Tweaking updated sprite dispatch flow

demos/tool/App.Main.s

@@ -38,8 +38,6 @@ Flips           equ   12
 
                 pea   #256
                 pea   #160
-;                pea   #176
-;                pea   #144
                 _GTESetScreenMode
 
 ; Load a tileset
@@ -95,58 +93,57 @@ HERO_SPRITE     equ   SPRITE_16X16+1
                 cpx   #MAX_SPRITES*2
                 bcc   :sloop
 
-; Manually fill in the 41x26 tiles of the TileStore with a test pattern of trees 
+; Manually fill in the 41x26 tiles of the TileStore with a test pattern of trees
-;
-;   Tile 65 Tile 66
-;   Tile 97 Tile 98
-                stz   0            ; X
-                stz   2            ; Y
 
-:tloop
+                jsr   _fillTileStore
-                ldx   0
-                ldy   2
 
-                phx
+                ldx   #0
-                phy
+                ldy   #0
-                pea   #65
+                jsr   _drawTree
-                
-                inx
-                phx
-                phy
-                pea   #66
 
-                cpy   #24
+                ldx   #3
-                beq   :skip_last
+                ldy   #0
-                iny
+                jsr   _drawTreeH
-                phx
-                phy
-                pea   #98
 
-                dex
+                ldx   #0
-                phx
+                ldy   #3
-                phy
+                jsr   _drawTreeV
-                pea   #97
 
-                _GTESetTile
+                ldx   #3
-                _GTESetTile
+                ldy   #3
-:skip_last
+                jsr   _drawTreeHV
-                _GTESetTile
-                _GTESetTile
 
-                lda   0
+                ldx   #9
-                inc
+                ldy   #0
-                inc
+                jsr   _drawTree
-                sta   0
-                cmp   #40
-                bcc   :tloop
 
-                stz   0
+                ldx   #9
-                lda   2
+                ldy   #3
-                inc
+                jsr   _drawTree
-                inc
+
-                sta   2
+                ldx   #12
-                cmp   #25
+                ldy   #0
-                bcc   :tloop
+                jsr   _drawTree
+
+                ldx   #12
+                ldy   #3
+                jsr   _drawTree
+
+                ldx   #6
+                ldy   #0
+                jsr   _drawTreeFront
+                ldx   #6
+                ldy   #3
+                jsr   _drawTreeFront
+                ldx   #6
+                ldy   #6
+                jsr   _drawTreeFront
+                ldx   #3
+                ldy   #6
+                jsr   _drawTreeFront
+                ldx   #0
+                ldy   #6
+                jsr   _drawTreeFront
 
 ; Initialize the frame counter
 
@@ -450,6 +447,162 @@ GTEStartUp
 :ok3
                 rts
 
+_fillTileStore
+                stz    Tmp0
+:oloop
+                stz    Tmp1
+:iloop
+                pei    Tmp1
+                pei    Tmp0
+                pea    #129
+                _GTESetTile
+
+                lda    Tmp1
+                inc
+                sta    Tmp1
+                cmp    #41
+                bcc    :iloop
+
+                lda    Tmp0
+                inc
+                sta    Tmp0
+                cmp    #26
+                bcc    :oloop
+                rts
+
+;   Tile 65 Tile 66
+;   Tile 97 Tile 98
+
+_drawTreeFront
+                phx
+                phy
+                pea   #65+TILE_PRIORITY_BIT
+                
+                inx
+                phx
+                phy
+                pea   #66+TILE_PRIORITY_BIT
+
+                iny
+                phx
+                phy
+                pea   #98+TILE_PRIORITY_BIT
+
+                dex
+                phx
+                phy
+                pea   #97+TILE_PRIORITY_BIT
+
+                _GTESetTile
+                _GTESetTile
+                _GTESetTile
+                _GTESetTile
+                rts
+
+_drawTree
+                phx
+                phy
+                pea   #65
+                
+                inx
+                phx
+                phy
+                pea   #66
+
+                iny
+                phx
+                phy
+                pea   #98
+
+                dex
+                phx
+                phy
+                pea   #97
+
+                _GTESetTile
+                _GTESetTile
+                _GTESetTile
+                _GTESetTile
+                rts
+
+_drawTreeH
+                phx
+                phy
+                pea   #66+TILE_HFLIP_BIT
+
+                inx
+                phx
+                phy
+                pea   #65+TILE_HFLIP_BIT
+
+                iny
+                phx
+                phy
+                pea   #97+TILE_HFLIP_BIT
+
+                dex
+                phx
+                phy
+                pea   #98+TILE_HFLIP_BIT
+
+                _GTESetTile
+                _GTESetTile
+                _GTESetTile
+                _GTESetTile
+                rts
+
+_drawTreeV
+                phx
+                phy
+                pea   #97+TILE_VFLIP_BIT
+                
+                inx
+                phx
+                phy
+                pea   #98+TILE_VFLIP_BIT
+
+                iny
+                phx
+                phy
+                pea   #66+TILE_VFLIP_BIT
+
+                dex
+                phx
+                phy
+                pea   #65+TILE_VFLIP_BIT
+
+                _GTESetTile
+                _GTESetTile
+                _GTESetTile
+                _GTESetTile
+                rts
+
+_drawTreeHV
+                phx
+                phy
+                pea   #98+TILE_VFLIP_BIT+TILE_HFLIP_BIT
+                
+                inx
+                phx
+                phy
+                pea   #97+TILE_VFLIP_BIT+TILE_HFLIP_BIT
+
+                iny
+                phx
+                phy
+                pea   #65+TILE_VFLIP_BIT+TILE_HFLIP_BIT
+
+                dex
+                phx
+                phy
+                pea   #66+TILE_VFLIP_BIT+TILE_HFLIP_BIT
+
+                _GTESetTile
+                _GTESetTile
+                _GTESetTile
+                _GTESetTile
+                rts
+
 MyUserId        ds    2
 ToolPath        str   '1/Tool160'
 FrameCount      ds    2

demos/tool/package.json

@@ -12,7 +12,7 @@
   },
   "scripts": {
     "test": "npm run build && build-image.bat %npm_package_config_cadius% && %npm_package_config_gsport%",
-    "debug": "%npm_package_config_crossrunner% GTEToolDemo -Source MAINSEG_Output.txt -Debug -CompatibilityLayer",
+    "debug": "%npm_package_config_crossrunner% GTEToolDemo -Source MAINSEG_Output.txt -Debug -CompatibilityLayer -Map App.s",
     "build": "npm run build:tool && npm run build:sys16",
     "build:sys16": "%npm_package_config_merlin32% -V %npm_package_config_macros% App.s",
     "build:tool": "%npm_package_config_merlin32% -V %npm_package_config_macros% ../../src/Master.s"

src/Tiles.s

@@ -217,28 +217,20 @@ _SetTile
                  lda  #4
                  sta  procIdx
 :low_priority
-
                  lda  #TILE_ID_MASK
                  bit  newTileId
-                 bne  :not_zero
+                 beq  :is_zero
                  lda  #2
                  tsb  procIdx
-:not_zero
+:is_zero
 
                  lda  #TILE_VFLIP_BIT
+                 bit  newTileId
                  beq  :no_vflip
                  lda  #1
                  tsb  procIdx
 :no_vflip
 
-; Multiple by 6 to get the correct table entry index
-
-                 asl  procIdx
-                 lda  procIdx
-                 asl
-                 adc  procIdx
-                 sta  procIdx
-
 ; Now integrate with the engine mode indicator
 
                  lda  EngineMode
@@ -296,12 +288,22 @@ _SetTile
 tblPtr  equ  blttmp
 _SetTileProcs
 
-; Set a long pointer to this bank
+; Multiple the proc index by 6 to get the correct table entry offset
-                 sty  tblPtr
+
-                 clc
+                 asl
+                 sta  tblPtr
+                 asl
                  adc  tblPtr
                  sta  tblPtr
 
+; Add this offset to the base table address
+
+                 tya
+                 adc  tblPtr
+                 sta  tblPtr
+
+; Set the pointer to this bank
+
                  phk
                  phk
                  pla
@@ -350,12 +352,16 @@ _SetTileProcs
 FastProcs
 FastOverZA   dw   _TBConstTile0,GenericOverZero,_OneSpriteFastOver0
 FastOverZV   dw   _TBConstTile0,GenericOverZero,_OneSpriteFastOver0
-FastOverNA   dw   _TBCopyDataFast,GenericOverAFast,_OneSpriteFastOverA
+FastOverNA   dw   _TBCopyDataAFast,GenericOverZero,_OneSpriteFastOverA
-FastOverNV   dw   _TBCopyDataVFast,GenericOverVFast,_OneSpriteFastOverV
+FastOverNV   dw   _TBCopyDataVFast,GenericOverZero,_OneSpriteFastOverV
+;FastOverNA   dw   _TBCopyDataAFast,GenericOverAFast,_OneSpriteFastOverA
+;FastOverNV   dw   _TBCopyDataVFast,GenericOverVFast,_OneSpriteFastOverV
 FastUnderZA  dw   _TBConstTile0,GenericUnderZero,GenericUnderZero
 FastUnderZV  dw   _TBConstTile0,GenericUnderZero,GenericUnderZero
-FastUnderNA  dw   _TBCopyDataFast,GenericUnderAFast,_OneSpriteFastUnderA
+FastUnderNA  dw   _TBConstTile0,GenericOverZero,_OneSpriteFastOver0
-FastUnderNV  dw   _TBCopyDataVFast,GenericUnderVFast,_OneSpriteFastUnderV
+FastUnderNV  dw   _TBConstTile0,GenericOverZero,_OneSpriteFastOver0
+;FastUnderNA  dw   _TBCopyDataFast,GenericUnderAFast,_OneSpriteFastUnderA
+;FastUnderNV  dw   _TBCopyDataVFast,GenericUnderVFast,_OneSpriteFastUnderV
 
 DynProcs
 DynOverZA
@@ -483,7 +489,7 @@ endbit1     mac
 
 ; OPTIMIZATION:
 ;
-;           bit     #$00FF                    ; Optimization to skip the first 8 bits if they are all zeros
+;           bit     #$00FF                    ; Skip the first 8 bits if they are all zeros
 ;           bne     norm_entry
 ;           xba
 ;           jmp     skip_entry

src/blitter/Tiles00000.s

@@ -41,6 +41,7 @@ _TBSolidTile_VH
 ; This is called via a JMP (abs,x) with an extra byte on the stack that holds the bank
 ; register value.  This must be restored prior to returning
 _TBCopyDataFast
+_TBCopyDataAFast
                  tax
 _TBCopyDataFastX
 ]line            equ             0

src/render/Fast.s

@@ -146,3 +146,45 @@ GenericUnderZero
 
             plb
             rts
+
+; Simple pair of routines that copies just the tile data to the direct page workspace.  Data Bank
+; must be set to the TileData bank in entry.
+;
+; Preserves the X-register
+FastCopyTileDataA
+            ldy   TileStore+TS_TILE_ADDR,x         ; load the tile address
+            pei   DP2_TILEDATA_AND_TILESTORE_BANKS
+            plb                                    ; set to the tiledata bank
+
+]line       equ   0
+            lup   8
+            lda   tiledata+{]line*4},y
+            sta   tmp_tile_data+{]line*4}
+
+            lda   tiledata+{]line*4}+2,y
+            sta   tmp_tile_data+{]line*4}+2
+]line       equ   ]line+1
+            --^
+
+            plb
+            rts
+
+FastCopyTileDataV
+            ldy   TileStore+TS_TILE_ADDR,x         ; load the tile address
+            pei   DP2_TILEDATA_AND_TILESTORE_BANKS
+            plb                                    ; set to the tiledata bank
+
+]src        equ   7
+]dest       equ   0
+            lup   8
+            lda   tiledata+{]src*4},y
+            sta   tmp_tile_data+{]dest*4}
+
+            lda   tiledata+{]src*4}+2,y
+            sta   tmp_tile_data+{]dest*4}+2
+]src        equ   ]src-1
+]dest       equ   ]dest+1
+            --^
+            
+            plb
+            rts

src/render/Render.s

@@ -65,7 +65,7 @@ CopyTileToDPSprite
 ; must be set to the TileData bank in entry.
 ;
 ; Preserves the X-register
-CopyTileDataToDP
+CopyTileDataToDPA
 ]line       equ   0
             lup   8
             lda   tiledata+{]line*4},y

src/render/Sprite1.s

@@ -34,11 +34,11 @@ _OneSpriteFastOver0
 ; The 1-sprite dispatch prserves the X-register, so it already points to the TileStore
 
 _OneSpriteFastOverV
-            jsr   CopyTileDataToDPV
+            jsr   FastCopyTileDataV
             bra   _OneSpriteFastOver
 
 _OneSpriteFastOverA
-            jsr   CopyTileDataToDP
+            jsr   FastCopyTileDataA
 
 _OneSpriteFastOver
             lda   TileStore+TS_CODE_ADDR_HIGH,x    ; load the bank of the target code field line
@@ -86,7 +86,7 @@ _OneSpriteSlowOver0
 ; Slow variant for regular tile.
 
 _OneSpriteSlowOver
-            jsr   CopyTileDataToDP
+            jsr   CopyTileDataToDPA
 
             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.