Merge branch 'RenderingOptimization3'

.gitignore

@@ -2,3 +2,4 @@
 /gscats.lst
 /GSCats.xcodeproj/xcuserdata/qd.xcuserdatad/xcdebugger
 /loader.lst
+/terrain_e1.lst

GSCats.xcodeproj/project.pbxproj

@@ -13,6 +13,7 @@
 		700F21E01F4A3A5500D7007D /* GenerateTrigTables.py */ = {isa = PBXFileReference; lastKnownFileType = text.script.python; path = GenerateTrigTables.py; sourceTree = "<group>"; };
 		700FFAFB1F40F3BF00A442DE /* font.s */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.asm; path = font.s; sourceTree = "<group>"; };
 		7059502B1F37A0BE00BBE90F /* GenerateVRAMTable.py */ = {isa = PBXFileReference; lastKnownFileType = text.script.python; path = GenerateVRAMTable.py; sourceTree = "<group>"; };
+		705AAFA920040B0D001BB0ED /* terrain_e1.s */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.asm; path = terrain_e1.s; sourceTree = "<group>"; };
 		706DF1641F2D39F700AA6680 /* loader.s */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.asm; path = loader.s; sourceTree = "<group>"; };
 		706DF1651F2D4A8100AA6680 /* terrain.s */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.asm; path = terrain.s; sourceTree = "<group>"; };
 		7088096D1F2ECE8D00D4C950 /* GenerateRenderSpans.py */ = {isa = PBXFileReference; lastKnownFileType = text.script.python; path = GenerateRenderSpans.py; sourceTree = "<group>"; };
@@ -20,6 +21,8 @@
 		7099E3841F41022100182A82 /* gameobject.s */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.asm; path = gameobject.s; sourceTree = "<group>"; };
 		7099E3851F4107B100182A82 /* GenerateVRAMYOffset.py */ = {isa = PBXFileReference; lastKnownFileType = text.script.python; path = GenerateVRAMYOffset.py; sourceTree = "<group>"; };
 		70A80FB01F43D7F200BD34C9 /* gamemanager.s */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.asm; path = gamemanager.s; sourceTree = "<group>"; };
+		70BDCBC92006AD5F00CB51F1 /* linkerConfig */ = {isa = PBXFileReference; lastKnownFileType = text; path = linkerConfig; sourceTree = "<group>"; };
+		70BDCBCA200A99F200CB51F1 /* ParseMapFile.py */ = {isa = PBXFileReference; lastKnownFileType = text.script.python; path = ParseMapFile.py; sourceTree = "<group>"; };
 		70C073091F5BAA3E009844A9 /* collision.s */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.asm; path = collision.s; sourceTree = "<group>"; };
 		70E266E31F6F262D005AC7E4 /* circleTable.s */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.asm; path = circleTable.s; sourceTree = "<group>"; };
 		70E554C41F807ADB00F3C871 /* spritebank.s */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.asm; path = spritebank.s; sourceTree = "<group>"; };
@@ -45,6 +48,7 @@
 				706DF1641F2D39F700AA6680 /* loader.s */,
 				700FFAFB1F40F3BF00A442DE /* font.s */,
 				706DF1651F2D4A8100AA6680 /* terrain.s */,
+				705AAFA920040B0D001BB0ED /* terrain_e1.s */,
 				70C073091F5BAA3E009844A9 /* collision.s */,
 				70F086A01F4230CB002446C3 /* utility.s */,
 				700C39C51F2E5CA800C24F9C /* tables.s */,
@@ -55,12 +59,14 @@
 				70E554C41F807ADB00F3C871 /* spritebank.s */,
 				70E9D8611F2BD95400555C19 /* gscats.s */,
 				70E9D8631F2BD95400555C19 /* Makefile */,
+				70BDCBC92006AD5F00CB51F1 /* linkerConfig */,
 				70FE79D21F8814A600E0095C /* MerlinToCA65.sh */,
 				7088096D1F2ECE8D00D4C950 /* GenerateRenderSpans.py */,
 				7059502B1F37A0BE00BBE90F /* GenerateVRAMTable.py */,
 				7099E3851F4107B100182A82 /* GenerateVRAMYOffset.py */,
 				700F21E01F4A3A5500D7007D /* GenerateTrigTables.py */,
 				709175C01F60D263008FAFAB /* GenerateCircles.py */,
+				70BDCBCA200A99F200CB51F1 /* ParseMapFile.py */,
 			);
 			sourceTree = "<group>";
 		};

GenerateSpanDataTable.py

@@ -0,0 +1,25 @@
+#!/usr/bin/env python3
+
+import sys
+
+
+def main(argv):
+	
+	print ("terrainSpanDataPtr:",end="")
+	rowCount = 19
+	MAXTERRAINHEIGHT = 100	# Must match assembly code!
+	
+	for jump in range(0,MAXTERRAINHEIGHT):
+		rowCount += 1
+		if (rowCount==20):
+			print ("\n\t.word ", end="")
+			rowCount=0
+		
+		print ("SPANROWSIZE*%d" % (jump), end="")
+		if (rowCount<19):
+			print (",", end="")
+
+
+if __name__ == "__main__":
+	main(sys.argv[1:])
+

GenerateVRAMYOffset.py

@@ -5,7 +5,6 @@ import sys
 
 def main(argv):
 
-	# Prologue
 	print ("vramYOffset:",end="")
 	rowCount = 19
 	
@@ -19,6 +18,19 @@ def main(argv):
 		if (rowCount<19):
 			print (",", end="")
 
+	print ("\n\vramRowInvertedSpanLookup:",end="")
+	rowCount = 19
+	
+	for jump in range(0,200):
+		rowCount += 1
+		if (rowCount==20):
+			print ("\n\t.word ", end="")
+			rowCount=0
+		
+		print ("$%04x" % ((200-jump)*160+8192), end="")
+		if (rowCount<19):
+			print (",", end="")
+
 
 if __name__ == "__main__":
 	main(sys.argv[1:])

Makefile

@@ -16,20 +16,23 @@ ADDR=800
 PGM=gscats
 MRSPRITE=../MrSprite/mrsprite
 CHROMA=00ff00
-PALETTE=a4dffb 008800 886611 cc9933 eebb44 dd6666 ff99aa 000000 0e7db1 ffff00 ffff00 ffff00 ffff00 ffff00 ffff00 ffffff
+PALETTE=a4dffb a4dffb 008800 886611 cc9933 eebb44 dd6666 ff99aa 0e7db1 000000 ffff00 ffff00 ffff00 ffff00 ffff00 ffffff
 SPRITES=SpriteBank
+REMOTESYMBOLS=-Wl $(shell ./ParseMapFile.py *.map)
 
-all: $(PGM) loader
+all: terrain_e1 $(PGM) loader
 
 
 $(PGM):
-	@PATH=$(PATH):/usr/local/bin; $(CL65) -t apple2enh --cpu 65816 --start-addr 0000 -l$(PGM).lst $(PGM).s
+	@echo $(REMOTESYMBOLS)
+	@PATH=$(PATH):/usr/local/bin; $(CL65) -t apple2enh -C linkerConfig --cpu 65816 --start-addr 0000 -l$(PGM).lst $(REMOTESYMBOLS) $(PGM).s
 	java -jar $(AC) -d $(PGM).2mg CODEBANK
 	java -jar $(AC) -p $(PGM).2mg CODEBANK BIN 0x0000 < $(PGM)
 	java -jar $(AC) -d $(PGM).2mg $(SPRITES)00
 	java -jar $(AC) -p $(PGM).2mg $(SPRITES)00 BIN 0x0000 < $(SPRITES)00.bin
 	rm -f $(PGM)
 	rm -f $(PGM).o
+	rm -f terrain_e1.map
 	osascript V2Make.scpt $(PROJECT_DIR) $(PGM)
 
 loader:
@@ -39,11 +42,21 @@ loader:
 	rm -f loader
 	rm -f loader.o
 
+terrain_e1:
+	@PATH=$(PATH):/usr/local/bin; $(CL65) -t apple2enh -C linkerConfig --cpu 65816 --start-addr $(ADDR) -l -vm -m terrain_e1.map terrain_e1.s
+	java -jar $(AC) -d $(PGM).2mg CODEBANKE1
+	java -jar $(AC) -p $(PGM).2mg CODEBANKE1 BIN 0x800 < terrain_e1
+	rm -f terrain_e1
+	rm -f terrain_e1.o
+
 clean:
 	rm -f $(PGM)
 	rm -f $(PGM).o
 	rm -f loader
 	rm -f loader.o
+	rm -f terrain_e1.o
+	rm -f terrain_e1.map
+	rm -f terrain_e1
 
 .PHONY: art
 art:

ParseMapFile.py

@@ -0,0 +1,64 @@
+#!/usr/bin/env python3
+
+# Outputs a list of exported symbols found in an ld65 link map file
+# suitable for passing as a commandline option to define those
+# same symbols in a cl65 compile/assemble operation. This allows
+# sharing symbols between code banks without having to actually
+# link them together. Sort of a low-budget DLL mechanism. :)
+
+import sys
+
+
+def parseInputFile(filename,bank):
+
+	inputFile = open(filename)
+	contents = inputFile.readlines()
+	output = ""
+
+	# Find exported symbols
+	lineNum = 0
+	for line in contents:
+		if line.find("Exports list") != -1:
+			break
+		lineNum+=1
+
+	# Found the exports table, so look for our symbol names
+	first=1
+	for line in contents[lineNum+1:]:
+		if line.startswith("_") or line.startswith("-"):
+			continue
+		if line.startswith("Imports") or len(line)<2:
+			break
+
+		if first!=1:
+			output += ","
+		first = 0
+
+		columns = line.split()
+		if len(columns) > 3:
+			output += ("-D,%s=0x%s,-D,%s=0x%s" % (columns[0],bank+columns[1][2:],columns[3],bank+columns[4][2:]))
+		else:
+			output += ("-D,%s=0x%s" % (columns[0],bank+columns[1][2:]))
+
+	inputFile.close()
+	return output
+
+
+def main(argv):
+	commandline = ""
+	first=1
+
+	for file in argv:
+
+		if first!=1:
+			commandline += ","
+		first = 0
+
+		bank = file[-6:-4]		# Assumes filename_XX.map
+		commandline += parseInputFile(file,bank)
+
+	print (commandline)
+
+
+if __name__ == "__main__":
+	main(sys.argv[1:])

collision.s

@@ -50,7 +50,7 @@ intersectRectTerrain:
 
 	; Check Y bottom
 intersectRectTerrainBottomLoop:
-	lda terrainData,y
+	lda terrainDataFar,y
 	cmp rectParams+6
 	bpl intersectRectTerrainYep
 	dey

equates.s

@@ -26,6 +26,30 @@ SCRATCH1		= $1a
 SCRATCHL		= $19		; 16-bit version of scratch
 SCRATCHL2		= $67		; Second 16-bit scratch
 PARAM24			= $67		; 24-bit param (This is almost certainly messing up AppleSoft, but meh)
+CACHEPTR		= $6A		; General purpose cache pointer (This is almost certainly messing up AppleSoft, but meh)
+CACHEDATA		= $6C		; General purpose cache data (This is almost certainly messing up AppleSoft, but meh)
+leftScreenEdge	= $6E		; Reserved for leftScreenEdge  (This is almost certainly messing up AppleSoft, but meh)
 STACKPTR		= $70		; Cache for stack pointer in fast graphics
 SHADOWREGISTER	= $72		; Cache for shadow register in fast graphics
 STACKREGISTER	= $73		; Cache for stack register in fast graphics
+lastCompiledTerrainY = $75	; The highest Y value that the compiled renderer must handle
+
+
+
+; Terrain constants
+TERRAINWIDTH = 640		; In pixels
+MAXTERRAINHEIGHT = 100	; In pixels
+COMPILEDTERRAINROW = TERRAINWIDTH/4+4	; In words, +4 to make room for clipping jump at end of row
+VISIBLETERRAINWIDTH = TERRAINWIDTH/4	; In words- width minus jump return padding
+VISIBLETERRAINWINDOW = 80				; In words
+MAXSPANSPERROW = 15
+SPANROWBYTES = MAXSPANSPERROW*2 + 2		; In bytes
+CLIPPEDTERRAINSTACK = $3fff				; Location of stack where clipped terrain data lives
+
+; Terrain data, stored as height values 2 pixels wide (bytes)
+terrainDataFar = $02f500
+terrainData = $f500
+;	.repeat TERRAINWIDTH/2
+;	.word 0
+;	.endrepeat
+terrainDataEnd = terrainData + (TERRAINWIDTH/2 * 2)

gamemanager.s

@@ -27,9 +27,8 @@ beginGameplay:
 	jsr colorFill
 
 	; Generate, compile, and clip terrain
+	stz leftScreenEdge
 	jsr generateTerrain
-	jsr compileTerrain
-	jsr clipTerrain
 
 	; Create players
 	lda #56
@@ -45,21 +44,21 @@ beginGameplay:
 
 	jsr protectPlayers
 	jsr protectProjectiles
+	jsr prepareRowRendering
+
+	jsr compileTerrain
+	jsr clipTerrain
+	jsl renderTerrainSpans
 
 gameplayLoop:
-
 	jsr syncVBL
+	BORDER_COLOR #$0
 
-	; Render the terrain if needed
+	;;;;;;;;;;;
-	lda terrainDirty
+	; Update
-	beq gameplayLoopKbd
+	;
-	jsr renderTerrain
+	lda #1
-	stz terrainDirty
+	sta projectilesDirty
-
-	; Render players
-	jsr renderPlayers
-
-gameplayLoopKbd:
 	lda projectileActive
 	bpl gameplayLoopShotTracking	; Skip input during shots
 
@@ -67,8 +66,13 @@ gameplayLoopKbd:
 	jsr kbdScan
 
 	; Check for pause
-	lda paused
+;	lda paused
-	bne gameplayLoopEndFrame
+;	bne gameplayLoopEndFrame
+
+	bra gameplayLoopScroll
+
+gameplayLoopShotTracking:
+	jsr trackActiveShot
 
 gameplayLoopScroll:
 
@@ -91,20 +95,44 @@ gameplayLoopPower:
 
 gameplayLoopFire:
 	lda fireRequested
-	beq gameplayLoopProjectiles
+	beq gameplayLoopRender
 	jsr fire
 
-gameplayLoopProjectiles:
+gameplayLoopRender:
 	sta KBDSTROBE
+
+	;;;;;;;;;;;
+	; Render
+	;
+
+	; Render the terrain if needed
+	lda terrainDirty
+	beq gameplayLoopProjectiles
+	BORDER_COLOR #$3
+	jsl renderTerrainSpans
+	jsr renderTerrain
+
+	stz terrainDirty
+	BORDER_COLOR #$1
+
+	; Render players
+	jsr renderPlayers
+
+gameplayLoopProjectiles:
+	lda projectilesDirty
+	beq gameplayLoopProjectilesSkip
+
 	jsr unrenderProjectiles
 	jsr updateProjectilePhysics
 	jsr protectProjectiles
 	jsr renderProjectiles
+
+gameplayLoopProjectilesSkip:
 	jsr updateProjectileCollisions
 
-	lda turnRequested
+;	lda turnRequested
-	beq gameplayLoopVictoryCondition
+;	beq gameplayLoopVictoryCondition
-	jsr endTurn
+;	jsr endTurn
 
 gameplayLoopVictoryCondition:
 	lda gameOver
@@ -113,12 +141,10 @@ gameplayLoopVictoryCondition:
 
 gameplayLoopEndFrame:
 	lda quitRequested
-	beq gameplayLoop
+	beq gameplayLoopContinue
 	jmp quitGame
-
+gameplayLoopContinue:
-gameplayLoopShotTracking:
+	jmp gameplayLoop
-	jsr trackActiveShot
-	bra gameplayLoopScroll
 
 
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
@@ -218,9 +244,12 @@ endGame:
 ;
 scrollMap:
 	jsr unclipTerrain
+	jsl unrenderTerrainSpans
 	jsr unrenderPlayers
 	jsr unrenderProjectiles
 
+	jsr updateProjectilePhysics	; Good idea?
+
 	sta mapScrollPos
 	asl
 	sta leftScreenEdge
@@ -235,8 +264,10 @@ scrollMap:
 	jsr protectPlayers
 	jsr protectProjectiles
 	jsr renderPlayers
+	jsr renderProjectiles		; Prevents flicker, but ads jitter to shot tracking
 	lda #1
 	sta terrainDirty
+	stz projectilesDirty
 	rts
 
 
@@ -287,7 +318,7 @@ fire:
 
 
 basePalette:
-	.word $0aef,$0080,$0861,$0c93,$0eb4,$0d66,$0f9a,$0000,$07b,$0000,$0000,$0000,$0000,$0000,$0000,$0FFF
+	.word $0aef,$0aef,$0080,$0861,$0c93,$0eb4,$0d66,$0f9a,$007b,$0000,$0000,$0000,$0000,$0000,$0000,$0FFF
 
 
 
@@ -305,6 +336,8 @@ turnRequested:
 	.word $0000
 terrainDirty:
 	.word 1
+projectilesDirty:
+	.word 1
 activePlayer:
 	.word 0
 currentPlayer:
@@ -323,7 +356,7 @@ paused:
 ; c) Byte-distance from left edge of logical terrain to right edge of visible screen minus game object width in words
 mapScrollPos:
 	.word 0
-leftScreenEdge:
+;leftScreenEdge = $6E		; Moved to zero page for speed and cross-bank convenience
-	.word 0
+;	.word 0
 rightScreenEdge:
 	.word 160-GAMEOBJECTWIDTH/4-1

graphics.s

@@ -80,6 +80,42 @@ initSCBsLoop:
 	rts
 
 
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; enableFillMode
+; Enables fill mode for a given scanline
+;
+; X = Scan line
+;
+enableFillMode:
+	SAVE_AXY
+	BITS8
+	lda $e19d00,x
+	ora #%00100000
+	sta $e19d00,x
+	BITS16
+	RESTORE_AXY
+	rts
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; disableFillMode
+; Disables fill mode for a given scanline
+;
+; X = Scan line
+;
+; Trashes A
+
+disableFillMode:
+	SAVE_AXY
+	BITS8
+	lda $e19d00,x
+	and #%11011111
+	sta $e19d00,x
+	BITS16
+	RESTORE_AXY
+	rts
+
+
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ; setPaletteColor
 ; Set a single color in a palette
@@ -155,6 +191,22 @@ drawNumber:
 	rts
 
 
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; BORDER_COLOR
+;
+; Trashes A
+;
+.macro BORDER_COLOR color
+	SAVE_AXY
+	BITS8
+	lda BORDERCOLOR
+	and #$f0
+	ora color
+	sta BORDERCOLOR
+	BITS16
+	RESTORE_AXY
+.endmacro
+
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ; Vertical blank checkers
 ;

gscats.s

@@ -8,12 +8,15 @@
 .include "macros.s"
 .include "equates.s"
 
+.import renderTerrainSpans
+.import unrenderTerrainSpans
+.import compileTerrainSpans
+
 .org $0000
 
 
 mainBank2:
 	SYNCDBR
-
 	BITS8
 	lda #$f0
 	sta TEXTCOLOR
@@ -34,10 +37,10 @@ quitGame:
 
 .include "graphics.s"
 .include "font.s"
+.include "player.s"
 .include "terrain.s"
 .include "collision.s"
 .include "gameobject.s"
-.include "player.s"
 .include "utility.s"
 .include "tables.s"
 .include "gamemanager.s"

input.s

@@ -48,7 +48,7 @@ kbdScanRightArrow:
 	lda mapScrollPos
 	cmp #VISIBLETERRAINWIDTH-VISIBLETERRAINWINDOW
 	beq kbdScanDone
-	inc
+;	inc
 	inc
 	sta mapScrollRequested
 	rts
@@ -59,7 +59,7 @@ kbdScanLeftArrow:
 	lda mapScrollPos
 	beq kbdScanDone
 	dec
-	dec
+;	dec
 	sta mapScrollRequested
 	rts
 

linkerConfig

@@ -0,0 +1,44 @@
+FEATURES {
+	STARTADDRESS: default = $0803;
+}
+
+MEMORY {
+	ZP:     start = $0080, size = $001A,                 define = yes;
+	HEADER: start = $0000, size = $0004,      file = %O;
+	RAM:    start = %S,    size = $9600 - %S, file = %O, define = yes;
+	MOVE:   start = $0000, size = $FFFF,      file = %O, define = yes;
+	LC:     start = $D400, size = $0C00,                 define = yes;
+}
+
+SEGMENTS {
+	ZEROPAGE: load = ZP,              type = zp;
+	EXEHDR:   load = HEADER,          type = ro;
+	STARTUP:  load = RAM,             type = ro;
+	LOWCODE:  load = RAM,             type = ro;
+	CODE:     load = RAM,             type = ro;
+	RODATA:   load = RAM,             type = ro;
+	DATA:     load = RAM,             type = rw;
+	ZPSAVE:   load = RAM,             type = bss, define   = yes;
+	BSS:      load = RAM,             type = bss, define   = yes;
+	INIT:     load = MOVE, run = RAM, type = ro,  define   = yes;
+	HIGHCODE: load = MOVE, run = LC,  type = ro,  optional = yes;
+}
+
+FEATURES {
+	CONDES: segment = INIT,
+	type    = constructor,
+	label   = __CONSTRUCTOR_TABLE__,
+	count   = __CONSTRUCTOR_COUNT__;
+	CONDES: segment = RODATA,
+	type    = destructor,
+	label   = __DESTRUCTOR_TABLE__,
+	count   = __DESTRUCTOR_COUNT__;
+	CONDES: type    = interruptor,
+	segment = RODATA,
+	label   = __INTERRUPTOR_TABLE__,
+	count   = __INTERRUPTOR_COUNT__;
+}
+SYMBOLS {
+	__STACKSIZE__: value = $0800, weak = yes;
+}
+

loader.s

@@ -40,10 +40,42 @@ main:
 	; Copy code into bank 2
 	ldx fileReadLen
 	lda #2
+	ldy #0
 	jsr copyBytes
 
 	EMULATION
 
+;;;;;;;;;;;;;;;;;;;;;
+
+	; Open the E1 code file
+	jsr PRODOS
+	.byte $c8
+	.addr fileOpenCodeE1
+	bne ioError
+
+	; Load the code into bank 0
+	jsr PRODOS
+	.byte $ca
+	.addr fileRead
+	bne ioError
+
+	; Close the file
+	jsr PRODOS
+	.byte $cc
+	.addr fileClose
+
+	NATIVE
+
+	; Copy code into bank E1
+	ldx fileReadLen
+	lda #$E1
+	ldy #$800		; Must match terrain_e1 .org
+	jsr copyBytes
+
+	EMULATION
+
+;;;;;;;;;;;;;;;;;;;;;
+
 	; Open the sprite bank file
 	jsr PRODOS
 	.byte $c8
@@ -66,6 +98,7 @@ main:
 	; Copy sprites into bank 3
 	ldx fileReadLen
 	lda #3
+	ldy #0
 	jsr copyBytes
 
 	; Set up a long jump into bank 2, and
@@ -73,6 +106,7 @@ main:
 	; properly to ProDOS 8
 	lda #returnToProDOS
 	sta PRODOSRETURN
+
 	jml MAINENTRY
 
 returnToProDOS:
@@ -90,9 +124,13 @@ ioError:
 ; bottom of any other bank. Must be in native mode.
 ;
 ; X = Length of data in bytes
+; Y = Origin within bank
 ; A = Bank number of destination
 ;
 copyBytes:
+	sty copyBytesDest+1
+	sty copyBytesDest2+1
+
 	phx
 	BITS8
 	sta copyBytesDest+3
@@ -135,6 +173,13 @@ fileOpenCode:
 	.byte 0					; Result (file handle)
 	.byte 0					; Padding
 
+fileOpenCodeE1:
+	.byte 3
+	.addr codePathE1
+	.addr $9200				; 1k below BASIC.SYSTEM
+	.byte 0					; Result (file handle)
+	.byte 0					; Padding
+
 fileRead:
 	.byte 4
 	.byte 1					; File handle (we know it's gonna be 1)
@@ -156,5 +201,7 @@ fileOpenSprites:
 
 codePath:
 	pstring "/GSAPP/CODEBANK"
+codePathE1:
+	pstring "/GSAPP/CODEBANKE1"
 spritePath:
 	pstring "/GSAPP/SPRITEBANK00"

macros.s

@@ -30,7 +30,7 @@
 
 .macro DBR bankNum
 	BITS8
-	lda bankNum
+	lda #bankNum
 	pha
 	plb
 	BITS16
@@ -42,6 +42,15 @@
 	plb
 .endmacro
 
+.macro SAVE_DBR
+	phb
+	phk
+	plb
+.endmacro
+
+.macro RESTORE_DBR
+	plb
+.endmacro
 
 .macro EMULATION
 	sec		; Enable 8-bit mode
@@ -100,7 +109,7 @@
 .macro FASTGRAPHICS			;34 cycles, 12 bytes
 	sei						;2
 	phd						;4
-	sep #%00100000	;		 3   16-bit A only, to preserve X/Y
+	sep #%00100000	;		 3   8-bit A only, to preserve X/Y
 	.a8
 
 	lda STACKCTL			;5
@@ -116,7 +125,7 @@
 
 
 .macro SLOWGRAPHICS			;28 cycles, 12 bytes
-	sep #%00100000	;        3    16-bit A only, to preserve X/Y
+	sep #%00100000	;        3    8-bit A only, to preserve X/Y
 	.a8
 
 	lda STACKREGISTER		;4
@@ -168,6 +177,25 @@ nobrk:
 	pla
 .endmacro
 
+.macro BREAK_NOSTACK
+	lda breakpoint
+	beq nobrk
+	lda #1
+	sta $e1c029
+	brk
+nobrk:	
+.endmacro
+
+
+.macro HARDBRK
+	pha
+	lda #1
+	sta $e1c029
+	pla
+	brk
+.endmacro
+
+
 ;;;;;;;;;;
 ; Stack Macros
 

terrain.s

@@ -5,29 +5,22 @@
 ;
 
 
-TERRAINWIDTH = 640		; In pixels
-MAXTERRAINHEIGHT = 100	; In pixels
-COMPILEDTERRAINROW = TERRAINWIDTH/4+3	; In words, +2 to make room for clipping jump at end of row
-VISIBLETERRAINWIDTH = TERRAINWIDTH/4	; In words- width minus jump return padding
-VISIBLETERRAINWINDOW = 80				; In words
-
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ; renderTerrain
 ;
 ; No stack operations permitted here!
 ;
-; Current implementation: Unknown cycles per row
 ; Trashes all registers
 ;
 renderTerrain:
 	FASTGRAPHICS
-	lda #MAXTERRAINHEIGHT
+	lda #0
 	sta SCRATCHL2		; Row counter
-	lda #$9cff			; 4   Point stack to end of VRAM
+	lda #$5f1f			; 4   Point stack to end of affected VRAM
 	tcs					; 2
 
 	sec
-	lda #compiledTerrainEnd-VISIBLETERRAINWINDOW-3
+	lda #compiledTerrainEnd-VISIBLETERRAINWINDOW-4
 	sbc mapScrollPos
 	sta PARAML0
 
@@ -39,13 +32,28 @@ renderTerrainLoop:
 	jmp (PARAML0)
 
 renderRowComplete:
+
 	lda PARAML0
 	sec
 	sbc #COMPILEDTERRAINROW
 	sta PARAML0
-	dec SCRATCHL2
-	bne renderTerrainLoop
 
+	tsc
+	clc
+	adc #320
+	tcs
+
+	lda SCRATCHL2
+	inc
+	cmp lastCompiledTerrainY
+	beq	renderRowCont
+	bcs renderTerrainDone
+
+renderRowCont:
+	sta SCRATCHL2
+	bra renderTerrainLoop
+
+renderTerrainDone:
 	SLOWGRAPHICS
 	rts
 
@@ -120,47 +128,55 @@ craterTerrainDone:
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ; clipTerrain
 ;
+; Saves the compiled terrain data that we overwrite into
+; a buffer at $E04000. We do this by shadowing the stack
+; to that area and pushing.
 ;
 clipTerrain:
 	SAVE_AXY
 
+	; Shadow stack into $E04000 to save clipped data fast
+	tsc
+	sta STACKPTR
+	lda #CLIPPEDTERRAINSTACK
+	tcs
+
 	sec
-	lda #COMPILEDTERRAINROW*MAXTERRAINHEIGHT-3
+	lda #COMPILEDTERRAINROW*MAXTERRAINHEIGHT-4
 	sbc mapScrollPos
 	tay
-	ldx #MAXTERRAINHEIGHT
+
+	ldx #0
 
 clipTerrainLoop:
-	clc		; Compute buffer to for saved data
-	txa
-	asl
-	asl
-	adc #clippedTerrainData-4
-	sta PARAML0
-
 	lda	compiledTerrain,y
-	sta (PARAML0)	; Preserve data we're overwriting
+	pha
-	inc PARAML0
+	lda #$4cea		; NOP followed by JMP
-	inc PARAML0
-
-	and #$ff00
-	ora #$004c	; jmp in low byte
 	sta compiledTerrain,y
-	iny
 
+	iny
+	iny
 	lda	compiledTerrain,y
-	sta (PARAML0)	; Preserve data we're overwriting
+	pha
 
 	lda #renderRowComplete
 	sta compiledTerrain,y
 
 	tya
 	sec
-	sbc #COMPILEDTERRAINROW+1
+	sbc #COMPILEDTERRAINROW+2
 	tay
 
-	dex
+	inx
-	bne clipTerrainLoop
+	cpx lastCompiledTerrainY
+	bcc clipTerrainLoop
+	beq clipTerrainLoop
+
+	; Put stack back where it belongs
+	tsc
+	sta clippedTerrainStackPtr
+	lda STACKPTR
+	tcs
 
 	RESTORE_AXY
 	rts
@@ -169,41 +185,53 @@ clipTerrainLoop:
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ; unclipTerrain
 ;
+; Restores the compiled terrain data that we wrote
+; to a buffer at $E04000. We do this by mapping the stack
+; to $4000, then using stack-relative addressing to pull the data.
+; We can't pop the data because it's all stored in reverse.
 ;
+; On first move-left unclip every second row is unclipped incorrectly
 unclipTerrain:
 	SAVE_AXY
 
+	phd
+	lda #(CLIPPEDTERRAINSTACK & $ff00)
+	pha
+	pld		; Point direct page at our clip data
+
 	sec
-	lda #COMPILEDTERRAINROW*MAXTERRAINHEIGHT-3
+	lda #COMPILEDTERRAINROW*MAXTERRAINHEIGHT-4
 	sbc mapScrollPos
 	tay
-	ldx #MAXTERRAINHEIGHT
+
+	lda clippedTerrainStackPtr
+	and #$00ff
+	tax
+	inx
 
 unclipTerrainLoop:
-	clc		; Compute buffer that saved data is in
+	lda	2,x
-	txa
-	asl
-	asl
-	adc #clippedTerrainData-4
-	sta PARAML0
-
-	lda	(PARAML0)
 	sta compiledTerrain,y
-	inc PARAML0
+	iny
-	inc PARAML0
 	iny
 
-	lda	(PARAML0)
+	lda 0,x
 	sta compiledTerrain,y
 
 	tya
 	sec
-	sbc #COMPILEDTERRAINROW+1
+	sbc #COMPILEDTERRAINROW+2
 	tay
 
-	dex
+	inx
+	inx
+	inx
+	inx
+	cpx #$100		; When x hits the top of the stack, we're done
 	bne unclipTerrainLoop
 
+	pld
+
 	RESTORE_AXY
 	rts
 
@@ -216,15 +244,16 @@ unclipTerrainLoop:
 compileTerrain:
 	SAVE_AY
 
-	ldy #MAXTERRAINHEIGHT-1
+	ldy #0
 	lda #compiledTerrain
 	sta PARAML0
 
 compileTerrainLoop:
 	sty PARAML1
 	jsr compileTerrainRow
-	dey
+	iny
-	bmi compileTerrainDone
+	cpy #MAXTERRAINHEIGHT
+	beq compileTerrainDone
 
 	clc
 	lda #COMPILEDTERRAINROW
@@ -234,6 +263,7 @@ compileTerrainLoop:
 	bra compileTerrainLoop
 
 compileTerrainDone:
+	jsl compileTerrainSpans
 	RESTORE_AY
 	rts
 
@@ -369,6 +399,48 @@ compileTerrainOpcode:
 
 
 
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; prepareRowRendering:
+;
+; Set SCBs to match rendering mode of each terrain line
+;
+; Trashes SCRATCHL, SCRATCHL2
+;
+prepareRowRendering:
+	SAVE_AXY
+
+	ldx #199
+	stz SCRATCHL2
+	stz lastCompiledTerrainY
+
+prepareRowRenderingLoop:
+	lda #0
+	PLAYERPTR_Y
+	sec
+	lda playerData+GO_POSY,y
+	sbc #GAMEOBJECTHEIGHT
+
+	cmp SCRATCHL2
+	bcc prepareRowRenderingCompileMode
+	beq prepareRowRenderingCompileMode
+
+	jsr enableFillMode
+	bra prepareRowRenderingLoopNext
+
+prepareRowRenderingCompileMode:
+	jsr disableFillMode
+	inc lastCompiledTerrainY
+
+prepareRowRenderingLoopNext:
+	inc SCRATCHL2
+	dex
+	cpx #200-MAXTERRAINHEIGHT
+	bne prepareRowRenderingLoop
+
+	RESTORE_AXY
+	rts
+
+
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ; generateTerrain
 ;
@@ -382,7 +454,6 @@ generateTerrain:
 	sta SCRATCHL
 
 generateTerrainLoop:
-
 	lda sineTable,x
 
 	lsr
@@ -414,23 +485,12 @@ generateTerrainLoop:
 	rts
 
 
-
-; Terrain data, stored as height values 2 pixels wide (bytes)
-
-terrainData:
-	.repeat TERRAINWIDTH/2
-	.word 0
-	.endrepeat
-terrainDataEnd:
-
 compiledTerrain:
 	.repeat COMPILEDTERRAINROW * MAXTERRAINHEIGHT
-	.byte 0
+	.byte $00
 	.endrepeat
 compiledTerrainEnd:
 
-clippedTerrainData:
+clippedTerrainStackPtr:
-	.repeat MAXTERRAINHEIGHT
+	.word 0
-	.byte 0,0,0,0	; xx,jmp,addr
-	.endrepeat
 

terrain_e1.s

@@ -0,0 +1,306 @@
+;
+;  terrain (in E1 bank)
+;
+;  Fill-mode terrain rendering code snippets
+;  that live in bank E1 for maximum speed.
+;
+;  Created by Quinn Dunki on 7/29/17
+;
+
+.include "equates.s"
+.include "macros.s"
+
+UNRENDERCACHESTACK = $7ff	; Stack downwards from bottom of VRAM. Mirrored from $01 into $E1!
+
+OP16
+
+.org $800
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; renderTerrainSpans:
+;
+;
+renderTerrainSpans:
+	SAVE_AXY
+	SAVE_DBR
+	FASTGRAPHICS
+
+	lda #UNRENDERCACHESTACK	; Prepare unrender cache
+	tcs
+
+	lda #MAXTERRAINHEIGHT-1
+	sec
+	sbc lastCompiledTerrainY
+
+renderTerrainSpansLoop:
+	sta PARAML1
+
+	; Find row data
+	lda PARAML1
+	asl		; Shifts must match SPANROWBYTES
+	asl
+	asl
+	asl
+	asl
+	tay
+	lda terrainSpanData,y
+	sta SCRATCHL			; Track span color
+
+	; Find VRAM row
+	lda PARAML1
+	asl
+	tax
+	lda vramRowInvertedSpanLookup,x
+	tax
+	
+	adc #160
+	sta PARAML0				; Watch for end of VRAM row
+
+	; Find span that intersects left logical edge
+	lda #0
+	clc
+
+renderTerrainRowSpansFindLeftLoop:
+	adc terrainSpanData+2,y
+	cmp leftScreenEdge
+	bpl renderTerrainRowSpansFoundLeft
+	iny
+	iny
+	sta CACHEDATA
+	lda SCRATCHL
+	eor #%110000	; Toggle span color cache
+	sta SCRATCHL
+	lda CACHEDATA
+	bra renderTerrainRowSpansFindLeftLoop
+
+renderTerrainRowSpansFoundLeft:
+	sec
+	sbc leftScreenEdge	; A now holds remainder of leftmost span
+						; and Y points to leftmost span
+
+renderTerrainRowSpansLoop:
+	sta SCRATCHL2
+
+	; Set fill mode byte for this span
+	lda SCRATCHL
+	sta VRAMBANK,x
+
+	; Cache the index we wrote to so we can erase later
+	phx
+
+	; Advance to end of span
+	clc
+	txa
+	adc SCRATCHL2
+	cmp PARAML0
+	bpl renderTerrainRowSpansDone
+	tax
+
+	; Prepare for next span
+	iny
+	iny
+	lda SCRATCHL
+	eor #%110000	; Toggle span color cache
+	sta SCRATCHL
+
+	lda terrainSpanData+2,y
+	bra renderTerrainRowSpansLoop
+
+renderTerrainRowSpansDone:
+	lda PARAML1
+	dec
+	bne renderTerrainSpansLoop
+
+	pea 0		; Null-terminate the unrender cache
+	SLOWGRAPHICS
+	RESTORE_DBR
+	RESTORE_AXY
+	rtl
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; unrenderTerrainSpans:
+;
+;
+
+unrenderTerrainSpans:
+	SAVE_AXY
+	SAVE_DBR
+	phb
+	pea $e1e1		; Work entirely in bank E1 for all local read/writes
+	plb
+	plb
+
+	ldy #UNRENDERCACHESTACK-1
+
+unrenderTerrainSpansLoop:
+	lda 0,y
+	beq unrenderTerrainSpansDone
+	tax
+	lda #0
+	sta a:0,x
+	dey
+	dey
+	bra unrenderTerrainSpansLoop
+
+unrenderTerrainSpansDone:
+	plb
+	RESTORE_DBR
+	RESTORE_AXY
+	rtl
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; compileTerrainSpans:
+;
+;
+compileTerrainSpans:
+	pha
+	SAVE_DBR
+	lda #0
+
+compileTerrainSpansLoop:
+	sta PARAML1
+	jsr compileTerrainSpansRow
+	lda PARAML1
+	inc
+	cmp #MAXTERRAINHEIGHT
+	bne compileTerrainSpansLoop
+
+	RESTORE_DBR
+	pla
+	rtl
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; compileTerrainSpansRow:
+;
+; PARAML1 = Row index (bottom relative)
+;
+; Trashes SCRATCHL
+;
+compileTerrainSpansRow:
+	SAVE_AXY
+
+	lda PARAML1
+	asl		; Shifts must match SPANROWBYTES
+	asl
+	asl
+	asl
+	asl
+	clc
+	adc #terrainSpanData
+	sta SCRATCHL
+
+	ldy #0
+	ldx #TERRAINWIDTH-2
+
+	lda terrainDataFar,x	; Figure out start value for row
+	cmp PARAML1
+	bcs compileTerrainSpansRowInitGreen
+
+compileTerrainSpansRowInitBlack:
+	lda #$0010		; First span is black
+	sta (SCRATCHL)	; Initialize the row
+	inc SCRATCHL
+	inc SCRATCHL
+	ldy #-1
+
+compileTerrainSpansRowBlackStart:
+	iny
+
+compileTerrainSpansRowBlackLoop:
+	lda terrainDataFar,x
+	cmp PARAML1
+	bcs compileTerrainSpansBlackEnd
+	dex
+	dex
+	iny
+	cpx #-2
+	bne compileTerrainSpansRowBlackLoop
+
+compileTerrainSpansBlackEnd:
+	tya				; Store this span's length
+	sta (SCRATCHL)
+	inc SCRATCHL
+	inc SCRATCHL
+
+	dex				; Begin searching for next span
+	dex
+	cpx #0
+	bmi compileTerrainSpansRowDone
+	ldy #0
+	bra compileTerrainSpansRowGreenStart
+
+compileTerrainSpansRowDone:
+	RESTORE_AXY
+	rts
+
+compileTerrainSpansRowInitGreen:
+	lda #$0020		; First span is green
+	sta (SCRATCHL)	; Initialize the row
+	inc SCRATCHL
+	inc SCRATCHL
+	ldy #-1
+
+compileTerrainSpansRowGreenStart:
+	iny
+
+compileTerrainSpansRowGreenLoop:
+	lda terrainDataFar,x
+	cmp PARAML1
+	bcc compileTerrainSpansGreenEnd
+	dex
+	dex
+	iny
+	cpx #-2
+	bne compileTerrainSpansRowGreenLoop
+
+compileTerrainSpansGreenEnd:
+	tya				; Store this span's length
+	sta (SCRATCHL)
+	inc SCRATCHL
+	inc SCRATCHL
+
+	dex				; Begin searching for next span
+	dex
+	cpx #0
+	bmi compileTerrainSpansRowDone
+	ldy #0
+	bra compileTerrainSpansRowBlackStart
+
+terrainSpanData:
+	.repeat MAXTERRAINHEIGHT
+		.word 0		; Start value (0=BG)
+		.repeat MAXSPANSPERROW
+		.word 0		; Length (in bytes)
+		.endrepeat
+	.endrepeat
+
+
+vramRowInvertedSpanLookup:
+	.word $9d00,$9c60,$9bc0,$9b20,$9a80,$99e0,$9940,$98a0,$9800,$9760,$96c0,$9620,$9580,$94e0,$9440,$93a0,$9300,$9260,$91c0,$9120
+	.word $9080,$8fe0,$8f40,$8ea0,$8e00,$8d60,$8cc0,$8c20,$8b80,$8ae0,$8a40,$89a0,$8900,$8860,$87c0,$8720,$8680,$85e0,$8540,$84a0
+	.word $8400,$8360,$82c0,$8220,$8180,$80e0,$8040,$7fa0,$7f00,$7e60,$7dc0,$7d20,$7c80,$7be0,$7b40,$7aa0,$7a00,$7960,$78c0,$7820
+	.word $7780,$76e0,$7640,$75a0,$7500,$7460,$73c0,$7320,$7280,$71e0,$7140,$70a0,$7000,$6f60,$6ec0,$6e20,$6d80,$6ce0,$6c40,$6ba0
+	.word $6b00,$6a60,$69c0,$6920,$6880,$67e0,$6740,$66a0,$6600,$6560,$64c0,$6420,$6380,$62e0,$6240,$61a0,$6100,$6060,$5fc0,$5f20
+	.word $5e80,$5de0,$5d40,$5ca0,$5c00,$5b60,$5ac0,$5a20,$5980,$58e0,$5840,$57a0,$5700,$5660,$55c0,$5520,$5480,$53e0,$5340,$52a0
+	.word $5200,$5160,$50c0,$5020,$4f80,$4ee0,$4e40,$4da0,$4d00,$4c60,$4bc0,$4b20,$4a80,$49e0,$4940,$48a0,$4800,$4760,$46c0,$4620
+	.word $4580,$44e0,$4440,$43a0,$4300,$4260,$41c0,$4120,$4080,$3fe0,$3f40,$3ea0,$3e00,$3d60,$3cc0,$3c20,$3b80,$3ae0,$3a40,$39a0
+	.word $3900,$3860,$37c0,$3720,$3680,$35e0,$3540,$34a0,$3400,$3360,$32c0,$3220,$3180,$30e0,$3040,$2fa0,$2f00,$2e60,$2dc0,$2d20
+	.word $2c80,$2be0,$2b40,$2aa0,$2a00,$2960,$28c0,$2820,$2780,$26e0,$2640,$25a0,$2500,$2460,$23c0,$2320,$2280,$21e0,$2140,$20a0
+
+
+
+
+.export renderTerrainSpans
+.export unrenderTerrainSpans
+.export compileTerrainSpans
+
+
+; Suppress some linker warnings - Must be the last thing in the file
+.SEGMENT "ZPSAVE"
+.SEGMENT "EXEHDR"
+.SEGMENT "STARTUP"
+.SEGMENT "INIT"
+.SEGMENT "LOWCODE"