Add optional DMA usage on flipbuf

This will copy an entire row of tiles instead of per tile copy, but for
applications that clear the screen on each frame it will be more
efficient.

1 files changed, 28 insertions, 26 deletions

diff --git a/src/ppu.c b/src/ppu.c
index d68ab8f..bc6f503 100644
--- a/src/ppu.c
+++ b/src/ppu.c
@@ -16,6 +16,7 @@ WITH REGARD TO THIS SOFTWARE.
 */
 
 #define NEW_PPU 1
+#define FLIPBUF_DMA 1
 
 #define FG_FRONT  ((u32*)(MEM_VRAM))
 #define BG_FRONT  ((u32*)(MEM_VRAM + KB(20)))
@@ -722,32 +723,33 @@ putfontchar(u32 *layer, u16 tile_x, u16 tile_y, u8 ch, u8 color) {
 IWRAM_CODE
 void
 flipbuf() {
-    // Copy the fg and bg layers fully with the DMA.
-    dma_copy((u32*)FG_FRONT, (u32*)FG_BACK, KB(40), 3);
-    // dma_sync_copy((u32*)FG_FRONT, (u32*)FG_BACK, KB(40));
-    //
-    // TODO: dma the dirty lines instead of per tile copying.
-    // dma_copy((u32*)BG_FRONT, (u32*)BG_BACK, KB(20), 3);
-    // dma_sync_copy((u32*)FG_FRONT, (u32*)FG_BACK, KB(20));
-    // dma_sync_copy((u32*)BG_FRONT, (u32*)BG_BACK, KB(20));
-    // Tile *mem_fg = FG_FRONT;
-    // Tile *mem_bg = BG_FRONT;
-    // for (size_t j = 0; j < 20; ++j) {
-    //     if (dirty_tiles[j] == 0) {
-    //         continue;
-    //     }
-
-    //     // size_t k = 1;
-    //     // for (size_t i = 0; i < 30; ++i, k <<= 1) {
-    //     //     if (dirty_tiles[j] & k) {
-    //     //         Tile *tile_fg = FG_BACK;
-    //     //         Tile *tile_bg = BG_BACK;
-    //     //         mem_fg[i + j * 32] = tile_fg[i + j * 32];
-    //     //         mem_bg[i + j * 32] = tile_bg[i + j * 32];
-    //     //     }
-    //     // }
-    //     dirty_tiles[j] = 0;
-    // }
+    u32 *fg_back = FG_BACK;
+    u32 *bg_back = BG_BACK;
+    u32 *bg_front = BG_FRONT;
+    u32 *fg_front = FG_FRONT;
+    for (size_t j = 0; j < 20; ++j) {
+        if (dirty_tiles[j] == 0) {
+            continue;
+        }
+#if FLIPBUF_DMA == 1
+        u32 offset = j * 32 * 8;
+        dma_copy(fg_front + offset, fg_back + offset, 32 * 8 * 4, 3);
+        dma_copy(bg_front + offset, bg_back + offset, 32 * 8 * 4, 3);
+#else
+        size_t k = 1;
+        for (size_t i = 0; i < 30; ++i, k <<= 1) {
+            if (dirty_tiles[j] & k) {
+                Tile *mem_fg = FG_FRONT;
+                Tile *mem_bg = BG_FRONT;
+                Tile *tile_fg = FG_BACK;
+                Tile *tile_bg = BG_BACK;
+                mem_fg[i + j * 32] = tile_fg[i + j * 32];
+                mem_bg[i + j * 32] = tile_bg[i + j * 32];
+            }
+        }
+#endif
+        dirty_tiles[j] = 0;
+    }
 }
 
 typedef struct KeyboardChar {