diff --git a/testbed.c b/testbed.c
index 61cf5a5..1059b1a 100755
--- a/testbed.c
+++ b/testbed.c
@@ -391,7 +391,7 @@ int tifLoad(FILE *f, TifImageMeta_t meta, unsigned int *planeBuf, int maxY, int
   unsigned int *rp = gp + planeStride;
   unsigned int *ip = rp + planeStride;
   unsigned int *mp = ip + planeStride;
-  unsigned char bv, gv, rv, iv;
+  unsigned int bv, gv, rv, iv;
 
   if (meta.width > MAX_WIDTH || (meta.width % 16) != 0 || planes < 4 || planes > 5) {
     return 0;
@@ -412,7 +412,7 @@ int tifLoad(FILE *f, TifImageMeta_t meta, unsigned int *planeBuf, int maxY, int
 	int gpair = (pixelpair & 0x02) >> 1 | (pixelpair & 0x20) >> 4;
 	int rpair = (pixelpair & 0x04) >> 2 | (pixelpair & 0x40) >> 5;
 	int ipair = (pixelpair & 0x08) >> 3 | (pixelpair & 0x80) >> 6;
-	int shift = (3 - (ipixelpair % 4)) << 1;
+	int shift = (7 - (ipixelpair % 8)) << 1;
 
 	bv |= bpair << shift;
 	gv |= gpair << shift;
@@ -425,7 +425,8 @@ int tifLoad(FILE *f, TifImageMeta_t meta, unsigned int *planeBuf, int maxY, int
 	  *rp++ = rv;
 	  *ip++ = iv;
 	  if (planes == 5) {
-	    *mp++ = bv & gv & rv & iv;
+	    iv = ~(bv & gv & rv & iv);
+	    *mp++ = iv;
 	  }
 	  bv = gv = rv = iv = 0;
 	}
@@ -448,6 +449,18 @@ int tifLoad(FILE *f, TifImageMeta_t meta, unsigned int *planeBuf, int maxY, int
 
 /*** T I L E S ***/
 
+// Tiles are 16x16 bitmaps, stored as arrays of words.
+// Each tile has 4 or 5 planes (depending on whether it is a tile or sprite)
+// which are stored adjacant to each other; ie. a 16-word array of blue,
+// followed by a 16-word array of green, etc.
+// Tiles in RAM are stored byte-swapped to aid in fast bit-shifting, and must
+// be byte-swapped before being written to video memory.
+
+// Because bit-shifting operations happen on little-endian words:
+// 01234567 89ABCDEF << 3 => 34567XXX BCDEF012
+// which is wrong. So instead we do:
+// 89ABCDEF 01234567 << 3 => BCDEFXXX 3456789A byteswap => 3456789A BCDEFXXX
+
 #define PAGE_TILES_W 21
 #define PAGE_TILES_H 13
 #define PAGE_TILES_COUNT (PAGE_TILES_H * PAGE_TILES_W)
@@ -503,6 +516,7 @@ void loadMap(unsigned char *map, unsigned int w, unsigned int h) {
   screen.map = map;
   screen.w = w;
   screen.h = h;
+  memset(screen.dirty, D_BGTILE, PAGE_TILES_COUNT * 2);
 }
 
 int prepareBuffer(int pageX, int pageY) {
@@ -523,6 +537,7 @@ int prepareBuffer(int pageX, int pageY) {
 
 void drawSpriteToBuf(unsigned int *sprite, int pageX, int pageY, int shift, int yStart) {
   unsigned int *buf, *mask;
+  unsigned int maskval;
   int y, h, plane;
   if (pageX < 0 || pageY < 0 ||
       pageX >= PAGE_TILES_W || pageY >= PAGE_TILES_H ||
@@ -532,19 +547,20 @@ void drawSpriteToBuf(unsigned int *sprite, int pageX, int pageY, int shift, int
   }
 
   buf = screen.buffer[prepareBuffer(pageX, pageY)];
-/*  if (yStart < 0) {
+  if (yStart < 0) {
     sprite = &sprite[-yStart];
     h = yStart + 16;
   } else {
     buf = &buf[yStart];
     h = 16 - yStart;
   }
-  mask = &sprite[16 * 4];
+  mask = &sprite[BUF_WSTRIDE * 4];
   if (shift < 0) {
     shift = -shift;
     for (plane = 0; plane < 4; plane ++) {
       for (y = 0; y < h; y ++) {
-	buf[y] = (buf[y] & ~(mask[y] << shift)) | (sprite[y] << shift);
+	maskval = mask[y] << shift;
+	buf[y] = (buf[y] & ~maskval) | ((sprite[y] << shift) & maskval);
       }
       sprite += BUF_WSTRIDE;
       buf += BUF_WSTRIDE;
@@ -552,12 +568,13 @@ void drawSpriteToBuf(unsigned int *sprite, int pageX, int pageY, int shift, int
   } else {
     for (plane = 0; plane < 4; plane ++) {
       for (y = 0; y < h; y ++) {
-	buf[y] = (buf[y] & ~(mask[y] >> shift)) | (sprite[y] >> shift);
+	maskval = mask[y] >> shift;
+	buf[y] = (buf[y] & ~maskval) | ((sprite[y] >> shift) & maskval);
       }
       sprite += BUF_WSTRIDE;
       buf += BUF_WSTRIDE;
     }
-  }*/
+  }
 }
 
 void drawSprite(unsigned int *sprite, int x, int y) {
@@ -572,15 +589,32 @@ void drawSprite(unsigned int *sprite, int x, int y) {
   drawSpriteToBuf(sprite, pageX + 1, pageY + 1, pageOffsetX - 16, pageOffsetY - 16);
 }
 
-void scroll(int x, int y) {
-  x = min(max(x, 0), (screen.w << 4) - 320);
-  y = min(max(y, 0), (screen.h << 4) - 176);
-  if ((screen.scrollX & 0xfff0) != (x & 0xfff0) ||
-      (screen.scrollY & 0xfff0) != (y & 0xfff0)) {
-    memset(screen.dirty, D_BGTILE, 2 * PAGE_TILES_COUNT);
+void scroll(int newX, int newY) {
+  newX = min(max(newX, 0), (screen.w << 4) - 320);
+  newY = min(max(newY, 0), (screen.h << 4) - 176);
+  if ((screen.scrollX & 0xfff0) != (newX & 0xfff0) ||
+      (screen.scrollY & 0xfff0) != (newY & 0xfff0)) {
+    int mapX, mapY;
+    unsigned char page;
+    for (page = 0; page < 2; page ++) {
+      int mapOffsetOld = (screen.scrollX >> 4) + ((screen.scrollY >> 4) * screen.w);
+      int mapOffsetNew = (newX >> 4) + ((newY >> 4) * screen.w);
+      unsigned char *dirty = screen.dirty[page];
+      for (mapY = 0; mapY < PAGE_TILES_H; mapY ++) {
+	for (mapX = 0; mapX < PAGE_TILES_W; mapX ++) {
+	  if (*dirty != D_NOTHING ||
+	    screen.map[mapOffsetOld + mapX] != screen.map[mapOffsetNew + mapX]) {
+	    *dirty = D_BGTILE;
+	  }
+	  dirty ++;
+	}
+	mapOffsetNew += screen.w;
+	mapOffsetOld += screen.w;
+      }
+    }
   }
-  screen.scrollX = x;
-  screen.scrollY = y;
+  screen.scrollX = newX;
+  screen.scrollY = newY;
 }
 
 void drawScreen() {
@@ -591,7 +625,7 @@ void drawScreen() {
   unsigned int drawOffset = screen.pageOffset[screen.currentPage];
   unsigned int scrollOffset = drawOffset + (offsetX >> 3) + (offsetY * PAGE_STRIDE);
   unsigned char *dirty = screen.dirty[screen.currentPage];
-  unsigned int x, y, di, plane;
+  unsigned int x, y, di, plane, bmp;
 
   setAllPlanes();
   setWriteMode(1);
@@ -600,7 +634,9 @@ void drawScreen() {
   for (y = startY; y < startY + PAGE_TILES_H; y ++) {
     for (x = startX; x < startX + PAGE_TILES_W; x ++) {
       if (dirty[di++] == D_BGTILE) {
-	blitTile(screen.tilesOffset + screen.map[x + (y * screen.w)], drawOffset);
+	blitTile(
+	  screen.tilesOffset + (screen.map[x + (y * screen.w)] << 5),
+	  drawOffset);
       }
       drawOffset += 2;
     }
@@ -612,7 +648,8 @@ void drawScreen() {
     for (di = screen.firstBuffer; di != screen.nextBuffer; di = nextBufferIndex(di)) {
       drawOffset = screen.bufferOffset[di] >> 1;
       for (y = 0; y < 16; y ++) {
-	WVID[drawOffset] = screen.buffer[di][y + (BUF_WSTRIDE * plane)];
+	bmp = screen.buffer[di][y + (BUF_WSTRIDE * plane)];
+	WVID[drawOffset] = (bmp << 8) | (bmp >> 8);
 	drawOffset += PAGE_STRIDE >> 1;
       }
     }
@@ -630,9 +667,10 @@ void drawScreen() {
 #define NUM_TILES 128
 #define NUM_SPRITES 64
 #define OFF_TILES 0x4840
-
-unsigned int tiles[NUM_TILES][4*16];
-unsigned int sprites[NUM_SPRITES][5*16];
+#define TILE_STRIDE 64
+#define SPRITE_STRIDE 80
+unsigned int tiles[NUM_TILES * TILE_STRIDE];
+unsigned int sprites[NUM_SPRITES * SPRITE_STRIDE];
 
 
 unsigned char map[10000];
@@ -643,7 +681,7 @@ void fillMap() {
 
   for (y = 0; y < 100; y ++) {
     for (x = 0; x < 100; x ++) {
-      map[x + (y * 100)] = (((x + y + z) >> 2) % 4) << 5;
+      map[x + (y * 100)] = ((x + y + z) >> 2) % 4;
     }
   }
 }
@@ -665,8 +703,8 @@ void game_init() {
 
   f = fopen("TILES.TIF", "rb");
   meta = tifLoadMeta(f);
-  tifLoadEGA(f, meta, OFF_TILES, NUM_TILES * 16, 16);
   tifLoad(f, meta, tiles, NUM_TILES * 16, 16, 4);
+  tifLoadEGA(f, meta, OFF_TILES, NUM_TILES * 16, 16);
   fclose(f);
 
   f = fopen("SPRITE.TIF", "rb");
@@ -691,13 +729,13 @@ int main() {
 
   while (!keyPressed(K_ESC)) {
 
-    if (keyPressed(K_LEFT))  x -= 1;
-    if (keyPressed(K_RIGHT)) x += 1;
-    if (keyPressed(K_UP))    y -= 1;
-    if (keyPressed(K_DOWN))  y += 1;
+    if (keyPressed(K_LEFT))  x -= 3;
+    if (keyPressed(K_RIGHT)) x += 3;
+    if (keyPressed(K_UP))    y -= 3;
+    if (keyPressed(K_DOWN))  y += 3;
     scroll(x - 152, y - 90);
-    drawSprite(sprites[0], 50, 50);
-    drawSprite(sprites[1], x, y);
+    drawSprite(&sprites[0 * SPRITE_STRIDE], 50, 50);
+    drawSprite(&sprites[1 * SPRITE_STRIDE], x, y);
     drawScreen();
 
     z++;
diff --git a/testbed.exe b/testbed.exe
index b39e2c7..78bdec1 100755
Binary files a/testbed.exe and b/testbed.exe differ