Rev 2552 | Rev 7276 | Go to most recent revision | Details | Compare with Previous | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
1547 | Asper | 1 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
7125 | dunkaist | 2 | ; A firework demo |
3 | ; Programmed by Yaniv LEVIATHAN |
||
4 | ; http://yaniv.leviathanonline.com |
||
5 | ; Converted to DexOS, By Dex |
||
1547 | Asper | 6 | ; Converted to KolibriOS, By Asper |
7 | ; Optimized for KolibriOS, By Diamond |
||
8 | ; Assemble with |
||
9 | ; c:fasm firework.asm firework.kex |
||
7125 | dunkaist | 10 | ; NOTE: Needs MMX & SSE, optionally AVX |
11 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
||
12 | use32 |
||
13 | org 0x0 |
||
1547 | Asper | 14 | |
7125 | dunkaist | 15 | db 'MENUET01' ; 8 byte id |
16 | dd 0x01 ; version |
||
17 | dd STARTAPP ; program start |
||
18 | dd I_END ; program image size |
||
19 | dd E_END ; required amount of memory |
||
20 | dd stacktop ; reserved=no extended header |
||
21 | dd 0, 0 |
||
1547 | Asper | 22 | |
2095 | dunkaist | 23 | include '../../../macros.inc' |
7125 | dunkaist | 24 | SCREEN_WIDTH = 320 |
25 | SCREEN_HEIGHT = 200 |
||
26 | SIMD equ SSE |
||
27 | SIMD_BYTES = 8 |
||
28 | assert SCREEN_WIDTH mod SIMD_BYTES = 0 |
||
1547 | Asper | 29 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
7125 | dunkaist | 30 | ; Global defines |
31 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
||
1547 | Asper | 32 | |
7125 | dunkaist | 33 | NUM_PARTS = 150 |
34 | X_OFFSET = 0 |
||
35 | Y_OFFSET = 4 |
||
36 | X_SPEED_OFFSET = 8 |
||
37 | Y_SPEED_OFFSET = 12 |
||
38 | COLOR_OFFSET = 16 |
||
39 | PART_SIZE = 20 |
||
1547 | Asper | 40 | |
7125 | dunkaist | 41 | macro shade |
1547 | Asper | 42 | { |
7125 | dunkaist | 43 | local .lop |
44 | if SIMD eq SSE |
||
45 | mov ecx, SCREEN_WIDTH * SCREEN_HEIGHT / SIMD_BYTES |
||
46 | mov edi, buffer |
||
47 | movq mm1, qword [sub_mask] |
||
48 | .lop: |
||
49 | movq mm0, [edi] |
||
50 | psubusb mm0, mm1 |
||
51 | movq [edi], mm0 |
||
52 | add edi, SIMD_BYTES |
||
53 | loop .lop |
||
54 | else if SIMD eq AVX |
||
55 | mov ecx, SCREEN_WIDTH * SCREEN_HEIGHT / SIMD_BYTES |
||
56 | mov edi, buffer |
||
57 | vmovdqa xmm1, xword [sub_mask] |
||
58 | .lop: |
||
59 | vmovdqa xmm0, [edi] |
||
60 | vpsubusb xmm0, xmm0, xmm1 |
||
61 | vmovdqa [edi], xmm0 |
||
62 | add edi, SIMD_BYTES |
||
63 | loop .lop |
||
64 | else if SIMD eq AVX2 |
||
65 | mov ecx, SCREEN_WIDTH * SCREEN_HEIGHT / SIMD_BYTES |
||
66 | mov edi, buffer |
||
67 | vmovdqa ymm1, yword [sub_mask] |
||
68 | .lop: |
||
69 | vmovdqa ymm0, [edi] |
||
70 | vpsubusb ymm0, ymm0, ymm1 |
||
71 | vmovdqa [edi], ymm0 |
||
72 | add edi, SIMD_BYTES |
||
73 | loop .lop |
||
74 | else if SIMD eq AVX512 |
||
75 | mov ecx, SCREEN_WIDTH * SCREEN_HEIGHT / SIMD_BYTES |
||
76 | mov edi, buffer |
||
77 | vmovdqa64 zmm1, zword [sub_mask] |
||
78 | .lop: |
||
79 | vmovdqa64 zmm0, [edi] |
||
80 | vpsubusb zmm0, zmm0, zmm1 |
||
81 | vmovdqa64 [edi], zmm0 |
||
82 | add edi, SIMD_BYTES |
||
83 | loop .lop |
||
84 | end if |
||
1547 | Asper | 85 | } |
86 | |||
7125 | dunkaist | 87 | macro blur_prepare |
1547 | Asper | 88 | { |
7125 | dunkaist | 89 | mov ecx, (SCREEN_WIDTH * SCREEN_HEIGHT - SCREEN_WIDTH * 2 - SIMD_BYTES*2) / SIMD_BYTES |
90 | mov edi, buffer + SCREEN_WIDTH + SIMD_BYTES |
||
1547 | Asper | 91 | } |
92 | |||
7125 | dunkaist | 93 | macro blur |
1547 | Asper | 94 | { |
95 | local .lop |
||
7125 | dunkaist | 96 | if SIMD eq SSE |
97 | .lop: |
||
98 | movq mm0, [edi] |
||
99 | movq mm1, [edi + 1] |
||
100 | movq mm2, [edi - 1] |
||
101 | movq mm3, mm0 |
||
102 | movq mm4, [edi - SCREEN_WIDTH] |
||
103 | movq mm5, [edi + SCREEN_WIDTH] |
||
1547 | Asper | 104 | |
7125 | dunkaist | 105 | pavgb mm0, mm1 ; mm0 = avg(cur,cur+1) |
106 | pavgb mm3, mm2 ; mm3 = avg(cur,cur-1) |
||
107 | pavgb mm4, mm5 ; mm4 = avg(cur+width,cur-width) |
||
108 | pavgb mm3, mm4 ; mm3 = avg(avg(cur,cur-1),avg(cur+width,cur-width)) |
||
109 | pavgb mm0, mm3 ; mm0 = avg(avg(cur,cur+1), |
||
1547 | Asper | 110 | |
7125 | dunkaist | 111 | movq [edi], mm0 |
112 | add edi, SIMD_BYTES |
||
113 | loop .lop |
||
114 | else if SIMD eq AVX |
||
115 | .lop: |
||
116 | vmovdqa xmm0, [edi] |
||
117 | vmovdqa xmm1, xmm0 |
||
118 | vmovdqa xmm2, [edi - SCREEN_WIDTH] |
||
119 | |||
120 | vpavgb xmm0, xmm0, [edi + 1] |
||
121 | vpavgb xmm1, xmm1, [edi - 1] |
||
122 | vpavgb xmm2, xmm2, [edi + SCREEN_WIDTH] |
||
123 | vpavgb xmm1, xmm1, xmm2 |
||
124 | vpavgb xmm0, xmm0, xmm1 |
||
125 | |||
126 | vmovdqa [edi], xmm0 |
||
127 | add edi, SIMD_BYTES |
||
128 | loop .lop |
||
129 | else if SIMD eq AVX2 |
||
130 | .lop: |
||
131 | vmovdqa ymm0, [edi] |
||
132 | vmovdqa ymm1, ymm0 |
||
133 | vmovdqa ymm2, [edi - SCREEN_WIDTH] |
||
134 | |||
135 | vpavgb ymm0, ymm0, [edi + 1] |
||
136 | vpavgb ymm1, ymm1, [edi - 1] |
||
137 | vpavgb ymm2, ymm2, [edi + SCREEN_WIDTH] |
||
138 | vpavgb ymm1, ymm1, ymm2 |
||
139 | vpavgb ymm0, ymm0, ymm1 |
||
140 | |||
141 | vmovdqa [edi], ymm0 |
||
142 | add edi, SIMD_BYTES |
||
143 | loop .lop |
||
144 | else if SIMD eq AVX512 |
||
145 | .lop: |
||
146 | vmovdqa64 zmm0, [edi] |
||
147 | vmovdqa64 zmm1, zmm0 |
||
148 | vmovdqa64 zmm2, [edi - SCREEN_WIDTH] |
||
149 | |||
150 | vpavgb zmm0, zmm0, [edi + 1] |
||
151 | vpavgb zmm1, zmm1, [edi - 1] |
||
152 | vpavgb zmm2, zmm2, [edi + SCREEN_WIDTH] |
||
153 | vpavgb zmm1, zmm1, zmm2 |
||
154 | vpavgb zmm0, zmm0, zmm1 |
||
155 | |||
156 | vmovdqa64 [edi], zmm0 |
||
157 | add edi, SIMD_BYTES |
||
158 | loop .lop |
||
159 | end if |
||
1547 | Asper | 160 | } |
161 | |||
7125 | dunkaist | 162 | macro blur_right |
1547 | Asper | 163 | { |
164 | local .lop |
||
7125 | dunkaist | 165 | if SIMD eq SSE |
166 | .lop: |
||
167 | movq mm0, [edi] |
||
168 | movq mm1, [edi + 1] |
||
169 | movq mm2, [edi + SCREEN_WIDTH] |
||
170 | movq mm3, [edi + SCREEN_WIDTH + 1] |
||
171 | pavgb mm0, mm1 |
||
172 | pavgb mm3, mm2 |
||
173 | pavgb mm0, mm3 |
||
174 | movq [edi], mm0 |
||
175 | add edi, SIMD_BYTES |
||
176 | loop .lop |
||
177 | else if SIMD eq AVX |
||
178 | .lop: |
||
179 | vmovdqa xmm0, [edi] |
||
180 | vmovdqu xmm1, [edi + SCREEN_WIDTH + 1] |
||
181 | vpavgb xmm2, xmm0, [edi + 1] |
||
182 | vpavgb xmm3, xmm1, [edi + SCREEN_WIDTH] |
||
183 | vpavgb xmm4, xmm2, xmm3 |
||
184 | vmovdqa [edi], xmm4 |
||
185 | add edi, SIMD_BYTES |
||
186 | loop .lop |
||
187 | else if SIMD eq AVX2 |
||
188 | .lop: |
||
189 | vmovdqa ymm0, [edi] |
||
190 | vmovdqu ymm1, [edi + SCREEN_WIDTH + 1] |
||
191 | vpavgb ymm2, ymm0, [edi + 1] |
||
192 | vpavgb ymm3, ymm1, [edi + SCREEN_WIDTH] |
||
193 | vpavgb ymm4, ymm2, ymm3 |
||
194 | vmovdqa [edi], ymm4 |
||
195 | add edi, SIMD_BYTES |
||
196 | loop .lop |
||
197 | else if SIMD eq AVX512 |
||
198 | .lop: |
||
199 | vmovdqa64 zmm0, [edi] |
||
200 | vmovdqu64 zmm1, [edi + SCREEN_WIDTH + 1] |
||
201 | vpavgb zmm2, zmm0, [edi + 1] |
||
202 | vpavgb zmm3, zmm1, [edi + SCREEN_WIDTH] |
||
203 | vpavgb zmm4, zmm2, zmm3 |
||
204 | vmovdqa64 [edi], zmm4 |
||
205 | add edi, SIMD_BYTES |
||
206 | loop .lop |
||
207 | end if |
||
1547 | Asper | 208 | } |
209 | |||
210 | STARTAPP: |
||
211 | |||
212 | init_palette: |
||
7125 | dunkaist | 213 | mov edi, pal |
214 | xor eax, eax |
||
1547 | Asper | 215 | red_loop: |
7125 | dunkaist | 216 | stosd |
217 | stosd |
||
218 | add eax, 0x040000 |
||
219 | and eax, 0xFFFFFF |
||
220 | jnz red_loop |
||
1547 | Asper | 221 | |
7125 | dunkaist | 222 | mov eax, 63*4 SHL 16 |
1547 | Asper | 223 | @@: |
7125 | dunkaist | 224 | stosd |
225 | stosd |
||
226 | add ax, 0x0404 |
||
227 | jnc @b |
||
228 | |||
229 | ;zero_buffer: |
||
230 | mov ecx, SCREEN_WIDTH * SCREEN_HEIGHT / 4 |
||
231 | ; mov edi, buffer |
||
232 | xor eax, eax |
||
233 | rep stosd |
||
234 | |||
235 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
||
236 | ; Main Functions |
||
237 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
||
1547 | Asper | 238 | virtual at esp |
7125 | dunkaist | 239 | global_x dd ? |
240 | global_y dd ? |
||
241 | seed dd ? |
||
1547 | Asper | 242 | end virtual |
243 | |||
7125 | dunkaist | 244 | rdtsc |
245 | push eax ; seed |
||
246 | push 100 * 64 ; global_y |
||
247 | push 160 * 64 ; global_x |
||
1547 | Asper | 248 | |
7125 | dunkaist | 249 | jmp MAIN |
1547 | Asper | 250 | |
251 | |||
252 | red: |
||
7125 | dunkaist | 253 | mcall 9, proc_info, -1 |
254 | x = 100 |
||
255 | y = 70 |
||
256 | xsize = SCREEN_WIDTH+9 |
||
257 | ysize = SCREEN_HEIGHT+4 |
||
258 | areacolor = 0x54224466 |
||
259 | mov eax, 12 ; function 12:tell os about windowdraw |
||
260 | mov ebx, 1 ; 1, start of draw |
||
261 | int 0x40 |
||
262 | mov eax, 48 |
||
263 | mov ebx, 4 |
||
264 | int 0x40 |
||
265 | lea ecx, [(y SHL 16) + ysize + eax] |
||
266 | xor eax, eax ; function 0 : define and draw window |
||
267 | mov ebx, (x SHL 16) + xsize ; [x start] *65536 + [x size] |
||
268 | mov edx, areacolor ; color of work area RRGGBB |
||
269 | mov edi, window_title |
||
270 | int 0x40 |
||
271 | mov eax, 12 ; end of redraw |
||
272 | mov ebx, 2 |
||
273 | int 0x40 |
||
274 | |||
1547 | Asper | 275 | MAIN: |
7125 | dunkaist | 276 | test [proc_info.wnd_state], 0x04 |
277 | jnz still |
||
278 | mov ecx, NUM_PARTS |
||
279 | mov ebp, particles |
||
280 | .advance_particles: |
||
281 | mov eax, [ebp + X_OFFSET] |
||
282 | mov ebx, [ebp + Y_OFFSET] |
||
1547 | Asper | 283 | |
7125 | dunkaist | 284 | sar eax, 6 |
285 | sar ebx, 6 |
||
1547 | Asper | 286 | |
7125 | dunkaist | 287 | cmp eax, 5 |
288 | jb .new_particle |
||
289 | cmp eax, SCREEN_WIDTH - 5 |
||
290 | jge .new_particle |
||
291 | cmp ebx, 5 |
||
292 | jb .new_particle |
||
293 | cmp ebx, SCREEN_HEIGHT - 5 |
||
294 | jl .part_ok |
||
1547 | Asper | 295 | |
7125 | dunkaist | 296 | .new_particle: |
297 | call init_particle |
||
298 | jmp .advance_particles |
||
1547 | Asper | 299 | |
7125 | dunkaist | 300 | .part_ok: |
301 | imul edi, ebx, SCREEN_WIDTH |
||
302 | mov dl, [ebp+COLOR_OFFSET] |
||
303 | mov [buffer+eax+edi], dl |
||
1547 | Asper | 304 | |
7125 | dunkaist | 305 | mov eax, [ebp+X_SPEED_OFFSET] |
306 | add [ebp+X_OFFSET], eax |
||
307 | mov eax, [ebp+Y_SPEED_OFFSET] |
||
308 | add [ebp+Y_OFFSET], eax |
||
1547 | Asper | 309 | |
7125 | dunkaist | 310 | rdtsc |
311 | and al, 0x7F |
||
312 | jnz .dont_inc_y_speed |
||
313 | inc dword [ebp+Y_SPEED_OFFSET] |
||
314 | .dont_inc_y_speed: |
||
1547 | Asper | 315 | |
7125 | dunkaist | 316 | add ebp, PART_SIZE |
317 | loop .advance_particles |
||
1547 | Asper | 318 | |
7125 | dunkaist | 319 | shade |
320 | ; jmp .copy_buffer_to_video |
||
321 | blur_prepare |
||
322 | test dword [blur_right_flag] , 0x800000 |
||
323 | jnz .do_blur_right |
||
324 | blur |
||
325 | rdtsc |
||
326 | and al, 1 |
||
327 | jz .blur_ok |
||
328 | jmp .dont_blur |
||
329 | .do_blur_right: |
||
330 | blur_right |
||
331 | .blur_ok: |
||
332 | add dword [blur_right_flag], 0x1000 |
||
333 | .dont_blur: |
||
1547 | Asper | 334 | |
7125 | dunkaist | 335 | .copy_buffer_to_video: |
1547 | Asper | 336 | |
7125 | dunkaist | 337 | mcall 48, 4 |
338 | lea edx, [(5 SHL 16) + eax] |
||
1547 | Asper | 339 | |
7125 | dunkaist | 340 | mov eax, 65 |
341 | mov ebx, buffer |
||
342 | mov ecx, (SCREEN_WIDTH SHL 16) + SCREEN_HEIGHT |
||
343 | push 8 |
||
344 | pop esi |
||
345 | mov edi, pal |
||
346 | xor ebp, ebp |
||
347 | int 0x40 |
||
1547 | Asper | 348 | |
7125 | dunkaist | 349 | |
1547 | Asper | 350 | still: |
7125 | dunkaist | 351 | mov eax, 11 ; Test if there is an event in the queue. |
352 | int 0x40 |
||
1547 | Asper | 353 | |
7125 | dunkaist | 354 | dec eax ; redraw request ? |
355 | jz red |
||
356 | dec eax ; key in buffer ? |
||
357 | jz key |
||
358 | dec eax ; button in buffer ? |
||
359 | jz button |
||
1547 | Asper | 360 | |
7125 | dunkaist | 361 | jmp MAIN |
1547 | Asper | 362 | |
363 | |||
364 | key: |
||
7125 | dunkaist | 365 | mov eax, 2 |
366 | int 0x40 |
||
367 | ; cmp ah, 1 ; Test Esc in Scan |
||
368 | ; je close_app |
||
369 | cmp ah, 27 ; Test Esc in ASCII |
||
370 | je close_app |
||
371 | jmp MAIN |
||
1547 | Asper | 372 | |
373 | button: |
||
374 | ; we have only one button, close |
||
375 | close_app: |
||
7125 | dunkaist | 376 | mov eax, -1 ; close this program |
377 | int 0x40 |
||
1547 | Asper | 378 | |
7125 | dunkaist | 379 | init_particle: |
380 | rdtsc |
||
381 | and al, 0x1F |
||
382 | jnz .dont_re_init_globals |
||
383 | ; init x |
||
384 | call rand |
||
385 | cdq |
||
386 | ;xor dx, dx |
||
387 | mov ebx, SCREEN_WIDTH |
||
388 | div ebx |
||
389 | shl edx, 6 |
||
390 | mov [4 + global_x], edx |
||
391 | ; init y |
||
392 | call rand |
||
393 | cdq |
||
394 | ;xor dx, dx |
||
395 | mov ebx, SCREEN_HEIGHT |
||
396 | div ebx |
||
397 | shl edx, 6 |
||
398 | mov [4 + global_y], edx |
||
399 | .dont_re_init_globals: |
||
400 | ; init x |
||
401 | mov eax, [4 + global_x] |
||
402 | mov [ebp + X_OFFSET], eax |
||
403 | ; init y |
||
404 | mov eax, [4 + global_y] |
||
405 | mov [ebp + Y_OFFSET], eax |
||
406 | ; init x speed |
||
407 | call rand |
||
408 | and eax, 31 |
||
409 | sub eax, 15 |
||
410 | ;shl ax, 6 |
||
411 | mov [ebp + X_SPEED_OFFSET], eax |
||
412 | ; init y speed |
||
413 | call rand |
||
414 | and eax, 31 |
||
415 | sub eax, 15 |
||
416 | ;shl ax, 6 |
||
417 | mov [ebp + Y_SPEED_OFFSET], eax |
||
418 | ; init color |
||
419 | mov [ebp + COLOR_OFFSET], dword 255 |
||
420 | ret |
||
1547 | Asper | 421 | |
7125 | dunkaist | 422 | rand: |
423 | mov eax, [8 + seed] |
||
424 | imul eax, 214013 |
||
425 | add eax, 2531011 |
||
426 | mov [8 + seed], eax |
||
427 | shr eax, 16 |
||
428 | ret |
||
1547 | Asper | 429 | |
430 | ; DATA AREA |
||
7125 | dunkaist | 431 | window_title db 'Firework demo',0 |
432 | align SIMD_BYTES |
||
433 | sub_mask db SIMD_BYTES dup 0x01 |
||
434 | ; x, y, x_speed, y_speed, color |
||
435 | particles: times NUM_PARTS dd 0, 0, 0, 0, 0 |
||
436 | blur_right_flag dd 0 |
||
1547 | Asper | 437 | I_END: |
7125 | dunkaist | 438 | proc_info process_information |
439 | align 16 |
||
440 | pal rb 256 * 4 |
||
441 | align SIMD_BYTES |
||
442 | buffer rb SCREEN_WIDTH * SCREEN_HEIGHT |
||
443 | E_END: |
||
444 | rd 0x200 |
||
445 | stacktop: |