doom-fire.txt

DOOM Fire
~~~

 function spreadFire(src) {
    var rand = Math.round(Math.random() * 3.0) & 3;
    var dst = src - rand + 1;
    firePixels[dst - FIRE_WIDTH ] = firePixels[src] - (rand & 1);
 }

 function doFire() {
    for(x=0 ; x < FIRE_WIDTH; x++) {
        for (y = 1; y < FIRE_HEIGHT; y++) {
            spreadFire(y * FIRE_WIDTH + x);
        }
    }
 }

If word is 8x pixels: screen word = 0x87654321
Then could propagate heat in one go
word -= rand & 0x11111111
But need to clamp to zero.

word & 0x88888888 >> 3 | word & 0x44444444 >> 2 | word & 0x22222222 >> 1 | word & 0x11111111
Would give us a rand_mask with zeros where pixels were 0.

bits_mask = 0x11111111

and temp, screen_word, bits_mask, lsl #3    ; 0x88888888
mov rand_mask, temp, lsr #3
and temp, screen_word, bits_mask, lsl #2    ; 0x44444444
orr rand_mask, rand_mask, temp, lsr #2
and temp, screen_word, bits_mask, lsl #1    ; 0x22222222
orr rand_mask, rand_mask, temp, lsr #1
and temp, screen_word, bits_mask
orr rand_mask, rand_mask, temp

and temp, rand, rand_mask                   ; rand & rand_mask
sub screen_word, screen_word, temp          ; screen_word -= rand & rand_mask

Seems like this would be faster.
In terms of writing pixels - can only move +1/-1 from current X so one word
worth of destination can only affect +1/-1 words either side.
Keep a running buffer of 3 words?!

	; and temp, screen_word, bits_mask, lsl #3    ; 0x88888888
	and r11, r5, r9, lsl #3
	; mov rand_mask, temp, lsr #3
	mov r7, r11, lsr #3

	and r11, r5, r9, lsl #2     ; 0x44444444
	orr r7, r7, r11, lsr #2
	and r11, r5, r9, lsl #1     ; 0x22222222
	orr r7, r7, r11, lsr #1
	and r11, r5, r9				; 0x11111111
	orr r7, r7, r11

	and r11, r8, r7				; rand & rand_mask
	sub r5, r5, r11				; screen_word -= rand & rand_mask

Reduce resolution to 160x64...
Would still only give about 15 cycles per pixel!