diff --git a/src/static/TileStoreDefs.s b/src/static/TileStoreDefs.s
index 1e5ada2..c6a41c2 100644
--- a/src/static/TileStoreDefs.s
+++ b/src/static/TileStoreDefs.s
@@ -17,7 +17,7 @@ TS_SCREEN_ADDR        equ  {TILE_STORE_SIZE*8}      ; cached value of on-screen
 
 ; TODO: Move these arrays into the K bank to support direct dispatch via jmp (abs,x)
 TS_BASE_TILE_COPY     equ  {TILE_STORE_SIZE*9}      ; derived from TS_TILE_ID to optimize tile copy to support sprite rendering
-TS_BASE_TILE_DISP     equ  {TILE_STORE_SIZE*10}     ; derived from TS_TILE_ID to optimize base (non-sprite) tile dispatch in the Render function
+; TS_BASE_TILE_DISP     equ  {TILE_STORE_SIZE*10}     ; derived from TS_TILE_ID to optimize base (non-sprite) tile dispatch in the Render function
 TS_DIRTY_TILE_DISP    equ  {TILE_STORE_SIZE*11}     ; derived from TS_TILE_ID to optimize dirty tile dispatch in the Render function
 
 TILE_STORE_NUM        equ  12                     ; Need this many parallel arrays
diff --git a/src/tiles/FastRenderer.s b/src/tiles/FastRenderer.s
index 4363625..14dcd90 100644
--- a/src/tiles/FastRenderer.s
+++ b/src/tiles/FastRenderer.s
@@ -13,13 +13,9 @@ NoSpriteFast
             lda   TileStore+TS_CODE_ADDR_HIGH,x    ; load the bank of the target code field line
             pha                                    ; and put on the stack for later. Has TileStore bank in high byte.
             ldy   TileStore+TS_CODE_ADDR_LOW,x     ; load the address of the code field
-;            lda   TileStore+TS_BASE_TILE_DISP,x    ; go to the tile copy routine (just basics)
-;            stal  nsf_patch+1
             lda   TileStore+TS_TILE_ADDR,x         ; load the address of this tile's data (pre-calculated)
             plb                                    ; set the code field bank
-            jmp   (K_TS_BASE_TILE_DISP,x)
-
-;nsf_patch   jmp   $0000
+            jmp   (K_TS_BASE_TILE_DISP,x)          ; go to the tile copy routine (just basics)
 
 ; The TS_BASE_TILE_DISP routines will come from this table when ENGINE_MODE_TWO_LAYER and
 ; ENGINE_MODE_DYN_TILES are both off.
@@ -32,8 +28,51 @@ FastTileProcs dw   _TBCopyDataFast,_TBCopyDataFast,_TBCopyDataFast,_TBCopyDataFa
 ; then the mask information must be copied as well....This is the last decision point.
 
 SpriteDispatch
+;            jmp   (K_TS_SPRITE_TILE_DISP,x)
+
             txy
-            SpriteBitsToVBuffAddrs OneSpriteFast;TwoSpritesFast;ThreeSpritesFast;FourSpritesFast
+            SpriteBitsToVBuffAddrs OneSpriteFastUnder;TwoSpritesFast;ThreeSpritesFast;FourSpritesFast
+
+; This handles sprite with the tile above
+OneSpriteFastUnder
+            tax
+            jsr   _CopySpriteDataToDP2             ; preserves Y
+
+            ldx   TileStore+TS_TILE_ADDR,y
+            lda   TileStore+TS_CODE_ADDR_HIGH,y    ; load the bank of the target code field line
+            pha                                    ; and put on the stack for later. Has TileStore bank in high byte.
+            lda   TileStore+TS_CODE_ADDR_LOW,y     ; load the address of the code field
+            tay
+            plb
+
+]line       equ   0
+            lup   8
+            lda   tmp_tile_data+{]line*4}
+            andl  tiledata+{]line*4}+32,x
+            oral  tiledata+{]line*4},x
+            sta:  $0004+{]line*$1000},y
+
+            lda   tmp_tile_data+{]line*4}+2
+            andl  tiledata+{]line*4}+32+2,x
+            oral  tiledata+{]line*4}+2,x
+            sta:  $0001+{]line*$1000},y
+]line       equ   ]line+1
+            --^
+            plb
+            rts 
+
+
+_CopySpriteDataToDP2
+]line       equ   0
+            lup   8
+            ldal  spritedata+{]line*SPRITE_PLANE_SPAN},x
+            sta   tmp_tile_data+{]line*4}
+
+            ldal   spritedata+{]line*SPRITE_PLANE_SPAN}+2,x
+            sta   tmp_tile_data+{]line*4}+2
+]line       equ   ]line+1
+            --^
+            rts
 
 ; Where there are sprites involved, the first step is to call a routine to copy the
 ; tile data into a temporary buffer.  Then the sprite data is merged and placed into
@@ -52,7 +91,6 @@ OneSpriteFast
             tay
             plb
 
-_TBApplySpriteData2
 ]line       equ   0
             lup   8
             lda   tmp_tile_data+{]line*4}