectrum::Spectrum;
-use shared::core::texture::{GPUFloatTexture, GPUSpectrumTexture};
-use shared::spectra::{DenselySampledSpectrum, DeviceStandardColorSpaces, RGBColorSpace};
-use shared::textures::*;
-use shared::utils::mesh::{DeviceBilinearPatchMesh, DeviceTriangleMesh};
-use shared::utils::sampling::{
-    DevicePiecewiseConstant1D, DevicePiecewiseConstant2D, DeviceWindowedPiecewiseConstant2D,
-};
-use shared::utils::Ptr;
+use parking_lot::Mutex;
+use shared::Ptr;
 use std::alloc::Layout;
 use std::collections::HashMap;
 use std::panic::Location;
-use std::slice::from_raw_parts;
 use std::sync::Arc;
 
-pub struct Arena<A: GpuAllocator> {
+struct Chunk {
+    ptr: *mut u8,
+    layout: Layout,
+}
+
+struct GpuBump<A: GpuAllocator> {
     allocator: A,
-    inner: Mutex<ArenaInner>,
+    current: *mut u8,
+    end: *mut u8,
+    chunks: Vec<Chunk>,
 }
 
-struct ArenaInner {
-    blocks: Vec<(*mut u8, Layout)>,
-    current_block: *mut u8,
-    current_offset: usize,
-    current_capacity: usize,
-    current_align: usize,
-    texture_cache: HashMap<usize, u64>,
+const CHUNK_SIZE: usize = 256 * 1024;
+
+impl<A: GpuAllocator> GpuBump<A> {
+    fn new(allocator: A) -> Self {
+        Self {
+            allocator,
+            current: std::ptr::null_mut(),
+            end: std::ptr::null_mut(),
+            chunks: Vec::new(),
+        }
+    }
+
+    fn alloc<T>(&mut self, value: T) -> *mut T {
+        let layout = Layout::new::<T>();
+        let ptr = self.alloc_layout(layout) as *mut T;
+        unsafe { ptr.write(value) };
+        ptr
+    }
+
+    fn alloc_slice<T: Copy>(&mut self, values: &[T]) -> (*mut T, usize) {
+        if values.is_empty() {
+            return (std::ptr::null_mut(), 0);
+        }
+        let layout = Layout::array::<T>(values.len()).unwrap();
+        let ptr = self.alloc_layout(layout) as *mut T;
+        unsafe {
+            std::ptr::copy_nonoverlapping(values.as_ptr(), ptr, values.len());
+        }
+        (ptr, values.len())
+    }
+
+    fn alloc_layout(&mut self, layout: Layout) -> *mut u8 {
+        let size = layout.size();
+        let align = layout.align();
+
+        if size == 0 {
+            return align as *mut u8;
+        }
+
+        // Fast path: bump from current chunk
+        let start = self.current as usize;
+        let aligned = (start + align - 1) & !(align - 1);
+        let end = aligned + size;
+
+        if end <= self.end as usize {
+            self.current = end as *mut u8;
+            return aligned as *mut u8;
+        }
+
+        let chunk_size = if size > CHUNK_SIZE {
+            size.next_multiple_of(align.max(16))
+        } else {
+            CHUNK_SIZE.max(size.next_multiple_of(align.max(16)))
+        };
+
+        let chunk_layout = Layout::from_size_align(chunk_size, align.max(16)).unwrap();
+        let chunk = unsafe { self.allocator.alloc(chunk_layout) };
+
+        let caller = Location::caller();
+        if chunk.is_null() {
+            panic!(
+                "GpuBump OOM {} {}: chunk_size={} align={} backend={}",
+                caller.file(),
+                caller.line(),
+                chunk_size,
+                chunk_layout.align(),
+                std::any::type_name::<A>()
+            );
+        }
+
+        self.chunks.push(Chunk {
+            ptr: chunk,
+            layout: chunk_layout,
+        });
+
+        // If this object alone fills the chunk, mark it consumed and return.
+        if size > CHUNK_SIZE {
+            self.current = unsafe { chunk.add(chunk_size) };
+            self.end = self.current;
+            return chunk;
+        }
+
+        // Set up bump pointers inside the new chunk.
+        self.current = chunk;
+        self.end = unsafe { chunk.add(chunk_size) };
+
+        let aligned = (self.current as usize + align - 1) & !(align - 1);
+        self.current = (aligned + size) as *mut u8;
+        aligned as *mut u8
+    }
 }
 
-const DEFAULT_BLOCK_SIZE: usize = 256 * 1024;
+impl<A: GpuAllocator> Drop for GpuBump<A> {
+    fn drop(&mut self) {
+        for chunk in self.chunks.drain(..) {
+            unsafe { self.allocator.dealloc(chunk.ptr, chunk.layout) };
+        }
+    }
+}
+
+unsafe impl<A: GpuAllocator> Send for GpuBump<A> {}
+unsafe impl<A: GpuAllocator> Sync for GpuBump<A> {}
+
+pub struct Arena<A: GpuAllocator> {
+    bump: Mutex<GpuBump<A>>,
+    texture_cache: Mutex<HashMap<usize, u64>>,
+}
 
 impl<A: GpuAllocator> Arena<A> {
     pub fn new(allocator: A) -> Self {
         Self {
-            allocator,
-            inner: Mutex::new(ArenaInner {
-                blocks: Vec::new(),
-                current_block: std::ptr::null_mut(),
-                current_offset: 0,
-                current_capacity: 0,
-                current_align: 1,
-                texture_cache: HashMap::new(),
-            }),
+            bump: Mutex::new(GpuBump::new(allocator)),
+            texture_cache: Mutex::new(HashMap::new()),
         }
     }
 
     pub fn alloc<T>(&self, value: T) -> Ptr<T> {
-        let layout = Layout::new::<T>();
-        let mut inner = self.inner.lock();
-
-        let aligned = (inner.current_offset + layout.align() - 1) & !(layout.align() - 1);
-
-        // Checking if current block alignment is sufficient
-        if aligned + layout.size() > inner.current_capacity || inner.current_align < layout.align()
-        {
-            let block_size = DEFAULT_BLOCK_SIZE.max(layout.size() * 2);
-            let block_layout = Layout::from_size_align(block_size, layout.align().max(16)).unwrap();
-            let block = unsafe { self.allocator.alloc(block_layout) };
-
-            // null check
-            if block.is_null() {
-                panic!(
-                    "Arena alloc failed at {}:{} — size={} align={}",
-                    caller.file(),
-                    caller.line(),
-                    block_layout.size(),
-                    block_layout.align()
-                );
-            }
-
-            inner.blocks.push((block, block_layout));
-            inner.current_block = block;
-            inner.current_offset = 0;
-            inner.current_capacity = block_size;
-            inner.current_align = block_layout.align(); // NEW
-
-            let aligned = (0 + layout.align() - 1) & !(layout.align() - 1);
-            let ptr = unsafe { inner.current_block.add(aligned) as *mut T };
-            unsafe { ptr.write(value) };
-            inner.current_offset = aligned + layout.size();
-            return Ptr::from_raw(ptr);
-        }
-
-        let ptr = unsafe { inner.current_block.add(aligned) as *mut T };
-        unsafe { ptr.write(value) };
-        inner.current_offset = aligned + layout.size();
+        let mut bump = self.bump.lock();
+        let ptr = bump.alloc(value);
         Ptr::from_raw(ptr)
     }
 
-    pub fn alloc_slice<T: Copy>(&self, values: &[T]) -> (Ptr<T>, usize) {
-        if values.is_empty() {
-            return (Ptr::null(), 0);
-        }
-        let layout = Layout::array::<T>(values.len()).unwrap();
-        let mut inner = self.inner.lock();
-
-        let aligned = (inner.current_offset + layout.align() - 1) & !(layout.align() - 1);
-
-        if aligned + layout.size() > inner.current_capacity || inner.current_align < layout.align()
-        {
-            let block_size = DEFAULT_BLOCK_SIZE.max(layout.size() * 2);
-            let block_layout = Layout::from_size_align(block_size, layout.align().max(16)).unwrap();
-            let block = unsafe { self.allocator.alloc(block_layout) };
-
-            if block.is_null() {
-                panic!(
-                    "Arena allocation failed: size={} align={}",
-                    block_layout.size(),
-                    block_layout.align()
-                );
-            }
-
-            inner.blocks.push((block, block_layout));
-            inner.current_block = block;
-            inner.current_offset = 0;
-            inner.current_capacity = block_size;
-            inner.current_align = block_layout.align(); // NEW
-
-            let aligned = 0;
-            let ptr = unsafe { inner.current_block.add(aligned) as *mut T };
-            unsafe { std::ptr::copy_nonoverlapping(values.as_ptr(), ptr, values.len()) };
-            inner.current_offset = aligned + layout.size();
-            return (Ptr::from_raw(ptr), values.len());
-        }
-
-        let ptr = unsafe { inner.current_block.add(aligned) as *mut T };
-        unsafe { std::ptr::copy_nonoverlapping(values.as_ptr(), ptr, values.len()) };
-        inner.current_offset = aligned + layout.size();
-        (Ptr::from_raw(ptr), values.len())
-    }
-
     pub fn alloc_opt<T>(&self, value: Option<T>) -> Ptr<T> {
         match value {
             Some(v) => self.alloc(v),
@@ -149,20 +150,20 @@ impl<A: GpuAllocator> Arena<A> {
         }
     }
 
+    pub fn alloc_slice<T: Copy>(&self, values: &[T]) -> (Ptr<T>, usize) {
+        let mut bump = self.bump.lock();
+        let (ptr, len) = bump.alloc_slice(values);
+        (Ptr::from_raw(ptr), len)
+    }
+
     pub fn get_texture_object(&self, mipmap: &Arc<MIPMap>) -> u64 {
         let key = Arc::as_ptr(mipmap) as usize;
-        let mut inner = self.inner.lock();
-
-        if let Some(&tex_obj) = inner.texture_cache.get(&key) {
+        let mut cache = self.texture_cache.lock();
+        if let Some(&tex_obj) = cache.get(&key) {
             return tex_obj;
         }
-
-        // TODO: Backend-specific texture object creation.
-        // CUDA: cudaCreateTextureObject
-        // Vulkan: VkImageView + VkSampler -> descriptor index
-        let tex_obj = 0u64;
-
-        inner.texture_cache.insert(key, tex_obj);
+        let tex_obj = 0u64; // TODO: backend-specific creation
+        cache.insert(key, tex_obj);
         tex_obj
     }
 }
@@ -173,76 +174,279 @@ impl<A: GpuAllocator + Default> Default for Arena<A> {
     }
 }
 
-impl<A: GpuAllocator> Drop for Arena<A> {
-    fn drop(&mut self) {
-        let inner = self.inner.get_mut();
-        for (ptr, layout) in inner.blocks.drain(..) {
-            unsafe { self.allocator.dealloc(ptr, layout) };
+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<A: GpuAllocator>(&self, arena: &Arena<A>) -> 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<A: GpuAllocator>(&self, arena: &Arena<A>) -> 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<A: GpuAllocator>(&self, arena: &Arena<A>) -> 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<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target> {
+        match self {
+            Some(val) => val.upload(arena),
+            None => Ptr::null(),
         }
     }
 }
 
-unsafe impl<A: GpuAllocator> Send for Arena<A> {}
-unsafe impl<A: GpuAllocator> Sync for Arena<A> {}
+impl<T: DeviceRepr> DeviceRepr for std::sync::Arc<T> {
+    type Target = T::Target;
 
-pub trait Upload {
-    type Target: Copy;
+    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> Self::Target {
+        (**self).upload_value(arena)
+    }
 
-    fn upload<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target>;
+    fn upload<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target> {
+        (**self).upload(arena)
+    }
 }
 
-impl Upload for Shape {
+impl<T: DeviceRepr> DeviceRepr for Box<T> {
+    type Target = T::Target;
+
+    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> Self::Target {
+        (**self).upload_value(arena)
+    }
+
+    fn upload<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target> {
+        (**self).upload(arena)
+    }
+}
+
+impl DeviceRepr for Shape {
     type Target = Shape;
-    fn upload<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target> {
-        arena.alloc(self.clone())
+    fn upload_value<A: GpuAllocator>(&self, _arena: &Arena<A>) -> Shape {
+        self.clone()
     }
 }
 
-impl Upload for Light {
+impl DeviceRepr for Light {
     type Target = Light;
-    fn upload<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target> {
-        arena.alloc(self.clone())
+    fn upload_value<A: GpuAllocator>(&self, _arena: &Arena<A>) -> Light {
+        self.clone()
     }
 }
 
-impl Upload for Image {
-    type Target = DeviceImage;
-    fn upload<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target> {
-        arena.alloc(*self.device())
-    }
-}
-
-impl Upload for Spectrum {
+impl DeviceRepr for Spectrum {
     type Target = Spectrum;
-    fn upload<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target> {
-        arena.alloc(self.clone())
+    fn upload_value<A: GpuAllocator>(&self, _arena: &Arena<A>) -> Spectrum {
+        self.clone()
     }
 }
 
-impl Upload for Material {
+impl DeviceRepr for Material {
     type Target = Material;
-    fn upload<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target> {
-        arena.alloc(self.clone())
+    fn upload_value<A: GpuAllocator>(&self, _arena: &Arena<A>) -> Material {
+        self.clone()
     }
 }
 
-impl Upload for DenselySampledSpectrumBuffer {
+// =============================================================================
+// Image → DeviceImage
+// =============================================================================
+
+impl DeviceRepr for Image {
+    type Target = DeviceImage;
+    fn upload_value<A: GpuAllocator>(&self, _arena: &Arena<A>) -> DeviceImage {
+        *self.device()
+    }
+}
+
+// =============================================================================
+// DenselySampledSpectrumBuffer → DenselySampledSpectrum
+// =============================================================================
+
+impl DeviceRepr for DenselySampledSpectrumBuffer {
     type Target = DenselySampledSpectrum;
-    fn upload<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target> {
-        arena.alloc(*&self.device())
+    fn upload_value<A: GpuAllocator>(&self, _arena: &Arena<A>) -> DenselySampledSpectrum {
+        self.device()
     }
 }
 
-impl Upload for SpectrumTexture {
+// =============================================================================
+// RGBToSpectrumTable — re-uploads Ptr fields into arena
+// =============================================================================
+
+impl DeviceRepr for RGBToSpectrumTable {
+    type Target = RGBToSpectrumTable;
+
+    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> 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,
+        }
+    }
+}
+
+// =============================================================================
+// RGBColorSpace — nested upload of spectrum table
+// =============================================================================
+
+impl DeviceRepr for RGBColorSpace {
+    type Target = RGBColorSpace;
+
+    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> 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,
+        }
+    }
+}
+
+// =============================================================================
+// DeviceStandardColorSpaces — composition of color space uploads
+// =============================================================================
+
+impl DeviceRepr for DeviceStandardColorSpaces {
+    type Target = DeviceStandardColorSpaces;
+
+    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> 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),
+        }
+    }
+}
+
+// =============================================================================
+// TriangleMesh → DeviceTriangleMesh
+// =============================================================================
+// TriangleMesh should own all its fields directly — no .device field.
+// The host struct holds Vec arrays + scalar metadata. derive(Device) handles it:
+//
+//   #[derive(Device)]
+//   #[device(name = "DeviceTriangleMesh")]
+//   pub struct TriangleMesh {
+//       pub vertex_indices: Vec<i32>,
+//       pub p: Vec<Point3f>,
+//       pub n: Vec<Normal3f>,
+//       pub s: Vec<Vector3f>,
+//       pub uv: Vec<Point2f>,
+//       pub face_indices: Vec<i32>,
+//       pub n_triangles: u32,
+//       pub reverse_orientation: bool,
+//       pub transform_swaps_handedness: bool,
+//   }
+//
+// Until the mesh struct is refactored to remove .device, here's a manual
+// impl that lists every field. This is the MIGRATION TARGET — once TriangleMesh
+// drops its .device field and puts scalars at the top level, switch to derive(Device).
+
+impl DeviceRepr for TriangleMesh {
+    type Target = DeviceTriangleMesh;
+
+    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> DeviceTriangleMesh {
+        let s = &self.storage;
+
+        let (vertex_indices_ptr, _) = arena.alloc_slice(&s.vertex_indices);
+        let (p_ptr, _) = arena.alloc_slice(&s.p);
+        let (n_ptr, _) = arena.alloc_slice(&s.n);
+        let (s_ptr, _) = arena.alloc_slice(&s.s);
+        let (uv_ptr, _) = arena.alloc_slice(&s.uv);
+        let (face_indices_ptr, _) = arena.alloc_slice(&s.face_indices);
+
+        DeviceTriangleMesh {
+            vertex_indices: vertex_indices_ptr,
+            p: p_ptr,
+            n: n_ptr,
+            s: s_ptr,
+            uv: uv_ptr,
+            face_indices: face_indices_ptr,
+            n_triangles: self.n_triangles,
+            reverse_orientation: self.reverse_orientation,
+            transform_swaps_handedness: self.transform_swaps_handedness,
+        }
+    }
+}
+
+// =============================================================================
+// BilinearPatchMesh → DeviceBilinearPatchMesh
+// =============================================================================
+// Same as TriangleMesh: once the host struct is refactored to own scalars
+// directly (no .device field), this switches to derive(Device).
+
+impl DeviceRepr for BilinearPatchMesh {
+    type Target = DeviceBilinearPatchMesh;
+
+    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> DeviceBilinearPatchMesh {
+        let s = &self.storage;
+
+        let (vertex_indices_ptr, _) = arena.alloc_slice(&s.vertex_indices);
+        let (p_ptr, _) = arena.alloc_slice(&s.p);
+        let (n_ptr, _) = arena.alloc_slice(&s.n);
+        let (uv_ptr, _) = arena.alloc_slice(&s.uv);
+
+        let image_dist_ptr = s.image_distribution.upload(arena);
+
+        DeviceBilinearPatchMesh {
+            vertex_indices: vertex_indices_ptr,
+            p: p_ptr,
+            n: n_ptr,
+            uv: uv_ptr,
+            image_distribution: image_dist_ptr,
+            n_patches: self.n_patches,
+            reverse_orientation: self.reverse_orientation,
+            transform_swaps_handedness: self.transform_swaps_handedness,
+        }
+    }
+}
+
+// =============================================================================
+// SpectrumTexture → GPUSpectrumTexture
+// =============================================================================
+
+impl DeviceRepr for SpectrumTexture {
     type Target = GPUSpectrumTexture;
-    fn upload<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target> {
-        let gpu_variant = match self {
+
+    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> GPUSpectrumTexture {
+        match self {
             SpectrumTexture::Constant(tex) => GPUSpectrumTexture::Constant(tex.clone()),
             SpectrumTexture::Checkerboard(tex) => GPUSpectrumTexture::Checkerboard(tex.clone()),
             SpectrumTexture::Dots(tex) => GPUSpectrumTexture::Dots(tex.clone()),
             SpectrumTexture::Image(tex) => {
                 let tex_obj = arena.get_texture_object(&tex.base.mipmap);
-                let gpu_img = GPUSpectrumImageTexture {
+                GPUSpectrumTexture::Image(GPUSpectrumImageTexture {
                     mapping: tex.base.mapping,
                     tex_obj,
                     scale: tex.base.scale,
@@ -255,55 +459,50 @@ impl Upload for SpectrumTexture {
                         .clone()
                         .unwrap_or_else(crate::spectra::default_colorspace),
                     spectrum_type: tex.spectrum_type,
-                };
-                GPUSpectrumTexture::Image(gpu_img)
+                })
             }
             SpectrumTexture::Bilerp(tex) => GPUSpectrumTexture::Bilerp(tex.clone()),
             SpectrumTexture::Scaled(tex) => {
                 let child_ptr = tex.tex.upload(arena);
-
-                let gpu_scaled = GPUSpectrumScaledTexture {
+                let scale_ptr = tex.scale.upload(arena);
+                GPUSpectrumTexture::Scaled(GPUSpectrumScaledTexture {
                     tex: child_ptr,
-                    scale: tex.scale.upload(arena),
-                };
-                GPUSpectrumTexture::Scaled(gpu_scaled)
+                    scale: scale_ptr,
+                })
             }
             SpectrumTexture::Marble(tex) => GPUSpectrumTexture::Marble(tex.clone()),
             SpectrumTexture::Mix(tex) => {
                 let tex1_ptr = tex.tex1.upload(arena);
                 let tex2_ptr = tex.tex2.upload(arena);
                 let amount_ptr = tex.amount.upload(arena);
-
-                let gpu_mix = GPUSpectrumMixTexture {
+                GPUSpectrumTexture::Mix(GPUSpectrumMixTexture {
                     tex1: tex1_ptr,
                     tex2: tex2_ptr,
                     amount: amount_ptr,
-                };
-
-                GPUSpectrumTexture::Mix(gpu_mix)
+                })
             }
             SpectrumTexture::DirectionMix(tex) => {
                 let tex1_ptr = tex.tex1.upload(arena);
                 let tex2_ptr = tex.tex2.upload(arena);
-
-                let gpu_mix = GPUSpectrumDirectionMixTexture {
+                GPUSpectrumTexture::DirectionMix(GPUSpectrumDirectionMixTexture {
                     tex1: tex1_ptr,
                     tex2: tex2_ptr,
                     dir: tex.dir,
-                };
-
-                GPUSpectrumTexture::DirectionMix(gpu_mix)
+                })
             }
-        };
-
-        arena.alloc(gpu_variant)
+        }
     }
 }
 
-impl Upload for FloatTexture {
+// =============================================================================
+// FloatTexture → GPUFloatTexture
+// =============================================================================
+
+impl DeviceRepr for FloatTexture {
     type Target = GPUFloatTexture;
-    fn upload<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target> {
-        let gpu_variant = match self {
+
+    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> GPUFloatTexture {
+        match self {
             FloatTexture::Constant(tex) => GPUFloatTexture::Constant(tex.clone()),
             FloatTexture::Checkerboard(tex) => GPUFloatTexture::Checkerboard(tex.clone()),
             FloatTexture::Dots(tex) => GPUFloatTexture::Dots(tex.clone()),
@@ -312,249 +511,40 @@ impl Upload for FloatTexture {
             FloatTexture::Wrinkled(tex) => GPUFloatTexture::Wrinkled(tex.clone()),
             FloatTexture::Scaled(tex) => {
                 let child_ptr = tex.tex.upload(arena);
-
-                let gpu_scaled = GPUFloatScaledTexture {
+                let scale_ptr = tex.scale.upload(arena);
+                GPUFloatTexture::Scaled(GPUFloatScaledTexture {
                     tex: child_ptr,
-                    scale: tex.scale.upload(arena),
-                };
-                GPUFloatTexture::Scaled(gpu_scaled)
+                    scale: scale_ptr,
+                })
             }
-
             FloatTexture::Mix(tex) => {
                 let tex1_ptr = tex.tex1.upload(arena);
                 let tex2_ptr = tex.tex2.upload(arena);
                 let amount_ptr = tex.amount.upload(arena);
-
-                let gpu_mix = GPUFloatMixTexture {
+                GPUFloatTexture::Mix(GPUFloatMixTexture {
                     tex1: tex1_ptr,
                     tex2: tex2_ptr,
                     amount: amount_ptr,
-                };
-                GPUFloatTexture::Mix(gpu_mix)
+                })
             }
-
             FloatTexture::DirectionMix(tex) => {
                 let tex1_ptr = tex.tex1.upload(arena);
                 let tex2_ptr = tex.tex2.upload(arena);
-                let gpu_dmix = GPUFloatDirectionMixTexture {
+                GPUFloatTexture::DirectionMix(GPUFloatDirectionMixTexture {
                     tex1: tex1_ptr,
                     tex2: tex2_ptr,
                     dir: tex.dir,
-                };
-                GPUFloatTexture::DirectionMix(gpu_dmix)
+                })
             }
-
             FloatTexture::Image(tex) => {
-                let gpu_image_tex = GPUFloatImageTexture {
+                GPUFloatTexture::Image(GPUFloatImageTexture {
                     mapping: tex.base.mapping,
                     tex_obj: tex.base.mipmap.texture_object(),
                     scale: tex.base.scale,
                     invert: tex.base.invert,
-                };
-                GPUFloatTexture::Image(gpu_image_tex)
+                })
             }
-
             FloatTexture::Bilerp(tex) => GPUFloatTexture::Bilerp(tex.clone()),
-        };
-
-        arena.alloc(gpu_variant)
-    }
-}
-
-impl Upload for RGBToSpectrumTable {
-    type Target = RGBToSpectrumTable;
-
-    fn upload<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target> {
-        let n_nodes = self.n_nodes as usize;
-        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 as usize).pow(3);
-        let coeffs_slice = unsafe { from_raw_parts(self.coeffs.as_raw(), n_coeffs) };
-        let (c_ptr, _) = arena.alloc_slice(coeffs_slice);
-
-        arena.alloc(RGBToSpectrumTable {
-            z_nodes: z_ptr,
-            coeffs: c_ptr,
-            n_nodes: self.n_nodes,
-        })
-    }
-}
-
-impl Upload for RGBColorSpace {
-    type Target = RGBColorSpace;
-
-    fn upload<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target> {
-        let table_ptr = self.rgb_to_spectrum_table.upload(arena);
-
-        let shared_space = 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,
-        };
-
-        arena.alloc(shared_space)
-    }
-}
-
-impl Upload for DeviceStandardColorSpaces {
-    type Target = DeviceStandardColorSpaces;
-
-    fn upload<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target> {
-        let srgb_ptr = self.srgb.upload(arena);
-        let dci_ptr = self.dci_p3.upload(arena);
-        let rec_ptr = self.rec2020.upload(arena);
-        let aces_ptr = self.aces2065_1.upload(arena);
-
-        let registry = DeviceStandardColorSpaces {
-            srgb: srgb_ptr,
-            dci_p3: dci_ptr,
-            rec2020: rec_ptr,
-            aces2065_1: aces_ptr,
-        };
-
-        arena.alloc(registry)
-    }
-}
-
-impl Upload for PiecewiseConstant2D {
-    type Target = DevicePiecewiseConstant2D;
-
-    fn upload<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target> {
-        let marginal_shared = self.marginal.to_shared(arena);
-
-        let conditionals_shared: Vec<DevicePiecewiseConstant1D> = self
-            .conditionals
-            .iter()
-            .map(|c| c.to_shared(arena))
-            .collect();
-
-        let (conditionals_ptr, _) = arena.alloc_slice(&conditionals_shared);
-
-        let shared_2d = DevicePiecewiseConstant2D {
-            conditionals: conditionals_ptr,
-            marginal: marginal_shared,
-            ..self.device
-        };
-
-        arena.alloc(shared_2d)
-    }
-}
-
-impl Upload for WindowedPiecewiseConstant2D {
-    type Target = DeviceWindowedPiecewiseConstant2D;
-    fn upload<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target> {
-        let specific = DeviceWindowedPiecewiseConstant2D {
-            sat: self.sat,
-            func: self.func,
-        };
-        arena.alloc(specific)
-    }
-}
-
-impl Upload for TriangleMesh {
-    type Target = DeviceTriangleMesh;
-
-    fn upload<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target> {
-        let storage = &self.storage;
-
-        // Upload all arrays to arena
-        let (vertex_indices_ptr, _) = arena.alloc_slice(&storage.vertex_indices);
-        let (p_ptr, _) = arena.alloc_slice(&storage.p);
-
-        let (n_ptr, _) = if storage.n.is_empty() {
-            (Ptr::null(), 0)
-        } else {
-            arena.alloc_slice(&storage.n)
-        };
-
-        let (s_ptr, _) = if storage.s.is_empty() {
-            (Ptr::null(), 0)
-        } else {
-            arena.alloc_slice(&storage.s)
-        };
-
-        let (uv_ptr, _) = if storage.uv.is_empty() {
-            (Ptr::null(), 0)
-        } else {
-            arena.alloc_slice(&storage.uv)
-        };
-
-        let (face_indices_ptr, _) = if storage.face_indices.is_empty() {
-            (Ptr::null(), 0)
-        } else {
-            arena.alloc_slice(&storage.face_indices)
-        };
-
-        let device = DeviceTriangleMesh {
-            vertex_indices: vertex_indices_ptr,
-            p: p_ptr,
-            n: n_ptr,
-            s: s_ptr,
-            uv: uv_ptr,
-            face_indices: face_indices_ptr,
-            ..self.device
-        };
-
-        arena.alloc(device)
-    }
-}
-
-impl Upload for BilinearPatchMesh {
-    type Target = DeviceBilinearPatchMesh;
-
-    fn upload<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target> {
-        let storage = &self.storage;
-
-        let (vertex_indices_ptr, _) = arena.alloc_slice(&storage.vertex_indices);
-        let (p_ptr, _) = arena.alloc_slice(&storage.p);
-
-        let (n_ptr, _) = if storage.n.is_empty() {
-            (Ptr::null(), 0)
-        } else {
-            arena.alloc_slice(&storage.n)
-        };
-
-        let (uv_ptr, _) = if storage.uv.is_empty() {
-            (Ptr::null(), 0)
-        } else {
-            arena.alloc_slice(&storage.uv)
-        };
-
-        let image_dist_ptr = storage.image_distribution.upload(arena);
-
-        let device = DeviceBilinearPatchMesh {
-            vertex_indices: vertex_indices_ptr,
-            p: p_ptr,
-            n: n_ptr,
-            uv: uv_ptr,
-            image_distribution: image_dist_ptr,
-            ..self.device
-        };
-
-        arena.alloc(device)
-    }
-}
-
-impl<T: Upload> Upload for Option<T> {
-    type Target = T::Target;
-    fn upload<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target> {
-        match self {
-            Some(val) => val.upload(arena),
-            None => Ptr::null(),
         }
     }
 }
-
-impl<T: Upload> Upload for Arc<T> {
-    type Target = T::Target;
-
-    fn upload<A: GpuAllocator>(&self, arena: &Arena<A>) -> Ptr<Self::Target> {
-        (**self).upload(arena)
-    }
-}

src/utils/backend.rs

@@ -1,18 +1,17 @@
 use std::alloc::Layout;
 
-pub trait GpuAllocator: Send + Sync {
-    /// Allocate `size` bytes with given alignment.
-    /// Returns a host-mapped pointer.
+pub trait GpuAllocator: Send + Sync + Clone {
     unsafe fn alloc(&self, layout: Layout) -> *mut u8;
     unsafe fn dealloc(&self, ptr: *mut u8, layout: Layout);
 }
 
-/// CPU fallback — standard system allocator.
+/// CPU fallback
+#[derive(Clone)]
 pub struct SystemAllocator;
 
 impl Default for SystemAllocator {
     fn default() -> Self {
-        Self {}
+        Self
     }
 }
 
@@ -21,101 +20,85 @@ impl GpuAllocator for SystemAllocator {
         if layout.size() == 0 {
             return layout.align() as *mut u8;
         }
-        unsafe { std::alloc::alloc(layout) }
+        std::alloc::alloc(layout)
     }
 
     unsafe fn dealloc(&self, ptr: *mut u8, layout: Layout) {
         if layout.size() > 0 {
-            unsafe {
-                std::alloc::dealloc(ptr, layout);
-            }
+            std::alloc::dealloc(ptr, layout);
         }
     }
 }
 
-/// CUDA unified memory backend using CudaAllocator
+/// CUDA unified memory. Still using CudaAllocator, might move over to
 #[cfg(feature = "cuda")]
 pub mod cuda {
     use super::GpuAllocator;
+    use cust::memory::{cuda_free_unified, cuda_malloc_unified, UnifiedPointer};
     use std::alloc::Layout;
 
+    #[derive(Clone)]
     pub struct CudaAllocator;
 
     impl Default for CudaAllocator {
         fn default() -> Self {
-            Self {}
+            Self
         }
     }
 
     impl GpuAllocator for CudaAllocator {
         unsafe fn alloc(&self, layout: Layout) -> *mut u8 {
-            use cust::memory::cuda_malloc_unified;
-            use cust_raw::driver_sys::*;
-
-            let size = layout.size().max(layout.align());
+            let size = layout.size();
             if size == 0 {
                 return layout.align() as *mut u8;
             }
 
-            let mut ctx: CUcontext = std::ptr::null_mut();
-            cuCtxGetCurrent(&mut ctx);
-            if ctx.is_null() {
-                let mut primary: CUcontext = std::ptr::null_mut();
-                cuDevicePrimaryCtxRetain(&mut primary, 0);
-                cuCtxSetCurrent(primary);
-            }
+            let ptr = cuda_malloc_unified::<u8>(size)
+                .expect("cuda_malloc_unified failed — is a CUDA context current?");
 
-            let mut unified_ptr =
-                unsafe { cuda_malloc_unified::<u8>(size).expect("cuda_malloc_unified failed") };
-            let raw = unified_ptr.as_raw_mut();
-            std::mem::forget(unified_ptr);
+            let raw = ptr.as_raw_mut();
+            std::mem::forget(ptr); // Leak RAII wrapper; Arena owns the raw pointer
             raw
         }
 
         unsafe fn dealloc(&self, ptr: *mut u8, layout: Layout) {
-            use cust::memory::{UnifiedPointer, cuda_free_unified};
-            if layout.size() > 0 {
-                let _ = unsafe { cuda_free_unified(UnifiedPointer::wrap(ptr)) };
+            if layout.size() > 0 && !ptr.is_null() {
+                let _ = cuda_free_unified(UnifiedPointer::wrap(ptr));
             }
         }
     }
 }
 
-/// Vulkan backend (gpu-allocator for now, there might be a better solution)
 #[cfg(feature = "vulkan")]
 pub mod vulkan {
     use super::GpuAllocator;
     use ash::vk;
-    use gpu_allocator::MemoryLocation;
     use gpu_allocator::vulkan::{
         Allocation, AllocationCreateDesc, AllocationScheme, Allocator, AllocatorCreateDesc,
     };
+    use gpu_allocator::MemoryLocation;
     use parking_lot::Mutex;
     use std::alloc::Layout;
     use std::collections::HashMap;
-    use std::sync::OnceLock;
+    use std::sync::Arc;
 
-    // So, having a static allocator seems like a terrible idea
-    // But I cant find a way to get a functioning generic Arena constructor
-    // That might not even be a necessity, since rust-gpu/rust-cuda might actually handle that
-    // differently
-    static VK_ALLOCATOR: OnceLock<VulkanAllocatorInner> = OnceLock::new();
-
-    struct VulkanAllocatorInner {
-        state: Mutex<VulkanState>,
+    #[derive(Clone)]
+    pub struct VulkanAllocator {
+        inner: Arc<Mutex<VulkanInner>>,
     }
 
-    struct VulkanState {
+    struct VulkanInner {
+        device: ash::Device,
         allocator: Allocator,
         allocations: HashMap<usize, Allocation>,
     }
 
-    pub fn init_vulkan(
-        instance: &ash::Instance,
-        device: &ash::Device,
-        physical_device: vk::PhysicalDevice,
-    ) {
-        VK_ALLOCATOR.get_or_init(|| {
+    impl VulkanAllocator {
+        pub fn new(
+            instance: &ash::Instance,
+            device: ash::Device,
+            physical_device: vk::PhysicalDevice,
+        ) -> Self {
             let allocator = Allocator::new(&AllocatorCreateDesc {
                 instance: instance.clone(),
                 device: device.clone(),
@@ -126,52 +109,29 @@ pub mod vulkan {
             })
             .expect("Failed to create Vulkan allocator");
 
-            VulkanAllocatorInner {
-                state: Mutex::new(VulkanState {
+            Self {
+                inner: Arc::new(Mutex::new(VulkanInner {
+                    device,
                     allocator,
                     allocations: HashMap::new(),
-                }),
+                })),
             }
-        });
-    }
-
-    fn inner() -> &'static VulkanAllocatorInner {
-        VK_ALLOCATOR
-            .get()
-            .expect("Vulkan not initialized — call init_vulkan() before arena creation")
-    }
-
-    impl Default for VulkanAllocator {
-        fn default() -> Self {
-            let _ = inner();
-            Self
         }
     }
 
-    pub struct VulkanAllocator;
-
-    // impl VulkanAllocator {
-    //     pub fn new(allocator: Allocator) -> Self {
-    //         Self {
-    //             allocator: Mutex::new(allocator),
-    //             allocations: Mutex::new(HashMap::new()),
-    //         }
-    //     }
-    // }
-
     impl GpuAllocator for VulkanAllocator {
         unsafe fn alloc(&self, layout: Layout) -> *mut u8 {
-            let size = layout.size().max(layout.align());
+            let size = layout.size();
             if size == 0 {
                 return layout.align() as *mut u8;
             }
 
-            let inner = inner();
-            let mut state = inner.state.lock();
-            let allocation = state
+            let mut inner = self.inner.lock();
+
+            let allocation = inner
                 .allocator
                 .allocate(&AllocationCreateDesc {
-                    name: "arena",
+                    name: "arena_chunk",
                     requirements: vk::MemoryRequirements {
                         size: size as u64,
                         alignment: layout.align() as u64,
@@ -188,18 +148,17 @@ pub mod vulkan {
                 .expect("Vulkan allocation not host-mapped")
                 .as_ptr() as *mut u8;
 
-            state.allocations.insert(ptr as usize, allocation);
+            inner.allocations.insert(ptr as usize, allocation);
             ptr
         }
 
         unsafe fn dealloc(&self, ptr: *mut u8, layout: Layout) {
-            if layout.size() == 0 {
+            if layout.size() == 0 || ptr.is_null() {
                 return;
             }
-            let inner = inner();
-            let mut state = inner.state.lock();
-            if let Some(allocation) = state.allocations.remove(&(ptr as usize)) {
-                state
+            let mut inner = self.inner.lock();
+            if let Some(allocation) = inner.allocations.remove(&(ptr as usize)) {
+                inner
                     .allocator
                     .free(allocation)
                     .expect("Vulkan free failed");

src/utils/containers.rs

@@ -34,6 +34,7 @@ where
 }
 
 #[derive(Debug, Clone)]
+#[derive(DeviceRepr)]
 pub struct Array2D<T> {
     pub device: DeviceArray2D<T>,
     pub values: Vec<T>,

src/utils/mod.rs

@@ -12,7 +12,7 @@ pub mod parser;
 pub mod sampling;
 pub mod strings;
 
-pub use arena::Upload;
+pub use arena::DeviceRepr;
 pub use error::FileLoc;
 pub use file::{read_float_file, resolve_filename};
 pub use parameters::{
@@ -28,3 +28,516 @@ pub type Arena = arena::Arena<backend::cuda::CudaAllocator>;
 
 #[cfg(not(any(feature = "cuda", feature = "vulkan")))]
 pub type Arena = arena::Arena<backend::SystemAllocator>;
+
+/// # Enum variant attributes
+///
+/// | Attribute | Effect |
+/// |-----------|--------|
+/// | *(none)* | Inner type has `DeviceRepr`; auto-call `upload_value` |
+/// | `#[device(clone)]` | Same type on both sides, just clone |
+/// | `#[device(custom = "method")]` | You provide `fn method(inner: &T, arena) -> DeviceT` |
+/// | `#[device(variant_type = "T")]` | Override the device-side variant's inner type |
+///
+/// # Container attribute
+///
+/// `#[device(name = "DeviceFoo")]` — override the generated type name (default: `Device{Name}`).
+#[proc_macro_derive(Device, attributes(device))]
+pub fn derive_device(input: TokenStream) -> TokenStream {
+    let input = parse_macro_input!(input as DeriveInput);
+    match derive_impl(input) {
+        Ok(tokens) => tokens.into(),
+        Err(e) => e.to_compile_error().into(),
+    }
+}
+
+fn derive_impl(input: DeriveInput) -> syn::Result<TokenStream2> {
+    match &input.data {
+        Data::Struct(_) => derive_struct(input),
+        Data::Enum(_) => derive_enum(input),
+        Data::Union(_) => Err(syn::Error::new_spanned(
+            &input.ident,
+            "Device derive does not support unions",
+        )),
+    }
+}
+
+// Struct derivation
+
+fn derive_struct(input: DeriveInput) -> syn::Result<TokenStream2> {
+    let host_name = &input.ident;
+    let vis = &input.vis;
+    let device_name = get_device_name(&input.attrs, host_name)?;
+
+    let fields = match &input.data {
+        Data::Struct(s) => match &s.fields {
+            Fields::Named(named) => &named.named,
+            _ => {
+                return Err(syn::Error::new_spanned(
+                    host_name,
+                    "Device derive only supports structs with named fields",
+                ))
+            }
+        },
+        _ => unreachable!(),
+    };
+
+    let mut device_fields = Vec::new();
+    let mut upload_stmts = Vec::new();
+    let mut device_field_inits = Vec::new();
+    let mut spread_expr: Option<Expr> = None;
+
+    for field in fields {
+        let field_name = field.ident.as_ref().unwrap();
+        let attrs = parse_field_attrs(&field.attrs)?;
+
+        if attrs.skip {
+            continue;
+        }
+
+        if let Some(ref expr_str) = attrs.spread {
+            spread_expr = Some(syn::parse_str(expr_str).map_err(|e| {
+                syn::Error::new_spanned(field, format!("invalid device(spread): {}", e))
+            })?);
+            continue;
+        }
+
+        if let Some(expr_str) = &attrs.expr {
+            let expr: Expr = syn::parse_str(expr_str).map_err(|e| {
+                syn::Error::new_spanned(field, format!("invalid device(expr): {}", e))
+            })?;
+            let ty = &field.ty;
+            device_fields.push(quote! { pub #field_name: #ty });
+            upload_stmts.push(quote! { let #field_name = #expr; });
+            device_field_inits.push(quote! { #field_name });
+            continue;
+        }
+
+        match classify_type(&field.ty) {
+            FieldClass::VecCopy(inner_ty) => {
+                let len_name = format_ident!("{}_len", field_name);
+                device_fields.push(quote! { pub #field_name: Ptr<#inner_ty> });
+                device_fields.push(quote! { pub #len_name: usize });
+                upload_stmts.push(quote! {
+                    let (#field_name, #len_name) = arena.alloc_slice(&self.#field_name);
+                });
+                device_field_inits.push(quote! { #field_name });
+                device_field_inits.push(quote! { #len_name });
+            }
+            FieldClass::VecUploadable(inner_ty) => {
+                let len_name = format_ident!("{}_len", field_name);
+                device_fields.push(quote! {
+                    pub #field_name: Ptr<<#inner_ty as DeviceRepr>::Target>
+                });
+                device_fields.push(quote! { pub #len_name: usize });
+                upload_stmts.push(quote! {
+                    let __up: Vec<<#inner_ty as DeviceRepr>::Target> = self.#field_name
+                        .iter()
+                        .map(|item| DeviceRepr::upload_value(item, arena))
+                        .collect();
+                    let (#field_name, #len_name) = arena.alloc_slice(&__up);
+                });
+                device_field_inits.push(quote! { #field_name });
+                device_field_inits.push(quote! { #len_name });
+            }
+            FieldClass::Option(inner_ty) => {
+                device_fields.push(quote! {
+                    pub #field_name: Ptr<<#inner_ty as DeviceRepr>::Target>
+                });
+                upload_stmts.push(quote! {
+                    let #field_name = match &self.#field_name {
+                        Some(val) => DeviceRepr::upload(val, arena),
+                        None => Ptr::null(),
+                    };
+                });
+                device_field_inits.push(quote! { #field_name });
+            }
+            FieldClass::Arc(inner_ty) => {
+                if attrs.flatten {
+                    device_fields.push(quote! {
+                        pub #field_name: <#inner_ty as DeviceRepr>::Target
+                    });
+                    upload_stmts.push(quote! {
+                        let #field_name = DeviceRepr::upload_value(&*self.#field_name, arena);
+                    });
+                } else {
+                    device_fields.push(quote! {
+                        pub #field_name: Ptr<<#inner_ty as DeviceRepr>::Target>
+                    });
+                    upload_stmts.push(quote! {
+                        let #field_name = DeviceRepr::upload(&*self.#field_name, arena);
+                    });
+                }
+                device_field_inits.push(quote! { #field_name });
+            }
+            FieldClass::Plain => {
+                let ty = &field.ty;
+                if attrs.copy_upload {
+                    device_fields.push(quote! { pub #field_name: #ty });
+                    upload_stmts.push(quote! {
+                        let #field_name = self.#field_name.clone();
+                    });
+                } else if attrs.flatten {
+                    device_fields.push(quote! {
+                        pub #field_name: <#ty as DeviceRepr>::Target
+                    });
+                    upload_stmts.push(quote! {
+                        let #field_name = DeviceRepr::upload_value(&self.#field_name, arena);
+                    });
+                } else if attrs.upload {
+                    device_fields.push(quote! {
+                        pub #field_name: Ptr<<#ty as DeviceRepr>::Target>
+                    });
+                    upload_stmts.push(quote! {
+                        let #field_name = DeviceRepr::upload(&self.#field_name, arena);
+                    });
+                } else {
+                    device_fields.push(quote! { pub #field_name: #ty });
+                    upload_stmts.push(quote! {
+                        let #field_name = self.#field_name;
+                    });
+                }
+                device_field_inits.push(quote! { #field_name });
+            }
+        }
+    }
+
+    let constructor = if let Some(spread) = spread_expr {
+        quote! {
+            #device_name {
+                #(#device_field_inits,)*
+                ..#spread
+            }
+        }
+    } else {
+        quote! {
+            #device_name {
+                #(#device_field_inits,)*
+            }
+        }
+    };
+
+    Ok(quote! {
+        #[repr(C)]
+        #[derive(Debug, Copy, Clone)]
+        #vis struct #device_name {
+            #(#device_fields,)*
+        }
+
+        unsafe impl Send for #device_name {}
+        unsafe impl Sync for #device_name {}
+
+        impl DeviceRepr for #host_name {
+            type Target = #device_name;
+
+            fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> Self::Target {
+                #(#upload_stmts)*
+                #constructor
+            }
+        }
+    })
+}
+
+// Enum derivation
+fn derive_enum(input: DeriveInput) -> syn::Result<TokenStream2> {
+    let host_name = &input.ident;
+    let vis = &input.vis;
+    let device_name = get_device_name(&input.attrs, host_name)?;
+
+    let variants = match &input.data {
+        Data::Enum(e) => &e.variants,
+        _ => unreachable!(),
+    };
+
+    let mut device_variants = Vec::new();
+    let mut match_arms = Vec::new();
+
+    for variant in variants {
+        let var_name = &variant.ident;
+        let var_attrs = parse_variant_attrs(&variant.attrs)?;
+        let inner_ty = get_variant_inner_type(variant)?;
+
+        // Determine the device-side inner type for this variant
+        let device_inner: Type = if let Some(ref ty_str) = var_attrs.variant_type {
+            syn::parse_str(ty_str).map_err(|e| {
+                syn::Error::new_spanned(variant, format!("invalid variant_type: {}", e))
+            })?
+        } else if var_attrs.clone_variant {
+            // clone: same type on both sides
+            inner_ty.clone()
+        } else {
+            // auto-upload: use DeviceRepr::Target
+            syn::parse_str(&format!("<{} as DeviceRepr>::Target", quote!(#inner_ty))).map_err(
+                |e| {
+                    syn::Error::new_spanned(variant, format!("cannot construct Target type: {}", e))
+                },
+            )?
+        };
+
+        device_variants.push(quote! { #var_name(#device_inner) });
+
+        if var_attrs.clone_variant {
+            match_arms.push(quote! {
+                #host_name::#var_name(inner) => #device_name::#var_name(inner.clone())
+            });
+        } else if let Some(ref method) = var_attrs.custom {
+            let method_ident = format_ident!("{}", method);
+            match_arms.push(quote! {
+                #host_name::#var_name(inner) => {
+                    #device_name::#var_name(Self::#method_ident(inner, arena))
+                }
+            });
+        } else {
+            // Default: inner implements DeviceRepr
+            match_arms.push(quote! {
+                #host_name::#var_name(inner) => {
+                    #device_name::#var_name(DeviceRepr::upload_value(inner, arena))
+                }
+            });
+        }
+    }
+
+    Ok(quote! {
+        #[repr(C)]
+        #[derive(Debug, Copy, Clone)]
+        #vis enum #device_name {
+            #(#device_variants,)*
+        }
+
+        unsafe impl Send for #device_name {}
+        unsafe impl Sync for #device_name {}
+
+        impl DeviceRepr for #host_name {
+            type Target = #device_name;
+
+            fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> Self::Target {
+                match self {
+                    #(#match_arms,)*
+                }
+            }
+        }
+    })
+}
+
+fn get_variant_inner_type(variant: &Variant) -> syn::Result<Type> {
+    match &variant.fields {
+        Fields::Unnamed(fields) if fields.unnamed.len() == 1 => {
+            Ok(fields.unnamed.first().unwrap().ty.clone())
+        }
+        Fields::Unit => Err(syn::Error::new_spanned(
+            variant,
+            "Device derive: enum variants must have exactly one field, e.g. Variant(Type)",
+        )),
+        _ => Err(syn::Error::new_spanned(
+            variant,
+            "Device derive: only single-field tuple variants supported, e.g. Variant(Type)",
+        )),
+    }
+}
+
+// Attribute parsing for variants
+struct VariantAttrs {
+    clone_variant: bool,
+    custom: Option<String>,
+    variant_type: Option<String>,
+}
+
+fn parse_variant_attrs(attrs: &[Attribute]) -> syn::Result<VariantAttrs> {
+    let mut result = VariantAttrs {
+        clone_variant: false,
+        custom: None,
+        variant_type: None,
+    };
+
+    for attr in attrs {
+        if !attr.path().is_ident("device") {
+            continue;
+        }
+        attr.parse_nested_meta(|meta| {
+            if meta.path.is_ident("clone") {
+                result.clone_variant = true;
+                Ok(())
+            } else if meta.path.is_ident("custom") {
+                let value = meta.value()?;
+                let lit: Lit = value.parse()?;
+                if let Lit::Str(s) = lit {
+                    result.custom = Some(s.value());
+                    Ok(())
+                } else {
+                    Err(meta.error("expected string literal"))
+                }
+            } else if meta.path.is_ident("variant_type") {
+                let value = meta.value()?;
+                let lit: Lit = value.parse()?;
+                if let Lit::Str(s) = lit {
+                    result.variant_type = Some(s.value());
+                    Ok(())
+                } else {
+                    Err(meta.error("expected string literal"))
+                }
+            } else {
+                Err(meta.error("unknown device variant attribute"))
+            }
+        })?;
+    }
+
+    Ok(result)
+}
+
+// Attribute parsing for fields
+struct FieldAttrs {
+    skip: bool,
+    expr: Option<String>,
+    copy_upload: bool,
+    flatten: bool,
+    upload: bool,
+    spread: Option<String>,
+}
+
+fn parse_field_attrs(attrs: &[Attribute]) -> syn::Result<FieldAttrs> {
+    let mut result = FieldAttrs {
+        skip: false,
+        expr: None,
+        copy_upload: false,
+        flatten: false,
+        upload: false,
+        spread: None,
+    };
+
+    for attr in attrs {
+        if !attr.path().is_ident("device") {
+            continue;
+        }
+        attr.parse_nested_meta(|meta| {
+            if meta.path.is_ident("skip") {
+                result.skip = true;
+                Ok(())
+            } else if meta.path.is_ident("expr") {
+                let value = meta.value()?;
+                let lit: Lit = value.parse()?;
+                if let Lit::Str(s) = lit {
+                    result.expr = Some(s.value());
+                    Ok(())
+                } else {
+                    Err(meta.error("expected string literal"))
+                }
+            } else if meta.path.is_ident("copy_upload") {
+                result.copy_upload = true;
+                Ok(())
+            } else if meta.path.is_ident("flatten") {
+                result.flatten = true;
+                Ok(())
+            } else if meta.path.is_ident("upload") {
+                result.upload = true;
+                Ok(())
+            } else if meta.path.is_ident("spread") {
+                let value = meta.value()?;
+                let lit: Lit = value.parse()?;
+                if let Lit::Str(s) = lit {
+                    result.spread = Some(s.value());
+                    Ok(())
+                } else {
+                    Err(meta.error("expected string literal"))
+                }
+            } else {
+                Err(meta.error("unknown device attribute"))
+            }
+        })?;
+    }
+
+    Ok(result)
+}
+
+// Container-level name attribute
+fn get_device_name(attrs: &[Attribute], host_name: &Ident) -> syn::Result<Ident> {
+    for attr in attrs {
+        if !attr.path().is_ident("device") {
+            continue;
+        }
+        let mut name = None;
+        attr.parse_nested_meta(|meta| {
+            if meta.path.is_ident("name") {
+                let value = meta.value()?;
+                let lit: Lit = value.parse()?;
+                if let Lit::Str(s) = lit {
+                    name = Some(format_ident!("{}", s.value()));
+                    Ok(())
+                } else {
+                    Err(meta.error("expected string literal"))
+                }
+            } else {
+                Ok(())
+            }
+        })?;
+        if let Some(n) = name {
+            return Ok(n);
+        }
+    }
+    Ok(format_ident!("Device{}", host_name))
+}
+
+// Type classification
+enum FieldClass {
+    VecCopy(Type),
+    VecUploadable(Type),
+    Option(Type),
+    Arc(Type),
+    Plain,
+}
+
+fn classify_type(ty: &Type) -> FieldClass {
+    if let Some(inner) = extract_generic_arg(ty, "Vec") {
+        if is_copy_primitive(&inner) {
+            FieldClass::VecCopy(inner)
+        } else {
+            FieldClass::VecUploadable(inner)
+        }
+    } else if let Some(inner) = extract_generic_arg(ty, "Option") {
+        FieldClass::Option(inner)
+    } else if let Some(inner) = extract_generic_arg(ty, "Arc") {
+        FieldClass::Arc(inner)
+    } else {
+        FieldClass::Plain
+    }
+}
+
+fn extract_generic_arg(ty: &Type, wrapper: &str) -> Option<Type> {
+    if let Type::Path(type_path) = ty {
+        let seg = type_path.path.segments.last()?;
+        if seg.ident != wrapper {
+            return None;
+        }
+        if let PathArguments::AngleBracketed(args) = &seg.arguments {
+            if let Some(GenericArgument::Type(inner)) = args.args.first() {
+                return Some(inner.clone());
+            }
+        }
+    }
+    None
+}
+
+fn is_copy_primitive(ty: &Type) -> bool {
+    if let Type::Path(type_path) = ty {
+        if let Some(seg) = type_path.path.segments.last() {
+            let name = seg.ident.to_string();
+            return matches!(
+                name.as_str(),
+                "f32"
+                    | "f64"
+                    | "u8"
+                    | "u16"
+                    | "u32"
+                    | "u64"
+                    | "i8"
+                    | "i16"
+                    | "i32"
+                    | "i64"
+                    | "usize"
+                    | "isize"
+                    | "bool"
+                    | "Float"
+            );
+        }
+    }
+    false
+}
+

src/utils/sampling.rs

@@ -13,43 +13,25 @@ use std::sync::Arc;
 
 #[derive(Debug, Clone)]
 pub struct PiecewiseConstant1D {
-    func: Box<[Float]>,
-    cdf: Box<[Float]>,
-    pub device: DevicePiecewiseConstant1D,
+    func: Vec<Float>,
+    cdf: Vec<Float>,
+    pub min: Float,
+    pub max: Float,
 }
 
 impl PiecewiseConstant1D {
-    // Constructors
     pub fn new(f: &[Float]) -> Self {
         Self::new_with_bounds(f.to_vec(), 0.0, 1.0)
     }
 
-    pub fn to_shared<A: GpuAllocator>(&self, arena: &Arena<A>) -> DevicePiecewiseConstant1D {
-        let (func_ptr, _) = arena.alloc_slice(&self.func);
-        let (cdf_ptr, _) = arena.alloc_slice(&self.cdf);
-
-        DevicePiecewiseConstant1D {
-            func: func_ptr,
-            cdf: cdf_ptr,
-            func_integral: self.func_integral,
-            n: self.func.len() as u32,
-            min: self.min,
-            max: self.max,
-        }
-    }
-
     pub fn from_func<F>(f: F, min: Float, max: Float, n: usize) -> Self
     where
         F: Fn(Float) -> Float,
     {
         let delta = (max - min) / n as Float;
         let values: Vec<Float> = (0..n)
-            .map(|i| {
-                let x = min + (i as Float + 0.5) * delta;
-                f(x)
-            })
+            .map(|i| f(min + (i as Float + 0.5) * delta))
             .collect();
-
         Self::new_with_bounds(values, min, max)
     }
 
@@ -64,74 +46,92 @@ impl PiecewiseConstant1D {
         }
 
         let func_integral = cdf[n];
-
         if func_integral > 0.0 {
             for c in &mut cdf {
                 *c /= func_integral;
             }
         }
 
-        // Convert to boxed slices (no more reallocation possible)
-        let func: Box<[Float]> = f.into_boxed_slice();
-        let cdf: Box<[Float]> = cdf.into_boxed_slice();
-
-        let device = DevicePiecewiseConstant1D {
-            func: func.as_ptr().into(),
-            cdf: cdf.as_ptr().into(),
-            min,
-            max,
-            n: n as u32,
-            func_integral,
-        };
-
-        Self { func, cdf, device }
+        Self { func: f, cdf, min, max }
     }
 
     // Accessors
-    pub fn min(&self) -> Float {
-        self.device.min
-    }
-    pub fn max(&self) -> Float {
-        self.device.max
-    }
-    pub fn n(&self) -> usize {
-        self.device.n as usize
-    }
+    pub fn n(&self) -> usize { self.func.len() }
+    pub fn func(&self) -> &[Float] { &self.func }
+    pub fn cdf(&self) -> &[Float] { &self.cdf }
+
     pub fn integral(&self) -> Float {
-        self.device.func_integral
+        // func_integral is the un-normalized sum. After normalization cdf[n] == 1.0,
+        // so we reconstruct from the last CDF entry before normalization.
+        // But since we normalized in-place, we need to store it. Let's compute it.
+        let n = self.func.len();
+        let delta = (self.max - self.min) / n as Float;
+        self.func.iter().sum::<Float>() * delta
     }
 
-    pub fn func(&self) -> &[Float] {
-        &self.func
+    /// Host-side sampling (for scene construction, not rendering).
+    /// During rendering, use the device struct via arena-uploaded Ptrs.
+    pub fn sample_host(&self, u: Float) -> (Float, Float, usize) {
+        let n = self.func.len();
+        let offset = self.find_interval_host(u);
+        let cdf_offset = self.cdf[offset];
+        let cdf_next = self.cdf[offset + 1];
+        let du = if cdf_next - cdf_offset > 0.0 {
+            (u - cdf_offset) / (cdf_next - cdf_offset)
+        } else {
+            0.0
+        };
+        let delta = (self.max - self.min) / n as Float;
+        let x = self.min + (offset as Float + du) * delta;
+        let func_integral = self.integral();
+        let pdf = if func_integral > 0.0 {
+            self.func[offset] / func_integral
+        } else {
+            0.0
+        };
+        (x, pdf, offset)
     }
-    pub fn cdf(&self) -> &[Float] {
-        &self.cdf
+
+    fn find_interval_host(&self, u: Float) -> usize {
+        let n = self.func.len();
+        let mut size = n;
+        let mut first = 0usize;
+        while size > 0 {
+            let half = size >> 1;
+            let middle = first + half;
+            if self.cdf[middle] <= u {
+                first = middle + 1;
+                size -= half + 1;
+            } else {
+                size = half;
+            }
+        }
+        first.saturating_sub(1).min(n - 1)
     }
 }
 
-impl std::ops::Deref for PiecewiseConstant1D {
-    type Target = DevicePiecewiseConstant1D;
-
-    fn deref(&self) -> &Self::Target {
-        &self.device
-    }
+#[derive(DeviceRepr)]
+#[device(name = "DevicePiecewiseConstant1D")]
+pub struct PiecewiseConstant1D {
+    pub func: Vec<Float>,
+    pub cdf: Vec<Float>,
+    pub min: Float,
+    pub max: Float,
+    pub n: u32,
+    #[device(expr = "self.integral()")]
+    pub func_integral: Float,
 }
 
-#[derive(Debug, Clone)]
+#[derive(Clone, Debug, DeviceRepr)]
+#[device(name = "DevicePiecewiseConstant2D")]
 pub struct PiecewiseConstant2D {
     pub conditionals: Vec<PiecewiseConstant1D>,
+    #[device(flatten)]
     pub marginal: PiecewiseConstant1D,
-    pub conditional_devices: Box<[DevicePiecewiseConstant1D]>,
-    pub device: DevicePiecewiseConstant2D,
+    pub n_u: u32,
+    pub n_v: u32,
 }
 
-impl std::ops::Deref for PiecewiseConstant2D {
-    type Target = DevicePiecewiseConstant2D;
-
-    fn deref(&self) -> &Self::Target {
-        &self.device
-    }
-}
 
 impl PiecewiseConstant2D {
     pub fn new(data: &Array2D<Float>) -> Self {
@@ -141,8 +141,8 @@ impl PiecewiseConstant2D {
     pub fn new_with_bounds(data: &Array2D<Float>, domain: Bounds2f) -> Self {
         Self::from_slice(
             data.as_slice(),
-            data.x_size() as usize,
-            data.y_size() as usize,
+            data.x_size(),
+            data.y_size(),
             domain,
         )
     }
@@ -154,36 +154,23 @@ impl PiecewiseConstant2D {
         let mut marginal_func = Vec::with_capacity(n_v);
 
         for v in 0..n_v {
-            let row_start = v * n_u;
-            let row: Vec<Float> = data[row_start..row_start + n_u].to_vec();
-            let conditional =
-                PiecewiseConstant1D::new_with_bounds(row, domain.p_min.x(), domain.p_max.x());
+            let row = data[v * n_u..(v + 1) * n_u].to_vec();
+            let conditional = PiecewiseConstant1D::new_with_bounds(
+                row,
+                domain.p_min.x(),
+                domain.p_max.x(),
+            );
             marginal_func.push(conditional.integral());
             conditionals.push(conditional);
         }
 
-        let marginal =
-            PiecewiseConstant1D::new_with_bounds(marginal_func, domain.p_min.y(), domain.p_max.y());
+        let marginal = PiecewiseConstant1D::new_with_bounds(
+            marginal_func,
+            domain.p_min.y(),
+            domain.p_max.y(),
+        );
 
-        let conditional_devices: Box<[DevicePiecewiseConstant1D]> = conditionals
-            .iter()
-            .map(|c| c.device)
-            .collect::<Vec<_>>()
-            .into_boxed_slice();
-
-        let device = DevicePiecewiseConstant2D {
-            conditionals: conditional_devices.as_ptr().into(),
-            marginal: marginal.device,
-            n_u: n_u as u32,
-            n_v: n_v as u32,
-        };
-
-        Self {
-            conditionals,
-            marginal,
-            conditional_devices,
-            device,
-        }
+        Self { conditionals, marginal, n_u, n_v }
     }
 
     pub fn from_image(image: &Image) -> Self {
@@ -192,12 +179,9 @@ impl PiecewiseConstant2D {
         let n_v = res.y() as usize;
 
         let mut data = Vec::with_capacity(n_u * n_v);
-
         for v in 0..n_v {
             for u in 0..n_u {
-                let p = Point2i::new(u as i32, v as i32);
-                let luminance = image.get_channels(p).average();
-                data.push(luminance);
+                data.push(image.get_channels(Point2i::new(u as i32, v as i32)).average());
             }
         }
 
@@ -217,10 +201,14 @@ struct PiecewiseLinear2DStorage<const N: usize> {
 }
 
 pub struct PiecewiseLinear2DHost<const N: usize> {
-    pub view: PiecewiseLinear2D<N>,
-    _storage: Arc<PiecewiseLinear2DStorage<N>>,
+    size: Vector2i,
+    inv_patch_size: Vector2f,
+    param_size: [u32; N],
+    param_strides: [u32; N],
+    storage: Arc<PiecewiseLinear2DStorage<N>>,
 }
 
+
 impl<const N: usize> PiecewiseLinear2DHost<N> {
     pub fn new(
         data: &[Float],
@@ -354,27 +342,49 @@ impl<const N: usize> PiecewiseLinear2DHost<N> {
     }
 }
 
+impl<const N: usize> DeviceRepr for PiecewiseLinear2DHost<N> {
+    type Target = PiecewiseLinear2D<N>;
+
+    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> PiecewiseLinear2D<N> {
+        let s = &self.storage;
+
+        let (data_ptr, _) = arena.alloc_slice(&s.data);
+        let (marginal_ptr, _) = arena.alloc_slice(&s.marginal_cdf);
+        let (conditional_ptr, _) = arena.alloc_slice(&s.conditional_cdf);
+
+        let param_ptrs: [Ptr<Float>; N] = std::array::from_fn(|i| {
+            let (ptr, _) = arena.alloc_slice(&s.param_values[i]);
+            ptr
+        });
+
+        PiecewiseLinear2D {
+            size: self.size,
+            inv_patch_size: self.inv_patch_size,
+            param_size: self.param_size,
+            param_strides: self.param_strides,
+            param_values: param_ptrs,
+            data: data_ptr,
+            marginal_cdf: marginal_ptr,
+            conditional_cdf: conditional_ptr,
+        }
+    }
+}
+
 #[derive(Debug, Clone)]
 pub struct AliasTableHost {
-    pub view: AliasTable,
-    _storage: Vec<Bin>,
+    bins: Vec<Bin>,
 }
 
 impl AliasTableHost {
     pub fn new(weights: &[Float]) -> Self {
         let n = weights.len();
         if n == 0 {
-            return Self {
-                view: AliasTable {
-                    bins: Ptr::null(),
-                    size: 0,
-                },
-                _storage: Vec::new(),
-            };
+            return Self { bins: Vec::new() };
         }
 
         let sum: f64 = weights.iter().map(|&w| w as f64).sum();
         assert!(sum > 0.0, "Sum of weights must be positive");
+
         let mut bins = Vec::with_capacity(n);
         for &w in weights {
             bins.push(Bin {
@@ -384,10 +394,7 @@ impl AliasTableHost {
             });
         }
 
-        struct Outcome {
-            p_hat: f64,
-            index: usize,
-        }
+        struct Outcome { p_hat: f64, index: usize }
 
         let mut under = Vec::with_capacity(n);
         let mut over = Vec::with_capacity(n);
@@ -409,17 +416,10 @@ impl AliasTableHost {
             bins[un.index].alias = ov.index as u32;
 
             let p_excess = un.p_hat + ov.p_hat - 1.0;
-
             if p_excess < 1.0 {
-                under.push(Outcome {
-                    p_hat: p_excess,
-                    index: ov.index,
-                });
+                under.push(Outcome { p_hat: p_excess, index: ov.index });
             } else {
-                over.push(Outcome {
-                    p_hat: p_excess,
-                    index: ov.index,
-                });
+                over.push(Outcome { p_hat: p_excess, index: ov.index });
             }
         }
 
@@ -427,51 +427,40 @@ impl AliasTableHost {
             bins[ov.index].q = 1.0;
             bins[ov.index].alias = ov.index as u32;
         }
-
         while let Some(un) = under.pop() {
             bins[un.index].q = 1.0;
             bins[un.index].alias = un.index as u32;
         }
 
-        let view = AliasTable {
-            bins: bins.as_ptr().into(),
-            size: bins.len() as u32,
-        };
-
-        Self {
-            view,
-            _storage: bins,
-        }
+        Self { bins }
     }
 
-    pub fn to_device<A: GpuAllocator>(&self, arena: &Arena<A>) -> AliasTable {
-        if self._storage.is_empty() {
-            return AliasTable {
-                bins: Ptr::null(),
-                size: 0,
-            };
+    pub fn size(&self) -> usize { self.bins.len() }
+    pub fn is_empty(&self) -> bool { self.bins.is_empty() }
+}
+
+impl DeviceRepr for AliasTableHost {
+    type Target = AliasTable;
+
+    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> AliasTable {
+        if self.bins.is_empty() {
+            return AliasTable { bins: Ptr::null(), size: 0 };
         }
-        let (bins_ptr, _) = arena.alloc_slice(&self._storage);
+        let (bins_ptr, _) = arena.alloc_slice(&self.bins);
         AliasTable {
             bins: bins_ptr,
-            size: self._storage.len() as u32,
+            size: self.bins.len() as u32,
         }
     }
 }
 
-#[derive(Clone, Debug)]
+#[derive(Clone, Debug, DeviceRepr)]
+#[device(name = "DeviceSummedAreaTable")]
 pub struct SummedAreaTable {
-    pub device: DeviceSummedAreaTable,
+    #[device(flatten)]
     sum: Array2D<f64>,
 }
 
-impl std::ops::Deref for SummedAreaTable {
-    type Target = DeviceSummedAreaTable;
-    fn deref(&self) -> &Self::Target {
-        &self.device
-    }
-}
-
 impl SummedAreaTable {
     pub fn new(values: &Array2D<Float>) -> Self {
         let width = values.x_size() as i32;
@@ -483,46 +472,194 @@ impl SummedAreaTable {
         for x in 1..width {
             sum[(x, 0)] = values[(x, 0)] as f64 + sum[(x - 1, 0)];
         }
-
         for y in 1..height {
             sum[(0, y)] = values[(0, y)] as f64 + sum[(0, y - 1)];
         }
-
         for y in 1..height {
             for x in 1..width {
-                let term = values[(x, y)] as f64;
-                let left = sum[(x - 1, y)];
-                let up = sum[(x, y - 1)];
-                let diag = sum[(x - 1, y - 1)];
-
-                sum[(x, y)] = term + left + up - diag;
+                sum[(x, y)] = values[(x, y)] as f64
+                    + sum[(x - 1, y)]
+                    + sum[(x, y - 1)]
+                    - sum[(x - 1, y - 1)];
             }
         }
 
-        let device = DeviceSummedAreaTable { sum: *sum };
-        Self { device, sum }
+        Self { sum }
     }
 }
 
-#[derive(Clone, Debug)]
+#[derive(Clone, Debug, DeviceRepr)]
+#[device(name = "DeviceWindowedPiecewiseConstant2D")]
 pub struct WindowedPiecewiseConstant2D {
-    pub device: DeviceWindowedPiecewiseConstant2D,
-    sat: DeviceSummedAreaTable,
+    #[device(flatten)]
+    sat: SummedAreaTable,
+    #[device(flatten)]
     func: Array2D<Float>,
 }
 
-impl std::ops::Deref for WindowedPiecewiseConstant2D {
-    type Target = DeviceWindowedPiecewiseConstant2D;
-    fn deref(&self) -> &Self::Target {
-        &self.device
-    }
-}
-
 impl WindowedPiecewiseConstant2D {
     pub fn new(func: Array2D<Float>) -> Self {
-        let sat = *SummedAreaTable::new(&func);
-        let device = DeviceWindowedPiecewiseConstant2D { sat, func: *func };
-
-        Self { sat, func, device }
+        let sat = SummedAreaTable::new(&func);
+        Self { sat, func }
+    }
+}
+
+struct PiecewiseLinear2DStorage<const N: usize> {
+    data: Vec<Float>,
+    marginal_cdf: Vec<Float>,
+    conditional_cdf: Vec<Float>,
+    param_values: [Vec<Float>; N],
+}
+
+pub struct PiecewiseLinear2DHost<const N: usize> {
+    size: Vector2i,
+    inv_patch_size: Vector2f,
+    param_size: [u32; N],
+    param_strides: [u32; N],
+    storage: Arc<PiecewiseLinear2DStorage<N>>,
+}
+
+impl<const N: usize> PiecewiseLinear2DHost<N> {
+    pub fn new(
+        data: &[Float],
+        x_size: i32,
+        y_size: i32,
+        param_res: [usize; N],
+        param_values: [&[Float]; N],
+        normalize: bool,
+        build_cdf: bool,
+    ) -> Self {
+        if build_cdf && !normalize {
+            panic!("PiecewiseLinear2D: build_cdf implies normalize=true");
+        }
+
+        let size = Vector2i::new(x_size, y_size);
+        let inv_patch_size = Vector2f::new(
+            1.0 / (x_size - 1) as Float,
+            1.0 / (y_size - 1) as Float,
+        );
+
+        let mut param_size = [0u32; N];
+        let mut param_strides = [0u32; N];
+        let owned_param_values: [Vec<Float>; N] = gpu_array_from_fn(|i| param_values[i].to_vec());
+
+        let mut slices: u32 = 1;
+        for i in (0..N).rev() {
+            assert!(param_res[i] >= 1, "Parameter resolution must be >= 1");
+            param_size[i] = param_res[i] as u32;
+            param_strides[i] = if param_res[i] > 1 { slices } else { 0 };
+            slices *= param_size[i];
+        }
+
+        let n_values = (x_size * y_size) as usize;
+        let mut new_data = vec![0.0; slices as usize * n_values];
+        let mut marginal_cdf = if build_cdf {
+            vec![0.0; slices as usize * y_size as usize]
+        } else {
+            Vec::new()
+        };
+        let mut conditional_cdf = if build_cdf {
+            vec![0.0; slices as usize * n_values]
+        } else {
+            Vec::new()
+        };
+
+        let mut data_offset = 0;
+        for slice in 0..slices as usize {
+            let slice_offset = slice * n_values;
+            let current_data = &data[data_offset..data_offset + n_values];
+            let mut sum = 0.0_f64;
+
+            if normalize {
+                for y in 0..(y_size - 1) {
+                    for x in 0..(x_size - 1) {
+                        let i = (y * x_size + x) as usize;
+                        let v00 = current_data[i] as f64;
+                        let v10 = current_data[i + 1] as f64;
+                        let v01 = current_data[i + x_size as usize] as f64;
+                        let v11 = current_data[i + 1 + x_size as usize] as f64;
+                        sum += 0.25 * (v00 + v10 + v01 + v11);
+                    }
+                }
+            }
+
+            let normalization = if normalize && sum > 0.0 {
+                1.0 / sum as Float
+            } else {
+                1.0
+            };
+            for k in 0..n_values {
+                new_data[slice_offset + k] = current_data[k] * normalization;
+            }
+
+            if build_cdf {
+                let marginal_slice_offset = slice * y_size as usize;
+                for y in 0..y_size as usize {
+                    let mut cdf_sum = 0.0;
+                    let i_base = y * x_size as usize;
+                    conditional_cdf[slice_offset + i_base] = 0.0;
+                    for x in 0..(x_size - 1) as usize {
+                        let i = i_base + x;
+                        cdf_sum += 0.5
+                            * (new_data[slice_offset + i] + new_data[slice_offset + i + 1]);
+                        conditional_cdf[slice_offset + i + 1] = cdf_sum;
+                    }
+                }
+                marginal_cdf[marginal_slice_offset] = 0.0;
+                let mut marginal_sum = 0.0;
+                for y in 0..(y_size - 1) as usize {
+                    let cdf1 =
+                        conditional_cdf[slice_offset + (y + 1) * x_size as usize - 1];
+                    let cdf2 =
+                        conditional_cdf[slice_offset + (y + 2) * x_size as usize - 1];
+                    marginal_sum += 0.5 * (cdf1 + cdf2);
+                    marginal_cdf[marginal_slice_offset + y + 1] = marginal_sum;
+                }
+            }
+            data_offset += n_values;
+        }
+
+        let storage = Arc::new(PiecewiseLinear2DStorage {
+            data: new_data,
+            marginal_cdf,
+            conditional_cdf,
+            param_values: owned_param_values,
+        });
+
+        Self {
+            size,
+            inv_patch_size,
+            param_size,
+            param_strides,
+            storage,
+        }
+    }
+}
+
+impl<const N: usize> DeviceRepr for PiecewiseLinear2DHost<N> {
+    type Target = PiecewiseLinear2D<N>;
+
+    fn upload_value<A: GpuAllocator>(&self, arena: &Arena<A>) -> PiecewiseLinear2D<N> {
+        let s = &self.storage;
+
+        let (data_ptr, _) = arena.alloc_slice(&s.data);
+        let (marginal_ptr, _) = arena.alloc_slice(&s.marginal_cdf);
+        let (conditional_ptr, _) = arena.alloc_slice(&s.conditional_cdf);
+
+        let param_ptrs: [Ptr<Float>; N] = std::array::from_fn(|i| {
+            let (ptr, _) = arena.alloc_slice(&s.param_values[i]);
+            ptr
+        });
+
+        PiecewiseLinear2D {
+            size: self.size,
+            inv_patch_size: self.inv_patch_size,
+            param_size: self.param_size,
+            param_strides: self.param_strides,
+            param_values: param_ptrs,
+            data: data_ptr,
+            marginal_cdf: marginal_ptr,
+            conditional_cdf: conditional_ptr,
+        }
     }
 }