microoptimization for update_cell

[lifeboot.git] / lifeboot.asm

diff --git a/lifeboot.asm b/lifeboot.asm
index 2d1542f7dfe16a5219f1eab056d2c3b0757a0d2b..67e7ce99b56247f0e328aefbec5388a59c8d71af 100644 (file)
--- a/lifeboot.asm
+++ b/lifeboot.asm
@@ -1,43 +1,46 @@
 bits 16
 bits 16

-org 0x7c00
+org 7c00h

 
 ; =============================================================================
 
 ; settings --------------------------------------------------------------------
 
 %define DEAD ' '            ; char to represent a dead cell
 
 ; =============================================================================
 
 ; settings --------------------------------------------------------------------
 
 %define DEAD ' '            ; char to represent a dead cell

-%define ALIVE '#'           ; char to represent an alive cell
-%define PRINT_COLOR 0x07    ; grey on black
-%define WAIT_DELAY 0x02     ; 0.131072 seconds
+%define ALIVE '.'           ; char to represent an alive cell
+%define PRINT_COLOR 07h     ; grey on black
+%define ITER_LIM 500        ; reset the simulation after this # iterations

 
 ; constants -------------------------------------------------------------------
 
 %define COLS 80
 %define ROWS 25
 
 ; constants -------------------------------------------------------------------
 
 %define COLS 80
 %define ROWS 25

-%define VGAPGSZ (COLS * ROWS * 2)
+%define VGAPGSZ 1000h       ; (bytes)

 
 ; memory layout ---------------------------------------------------------------
 
 
 ; memory layout ---------------------------------------------------------------
 

-%define ORG 0x7c00
-%define VGA 0xb8000
+%define ORG 7c00h
+%define VGA 0b8000h

 %define DAT ORG + 512
 
 ; static variables ------------------------------------------------------------
 
 %define xss DAT             ; xs() state (word)
 %define DAT ORG + 512
 
 ; static variables ------------------------------------------------------------
 
 %define xss DAT             ; xs() state (word)

-%define currvgapg xss + 2   ; current vga page near pointer (word)
+%define currvgapg DAT + 2   ; current vga page near pointer (word)
+%define iter DAT + 4        ; current iteration # (word)

 
 ; =============================================================================
 
 ; entry point -----------------------------------------------------------------
 
 
 ; =============================================================================
 
 ; entry point -----------------------------------------------------------------
 

+entry:
+

 ; initialize segment registers
 mov ax, VGA >> 4
 ; initialize segment registers
 mov ax, VGA >> 4

-mov ds, ax
-mov es, ax
+mov es, ax                  ; es:0 -> video memory

 
 xor ax, ax
 
 xor ax, ax

-mov ss, ax
+mov ds, ax                  ; ds = 0
+mov ss, ax                  ; ss = 0

 
 ; set stack pointers
 mov bp, ORG
 
 ; set stack pointers
 mov bp, ORG


@@ -47,63 +50,65 @@ mov sp, bp
 cld
 
 ; disable cursor
 cld
 
 ; disable cursor

-mov ch, 0x3f                ; cursor start and options
-mov ah, 0x01                ; set text-mode cursor shape
-int 0x10                    ; video services
+mov ch, 3fh                 ; cursor start and options
+mov ah, 01h                 ; set text-mode cursor shape
+int 10h                     ; bios video services

 
 ; initialize xss
 
 ; initialize xss

-mov ah, 0x00                ; get
-int 0x1a                    ;  system time
-mov [xss], dx               ; cx:dx = number of clock ticks since midnight
+.set_xss:
+    mov ah, 00h             ; get system time
+    int 1ah                 ; bios time services
+                            ; cx:dx = # clock ticks since midnight
+    or dx, dx
+    jz .set_xss             ; wait for something meaningful
+
+mov [xss], dx

 
 ; initialize currvgapg
 
 ; initialize currvgapg

-mov word [currvgapg], VGAPGSZ
+mov word [currvgapg], 0

 
 

-; setup -----------------------------------------------------------------------
+; start simulation from random state ------------------------------------------

 
 

-setup:
+start:

 
 call init_grid
 
 call init_grid

-call flip_vga_page
-call delay
+mov word [iter], 0

 
 

-jmp setup
+.next_state:

 
 

-; =============================================================================
-
-; functions -------------------------------------------------------------------
+    call vsync_wait
+    call update_grid

 
 

-; delay() - suspend program execution temporarily -----------------------------
+    call vsync_wait
+    call flip_vgapg

 
 

-; clobbers ah, cx, dx
+    inc word [iter]
+    cmp word [iter], ITER_LIM
+    jl .next_state

 
 

-delay:
+jmp start                   ; reset

 
 

-mov cx, WAIT_DELAY          ; cx:dx = interval in microseconds
-mov dx, 0
-
-mov ah, 0x86
-int 0x15                    ; wait
+; =============================================================================

 
 

-ret
+; functions -------------------------------------------------------------------

 
 

-; vsync() - wait for display to enter the next VBlank cycle -------------------
+; vsync_wait() - wait for display to enter the next retrace cycle -------------

 
 

-; clobbers ax, dx
+; clobbers al, dx

 
 

-vsync:
+vsync_wait:

 
 

-mov dx, 0x3da               ; input status #1 register
+mov dx, 3dah                ; input status #1 register

 
 

-.wait_on:
+.wait_retrace_end:

     in al, dx
     in al, dx

-    test al, 0x08           ; vertical retrace bit
-    jnz .wait_on
+    test al, 08h            ; vertical retrace bit
+    jnz .wait_retrace_end

 
 

-.wait_off:
+.wait_retrace_start:

     in al, dx
     in al, dx

-    test al, 0x08
-    jz .wait_off
+    test al, 08h
+    jz .wait_retrace_start

 
 ret
 
 
 ret
 


@@ -111,6 +116,7 @@ ret
 
 ; output:
 ; [xss]     = updated state
 
 ; output:
 ; [xss]     = updated state

+; bx        = [xss]

 
 ; clobbers bx, dx
 
 
 ; clobbers bx, dx
 


@@ -144,34 +150,189 @@ mov di, [currvgapg]         ; es:di -> grid start
 mov cx, COLS * ROWS         ; for each cell
 
 .write_cell:                ; do {
 mov cx, COLS * ROWS         ; for each cell
 
 .write_cell:                ; do {

-    call xs                 ;     [xss] = rand
+    call xs                 ;     bx = random value

     mov al, DEAD            ;     al = DEAD (likely)
 
     mov al, DEAD            ;     al = DEAD (likely)
 

-    test word [xss], 0b11
+    test bx, 0b11

 
 

-    jnz  .nz                ;     if ([xss] % 4 == 0)
+    jnz  .nz                ;     if (bx % 4 == 0)

     mov al, ALIVE           ;         al = ALIVE
 
 .nz:
     mov al, ALIVE           ;         al = ALIVE
 
 .nz:

-    stosw                   ;     [es:di] = ax ; di += 2
+    stosw                   ;     [es:di] = ax, di += 2

 
     loop .write_cell        ; } while (--cx)
 
 ret
 
 
     loop .write_cell        ; } while (--cx)
 
 ret
 

-; flip_vga_page() - flip active display page ----------------------------------
+; flip_vgapg() - flip active display page -------------------------------------
+
+; output:
+; [currvgapg] ^= VGAPGSZ
+
+; clobbers ax
+
+flip_vgapg:
+
+xor word [currvgapg], VGAPGSZ
+
+setnz al                    ; al = !(currvgapg == 0)
+mov ah, 05h                 ; select active display page
+int 10h                     ; bios video services
+
+ret
+
+; alive_neighbours() - get number of alive adjacent cells ---------------------
+
+; input:
+; es:si     -> current grid
+; cx        -> cell index
+
+; output:
+; ax = # alive neighbours
+
+; clobbers ax, bx, dx
+
+alive_neighbours:
+enter 5, 0
+
+; [bp - 1]: row
+; [bp - 2]: col
+; [bp - 3]: neighbours
+; [bp - 4]: i
+; [bp - 5]: j
+
+mov ax, cx                  ; ax = idx
+mov bl, COLS                ; bl = COLS
+div bl                      ; al = idx / COLS, ah = idx % COLS
+
+mov [bp - 1], al            ; row = idx / COLS
+mov [bp - 2], ah            ; row = idx % COLS
+mov byte [bp - 3], 0        ; neighbours = 0
+
+mov byte [bp - 4], -1       ; i = -1
+.i:                         ; do {
+
+    mov byte [bp - 5], -1           ; j = -1
+.j:                                 ; do {
+
+        mov al, [bp - 4]
+        or al, [bp - 5]                 ; if (!i && !j)
+        jz .continue                    ;     continue
+
+        mov al, [bp - 1]
+        add al, [bp - 4]                ; al = row + i
+
+        cmp al, 0                       ; if (row + i < 0)
+        jl .continue                    ;     continue
+
+        cmp al, ROWS                    ; if (row + i >= ROWS)
+        jge .continue                   ;     continue
+
+        mov ah, [bp - 2]
+        add ah, [bp - 5]                ; ah = col + j
+
+        cmp ah, 0                       ; if (col + j < 0)
+        jl .continue                    ;     continue
+
+        cmp ah, COLS                    ; if (col + j >= COLS)
+        jge .continue                   ;     continue
+
+        movzx bx, ah                    ; bx = col + j
+
+        mov ah, COLS
+        mul ah                          ; ax = (row + i) * COLS
+
+        add bx, ax                      ; bx = cell index
+        shl bx, 1                       ; bx = grid offset
+
+        mov dx, [es:si + bx]            ; dl = cell state
+
+        cmp dl, ALIVE                   ; if (!ALIVE)
+        jne .continue                   ;     continue
+
+        inc byte [bp - 3]               ; neighbours++
+
+.continue:
+        inc byte [bp - 5]
+        cmp byte [bp - 5], 1
+        jle .j                      ; } while (++j <= 1)
+
+    inc byte [bp - 4]
+    cmp byte [bp - 4], 1
+    jle .i                  ; } while (++i <= 1)
+
+movzx ax, byte [bp - 3]     ; ax = # alive neighbours
+
+leave
+ret
+
+; update_cell() - write next cell state ---------------------------------------
+
+; input:
+; es:si     -> current grid
+; es:di     -> next grid
+; cx        -> cell index
+
+; output:
+; [es:di + cx * 2] = updated cell state
+
+; clobbers ax, bx, dx
+
+update_cell:
+
+call alive_neighbours       ; ax = # alive neighbours
+
+mov bx, cx
+shl bx, 1
+
+mov dx, [es:si + bx]        ; dl = cell current state
+
+cmp dl, ALIVE               ; if (ALIVE) {
+jne .else
+
+shr ax, 1
+xor ax, 1                   ;     if (n != 2 || n != 3)
+jnz .dead                   ;         return DEAD (likely)
+
+jmp .alive                  ;     return ALIVE
+
+.else:                      ; }
+
+cmp ax, 3                   ; if (n != 3)
+jne .dead                   ;     return DEAD
+
+.alive:                     ; return ALIVE
+mov dl, ALIVE
+jmp .write
+
+.dead:
+mov dl, DEAD
+
+.write:
+mov [es:di + bx], dx
+
+ret
+
+; update_grid() - write next grid (inactive vga page) -------------------------
+
+; clobbers di, si, ax, bx, cx, dx
+
+update_grid:
+
+mov si, [currvgapg]         ; es:si -> current page

 
 

-; clobbers ax, dx
+mov di, VGAPGSZ
+xor di, si                  ; es:di -> next page

 
 

-flip_vga_page:
+mov cx, COLS * ROWS - 1     ; for each cell (index)

 
 

-call vsync
+.update_loop:               ; do {

 
 

-xor word [currvgapg], VGAPGSZ   ; flip currvgapg
-setnz al                        ; al = !(currvgapg == 0)
+    call update_cell

 
 

-mov ah, 0x05                ; select active display page
-int 0x10                    ; video services
+    dec cx
+    jns .update_loop        ; } while (--cx > 0)

 
 ret
 
 
 ret
 


@@ -186,4 +347,4 @@ jmp halt
 ; =============================================================================
 
 times 510 - ($ - $$) db 0   ; fill remaining bytes with zeroes
 ; =============================================================================
 
 times 510 - ($ - $$) db 0   ; fill remaining bytes with zeroes

-dw 0xaa55                   ; mbr magic byte
+dw 0aa55h                   ; mbr magic byte