diff --git a/Cargo.lock b/Cargo.lock
index ec2aebd..eb8a5a7 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -101,6 +101,13 @@ dependencies = [
  "ash",
 ]
 
+[[package]]
+name = "piet-gpu-types"
+version = "0.0.0"
+dependencies = [
+ "piet-gpu-derive",
+]
+
 [[package]]
 name = "png"
 version = "0.16.2"
diff --git a/Cargo.toml b/Cargo.toml
index 3098acc..f71f2de 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -3,5 +3,6 @@
 members = [
     "piet-gpu",
     "piet-gpu-derive",
-    "piet-gpu-hal"
+    "piet-gpu-hal",
+    "piet-gpu-types"
 ]
diff --git a/piet-gpu-derive/src/lib.rs b/piet-gpu-derive/src/lib.rs
index 220def1..43177de 100644
--- a/piet-gpu-derive/src/lib.rs
+++ b/piet-gpu-derive/src/lib.rs
@@ -19,7 +19,7 @@ pub fn piet_gpu(input: TokenStream) -> TokenStream {
     let glsl = glsl::gen_glsl(&layout);
     let gen_gpu_fn = format_ident!("gen_gpu_{}", layout.name);
     let mut expanded = quote! {
-        fn #gen_gpu_fn() -> String {
+        pub fn #gen_gpu_fn() -> String {
             #glsl.into()
         }
     };
diff --git a/piet-gpu-hal/src/lib.rs b/piet-gpu-hal/src/lib.rs
index 7fa133a..54bced7 100644
--- a/piet-gpu-hal/src/lib.rs
+++ b/piet-gpu-hal/src/lib.rs
@@ -75,5 +75,7 @@ pub trait CmdBuf<D: Device> {
 }
 
 pub trait MemFlags: Sized + Clone + Copy {
+    fn device_local() -> Self;
+
     fn host_coherent() -> Self;
 }
diff --git a/piet-gpu-hal/src/vulkan.rs b/piet-gpu-hal/src/vulkan.rs
index f250abc..ab3e724 100644
--- a/piet-gpu-hal/src/vulkan.rs
+++ b/piet-gpu-hal/src/vulkan.rs
@@ -455,6 +455,10 @@ impl crate::CmdBuf<VkDevice> for CmdBuf {
 }
 
 impl crate::MemFlags for MemFlags {
+    fn device_local() -> Self {
+        MemFlags(vk::MemoryPropertyFlags::DEVICE_LOCAL)
+    }
+
     fn host_coherent() -> Self {
         MemFlags(vk::MemoryPropertyFlags::HOST_VISIBLE | vk::MemoryPropertyFlags::HOST_COHERENT)
     }
diff --git a/piet-gpu-types/Cargo.toml b/piet-gpu-types/Cargo.toml
new file mode 100644
index 0000000..6de92a5
--- /dev/null
+++ b/piet-gpu-types/Cargo.toml
@@ -0,0 +1,11 @@
+[package]
+name = "piet-gpu-types"
+version = "0.0.0"
+authors = ["Raph Levien <raph.levien@gmail.com>"]
+description = "The scene graph and internal GPU types for piet-gpu."
+license = "MIT/Apache-2.0"
+edition = "2018"
+keywords = ["graphics", "2d"]
+
+[dependencies]
+piet-gpu-derive = { path = "../piet-gpu-derive" }
diff --git a/piet-gpu-types/src/encoder.rs b/piet-gpu-types/src/encoder.rs
new file mode 100644
index 0000000..7ac8bbb
--- /dev/null
+++ b/piet-gpu-types/src/encoder.rs
@@ -0,0 +1,151 @@
+//  Copyright 2020 The xi-editor authors.
+
+//! New-style encoders (supporting proc macros)
+
+pub struct A;
+
+/// A reference to an encoded object within a buffer
+#[derive(Clone, Copy, Debug)]
+pub struct Ref<T> {
+    offset: u32,
+    _phantom: std::marker::PhantomData<T>,
+}
+
+pub struct Encoder {
+    buf: Vec<u8>,
+}
+
+// TODO: we probably do want to encode slices, get rid of Sized bound
+pub trait Encode: Sized {
+    /// Size if it's a fixed-size object, otherwise 0.
+    fn fixed_size() -> usize;
+
+    /// Encoded size, for both fixed and variable sized objects.
+    fn encoded_size(&self) -> usize {
+        Self::fixed_size()
+    }
+
+    /// Encode into a buffer; panics if not appropriately sized.
+    fn encode_to(&self, buf: &mut [u8]);
+
+    /// Allocate a chunk and encode, returning a reference.
+    fn encode(&self, encoder: &mut Encoder) -> Ref<Self> {
+        let size = self.encoded_size();
+        let (offset, buf) = encoder.alloc_chunk(size as u32);
+        self.encode_to(buf);
+        Ref::new(offset)
+    }
+}
+
+impl<T> Ref<T> {
+    fn new(offset: u32) -> Ref<T> {
+        Ref {
+            offset,
+            _phantom: Default::default(),
+        }
+    }
+
+    pub fn offset(&self) -> u32 {
+        self.offset
+    }
+
+    pub fn transmute<U>(&self) -> Ref<U> {
+        Ref::new(self.offset)
+    }
+}
+
+impl Encoder {
+    pub fn new() -> Encoder {
+        Encoder { buf: Vec::new() }
+    }
+
+    pub fn alloc_chunk(&mut self, size: u32) -> (u32, &mut [u8]) {
+        let offset = self.buf.len();
+        self.buf.resize(size as usize + offset, 0);
+        (offset as u32, &mut self.buf[offset..])
+    }
+
+    pub fn buf(&self) -> &[u8] {
+        &self.buf
+    }
+
+    pub fn buf_mut(&mut self) -> &mut [u8] {
+        &mut self.buf
+    }
+}
+
+impl<T> Encode for Ref<T> {
+    fn fixed_size() -> usize {
+        4
+    }
+
+    fn encode_to(&self, buf: &mut [u8]) {
+        buf[0..4].copy_from_slice(&self.offset.to_le_bytes());
+    }
+}
+
+// Encode impls for scalar and small vector types are as needed; it's a finite set of
+// possibilities, so we could do it all with macros, but by hand is expedient.
+
+impl Encode for u32 {
+    fn fixed_size() -> usize {
+        4
+    }
+
+    fn encode_to(&self, buf: &mut [u8]) {
+        buf[0..4].copy_from_slice(&self.to_le_bytes());
+    }
+}
+
+impl Encode for f32 {
+    fn fixed_size() -> usize {
+        4
+    }
+
+    fn encode_to(&self, buf: &mut [u8]) {
+        buf[0..4].copy_from_slice(&self.to_le_bytes());
+    }
+}
+
+impl Encode for [u16; 4] {
+    fn fixed_size() -> usize {
+        8
+    }
+
+    fn encode_to(&self, buf: &mut [u8]) {
+        buf[0..2].copy_from_slice(&self[0].to_le_bytes());
+        buf[2..4].copy_from_slice(&self[1].to_le_bytes());
+        buf[4..6].copy_from_slice(&self[2].to_le_bytes());
+        buf[6..8].copy_from_slice(&self[3].to_le_bytes());
+    }
+}
+
+impl Encode for [f32; 2] {
+    fn fixed_size() -> usize {
+        8
+    }
+
+    fn encode_to(&self, buf: &mut [u8]) {
+        buf[0..4].copy_from_slice(&self[0].to_le_bytes());
+        buf[4..8].copy_from_slice(&self[1].to_le_bytes());
+    }
+}
+
+// TODO: make this work for slices too, but need to deal with Sized bound
+//
+// Note: only works for vectors of fixed size objects.
+impl<T: Encode> Encode for Vec<T> {
+    fn fixed_size() -> usize {
+        0
+    }
+    fn encoded_size(&self) -> usize {
+        self.len() * T::fixed_size()
+    }
+
+    fn encode_to(&self, buf: &mut [u8]) {
+        let size = T::fixed_size();
+        for (ix, val) in self.iter().enumerate() {
+            val.encode_to(&mut buf[ix * size..]);
+        }
+    }
+}
diff --git a/piet-gpu-types/src/lib.rs b/piet-gpu-types/src/lib.rs
new file mode 100644
index 0000000..60c11ab
--- /dev/null
+++ b/piet-gpu-types/src/lib.rs
@@ -0,0 +1,3 @@
+pub mod encoder;
+pub mod ptcl;
+pub mod scene;
diff --git a/piet-gpu-types/src/main.rs b/piet-gpu-types/src/main.rs
new file mode 100644
index 0000000..00a5d0b
--- /dev/null
+++ b/piet-gpu-types/src/main.rs
@@ -0,0 +1,8 @@
+fn main() {
+    let mod_name = std::env::args().skip(1).next().expect("provide a module name");
+    match mod_name.as_str() {
+        "scene" => print!("{}", piet_gpu_types::scene::gen_gpu_scene()),
+        "ptcl" => print!("{}", piet_gpu_types::ptcl::gen_gpu_ptcl()),
+        _ => println!("Oops, unknown module name"),
+    }
+}
diff --git a/piet-gpu-types/src/ptcl.rs b/piet-gpu-types/src/ptcl.rs
new file mode 100644
index 0000000..f5e42af
--- /dev/null
+++ b/piet-gpu-types/src/ptcl.rs
@@ -0,0 +1,47 @@
+use piet_gpu_derive::piet_gpu;
+
+piet_gpu! {
+    #[gpu_write]
+    mod ptcl {
+        struct CmdCircle {
+            // In existing code, this is packed; we might need an annotation for this.
+            bbox: [u16; 4],
+        }
+        struct CmdLine {
+            start: [f32; 2],
+            end: [f32; 2],
+        }
+        struct CmdStroke {
+            // In existing code, this is f16. Should we have support?
+            halfWidth: f32,
+            rgba_color: u32,
+        }
+        struct CmdFill {
+            start: [f32; 2],
+            end: [f32; 2],
+        }
+        struct CmdFillEdge {
+            // The sign is only one bit.
+            sign: i32,
+            y: f32,
+        }
+        struct CmdDrawFill {
+            backdrop: i32,
+            rgba_color: u32,
+        }
+        struct CmdSolid {
+            rgba_color: u32,
+        }
+        enum Cmd {
+            End,
+            Circle(CmdCircle),
+            Line(CmdLine),
+            Fill(CmdFill),
+            Stroke(CmdStroke),
+            FillEdge(CmdFillEdge),
+            DrawFill(CmdDrawFill),
+            Solid(CmdSolid),
+            Bail,
+        }
+    }
+}
diff --git a/piet-gpu-types/src/scene.rs b/piet-gpu-types/src/scene.rs
new file mode 100644
index 0000000..d2cc466
--- /dev/null
+++ b/piet-gpu-types/src/scene.rs
@@ -0,0 +1,49 @@
+use piet_gpu_derive::piet_gpu;
+
+pub use self::scene::{
+    Bbox, PietFill, PietItem, PietStrokeLine, PietStrokePolyLine, Point, SimpleGroup,
+};
+
+piet_gpu! {
+    #[rust_encode]
+    mod scene {
+        struct Bbox {
+            // TODO: this should be i16
+            bbox: [u16; 4],
+        }
+        struct Point {
+            xy: [f32; 2],
+        }
+        struct SimpleGroup {
+            n_items: u32,
+            // Note: both of the following items are actually arrays
+            items: Ref<PietItem>,
+            bboxes: Ref<Bbox>,
+        }
+        struct PietStrokeLine {
+            flags: u32,
+            rgba_color: u32,
+            width: f32,
+            start: Point,
+            end: Point,
+        }
+        struct PietFill {
+            flags: u32,
+            rgba_color: u32,
+            n_points: u32,
+            points: Ref<Point>,
+        }
+        struct PietStrokePolyLine {
+            rgba_color: u32,
+            width: f32,
+            n_points: u32,
+            points: Ref<Point>,
+        }
+        enum PietItem {
+            Circle(),
+            Line(PietStrokeLine),
+            Fill(PietFill),
+            Poly(PietStrokePolyLine),
+        }
+    }
+}
diff --git a/piet-gpu/shader/image.comp b/piet-gpu/shader/image.comp
index 2c35e74..5487366 100644
--- a/piet-gpu/shader/image.comp
+++ b/piet-gpu/shader/image.comp
@@ -4,6 +4,8 @@
 // plan is to use a texture. This is because of limited support.
 
 #version 450
+#extension GL_GOOGLE_include_directive : enable
+
 layout(local_size_x = 16, local_size_y = 16) in;
 
 layout(set = 0, binding = 0) readonly buffer SceneBuf {
@@ -14,6 +16,8 @@ layout(set = 0, binding = 1) buffer ImageBuf {
     uint[] image;
 };
 
+#include "scene.h"
+
 // TODO: make the image size dynamic.
 #define IMAGE_WIDTH 2048
 #define IMAGE_HEIGHT 1535
diff --git a/piet-gpu/shader/image.spv b/piet-gpu/shader/image.spv
index dd4bbc8..e05c8c0 100644
Binary files a/piet-gpu/shader/image.spv and b/piet-gpu/shader/image.spv differ
diff --git a/piet-gpu/shader/scene.h b/piet-gpu/shader/scene.h
new file mode 100644
index 0000000..eb99bd4
--- /dev/null
+++ b/piet-gpu/shader/scene.h
@@ -0,0 +1,200 @@
+// Code auto-generated by piet-gpu-derive
+
+struct BboxRef {
+    uint offset;
+};
+
+struct PointRef {
+    uint offset;
+};
+
+struct SimpleGroupRef {
+    uint offset;
+};
+
+struct PietStrokeLineRef {
+    uint offset;
+};
+
+struct PietFillRef {
+    uint offset;
+};
+
+struct PietStrokePolyLineRef {
+    uint offset;
+};
+
+struct PietItemRef {
+    uint offset;
+};
+
+struct Bbox {
+    uvec4 bbox;
+};
+
+#define Bbox_size 8
+
+BboxRef Bbox_index(BboxRef ref, uint index) {
+    return BboxRef(ref.offset + index * Bbox_size);
+}
+
+struct Point {
+    vec2 xy;
+};
+
+#define Point_size 8
+
+PointRef Point_index(PointRef ref, uint index) {
+    return PointRef(ref.offset + index * Point_size);
+}
+
+struct SimpleGroup {
+    uint n_items;
+    PietItemRef items;
+    BboxRef bboxes;
+};
+
+#define SimpleGroup_size 12
+
+SimpleGroupRef SimpleGroup_index(SimpleGroupRef ref, uint index) {
+    return SimpleGroupRef(ref.offset + index * SimpleGroup_size);
+}
+
+struct PietStrokeLine {
+    uint flags;
+    uint rgba_color;
+    float width;
+    Point start;
+    Point end;
+};
+
+#define PietStrokeLine_size 28
+
+PietStrokeLineRef PietStrokeLine_index(PietStrokeLineRef ref, uint index) {
+    return PietStrokeLineRef(ref.offset + index * PietStrokeLine_size);
+}
+
+struct PietFill {
+    uint flags;
+    uint rgba_color;
+    uint n_points;
+    PointRef points;
+};
+
+#define PietFill_size 16
+
+PietFillRef PietFill_index(PietFillRef ref, uint index) {
+    return PietFillRef(ref.offset + index * PietFill_size);
+}
+
+struct PietStrokePolyLine {
+    uint rgba_color;
+    float width;
+    uint n_points;
+    PointRef points;
+};
+
+#define PietStrokePolyLine_size 16
+
+PietStrokePolyLineRef PietStrokePolyLine_index(PietStrokePolyLineRef ref, uint index) {
+    return PietStrokePolyLineRef(ref.offset + index * PietStrokePolyLine_size);
+}
+
+#define PietItem_Circle 0
+#define PietItem_Line 1
+#define PietItem_Fill 2
+#define PietItem_Poly 3
+#define PietItem_size 32
+
+PietItemRef PietItem_index(PietItemRef ref, uint index) {
+    return PietItemRef(ref.offset + index * PietItem_size);
+}
+
+Bbox Bbox_read(BboxRef ref) {
+    uint ix = ref.offset >> 2;
+    uint raw0 = scene[ix + 0];
+    uint raw1 = scene[ix + 1];
+    Bbox s;
+    s.bbox = uvec4(raw0 & 0xffff, raw0 >> 16, raw1 & 0xffff, raw1 >> 16);
+    return s;
+}
+
+Point Point_read(PointRef ref) {
+    uint ix = ref.offset >> 2;
+    uint raw0 = scene[ix + 0];
+    uint raw1 = scene[ix + 1];
+    Point s;
+    s.xy = vec2(uintBitsToFloat(raw0), uintBitsToFloat(raw1));
+    return s;
+}
+
+SimpleGroup SimpleGroup_read(SimpleGroupRef ref) {
+    uint ix = ref.offset >> 2;
+    uint raw0 = scene[ix + 0];
+    uint raw1 = scene[ix + 1];
+    uint raw2 = scene[ix + 2];
+    SimpleGroup s;
+    s.n_items = raw0;
+    s.items = PietItemRef(raw1);
+    s.bboxes = BboxRef(raw2);
+    return s;
+}
+
+PietStrokeLine PietStrokeLine_read(PietStrokeLineRef ref) {
+    uint ix = ref.offset >> 2;
+    uint raw0 = scene[ix + 0];
+    uint raw1 = scene[ix + 1];
+    uint raw2 = scene[ix + 2];
+    PietStrokeLine s;
+    s.flags = raw0;
+    s.rgba_color = raw1;
+    s.width = uintBitsToFloat(raw2);
+    s.start = Point_read(PointRef(ref.offset + 12));
+    s.end = Point_read(PointRef(ref.offset + 20));
+    return s;
+}
+
+PietFill PietFill_read(PietFillRef ref) {
+    uint ix = ref.offset >> 2;
+    uint raw0 = scene[ix + 0];
+    uint raw1 = scene[ix + 1];
+    uint raw2 = scene[ix + 2];
+    uint raw3 = scene[ix + 3];
+    PietFill s;
+    s.flags = raw0;
+    s.rgba_color = raw1;
+    s.n_points = raw2;
+    s.points = PointRef(raw3);
+    return s;
+}
+
+PietStrokePolyLine PietStrokePolyLine_read(PietStrokePolyLineRef ref) {
+    uint ix = ref.offset >> 2;
+    uint raw0 = scene[ix + 0];
+    uint raw1 = scene[ix + 1];
+    uint raw2 = scene[ix + 2];
+    uint raw3 = scene[ix + 3];
+    PietStrokePolyLine s;
+    s.rgba_color = raw0;
+    s.width = uintBitsToFloat(raw1);
+    s.n_points = raw2;
+    s.points = PointRef(raw3);
+    return s;
+}
+
+uint PietItem_tag(PietItemRef ref) {
+    return scene[ref.offset >> 2];
+}
+
+PietStrokeLine PietItem_Line_read(PietItemRef ref) {
+    return PietStrokeLine_read(PietStrokeLineRef(ref.offset + 4));
+}
+
+PietFill PietItem_Fill_read(PietItemRef ref) {
+    return PietFill_read(PietFillRef(ref.offset + 4));
+}
+
+PietStrokePolyLine PietItem_Poly_read(PietItemRef ref) {
+    return PietStrokePolyLine_read(PietStrokePolyLineRef(ref.offset + 4));
+}
+