diff --git a/piet-gpu/src/encoder.rs b/piet-gpu/src/encoder.rs
index 2f4b85e..a24ddbc 100644
--- a/piet-gpu/src/encoder.rs
+++ b/piet-gpu/src/encoder.rs
@@ -37,6 +37,147 @@ pub struct Encoder {
     n_clip: u32,
 }
 
+#[derive(Copy, Clone, Debug)]
+pub struct EncodedSceneRef<'a, T: Copy + Pod> {
+    pub transform_stream: &'a [T],
+    pub tag_stream: &'a [u8],
+    pub pathseg_stream: &'a [u8],
+    pub linewidth_stream: &'a [f32],
+    pub drawtag_stream: &'a [u32],
+    pub drawdata_stream: &'a [u8],
+    pub n_path: u32,
+    pub n_pathseg: u32,
+    pub n_clip: u32,
+    pub ramp_data: &'a [u32],
+}
+
+impl<'a, T: Copy + Pod> EncodedSceneRef<'a, T> {
+    /// Return a config for the element processing pipeline.
+    ///
+    /// This does not include further pipeline processing. Also returns the
+    /// beginning of free memory.
+    pub fn stage_config(&self) -> (Config, usize) {
+        // Layout of scene buffer
+        let drawtag_offset = 0;
+        let n_drawobj = self.n_drawobj();
+        let n_drawobj_padded = align_up(n_drawobj, DRAW_PART_SIZE as usize);
+        let drawdata_offset = drawtag_offset + n_drawobj_padded * DRAWTAG_SIZE;
+        let trans_offset = drawdata_offset + self.drawdata_stream.len();
+        let n_trans = self.transform_stream.len();
+        let n_trans_padded = align_up(n_trans, TRANSFORM_PART_SIZE as usize);
+        let linewidth_offset = trans_offset + n_trans_padded * TRANSFORM_SIZE;
+        let n_linewidth = self.linewidth_stream.len();
+        let pathtag_offset = linewidth_offset + n_linewidth * LINEWIDTH_SIZE;
+        let n_pathtag = self.tag_stream.len();
+        let n_pathtag_padded = align_up(n_pathtag, PATHSEG_PART_SIZE as usize);
+        let pathseg_offset = pathtag_offset + n_pathtag_padded;
+
+        // Layout of memory
+        let mut alloc = 0;
+        let trans_alloc = alloc;
+        alloc += trans_alloc + n_trans_padded * TRANSFORM_SIZE;
+        let pathseg_alloc = alloc;
+        alloc += pathseg_alloc + self.n_pathseg as usize * PATHSEG_SIZE;
+        let path_bbox_alloc = alloc;
+        let n_path = self.n_path as usize;
+        alloc += path_bbox_alloc + n_path * PATH_BBOX_SIZE;
+        let drawmonoid_alloc = alloc;
+        alloc += n_drawobj_padded * DRAWMONOID_SIZE;
+        let anno_alloc = alloc;
+        alloc += n_drawobj * ANNOTATED_SIZE;
+        let clip_alloc = alloc;
+        let n_clip = self.n_clip as usize;
+        const CLIP_SIZE: usize = 4;
+        alloc += n_clip * CLIP_SIZE;
+        let clip_bic_alloc = alloc;
+        const CLIP_BIC_SIZE: usize = 8;
+        // This can round down, as we only reduce the prefix
+        alloc += (n_clip / CLIP_PART_SIZE as usize) * CLIP_BIC_SIZE;
+        let clip_stack_alloc = alloc;
+        const CLIP_EL_SIZE: usize = 20;
+        alloc += n_clip * CLIP_EL_SIZE;
+        let clip_bbox_alloc = alloc;
+        const CLIP_BBOX_SIZE: usize = 16;
+        alloc += align_up(n_clip as usize, CLIP_PART_SIZE as usize) * CLIP_BBOX_SIZE;
+        let draw_bbox_alloc = alloc;
+        alloc += n_drawobj * DRAW_BBOX_SIZE;
+        let drawinfo_alloc = alloc;
+        // TODO: not optimized; it can be accumulated during encoding or summed from drawtags
+        const MAX_DRAWINFO_SIZE: usize = 44;
+        alloc += n_drawobj * MAX_DRAWINFO_SIZE;
+
+        let config = Config {
+            n_elements: n_drawobj as u32,
+            n_pathseg: self.n_pathseg,
+            pathseg_alloc: pathseg_alloc as u32,
+            anno_alloc: anno_alloc as u32,
+            trans_alloc: trans_alloc as u32,
+            path_bbox_alloc: path_bbox_alloc as u32,
+            drawmonoid_alloc: drawmonoid_alloc as u32,
+            clip_alloc: clip_alloc as u32,
+            clip_bic_alloc: clip_bic_alloc as u32,
+            clip_stack_alloc: clip_stack_alloc as u32,
+            clip_bbox_alloc: clip_bbox_alloc as u32,
+            draw_bbox_alloc: draw_bbox_alloc as u32,
+            drawinfo_alloc: drawinfo_alloc as u32,
+            n_trans: n_trans as u32,
+            n_path: self.n_path,
+            n_clip: self.n_clip,
+            trans_offset: trans_offset as u32,
+            linewidth_offset: linewidth_offset as u32,
+            pathtag_offset: pathtag_offset as u32,
+            pathseg_offset: pathseg_offset as u32,
+            drawtag_offset: drawtag_offset as u32,
+            drawdata_offset: drawdata_offset as u32,
+            ..Default::default()
+        };
+        (config, alloc)
+    }
+
+    pub fn write_scene(&self, buf: &mut BufWrite) {
+        buf.extend_slice(&self.drawtag_stream);
+        let n_drawobj = self.drawtag_stream.len();
+        buf.fill_zero(padding(n_drawobj, DRAW_PART_SIZE as usize) * DRAWTAG_SIZE);
+        buf.extend_slice(&self.drawdata_stream);
+        buf.extend_slice(&self.transform_stream);
+        let n_trans = self.transform_stream.len();
+        buf.fill_zero(padding(n_trans, TRANSFORM_PART_SIZE as usize) * TRANSFORM_SIZE);
+        buf.extend_slice(&self.linewidth_stream);
+        buf.extend_slice(&self.tag_stream);
+        let n_pathtag = self.tag_stream.len();
+        buf.fill_zero(padding(n_pathtag, PATHSEG_PART_SIZE as usize));
+        buf.extend_slice(&self.pathseg_stream);
+    }
+
+    /// The number of draw objects in the draw object stream.
+    pub(crate) fn n_drawobj(&self) -> usize {
+        self.drawtag_stream.len()
+    }
+
+    /// The number of paths.
+    pub(crate) fn n_path(&self) -> u32 {
+        self.n_path
+    }
+
+    /// The number of path segments.
+    pub(crate) fn n_pathseg(&self) -> u32 {
+        self.n_pathseg
+    }
+
+    pub(crate) fn n_transform(&self) -> usize {
+        self.transform_stream.len()
+    }
+
+    /// The number of tags in the path stream.
+    pub(crate) fn n_pathtag(&self) -> usize {
+        self.tag_stream.len()
+    }
+
+    pub(crate) fn n_clip(&self) -> u32 {
+        self.n_clip
+    }
+}
+
 /// A scene fragment encoding a glyph.
 ///
 /// This is a reduced version of the full encoder.
diff --git a/piet-gpu/src/lib.rs b/piet-gpu/src/lib.rs
index 773007d..d32a9c5 100644
--- a/piet-gpu/src/lib.rs
+++ b/piet-gpu/src/lib.rs
@@ -8,9 +8,11 @@ pub mod stages;
 pub mod test_scenes;
 mod text;
 
+use bytemuck::Pod;
 use std::convert::TryInto;
 
 pub use blend::{Blend, BlendMode, CompositionMode};
+pub use encoder::EncodedSceneRef;
 pub use render_ctx::PietGpuRenderContext;
 pub use gradient::Colrv1RadialGradient;
 
@@ -406,6 +408,61 @@ impl Renderer {
         Ok(())
     }
 
+    pub fn upload_scene<T: Copy + Pod>(
+        &mut self,
+        scene: &EncodedSceneRef<T>,
+        buf_ix: usize,
+    ) -> Result<(), Error> {
+        let (mut config, mut alloc) = scene.stage_config();
+        let n_drawobj = scene.n_drawobj();
+        // TODO: be more consistent in size types
+        let n_path = scene.n_path() as usize;
+        self.n_paths = n_path;
+        self.n_transform = scene.n_transform();
+        self.n_drawobj = scene.n_drawobj();
+        self.n_pathseg = scene.n_pathseg() as usize;
+        self.n_pathtag = scene.n_pathtag();
+        self.n_clip = scene.n_clip();
+
+        // These constants depend on encoding and may need to be updated.
+        // Perhaps we can plumb these from piet-gpu-derive?
+        const PATH_SIZE: usize = 12;
+        const BIN_SIZE: usize = 8;
+        let width_in_tiles = self.width / TILE_W;
+        let height_in_tiles = self.height / TILE_H;
+        let tile_base = alloc;
+        alloc += ((n_path + 3) & !3) * PATH_SIZE;
+        let bin_base = alloc;
+        alloc += ((n_drawobj + 255) & !255) * BIN_SIZE;
+        let ptcl_base = alloc;
+        alloc += width_in_tiles * height_in_tiles * PTCL_INITIAL_ALLOC;
+
+        config.width_in_tiles = width_in_tiles as u32;
+        config.height_in_tiles = height_in_tiles as u32;
+        config.tile_alloc = tile_base as u32;
+        config.bin_alloc = bin_base as u32;
+        config.ptcl_alloc = ptcl_base as u32;
+        unsafe {
+            // TODO: reallocate scene buffer if size is inadequate
+            {
+                let mut mapped_scene = self.scene_bufs[buf_ix].map_write(..)?;
+                scene.write_scene(&mut mapped_scene);
+            }
+            self.config_bufs[buf_ix].write(&[config])?;
+            self.memory_buf_host[buf_ix].write(&[alloc as u32, 0 /* Overflow flag */])?;
+
+            // Upload gradient data.
+            if !scene.ramp_data.is_empty() {
+                assert!(
+                    self.gradient_bufs[buf_ix].size() as usize
+                        >= std::mem::size_of_val(&*scene.ramp_data)
+                );
+                self.gradient_bufs[buf_ix].write(scene.ramp_data)?;
+            }
+        }
+        Ok(())
+    }
+
     pub unsafe fn record(&self, cmd_buf: &mut CmdBuf, query_pool: &QueryPool, buf_ix: usize) {
         cmd_buf.copy_buffer(&self.config_bufs[buf_ix], &self.config_buf);
         cmd_buf.copy_buffer(&self.memory_buf_host[buf_ix], &self.memory_buf_dev);
diff --git a/piet-scene/src/main.rs b/piet-scene/src/main.rs
index ff4030f..b1b777b 100644
--- a/piet-scene/src/main.rs
+++ b/piet-scene/src/main.rs
@@ -57,7 +57,7 @@ fn main() -> Result<(), Error> {
             .map(|_| session.create_semaphore())
             .collect::<Result<Vec<_>, Error>>()?;
         let query_pools = (0..NUM_FRAMES)
-            .map(|_| session.create_query_pool(8))
+            .map(|_| session.create_query_pool(12))
             .collect::<Result<Vec<_>, Error>>()?;
         let mut cmd_bufs: [Option<CmdBuf>; NUM_FRAMES] = Default::default();
         let mut submitted: [Option<SubmittedCmdBuf>; NUM_FRAMES] = Default::default();
@@ -99,72 +99,26 @@ fn main() -> Result<(), Error> {
                         if !ts.is_empty() {
                             info_string = format!(
                                 "{:.3}ms :: e:{:.3}ms|alloc:{:.3}ms|cp:{:.3}ms|bd:{:.3}ms|bin:{:.3}ms|cr:{:.3}ms|r:{:.3}ms",
-                                ts[6] * 1e3,
+                                ts[10] * 1e3,
                                 ts[0] * 1e3,
                                 (ts[1] - ts[0]) * 1e3,
                                 (ts[2] - ts[1]) * 1e3,
-                                (ts[3] - ts[2]) * 1e3,
                                 (ts[4] - ts[3]) * 1e3,
-                                (ts[5] - ts[4]) * 1e3,
                                 (ts[6] - ts[5]) * 1e3,
+                                (ts[8] - ts[7]) * 1e3,
+                                (ts[10] - ts[9]) * 1e3,
                             );
                         }
                     }
 
                     let mut ctx = PietGpuRenderContext::new();
-                    if let Some(input) = matches.value_of("INPUT") {
-                        let mut scale = matches
-                            .value_of("scale")
-                            .map(|scale| scale.parse().unwrap())
-                            .unwrap_or(8.0);
-                        if matches.is_present("flip") {
-                            scale = -scale;
-                        }
-                        test_scenes::render_svg(&mut ctx, input, scale);
-                    } else {
-                        use piet_gpu::{Blend, BlendMode::*, CompositionMode::*};
-                        let blends = [
-                            Blend::new(Normal, SrcOver),
-                            Blend::new(Multiply, SrcOver),
-                            Blend::new(Screen, SrcOver),
-                            Blend::new(Overlay, SrcOver),
-                            Blend::new(Darken, SrcOver),
-                            Blend::new(Lighten, SrcOver),
-                            Blend::new(ColorDodge, SrcOver),
-                            Blend::new(ColorBurn, SrcOver),
-                            Blend::new(HardLight, SrcOver),
-                            Blend::new(SoftLight, SrcOver),
-                            Blend::new(Difference, SrcOver),
-                            Blend::new(Exclusion, SrcOver),
-                            Blend::new(Hue, SrcOver),
-                            Blend::new(Saturation, SrcOver),
-                            Blend::new(Color, SrcOver),
-                            Blend::new(Luminosity, SrcOver),
-                            Blend::new(Normal, Clear),
-                            Blend::new(Normal, Copy),
-                            Blend::new(Normal, Dest),
-                            Blend::new(Normal, SrcOver),
-                            Blend::new(Normal, DestOver),
-                            Blend::new(Normal, SrcIn),
-                            Blend::new(Normal, DestIn),
-                            Blend::new(Normal, SrcOut),
-                            Blend::new(Normal, DestOut),
-                            Blend::new(Normal, SrcAtop),
-                            Blend::new(Normal, DestAtop),
-                            Blend::new(Normal, Xor),
-                            Blend::new(Normal, Plus),
-                        ];
-                        let blend = blends[mode % blends.len()];
-                        test_scenes::render_blend_test(&mut ctx, current_frame, blend);
-                        info_string = format!("{:?}", blend);
-                    }
                     render_info_string(&mut ctx, &info_string);
 
                     ctx = PietGpuRenderContext::new();
                     test_scene1_old(&mut ctx);
-                    let mut encoded_scene_old = ctx.encoded_scene();
-                    let ramp_data = ctx.get_ramp_data();
-                    encoded_scene_old.ramp_data = &ramp_data;
+                    //let mut encoded_scene_old = ctx.encoded_scene();
+                    // let ramp_data = ctx.get_ramp_data();
+                    //encoded_scene_old.ramp_data = &ramp_data;
                     test_scene1(&mut scene, &mut rcx);
                     let encoded_scene = scene_to_encoded_scene(&scene, &rcx);
                     // println!("{:?}\n============\n{:?}", encoded_scene_old, encoded_scene);