⚡ Soft-drain GPU mid-walk on progressive Partials (#11127)

Release packs far more cheap Current draws (e.g. fills_none paths)
into one Partial than debug; a single end-of-Partial
flush_and_submit then stalls the browser. Soft-flush every N walker
nodes (and on Partial yield) keeps ops buffers bounded while Full
still submits via present_frame.

render-wasm/src/render.rs

@@ -1386,8 +1386,8 @@ impl RenderState {
         // the layered path without Fills/Strokes blits.
         // Non-SrcOver blend, frame clip blur, and masked groups stay layered.
         // Stroke-only (fills_none) can go direct: empty fills are a no-op and
-        // strokes paint into Current. Requires Partial GPU drain (dc1ab) so
+        // strokes paint into Current. Large files need mid-walk GPU drains so
-        // large SVG-icon files do not backlog commands until Full present.
+        // release builds do not backlog a huge ops buffer in one Partial.
         let can_render_directly = apply_to_current_surface
             && offset.is_none()
             && parent_shadows.is_none()
@@ -2461,12 +2461,9 @@ impl RenderState {
                 panic!("FrameType::None");
             }
             FrameType::Partial => {
-                // Drain tile GPU work (Current / tile atlas / cache) without
+                // Final soft drain for this yield (mid-walk also drains; see
-                // presenting Target and without re-snapshotting the tile atlas —
+                // `drain_partial_gpu_soft`). Full still submits via present_frame.
-                // composition stays deferred to Full. A Backbuffer flush alone
+                Self::drain_partial_gpu_soft();
-                // left commands queued until present_frame's flush_and_submit,
-                // which stalled the browser on large files.
-                crate::get_gpu_state().context.flush_and_submit();
             }
             FrameType::Full => {
                 // A full-quality frame is now complete. Rebuild the per-shape crop
@@ -2669,6 +2666,15 @@ impl RenderState {
         true
     }
 
+    /// Soft-drain GPU command buffers during progressive tile walks.
+    /// Release packs far more cheap Current draws (e.g. fills_none paths) into
+    /// one Partial than debug; flushing only at Partial end then stalls. Call
+    /// periodically so each flush stays small. Full present still submits.
+    #[inline]
+    fn drain_partial_gpu_soft() {
+        crate::get_gpu_state().context.flush(None);
+    }
+
     /// Skip all drop/inner shadows in fast mode, or when even a large design-space
     /// shadow would be subpixel. Otherwise filter per shadow via
     /// [`Shadow::is_perceptible_at_scale_for`] (stricter for recursive shapes).
@@ -3775,6 +3781,12 @@ impl RenderState {
             if allow_stop && self.should_stop_rendering(iteration, timestamp) {
                 return Ok((is_empty, true));
             }
+            // Keep GPU ops buffers bounded when many shapes paint cheaply to
+            // Current (release packs far more per Partial than debug).
+            let drain_every = self.options.partial_gpu_drain_every_n;
+            if allow_stop && drain_every > 0 && iteration > 0 && iteration % drain_every == 0 {
+                Self::drain_partial_gpu_soft();
+            }
             iteration += 1;
         }
 

render-wasm/src/render/options.rs

@@ -11,6 +11,10 @@ const VIEWPORT_INTEREST_AREA_THRESHOLD: i32 = 1;
 const MIN_DPR_VIEWPORT_INTEREST_AREA_THRESHOLD: i32 = 2;
 const MAX_BLOCKING_TIME_MS: i32 = 32;
 const NODE_BATCH_THRESHOLD: i32 = 3;
+/// Soft-drain GPU every N walker nodes on progressive Partials. Keeps ops
+/// buffers bounded when many shapes paint cheaply to Current (release packs
+/// far more per budget than debug).
+const PARTIAL_GPU_DRAIN_EVERY_N: i32 = 64;
 const BLUR_DOWNSCALE_THRESHOLD: f32 = 8.0;
 const ANTIALIAS_THRESHOLD: f32 = 7.0;
 #[derive(Debug, Copy, Clone, PartialEq)]
@@ -29,6 +33,9 @@ pub struct RenderOptions {
     pub dpr_viewport_interest_area_threshold: i32,
     pub max_blocking_time_ms: i32,
     pub node_batch_threshold: i32,
+    /// Soft-flush GPU every N nodes during progressive tile walks (see
+    /// [`PARTIAL_GPU_DRAIN_EVERY_N`]).
+    pub partial_gpu_drain_every_n: i32,
     pub blur_downscale_threshold: f32,
     pub capture_frames: i32,
 }
@@ -45,6 +52,7 @@ impl Default for RenderOptions {
             dpr_viewport_interest_area_threshold: VIEWPORT_INTEREST_AREA_THRESHOLD,
             max_blocking_time_ms: MAX_BLOCKING_TIME_MS,
             node_batch_threshold: NODE_BATCH_THRESHOLD,
+            partial_gpu_drain_every_n: PARTIAL_GPU_DRAIN_EVERY_N,
             blur_downscale_threshold: BLUR_DOWNSCALE_THRESHOLD,
             capture_frames: 0,
         }