Merge pull request #96 from cellos51/hblank_affine_background_effects_optimizations

Optimized affine background HBLANK effect (still not good enough)

include/affine_background.h

@@ -7,10 +7,11 @@
 
 #define AFFINE_BG_PAL_LEN 5
 #define AFFINE_BG_PB (PAL_ROW_LEN * 15) // This isn't really a palette bank, just the starting index of the palette
+#define AFFINE_BG_IDX 2 // The index of the affine background BGCNT register etc.
 
 void affine_background_init();
-void affine_background_hblank();
-void affine_background_update();
+IWRAM_CODE void affine_background_hblank();
+IWRAM_CODE void affine_background_update();
 void affine_background_set_color(COLOR color);
 // Must be called with an array of size at least  AFFINE_BG_PAL_LEN
 void affine_background_load_palette(const u16 *src);

source/affine_background.c

@@ -3,11 +3,10 @@
 
 #include "graphic_utils.h"
 
-BG_AFFINE bgaff;
-AFF_SRC_EX asx = {32<<8, 64<<8, 120, 80, 0x0100, 0x0100, 0};
+BG_AFFINE bgaff_arr[SCREEN_HEIGHT + 1];
+
 static uint timer = 0;
 
-#define TOP_SCANLINE_OFFSET 228
 #define ANIMATION_SPEED_DIVISOR 16
 
 
@@ -17,41 +16,57 @@ void affine_background_init()
     GRIT_CPY(&se_mem[AFFINE_BG_SBB], affine_background_gfxMap);
     affine_background_load_palette(affine_background_gfxPal);
 
-    bgaff = bg_aff_default;
+    REG_BG_AFFINE[AFFINE_BG_IDX] = bg_aff_default;
 }
 
-void affine_background_hblank()
+// Pre-computes the affine matrices values for each scanline 
+// and stores in bgaff_arr. 
+// This is to be done in VBLANK so the HBLANK code 
+// can just fetch the values quickly.
+IWRAM_CODE void affine_background_prep_bgaff_arr()
+{
+    for (u16 vcount = 0; vcount < SCREEN_HEIGHT; vcount++)
+    {
+        AFF_SRC_EX asx = {32<<8, 64<<8, 120, 80, 0x0100, 0x0100, 0};
+        const s32 timer_s32 = timer << 8;
+        const s32 vcount_s32 = vcount << 8;
+        const s16 vcount_s16 = vcount;
+        const s32 vcount_sine = lu_sin(vcount_s32 + timer_s32 / ANIMATION_SPEED_DIVISOR); // dividing the timer by 16 to make the animation slower
+
+        asx.scr_x = (SCREEN_WIDTH / 2); // 128 on x and y is an offset used to center the rotation
+        asx.scr_y = vcount_s16 - (SCREEN_HEIGHT / 2); // scr_y must equal vcount otherwise the background will have no vertical difference
+        asx.tex_x = vcount_sine;
+        asx.tex_y = vcount_sine / 64;
+        asx.sx = vcount_sine / 32;
+        asx.sy = vcount_sine / 16;
+        asx.alpha = vcount_sine + (timer_s32 / ANIMATION_SPEED_DIVISOR);
+
+        bg_rotscale_ex(&bgaff_arr[vcount], &asx);
+    }
+
+    /* HBLANK occurs after the scanline so REG_VCOUNT represents the 
+     * the scanline that just passed, so when it's SCREEN_HEIGHT we will
+     * actually be updating the first line
+     */
+    bgaff_arr[SCREEN_HEIGHT] = bgaff_arr[0];
+}
+
+IWRAM_CODE void affine_background_hblank()
 {
     vu16 vcount = REG_VCOUNT;
 
-    if ((vcount > 160 && vcount < 225) || vcount > TOP_SCANLINE_OFFSET) // Exit the function if the scanline is outside the valid range
+    if ((vcount >= SCREEN_HEIGHT)) // Exit the function if the scanline is outside the screen
     {
         return;
     }
-    else if (vcount > TOP_SCANLINE_OFFSET - 3) // This fixes a visual issue where the top 3 pixels of the scanline are out of sync
-    {
-        vcount -= TOP_SCANLINE_OFFSET;
-    }
 
-    const s32 timer_s32 = timer << 8;
-    const s32 vcount_s32 = vcount << 8;
-    const s16 vcount_s16 = vcount;
-    const s32 vcount_sine = lu_sin(vcount_s32 + timer_s32 / ANIMATION_SPEED_DIVISOR); // dividing the timer by 16 to make the animation slower
-
-    asx.scr_x = (SCREEN_WIDTH / 2); // 128 on x and y is an offset used to center the rotation
-    asx.scr_y = vcount_s16 - (SCREEN_HEIGHT / 2); // scr_y must equal vcount otherwise the background will have no vertical difference
-    asx.tex_x = vcount_sine;
-    asx.tex_y = vcount_sine / 64;
-    asx.sx = vcount_sine / 32;
-    asx.sy = vcount_sine / 16;
-    asx.alpha = vcount_sine + (timer_s32 / ANIMATION_SPEED_DIVISOR);
-    
-    bg_rotscale_ex(&bgaff, &asx);
-    REG_BG_AFFINE[2] = bgaff;
+    // See comment in affine_background_prep_bgaff_arr()
+    REG_BG_AFFINE[AFFINE_BG_IDX] = bgaff_arr[vcount + 1];
 }
 
 void affine_background_update()
 {
+    affine_background_prep_bgaff_arr();
     timer++;
 }