WebSVN – Kolibri OS – Diff – /programs/demos/firework/trunk/firework.asm

 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ; A firework demo
 ; Programmed by Yaniv LEVIATHAN
 ; http://yaniv.leviathanonline.com
 ; Converted to DexOS, By Dex
 ; Converted to KolibriOS, By Asper
 ; Optimized for KolibriOS, By Diamond
 ; Assemble with
 ; c:fasm firework.asm firework.kex
-; NOTE: Needs MMX & SSE
+; NOTE: Needs MMX & SSE, optionally AVX
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 use32
-	org	0x0
+        org     0x0
-	db	'MENUET00'	; 8 byte id
+        db      'MENUET01'      ; 8 byte id
-	dd	38		; required os
+        dd      0x01            ; version
-	dd	STARTAPP	; program start
+        dd      STARTAPP        ; program start
-	dd	I_END		; program image size
+        dd      I_END           ; program image size
-	dd	0x100000	; required amount of memory
+        dd      E_END           ; required amount of memory
-	dd	0x00000000	; reserved=no extended header
+        dd      stacktop        ; reserved=no extended header
+        dd      0, 0
-include '../../../macros.inc'
-include "aspapi.inc"
+include '../../../macros.inc'
-SCREEN_WIDTH   equ    320
+SCREEN_WIDTH   = 320
+SCREEN_HEIGHT  = 200
+SIMD equ SSE
+SIMD_BYTES = 8
+assert SCREEN_WIDTH mod SIMD_BYTES = 0
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-SCREEN_HEIGHT  equ    200
+; Global defines
-;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-; Global defines
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-   NUM_PARTS   = 150
+NUM_PARTS      = 150
-   X_OFFSET   = 0
+X_OFFSET       = 0
-   Y_OFFSET   = 4
+Y_OFFSET       = 4
-   X_SPEED_OFFSET   = 8
+X_SPEED_OFFSET = 8
-   Y_SPEED_OFFSET   = 12
+Y_SPEED_OFFSET = 12
-   COLOR_OFFSET   = 16
+COLOR_OFFSET   = 16
-   PART_SIZE   = 20
+PART_SIZE      = 20
+macro shade
+{
+local .lop
+if SIMD eq SSE
-macro draw_window
+        mov     ecx, SCREEN_WIDTH * SCREEN_HEIGHT / SIMD_BYTES
-{
+        mov     edi, buffer
-local x, xsize, y, ysize, areacolor, caption
+        movq    mm1, qword [sub_mask]
-x = 100
+  .lop:
+        movq    mm0, [edi]
-y = 70
+        psubusb mm0, mm1
-xsize = SCREEN_WIDTH+9
+        movq    [edi], mm0
-ysize = SCREEN_HEIGHT+4
+        add     edi, SIMD_BYTES
+        loop    .lop
-areacolor = 0x54224466
+else if SIMD eq AVX
-caption = labelt
+        mov     ecx, SCREEN_WIDTH * SCREEN_HEIGHT / SIMD_BYTES
-	mov	eax, 12 		; function 12:tell os about windowdraw
+        mov     edi, buffer
-	mov	ebx, 1			; 1, start of draw
+        vmovdqa xmm1, xword [sub_mask]
-	int	0x40
+  .lop:
+        vmovdqa xmm0, [edi]
-	; DRAW WINDOW
+        vpsubusb xmm0, xmm0, xmm1
-	mov	eax, 48
+        vmovdqa [edi], xmm0
-	mov	ebx, 4
+        add     edi, SIMD_BYTES
-	int	0x40
+        loop    .lop
-	lea	ecx, [y*65536+ysize+eax]
+else if SIMD eq AVX2
+        mov     ecx, SCREEN_WIDTH * SCREEN_HEIGHT / SIMD_BYTES
-	xor	eax, eax		; function 0 : define and draw window
+        mov     edi, buffer
+        vmovdqa ymm1, yword [sub_mask]
-	mov	ebx, x*65536+xsize	; [x start] *65536 + [x size]
+  .lop:
-	mov	edx, areacolor		 ; color of work area RRGGBB
+        vmovdqa ymm0, [edi]
-	mov	edi, caption
+        vpsubusb ymm0, ymm0, ymm1
-	int	0x40
-;  start_draw_window 100,70,SCREEN_WIDTH+9,SCREEN_HEIGHT+29,0x04224466,labelt;, 14;labellen-labelt
-  end_draw_window
+        vmovdqa [edi], ymm0
-}
+        add     edi, SIMD_BYTES
-macro mmx_shade
+        loop    .lop
-{
+else if SIMD eq AVX512
-      mov ecx, SCREEN_WIDTH*SCREEN_HEIGHT/8
+        mov     ecx, SCREEN_WIDTH * SCREEN_HEIGHT / SIMD_BYTES
-      mov  edi,buffer
+        mov     edi, buffer
-      movq mm1, [sub_mask]
+        vmovdqa64 zmm1, zword [sub_mask]
-.lop:
+  .lop:
-      movq mm0, [edi]
+        vmovdqa64 zmm0, [edi]
-      psubusb mm0, mm1
+        vpsubusb zmm0, zmm0, zmm1
+        vmovdqa64 [edi], zmm0
-      movq [edi], mm0
+        add     edi, SIMD_BYTES
-      add edi, 8
+        loop    .lop
-      loop .lop
+end if
 }
-macro mmx_blur_prepare
+macro blur_prepare
 {
-      mov ecx, (SCREEN_WIDTH*SCREEN_HEIGHT-330*2)/8
+        mov     ecx, (SCREEN_WIDTH * SCREEN_HEIGHT - SCREEN_WIDTH * 2 - SIMD_BYTES*2) / SIMD_BYTES
-      mov edi,buffer + 328
+        mov     edi, buffer + SCREEN_WIDTH + SIMD_BYTES
+}
-}
+macro blur
-macro mmx_blur
+{
-{
+local .lop
-local .lop
+if SIMD eq SSE
 .lop:
-      movq mm0, [edi]
+        movq    mm0, [edi]
-      movq mm1, [edi+1]
+        movq    mm1, [edi + 1]
-      movq mm2, [edi-1]
+        movq    mm2, [edi - 1]
-      movq mm3, mm0
+        movq    mm3, mm0
-      movq mm4, [edi-SCREEN_WIDTH]
+        movq    mm4, [edi - SCREEN_WIDTH]
-      movq mm5, [edi+SCREEN_WIDTH]
+        movq    mm5, [edi + SCREEN_WIDTH]
-      pavgb mm0, mm1 ; mm0 = avg(cur,cur+1)
+        pavgb   mm0, mm1 ; mm0 = avg(cur,cur+1)
-      pavgb mm3, mm2 ; mm3 = avg(cur,cur-1)
+        pavgb   mm3, mm2 ; mm3 = avg(cur,cur-1)
+        pavgb   mm4, mm5 ; mm4 = avg(cur+width,cur-width)
+        pavgb   mm3, mm4 ; mm3 = avg(avg(cur,cur-1),avg(cur+width,cur-width))
+        pavgb   mm0, mm3 ; mm0 = avg(avg(cur,cur+1),
+        movq    [edi], mm0
+        add     edi, SIMD_BYTES
+        loop    .lop
+else if SIMD eq AVX
+.lop:
+        vmovdqa xmm0, [edi]
+        vmovdqa xmm1, xmm0
+        vmovdqa xmm2, [edi - SCREEN_WIDTH]
+        vpavgb  xmm0, xmm0, [edi + 1]
-      pavgb mm4, mm5 ; mm4 = avg(cur+320,cur-320)
+        vpavgb  xmm1, xmm1, [edi - 1]
+        vpavgb  xmm2, xmm2, [edi + SCREEN_WIDTH]
+        vpavgb  xmm1, xmm1, xmm2
+        vpavgb  xmm0, xmm0, xmm1
+        vmovdqa [edi], xmm0
+        add     edi, SIMD_BYTES
+        loop    .lop
+else if SIMD eq AVX2
+.lop:
+        vmovdqa ymm0, [edi]
+        vmovdqa ymm1, ymm0
+        vmovdqa ymm2, [edi - SCREEN_WIDTH]
+        vpavgb  ymm0, ymm0, [edi + 1]
+        vpavgb  ymm1, ymm1, [edi - 1]
+        vpavgb  ymm2, ymm2, [edi + SCREEN_WIDTH]
+        vpavgb  ymm1, ymm1, ymm2
+        vpavgb  ymm0, ymm0, ymm1
+        vmovdqa [edi], ymm0
+        add     edi, SIMD_BYTES
+        loop    .lop
+else if SIMD eq AVX512
+.lop:
+        vmovdqa64 zmm0, [edi]
+        vmovdqa64 zmm1, zmm0
+        vmovdqa64 zmm2, [edi - SCREEN_WIDTH]
+        vpavgb  zmm0, zmm0, [edi + 1]
+        vpavgb  zmm1, zmm1, [edi - 1]
+        vpavgb  zmm2, zmm2, [edi + SCREEN_WIDTH]
-      pavgb mm3, mm4 ; mm3 = avg(avg(cur,cur-1),avg(cur+320,cur-320))
+        vpavgb  zmm1, zmm1, zmm2
+        vpavgb  zmm0, zmm0, zmm1
       pavgb mm0, mm3 ; mm0 = avg(avg(cur,cur+1),
-      movq [edi], mm0
+        vmovdqa64 [edi], zmm0
-      add edi, 8
+        add     edi, SIMD_BYTES
-      loop .lop
+        loop    .lop
+end if
 }
+macro blur_right
-macro mmx_blur_right
+{
-{
+local .lop
-local .lop
+if SIMD eq SSE
-.lop:
+  .lop:
-      movq mm0, [edi]
+        movq    mm0, [edi]
-      movq mm1, [edi+1]
+        movq    mm1, [edi + 1]
+        movq    mm2, [edi + SCREEN_WIDTH]
+        movq    mm3, [edi + SCREEN_WIDTH + 1]
+        pavgb   mm0, mm1
+        pavgb   mm3, mm2
+        pavgb   mm0, mm3
+        movq    [edi], mm0
+        add     edi, SIMD_BYTES
+        loop    .lop
+else if SIMD eq AVX
+  .lop:
+        vmovdqa xmm0, [edi]
+        vmovdqu xmm1, [edi + SCREEN_WIDTH + 1]
+        vpavgb  xmm2, xmm0, [edi + 1]
+        vpavgb  xmm3, xmm1, [edi + SCREEN_WIDTH]
+        vpavgb  xmm4, xmm2, xmm3
+        vmovdqa [edi], xmm4
+        add     edi, SIMD_BYTES
+        loop    .lop
+else if SIMD eq AVX2
-      movq mm2, [edi+SCREEN_WIDTH]
+  .lop:
+        vmovdqa ymm0, [edi]
+        vmovdqu ymm1, [edi + SCREEN_WIDTH + 1]
+        vpavgb  ymm2, ymm0, [edi + 1]
+        vpavgb  ymm3, ymm1, [edi + SCREEN_WIDTH]
+        vpavgb  ymm4, ymm2, ymm3
+        vmovdqa [edi], ymm4
+        add     edi, SIMD_BYTES
+        loop    .lop
+else if SIMD eq AVX512
+  .lop:
+        vmovdqa64 zmm0, [edi]
-      movq mm3, [edi+SCREEN_WIDTH+1]
+        vmovdqu64 zmm1, [edi + SCREEN_WIDTH + 1]
+        vpavgb  zmm2, zmm0, [edi + 1]
-      pavgb mm0, mm1
+        vpavgb  zmm3, zmm1, [edi + SCREEN_WIDTH]
-      pavgb mm3, mm2
+        vpavgb  zmm4, zmm2, zmm3
-      pavgb mm0, mm3
+        vmovdqa64 [edi], zmm4
-      movq [edi], mm0
+        add     edi, SIMD_BYTES
-      add edi, 8
+        loop    .lop
-      loop .lop
+end if
 }
 STARTAPP:
 init_palette:
-   mov edi, pal
+        mov     edi, pal
-	xor	eax, eax
+        xor     eax, eax
 red_loop:
-	stosd
+        stosd
-	stosd
+        stosd
-	add	eax, 0x040000
+        add     eax, 0x040000
-	and	eax, 0xFFFFFF
+        and     eax, 0xFFFFFF
-	jnz	red_loop
+        jnz     red_loop
-	mov	eax, 63*4*65536
+        mov     eax, 63*4 SHL 16
 @@:
-	stosd
+        stosd
+        stosd
-	stosd
+        add     ax, 0x0404
-	add	ax, 0x0404
+        jnc     @b
-	jnc	@b
 ;zero_buffer:
-;zero_buffer:
+        mov     ecx, SCREEN_WIDTH * SCREEN_HEIGHT / 4
-   mov ecx, SCREEN_WIDTH*SCREEN_HEIGHT/4
+;       mov     edi, buffer
-;   mov edi,buffer
+        xor     eax, eax
-   xor eax, eax
+        rep     stosd
-   rep stosd
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ; Main Functions
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 virtual at esp
-global_x	dd	?
+        global_x        dd ?
-global_y	dd	?
+        global_y        dd ?
+        seed            dd ?
+end virtual
+        rdtsc
+        push    eax      ; seed
+        push    100 * 64 ; global_y
+        push    160 * 64 ; global_x
+   jmp MAIN
+red:
+        mcall   9, proc_info, -1
+x = 100
+y = 70
+xsize = SCREEN_WIDTH+9
+ysize = SCREEN_HEIGHT+4
+areacolor = 0x54224466
+        mov     eax, 12                 ; function 12:tell os about windowdraw
-seed		dd	?
+        mov     ebx, 1                  ; 1, start of draw
+        int     0x40
-end virtual
+        mov     eax, 48
+        mov     ebx, 4
-   db 0x0f, 0x31
+        int     0x40
-	push	eax	; seed
+        lea     ecx, [(y SHL 16) + ysize + eax]
-	push	100*64	; global_y
+        xor     eax, eax                ; function 0 : define and draw window
-	push	160*64	; global_x
+        mov     ebx, (x SHL 16) + xsize ; [x start] *65536 + [x size]
+        mov     edx, areacolor          ; color of work area RRGGBB
-   jmp MAIN
+        mov     edi, window_title
+        int     0x40
+        mov     eax, 12                 ; end of redraw
-red:
+        mov     ebx, 2
-	mcall	9,proc_info,-1
+        int     0x40
-   draw_window
 MAIN:
-	test	[proc_info.wnd_state], 0x04
+        test    [proc_info.wnd_state], 0x04
-	jnz	still
+        jnz     still
-   mov ecx, NUM_PARTS
+        mov     ecx, NUM_PARTS
-   mov ebp, particles
+        mov     ebp, particles
-   .advance_particles:
+  .advance_particles:
-      mov eax, [ebp+X_OFFSET]
+        mov     eax, [ebp + X_OFFSET]
-      mov ebx, [ebp+Y_OFFSET]
+        mov     ebx, [ebp + Y_OFFSET]
-      sar eax, 6
+        sar     eax, 6
-      sar ebx, 6
+        sar     ebx, 6
-      cmp eax, 5
+        cmp     eax, 5
-      jb .new_particle
-      cmp eax, SCREEN_WIDTH-5;315
-      jge .new_particle
-      cmp ebx, 5
-      jb .new_particle
+        jb      .new_particle
-      cmp ebx, SCREEN_HEIGHT-5;195
+        cmp     eax, SCREEN_WIDTH - 5
-      jl .part_ok
+        jge     .new_particle
+        cmp     ebx, 5
-   .new_particle:
-      call init_particle
-      jmp .advance_particles
+        jb      .new_particle
-   .part_ok:
+        cmp     ebx, SCREEN_HEIGHT - 5
-;      mov edi, eax
- ;     add edi,buffer
-;      mov eax, SCREEN_WIDTH
-;      mul ebx
+        jl      .part_ok
-	imul	edi, ebx, SCREEN_WIDTH
-      mov dl, [ebp+COLOR_OFFSET]
+  .new_particle:
-      mov [buffer+eax+edi], dl
+        call    init_particle
+        jmp     .advance_particles
-;      mov eax, [ebp+X_OFFSET]
-;      mov ebx, [ebp+Y_OFFSET]
+  .part_ok:
-;      add eax, [ebp+X_SPEED_OFFSET]
+        imul    edi, ebx, SCREEN_WIDTH
-;      add ebx, [ebp+Y_SPEED_OFFSET]
+        mov     dl, [ebp+COLOR_OFFSET]
-;      mov [ebp+X_OFFSET], eax
+        mov     [buffer+eax+edi], dl
-;      mov [ebp+Y_OFFSET], ebx
-	mov	eax, [ebp+X_SPEED_OFFSET]
+        mov     eax, [ebp+X_SPEED_OFFSET]
-	add	[ebp+X_OFFSET], eax
+        add     [ebp+X_OFFSET], eax
-	mov	eax, [ebp+Y_SPEED_OFFSET]
+        mov     eax, [ebp+Y_SPEED_OFFSET]
-	add	[ebp+Y_OFFSET], eax
+        add     [ebp+Y_OFFSET], eax
-      db 0x0f, 0x31
+        rdtsc
-      and al, 0x7F
+        and     al, 0x7F
-      jnz .dont_inc_y_speed
+        jnz     .dont_inc_y_speed
-      inc dword [ebp+Y_SPEED_OFFSET]
+        inc     dword [ebp+Y_SPEED_OFFSET]
-      .dont_inc_y_speed:
+    .dont_inc_y_speed:
-      add ebp, PART_SIZE
+        add     ebp, PART_SIZE
-   loop .advance_particles
+        loop    .advance_particles
-   mmx_shade
+        shade
-; jmp .copy_buffer_to_video
+;       jmp     .copy_buffer_to_video
-   mmx_blur_prepare
+        blur_prepare
-   test dword [blur_right_flag] , 0x800000
+        test    dword [blur_right_flag] , 0x800000
-   jnz .do_blur_right
-   mmx_blur
-   db 0x0f, 0x31
-   and al, 1
+        jnz     .do_blur_right
-   jz .blur_ok
+        blur
-   jmp .dont_blur
-   .do_blur_right:
-   mmx_blur_right
+        rdtsc
-   .blur_ok:
+        and     al, 1
-   add dword [blur_right_flag], 0x1000
+        jz      .blur_ok
-   .dont_blur:
+        jmp     .dont_blur
+    .do_blur_right:
-   .copy_buffer_to_video:
-    ;    mov     eax, 18 ;@WAITVSYNC();
+        blur_right
-    ;    mov     ebx, 14
+    .blur_ok:
-    ;    int     0x40
+        add     dword [blur_right_flag], 0x1000
-	mov	eax, 48
+    .dont_blur:
-	mov	ebx, 4
-	int	0x40
+    .copy_buffer_to_video:
-	lea	edx, [5*65536+eax]
+        mcall   48, 4
-	mov	eax, 65 ;copyfard(0xA000,0,screen,0,16000);
+        lea     edx, [(5 SHL 16) + eax]
-	mov	ebx, buffer;dword [screen]
-	mov	ecx, SCREEN_WIDTH*65536+SCREEN_HEIGHT ;ecx = w*65536+h
+        mov     eax, 65
-;	mov	edx, 5*65536+25 ;edx = x*65536+y
+        mov     ebx, buffer
-	push	8
+        mov     ecx, (SCREEN_WIDTH SHL 16) + SCREEN_HEIGHT
-	pop	esi
+        push    8
-	;mov	esi, 8
+        pop     esi
-	mov	edi, pal
+        mov     edi, pal
-	xor	ebp, ebp
+        xor     ebp, ebp
-	int	0x40
+        int     0x40
 still:
-	mov	eax, 11 	    ; Test if there is an event in the queue.
+        mov     eax, 11             ; Test if there is an event in the queue.
-	int	0x40
+        int     0x40
-	dec	eax		      ; redraw request ?
+        dec     eax                   ; redraw request ?
-	jz	red
-	dec	eax		      ; key in buffer ?
-	jz	key
-	dec	eax		      ; button in buffer ?
-	jz	button
-	jmp	MAIN
-key:
+        jz      red
-	mov	eax, 2
+        dec     eax                   ; key in buffer ?
-	int	0x40
+        jz      key
-;	cmp	ah, 1		    ; Test Esc in Scan
-;	je	close_app
-	cmp	ah, 27		    ; Test Esc in ASCII
-	je	close_app
-	jmp	MAIN
-button:
-; we have only one button, close
-;	mov	eax, 17 	    ; Get pressed button code
-;	int	0x40
-;	cmp	ah, 1		    ; Test x button
-;	je	close_app
-;	jmp	MAIN
-; fall through to close_app
-fail:
-	; Type something here.
-close_app:
-    mov  eax,-1 		 ; close this program
-    int  0x40
-init_particle:
-   db 0x0f, 0x31
-   and al, 0x1F
-jnz .dont_re_init_globals
-   ; init x
-   call rand
-   cdq
-   ;xor dx, dx
-   mov ebx, SCREEN_WIDTH
-   div ebx
-   shl edx, 6
-   mov [4+global_x], edx
-   ; init y
-   call rand
-   cdq
-   ;xor dx, dx
-   mov ebx, SCREEN_HEIGHT
-   div ebx
-   shl edx, 6
-   mov [4+global_y], edx
-.dont_re_init_globals:
-   ; init x
-   mov eax, [4+global_x]
-   mov [ebp+X_OFFSET], eax
-   ; init y
-   mov eax, [4+global_y]
-   mov [ebp+Y_OFFSET], eax
-   ; init x speed
-   call rand
+        dec     eax                   ; button in buffer ?
+        jz      button
+        jmp     MAIN
+key:
+        mov     eax, 2
+        int     0x40
+;       cmp     ah, 1               ; Test Esc in Scan
+;       je      close_app
+        cmp     ah, 27              ; Test Esc in ASCII
+        je      close_app
+        jmp     MAIN
+button:
+; we have only one button, close
+close_app:
+        mov     eax, -1         ; close this program
+        int     0x40
+init_particle:
+        rdtsc
+        and     al, 0x1F
+        jnz     .dont_re_init_globals
+        ; init x
+        call    rand
+        cdq
+        ;xor dx, dx
+        mov     ebx, SCREEN_WIDTH
+        div     ebx
+        shl     edx, 6
+        mov     [4 + global_x], edx
+        ; init y
+        call    rand
+        cdq
+        ;xor    dx, dx
+        mov     ebx, SCREEN_HEIGHT
+        div     ebx
+        shl     edx, 6
+        mov     [4 + global_y], edx
+  .dont_re_init_globals:
-   and eax, 31
+        ; init x
-   sub eax, 15
+        mov     eax, [4 + global_x]
-   ;shl ax, 6
+        mov     [ebp + X_OFFSET], eax
-   mov [ebp+X_SPEED_OFFSET], eax
+        ; init y
-   ; init y speed
+        mov     eax, [4 + global_y]
-   call rand
+        mov     [ebp + Y_OFFSET], eax
-   and eax, 31
+        ; init x speed
-   sub eax, 15
+        call    rand
-   ;shl ax, 6
+        and     eax, 31
-   mov [ebp+Y_SPEED_OFFSET], eax
-   ; init color
-;   mov ax, 255
+        sub     eax, 15
-   ;call rand
-   ;and ax, 0xFF
-   mov [ebp+COLOR_OFFSET], dword 255;ax
+        ;shl    ax, 6
-   ret
-;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+        mov     [ebp + X_SPEED_OFFSET], eax
-; Misc. Functions
+        ; init y speed
-;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+        call    rand
+        and     eax, 31
-rand:
+        sub     eax, 15
-      mov eax, [8+seed]
+        ;shl    ax, 6
-      imul eax, 214013
+        mov     [ebp + Y_SPEED_OFFSET], eax
-      add eax, 2531011
+        ; init color
-      mov [8+seed], eax
+        mov     [ebp + COLOR_OFFSET], dword 255
-      shr eax, 16
+        ret
-      ret
+rand:
+        mov     eax, [8 + seed]

Subversion Repositories Kolibri OS

(root)/programs/demos/firework/trunk/firework.asm @ 7276 – Rev 2552 → 7125