diff --git a/mandel.s b/mandel.s
index d824193..a8f3cac 100644
--- a/mandel.s
+++ b/mandel.s
@@ -433,7 +433,7 @@ fill_masks:
 
 ; lookup table for top byte -> PORTB value for bank-switch
 ;.align 256 ; warning - if not aligned this will cost an extra cycle sometimes
-bankswitch:
+bank_switch_table:
     .repeat 256, i
         .byte ((i & $c0) >> 5) | $c1
     .endrepeat
@@ -460,9 +460,9 @@ bankswitch:
     sta ptr + 1  ; 3 cyc
     
     ; top 2 bits are the table bank selector
-    ldx arg2          ; 3 cyc
-    lda bank_switch,x ; 4 cyc
-    sta PORTB         ; 4 cyc
+    ldx arg2                ; 3 cyc
+    lda bank_switch_table,x ; 4 cyc
+    sta PORTB               ; 4 cyc
 
 
     ; copy the entry into output
@@ -612,20 +612,20 @@ inner_loop:
     ; h1*256*(h2*256 + l2) + l1*(h2*256 + l2)
     ; h1*h2*256*256 + h1*l2*256 + h2*l1*256 + l1*l2
 
-    imul8 result, arg1, arg2
+    imul8xe result, arg1, arg2
     lda #0
     sta result + 2
     sta result + 3
 
-    imul8 inter, arg1 + 1, arg2
+    imul8xe inter, arg1 + 1, arg2
     add16 result + 1, result + 1, inter
     add_carry result + 3
 
-    imul8 inter, arg1, arg2 + 1
+    imul8xe inter, arg1, arg2 + 1
     add16 result + 1, result + 1, inter
     add_carry result + 3
 
-    imul8 inter, arg1 + 1, arg2 + 1
+    imul8xe inter, arg1 + 1, arg2 + 1
     add16 result + 2, result + 2, inter
 
     ; In case of negative inputs, adjust high word
@@ -1147,6 +1147,8 @@ zero_byte_loop:
 
 .proc start
 
+    jsr imul8xe_init
+
     ; ox = 0; oy = 0; zoom = 0
     ; count_frames = 0; count_pixels = 0
     lda #0