Rev 7125 | Rev 7277 | 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 |
||
7276 | dunkaist | 10 | ; NOTE: Needs MMX & SSE, |
11 | ; optionally AVX, AVX2, AVX512 |
||
7125 | dunkaist | 12 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
13 | use32 |
||
14 | org 0x0 |
||
1547 | Asper | 15 | |
7125 | dunkaist | 16 | db 'MENUET01' ; 8 byte id |
17 | dd 0x01 ; version |
||
18 | dd STARTAPP ; program start |
||
19 | dd I_END ; program image size |
||
20 | dd E_END ; required amount of memory |
||
21 | dd stacktop ; reserved=no extended header |
||
22 | dd 0, 0 |
||
1547 | Asper | 23 | |
2095 | dunkaist | 24 | include '../../../macros.inc' |
7125 | dunkaist | 25 | SCREEN_WIDTH = 320 |
26 | SCREEN_HEIGHT = 200 |
||
7276 | dunkaist | 27 | SIMD equ AVX |
28 | SIMD_BYTES = 16 |
||
29 | ; SSE 8 |
||
30 | ; AVX 16 |
||
31 | ; AVX2 32 |
||
32 | ; AVX512 64 |
||
7125 | dunkaist | 33 | assert SCREEN_WIDTH mod SIMD_BYTES = 0 |
1547 | Asper | 34 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
7125 | dunkaist | 35 | ; Global defines |
36 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
||
1547 | Asper | 37 | |
7125 | dunkaist | 38 | NUM_PARTS = 150 |
39 | X_OFFSET = 0 |
||
40 | Y_OFFSET = 4 |
||
41 | X_SPEED_OFFSET = 8 |
||
42 | Y_SPEED_OFFSET = 12 |
||
43 | COLOR_OFFSET = 16 |
||
44 | PART_SIZE = 20 |
||
1547 | Asper | 45 | |
7125 | dunkaist | 46 | macro shade |
1547 | Asper | 47 | { |
7125 | dunkaist | 48 | local .lop |
49 | if SIMD eq SSE |
||
50 | mov ecx, SCREEN_WIDTH * SCREEN_HEIGHT / SIMD_BYTES |
||
51 | mov edi, buffer |
||
52 | movq mm1, qword [sub_mask] |
||
53 | .lop: |
||
54 | movq mm0, [edi] |
||
55 | psubusb mm0, mm1 |
||
56 | movq [edi], mm0 |
||
57 | add edi, SIMD_BYTES |
||
58 | loop .lop |
||
59 | else if SIMD eq AVX |
||
60 | mov ecx, SCREEN_WIDTH * SCREEN_HEIGHT / SIMD_BYTES |
||
61 | mov edi, buffer |
||
62 | vmovdqa xmm1, xword [sub_mask] |
||
63 | .lop: |
||
64 | vmovdqa xmm0, [edi] |
||
65 | vpsubusb xmm0, xmm0, xmm1 |
||
66 | vmovdqa [edi], xmm0 |
||
67 | add edi, SIMD_BYTES |
||
68 | loop .lop |
||
69 | else if SIMD eq AVX2 |
||
70 | mov ecx, SCREEN_WIDTH * SCREEN_HEIGHT / SIMD_BYTES |
||
71 | mov edi, buffer |
||
72 | vmovdqa ymm1, yword [sub_mask] |
||
73 | .lop: |
||
74 | vmovdqa ymm0, [edi] |
||
75 | vpsubusb ymm0, ymm0, ymm1 |
||
76 | vmovdqa [edi], ymm0 |
||
77 | add edi, SIMD_BYTES |
||
78 | loop .lop |
||
79 | else if SIMD eq AVX512 |
||
80 | mov ecx, SCREEN_WIDTH * SCREEN_HEIGHT / SIMD_BYTES |
||
81 | mov edi, buffer |
||
82 | vmovdqa64 zmm1, zword [sub_mask] |
||
83 | .lop: |
||
84 | vmovdqa64 zmm0, [edi] |
||
85 | vpsubusb zmm0, zmm0, zmm1 |
||
86 | vmovdqa64 [edi], zmm0 |
||
87 | add edi, SIMD_BYTES |
||
88 | loop .lop |
||
89 | end if |
||
1547 | Asper | 90 | } |
91 | |||
7125 | dunkaist | 92 | macro blur_prepare |
1547 | Asper | 93 | { |
7125 | dunkaist | 94 | mov ecx, (SCREEN_WIDTH * SCREEN_HEIGHT - SCREEN_WIDTH * 2 - SIMD_BYTES*2) / SIMD_BYTES |
95 | mov edi, buffer + SCREEN_WIDTH + SIMD_BYTES |
||
1547 | Asper | 96 | } |
97 | |||
7125 | dunkaist | 98 | macro blur |
1547 | Asper | 99 | { |
100 | local .lop |
||
7125 | dunkaist | 101 | if SIMD eq SSE |
102 | .lop: |
||
103 | movq mm0, [edi] |
||
104 | movq mm1, [edi + 1] |
||
105 | movq mm2, [edi - 1] |
||
106 | movq mm3, mm0 |
||
107 | movq mm4, [edi - SCREEN_WIDTH] |
||
108 | movq mm5, [edi + SCREEN_WIDTH] |
||
1547 | Asper | 109 | |
7125 | dunkaist | 110 | pavgb mm0, mm1 ; mm0 = avg(cur,cur+1) |
111 | pavgb mm3, mm2 ; mm3 = avg(cur,cur-1) |
||
112 | pavgb mm4, mm5 ; mm4 = avg(cur+width,cur-width) |
||
113 | pavgb mm3, mm4 ; mm3 = avg(avg(cur,cur-1),avg(cur+width,cur-width)) |
||
114 | pavgb mm0, mm3 ; mm0 = avg(avg(cur,cur+1), |
||
1547 | Asper | 115 | |
7125 | dunkaist | 116 | movq [edi], mm0 |
117 | add edi, SIMD_BYTES |
||
118 | loop .lop |
||
119 | else if SIMD eq AVX |
||
120 | .lop: |
||
121 | vmovdqa xmm0, [edi] |
||
122 | vmovdqa xmm1, xmm0 |
||
123 | vmovdqa xmm2, [edi - SCREEN_WIDTH] |
||
124 | |||
125 | vpavgb xmm0, xmm0, [edi + 1] |
||
126 | vpavgb xmm1, xmm1, [edi - 1] |
||
127 | vpavgb xmm2, xmm2, [edi + SCREEN_WIDTH] |
||
128 | vpavgb xmm1, xmm1, xmm2 |
||
129 | vpavgb xmm0, xmm0, xmm1 |
||
130 | |||
131 | vmovdqa [edi], xmm0 |
||
132 | add edi, SIMD_BYTES |
||
133 | loop .lop |
||
134 | else if SIMD eq AVX2 |
||
135 | .lop: |
||
136 | vmovdqa ymm0, [edi] |
||
137 | vmovdqa ymm1, ymm0 |
||
138 | vmovdqa ymm2, [edi - SCREEN_WIDTH] |
||
139 | |||
140 | vpavgb ymm0, ymm0, [edi + 1] |
||
141 | vpavgb ymm1, ymm1, [edi - 1] |
||
142 | vpavgb ymm2, ymm2, [edi + SCREEN_WIDTH] |
||
143 | vpavgb ymm1, ymm1, ymm2 |
||
144 | vpavgb ymm0, ymm0, ymm1 |
||
145 | |||
146 | vmovdqa [edi], ymm0 |
||
147 | add edi, SIMD_BYTES |
||
148 | loop .lop |
||
149 | else if SIMD eq AVX512 |
||
150 | .lop: |
||
151 | vmovdqa64 zmm0, [edi] |
||
152 | vmovdqa64 zmm1, zmm0 |
||
153 | vmovdqa64 zmm2, [edi - SCREEN_WIDTH] |
||
154 | |||
155 | vpavgb zmm0, zmm0, [edi + 1] |
||
156 | vpavgb zmm1, zmm1, [edi - 1] |
||
157 | vpavgb zmm2, zmm2, [edi + SCREEN_WIDTH] |
||
158 | vpavgb zmm1, zmm1, zmm2 |
||
159 | vpavgb zmm0, zmm0, zmm1 |
||
160 | |||
161 | vmovdqa64 [edi], zmm0 |
||
162 | add edi, SIMD_BYTES |
||
163 | loop .lop |
||
164 | end if |
||
1547 | Asper | 165 | } |
166 | |||
7125 | dunkaist | 167 | macro blur_right |
1547 | Asper | 168 | { |
169 | local .lop |
||
7125 | dunkaist | 170 | if SIMD eq SSE |
171 | .lop: |
||
172 | movq mm0, [edi] |
||
173 | movq mm1, [edi + 1] |
||
174 | movq mm2, [edi + SCREEN_WIDTH] |
||
175 | movq mm3, [edi + SCREEN_WIDTH + 1] |
||
176 | pavgb mm0, mm1 |
||
177 | pavgb mm3, mm2 |
||
178 | pavgb mm0, mm3 |
||
179 | movq [edi], mm0 |
||
180 | add edi, SIMD_BYTES |
||
181 | loop .lop |
||
182 | else if SIMD eq AVX |
||
183 | .lop: |
||
184 | vmovdqa xmm0, [edi] |
||
185 | vmovdqu xmm1, [edi + SCREEN_WIDTH + 1] |
||
186 | vpavgb xmm2, xmm0, [edi + 1] |
||
187 | vpavgb xmm3, xmm1, [edi + SCREEN_WIDTH] |
||
188 | vpavgb xmm4, xmm2, xmm3 |
||
189 | vmovdqa [edi], xmm4 |
||
190 | add edi, SIMD_BYTES |
||
191 | loop .lop |
||
192 | else if SIMD eq AVX2 |
||
193 | .lop: |
||
194 | vmovdqa ymm0, [edi] |
||
195 | vmovdqu ymm1, [edi + SCREEN_WIDTH + 1] |
||
196 | vpavgb ymm2, ymm0, [edi + 1] |
||
197 | vpavgb ymm3, ymm1, [edi + SCREEN_WIDTH] |
||
198 | vpavgb ymm4, ymm2, ymm3 |
||
199 | vmovdqa [edi], ymm4 |
||
200 | add edi, SIMD_BYTES |
||
201 | loop .lop |
||
202 | else if SIMD eq AVX512 |
||
203 | .lop: |
||
204 | vmovdqa64 zmm0, [edi] |
||
205 | vmovdqu64 zmm1, [edi + SCREEN_WIDTH + 1] |
||
206 | vpavgb zmm2, zmm0, [edi + 1] |
||
207 | vpavgb zmm3, zmm1, [edi + SCREEN_WIDTH] |
||
208 | vpavgb zmm4, zmm2, zmm3 |
||
209 | vmovdqa64 [edi], zmm4 |
||
210 | add edi, SIMD_BYTES |
||
211 | loop .lop |
||
212 | end if |
||
1547 | Asper | 213 | } |
214 | |||
215 | STARTAPP: |
||
216 | |||
217 | init_palette: |
||
7125 | dunkaist | 218 | mov edi, pal |
219 | xor eax, eax |
||
1547 | Asper | 220 | red_loop: |
7125 | dunkaist | 221 | stosd |
222 | stosd |
||
223 | add eax, 0x040000 |
||
224 | and eax, 0xFFFFFF |
||
225 | jnz red_loop |
||
1547 | Asper | 226 | |
7125 | dunkaist | 227 | mov eax, 63*4 SHL 16 |
1547 | Asper | 228 | @@: |
7125 | dunkaist | 229 | stosd |
230 | stosd |
||
231 | add ax, 0x0404 |
||
232 | jnc @b |
||
233 | |||
234 | ;zero_buffer: |
||
235 | mov ecx, SCREEN_WIDTH * SCREEN_HEIGHT / 4 |
||
236 | ; mov edi, buffer |
||
237 | xor eax, eax |
||
238 | rep stosd |
||
239 | |||
240 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
||
241 | ; Main Functions |
||
242 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
||
1547 | Asper | 243 | virtual at esp |
7125 | dunkaist | 244 | global_x dd ? |
245 | global_y dd ? |
||
246 | seed dd ? |
||
1547 | Asper | 247 | end virtual |
248 | |||
7125 | dunkaist | 249 | rdtsc |
250 | push eax ; seed |
||
251 | push 100 * 64 ; global_y |
||
252 | push 160 * 64 ; global_x |
||
1547 | Asper | 253 | |
7125 | dunkaist | 254 | jmp MAIN |
1547 | Asper | 255 | |
256 | |||
257 | red: |
||
7125 | dunkaist | 258 | mcall 9, proc_info, -1 |
259 | x = 100 |
||
260 | y = 70 |
||
261 | xsize = SCREEN_WIDTH+9 |
||
262 | ysize = SCREEN_HEIGHT+4 |
||
263 | areacolor = 0x54224466 |
||
264 | mov eax, 12 ; function 12:tell os about windowdraw |
||
265 | mov ebx, 1 ; 1, start of draw |
||
266 | int 0x40 |
||
267 | mov eax, 48 |
||
268 | mov ebx, 4 |
||
269 | int 0x40 |
||
270 | lea ecx, [(y SHL 16) + ysize + eax] |
||
271 | xor eax, eax ; function 0 : define and draw window |
||
272 | mov ebx, (x SHL 16) + xsize ; [x start] *65536 + [x size] |
||
273 | mov edx, areacolor ; color of work area RRGGBB |
||
274 | mov edi, window_title |
||
275 | int 0x40 |
||
276 | mov eax, 12 ; end of redraw |
||
277 | mov ebx, 2 |
||
278 | int 0x40 |
||
279 | |||
1547 | Asper | 280 | MAIN: |
7125 | dunkaist | 281 | test [proc_info.wnd_state], 0x04 |
282 | jnz still |
||
283 | mov ecx, NUM_PARTS |
||
284 | mov ebp, particles |
||
285 | .advance_particles: |
||
286 | mov eax, [ebp + X_OFFSET] |
||
287 | mov ebx, [ebp + Y_OFFSET] |
||
1547 | Asper | 288 | |
7125 | dunkaist | 289 | sar eax, 6 |
290 | sar ebx, 6 |
||
1547 | Asper | 291 | |
7125 | dunkaist | 292 | cmp eax, 5 |
293 | jb .new_particle |
||
294 | cmp eax, SCREEN_WIDTH - 5 |
||
295 | jge .new_particle |
||
296 | cmp ebx, 5 |
||
297 | jb .new_particle |
||
298 | cmp ebx, SCREEN_HEIGHT - 5 |
||
299 | jl .part_ok |
||
1547 | Asper | 300 | |
7125 | dunkaist | 301 | .new_particle: |
302 | call init_particle |
||
303 | jmp .advance_particles |
||
1547 | Asper | 304 | |
7125 | dunkaist | 305 | .part_ok: |
306 | imul edi, ebx, SCREEN_WIDTH |
||
307 | mov dl, [ebp+COLOR_OFFSET] |
||
308 | mov [buffer+eax+edi], dl |
||
1547 | Asper | 309 | |
7125 | dunkaist | 310 | mov eax, [ebp+X_SPEED_OFFSET] |
311 | add [ebp+X_OFFSET], eax |
||
312 | mov eax, [ebp+Y_SPEED_OFFSET] |
||
313 | add [ebp+Y_OFFSET], eax |
||
1547 | Asper | 314 | |
7125 | dunkaist | 315 | rdtsc |
316 | and al, 0x7F |
||
317 | jnz .dont_inc_y_speed |
||
318 | inc dword [ebp+Y_SPEED_OFFSET] |
||
319 | .dont_inc_y_speed: |
||
1547 | Asper | 320 | |
7125 | dunkaist | 321 | add ebp, PART_SIZE |
322 | loop .advance_particles |
||
1547 | Asper | 323 | |
7125 | dunkaist | 324 | shade |
325 | ; jmp .copy_buffer_to_video |
||
326 | blur_prepare |
||
327 | test dword [blur_right_flag] , 0x800000 |
||
328 | jnz .do_blur_right |
||
329 | blur |
||
330 | rdtsc |
||
331 | and al, 1 |
||
332 | jz .blur_ok |
||
333 | jmp .dont_blur |
||
334 | .do_blur_right: |
||
335 | blur_right |
||
336 | .blur_ok: |
||
337 | add dword [blur_right_flag], 0x1000 |
||
338 | .dont_blur: |
||
1547 | Asper | 339 | |
7125 | dunkaist | 340 | .copy_buffer_to_video: |
1547 | Asper | 341 | |
7125 | dunkaist | 342 | mcall 48, 4 |
343 | lea edx, [(5 SHL 16) + eax] |
||
1547 | Asper | 344 | |
7125 | dunkaist | 345 | mov eax, 65 |
346 | mov ebx, buffer |
||
347 | mov ecx, (SCREEN_WIDTH SHL 16) + SCREEN_HEIGHT |
||
348 | push 8 |
||
349 | pop esi |
||
350 | mov edi, pal |
||
351 | xor ebp, ebp |
||
352 | int 0x40 |
||
1547 | Asper | 353 | |
7125 | dunkaist | 354 | |
1547 | Asper | 355 | still: |
7125 | dunkaist | 356 | mov eax, 11 ; Test if there is an event in the queue. |
357 | int 0x40 |
||
1547 | Asper | 358 | |
7125 | dunkaist | 359 | dec eax ; redraw request ? |
360 | jz red |
||
361 | dec eax ; key in buffer ? |
||
362 | jz key |
||
363 | dec eax ; button in buffer ? |
||
364 | jz button |
||
1547 | Asper | 365 | |
7125 | dunkaist | 366 | jmp MAIN |
1547 | Asper | 367 | |
368 | |||
369 | key: |
||
7125 | dunkaist | 370 | mov eax, 2 |
371 | int 0x40 |
||
372 | ; cmp ah, 1 ; Test Esc in Scan |
||
373 | ; je close_app |
||
374 | cmp ah, 27 ; Test Esc in ASCII |
||
375 | je close_app |
||
376 | jmp MAIN |
||
1547 | Asper | 377 | |
378 | button: |
||
379 | ; we have only one button, close |
||
380 | close_app: |
||
7125 | dunkaist | 381 | mov eax, -1 ; close this program |
382 | int 0x40 |
||
1547 | Asper | 383 | |
7125 | dunkaist | 384 | init_particle: |
385 | rdtsc |
||
386 | and al, 0x1F |
||
387 | jnz .dont_re_init_globals |
||
388 | ; init x |
||
389 | call rand |
||
390 | cdq |
||
391 | ;xor dx, dx |
||
392 | mov ebx, SCREEN_WIDTH |
||
393 | div ebx |
||
394 | shl edx, 6 |
||
395 | mov [4 + global_x], edx |
||
396 | ; init y |
||
397 | call rand |
||
398 | cdq |
||
399 | ;xor dx, dx |
||
400 | mov ebx, SCREEN_HEIGHT |
||
401 | div ebx |
||
402 | shl edx, 6 |
||
403 | mov [4 + global_y], edx |
||
404 | .dont_re_init_globals: |
||
405 | ; init x |
||
406 | mov eax, [4 + global_x] |
||
407 | mov [ebp + X_OFFSET], eax |
||
408 | ; init y |
||
409 | mov eax, [4 + global_y] |
||
410 | mov [ebp + Y_OFFSET], eax |
||
411 | ; init x speed |
||
412 | call rand |
||
413 | and eax, 31 |
||
414 | sub eax, 15 |
||
415 | ;shl ax, 6 |
||
416 | mov [ebp + X_SPEED_OFFSET], eax |
||
417 | ; init y speed |
||
418 | call rand |
||
419 | and eax, 31 |
||
420 | sub eax, 15 |
||
421 | ;shl ax, 6 |
||
422 | mov [ebp + Y_SPEED_OFFSET], eax |
||
423 | ; init color |
||
424 | mov [ebp + COLOR_OFFSET], dword 255 |
||
425 | ret |
||
1547 | Asper | 426 | |
7125 | dunkaist | 427 | rand: |
428 | mov eax, [8 + seed] |
||
429 | imul eax, 214013 |
||
430 | add eax, 2531011 |
||
431 | mov [8 + seed], eax |
||
432 | shr eax, 16 |
||
433 | ret |
||
1547 | Asper | 434 | |
435 | ; DATA AREA |
||
7125 | dunkaist | 436 | window_title db 'Firework demo',0 |
437 | align SIMD_BYTES |
||
438 | sub_mask db SIMD_BYTES dup 0x01 |
||
439 | ; x, y, x_speed, y_speed, color |
||
440 | particles: times NUM_PARTS dd 0, 0, 0, 0, 0 |
||
441 | blur_right_flag dd 0 |
||
1547 | Asper | 442 | I_END: |
7125 | dunkaist | 443 | proc_info process_information |
444 | align 16 |
||
445 | pal rb 256 * 4 |
||
446 | align SIMD_BYTES |
||
447 | buffer rb SCREEN_WIDTH * SCREEN_HEIGHT |
||
448 | E_END: |
||
449 | rd 0x200 |
||
450 | stacktop: |