Rev 9172 | Details | Compare with Previous | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
8210 | maxcodehac | 1 | ; |
2 | ; x86 format converters for HERMES |
||
3 | ; Copyright (c) 1998 Glenn Fielder (gaffer@gaffer.org) |
||
4 | ; This source code is licensed under the GNU LGPL |
||
5 | ; |
||
6 | ; Please refer to the file COPYING.LIB contained in the distribution for |
||
7 | ; licensing conditions |
||
8 | ; |
||
9 | ; Routines adjusted for Hermes by Christian Nentwich (brn@eleet.mcb.at) |
||
10 | ; Used with permission. |
||
11 | ; |
||
12 | |||
13 | BITS 32 |
||
14 | |||
9172 | turbocat | 15 | %include "common.inc" |
8210 | maxcodehac | 16 | |
9172 | turbocat | 17 | SDL_FUNC _ConvertX86p16_16BGR565 |
18 | SDL_FUNC _ConvertX86p16_16RGB555 |
||
19 | SDL_FUNC _ConvertX86p16_16BGR555 |
||
20 | SDL_FUNC _ConvertX86p16_8RGB332 |
||
21 | |||
9202 | turbocat | 22 | SDL_FUNC ConvertX86p16_16BGR565 |
23 | SDL_FUNC ConvertX86p16_16RGB555 |
||
24 | SDL_FUNC ConvertX86p16_16BGR555 |
||
25 | SDL_FUNC ConvertX86p16_8RGB332 |
||
26 | |||
27 | |||
8210 | maxcodehac | 28 | EXTERN _ConvertX86 |
9202 | turbocat | 29 | EXTERN ConvertX86 |
8210 | maxcodehac | 30 | |
31 | SECTION .text |
||
32 | |||
9202 | turbocat | 33 | ConvertX86p16_16BGR565: |
8210 | maxcodehac | 34 | _ConvertX86p16_16BGR565: |
35 | |||
36 | ; check short |
||
37 | cmp ecx,BYTE 16 |
||
38 | ja .L3 |
||
39 | |||
40 | |||
9172 | turbocat | 41 | .L1: ; short loop |
8210 | maxcodehac | 42 | mov al,[esi] |
43 | mov ah,[esi+1] |
||
44 | mov ebx,eax |
||
45 | mov edx,eax |
||
46 | shr eax,11 |
||
47 | and eax,BYTE 11111b |
||
48 | and ebx,11111100000b |
||
49 | shl edx,11 |
||
50 | add eax,ebx |
||
51 | add eax,edx |
||
52 | mov [edi],al |
||
53 | mov [edi+1],ah |
||
54 | add esi,BYTE 2 |
||
55 | add edi,BYTE 2 |
||
56 | dec ecx |
||
57 | jnz .L1 |
||
9172 | turbocat | 58 | .L2: |
59 | retn |
||
8210 | maxcodehac | 60 | |
9172 | turbocat | 61 | .L3: ; head |
8210 | maxcodehac | 62 | mov eax,edi |
63 | and eax,BYTE 11b |
||
64 | jz .L4 |
||
65 | mov al,[esi] |
||
66 | mov ah,[esi+1] |
||
67 | mov ebx,eax |
||
68 | mov edx,eax |
||
69 | shr eax,11 |
||
70 | and eax,BYTE 11111b |
||
71 | and ebx,11111100000b |
||
72 | shl edx,11 |
||
73 | add eax,ebx |
||
74 | add eax,edx |
||
75 | mov [edi],al |
||
76 | mov [edi+1],ah |
||
77 | add esi,BYTE 2 |
||
78 | add edi,BYTE 2 |
||
79 | dec ecx |
||
80 | |||
9172 | turbocat | 81 | .L4: ; save count |
8210 | maxcodehac | 82 | push ecx |
83 | |||
84 | ; unroll twice |
||
85 | shr ecx,1 |
||
86 | |||
87 | ; point arrays to end |
||
88 | lea esi,[esi+ecx*4] |
||
89 | lea edi,[edi+ecx*4] |
||
90 | |||
91 | ; negative counter |
||
92 | neg ecx |
||
93 | jmp SHORT .L6 |
||
94 | |||
9172 | turbocat | 95 | .L5: mov [edi+ecx*4-4],eax |
96 | .L6: mov eax,[esi+ecx*4] |
||
8210 | maxcodehac | 97 | |
98 | mov ebx,[esi+ecx*4] |
||
99 | and eax,07E007E0h |
||
100 | |||
101 | mov edx,[esi+ecx*4] |
||
102 | and ebx,0F800F800h |
||
103 | |||
104 | shr ebx,11 |
||
105 | and edx,001F001Fh |
||
106 | |||
107 | shl edx,11 |
||
108 | add eax,ebx |
||
109 | |||
110 | add eax,edx |
||
111 | inc ecx |
||
112 | |||
113 | jnz .L5 |
||
114 | |||
115 | mov [edi+ecx*4-4],eax |
||
116 | |||
117 | ; tail |
||
118 | pop ecx |
||
119 | and ecx,BYTE 1 |
||
120 | jz .L7 |
||
121 | mov al,[esi] |
||
122 | mov ah,[esi+1] |
||
123 | mov ebx,eax |
||
124 | mov edx,eax |
||
125 | shr eax,11 |
||
126 | and eax,BYTE 11111b |
||
127 | and ebx,11111100000b |
||
128 | shl edx,11 |
||
129 | add eax,ebx |
||
130 | add eax,edx |
||
131 | mov [edi],al |
||
132 | mov [edi+1],ah |
||
133 | add esi,BYTE 2 |
||
134 | add edi,BYTE 2 |
||
135 | |||
9172 | turbocat | 136 | .L7: |
137 | retn |
||
8210 | maxcodehac | 138 | |
9202 | turbocat | 139 | ConvertX86p16_16RGB555: |
8210 | maxcodehac | 140 | _ConvertX86p16_16RGB555: |
141 | |||
142 | ; check short |
||
143 | cmp ecx,BYTE 32 |
||
144 | ja .L3 |
||
145 | |||
146 | |||
9172 | turbocat | 147 | .L1: ; short loop |
8210 | maxcodehac | 148 | mov al,[esi] |
149 | mov ah,[esi+1] |
||
150 | mov ebx,eax |
||
151 | shr ebx,1 |
||
152 | and ebx, 0111111111100000b |
||
153 | and eax,BYTE 0000000000011111b |
||
154 | add eax,ebx |
||
155 | mov [edi],al |
||
156 | mov [edi+1],ah |
||
157 | add esi,BYTE 2 |
||
158 | add edi,BYTE 2 |
||
159 | dec ecx |
||
160 | jnz .L1 |
||
9172 | turbocat | 161 | .L2: |
162 | retn |
||
8210 | maxcodehac | 163 | |
9172 | turbocat | 164 | .L3: ; head |
8210 | maxcodehac | 165 | mov eax,edi |
166 | and eax,BYTE 11b |
||
167 | jz .L4 |
||
168 | mov al,[esi] |
||
169 | mov ah,[esi+1] |
||
170 | mov ebx,eax |
||
171 | shr ebx,1 |
||
172 | and ebx, 0111111111100000b |
||
173 | and eax,BYTE 0000000000011111b |
||
174 | add eax,ebx |
||
175 | mov [edi],al |
||
176 | mov [edi+1],ah |
||
177 | add esi,BYTE 2 |
||
178 | add edi,BYTE 2 |
||
179 | dec ecx |
||
180 | |||
9172 | turbocat | 181 | .L4: ; save ebp |
8210 | maxcodehac | 182 | push ebp |
183 | |||
184 | ; save count |
||
185 | push ecx |
||
186 | |||
187 | ; unroll four times |
||
188 | shr ecx,2 |
||
189 | |||
190 | ; point arrays to end |
||
191 | lea esi,[esi+ecx*8] |
||
192 | lea edi,[edi+ecx*8] |
||
193 | |||
194 | ; negative counter |
||
195 | xor ebp,ebp |
||
196 | sub ebp,ecx |
||
197 | |||
9172 | turbocat | 198 | .L5: mov eax,[esi+ebp*8] ; agi? |
8210 | maxcodehac | 199 | mov ecx,[esi+ebp*8+4] |
200 | |||
201 | mov ebx,eax |
||
202 | mov edx,ecx |
||
203 | |||
204 | and eax,0FFC0FFC0h |
||
205 | and ecx,0FFC0FFC0h |
||
206 | |||
207 | shr eax,1 |
||
208 | and ebx,001F001Fh |
||
209 | |||
210 | shr ecx,1 |
||
211 | and edx,001F001Fh |
||
212 | |||
213 | add eax,ebx |
||
214 | add ecx,edx |
||
215 | |||
216 | mov [edi+ebp*8],eax |
||
217 | mov [edi+ebp*8+4],ecx |
||
218 | |||
219 | inc ebp |
||
220 | jnz .L5 |
||
221 | |||
222 | ; tail |
||
223 | pop ecx |
||
9172 | turbocat | 224 | .L6: and ecx,BYTE 11b |
8210 | maxcodehac | 225 | jz .L7 |
226 | mov al,[esi] |
||
227 | mov ah,[esi+1] |
||
228 | mov ebx,eax |
||
229 | shr ebx,1 |
||
230 | and ebx, 0111111111100000b |
||
231 | and eax,BYTE 0000000000011111b |
||
232 | add eax,ebx |
||
233 | mov [edi],al |
||
234 | mov [edi+1],ah |
||
235 | add esi,BYTE 2 |
||
236 | add edi,BYTE 2 |
||
237 | dec ecx |
||
238 | jmp SHORT .L6 |
||
239 | |||
9172 | turbocat | 240 | .L7: pop ebp |
241 | retn |
||
8210 | maxcodehac | 242 | |
9202 | turbocat | 243 | ConvertX86p16_16BGR555: |
8210 | maxcodehac | 244 | _ConvertX86p16_16BGR555: |
245 | |||
246 | ; check short |
||
247 | cmp ecx,BYTE 16 |
||
248 | ja .L3 |
||
249 | |||
250 | |||
9172 | turbocat | 251 | .L1: ; short loop |
8210 | maxcodehac | 252 | mov al,[esi] |
253 | mov ah,[esi+1] |
||
254 | mov ebx,eax |
||
255 | mov edx,eax |
||
256 | shr eax,11 |
||
257 | and eax,BYTE 11111b |
||
258 | shr ebx,1 |
||
259 | and ebx,1111100000b |
||
260 | shl edx,10 |
||
261 | and edx,0111110000000000b |
||
262 | add eax,ebx |
||
263 | add eax,edx |
||
264 | mov [edi],al |
||
265 | mov [edi+1],ah |
||
266 | add esi,BYTE 2 |
||
267 | add edi,BYTE 2 |
||
268 | dec ecx |
||
269 | jnz .L1 |
||
9172 | turbocat | 270 | .L2: |
271 | retn |
||
8210 | maxcodehac | 272 | |
9172 | turbocat | 273 | .L3: ; head |
8210 | maxcodehac | 274 | mov eax,edi |
275 | and eax,BYTE 11b |
||
276 | jz .L4 |
||
277 | mov al,[esi] |
||
278 | mov ah,[esi+1] |
||
279 | mov ebx,eax |
||
280 | mov edx,eax |
||
281 | shr eax,11 |
||
282 | and eax,BYTE 11111b |
||
283 | shr ebx,1 |
||
284 | and ebx,1111100000b |
||
285 | shl edx,10 |
||
286 | and edx,0111110000000000b |
||
287 | add eax,ebx |
||
288 | add eax,edx |
||
289 | mov [edi],al |
||
290 | mov [edi+1],ah |
||
291 | add esi,BYTE 2 |
||
292 | add edi,BYTE 2 |
||
293 | dec ecx |
||
294 | |||
9172 | turbocat | 295 | .L4: ; save count |
8210 | maxcodehac | 296 | push ecx |
297 | |||
298 | ; unroll twice |
||
299 | shr ecx,1 |
||
300 | |||
301 | ; point arrays to end |
||
302 | lea esi,[esi+ecx*4] |
||
303 | lea edi,[edi+ecx*4] |
||
304 | |||
305 | ; negative counter |
||
306 | neg ecx |
||
307 | jmp SHORT .L6 |
||
308 | |||
9172 | turbocat | 309 | .L5: mov [edi+ecx*4-4],eax |
310 | .L6: mov eax,[esi+ecx*4] |
||
8210 | maxcodehac | 311 | |
312 | shr eax,1 |
||
313 | mov ebx,[esi+ecx*4] |
||
314 | |||
315 | and eax,03E003E0h |
||
316 | mov edx,[esi+ecx*4] |
||
317 | |||
318 | and ebx,0F800F800h |
||
319 | |||
320 | shr ebx,11 |
||
321 | and edx,001F001Fh |
||
322 | |||
323 | shl edx,10 |
||
324 | add eax,ebx |
||
325 | |||
326 | add eax,edx |
||
327 | inc ecx |
||
328 | |||
329 | jnz .L5 |
||
330 | |||
331 | mov [edi+ecx*4-4],eax |
||
332 | |||
333 | ; tail |
||
334 | pop ecx |
||
335 | and ecx,BYTE 1 |
||
336 | jz .L7 |
||
337 | mov al,[esi] |
||
338 | mov ah,[esi+1] |
||
339 | mov ebx,eax |
||
340 | mov edx,eax |
||
341 | shr eax,11 |
||
342 | and eax,BYTE 11111b |
||
343 | shr ebx,1 |
||
344 | and ebx,1111100000b |
||
345 | shl edx,10 |
||
346 | and edx,0111110000000000b |
||
347 | add eax,ebx |
||
348 | add eax,edx |
||
349 | mov [edi],al |
||
350 | mov [edi+1],ah |
||
351 | add esi,BYTE 2 |
||
352 | add edi,BYTE 2 |
||
353 | |||
9172 | turbocat | 354 | .L7: |
355 | retn |
||
8210 | maxcodehac | 356 | |
357 | |||
358 | |||
359 | |||
360 | |||
9202 | turbocat | 361 | ConvertX86p16_8RGB332: |
8210 | maxcodehac | 362 | _ConvertX86p16_8RGB332: |
363 | |||
364 | ; check short |
||
365 | cmp ecx,BYTE 16 |
||
366 | ja .L3 |
||
367 | |||
368 | |||
9172 | turbocat | 369 | .L1: ; short loop |
8210 | maxcodehac | 370 | mov al,[esi+0] |
371 | mov ah,[esi+1] |
||
372 | mov ebx,eax |
||
373 | mov edx,eax |
||
374 | and eax,BYTE 11000b ; blue |
||
375 | shr eax,3 |
||
376 | and ebx,11100000000b ; green |
||
377 | shr ebx,6 |
||
378 | and edx,1110000000000000b ; red |
||
379 | shr edx,8 |
||
380 | add eax,ebx |
||
381 | add eax,edx |
||
382 | mov [edi],al |
||
383 | add esi,BYTE 2 |
||
384 | inc edi |
||
385 | dec ecx |
||
386 | jnz .L1 |
||
9172 | turbocat | 387 | .L2: |
388 | retn |
||
8210 | maxcodehac | 389 | |
9172 | turbocat | 390 | .L3: mov eax,edi |
8210 | maxcodehac | 391 | and eax,BYTE 11b |
392 | jz .L4 |
||
393 | mov al,[esi+0] |
||
394 | mov ah,[esi+1] |
||
395 | mov ebx,eax |
||
396 | mov edx,eax |
||
397 | and eax,BYTE 11000b ; blue |
||
398 | shr eax,3 |
||
399 | and ebx,11100000000b ; green |
||
400 | shr ebx,6 |
||
401 | and edx,1110000000000000b ; red |
||
402 | shr edx,8 |
||
403 | add eax,ebx |
||
404 | add eax,edx |
||
405 | mov [edi],al |
||
406 | add esi,BYTE 2 |
||
407 | inc edi |
||
408 | dec ecx |
||
409 | jmp SHORT .L3 |
||
410 | |||
9172 | turbocat | 411 | .L4: ; save ebp |
8210 | maxcodehac | 412 | push ebp |
413 | |||
414 | ; save count |
||
415 | push ecx |
||
416 | |||
417 | ; unroll 4 times |
||
418 | shr ecx,2 |
||
419 | |||
420 | ; prestep |
||
421 | mov dl,[esi+0] |
||
422 | mov bl,[esi+1] |
||
423 | mov dh,[esi+2] |
||
424 | |||
9172 | turbocat | 425 | .L5: shl edx,16 |
8210 | maxcodehac | 426 | mov bh,[esi+3] |
427 | |||
428 | shl ebx,16 |
||
429 | mov dl,[esi+4] |
||
430 | |||
431 | mov dh,[esi+6] |
||
432 | mov bl,[esi+5] |
||
433 | |||
434 | and edx,00011000000110000001100000011000b |
||
435 | mov bh,[esi+7] |
||
436 | |||
437 | ror edx,16+3 |
||
438 | mov eax,ebx ; setup eax for reds |
||
439 | |||
440 | and ebx,00000111000001110000011100000111b |
||
441 | and eax,11100000111000001110000011100000b ; reds |
||
442 | |||
443 | ror ebx,16-2 |
||
444 | add esi,BYTE 8 |
||
445 | |||
446 | ror eax,16 |
||
447 | add edi,BYTE 4 |
||
448 | |||
449 | add eax,ebx |
||
450 | mov bl,[esi+1] ; greens |
||
451 | |||
452 | add eax,edx |
||
453 | mov dl,[esi+0] ; blues |
||
454 | |||
455 | mov [edi-4],eax |
||
456 | mov dh,[esi+2] |
||
457 | |||
458 | dec ecx |
||
459 | jnz .L5 |
||
460 | |||
461 | ; check tail |
||
462 | pop ecx |
||
463 | and ecx,BYTE 11b |
||
464 | jz .L7 |
||
465 | |||
9172 | turbocat | 466 | .L6: ; tail |
8210 | maxcodehac | 467 | mov al,[esi+0] |
468 | mov ah,[esi+1] |
||
469 | mov ebx,eax |
||
470 | mov edx,eax |
||
471 | and eax,BYTE 11000b ; blue |
||
472 | shr eax,3 |
||
473 | and ebx,11100000000b ; green |
||
474 | shr ebx,6 |
||
475 | and edx,1110000000000000b ; red |
||
476 | shr edx,8 |
||
477 | add eax,ebx |
||
478 | add eax,edx |
||
479 | mov [edi],al |
||
480 | add esi,BYTE 2 |
||
481 | inc edi |
||
482 | dec ecx |
||
483 | jnz .L6 |
||
484 | |||
9172 | turbocat | 485 | .L7: pop ebp |
486 | retn |
||
8210 | maxcodehac | 487 | |
9172 | turbocat | 488 | %ifidn __OUTPUT_FORMAT__,elf32 |
489 | section .note.GNU-stack noalloc noexec nowrite progbits |
||
490 | %endif |