working?

dither4.s

@@ -129,42 +129,10 @@ wait_loop:
     sta scanline
 
 each_scanline:
-    ; it'll fire on unused lines, but harmlessly
-
-    .macro fake1
-        jmp fake
-        ; audio sample; low nybble
-        ldx #0
-        lda (sample_ptr,x)
-        sta temp2
-
-        ; high nybble
-        lsr a
-        lsr a
-        lsr a
-        lsr a
-        ; set the volume-only bit
-        ora #$10
-        pha
-
-        ; low nybble
-        lda temp2
-        and #$0f
-        ; set the volume-only bit
-        ora #$10
-        lda #0
-        pha
-    .endmacro
-
-    .macro audio_play
-        ldy #0
-        lda (sample_ptr),y
-        sta AUDC1
-    .endmacro
-
     .macro inner_scanline
-        ldy scanline
-        inc scanline
+        ; it'll fire on unused lines, but harmlessly
+        ldy scanline ; 3 cyc
+        inc scanline ; 5 cyc
 
         ; Leisurely memory fetches
         lda palette1,y
@@ -183,16 +151,25 @@ each_scanline:
         sty COLPF2
     .endmacro
 
+    .macro audio_play
+        ldy VCOUNT ; 4 cycles
+        lda (sample_ptr),y ; 5/6 cyc
+        sta AUDC1 ; 4 cyc
+    .endmacro
+
     inner_scanline
     audio_play
 
     .macro audio_inc
         ; Increment sample ptr
-        inc sample_ptrl
-        bne audio_cont
-
-        inc sample_ptrh
+        clc
+        lda sample_ptrl
+        adc #130
+        sta sample_ptrl
         lda sample_ptrh
+        adc #0
+        sta sample_ptrh
+
         cmp #.hibyte(audio_samples_end)
         bne audio_cont
 
@@ -210,15 +187,12 @@ each_scanline:
 
 
     inner_scanline
-    audio_inc
 
     lda VCOUNT
     cmp #130 ;#130
     bne each_scanline
-    ;bne foo
+    audio_inc
     jmp wait_start
-    ;foo:
-    ;jmp each_scanline
 
 .endproc