More performance stuff

Still can't get 50fps on an A500 but this is good enough for now.

bun.c

@@ -56,6 +56,33 @@ struct BunClear {
 
 uint8_t *coolbunPlaneStarts[3][2];
 
+uint16_t bunRightSide_bltalwmPrecalc[16];
+uint16_t bunLeftSide_bltalwmPrecalc[16];
+
+void precalculateBunBLTALWM() {
+  int shift, j;
+  uint16_t bltalwm;
+
+  for (shift = 0; shift < 16; ++shift) {
+    // right
+    bltalwm = 0x0000;
+
+    for (j = 0; j < 15 - shift; ++j) {
+      bltalwm += (1 << (15 - j));
+    }
+
+    bunRightSide_bltalwmPrecalc[shift] = bltalwm;
+
+    //left
+    bltalwm = 0x0000;
+    for (j = 0; j <= 15 - shift; ++j) {
+      bltalwm += (1 << j);
+    }
+
+    bunLeftSide_bltalwmPrecalc[shift] = bltalwm;
+  }
+}
+
 void bun_offRightSide(
   int plusXValue,
   int y,
@@ -75,10 +102,14 @@ void bun_offRightSide(
   bunClear->widthWords = 2 - wordShift;
   bunClear->direction = BLITTER_ASCENDING;
 
+  bltalwm = bunRightSide_bltalwmPrecalc[shift];
+
+  /*
   bltalwm = 0x0000;
   for (i = 0; i <= 15 - shift; ++i) {
     bltalwm += (1 << (15 - i));
   }
+  */
 
   bltcon0 = BLTCON0(0xf0, 1, 0, 0, 1, shift);
   bltcon1 = BLTCON1(bunClear->direction, shift);
@@ -130,10 +161,14 @@ void bun_offLeftSide(
   bunClear->widthWords = 2 - wordShift;
   bunClear->direction = BLITTER_DESCENDING;
 
+  bltalwm = bunLeftSide_bltalwmPrecalc[shift];
+
+  /*
   bltalwm = 0x0000;
   for (i = 0; i <= 15 - shift; ++i) {
     bltalwm += (1 << i);
   }
+  */
 
   bltcon0 = BLTCON0(0xf0, 1, 0, 0, 1, shift);
   bltcon1 = BLTCON1(bunClear->direction, shift);
@@ -215,6 +250,8 @@ void bun_anywhere(
   }
 }
 
+#define ENABLE_RENDER_BUN_CHECKING (0)
+
 void renderBun(
   int x,
   int y,
@@ -224,14 +261,17 @@ void renderBun(
 ) {
   /**
     * Conditions that will cause the program to crash if met. If your bun
-    * isn't rendering, it's due to here.
+    * isn't rendering, it's due to here. Turn off checking once you know
+    * everything is solid to save some CPU.
     *
     * TODO: Handle top/bottom off-screen as well.
     */
+  if (ENABLE_RENDER_BUN_CHECKING) {
     if (x < -31) return;
     if (x > screenDefinition->width + 31) return;
     if (y < 1) return;
     if (y > screenDefinition->height - COOL_BUN_HEIGHT - 1) return;
+  }
 
   if (x < 0) {
     bun_offLeftSide(
@@ -348,6 +388,7 @@ void setupBunRenderer(
   buildBunAngleAdjustments();
   calculateAllBunPositions(screenDefinition);
   precalculateBunRenderInfo();
+  precalculateBunBLTALWM();
 }
 
 /**

main.c

@@ -30,8 +30,11 @@
 #include "bun.h"
 
 /**
- * This barely gets 50fps but I'm leaving it for now. Potential improvements:
+ * This barely gets 50fps on an A1200 but I'm leaving it for now. Potential improvements:
  *
+ * [ ] Precalculate even more of bun.c. This may get 50fps on an A500 if I do this.
+ *     It would require more significant caching of as much stuff as possible.
+ *     The A500 is spending entirely too much time on CPU tasks.
  * [ ] Topaz re-rendering is limited to squares or rows that changed in the last frame.
  *     I tried this once but I need a more precise way of redrawing those areas.
  * [ ] Cool bun clears and re-renders happen in the same pass.
@@ -368,7 +371,7 @@ int main(void) {
     if ((ciaa.ciapra >> 6) != 3) break;
 
     i++;
-    i %= FRAMES_FOR_SCREEN;
+    if (i > FRAMES_FOR_SCREEN) i = 0;
   }
 
   giveBackSystem();

screen.c

@@ -21,6 +21,7 @@ void allocateDoubleBufferedScreenMemory(
   uint8_t bitplanes
 ) {
   unsigned char *memory;
+  int buffer, plane;
 
   screenDefinition->width = width;
   screenDefinition->height = height;
@@ -38,6 +39,14 @@ void allocateDoubleBufferedScreenMemory(
   screenDefinition->nextBitplaneAdvance = screenDefinition->byteWidth * height;
   screenDefinition->nextBufferAdvance = screenDefinition->nextBitplaneAdvance * bitplanes;
 
+  for (buffer = 0; buffer < 2; ++buffer) {
+    for (plane = 0; plane < bitplanes; ++plane) {
+      screenDefinition->bufferPlanes[buffer][plane] = screenDefinition->memoryStart +
+        buffer * screenDefinition->nextBufferAdvance +
+        plane * screenDefinition->nextBitplaneAdvance;
+    }
+  }
+
   setActiveScreenBuffer(screenDefinition, activeScreenBufferDetails, 0);
 }
 
@@ -59,9 +68,7 @@ void setActiveScreenBuffer(
   currentScreen->currentBuffer = buffer;
 
   for (plane = 0; plane < screenDefinition->bitplanes; ++plane) {
-    currentScreen->planes[plane] = screenDefinition->memoryStart +
-      buffer * screenDefinition->nextBufferAdvance +
-      plane * screenDefinition->nextBitplaneAdvance;
+    currentScreen->planes[plane] = screenDefinition->bufferPlanes[buffer][plane];
   }
 }
 

screen.h

@@ -17,6 +17,7 @@ struct ScreenDefinition {
   uint16_t byteWidth;
   uint16_t nextBitplaneAdvance;
   uint16_t nextBufferAdvance;
+  unsigned char *bufferPlanes[2][8];
 };
 
 struct ActiveScreenBufferDetails {