diff --git a/bin/vasmarm_std_win32.exe b/bin/vasmarm_std_win32.exe
new file mode 100644
index 0000000..678a256
Binary files /dev/null and b/bin/vasmarm_std_win32.exe differ
diff --git a/doom-fire.asm b/doom-fire.asm
index 155f269..68dfa81 100644
--- a/doom-fire.asm
+++ b/doom-fire.asm
@@ -479,31 +479,22 @@ plot_horizontal_line:
 	add r0, r2, r7
 	subs r0, r0, #1				; dest_x += (rnd & 3)-1
 	movlt r0, #0				; or MOD ScreenWidth
-
-	; update colour with some randomness
-;	and r7, r7, #1				; rnd & 1
 	subs r4, r4, r7, lsr #1		; colour -= rnd & 1
 .endm
 
 .macro PLOT_PIXEL_BIT
-	add r6, r12, r0, lsr #1		; r6 = dest_start + x DIV 2
-
-	ldrb r11, [r6]				; load screen byte
-	tst r0, #1					; odd or even pixel?
-	andeq r11, r11, #0xF0		; mask out left hand pixel
-	orreq r11, r11, r4			; mask in colour as left hand pixel
-	andne r11, r11, #0x0F		; mask out right hand pixel
-	orrne r11, r11, r4, lsl #4	; mask in colour as right hand pixel
-	strb r11, [r6]				; store screen byte
+	orr r11, r4, r4, lsl #4
+	strb r11, [r12, r0]			; dest_start + x
+	strb r11, [r6, r0]			; dest_start + x
 .endm
 
 ; DOOM FIRE!
 ; R0 = x
 ; R1 = y
 ; R2 = x
-; R3 = y
+; R3 = end of screen address (for loop termination)
 ; R4 = pixel colour
-; R5 = screen byte
+; R5 = screen word
 ; R6 = ptr
 ; R7 = rnd temp
 ; R8 = seed
@@ -515,9 +506,9 @@ do_fire:
 	str lr, [sp, #-4]!
 
 	ldr r8, rnd_seed			; seed
-	mov r9, #1					; bit
+	ldr r9, bits_mask			; bit
 
-	mov r3, #Screen_Height - 63
+	mov r3, #Screen_Height - 123
 
 	; R10 = ptr to start of source line
 	add r10, r12, r3, lsl #7	; r10 = screen_addr + y * 128
@@ -529,7 +520,8 @@ do_fire:
 	add r3, r3, r2, lsl #5	; r10 += y * 32 = y * 160
 
 	; R12 = ptr to start of dest line
-	sub r12, r10, #Screen_Stride
+	sub r12, r10, #Screen_Stride*2
+	add r6, r12, #Screen_Stride
 
 	.1:
 	mov r2, #0
@@ -539,7 +531,7 @@ do_fire:
 	RND
 
 	; source is contiguous
-	; read source word = 8x pixels
+	; read source word = 4x pixels
 	ldr r5, [r10], #4
 
 	; Byte 0 Left pixel
@@ -551,17 +543,6 @@ do_fire:
 	PLOT_PIXEL_BIT
 	add r2, r2, #1
 
-	; Byte 0 Right pixel
-	mov r4, r5, lsr #4
-	ands r4, r4, #0x0F
-	moveq r0, r2
-	beq .4
-	mov r7, r8, lsr #4
-	DO_RANDOM_BIT
-	.4:
-	PLOT_PIXEL_BIT
-	add r2, r2, #1
-
 	; Byte 1 Left pixel
 	mov r4, r5, lsr #8
 	ands r4, r4, #0x0F
@@ -573,17 +554,6 @@ do_fire:
 	PLOT_PIXEL_BIT
 	add r2, r2, #1
 
-	; Byte 1 Right pixel
-	mov r4, r5, lsr #12
-	ands r4, r4, #0x0F
-	moveq r0, r2
-	beq .6
-	mov r7, r8, lsr #12
-	DO_RANDOM_BIT
-	.6:
-	PLOT_PIXEL_BIT
-	add r2, r2, #1
-
 	; Byte 2 Left pixel
 	mov r4, r5, lsr #16
 	ands r4, r4, #0x0F
@@ -595,17 +565,6 @@ do_fire:
 	PLOT_PIXEL_BIT
 	add r2, r2, #1
 
-	; Byte 2 Right pixel
-	mov r4, r5, lsr #20
-	ands r4, r4, #0x0F
-	moveq r0, r2
-	beq .8
-	mov r7, r8, lsr #20
-	DO_RANDOM_BIT
-	.8:
-	PLOT_PIXEL_BIT
-	add r2, r2, #1
-
 	; Byte 3 Left pixel
 	mov r4, r5, lsr #24
 	ands r4, r4, #0x0F
@@ -617,21 +576,13 @@ do_fire:
 	PLOT_PIXEL_BIT
 	add r2, r2, #1
 
-	; Byte 2 Right pixel
-	movs r4, r5, lsr #28
-	moveq r0, r2
-	beq .10
-	mov r7, r8, lsr #28
-	DO_RANDOM_BIT
-	.10:
-	PLOT_PIXEL_BIT
-	add r2, r2, #1
-
-	cmp r2, #Screen_Width
+	cmp r2, #Screen_Stride
 	blt .2
 
 	; Next line
-	add r12, r12, #Screen_Stride
+	add r10, r10, #Screen_Stride
+	add r6, r6, #2*Screen_Stride
+	add r12, r12, #2*Screen_Stride
 	cmp r12, r3
 	blt .1
 
@@ -642,6 +593,9 @@ do_fire:
 rnd_seed:
 	.long 0x87654321
 
+bits_mask:
+	.long 0x11111111
+
 rnd:
 ; enter with seed in R0 (32 bits), R1 (1 bit in least significant bit)
 ; R2 is used as a temporary register.
diff --git a/doom-fire.txt b/doom-fire.txt
new file mode 100644
index 0000000..c03029a
--- /dev/null
+++ b/doom-fire.txt
@@ -0,0 +1,61 @@
+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!
diff --git a/make.bat b/make.bat
index 687a106..45bbaec 100644
--- a/make.bat
+++ b/make.bat
@@ -1,3 +1,3 @@
 @echo off
-..\..\vasm\vasmarm_std_win32.exe -L compile.txt -m250 -Fbin -opt-adr -o build\doomfire.bin doom-fire.asm
+bin\vasmarm_std_win32.exe -L compile.txt -m250 -Fbin -opt-adr -o build\doomfire.bin doom-fire.asm
 if %ERRORLEVEL%==0 copy build\doomfire.bin "..\..\Arculator_V2.0_Windows\hostfs\doomfire,ff8"