From 1cf1a42b0eaf807e69608979772d09dbfe00b9fe Mon Sep 17 00:00:00 2001
From: "B. Watson" <urchlay@slackware.uk>
Date: Tue, 1 Nov 2022 01:27:07 -0400
Subject: fastmasks, ~15% speedup

---
 Makefile |  2 +-
 dla.s    | 41 ++++++++++++++++++++++++++++++-----------
 2 files changed, 31 insertions(+), 12 deletions(-)

diff --git a/Makefile b/Makefile
index fb969a8..3109cd9 100644
--- a/Makefile
+++ b/Makefile
@@ -14,7 +14,7 @@ clean:
 	rm -f dla.xex dlatbl.s dla.list dla.labels *.o
 
 test: all
-	atari800 -turbo dla.xex
+	atari800 dla.xex
 
 %.xex: %.s
 	$(CL65) $(CL65FLAGS) -t none -o $@ $<
diff --git a/dla.s b/dla.s
index 7904af7..3c61064 100644
--- a/dla.s
+++ b/dla.s
@@ -495,6 +495,8 @@ unplot:
 ;  sta part_y
 ;  rts
 
+oob:
+ rts
 ;;; Subroutine: drunkwalk
 ;;; Walk the point around randomly until it either is
 ;;; adjacent to a set pixel or goes out of bounds.
@@ -547,33 +549,46 @@ dontplot:
 
  ; check neighbors. used to be a subroutine, inlined it.
  ; also inlined plotsetup here.
- ; (-1,0)
- dec cursor_x
  ldx cursor_y
  lda lineaddrs_l,x
  sta pixptr
  lda lineaddrs_h,x
  sta pixptr+1
+
+ ; 3/4 of the time, we can use a faster code path, check
+ ; (-1,0) and (1,0) at the same time. this happens only when
+ ; both pixels lie within the same byte.
+ ldx cursor_x
+ lda fastmasks,x
+ beq slow_x
+ ldy xoffsets,x
+ and (pixptr),y
+ bne stick
+ beq check_y
+slow_x:
+ ; (-1,0)
  ldx cursor_x
+ dex
  ldy xoffsets,x
  lda xmasks,x
  and (pixptr),y
  bne stick
  ; (1,0)
  inx
- stx cursor_x
  inx
  ldy xoffsets,x
  lda xmasks,x
  and (pixptr),y
  bne stick
+check_y:
  ; (0,-1)
- dec cursor_y
- ldx cursor_y
- lda lineaddrs_l,x
- sta pixptr
- lda lineaddrs_h,x
- sta pixptr+1
+ ldx cursor_y      ;3
+ dex               ;2
+ lda lineaddrs_l,x ;5
+ sta pixptr        ;3
+ lda lineaddrs_h,x ;5
+ sta pixptr+1      ;3
+                   ;=21
  ldx cursor_x
  ldy xoffsets,x
  lda xmasks,x
@@ -582,7 +597,7 @@ dontplot:
  bne stick
  ; (0,1)
  tya
- ora #$40 ; add 64
+ ora #$40 ; add 64, AKA 2 screen lines
  tay
  lda (pixptr),y
  and pixmask
@@ -590,7 +605,6 @@ dontplot:
  jmp drunkwalk ; too far for a branch
 
 stick:
-oob:
  rts
 
 ;;; Subroutine: drawseed
@@ -800,6 +814,11 @@ xmasks:
   .byte $80,$40,$20,$10,$08,$04,$02,$01
  .endrep
 
+fastmasks:
+ .repeat 32
+  .byte $00,$a0,$50,$28,$14,$0a,$05,$00
+ .endrep
+
 ;;; display list
  ; ANTIC opcodes
  blank8 = $70
-- 
cgit v1.2.3