From fec401597834a22bf9d8554bcbe38278bf9734f1 Mon Sep 17 00:00:00 2001
From: Gwilym Kuiper <gw@ilym.me>
Date: Sat, 23 Apr 2022 14:30:46 +0100
Subject: [PATCH] Remove hash map lookup for every pixel rendered

---
 agb/src/display/font.rs | 78 ++++++++++++++++++++++++++++-------------
 1 file changed, 53 insertions(+), 25 deletions(-)

diff --git a/agb/src/display/font.rs b/agb/src/display/font.rs
index badafc56..c5a71ace 100644
--- a/agb/src/display/font.rs
+++ b/agb/src/display/font.rs
@@ -86,7 +86,7 @@ pub struct TextRenderer<'a> {
     bg: &'a mut RegularMap,
     background_colour: u8,
     foreground_colour: u8,
-    tiles: HashMap<(usize, usize), DynamicTile<'a>>,
+    tiles: HashMap<(i32, i32), DynamicTile<'a>>,
 }
 
 impl<'a> Write for TextRenderer<'a> {
@@ -96,23 +96,6 @@ impl<'a> Write for TextRenderer<'a> {
         let foreground_colour = self.foreground_colour;
         let background_colour = self.background_colour;
 
-        let mut render_pixel = |x: u16, y: u16| {
-            let tile_x = (x / 8) as usize;
-            let tile_y = (y / 8) as usize;
-            let inner_x = x % 8;
-            let inner_y = y % 8;
-
-            let colour = foreground_colour as u32;
-
-            let index = (inner_x + inner_y * 8) as usize;
-
-            let tile = tiles
-                .entry((tile_x, tile_y))
-                .or_insert_with(|| vram_manager.new_dynamic_tile().fill_with(background_colour));
-
-            tile.tile_data[index / 8] |= colour << ((index % 8) * 4);
-        };
-
         for c in text.chars() {
             if c == '\n' {
                 self.current_y_pos += self.font.line_height;
@@ -126,15 +109,56 @@ impl<'a> Write for TextRenderer<'a> {
             let y_start =
                 self.current_y_pos + self.font.ascent - letter.height as i32 - letter.ymin as i32;
 
-            for letter_y in 0..(letter.height as i32) {
-                for letter_x in 0..(letter.width as i32) {
-                    let x = x_start + letter_x;
-                    let y = y_start + letter_y;
+            let x_tile_start = x_start / 8;
+            let y_tile_start = y_start / 8;
 
-                    let px = letter.data[(letter_x + letter_y * letter.width as i32) as usize];
+            let letter_offset_x = x_start.rem_euclid(8);
+            let letter_offset_y = y_start.rem_euclid(8);
 
-                    if px > 100 {
-                        render_pixel(x as u16, y as u16);
+            let x_tiles = div_ceil(letter.width as i32 + letter_offset_x, 8);
+            let y_tiles = div_ceil(letter.height as i32 + letter_offset_y, 8);
+
+            for letter_y_tile in 0..(y_tiles + 1) {
+                let letter_y_start = 0.max(letter_offset_y - 8 * letter_y_tile) + 8 * letter_y_tile;
+                let letter_y_end =
+                    (letter_offset_y + letter.height as i32).min((letter_y_tile + 1) * 8);
+
+                let tile_y = y_tile_start + letter_y_tile;
+
+                for letter_x_tile in 0..(x_tiles + 1) {
+                    let letter_x_start =
+                        0.max(letter_offset_x - 8 * letter_x_tile) + 8 * letter_x_tile;
+                    let letter_x_end =
+                        (letter_offset_x + letter.width as i32).min((letter_x_tile + 1) * 8);
+
+                    let tile_x = x_tile_start + letter_x_tile;
+
+                    let mut masks = [0u32; 8];
+                    let mut zero = true;
+
+                    for letter_y in letter_y_start..letter_y_end {
+                        let y = letter_y - letter_offset_y;
+
+                        for letter_x in letter_x_start..letter_x_end {
+                            let x = letter_x - letter_offset_x;
+                            let px = letter.data[(x + y * letter.width as i32) as usize];
+
+                            if px > 100 {
+                                masks[(letter_y & 7) as usize] |=
+                                    (foreground_colour as u32) << ((letter_x & 7) * 4);
+                                zero = false;
+                            }
+                        }
+                    }
+
+                    if !zero {
+                        let tile = tiles.entry((tile_x, tile_y)).or_insert_with(|| {
+                            vram_manager.new_dynamic_tile().fill_with(background_colour)
+                        });
+
+                        for (i, tile_data_line) in tile.tile_data.iter_mut().enumerate() {
+                            *tile_data_line |= masks[i];
+                        }
                     }
                 }
             }
@@ -146,6 +170,10 @@ impl<'a> Write for TextRenderer<'a> {
     }
 }
 
+fn div_ceil(quotient: i32, divisor: i32) -> i32 {
+    (quotient + divisor - 1) / divisor
+}
+
 impl<'a> TextRenderer<'a> {
     pub fn commit(mut self) {
         let tiles = core::mem::take(&mut self.tiles);