diff --git a/piet-gpu/bin/cli.rs b/piet-gpu/bin/cli.rs
index f37f0cd..c5c0b6b 100644
--- a/piet-gpu/bin/cli.rs
+++ b/piet-gpu/bin/cli.rs
@@ -34,6 +34,47 @@ fn dump_state(buf: &[u8]) {
 
 }
 
+/// Interpret the output of the binning stage, for diagnostic purposes.
+#[allow(unused)]
+fn trace_merge(buf: &[u32]) {
+    for bin in 0..256 {
+        println!("bin {}:", bin);
+        let mut starts = (0..16).map(|i| Some((bin * 16 + i) * 64)).collect::<Vec<Option<usize>>>();
+        loop {
+            let min_start = starts.iter().map(|st|
+                st.map(|st|
+                    if buf[st / 4] == 0 {
+                        !0
+                    } else {
+                        buf[st / 4 + 2]
+                    }).unwrap_or(!0)).min().unwrap();
+            if min_start == !0 {
+                break;
+            }
+            let mut selected = !0;
+            for i in 0..16 {
+                if let Some(st) = starts[i] {
+                    if buf[st/4] != 0 && buf[st/4 + 2] == min_start {
+                        selected = i;
+                        break;
+                    }
+                }
+            }
+            let st = starts[selected].unwrap();
+            println!("selected {}, start {:x}", selected, st);
+            for j in 0..buf[st/4] {
+                println!("{:x}", buf[st/4 + 2 + j as usize])
+            }
+            if buf[st/4 + 1] == 0 {
+                starts[selected] = None;
+            } else {
+                starts[selected] = Some(buf[st/4 + 1] as usize);
+            }
+        }
+
+    }
+}
+
 fn main() -> Result<(), Error> {
     let (instance, _) = VkInstance::new(None)?;
     unsafe {
@@ -66,12 +107,9 @@ fn main() -> Result<(), Error> {
 
         /*
         let mut data: Vec<u32> = Default::default();
-        device.read_buffer(&renderer.ptcl_buf, &mut data).unwrap();
-        piet_gpu::dump_k1_data(&data);
-
-        let mut data: Vec<u32> = Default::default();
-        device.read_buffer(&renderer.anno_buf, &mut data).unwrap();
-        piet_gpu::dump_k1_data(&data);
+        device.read_buffer(&renderer.bin_buf, &mut data).unwrap();
+        //piet_gpu::dump_k1_data(&data);
+        //trace_merge(&data);
         */
 
         let mut img_data: Vec<u8> = Default::default();
diff --git a/piet-gpu/shader/coarse.comp b/piet-gpu/shader/coarse.comp
index f94dc6b..2ca0cff 100644
--- a/piet-gpu/shader/coarse.comp
+++ b/piet-gpu/shader/coarse.comp
@@ -109,8 +109,8 @@ void main() {
         sh_first_el[th_ix] = chunk.n > 0 ?
             BinInstance_read(BinInstanceRef(start_chunk + BinChunk_size)).element_ix : ~0;
     }
-    uint probe = 0; // for debugging
-    do {
+    uint count = 0;
+    while (true) {
         for (uint i = 0; i < N_SLICE; i++) {
             sh_bitmaps[i][th_ix] = 0;
         }
@@ -121,18 +121,18 @@ void main() {
             if (th_ix < N_WG) {
                 if (th_ix == 0) {
                     sh_selected_n = 0;
-                    sh_min_buf = ~1;
+                    sh_min_buf = ~0;
                 }
             }
             barrier();
             // Tempting to do this with subgroups, but atomic should be good enough.
-            my_min = sh_first_el[th_ix];
             if (th_ix < N_WG) {
+                my_min = sh_first_el[th_ix];
                 atomicMin(sh_min_buf, my_min);
             }
             barrier();
             if (th_ix < N_WG) {
-                if (sh_first_el[th_ix] == sh_min_buf) {
+                if (my_min == sh_min_buf && my_min != ~0) {
                     sh_elements_ref = sh_chunk[th_ix] + BinChunk_size;
                     uint selected_n = sh_chunk_n[th_ix];
                     sh_selected_n = selected_n;
@@ -162,6 +162,7 @@ void main() {
             }
             wr_ix += chunk_n;
         }
+        barrier();
 
         // We've done the merge and filled the buffer.
 
@@ -272,8 +273,11 @@ void main() {
             // clear LSB
             bitmap &= bitmap - 1;
         }
+        barrier();
 
         rd_ix += N_TILE;
-        break;
-    } while (wr_ix > rd_ix);
+        // The second disjunct is there as a strange workaround on Nvidia. If it is
+        // removed, then the kernel fails with ERROR_DEVICE_LOST.
+        if (rd_ix >= wr_ix || bin_ix == ~0) break;
+    }
 }
diff --git a/piet-gpu/shader/coarse.spv b/piet-gpu/shader/coarse.spv
index 92ecbae..b0bec3f 100644
Binary files a/piet-gpu/shader/coarse.spv and b/piet-gpu/shader/coarse.spv differ
diff --git a/piet-gpu/src/lib.rs b/piet-gpu/src/lib.rs
index bf7a7c7..2dca39d 100644
--- a/piet-gpu/src/lib.rs
+++ b/piet-gpu/src/lib.rs
@@ -65,7 +65,7 @@ pub fn render_scene(rc: &mut impl RenderContext) {
 
 #[allow(unused)]
 fn render_cardioid(rc: &mut impl RenderContext) {
-    let n = 91;
+    let n = 601;
     let dth = std::f64::consts::PI * 2.0 / (n as f64);
     let center = Point::new(1024.0, 768.0);
     let r = 750.0;
@@ -73,7 +73,7 @@ fn render_cardioid(rc: &mut impl RenderContext) {
     for i in 1..n {
         let p0 = center + Vec2::from_angle(i as f64 * dth) * r;
         let p1 = center + Vec2::from_angle(((i * 2) % n) as f64 * dth) * r;
-        rc.fill(&Circle::new(p0, 8.0), &Color::WHITE);
+        //rc.fill(&Circle::new(p0, 8.0), &Color::WHITE);
         path.move_to(p0);
         path.line_to(p1);
         //rc.stroke(Line::new(p0, p1), &Color::BLACK, 2.0);
diff --git a/piet-gpu/src/pico_svg.rs b/piet-gpu/src/pico_svg.rs
index a630f70..9cf5cc3 100644
--- a/piet-gpu/src/pico_svg.rs
+++ b/piet-gpu/src/pico_svg.rs
@@ -48,7 +48,7 @@ impl PicoSvg {
                 }
                 if let Some(stroke_color) = el.attribute("stroke") {
                     let width =
-                        f64::from_str(el.attribute("stroke-width").ok_or("missing width")?)?;
+                        scale * f64::from_str(el.attribute("stroke-width").ok_or("missing width")?)?;
                     let color = parse_color(stroke_color);
                     items.push(Item::Stroke(StrokeItem { width, color, path }));
                 }