Index of every page on this site: llms.txt
All skills
geoffjay avatar

/gpui-performance

@0ab5ae1
Run it yourself

Performance optimization techniques for GPUI including rendering optimization, layout performance, memory management, and profiling strategies. Use when user needs to optimize GPUI application performance or debug performance issues.

Use this Skill: https://skilld.dev/gh/geoffjay/claude-plugins/gpui-performance

Nothing lands on disk. Nothing to clean up.

Fork this Skill

Edit a local copy. It keeps the author and licence.

SKILL.md

≈63 tokens for metadata: the name and description. ≈3.6k when used: this file.

Description uses 3.1% of example budget

Before choosing a Skill, your Agent reads its name and description. All available Skills share that space.

  • A shorter description leaves more room for other Skills. This entry exceeds our 1% size suggestion.

In our Claude Code example, all Skill names and descriptions share 8,000 characters. This Skill uses ≈252 characters, or 3.1%.

The 1% threshold is a size suggestion. Longer descriptions can still fit.

Your model, settings, and other Skills decide how much text your Agent can read.

Example settings and source

The example uses a 200k-token context and default Claude Code settings. The count includes the name, description, separators, and when_to_use when present. Codex also counts local file paths.

Skit's source and limits: Codex 0.160.1, Claude Code 2.1.292.

GPUI Performance Optimization

Metadata

This skill provides comprehensive guidance on optimizing GPUI applications for rendering performance, memory efficiency, and overall runtime speed.

Instructions

Rendering Optimization

Understanding the Render Cycle
State Change → cx.notify() → Render → Layout → Paint → Display

Key Points:

  • Only call cx.notify() when state actually changes
  • Minimize work in render() method
  • Cache expensive computations
  • Reduce element count and nesting
Avoiding Unnecessary Renders
// BAD: Renders on every frame
impl MyComponent {
    fn start_animation(&mut self, cx: &mut ViewContext<Self>) {
        cx.spawn(|this, mut cx| async move {
            loop {
                cx.update(|_, cx| cx.notify()).ok();  // Forces rerender!
                Timer::after(Duration::from_millis(16)).await;
            }
        }).detach();
    }
}

// GOOD: Only render when state changes
impl MyComponent {
    fn update_value(&mut self, new_value: i32, cx: &mut ViewContext<Self>) {
        if self.value != new_value {
            self.value = new_value;
            cx.notify();  // Only notify on actual change
        }
    }
}
Optimize Subscription Updates
// BAD: Always rerenders on model change
let _subscription = cx.observe(&model, |_, _, cx| {
    cx.notify();  // Rerenders even if nothing relevant changed
});

// GOOD: Selective updates
let _subscription = cx.observe(&model, |this, model, cx| {
    let data = model.read(cx);

    // Only rerender if relevant field changed
    if data.relevant_field != this.cached_field {
        this.cached_field = data.relevant_field.clone();
        cx.notify();
    }
});
Memoization Pattern
use std::cell::RefCell;
use std::collections::hash_map::DefaultHasher;
use std::hash::{Hash, Hasher};

struct MemoizedComponent {
    model: Model<Data>,
    cached_result: RefCell<Option<(u64, String)>>,  // (hash, result)
}

impl MemoizedComponent {
    fn expensive_computation(&self, cx: &ViewContext<Self>) -> String {
        let data = self.model.read(cx);

        // Calculate hash of input
        let mut hasher = DefaultHasher::new();
        data.relevant_fields.hash(&mut hasher);
        let hash = hasher.finish();

        // Return cached if unchanged
        if let Some((cached_hash, cached_result)) = &*self.cached_result.borrow() {
            if *cached_hash == hash {
                return cached_result.clone();
            }
        }

        // Compute and cache
        let result = perform_expensive_computation(&data);
        *self.cached_result.borrow_mut() = Some((hash, result.clone()));
        result
    }
}

Layout Performance

Minimize Layout Complexity
// BAD: Deep nesting
div()
    .flex()
    .child(
        div()
            .flex()
            .child(
                div()
                    .flex()
                    .child(
                        div().child("Content")
                    )
            )
    )

// GOOD: Flat structure
div()
    .flex()
    .flex_col()
    .gap_4()
    .child("Header")
    .child("Content")
    .child("Footer")
Use Fixed Sizing When Possible
// BETTER: Fixed sizes (no layout calculation)
div()
    .w(px(200.))
    .h(px(100.))
    .child("Fixed size")

// SLOWER: Dynamic sizing (requires layout calculation)
div()
    .w_full()
    .h_full()
    .child("Dynamic size")
Avoid Layout Thrashing
// BAD: Reading layout during render
impl Render for BadComponent {
    fn render(&mut self, cx: &mut ViewContext<Self>) -> impl IntoElement {
        let width = cx.window_bounds().get_bounds().size.width;
        // Using width immediately causes layout thrashing
        div().w(width)
    }
}

// GOOD: Cache layout-dependent values
struct GoodComponent {
    cached_width: Pixels,
}

impl GoodComponent {
    fn on_window_resize(&mut self, cx: &mut ViewContext<Self>) {
        let width = cx.window_bounds().get_bounds().size.width;
        if self.cached_width != width {
            self.cached_width = width;
            cx.notify();
        }
    }
}
Virtual Scrolling for Long Lists
struct VirtualList {
    items: Vec<String>,
    scroll_offset: f32,
    viewport_height: f32,
    item_height: f32,
}

impl Render for VirtualList {
    fn render(&mut self, cx: &mut ViewContext<Self>) -> impl IntoElement {
        // Calculate visible range
        let start_index = (self.scroll_offset / self.item_height).floor() as usize;
        let visible_count = (self.viewport_height / self.item_height).ceil() as usize;
        let end_index = (start_index + visible_count).min(self.items.len());

        // Only render visible items
        div()
            .h(px(self.viewport_height))
            .overflow_y_scroll()
            .on_scroll(cx.listener(|this, event, cx| {
                this.scroll_offset = event.scroll_offset.y;
                cx.notify();
            }))
            .child(
                div()
                    .h(px(self.items.len() as f32 * self.item_height))
                    .child(
                        div()
                            .absolute()
                            .top(px(start_index as f32 * self.item_height))
                            .children(
                                self.items[start_index..end_index]
                                    .iter()
                                    .map(|item| {
                                        div()
                                            .h(px(self.item_height))
                                            .child(item.as_str())
                                    })
                            )
                    )
            )
    }
}

Memory Management

Preventing Memory Leaks
// LEAK: Subscription not stored
impl BadView {
    fn new(model: Model<Data>, cx: &mut ViewContext<Self>) -> Self {
        cx.observe(&model, |_, _, cx| cx.notify());  // Leak!
        Self { model }
    }
}

// CORRECT: Store subscription
struct GoodView {
    model: Model<Data>,
    _subscription: Subscription,  // Cleaned up on Drop
}

impl GoodView {
    fn new(model: Model<Data>, cx: &mut ViewContext<Self>) -> Self {
        let _subscription = cx.observe(&model, |_, _, cx| cx.notify());
        Self { model, _subscription }
    }
}
Avoid Circular References
// BAD: Circular reference
struct CircularRef {
    self_view: Option<View<Self>>,  // Circular!
}

// GOOD: Use weak references or redesign
struct NoCycle {
    other_view: View<OtherView>,  // No cycle
}
Bounded Collections
use std::collections::VecDeque;

const MAX_HISTORY: usize = 100;

struct BoundedHistory {
    items: VecDeque<Item>,
}

impl BoundedHistory {
    fn add_item(&mut self, item: Item) {
        self.items.push_back(item);

        // Maintain size limit
        while self.items.len() > MAX_HISTORY {
            self.items.pop_front();
        }
    }
}
Reuse Allocations
struct BufferedComponent {
    buffer: String,  // Reused across operations
}

impl BufferedComponent {
    fn format_data(&mut self, data: &[Item]) -> &str {
        self.buffer.clear();  // Reuse allocation

        for item in data {
            use std::fmt::Write;
            write!(&mut self.buffer, "{}\n", item.name).ok();
        }

        &self.buffer
    }
}

Profiling Strategies

CPU Profiling with cargo-flamegraph
# Install
cargo install flamegraph

# Profile application
cargo flamegraph --bin your-app

# With specific features
cargo flamegraph --bin your-app --features profiling

# Opens flamegraph.svg showing CPU time distribution
Memory Profiling
# valgrind (Linux)
valgrind --tool=massif --massif-out-file=massif.out ./target/release/your-app
ms_print massif.out

# heaptrack (Linux)
heaptrack ./target/release/your-app
heaptrack_gui heaptrack.your-app.*.gz

# Instruments (macOS)
instruments -t "Allocations" ./target/release/your-app
Custom Performance Monitoring
use std::time::Instant;

struct PerformanceMonitor {
    frame_times: VecDeque<Duration>,
    max_samples: usize,
}

impl PerformanceMonitor {
    fn new() -> Self {
        Self {
            frame_times: VecDeque::with_capacity(100),
            max_samples: 100,
        }
    }

    fn record_frame(&mut self, duration: Duration) {
        self.frame_times.push_back(duration);

        if self.frame_times.len() > self.max_samples {
            self.frame_times.pop_front();
        }

        // Warn if frame is slow (> 16ms for 60fps)
        if duration.as_millis() > 16 {
            eprintln!("⚠️  Slow frame: {}ms", duration.as_millis());
        }
    }

    fn average_fps(&self) -> f64 {
        if self.frame_times.is_empty() {
            return 0.0;
        }

        let total: Duration = self.frame_times.iter().sum();
        let avg = total / self.frame_times.len() as u32;
        1000.0 / avg.as_millis() as f64
    }

    fn percentile(&self, p: f64) -> Duration {
        let mut sorted: Vec<_> = self.frame_times.iter().copied().collect();
        sorted.sort();

        let index = (sorted.len() as f64 * p) as usize;
        sorted[index.min(sorted.len() - 1)]
    }
}

// Usage in component
impl MyView {
    fn measure_render<F>(&mut self, f: F, cx: &mut ViewContext<Self>)
    where
        F: FnOnce(&mut Self, &mut ViewContext<Self>)
    {
        let start = Instant::now();
        f(self, cx);
        let elapsed = start.elapsed();

        self.perf_monitor.record_frame(elapsed);

        // Log stats periodically
        if self.frame_count % 60 == 0 {
            println!(
                "Avg FPS: {:.1}, p95: {}ms, p99: {}ms",
                self.perf_monitor.average_fps(),
                self.perf_monitor.percentile(0.95).as_millis(),
                self.perf_monitor.percentile(0.99).as_millis(),
            );
        }
    }
}
Benchmark with Criterion
// benches/component_bench.rs
use criterion::{black_box, criterion_group, criterion_main, Criterion, BenchmarkId};

fn render_benchmark(c: &mut Criterion) {
    let mut group = c.benchmark_group("rendering");

    for size in [10, 100, 1000].iter() {
        group.bench_with_input(
            BenchmarkId::from_parameter(size),
            size,
            |b, &size| {
                b.iter(|| {
                    App::test(|cx| {
                        let items = vec![Item::default(); size];
                        let view = cx.new_view(|cx| {
                            ListView::new(items, cx)
                        });

                        view.update(cx, |view, cx| {
                            black_box(view.render(cx));
                        });
                    });
                });
            }
        );
    }

    group.finish();
}

criterion_group!(benches, render_benchmark);
criterion_main!(benches);

Batching Updates

// BAD: Multiple individual updates
for item in items {
    self.model.update(cx, |model, cx| {
        model.add_item(item);  // Triggers rerender each time!
        cx.notify();
    });
}

// GOOD: Batch into single update
self.model.update(cx, |model, cx| {
    for item in items {
        model.add_item(item);
    }
    cx.notify();  // Single rerender
});

Async Rendering Optimization

struct AsyncView {
    loading_state: Model<LoadingState>,
}

impl AsyncView {
    fn load_data(&mut self, cx: &mut ViewContext<Self>) {
        let loading_state = self.loading_state.clone();

        // Show loading immediately
        self.loading_state.update(cx, |state, cx| {
            *state = LoadingState::Loading;
            cx.notify();
        });

        // Load asynchronously
        cx.spawn(|_, mut cx| async move {
            // Fetch data
            let data = fetch_data().await?;

            // Update state once
            cx.update_model(&loading_state, |state, cx| {
                *state = LoadingState::Loaded(data);
                cx.notify();
            })?;

            Ok::<_, anyhow::Error>(())
        }).detach();
    }
}

Caching Strategies

Result Caching
use std::collections::HashMap;

struct CachedRenderer {
    cache: RefCell<HashMap<String, CachedElement>>,
}

impl CachedRenderer {
    fn render_cached(
        &self,
        key: String,
        render_fn: impl FnOnce() -> AnyElement,
    ) -> AnyElement {
        let mut cache = self.cache.borrow_mut();

        cache.entry(key)
            .or_insert_with(|| CachedElement::new(render_fn()))
            .element
            .clone()
    }

    fn invalidate(&self, key: &str) {
        self.cache.borrow_mut().remove(key);
    }
}

Resources

Performance Targets

Rendering:

  • Target: 60 FPS (16.67ms per frame)
  • Render + Layout: ~10ms
  • Paint: ~6ms
  • Warning: Any frame > 16ms

Memory:

  • Monitor heap growth
  • Warning: Steady increase (leak)
  • Target: Stable after initialization

Startup:

  • Window display: < 100ms
  • Fully interactive: < 500ms

Profiling Tools

CPU Profiling:

  • cargo-flamegraph: Visualize CPU time
  • perf (Linux): System-level profiling
  • Instruments (macOS): Apple's profiler

Memory Profiling:

  • valgrind/massif: Memory usage tracking
  • heaptrack: Heap allocation tracking
  • Instruments: Memory allocations

Benchmarking:

  • criterion: Statistical benchmarking
  • cargo bench: Built-in benchmarks
  • hyperfine: Command-line tool benchmarking

Best Practices

  1. Measure First: Profile before optimizing
  2. Minimize Renders: Only cx.notify() when necessary
  3. Cache Results: Memoize expensive computations
  4. Batch Updates: Group state changes
  5. Virtual Scrolling: For long lists
  6. Flat Layouts: Avoid deep nesting
  7. Fixed Sizing: When possible
  8. Monitor Memory: Watch for leaks
  9. Async Loading: Don't block UI
  10. Test Performance: Include benchmarks

Common Bottlenecks

  • Subscription in render (memory leak)
  • Expensive computation in render
  • Deep component nesting
  • Unnecessary rerenders
  • Layout thrashing
  • Large lists without virtualization
  • Memory leaks from circular refs
  • Unbounded collections

Source: SKILL.md on GitHub

skilld matched fixed text patterns in SKILL.md and file names. Patterns miss obfuscated code.

skilld run checks every file with the same patterns. It asks for approval before it loads a Skill with a behavior marked Needs approval.

1 warning16d4 checks · Risk SAFE
  • Gen Agent Trust Hub16d

    This skill provides comprehensive technical guidance and code patterns for optimizing performance in GPUI-based applications. It includes legitimate advice on rendering, memory management, and profiling using standard developer tools.

  • Socket16d

    No alerts

  • Snyk16d

    Risk: LOW · No issues

  • Runlayer9mo

    1/1 file flagged

Signed by skilld at 0ab5ae1. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub last month.

Dormantupdated last year

README badge

README badge for geoffjay/claude-plugins/gpui-performance