diff --git a/piet-gpu-types/src/annotated.rs b/piet-gpu-types/src/annotated.rs
new file mode 100644
index 0000000..247ab12
--- /dev/null
+++ b/piet-gpu-types/src/annotated.rs
@@ -0,0 +1,46 @@
+use piet_gpu_derive::piet_gpu;
+
+piet_gpu! {
+    #[gpu_write]
+    mod annotated {
+        struct AnnoLineSeg {
+            p0: [f32; 2],
+            p1: [f32; 2],
+            // halfwidth in both x and y for binning
+            stroke: [f32; 2],
+        }
+        struct AnnoQuadSeg {
+            p0: [f32; 2],
+            p1: [f32; 2],
+            p2: [f32; 2],
+            stroke: [f32; 2],
+        }
+        struct AnnoCubicSeg {
+            p0: [f32; 2],
+            p1: [f32; 2],
+            p2: [f32; 2],
+            p3: [f32; 2],
+            stroke: [f32; 2],
+        }
+        struct AnnoFill {
+            rgba_color: u32,
+            bbox: [f32; 4],
+        }
+        struct AnnoStroke {
+            rgba_color: u32,
+            bbox: [f32; 4],
+            // For the nonuniform scale case, this needs to be a 2x2 matrix.
+            // That's expected to be uncommon, so we could special-case it.
+            linewidth: f32,
+        }
+        enum Annotated {
+            Nop,
+            // The segments need a flag to indicate fill/stroke
+            Line(AnnoLineSeg),
+            Quad(AnnoQuadSeg),
+            Cubic(AnnoCubicSeg),
+            Stroke(AnnoStroke),
+            Fill(AnnoFill),
+        }
+    }
+}
diff --git a/piet-gpu-types/src/lib.rs b/piet-gpu-types/src/lib.rs
index 288f71c..1759c4d 100644
--- a/piet-gpu-types/src/lib.rs
+++ b/piet-gpu-types/src/lib.rs
@@ -1,3 +1,4 @@
+pub mod annotated;
 pub mod encoder;
 pub mod fill_seg;
 pub mod ptcl;
diff --git a/piet-gpu-types/src/main.rs b/piet-gpu-types/src/main.rs
index 033bec4..68e6487 100644
--- a/piet-gpu-types/src/main.rs
+++ b/piet-gpu-types/src/main.rs
@@ -6,6 +6,7 @@ fn main() {
     match mod_name.as_str() {
         "scene" => print!("{}", piet_gpu_types::scene::gen_gpu_scene()),
         "state" => print!("{}", piet_gpu_types::state::gen_gpu_state()),
+        "annotated" => print!("{}", piet_gpu_types::annotated::gen_gpu_annotated()),
         "tilegroup" => print!("{}", piet_gpu_types::tilegroup::gen_gpu_tilegroup()),
         "segment" => print!("{}", piet_gpu_types::segment::gen_gpu_segment()),
         "fill_seg" => print!("{}", piet_gpu_types::fill_seg::gen_gpu_fill_seg()),
diff --git a/piet-gpu/shader/annotated.h b/piet-gpu/shader/annotated.h
new file mode 100644
index 0000000..a3fc464
--- /dev/null
+++ b/piet-gpu/shader/annotated.h
@@ -0,0 +1,290 @@
+// Code auto-generated by piet-gpu-derive
+
+struct AnnoLineSegRef {
+    uint offset;
+};
+
+struct AnnoQuadSegRef {
+    uint offset;
+};
+
+struct AnnoCubicSegRef {
+    uint offset;
+};
+
+struct AnnoFillRef {
+    uint offset;
+};
+
+struct AnnoStrokeRef {
+    uint offset;
+};
+
+struct AnnotatedRef {
+    uint offset;
+};
+
+struct AnnoLineSeg {
+    vec2 p0;
+    vec2 p1;
+    vec2 stroke;
+};
+
+#define AnnoLineSeg_size 24
+
+AnnoLineSegRef AnnoLineSeg_index(AnnoLineSegRef ref, uint index) {
+    return AnnoLineSegRef(ref.offset + index * AnnoLineSeg_size);
+}
+
+struct AnnoQuadSeg {
+    vec2 p0;
+    vec2 p1;
+    vec2 p2;
+    vec2 stroke;
+};
+
+#define AnnoQuadSeg_size 32
+
+AnnoQuadSegRef AnnoQuadSeg_index(AnnoQuadSegRef ref, uint index) {
+    return AnnoQuadSegRef(ref.offset + index * AnnoQuadSeg_size);
+}
+
+struct AnnoCubicSeg {
+    vec2 p0;
+    vec2 p1;
+    vec2 p2;
+    vec2 p3;
+    vec2 stroke;
+};
+
+#define AnnoCubicSeg_size 40
+
+AnnoCubicSegRef AnnoCubicSeg_index(AnnoCubicSegRef ref, uint index) {
+    return AnnoCubicSegRef(ref.offset + index * AnnoCubicSeg_size);
+}
+
+struct AnnoFill {
+    uint rgba_color;
+    vec4 bbox;
+};
+
+#define AnnoFill_size 20
+
+AnnoFillRef AnnoFill_index(AnnoFillRef ref, uint index) {
+    return AnnoFillRef(ref.offset + index * AnnoFill_size);
+}
+
+struct AnnoStroke {
+    uint rgba_color;
+    vec4 bbox;
+    float linewidth;
+};
+
+#define AnnoStroke_size 24
+
+AnnoStrokeRef AnnoStroke_index(AnnoStrokeRef ref, uint index) {
+    return AnnoStrokeRef(ref.offset + index * AnnoStroke_size);
+}
+
+#define Annotated_Nop 0
+#define Annotated_Line 1
+#define Annotated_Quad 2
+#define Annotated_Cubic 3
+#define Annotated_Stroke 4
+#define Annotated_Fill 5
+#define Annotated_size 44
+
+AnnotatedRef Annotated_index(AnnotatedRef ref, uint index) {
+    return AnnotatedRef(ref.offset + index * Annotated_size);
+}
+
+AnnoLineSeg AnnoLineSeg_read(AnnoLineSegRef ref) {
+    uint ix = ref.offset >> 2;
+    uint raw0 = annotated[ix + 0];
+    uint raw1 = annotated[ix + 1];
+    uint raw2 = annotated[ix + 2];
+    uint raw3 = annotated[ix + 3];
+    uint raw4 = annotated[ix + 4];
+    uint raw5 = annotated[ix + 5];
+    AnnoLineSeg s;
+    s.p0 = vec2(uintBitsToFloat(raw0), uintBitsToFloat(raw1));
+    s.p1 = vec2(uintBitsToFloat(raw2), uintBitsToFloat(raw3));
+    s.stroke = vec2(uintBitsToFloat(raw4), uintBitsToFloat(raw5));
+    return s;
+}
+
+void AnnoLineSeg_write(AnnoLineSegRef ref, AnnoLineSeg s) {
+    uint ix = ref.offset >> 2;
+    annotated[ix + 0] = floatBitsToUint(s.p0.x);
+    annotated[ix + 1] = floatBitsToUint(s.p0.y);
+    annotated[ix + 2] = floatBitsToUint(s.p1.x);
+    annotated[ix + 3] = floatBitsToUint(s.p1.y);
+    annotated[ix + 4] = floatBitsToUint(s.stroke.x);
+    annotated[ix + 5] = floatBitsToUint(s.stroke.y);
+}
+
+AnnoQuadSeg AnnoQuadSeg_read(AnnoQuadSegRef ref) {
+    uint ix = ref.offset >> 2;
+    uint raw0 = annotated[ix + 0];
+    uint raw1 = annotated[ix + 1];
+    uint raw2 = annotated[ix + 2];
+    uint raw3 = annotated[ix + 3];
+    uint raw4 = annotated[ix + 4];
+    uint raw5 = annotated[ix + 5];
+    uint raw6 = annotated[ix + 6];
+    uint raw7 = annotated[ix + 7];
+    AnnoQuadSeg s;
+    s.p0 = vec2(uintBitsToFloat(raw0), uintBitsToFloat(raw1));
+    s.p1 = vec2(uintBitsToFloat(raw2), uintBitsToFloat(raw3));
+    s.p2 = vec2(uintBitsToFloat(raw4), uintBitsToFloat(raw5));
+    s.stroke = vec2(uintBitsToFloat(raw6), uintBitsToFloat(raw7));
+    return s;
+}
+
+void AnnoQuadSeg_write(AnnoQuadSegRef ref, AnnoQuadSeg s) {
+    uint ix = ref.offset >> 2;
+    annotated[ix + 0] = floatBitsToUint(s.p0.x);
+    annotated[ix + 1] = floatBitsToUint(s.p0.y);
+    annotated[ix + 2] = floatBitsToUint(s.p1.x);
+    annotated[ix + 3] = floatBitsToUint(s.p1.y);
+    annotated[ix + 4] = floatBitsToUint(s.p2.x);
+    annotated[ix + 5] = floatBitsToUint(s.p2.y);
+    annotated[ix + 6] = floatBitsToUint(s.stroke.x);
+    annotated[ix + 7] = floatBitsToUint(s.stroke.y);
+}
+
+AnnoCubicSeg AnnoCubicSeg_read(AnnoCubicSegRef ref) {
+    uint ix = ref.offset >> 2;
+    uint raw0 = annotated[ix + 0];
+    uint raw1 = annotated[ix + 1];
+    uint raw2 = annotated[ix + 2];
+    uint raw3 = annotated[ix + 3];
+    uint raw4 = annotated[ix + 4];
+    uint raw5 = annotated[ix + 5];
+    uint raw6 = annotated[ix + 6];
+    uint raw7 = annotated[ix + 7];
+    uint raw8 = annotated[ix + 8];
+    uint raw9 = annotated[ix + 9];
+    AnnoCubicSeg s;
+    s.p0 = vec2(uintBitsToFloat(raw0), uintBitsToFloat(raw1));
+    s.p1 = vec2(uintBitsToFloat(raw2), uintBitsToFloat(raw3));
+    s.p2 = vec2(uintBitsToFloat(raw4), uintBitsToFloat(raw5));
+    s.p3 = vec2(uintBitsToFloat(raw6), uintBitsToFloat(raw7));
+    s.stroke = vec2(uintBitsToFloat(raw8), uintBitsToFloat(raw9));
+    return s;
+}
+
+void AnnoCubicSeg_write(AnnoCubicSegRef ref, AnnoCubicSeg s) {
+    uint ix = ref.offset >> 2;
+    annotated[ix + 0] = floatBitsToUint(s.p0.x);
+    annotated[ix + 1] = floatBitsToUint(s.p0.y);
+    annotated[ix + 2] = floatBitsToUint(s.p1.x);
+    annotated[ix + 3] = floatBitsToUint(s.p1.y);
+    annotated[ix + 4] = floatBitsToUint(s.p2.x);
+    annotated[ix + 5] = floatBitsToUint(s.p2.y);
+    annotated[ix + 6] = floatBitsToUint(s.p3.x);
+    annotated[ix + 7] = floatBitsToUint(s.p3.y);
+    annotated[ix + 8] = floatBitsToUint(s.stroke.x);
+    annotated[ix + 9] = floatBitsToUint(s.stroke.y);
+}
+
+AnnoFill AnnoFill_read(AnnoFillRef ref) {
+    uint ix = ref.offset >> 2;
+    uint raw0 = annotated[ix + 0];
+    uint raw1 = annotated[ix + 1];
+    uint raw2 = annotated[ix + 2];
+    uint raw3 = annotated[ix + 3];
+    uint raw4 = annotated[ix + 4];
+    AnnoFill s;
+    s.rgba_color = raw0;
+    s.bbox = vec4(uintBitsToFloat(raw1), uintBitsToFloat(raw2), uintBitsToFloat(raw3), uintBitsToFloat(raw4));
+    return s;
+}
+
+void AnnoFill_write(AnnoFillRef ref, AnnoFill s) {
+    uint ix = ref.offset >> 2;
+    annotated[ix + 0] = s.rgba_color;
+    annotated[ix + 1] = floatBitsToUint(s.bbox.x);
+    annotated[ix + 2] = floatBitsToUint(s.bbox.y);
+    annotated[ix + 3] = floatBitsToUint(s.bbox.z);
+    annotated[ix + 4] = floatBitsToUint(s.bbox.w);
+}
+
+AnnoStroke AnnoStroke_read(AnnoStrokeRef ref) {
+    uint ix = ref.offset >> 2;
+    uint raw0 = annotated[ix + 0];
+    uint raw1 = annotated[ix + 1];
+    uint raw2 = annotated[ix + 2];
+    uint raw3 = annotated[ix + 3];
+    uint raw4 = annotated[ix + 4];
+    uint raw5 = annotated[ix + 5];
+    AnnoStroke s;
+    s.rgba_color = raw0;
+    s.bbox = vec4(uintBitsToFloat(raw1), uintBitsToFloat(raw2), uintBitsToFloat(raw3), uintBitsToFloat(raw4));
+    s.linewidth = uintBitsToFloat(raw5);
+    return s;
+}
+
+void AnnoStroke_write(AnnoStrokeRef ref, AnnoStroke s) {
+    uint ix = ref.offset >> 2;
+    annotated[ix + 0] = s.rgba_color;
+    annotated[ix + 1] = floatBitsToUint(s.bbox.x);
+    annotated[ix + 2] = floatBitsToUint(s.bbox.y);
+    annotated[ix + 3] = floatBitsToUint(s.bbox.z);
+    annotated[ix + 4] = floatBitsToUint(s.bbox.w);
+    annotated[ix + 5] = floatBitsToUint(s.linewidth);
+}
+
+uint Annotated_tag(AnnotatedRef ref) {
+    return annotated[ref.offset >> 2];
+}
+
+AnnoLineSeg Annotated_Line_read(AnnotatedRef ref) {
+    return AnnoLineSeg_read(AnnoLineSegRef(ref.offset + 4));
+}
+
+AnnoQuadSeg Annotated_Quad_read(AnnotatedRef ref) {
+    return AnnoQuadSeg_read(AnnoQuadSegRef(ref.offset + 4));
+}
+
+AnnoCubicSeg Annotated_Cubic_read(AnnotatedRef ref) {
+    return AnnoCubicSeg_read(AnnoCubicSegRef(ref.offset + 4));
+}
+
+AnnoStroke Annotated_Stroke_read(AnnotatedRef ref) {
+    return AnnoStroke_read(AnnoStrokeRef(ref.offset + 4));
+}
+
+AnnoFill Annotated_Fill_read(AnnotatedRef ref) {
+    return AnnoFill_read(AnnoFillRef(ref.offset + 4));
+}
+
+void Annotated_Nop_write(AnnotatedRef ref) {
+    annotated[ref.offset >> 2] = Annotated_Nop;
+}
+
+void Annotated_Line_write(AnnotatedRef ref, AnnoLineSeg s) {
+    annotated[ref.offset >> 2] = Annotated_Line;
+    AnnoLineSeg_write(AnnoLineSegRef(ref.offset + 4), s);
+}
+
+void Annotated_Quad_write(AnnotatedRef ref, AnnoQuadSeg s) {
+    annotated[ref.offset >> 2] = Annotated_Quad;
+    AnnoQuadSeg_write(AnnoQuadSegRef(ref.offset + 4), s);
+}
+
+void Annotated_Cubic_write(AnnotatedRef ref, AnnoCubicSeg s) {
+    annotated[ref.offset >> 2] = Annotated_Cubic;
+    AnnoCubicSeg_write(AnnoCubicSegRef(ref.offset + 4), s);
+}
+
+void Annotated_Stroke_write(AnnotatedRef ref, AnnoStroke s) {
+    annotated[ref.offset >> 2] = Annotated_Stroke;
+    AnnoStroke_write(AnnoStrokeRef(ref.offset + 4), s);
+}
+
+void Annotated_Fill_write(AnnotatedRef ref, AnnoFill s) {
+    annotated[ref.offset >> 2] = Annotated_Fill;
+    AnnoFill_write(AnnoFillRef(ref.offset + 4), s);
+}
+
diff --git a/piet-gpu/shader/elements.comp b/piet-gpu/shader/elements.comp
index 5cede7c..1061fab 100644
--- a/piet-gpu/shader/elements.comp
+++ b/piet-gpu/shader/elements.comp
@@ -12,13 +12,21 @@ layout(set = 0, binding = 0) readonly buffer SceneBuf {
     uint[] scene;
 };
 
-// This will be used for inter-wprkgroup aggregates
+// This will be used for inter-workgroup aggregates. In the
+// meantime, for development, it has been used to store the
+// scan of the state objects.
 layout(set = 0, binding = 1) buffer StateBuf {
     uint[] state;
 };
 
+// The annotated results are stored here.
+layout(set = 0, binding = 2) buffer AnnotatedBuf {
+    uint[] annotated;
+};
+
 #include "scene.h"
 #include "state.h"
+#include "annotated.h"
 
 #define FLAG_SET_LINEWIDTH 1
 #define FLAG_RESET_BBOX 2
@@ -93,6 +101,12 @@ State map_element(ElementRef ref) {
     return c;
 }
 
+// Get the bounding box of a circle transformed by the matrix into an ellipse.
+vec2 get_linewidth(State st) {
+    // See https://www.iquilezles.org/www/articles/ellipses/ellipses.htm
+    return 0.5 * st.linewidth * vec2(length(st.mat.xz), length(st.mat.yw));
+}
+
 // We should be able to use an array of structs but the NV shader compiler
 // doesn't seem to like it :/
 //shared State sh_state[WG_SIZE];
@@ -165,9 +179,38 @@ void main() {
         row = combine_state(row, other);
     }
     for (uint i = 0; i < N_ROWS; i++) {
-        State this_state = combine_state(row, th_state[i]);
+        State st = combine_state(row, th_state[i]);
         // We write the state now for development purposes, but the
         // actual goal is to write transformed and annotated elements.
-        State_write(StateRef((ix + i) * State_size), this_state);
+        //State_write(StateRef((ix + i) * State_size), st);
+
+        // Here we read again from the original scene. There may be
+        // gains to be had from stashing in shared memory or possibly
+        // registers (though register pressure is an issue).
+        ElementRef this_ref = Element_index(ref, i);
+        AnnotatedRef out_ref = AnnotatedRef((ix + i) * Annotated_size);
+        uint tag = Element_tag(this_ref);
+        switch (tag) {
+        case Element_Line:
+            LineSeg line = Element_Line_read(this_ref);
+            AnnoLineSeg anno_line;
+            anno_line.p0 = st.mat.xz * line.p0.x + st.mat.yw * line.p0.y + st.translate;
+            anno_line.p1 = st.mat.xz * line.p1.x + st.mat.yw * line.p1.y + st.translate;
+            anno_line.stroke = get_linewidth(st);
+            Annotated_Line_write(out_ref, anno_line);
+            break;
+        case Element_Stroke:
+            Stroke stroke = Element_Stroke_read(this_ref);
+            AnnoStroke anno_stroke;
+            anno_stroke.rgba_color = stroke.rgba_color;
+            vec2 lw = get_linewidth(st);
+            anno_stroke.bbox = st.bbox + vec4(-lw, lw);
+            anno_stroke.linewidth = st.linewidth * sqrt(st.mat.x * st.mat.w - st.mat.y * st.mat.z);
+            Annotated_Stroke_write(out_ref, anno_stroke);
+            break;
+        default:
+            Annotated_Nop_write(out_ref);
+            break;
+        }
     }
 }
diff --git a/piet-gpu/shader/elements.spv b/piet-gpu/shader/elements.spv
index e97226c..41d9bc1 100644
Binary files a/piet-gpu/shader/elements.spv and b/piet-gpu/shader/elements.spv differ
diff --git a/piet-gpu/src/lib.rs b/piet-gpu/src/lib.rs
index 82b20c8..0753054 100644
--- a/piet-gpu/src/lib.rs
+++ b/piet-gpu/src/lib.rs
@@ -113,6 +113,7 @@ pub struct Renderer<D: Device> {
     scene_dev: D::Buffer,
 
     pub state_buf: D::Buffer,
+    pub anno_buf: D::Buffer,
 
     el_pipeline: D::Pipeline,
     el_ds: D::DescriptorSet,
@@ -156,14 +157,15 @@ impl<D: Device> Renderer<D> {
             .unwrap();
         device.write_buffer(&scene_buf, &scene)?;
 
-        let state_buf = device.create_buffer(4 * 1024 * 1024, dev)?;
+        let state_buf = device.create_buffer(64 * 1024 * 1024, dev)?;
+        let anno_buf = device.create_buffer(64 * 1024 * 1024, dev)?;
         let image_dev = device.create_image2d(WIDTH as u32, HEIGHT as u32, dev)?;
 
         let el_code = include_bytes!("../shader/elements.spv");
-        let el_pipeline = device.create_simple_compute_pipeline(el_code, 2, 0)?;
+        let el_pipeline = device.create_simple_compute_pipeline(el_code, 3, 0)?;
         let el_ds = device.create_descriptor_set(
             &el_pipeline,
-            &[&scene_dev, &state_buf],
+            &[&scene_dev, &state_buf, &anno_buf],
             &[],
         )?;
 
@@ -252,6 +254,7 @@ impl<D: Device> Renderer<D> {
             el_pipeline,
             el_ds,
             state_buf,
+            anno_buf,
             n_elements,
         })
     }
diff --git a/piet-gpu/src/render_ctx.rs b/piet-gpu/src/render_ctx.rs
index ad84a60..e01a6ae 100644
--- a/piet-gpu/src/render_ctx.rs
+++ b/piet-gpu/src/render_ctx.rs
@@ -199,7 +199,7 @@ impl RenderContext for PietGpuRenderContext {
 
 impl PietGpuRenderContext {
     fn encode_path(&mut self, path: impl Iterator<Item = PathEl>) {
-        let flatten = false;
+        let flatten = true;
         if flatten {
             let mut start_pt = None;
             let mut last_pt = None;