From 54e68df295d772e068506e53551786051a6fda02 Mon Sep 17 00:00:00 2001
From: Hibryda <hibryda@protonmail.com>
Date: Thu, 19 Mar 2026 23:47:39 +0100
Subject: [PATCH] docs: deep dive into GPUI dirty propagation, SharedBlink,
 hierarchy flattening results

---
 docs/architecture/gpui-findings.md | 55 ++++++++++++++++++++++++++++++
 1 file changed, 55 insertions(+)

diff --git a/docs/architecture/gpui-findings.md b/docs/architecture/gpui-findings.md
index c457c41..56aede7 100644
--- a/docs/architecture/gpui-findings.md
+++ b/docs/architecture/gpui-findings.md
@@ -26,6 +26,61 @@ Key mechanisms:
 3. **`mark_view_dirty()` walks ancestors** — this is GPUI's fundamental constraint
 4. **`refreshing=true` on cache miss** cascades down → children can't cache if parent is dirty
 
+## DEEP DIVE: Why 3% is the floor for div-based views (2026-03-19 late)
+
+### The full dirty propagation chain (confirmed by source analysis)
+
+```
+cx.notify(entity_id)
+  → App::notify → WindowInvalidator::invalidate_view(entity_id)
+    → dirty_views.insert(entity_id)
+  → draw() → invalidate_entities() → mark_view_dirty(entity_id)
+    → dispatch_tree.view_path_reversed(entity_id) → walks parent chain
+    → inserts entity + ALL render-time ancestors into dirty_views
+  → AnyView::prepaint checks: dirty_views.contains(MY entity_id)
+    → if true: cache MISS, re-render
+    → if false: cache HIT, replay GPU commands
+```
+
+### Why Zed achieves <1% with the SAME mechanism
+
+Zed's BlinkManager → observer → cx.notify(Editor) → Editor + Pane + Workspace ALL in dirty_views.
+ALL re-render. But:
+- Workspace::render() = ~5 div builders = <0.1ms
+- Pane::render() = ~10 div builders = <0.2ms
+- Editor::render() = returns EditorElement struct = <0.01ms
+- EditorElement::prepaint/paint = real work, bounded to visible rows
+
+Total: <1ms per blink frame. Our ProjectBox::render() = ~9 divs + 3 tab buttons = ~15ms.
+The 15ms includes Taffy layout + element prepaint + paint pipeline overhead per div.
+
+### SharedBlink pattern (Arc<AtomicBool>)
+
+Replaced Entity<BlinkState> with `Arc<AtomicBool>`:
+- Background timer toggles atomic bool every 500ms
+- Timer calls cx.notify() on ProjectBox directly (no intermediate entity)
+- ProjectBox::render() reads bool atomically — no Entity subscription overhead
+- Same 3% CPU — confirms cost is in render() itself, not entity machinery
+
+### Hierarchy flattening (4 levels → 2 levels)
+
+Removed ProjectGrid entity level. ProjectBoxes render directly from Workspace.
+Dispatch tree: Workspace → ProjectBox (2 levels, same as Zed's Workspace → Pane).
+CPU unchanged at 3% — confirms cost is per-frame render(), not ancestor walk count.
+
+### Path to <1%: Custom Element for ProjectBox header
+
+Convert header (accent stripe + dot + name + CWD + tab bar) from div tree to
+custom `impl Element` that paints directly via `window.paint_quad()` +
+`window.text_system().shape_line().paint()`. This eliminates:
+- 9 div() constructor calls
+- 9 Taffy layout nodes
+- 9 prepaint traversals
+- 9 paint traversals
+
+Replaced by ~6 direct GPU primitive insertions (paint_quad × 4 + shape_line × 2).
+Expected reduction: 15ms → <1ms per frame.
+
 ## Overview
 
 GPUI is the GPU-accelerated UI framework powering the Zed editor. Apache-2.0 licensed (the crate itself; Zed app is GPL-3.0). Pre-1.0 with breaking changes every 2-3 months. 76.7k stars (Zed repo), $32M Sequoia funding.