diff --git a/piet-gpu/shader/coarse.comp b/piet-gpu/shader/coarse.comp
index 860be2e..a4837bd 100644
--- a/piet-gpu/shader/coarse.comp
+++ b/piet-gpu/shader/coarse.comp
@@ -324,16 +324,18 @@ void main() {
                         if (tile.tile.offset != 0) {
                             CmdFill cmd_fill = CmdFill(tile.tile.offset, tile.backdrop);
                             Cmd_Fill_write(cmd_alloc, cmd_ref, cmd_fill);
+                            cmd_ref.offset += 4 + CmdFill_size;
                         } else {
                             Cmd_Solid_write(cmd_alloc, cmd_ref);
+                            cmd_ref.offset += 4;
                         }
                     } else {
                         CmdStroke cmd_stroke = CmdStroke(tile.tile.offset, 0.5 * fill.linewidth);
                         Cmd_Stroke_write(cmd_alloc, cmd_ref, cmd_stroke);
+                        cmd_ref.offset += 4 + CmdStroke_size;
                     }
-                    cmd_ref.offset += Cmd_size;
                     Cmd_Color_write(cmd_alloc, cmd_ref, CmdColor(fill.rgba_color));
-                    cmd_ref.offset += Cmd_size;
+                    cmd_ref.offset += 4 + CmdColor_size;
                     break;
                 case Annotated_Image:
                     tile = Tile_read(read_tile_alloc(element_ref_ix), TileRef(sh_tile_base[element_ref_ix]
@@ -346,16 +348,18 @@ void main() {
                         if (tile.tile.offset != 0) {
                             CmdFill cmd_fill = CmdFill(tile.tile.offset, tile.backdrop);
                             Cmd_Fill_write(cmd_alloc, cmd_ref, cmd_fill);
+                            cmd_ref.offset += 4 + CmdFill_size;
                         } else {
                             Cmd_Solid_write(cmd_alloc, cmd_ref);
+                            cmd_ref.offset += 4;
                         }
                     } else {
                         CmdStroke cmd_stroke = CmdStroke(tile.tile.offset, 0.5 * fill_img.linewidth);
                         Cmd_Stroke_write(cmd_alloc, cmd_ref, cmd_stroke);
+                        cmd_ref.offset += 4 + CmdStroke_size;
                     }
-                    cmd_ref.offset += Cmd_size;
                     Cmd_Image_write(cmd_alloc, cmd_ref, CmdImage(fill_img.index, fill_img.offset));
-                    cmd_ref.offset += Cmd_size;
+                    cmd_ref.offset += 4 + CmdImage_size;
                     break;
                 case Annotated_BeginClip:
                     tile = Tile_read(read_tile_alloc(element_ref_ix), TileRef(sh_tile_base[element_ref_ix]
@@ -373,18 +377,20 @@ void main() {
                             if (tile.tile.offset != 0) {
                                 CmdFill cmd_fill = CmdFill(tile.tile.offset, tile.backdrop);
                                 Cmd_Fill_write(cmd_alloc, cmd_ref, cmd_fill);
+                                cmd_ref.offset += 4 + CmdFill_size;
                             } else {
                                 // TODO: here is where a bunch of optimization magic should happen
                                 float alpha = tile.backdrop == 0 ? 0.0 : 1.0;
                                 Cmd_Alpha_write(cmd_alloc, cmd_ref, CmdAlpha(alpha));
+                                cmd_ref.offset += 4 + CmdAlpha_size;
                             }
                         } else {
                             CmdStroke cmd_stroke = CmdStroke(tile.tile.offset, 0.5 * begin_clip.linewidth);
                             Cmd_Stroke_write(cmd_alloc, cmd_ref, cmd_stroke);
+                            cmd_ref.offset += 4 + CmdStroke_size;
                         }
-                        cmd_ref.offset += Cmd_size;
                         Cmd_BeginClip_write(cmd_alloc, cmd_ref);
-                        cmd_ref.offset += Cmd_size;
+                        cmd_ref.offset += 4;
                         if (clip_depth < 32) {
                             clip_one_mask &= ~(1 << clip_depth);
                         }
@@ -398,9 +404,9 @@ void main() {
                             break;
                         }
                         Cmd_Solid_write(cmd_alloc, cmd_ref);
-                        cmd_ref.offset += Cmd_size;
+                        cmd_ref.offset += 4;
                         Cmd_EndClip_write(cmd_alloc, cmd_ref);
-                        cmd_ref.offset += Cmd_size;
+                        cmd_ref.offset += 4;
                     }
                     break;
                 }
diff --git a/piet-gpu/shader/coarse.spv b/piet-gpu/shader/coarse.spv
index 7e85c00..763da55 100644
Binary files a/piet-gpu/shader/coarse.spv and b/piet-gpu/shader/coarse.spv differ
diff --git a/piet-gpu/shader/kernel4.comp b/piet-gpu/shader/kernel4.comp
index 8e26e1b..b9f59dc 100644
--- a/piet-gpu/shader/kernel4.comp
+++ b/piet-gpu/shader/kernel4.comp
@@ -153,6 +153,7 @@ void main() {
             for (uint k = 0; k < CHUNK; k++) {
                 area[k] = clamp(stroke.half_width + 0.5 - df[k], 0.0, 1.0);
             }
+            cmd_ref.offset += 4 + CmdStroke_size;
             break;
         case Cmd_Fill:
             CmdFill fill = Cmd_Fill_read(cmd_alloc, cmd_ref);
@@ -184,17 +185,20 @@ void main() {
             for (uint k = 0; k < CHUNK; k++) {
                 area[k] = min(abs(area[k]), 1.0);
             }
+            cmd_ref.offset += 4 + CmdFill_size;
             break;
         case Cmd_Solid:
             for (uint k = 0; k < CHUNK; k++) {
                 area[k] = 1.0;
             }
+            cmd_ref.offset += 4;
             break;
         case Cmd_Alpha:
             CmdAlpha alpha = Cmd_Alpha_read(cmd_alloc, cmd_ref);
             for (uint k = 0; k < CHUNK; k++) {
                 area[k] = alpha.alpha;
             }
+            cmd_ref.offset += 4 + CmdAlpha_size;
             break;
         case Cmd_Color:
             CmdColor color = Cmd_Color_read(cmd_alloc, cmd_ref);
@@ -202,6 +206,7 @@ void main() {
             for (uint k = 0; k < CHUNK; k++) {
                 rgb[k] = mix(rgb[k], fg_rgba.rgb, mask[k] * area[k] * fg_rgba.a);
             }
+            cmd_ref.offset += 4 + CmdColor_size;
             break;
         case Cmd_Image:
             CmdImage fill_img = Cmd_Image_read(cmd_alloc, cmd_ref);
@@ -209,6 +214,7 @@ void main() {
             for (uint k = 0; k < CHUNK; k++) {
                 rgb[k] = mix(rgb[k], rgba[k].rgb, mask[k] * area[k] * rgba[k].a);
             }
+            cmd_ref.offset += 4 + CmdImage_size;
             break;
         case Cmd_BeginClip:
             uint blend_slot = blend_sp % BLEND_STACK_SIZE;
@@ -229,6 +235,7 @@ void main() {
                 blend_stack[blend_slot][k] = packsRGB(vec4(rgb[k], clamp(abs(area[k]), 0.0, 1.0)));
             }
             blend_sp++;
+            cmd_ref.offset += 4;
             break;
         case Cmd_EndClip:
             blend_slot = (blend_sp - 1) % BLEND_STACK_SIZE;
@@ -245,13 +252,13 @@ void main() {
                 vec4 rgba = unpacksRGB(blend_stack[blend_slot][k]);
                 rgb[k] = mix(rgba.rgb, rgb[k], area[k] * rgba.a);
             }
+            cmd_ref.offset += 4;
             break;
         case Cmd_Jump:
             cmd_ref = CmdRef(Cmd_Jump_read(cmd_alloc, cmd_ref).new_ref);
             cmd_alloc.offset = cmd_ref.offset;
             continue;
         }
-        cmd_ref.offset += Cmd_size;
     }
 
     for (uint i = 0; i < CHUNK; i++) {
diff --git a/piet-gpu/shader/kernel4.spv b/piet-gpu/shader/kernel4.spv
index 91e54dd..5f776e4 100644
Binary files a/piet-gpu/shader/kernel4.spv and b/piet-gpu/shader/kernel4.spv differ