Mode3::clear_to is a very fast and easily-proven-correct clear routine.

examples/paddle_ball.rs

@@ -1,24 +1,15 @@
-#![allow(unused_imports)]
-/*
- * Made by Evan Goemer
- * Discord: @evangoemer
- */
 #![no_std]
 #![no_main]
 
-use core::ptr::addr_of;
+//! Made by Evan Goemer, Discord: @evangoemer
 
 use gba::{
   asm_runtime::USER_IRQ_HANDLER,
   bios::VBlankIntrWait,
-  dma::{DmaControl, DmaSrcAddr},
   gba_cell::GbaCell,
   mem::bulk_memory_set,
-  mmio::{
-    DISPCNT, DISPSTAT, DMA3_CONTROL, DMA3_DESTINATION, DMA3_SOURCE,
-    DMA3_TRANSFER_COUNT, IE, IME, KEYINPUT, MODE3_VRAM, OBJ_PALRAM,
-  },
-  video::{Color, DisplayControl, DisplayStatus},
+  mmio::{DISPCNT, DISPSTAT, IE, IME, KEYINPUT, MODE3_VRAM},
+  video::{Color, DisplayControl, DisplayStatus, Mode3},
   IrqBits, KeyInput,
 };
 
@@ -151,29 +142,7 @@ fn main() -> ! {
 }
 
 extern "C" fn draw_sprites(_bits: IrqBits) {
-  /*
-  unsafe {
-    let color = OBJ_PALRAM.index(0).read();
-    let c: u32 = color.0 as u32 | (color.0 as u32) << 16;
-    DMA3_SOURCE.write(addr_of!(c).cast());
-    DMA3_DESTINATION.write(MODE3_VRAM.as_usize() as _);
-    DMA3_TRANSFER_COUNT.write(SCREEN_WIDTH * SCREEN_HEIGHT / 2);
-    DMA3_CONTROL.write(
-      DmaControl::new()
-        .with_src_addr(DmaSrcAddr::Fixed)
-        .with_u32_transfer(true)
-        .with_enabled(true),
-    );
-  }
-  // */
-  unsafe {
-    let dest = MODE3_VRAM.as_usize() as *mut u32;
-    let byte_count = SCREEN_WIDTH as usize
-      * SCREEN_HEIGHT as usize
-      * core::mem::size_of::<Color>();
-    let r2 = 0;
-    bulk_memory_set(dest, byte_count, r2);
-  }
+  Mode3.clear_to(Color::BLACK);
 
   draw_rect(
     SPRITE_POSITIONS[0].read(),
@@ -198,6 +167,8 @@ extern "C" fn draw_sprites(_bits: IrqBits) {
   );
 }
 
+// we out-line this because otherwise `draw_sprites` overloads the stack when it
+// holds 3 copies of this function.
 #[inline(never)]
 fn draw_rect(x: u16, y: u16, width: u16, height: u16, color: Color) {
   for i in 0..width {

src/video.rs

@@ -1,4 +1,4 @@
-//!
+//! Module for screen-related types and functions.
 
 use bitfrob::{u16_get_bit, u16_with_bit, u16_with_value};
 
@@ -326,3 +326,59 @@ pub struct Tile4bpp(pub [u32; 8]);
 #[derive(Clone, Copy, Default)]
 #[repr(transparent)]
 pub struct Tile8bpp(pub [u32; 16]);
+
+/// A zero-sized type that gives a namespace for Mode 3 related things.
+#[derive(Clone, Copy)]
+pub struct Mode3;
+impl Mode3 {
+  /// Width, in pixels, of the Mode 3 bitmap.
+  pub const WIDTH: usize = 240;
+
+  /// Height, in pixels, of the Mode 3 bitmap.
+  pub const HEIGHT: usize = 160;
+
+  /// The size, in bytes, of the Mode 3 bitmap.
+  pub const BYTES: usize =
+    Self::WIDTH * Self::HEIGHT * core::mem::size_of::<Color>();
+
+  /// Clears the entire bitmap to a color of your choosing.
+  #[cfg_attr(feature = "on_gba", instruction_set(arm::a32))]
+  #[cfg_attr(feature = "on_gba", link_section = ".iwram.mode3.clear_to")]
+  pub fn clear_to(self, color: Color) {
+    on_gba_or_unimplemented!(unsafe {
+      let x: u32 = color.0 as u32 | ((color.0 as u32) << 16);
+      // now we spam out that `u32`, 10 stm per loop, 8 times per stm.
+      core::arch::asm!(
+        "1:",
+        "stm  {ptr}!, {{r0-r5,r7-r8}}",
+        "stm  {ptr}!, {{r0-r5,r7-r8}}",
+        "stm  {ptr}!, {{r0-r5,r7-r8}}",
+        "stm  {ptr}!, {{r0-r5,r7-r8}}",
+        "stm  {ptr}!, {{r0-r5,r7-r8}}",
+        "stm  {ptr}!, {{r0-r5,r7-r8}}",
+        "stm  {ptr}!, {{r0-r5,r7-r8}}",
+        "stm  {ptr}!, {{r0-r5,r7-r8}}",
+        "stm  {ptr}!, {{r0-r5,r7-r8}}",
+        "stm  {ptr}!, {{r0-r5,r7-r8}}",
+        "subs {count}, {count}, #1",
+        "bne 1b",
+
+        // The assembler will give us a warning (that we can't easily disable)
+        // if the reg_list for `stm` doesn't give the registers in order from
+        // low to high, so we just manually pick registers. The count register
+        // and the pointer register can be anything else.
+        in("r0") x,
+        in("r1") x,
+        in("r2") x,
+        in("r3") x,
+        in("r4") x,
+        in("r5") x,
+        in("r7") x,
+        in("r8") x,
+        count = inout(reg) 240 => _,
+        ptr = inout(reg) crate::mmio::MODE3_VRAM.as_usize() => _,
+        options(nostack),
+      )
+    });
+  }
+}