macros and improvements

Cargo.lock

@@ -19,6 +19,7 @@ dependencies = [
  "abi_sys",
  "embedded-graphics",
  "embedded-sdmmc",
+ "main_proc_macro",
  "once_cell",
  "rand_core 0.9.3",
 ]
@@ -350,7 +351,7 @@ checksum = "e37549a379a9e0e6e576fd208ee60394ccb8be963889eebba3ffe0980364f472"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.104",
+ "syn 2.0.108",
 ]
 
 [[package]]
@@ -447,7 +448,7 @@ dependencies = [
  "proc-macro2",
  "quote",
  "strsim 0.11.1",
- "syn 2.0.104",
+ "syn 2.0.108",
 ]
 
 [[package]]
@@ -458,7 +459,7 @@ checksum = "fc34b93ccb385b40dc71c6fceac4b2ad23662c7eeb248cf10d529b7e055b6ead"
 dependencies = [
  "darling_core",
  "quote",
- "syn 2.0.104",
+ "syn 2.0.108",
 ]
 
 [[package]]
@@ -496,7 +497,7 @@ dependencies = [
  "proc-macro-error2",
  "proc-macro2",
  "quote",
- "syn 2.0.104",
+ "syn 2.0.108",
 ]
 
 [[package]]
@@ -645,7 +646,7 @@ dependencies = [
  "darling",
  "proc-macro2",
  "quote",
- "syn 2.0.104",
+ "syn 2.0.108",
 ]
 
 [[package]]
@@ -1019,7 +1020,7 @@ checksum = "4f6e621fe4c7e05b695274b722dc0a60bacd1c8696b58191baa0154713d52400"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.104",
+ "syn 2.0.108",
 ]
 
 [[package]]
@@ -1212,7 +1213,7 @@ checksum = "162ee34ebcb7c64a8abebc059ce0fee27c2262618d7b60ed8faf72fef13c3650"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.104",
+ "syn 2.0.108",
 ]
 
 [[package]]
@@ -1648,6 +1649,14 @@ version = "0.4.27"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "13dc2df351e3202783a1fe0d44375f7295ffb4049267b0f3018346dc122a1d94"
 
+[[package]]
+name = "main_proc_macro"
+version = "0.1.0"
+dependencies = [
+ "quote",
+ "syn 2.0.108",
+]
+
 [[package]]
 name = "memchr"
 version = "2.7.5"
@@ -1728,7 +1737,7 @@ checksum = "77e878c846a8abae00dd069496dbe8751b16ac1c3d6bd2a7283a938e8228f90d"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.104",
+ "syn 2.0.108",
 ]
 
 [[package]]
@@ -1920,7 +1929,7 @@ dependencies = [
  "proc-macro-error2",
  "proc-macro2",
  "quote",
- "syn 2.0.104",
+ "syn 2.0.108",
 ]
 
 [[package]]
@@ -1987,7 +1996,7 @@ dependencies = [
  "proc-macro-error-attr2",
  "proc-macro2",
  "quote",
- "syn 2.0.104",
+ "syn 2.0.108",
 ]
 
 [[package]]
@@ -2001,9 +2010,9 @@ dependencies = [
 
 [[package]]
 name = "quote"
-version = "1.0.40"
+version = "1.0.41"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1885c039570dc00dcb4ff087a89e185fd56bae234ddc7f056a945bf36467248d"
+checksum = "ce25767e7b499d1b604768e7cde645d14cc8584231ea6b295e9c9eb22c02e1d1"
 dependencies = [
  "proc-macro2",
 ]
@@ -2233,7 +2242,7 @@ checksum = "d540f220d3187173da220f885ab66608367b6574e925011a9353e4badda91d79"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.104",
+ "syn 2.0.108",
 ]
 
 [[package]]
@@ -2400,7 +2409,7 @@ dependencies = [
  "heck 0.5.0",
  "proc-macro2",
  "quote",
- "syn 2.0.104",
+ "syn 2.0.108",
 ]
 
 [[package]]
@@ -2429,9 +2438,9 @@ dependencies = [
 
 [[package]]
 name = "syn"
-version = "2.0.104"
+version = "2.0.108"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "17b6f705963418cdb9927482fa304bc562ece2fdd4f616084c50b7023b435a40"
+checksum = "da58917d35242480a05c2897064da0a80589a2a0476c9a3f2fdc83b53502e917"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -2528,7 +2537,7 @@ checksum = "4fee6c4efc90059e10f81e6d42c60a18f76588c3d74cb83a0b242a2b6c7504c1"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.104",
+ "syn 2.0.108",
 ]
 
 [[package]]
@@ -2539,7 +2548,7 @@ checksum = "7f7cf42b4507d8ea322120659672cf1b9dbb93f8f2d4ecfd6e51350ff5b17a1d"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.104",
+ "syn 2.0.108",
 ]
 
 [[package]]
@@ -2607,7 +2616,7 @@ dependencies = [
  "darling",
  "proc-macro2",
  "quote",
- "syn 2.0.104",
+ "syn 2.0.108",
  "uuid",
 ]
 
@@ -2766,7 +2775,7 @@ dependencies = [
  "log",
  "proc-macro2",
  "quote",
- "syn 2.0.104",
+ "syn 2.0.108",
  "wasm-bindgen-shared",
 ]
 
@@ -2788,7 +2797,7 @@ checksum = "8ae87ea40c9f689fc23f209965b6fb8a99ad69aeeb0231408be24920604395de"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.104",
+ "syn 2.0.108",
  "wasm-bindgen-backend",
  "wasm-bindgen-shared",
 ]
@@ -2938,5 +2947,5 @@ checksum = "9ecf5b4cc5364572d7f4c329661bcc82724222973f2cab6f050a4e5c22f75181"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.104",
+ "syn 2.0.108",
 ]

Cargo.toml

@@ -8,6 +8,7 @@ members = [
   "user-apps/snake",
   "user-apps/gallery",
   "user-apps/gif",
+  "main_proc_macro",
 ]
 
 [profile.release]

abi/Cargo.toml

@@ -4,8 +4,9 @@ version = "0.1.0"
 edition = "2024"
 
 [dependencies]
-embedded-sdmmc = { version = "0.9.0", default-features = false }
-embedded-graphics = "0.8.1"
 abi_sys = { path = "../abi_sys" }
+main_proc_macro = { path = "../main_proc_macro" }
+embedded-sdmmc = { version = "0.9", default-features = false }
+embedded-graphics = "0.8.1"
 once_cell = { version = "1", default-features = false }
 rand_core = "0.9.3"

abi/src/lib.rs

@@ -5,10 +5,20 @@ extern crate alloc;
 
 pub use abi_sys::{self, keyboard};
 use abi_sys::{RngRequest, alloc, dealloc, keyboard::KeyEvent};
-pub use alloc::format;
-use core::alloc::{GlobalAlloc, Layout};
+use alloc::format;
+use core::{
+    alloc::{GlobalAlloc, Layout},
+    panic::PanicInfo,
+};
+pub use main_proc_macro::main;
 use rand_core::RngCore;
 
+#[panic_handler]
+fn panic(info: &PanicInfo) -> ! {
+    print!("user panic: {} @ {:?}", info.message(), info.location(),);
+    loop {}
+}
+
 #[global_allocator]
 static ALLOC: Alloc = Alloc;
 
@@ -27,7 +37,7 @@ unsafe impl GlobalAlloc for Alloc {
 #[macro_export]
 macro_rules! print {
     ($($arg:tt)*) => {{
-        let s = $crate::format!($($arg)*);
+        let s = format!($($arg)*);
         $crate::abi_sys::print(s.as_ptr(), s.len());
     }};
 }
@@ -46,7 +56,6 @@ pub fn get_key() -> KeyEvent {
 
 pub mod display {
     use abi_sys::CPixel;
-    use alloc::{vec, vec::Vec};
     use embedded_graphics::{
         Pixel,
         geometry::{Dimensions, Point},
@@ -54,7 +63,6 @@ pub mod display {
         prelude::{DrawTarget, Size},
         primitives::Rectangle,
     };
-    use once_cell::unsync::Lazy;
 
     pub const SCREEN_WIDTH: usize = 320;
     pub const SCREEN_HEIGHT: usize = 320;

kernel/Cargo.toml

@@ -11,7 +11,7 @@ doctest = false
 bench = false
 
 [features]
-default = ["rp235x", "defmt", "pimoroni2w"]
+default = ["rp235x", "defmt"]
 pimoroni2w = ["rp235x"]
 rp2040 = ["embassy-rp/rp2040"]
 rp235x = ["embassy-rp/rp235xb"]

kernel/src/abi.rs

@@ -81,7 +81,7 @@ pub extern "C" fn draw_iter(cpixels: *const CPixel, len: usize) {
     let iter = cpixels.iter().copied().map(|c: CPixel| c.into());
 
     FB_PAUSED.store(true, Ordering::Release);
-    unsafe { FRAMEBUFFER.draw_iter(iter).unwrap() }
+    unsafe { FRAMEBUFFER.as_mut().unwrap().draw_iter(iter).unwrap() }
     FB_PAUSED.store(false, Ordering::Release);
 }
 

kernel/src/display.rs

@@ -1,6 +1,6 @@
 use crate::framebuffer::{self, AtomicFrameBuffer, FB_PAUSED};
-use alloc::vec;
-use core::sync::atomic::{AtomicBool, Ordering};
+use core::alloc::{GlobalAlloc, Layout};
+use core::sync::atomic::Ordering;
 use embassy_rp::{
     Peri,
     gpio::{Level, Output},
@@ -8,8 +8,11 @@ use embassy_rp::{
     spi::{Async, Spi},
 };
 use embassy_time::{Delay, Timer};
+use embedded_graphics::{
+    pixelcolor::Rgb565,
+    prelude::{DrawTarget, RgbColor},
+};
 use embedded_hal_bus::spi::ExclusiveDevice;
-use once_cell::unsync::Lazy;
 use st7365p_lcd::ST7365P;
 
 type DISPLAY = ST7365P<
@@ -22,10 +25,24 @@ type DISPLAY = ST7365P<
 pub const SCREEN_WIDTH: usize = 320;
 pub const SCREEN_HEIGHT: usize = 320;
 
-pub static mut FRAMEBUFFER: Lazy<AtomicFrameBuffer> = Lazy::new(|| {
-    static mut BUF: [u16; framebuffer::SIZE] = [0; framebuffer::SIZE];
-    AtomicFrameBuffer::new(unsafe { &mut BUF })
-});
+pub static mut FRAMEBUFFER: Option<AtomicFrameBuffer> = None;
+
+fn init_fb() {
+    unsafe {
+        FRAMEBUFFER = Some(if cfg!(not(feature = "pimoroni2w")) {
+            static mut BUF: [u16; framebuffer::SIZE] = [0; framebuffer::SIZE];
+            AtomicFrameBuffer::new(&mut BUF)
+        } else {
+            let slab = crate::heap::HEAP.alloc(Layout::array::<u16>(framebuffer::SIZE).unwrap())
+                as *mut u16;
+            let buf = core::slice::from_raw_parts_mut(slab, framebuffer::SIZE);
+
+            let mut fb = AtomicFrameBuffer::new(buf);
+            fb.clear(Rgb565::BLACK).unwrap();
+            fb
+        });
+    }
+}
 
 pub async fn init_display(
     spi: Spi<'static, SPI1, Async>,
@@ -33,6 +50,8 @@ pub async fn init_display(
     data: Peri<'static, PIN_14>,
     reset: Peri<'static, PIN_15>,
 ) -> DISPLAY {
+    init_fb();
+
     let spi_device = ExclusiveDevice::new(spi, Output::new(cs, Level::Low), Delay).unwrap();
     let mut display = ST7365P::new(
         spi_device,
@@ -44,7 +63,14 @@ pub async fn init_display(
     );
     display.init().await.unwrap();
     display.set_custom_orientation(0x40).await.unwrap();
-    unsafe { FRAMEBUFFER.draw(&mut display).await.unwrap() }
+    unsafe {
+        FRAMEBUFFER
+            .as_mut()
+            .unwrap()
+            .draw(&mut display)
+            .await
+            .unwrap()
+    }
     display.set_on().await.unwrap();
 
     display
@@ -54,7 +80,14 @@ pub async fn init_display(
 pub async fn display_handler(mut display: DISPLAY) {
     loop {
         if !FB_PAUSED.load(Ordering::Acquire) {
-            unsafe { FRAMEBUFFER.safe_draw(&mut display).await.unwrap() };
+            unsafe {
+                FRAMEBUFFER
+                    .as_mut()
+                    .unwrap()
+                    .partial_draw(&mut display)
+                    .await
+                    .unwrap()
+            };
         }
 
         // small yield to allow other tasks to run

kernel/src/framebuffer.rs

@@ -1,6 +1,5 @@
 use crate::display::{SCREEN_HEIGHT, SCREEN_WIDTH};
 use core::sync::atomic::{AtomicBool, Ordering};
-use embassy_sync::lazy_lock::LazyLock;
 use embedded_graphics::{
     draw_target::DrawTarget,
     pixelcolor::{
@@ -14,32 +13,33 @@ use embedded_hal_2::digital::OutputPin;
 use embedded_hal_async::{delay::DelayNs, spi::SpiDevice};
 use st7365p_lcd::ST7365P;
 
-pub const TILE_SIZE: usize = 16; // 16x16 tile
-pub const TILE_COUNT: usize = (SCREEN_WIDTH / TILE_SIZE) * (SCREEN_HEIGHT / TILE_SIZE); // 400 tiles
+const TILE_SIZE: usize = 16; // 16x16 tile
+const TILE_COUNT: usize = (SCREEN_WIDTH / TILE_SIZE) * (SCREEN_HEIGHT / TILE_SIZE); // 400 tiles
+const NUM_TILE_ROWS: usize = SCREEN_WIDTH / TILE_SIZE;
+const NUM_TILE_COLS: usize = SCREEN_WIDTH / TILE_SIZE;
 
-// Group of tiles for batching
-pub const MAX_META_TILES: usize = (SCREEN_WIDTH / TILE_SIZE) * 2; // max number of meta tiles in buffer
-type MetaTileVec = heapless::Vec<Rectangle, { TILE_COUNT / MAX_META_TILES }>;
+const MAX_BATCH_TILES: usize = (SCREEN_WIDTH / TILE_SIZE) * 2;
+type BatchTileBuf = [u16; MAX_BATCH_TILES * TILE_SIZE * TILE_SIZE];
 
 pub const SIZE: usize = SCREEN_HEIGHT * SCREEN_WIDTH;
 
 pub static FB_PAUSED: AtomicBool = AtomicBool::new(false);
 
-static mut DIRTY_TILES: LazyLock<heapless::Vec<AtomicBool, TILE_COUNT>> = LazyLock::new(|| {
-    let mut tiles = heapless::Vec::new();
-    for _ in 0..TILE_COUNT {
-        tiles.push(AtomicBool::new(true)).unwrap();
-    }
-    tiles
-});
-
 #[allow(dead_code)]
-pub struct AtomicFrameBuffer<'a>(&'a mut [u16]);
+pub struct AtomicFrameBuffer<'a> {
+    fb: &'a mut [u16],
+    dirty_tiles: [AtomicBool; TILE_COUNT],
+    batch_tile_buf: BatchTileBuf,
+}
 
 impl<'a> AtomicFrameBuffer<'a> {
     pub fn new(buffer: &'a mut [u16]) -> Self {
         assert!(buffer.len() == SIZE);
-        Self(buffer)
+        Self {
+            fb: buffer,
+            dirty_tiles: core::array::from_fn(|_| AtomicBool::new(true)),
+            batch_tile_buf: [0; MAX_BATCH_TILES * TILE_SIZE * TILE_SIZE],
+        }
     }
 
     fn mark_tiles_dirty(&mut self, rect: Rectangle) {
@@ -52,7 +52,7 @@ impl<'a> AtomicFrameBuffer<'a> {
         for ty in start_ty..=end_ty {
             for tx in start_tx..=end_tx {
                 let tile_idx = ty * tiles_x + tx;
-                unsafe { DIRTY_TILES.get_mut()[tile_idx].store(true, Ordering::Release) };
+                self.dirty_tiles[tile_idx].store(true, Ordering::Release);
             }
         }
     }
@@ -78,7 +78,7 @@ impl<'a> AtomicFrameBuffer<'a> {
         for y in sy..=ey {
             for x in sx..=ex {
                 if let Some(color) = color_iter.next() {
-                    self.0[(y as usize * SCREEN_WIDTH) + x as usize] = color;
+                    self.fb[(y as usize * SCREEN_WIDTH) + x as usize] = color;
                 } else {
                     return Err(()); // Not enough data
                 }
@@ -93,60 +93,17 @@ impl<'a> AtomicFrameBuffer<'a> {
         Ok(())
     }
 
-    // walk the dirty tiles and mark groups of tiles(meta-tiles) for batched updates
-    fn find_meta_tiles(&mut self, tiles_x: usize, tiles_y: usize) -> MetaTileVec {
-        let mut meta_tiles: MetaTileVec = heapless::Vec::new();
+    // Checks if a full draw would be faster than individual tile batches
+    fn should_full_draw(&self) -> bool {
+        let threshold_pixels = SIZE * 80 / 100;
+        let mut dirty_pixels = 0;
 
-        for ty in 0..tiles_y {
-            let mut tx = 0;
-            while tx < tiles_x {
-                let idx = ty * tiles_x + tx;
-                if !unsafe { DIRTY_TILES.get()[idx].load(Ordering::Acquire) } {
-                    tx += 1;
-                    continue;
-                }
-
-                // Start meta-tile at this tile
-                let mut width_tiles = 1;
-                let height_tiles = 1;
-
-                // Grow horizontally, but keep under MAX_TILES_PER_METATILE
-                while tx + width_tiles < tiles_x
-                    && unsafe {
-                        DIRTY_TILES.get()[ty * tiles_x + tx + width_tiles].load(Ordering::Acquire)
-                    }
-                    && (width_tiles + height_tiles) <= MAX_META_TILES
-                {
-                    width_tiles += 1;
-                }
-
-                // TODO: for simplicity, skipped vertical growth
-
-                for x_off in 0..width_tiles {
-                    unsafe {
-                        DIRTY_TILES.get()[ty * tiles_x + tx + x_off]
-                            .store(false, Ordering::Release);
-                    };
-                }
-
-                // new meta-tile pos
-                let rect = Rectangle::new(
-                    Point::new((tx * TILE_SIZE) as i32, (ty * TILE_SIZE) as i32),
-                    Size::new(
-                        (width_tiles * TILE_SIZE) as u32,
-                        (height_tiles * TILE_SIZE) as u32,
-                    ),
-                );
-
-                if meta_tiles.push(rect).is_err() {
-                    return meta_tiles;
-                };
-
-                tx += width_tiles;
+        self.dirty_tiles.iter().any(|tile| {
+            if tile.load(Ordering::Acquire) {
+                dirty_pixels += TILE_SIZE * TILE_SIZE;
             }
-        }
-
-        meta_tiles
+            dirty_pixels >= threshold_pixels
+        })
     }
 
     /// Sends the entire framebuffer to the display
@@ -165,71 +122,40 @@ impl<'a> AtomicFrameBuffer<'a> {
                 0,
                 self.size().width as u16 - 1,
                 self.size().height as u16 - 1,
-                &self.0[..],
+                &self.fb[..],
             )
             .await?;
 
-        unsafe {
-            for tile in DIRTY_TILES.get_mut().iter() {
-                tile.store(false, Ordering::Release);
-            }
-        };
-
-        Ok(())
-    }
-
-    pub async fn safe_draw<SPI, DC, RST, DELAY>(
-        &mut self,
-        display: &mut ST7365P<SPI, DC, RST, DELAY>,
-    ) -> Result<(), ()>
-    where
-        SPI: SpiDevice,
-        DC: OutputPin,
-        RST: OutputPin,
-        DELAY: DelayNs,
-    {
-        let tiles_x = SCREEN_WIDTH / TILE_SIZE;
-        let _tiles_y = SCREEN_HEIGHT / TILE_SIZE;
-
-        let tiles = unsafe { DIRTY_TILES.get_mut() };
-        let mut pixel_buffer: heapless::Vec<u16, { TILE_SIZE * TILE_SIZE }> = heapless::Vec::new();
-
-        for tile_idx in 0..TILE_COUNT {
-            if tiles[tile_idx].swap(false, Ordering::AcqRel) {
-                let tx = tile_idx % tiles_x;
-                let ty = tile_idx / tiles_x;
-
-                let x_start = tx * TILE_SIZE;
-                let y_start = ty * TILE_SIZE;
-
-                let x_end = (x_start + TILE_SIZE).min(SCREEN_WIDTH);
-                let y_end = (y_start + TILE_SIZE).min(SCREEN_HEIGHT);
-
-                pixel_buffer.clear();
-
-                for y in y_start..y_end {
-                    let start = y * SCREEN_WIDTH + x_start;
-                    let end = y * SCREEN_WIDTH + x_end;
-                    pixel_buffer.extend_from_slice(&self.0[start..end]).unwrap();
-                }
-
-                display
-                    .set_pixels_buffered(
-                        x_start as u16,
-                        y_start as u16,
-                        (x_end - 1) as u16,
-                        (y_end - 1) as u16,
-                        &pixel_buffer,
-                    )
-                    .await
-                    .unwrap();
-            }
+        for tile in self.dirty_tiles.iter() {
+            tile.store(false, Ordering::Release);
         }
 
         Ok(())
     }
 
-    /// Sends only dirty tiles (16x16px) in batches to the display
+    // copy N tiles horizontally to the right into batch tile buf
+    fn append_tiles_to_batch(
+        &mut self,
+        tile_x: u16,
+        tile_y: u16,
+        total_tiles: u16, // number of tiles being written to buf
+    ) {
+        debug_assert!(total_tiles as usize <= NUM_TILE_COLS);
+        for batch_row_num in 0..TILE_SIZE {
+            let batch_row_offset = batch_row_num * total_tiles as usize * TILE_SIZE;
+            let batch_row = &mut self.batch_tile_buf
+                [batch_row_offset..batch_row_offset + (total_tiles as usize * TILE_SIZE)];
+
+            let fb_row_offset = (tile_y as usize * TILE_SIZE + batch_row_num) * SCREEN_WIDTH
+                + tile_x as usize * TILE_SIZE;
+            let fb_row =
+                &self.fb[fb_row_offset..fb_row_offset + (total_tiles as usize * TILE_SIZE)];
+
+            batch_row.copy_from_slice(fb_row);
+        }
+    }
+
+    // Pushes tiles to the display in batches to avoid full frame pushes (unless needed)
     pub async fn partial_draw<SPI, DC, RST, DELAY>(
         &mut self,
         display: &mut ST7365P<SPI, DC, RST, DELAY>,
@@ -240,55 +166,53 @@ impl<'a> AtomicFrameBuffer<'a> {
         RST: OutputPin,
         DELAY: DelayNs,
     {
-        if unsafe { DIRTY_TILES.get().iter().any(|p| p.load(Ordering::Acquire)) } {
-            let tiles_x = (SCREEN_WIDTH + TILE_SIZE - 1) / TILE_SIZE;
-            let tiles_y = (SCREEN_HEIGHT + TILE_SIZE - 1) / TILE_SIZE;
+        if self.should_full_draw() {
+            return self.draw(display).await;
+        }
 
-            let meta_tiles = self.find_meta_tiles(tiles_x, tiles_y);
+        for tile_row in 0..NUM_TILE_ROWS {
+            let row_start_idx = tile_row * NUM_TILE_COLS;
+            let mut col = 0;
 
-            // buffer for copying meta tiles before sending to display
-            let mut pixel_buffer: heapless::Vec<u16, { MAX_META_TILES * TILE_SIZE * TILE_SIZE }> =
-                heapless::Vec::new();
+            while col < NUM_TILE_COLS {
+                // Check for dirty tile
+                if self.dirty_tiles[row_start_idx + col].swap(false, Ordering::Acquire) {
+                    let run_start = col;
+                    let mut run_len = 1;
 
-            for rect in meta_tiles {
-                let rect_width = rect.size.width as usize;
-                let rect_height = rect.size.height as usize;
-                let rect_x = rect.top_left.x as usize;
-                let rect_y = rect.top_left.y as usize;
-
-                pixel_buffer.clear();
-
-                for row in 0..rect_height {
-                    let y = rect_y + row;
-                    let start = y * SCREEN_WIDTH + rect_x;
-                    let end = start + rect_width;
-
-                    // Safe: we guarantee buffer will not exceed MAX_META_TILE_PIXELS
-                    pixel_buffer.extend_from_slice(&self.0[start..end]).unwrap();
-                }
-
-                display
-                    .set_pixels_buffered(
-                        rect_x as u16,
-                        rect_y as u16,
-                        (rect_x + rect_width - 1) as u16,
-                        (rect_y + rect_height - 1) as u16,
-                        &pixel_buffer,
-                    )
-                    .await?;
-
-                // walk the meta-tile and set as clean
-                let start_tx = rect_x / TILE_SIZE;
-                let start_ty = rect_y / TILE_SIZE;
-                let end_tx = (rect_x + rect_width - 1) / TILE_SIZE;
-                let end_ty = (rect_y + rect_height - 1) / TILE_SIZE;
-
-                for ty in start_ty..=end_ty {
-                    for tx in start_tx..=end_tx {
-                        let tile_idx = ty * tiles_x + tx;
-                        unsafe { DIRTY_TILES.get_mut()[tile_idx].store(false, Ordering::Release) };
+                    // Extend run while contiguous dirty tiles and within MAX_BATCH_TILES
+                    while col + 1 < NUM_TILE_COLS
+                        && self.dirty_tiles[row_start_idx + col + 1].load(Ordering::Acquire)
+                        && run_len < MAX_BATCH_TILES
+                    {
+                        col += 1;
+                        run_len += 1;
                     }
+
+                    // Copy the whole horizontal run into the batch buffer in one call
+                    let tile_x = run_start;
+                    let tile_y = tile_row;
+                    self.append_tiles_to_batch(tile_x as u16, tile_y as u16, run_len as u16);
+
+                    // Compute coordinates for display write
+                    let start_x = tile_x * TILE_SIZE;
+                    let end_x = start_x + run_len * TILE_SIZE - 1;
+                    let start_y = tile_y * TILE_SIZE;
+                    let end_y = start_y + TILE_SIZE - 1;
+
+                    // Send batch to display
+                    display
+                        .set_pixels_buffered(
+                            start_x as u16,
+                            start_y as u16,
+                            end_x as u16,
+                            end_y as u16,
+                            &self.batch_tile_buf[..run_len * TILE_SIZE * TILE_SIZE],
+                        )
+                        .await?;
                 }
+
+                col += 1;
             }
         }
 
@@ -309,14 +233,14 @@ impl<'a> DrawTarget for AtomicFrameBuffer<'a> {
 
         for Pixel(coord, color) in pixels {
             if coord.x >= 0 && coord.y >= 0 {
-                let x = coord.x as i32;
-                let y = coord.y as i32;
+                let x = coord.x;
+                let y = coord.y;
 
                 if (x as usize) < SCREEN_WIDTH && (y as usize) < SCREEN_HEIGHT {
                     let idx = (y as usize) * SCREEN_WIDTH + (x as usize);
                     let raw_color = RawU16::from(color).into_inner();
-                    if self.0[idx] != raw_color {
-                        self.0[idx] = raw_color;
+                    if self.fb[idx] != raw_color {
+                        self.fb[idx] = raw_color;
                         changed = true;
                     }
 
@@ -365,8 +289,8 @@ impl<'a> DrawTarget for AtomicFrameBuffer<'a> {
                         if let Some(color) = colors.next() {
                             let idx = (p.y as usize * SCREEN_WIDTH) + (p.x as usize);
                             let raw_color = RawU16::from(color).into_inner();
-                            if self.0[idx] != raw_color {
-                                self.0[idx] = raw_color;
+                            if self.fb[idx] != raw_color {
+                                self.fb[idx] = raw_color;
                                 changed = true;
                             }
                         } else {
@@ -404,7 +328,7 @@ impl<'a> DrawTarget for AtomicFrameBuffer<'a> {
                 .take((self.size().width * self.size().height) as usize),
         )?;
 
-        for tile in unsafe { DIRTY_TILES.get_mut() }.iter() {
+        for tile in self.dirty_tiles.iter() {
             tile.store(true, Ordering::Release);
         }
 

kernel/src/main.rs

@@ -12,8 +12,10 @@ mod abi;
 mod display;
 mod elf;
 mod framebuffer;
+#[allow(unused)]
 mod heap;
 mod peripherals;
+#[allow(unused)]
 mod psram;
 mod scsi;
 mod storage;
@@ -80,7 +82,10 @@ static mut CORE1_STACK: Stack<16384> = Stack::new();
 static EXECUTOR0: StaticCell<Executor> = StaticCell::new();
 static EXECUTOR1: StaticCell<Executor> = StaticCell::new();
 
+#[cfg(not(feature = "pimoroni2w"))]
 static mut ARENA: [u8; 250 * 1024] = [0; 250 * 1024];
+#[cfg(feature = "pimoroni2w")]
+static mut ARENA: [u8; 400 * 1024] = [0; 400 * 1024];
 
 #[global_allocator]
 static ALLOCATOR: Talck<spin::Mutex<()>, ClaimOnOom> =
@@ -90,11 +95,12 @@ static ALLOCATOR: Talck<spin::Mutex<()>, ClaimOnOom> =
 #[embassy_executor::task]
 async fn watchdog_task(mut watchdog: Watchdog) {
     if let Some(reason) = watchdog.reset_reason() {
-        let reason = match reason {
+        let _reason = match reason {
             ResetReason::Forced => "forced",
             ResetReason::TimedOut => "timed out",
         };
-        defmt::error!("Watchdog reset reason: {}", reason);
+        #[cfg(feature = "debug")]
+        defmt::error!("Watchdog reset reason: {}", _reason);
     }
 
     watchdog.start(Duration::from_secs(3));
@@ -194,7 +200,7 @@ async fn userland_task() {
         {
             ENABLE_UI.store(true, Ordering::Release);
             UI_CHANGE.signal(());
-            unsafe { FRAMEBUFFER.clear(Rgb565::BLACK).unwrap() };
+            unsafe { FRAMEBUFFER.as_mut().unwrap().clear(Rgb565::BLACK).unwrap() };
 
             let mut selections = SELECTIONS.lock().await;
             selections.set_changed(true);
@@ -221,6 +227,7 @@ struct Sd {
     cs: Peri<'static, PIN_17>,
     det: Peri<'static, PIN_22>,
 }
+#[allow(dead_code)]
 struct Psram {
     pio: Peri<'static, PIO0>,
     sclk: Peri<'static, PIN_21>,
@@ -279,20 +286,18 @@ async fn setup_psram(psram: Psram) {
 
 #[cfg(feature = "pimoroni2w")]
 async fn setup_qmi_psram() {
-    let mut tries = 5;
     Timer::after_millis(250).await;
-    while tries > 1 {
-        let psram_qmi_size = init_psram_qmi(&embassy_rp::pac::QMI, &embassy_rp::pac::XIP_CTRL);
-        defmt::info!("size:  {}", psram_qmi_size);
-        Timer::after_millis(100).await;
-        if psram_qmi_size > 0 {
-            init_qmi_psram_heap(psram_qmi_size);
-            return;
-        }
-        defmt::info!("failed to init qmi psram... trying again");
-        tries -= 1;
+    let psram_qmi_size = init_psram_qmi(&embassy_rp::pac::QMI, &embassy_rp::pac::XIP_CTRL);
+    #[cfg(feature = "debug")]
+    defmt::info!("size:  {}", psram_qmi_size);
+    Timer::after_millis(100).await;
+
+    if psram_qmi_size > 0 {
+        init_qmi_psram_heap(psram_qmi_size);
+        return;
+    } else {
+        panic!("qmi psram not initialized");
     }
-    panic!("qmi psram not initialized");
 }
 
 async fn setup_sd(sd: Sd) {
@@ -316,24 +321,30 @@ async fn kernel_task(
     watchdog: Peri<'static, WATCHDOG>,
     display: Display,
     sd: Sd,
-    psram: Psram,
+    _psram: Psram,
     mcu: Mcu,
     usb: Peri<'static, USB>,
 ) {
     spawner
         .spawn(watchdog_task(Watchdog::new(watchdog)))
         .unwrap();
+
     setup_mcu(mcu).await;
+
+    defmt::info!("setting up psram");
+    Timer::after_millis(100).await;
+
+    // setup_psram(psram).await;
+    #[cfg(feature = "pimoroni2w")]
+    setup_qmi_psram().await;
+
+    Timer::after_millis(100).await;
     setup_display(display, spawner).await;
     setup_sd(sd).await;
 
     let _usb = embassy_rp_usb::Driver::new(usb, Irqs);
     // spawner.spawn(usb_handler(usb)).unwrap();
 
-    // setup_psram(psram).await;
-    #[cfg(feature = "pimoroni2w")]
-    setup_qmi_psram().await;
-
     loop {
         let ui_enabled = ENABLE_UI.load(Ordering::Relaxed);
         if ui_enabled {

kernel/src/psram.rs

@@ -526,6 +526,7 @@ pub fn init_psram_qmi(
     let psram_size = detect_psram_qmi(qmi);
 
     if psram_size == 0 {
+        #[cfg(feature = "debug")]
         defmt::error!("qmi psram size 0");
         return 0;
     }
@@ -548,13 +549,13 @@ pub fn init_psram_qmi(
     let min_deselect: u32 = ((18 * 1_000_000 + (clock_period_fs - 1)) / clock_period_fs
         - u64::from(divisor + 1) / 2) as u32;
 
-    #[cfg(feature = "defmt")]
-    defmt::info!(
-        "clock_period_fs={} max_select={} min_deselect={}",
-        clock_period_fs,
-        max_select,
-        min_deselect
-    );
+    // #[cfg(feature = "defmt")]
+    // defmt::info!(
+    //     "clock_period_fs={} max_select={} min_deselect={}",
+    //     clock_period_fs,
+    //     max_select,
+    //     min_deselect
+    // );
 
     qmi.direct_csr().write(|w| {
         w.set_clkdiv(10);

kernel/src/ui.rs

@@ -68,7 +68,7 @@ pub async fn clear_selection() {
     if let Some(area) = sel.last_bounds {
         Rectangle::new(area.top_left, area.size)
             .into_styled(PrimitiveStyle::with_fill(Rgb565::BLACK))
-            .draw(unsafe { &mut *FRAMEBUFFER })
+            .draw(unsafe { &mut *FRAMEBUFFER.as_mut().unwrap() })
             .unwrap();
     }
 }
@@ -78,7 +78,7 @@ async fn draw_selection() {
     let file_names = &guard.selections.clone();
 
     let text_style = MonoTextStyle::new(&FONT_10X20, Rgb565::WHITE);
-    let display_area = unsafe { FRAMEBUFFER.bounding_box() };
+    let display_area = unsafe { FRAMEBUFFER.as_mut().unwrap().bounding_box() };
 
     const NO_BINS: &str = "No Programs found on SD Card. Ensure programs end with '.bin', and are located in the root directory";
     let no_bins = String::from_str(NO_BINS).unwrap();
@@ -94,7 +94,7 @@ async fn draw_selection() {
             ),
             text_style,
         )
-        .draw(unsafe { &mut *FRAMEBUFFER })
+        .draw(unsafe { &mut *FRAMEBUFFER.as_mut().unwrap() })
         .unwrap();
     } else {
         let mut views: alloc::vec::Vec<Text<MonoTextStyle<Rgb565>>> = Vec::new();
@@ -119,12 +119,14 @@ async fn draw_selection() {
             .bounding_box();
         Rectangle::new(selected_bounds.top_left, selected_bounds.size)
             .into_styled(PrimitiveStyle::with_stroke(Rgb565::WHITE, 1))
-            .draw(unsafe { &mut *FRAMEBUFFER })
+            .draw(unsafe { &mut *FRAMEBUFFER.as_mut().unwrap() })
             .unwrap();
 
         guard.last_bounds = Some(layout.bounds());
 
-        layout.draw(unsafe { &mut *FRAMEBUFFER }).unwrap();
+        layout
+            .draw(unsafe { &mut *FRAMEBUFFER.as_mut().unwrap() })
+            .unwrap();
     }
 
     guard.changed = false;

main_proc_macro/Cargo.toml

@@ -0,0 +1,11 @@
+[package]
+name = "main_proc_macro"
+version = "0.1.0"
+edition = "2024"
+
+[lib]
+proc-macro = true
+
+[dependencies]
+quote = "1.0.41"
+syn = "2.0.108"

main_proc_macro/src/lib.rs

@@ -0,0 +1,21 @@
+use proc_macro::TokenStream;
+use quote::quote;
+use syn::{ItemFn, parse_macro_input};
+
+#[proc_macro_attribute]
+pub fn main(_attr: TokenStream, item: TokenStream) -> TokenStream {
+    let input = parse_macro_input!(item as ItemFn);
+    let name = &input.sig.ident;
+
+    // ensure we emit _start in the same module as the fn
+    let expanded = quote! {
+        #input
+
+        #[unsafe(no_mangle)]
+        pub extern "Rust" fn _start() {
+            #name();
+        }
+    };
+
+    expanded.into()
+}

user-apps/calculator/src/main.rs

@@ -9,7 +9,6 @@ use abi::{
     print,
 };
 use alloc::{format, string::String, vec, vec::Vec};
-use core::panic::PanicInfo;
 use embedded_graphics::{
     Drawable,
     geometry::{Dimensions, Point},
@@ -26,12 +25,6 @@ use embedded_layout::{
     prelude::Chain,
 };
 
-#[panic_handler]
-fn panic(info: &PanicInfo) -> ! {
-    print!("user panic: {} @ {:?}", info.message(), info.location(),);
-    loop {}
-}
-
 #[unsafe(no_mangle)]
 pub extern "Rust" fn _start() {
     main()

user-apps/gallery/src/main.rs

@@ -8,27 +8,16 @@ use abi::{
     fs::{list_dir, read_file},
     get_key,
     keyboard::{KeyCode, KeyState},
-    print,
+    main, print,
 };
 use alloc::{format, string::ToString, vec};
-use core::panic::PanicInfo;
 use embedded_graphics::{
     Drawable, image::Image, mono_font::MonoTextStyle, mono_font::ascii::FONT_6X10,
     pixelcolor::Rgb565, prelude::*, text::Text,
 };
 use tinybmp::Bmp;
 
-#[panic_handler]
-fn panic(info: &PanicInfo) -> ! {
-    print!("user panic: {} @ {:?}", info.message(), info.location());
-    loop {}
-}
-
-#[unsafe(no_mangle)]
-pub extern "Rust" fn _start() {
-    main()
-}
-
+#[main]
 pub fn main() {
     print!("Starting Gallery app");
     let mut bmp_buf = vec![0_u8; 100_000];

user-apps/gif/src/main.rs

@@ -7,26 +7,15 @@ use abi::{
     fs::{file_len, read_file},
     get_key, get_ms,
     keyboard::{KeyCode, KeyState},
-    print, sleep,
+    main, print, sleep,
 };
-use alloc::vec;
-use core::panic::PanicInfo;
+use alloc::{format, vec};
 use embedded_graphics::{
     image::ImageDrawable, pixelcolor::Rgb565, prelude::Point, transform::Transform,
 };
 use tinygif::Gif;
 
-#[panic_handler]
-fn panic(info: &PanicInfo) -> ! {
-    print!("user panic: {} @ {:?}", info.message(), info.location(),);
-    loop {}
-}
-
-#[unsafe(no_mangle)]
-pub extern "Rust" fn _start() {
-    main()
-}
-
+#[main]
 pub fn main() {
     print!("Starting Gif app");
     let mut display = Display;
@@ -51,11 +40,14 @@ pub fn main() {
                 .unwrap();
             frame_num += 1;
 
-            if frame_num % 100 == 0 {
+            if frame_num % 5 == 0 {
                 let event = get_key();
                 if event.state != KeyState::Idle {
                     match event.key {
-                        KeyCode::Esc => return,
+                        KeyCode::Esc => {
+                            drop(buf);
+                            return;
+                        }
                         _ => (),
                     };
                 };

user-apps/snake/src/main.rs

@@ -9,16 +9,10 @@ use abi::{
     keyboard::{KeyCode, KeyState},
     print, sleep,
 };
-use core::panic::PanicInfo;
+use alloc::format;
 use embedded_graphics::{pixelcolor::Rgb565, prelude::RgbColor};
 use embedded_snake::{Direction, SnakeGame};
 
-#[panic_handler]
-fn panic(info: &PanicInfo) -> ! {
-    print!("user panic: {} @ {:?}", info.message(), info.location(),);
-    loop {}
-}
-
 #[unsafe(no_mangle)]
 pub extern "Rust" fn _start() {
     main()