microoptimization for update_cell

[lifeboot.git] / lifeboot.asm

diff --git a/lifeboot.asm b/lifeboot.asm
index 07d9c5a56e900f5b895e521b4e3a763bb6b782ae..67e7ce99b56247f0e328aefbec5388a59c8d71af 100644 (file)
--- a/lifeboot.asm
+++ b/lifeboot.asm
@@ -1,83 +1,128 @@
 bits 16
 bits 16

-org 0x7c00
+org 7c00h

 
 

-%define ORG 0x7c00
-%define VGA 0xb8000
+; =============================================================================
+
+; settings --------------------------------------------------------------------
+
+%define DEAD ' '            ; char to represent a dead cell
+%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
 
 %define COLS 80
 %define ROWS 25

+%define VGAPGSZ 1000h       ; (bytes)

 
 

-%define DEAD ' '            ; char to represent a dead cell
-%define ALIVE '#'           ; char to represent an alive cell
-%define PRINT_COLOR 0x07    ; grey on black
+; memory layout ---------------------------------------------------------------
+
+%define ORG 7c00h
+%define VGA 0b8000h
+%define DAT ORG + 512
+
+; static variables ------------------------------------------------------------

 
 

-%define GRID        ORG + 512               ; current cell grid (80 * 25 bytes)
-%define NEXT_GRID   GRID + COLS * ROWS      ; next cell grid (80 * 25 bytes)
-%define XSS         NEXT_GRID + COLS * ROWS ; xs() state
+%define xss DAT             ; xs() state (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
 ; initialize segment registers

+mov ax, VGA >> 4
+mov es, ax                  ; es:0 -> video memory
+

 xor ax, ax
 xor ax, ax

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

 
 ; set stack pointers
 
 ; set stack pointers

-mov sp, ORG

 mov bp, ORG
 mov bp, ORG

+mov sp, bp

 
 ; clear direction flag
 cld
 
 ; disable cursor
 
 ; clear direction flag
 cld
 
 ; disable cursor

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

 
 

-; setup -----------------------------------------------------------------------
+; initialize xss
+.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

 
 

-setup:
+mov [xss], dx

 
 

-; initialize xss
-mov ah, 0x00                ; get
-int 0x1a                    ;  system time
-mov [XSS], dx               ; cx:dx = number of clock ticks since midnight
+; initialize currvgapg
+mov word [currvgapg], 0

 
 

-; initialize grid
-mov di, GRID                ; es:di -> GRID
-mov cx, COLS * ROWS         ; for each cell
+; start simulation from random state ------------------------------------------

 
 

-.initgrid:                  ; do {
-    call xs                 ;     [XSS] = rand
-    mov al, DEAD            ;     al = DEAD (likely)
+start:

 
 

-    test [XSS], 0b11
+call init_grid
+mov word [iter], 0

 
 

-    jnz  .initgrid_nz       ;     if ([XSS] % 4 == 0)
-    mov al, ALIVE           ;         al = ALIVE
+.next_state:

 
 

-.initgrid_nz:
-    stosb                   ;     [es:(di++)] = al
+    call vsync_wait
+    call update_grid

 
 

-    loop .initgrid          ; } while (--cx)
+    call vsync_wait
+    call flip_vgapg

 
 

-call print_grid
+    inc word [iter]
+    cmp word [iter], ITER_LIM
+    jl .next_state

 
 

-jmp halt
+jmp start                   ; reset
+
+; =============================================================================

 
 ; functions -------------------------------------------------------------------
 
 
 ; functions -------------------------------------------------------------------
 

+; vsync_wait() - wait for display to enter the next retrace cycle -------------
+
+; clobbers al, dx
+
+vsync_wait:
+
+mov dx, 3dah                ; input status #1 register
+
+.wait_retrace_end:
+    in al, dx
+    test al, 08h            ; vertical retrace bit
+    jnz .wait_retrace_end
+
+.wait_retrace_start:
+    in al, dx
+    test al, 08h
+    jz .wait_retrace_start
+
+ret
+

 ; xs() - xorshift pseudorandom number generator -------------------------------
 
 ; output:
 ; xs() - xorshift pseudorandom number generator -------------------------------
 
 ; output:

-; [XSS]     = updated state
+; [xss]     = updated state
+; bx        = [xss]

 
 ; clobbers bx, dx
 
 xs:
 
 
 ; clobbers bx, dx
 
 xs:
 

-mov bx, [XSS]
+mov bx, [xss]

 mov dx, bx
 
 shl dx, 1                   ; dx = xss << 1
 mov dx, bx
 
 shl dx, 1                   ; dx = xss << 1


@@ -90,38 +135,205 @@ mov dx, bx
 
 shl dx, 10                  ; dx = xss'' << 10
 xor bx, dx                  ; bx = xss'' ^ (xss'' << 10)
 
 shl dx, 10                  ; dx = xss'' << 10
 xor bx, dx                  ; bx = xss'' ^ (xss'' << 10)

-mov [XSS], bx
+mov [xss], bx

 
 ret
 
 
 ret
 

-; print_grid() - print GRID cells to screen -----------------------------------
+; init_grid() - initialize grid cells randomly --------------------------------

 
 

-; (*) this function expects ds = 0 when called and sets es = 0 before
-;     returning. this is simpler because in the rest of the program we want the
-;     segment registers to be zeroed anyway.
+; clobbers ax, bx, cx, dx, di

 
 

-; clobbers ax, cx, di, si
+init_grid:

 
 

-print_grid:
+mov ah, PRINT_COLOR         ; ah = color attribute
+mov di, [currvgapg]         ; es:di -> grid start
+mov cx, COLS * ROWS         ; for each cell

 
 

-mov ax, VGA >> 4
-xor di, di
+.write_cell:                ; do {
+    call xs                 ;     bx = random value
+    mov al, DEAD            ;     al = DEAD (likely)

 
 

-mov es, ax                  ; es:di -> VGA
-mov si, GRID                ; ds:si -> GRID
+    test bx, 0b11

 
 

-mov ah, PRINT_COLOR         ; ah = color attribute
+    jnz  .nz                ;     if (bx % 4 == 0)
+    mov al, ALIVE           ;         al = ALIVE

 
 

-mov cx, COLS * ROWS         ; for each cell
+.nz:
+    stosw                   ;     [es:di] = ax, di += 2
+
+    loop .write_cell        ; } while (--cx)
+
+ret
+
+; 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
+
+mov di, VGAPGSZ
+xor di, si                  ; es:di -> next page
+
+mov cx, COLS * ROWS - 1     ; for each cell (index)

 
 

-.print_cell:                ; do {
+.update_loop:               ; do {

 
 

-    lodsb                   ;     al = [ds:(si++)]
-    stosw                   ;     [es:di] = ax ; di += 2
+    call update_cell

 
 

-    loop .print_cell        ; } while (--cx)
+    dec cx
+    jns .update_loop        ; } while (--cx > 0)

 
 

-mov es, cx                  ; es = 0

 ret
 
 ; halt() - stop program execution ---------------------------------------------
 ret
 
 ; halt() - stop program execution ---------------------------------------------


@@ -132,7 +344,7 @@ cli                         ; disable interrupts
 hlt
 jmp halt
 
 hlt
 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