Update hline with high performance method

2 files changed, 36 insertions, 7 deletions

diff --git a/src/renderer_m0.c b/src/renderer_m0.c
index 5c57278..c0416ea 100644
--- a/src/renderer_m0.c
+++ b/src/renderer_m0.c
@@ -85,10 +85,40 @@ void
 draw_hline(size_t x0, size_t x1, size_t y0, u8 clr) {
     BOUNDCHECK_SCREEN(x0, y0);
     BOUNDCHECK_SCREEN(x1, y0);
-
-    // TODO: perf
-    for (size_t x = x0; x <= x1; x++) {
-        draw_pixel(x, y0, clr);
+    // Find row positions for the given x/y coordinates.
+    size_t tile_x0 = x0 / 8;
+    size_t tile_x1 = x1 / 8;
+    size_t tile_y = y0 / 8;
+    size_t start_col = x0 % 8;
+    size_t end_col = x1 % 8;
+    size_t start_row = y0 % 8;
+
+    // Horizontal line. There are 3 cases:
+    //     1. Lines fit on a single tile.
+    //     2. Lines go through 2 tiles, both require partial row updates.
+    //     3. Lines go through 3 or more tiles, first and last tiles use
+    //        partial row updates, rows in the middle can write the entire
+    //        row.
+    size_t dx = tile_x1 - tile_x0;
+    u32 *dst = &backbuf[tile_x0 * 8 + tile_y * 8 * 32 + start_row];
+    if (dx < 1) {
+        u32 mask = 0xFFFFFFFF;
+        mask >>= (7 - end_col - dx) * 4;
+        mask &= 0xFFFFFFFF << start_col * 4;
+        u32 row = (0x11111111 * clr) & mask;
+        *dst = (*dst & ~mask) | row;
+    } else {
+        size_t shift_left = start_col * 4;
+        size_t shift_right = (7 - end_col) * 4;
+        u32 mask = 0xFFFFFFFF;
+        u32 row = 0x11111111 * clr;
+        *dst = (*dst & ~(mask << shift_left)) | (row << shift_left);
+        dst += 8;
+        for (size_t i = 1; i < dx; i++) {
+            *dst = (*dst & ~mask) | row;
+            dst += 8;
+        }
+        *dst = (*dst & ~(mask >> shift_right)) | (row >> shift_right);
     }
 }
 
@@ -407,8 +437,7 @@ draw_2bpp_row(size_t x, size_t y, u8 a, u8 b, u8 flip_x) {
     size_t shift_left = start_col * 4;
     size_t shift_right = (8 - start_col) * 4;
 
-    u32 *dst = &backbuf[tile_x * 8 + tile_y * 8 * 32];
-    dst += start_row;
+    u32 *dst = &backbuf[tile_x * 8 + tile_y * 8 * 32 + start_row];
     if (start_col == 0) {
         u32 clr_a = decode_1bpp(a, flip_x);
         u32 clr_b = decode_1bpp(b, flip_x);
diff --git a/src/renderer_m4.c b/src/renderer_m4.c
index 4cb526b..1539ccb 100644
--- a/src/renderer_m4.c
+++ b/src/renderer_m4.c
@@ -70,7 +70,7 @@ draw_hline(size_t x0, size_t x1, size_t y0, u8 clr) {
         u64 row_mask = 0xFFFFFFFFFFFFFFFF;
         row_mask >>= (7 - end_col - dx) * 8;
         row_mask &= 0xFFFFFFFFFFFFFFFF << start_col * 8;
-        u64 row = (0x0101010101010101 * (u8)clr) & row_mask;
+        u64 row = (0x0101010101010101 * clr) & row_mask;
         *dst = (*dst & ~row_mask) | row;
     } else {
         size_t shift_left = start_col * 8;