diff --git a/mandel.s b/mandel.s
index 10782b8..f4862c6 100644
--- a/mandel.s
+++ b/mandel.s
@@ -254,10 +254,10 @@ color_map:
     ; 5 cycles either way
     .if bitnum < 8
         lda arg1                 ; 3 cyc
-        and #(1 << bitnum)       ; 2 cyc
+        and #(1 << (bitnum))       ; 2 cyc
     .else
         lda arg1 + 1             ; 3 cyc
-        and #(1 << (bitnum - 8)) ; 2 cyc
+        and #(1 << ((bitnum) - 8)) ; 2 cyc
     .endif
     bne one ; 2 cyc
 
@@ -434,43 +434,48 @@ loop:
     rts
 keep_going:
 
+    .macro quick_exit arg
+        .local keep_going
+        lda arg + 1
+        cmp #(4 << 4)
+        bmi keep_going
+        rts
+    keep_going:
+    .endmacro
+
     ; 4.12: (-8 .. +7.9)
     ; zx = zx_2  - zy_2  + cx
-    sub16 zx, zx_2 + 2, zy_2 + 2
+    sub16 zx, zx_2, zy_2
     add16 zx, zx, cx
+    quick_exit zx
 
     ; zy = zx_zy + zx_zy + cy
-    add16 zy, zx_zy + 2, zx_zy + 2
+    add16 zy, zx_zy, zx_zy
     add16 zy, zy, cy
 
-    ; 8.24: (-128 .. +127.9)
     ; zx_2 = zx * zx
-    imul16 zx_2, zx, zx
+    imul16_round zx_2, zx, zx, 4
+    quick_exit dist
 
     ; zy_2 = zy * zy
-    imul16 zy_2, zy, zy
+    imul16_round zy_2, zy, zy, 4
+    quick_exit dist
 
     ; zx_zy = zx * zy
-    imul16 zx_zy, zx, zy
+    imul16_round zx_zy, zx, zy, 4
+    quick_exit dist
 
     ; dist = zx_2 + zy_2
-    add32 dist, zx_2, zy_2
-
-    ; if dist >= 4 break, else continue iterating
-    lda dist + 3
-    cmp #4
-    bmi still_in
-    rts
-still_in:
-
-    shift_round_16 zx_2, 4
-    shift_round_16 zy_2, 4
-    shift_round_16 zx_zy, 4
+    add16 dist, zx_2, zy_2
+    quick_exit dist
 
     ; if may be in the lake, look for looping output with a small buffer
     ; as an optimization vs running to max iters
     jmp loop
 
+peace_out:
+    rts
+
 .endproc
 
 .macro zoom_factor dest, src, zoom, aspect
@@ -607,9 +612,13 @@ done:
     sta ox + 1
     sta oy
     sta oy + 1
+
+    ; zoom = 2x
+    lda #1
     sta zoom
 
     ; Disable display DMA
+    lda #0
     sta DMACTL
 
     ; zero the range from framebuffer_top to framebuffer_end