ca482e5d |
1 | @ vim:filetype=armasm\r |
cc68a136 |
2 | @ some color conversion and blitting routines\r |
3 | \r |
4 | @ (c) Copyright 2006, notaz\r |
5 | @ All Rights Reserved\r |
6 | \r |
7 | \r |
8 | @ Convert 0000bbb0 ggg0rrr0 0000bbb0 ggg0rrr0\r |
9 | @ to 00000000 rrr00000 ggg00000 bbb00000 ...\r |
10 | \r |
11 | @ lr = 0x00e000e0, out: r3=lower_pix, r2=higher_pix; trashes rin\r |
12 | @ if sh==2, r8=0x00404040 (sh!=0 destroys flags!)\r |
13 | .macro convRGB32_2 rin sh=0\r |
14 | and r2, lr, \rin, lsr #4 @ blue\r |
15 | and r3, \rin, lr\r |
16 | orr r2, r2, r3, lsl #8 @ g0b0g0b0\r |
17 | \r |
18 | mov r3, r2, lsl #16 @ g0b00000\r |
19 | and \rin,lr, \rin, ror #12 @ 00r000r0 (reversed)\r |
20 | orr r3, r3, \rin, lsr #16 @ g0b000r0\r |
21 | .if \sh == 1\r |
22 | mov r3, r3, ror #17 @ shadow mode\r |
23 | .elseif \sh == 2\r |
24 | adds r3, r3, #0x40000000 @ green\r |
25 | orrcs r3, r3, #0xe0000000\r |
26 | mov r3, r3, ror #8\r |
27 | adds r3, r3, #0x40000000\r |
28 | orrcs r3, r3, #0xe0000000\r |
29 | mov r3, r3, ror #16\r |
30 | adds r3, r3, #0x40000000\r |
31 | orrcs r3, r3, #0xe0000000\r |
32 | mov r3, r3, ror #24\r |
33 | orr r3, r3, r3, lsr #3\r |
34 | .else\r |
35 | mov r3, r3, ror #16 @ r3=low\r |
36 | orr r3, r3, r3, lsr #3\r |
37 | .endif\r |
38 | \r |
39 | str r3, [r0], #4\r |
40 | \r |
41 | mov r2, r2, lsr #16\r |
42 | orr r2, r2, \rin, lsl #16\r |
43 | .if \sh == 1\r |
44 | mov r2, r2, lsr #1\r |
45 | .elseif \sh == 2\r |
46 | mov r2, r2, ror #8\r |
47 | adds r2, r2, #0x40000000 @ blue\r |
48 | orrcs r2, r2, #0xe0000000\r |
49 | mov r2, r2, ror #8\r |
50 | adds r2, r2, #0x40000000\r |
51 | orrcs r2, r2, #0xe0000000\r |
52 | mov r2, r2, ror #8\r |
53 | adds r2, r2, #0x40000000\r |
54 | orrcs r2, r2, #0xe0000000\r |
55 | mov r2, r2, ror #8\r |
56 | orr r2, r2, r2, lsr #3\r |
57 | .else\r |
58 | orr r2, r2, r2, lsr #3\r |
59 | .endif\r |
60 | \r |
61 | str r2, [r0], #4\r |
62 | .endm\r |
63 | \r |
64 | \r |
65 | .global vidConvCpyRGB32 @ void *to, void *from, int pixels\r |
66 | \r |
67 | vidConvCpyRGB32:\r |
68 | stmfd sp!, {r4-r7,lr}\r |
69 | \r |
70 | mov r12, r2, lsr #3 @ repeats\r |
71 | mov lr, #0x00e00000\r |
72 | orr lr, lr, #0x00e0\r |
73 | \r |
74 | .loopRGB32:\r |
75 | subs r12, r12, #1\r |
76 | \r |
77 | ldmia r1!, {r4-r7}\r |
78 | convRGB32_2 r4\r |
79 | convRGB32_2 r5\r |
80 | convRGB32_2 r6\r |
81 | convRGB32_2 r7\r |
82 | \r |
83 | bgt .loopRGB32\r |
84 | \r |
85 | ldmfd sp!, {r4-r7,lr}\r |
86 | bx lr\r |
87 | \r |
88 | \r |
89 | .global vidConvCpyRGB32sh @ void *to, void *from, int pixels\r |
90 | \r |
91 | vidConvCpyRGB32sh:\r |
92 | stmfd sp!, {r4-r7,lr}\r |
93 | \r |
94 | mov r12, r2, lsr #3 @ repeats\r |
95 | mov lr, #0x00e00000\r |
96 | orr lr, lr, #0x00e0\r |
97 | \r |
98 | .loopRGB32sh:\r |
99 | subs r12, r12, #1\r |
100 | \r |
101 | ldmia r1!, {r4-r7}\r |
102 | convRGB32_2 r4, 1\r |
103 | convRGB32_2 r5, 1\r |
104 | convRGB32_2 r6, 1\r |
105 | convRGB32_2 r7, 1\r |
106 | \r |
107 | bgt .loopRGB32sh\r |
108 | \r |
109 | ldmfd sp!, {r4-r7,lr}\r |
110 | bx lr\r |
111 | \r |
112 | \r |
113 | .global vidConvCpyRGB32hi @ void *to, void *from, int pixels\r |
114 | \r |
115 | vidConvCpyRGB32hi:\r |
116 | stmfd sp!, {r4-r7,lr}\r |
117 | \r |
118 | mov r12, r2, lsr #3 @ repeats\r |
119 | mov lr, #0x00e00000\r |
120 | orr lr, lr, #0x00e0\r |
121 | \r |
122 | .loopRGB32hi:\r |
123 | ldmia r1!, {r4-r7}\r |
124 | convRGB32_2 r4, 2\r |
125 | convRGB32_2 r5, 2\r |
126 | convRGB32_2 r6, 2\r |
127 | convRGB32_2 r7, 2\r |
128 | \r |
129 | subs r12, r12, #1\r |
130 | bgt .loopRGB32hi\r |
131 | \r |
132 | ldmfd sp!, {r4-r7,lr}\r |
133 | bx lr\r |
134 | \r |
135 | \r |
136 | @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@\r |
137 | \r |
138 | @ -------- M2 stuff ---------\r |
139 | /*\r |
140 | .global vidConvCpy_90 @ void *to, void *from, int width\r |
141 | \r |
142 | vidConvCpy_90:\r |
143 | stmfd sp!, {r4-r10,lr}\r |
144 | \r |
145 | mov lr, #0x00F00000\r |
146 | orr lr, lr, #0x00F0\r |
147 | \r |
148 | mov r12, #224/4 @ row counter\r |
149 | mov r10, r2, lsl #2 @ we do 2 pixel wide copies\r |
150 | \r |
151 | add r8, r0, #256*4 @ parallel line\r |
152 | add r1, r1, #0x23000\r |
153 | add r1, r1, #0x00B80 @ r1+=328*223*2+8*2\r |
154 | mov r9, r1\r |
155 | \r |
156 | mov r4, #0 @ fill bottom border\r |
157 | mov r5, #0\r |
158 | mov r6, #0\r |
159 | mov r7, #0\r |
160 | stmia r0!, {r4-r7}\r |
161 | stmia r0!, {r4-r7}\r |
162 | stmia r8!, {r4-r7}\r |
163 | stmia r8!, {r4-r7}\r |
164 | \r |
165 | .loopM2RGB32_90:\r |
166 | subs r12, r12, #1\r |
167 | \r |
168 | @ at first this loop was written differently: src pixels were fetched with ldm's and\r |
169 | @ dest was not sequential. It ran nearly 2 times slower. It seems it is very important\r |
170 | @ to do sequential memory access on those items, which we have more (to offload addressing bus?).\r |
171 | \r |
172 | ldr r4, [r1], #-328*2\r |
173 | ldr r5, [r1], #-328*2\r |
174 | ldr r6, [r1], #-328*2\r |
175 | ldr r7, [r1], #-328*2\r |
176 | \r |
177 | convRGB32_2 r4, 1\r |
178 | convRGB32_2 r5, 1\r |
179 | convRGB32_2 r6, 1\r |
180 | convRGB32_2 r7, 1\r |
181 | \r |
182 | str r4, [r8], #4\r |
183 | str r5, [r8], #4\r |
184 | str r6, [r8], #4\r |
185 | str r7, [r8], #4\r |
186 | \r |
187 | bne .loopM2RGB32_90\r |
188 | \r |
189 | mov r4, #0 @ top border\r |
190 | mov r5, #0\r |
191 | mov r6, #0\r |
192 | stmia r0!, {r4-r6,r12}\r |
193 | stmia r0!, {r4-r6,r12}\r |
194 | stmia r8!, {r4-r6,r12}\r |
195 | stmia r8!, {r4-r6,r12}\r |
196 | \r |
197 | subs r10, r10, #1\r |
198 | ldmeqfd sp!, {r4-r10,pc} @ return\r |
199 | \r |
200 | add r0, r8, #16*4 @ set new dst pointer\r |
201 | add r8, r0, #256*4\r |
202 | add r9, r9, #2*2 @ fix src pointer\r |
203 | mov r1, r9\r |
204 | \r |
205 | stmia r0!, {r4-r6,r12} @ bottom border\r |
206 | stmia r0!, {r4-r6,r12}\r |
207 | stmia r8!, {r4-r6,r12}\r |
208 | stmia r8!, {r4-r6,r12}\r |
209 | \r |
210 | mov r12, #224/4 @ restore row counter\r |
211 | b .loopM2RGB32_90\r |
212 | \r |
213 | \r |
214 | \r |
215 | @ converter for vidConvCpy_270\r |
216 | @ lr = 0x00F000F0, out: r3=lower_pix, r2=higher_pix; trashes rin\r |
217 | .macro convRGB32_3 rin\r |
218 | and r2, lr, \rin, lsr #4 @ blue\r |
219 | and r3, \rin, lr\r |
220 | orr r2, r2, r3, lsl #8 @ g0b0g0b0\r |
221 | \r |
222 | mov r3, r2, lsl #16 @ g0b00000\r |
223 | and \rin,lr, \rin, ror #12 @ 00r000r0 (reversed)\r |
224 | orr r3, r3, \rin, lsr #16 @ g0b000r0\r |
225 | \r |
226 | mov r2, r2, lsr #16\r |
227 | orr r2, r2, \rin, lsl #16\r |
228 | str r2, [r0], #4\r |
229 | \r |
230 | mov \rin,r3, ror #16 @ r3=low\r |
231 | .endm\r |
232 | */\r |
233 | @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@\r |
234 | \r |
235 | \r |
236 | @ takes byte-sized pixels from r3-r6, fetches from pal and stores to r7,r8,r10,lr\r |
237 | @ r2=pal\r |
238 | .macro mode2_4pix shift\r |
239 | and r7, r11, r3, lsr #\shift\r |
240 | ldr r7, [r2, r7, lsl #2]\r |
241 | \r |
242 | and r8, r11, r4, lsr #\shift\r |
243 | ldr r8, [r2, r8, lsl #2]\r |
244 | \r |
245 | and r10,r11, r5, lsr #\shift\r |
246 | ldr r10,[r2, r10,lsl #2]\r |
247 | \r |
248 | and lr, r11, r6, lsr #\shift\r |
249 | ldr lr, [r2, lr, lsl #2]\r |
250 | .endm\r |
251 | \r |
252 | @ r2=pal, r11=0xff\r |
253 | .macro mode2_4pix_getpix0 dreg sreg\r |
254 | and \dreg, r11, \sreg\r |
255 | ldr \dreg, [r2, \dreg, lsl #2]\r |
256 | .endm\r |
257 | \r |
258 | .macro mode2_4pix_getpix1 dreg sreg\r |
259 | and \dreg, r11, \sreg, lsr #8\r |
260 | ldr \dreg, [r2, \dreg, lsl #2]\r |
261 | .endm\r |
262 | \r |
263 | .macro mode2_4pix_getpix2 dreg sreg\r |
264 | and \dreg, r11, \sreg, lsr #16\r |
265 | ldr \dreg, [r2, \dreg, lsl #2]\r |
266 | .endm\r |
267 | \r |
268 | .macro mode2_4pix_getpix3 dreg sreg\r |
269 | and \dreg, r11, \sreg, lsr #24\r |
270 | ldr \dreg, [r2, \dreg, lsl #2]\r |
271 | .endm\r |
272 | \r |
273 | @ takes byte-sized pixels from reg, fetches from pal and stores to r3-r6\r |
274 | @ r11=0xFF, r2=pal\r |
275 | .macro mode2_4pix2_0 reg\r |
276 | mode2_4pix_getpix0 r3, \reg\r |
277 | mode2_4pix_getpix1 r4, \reg\r |
278 | mode2_4pix_getpix2 r5, \reg\r |
279 | mode2_4pix_getpix3 r6, \reg\r |
280 | .endm\r |
281 | \r |
282 | @ ...\r |
283 | .macro mode2_4pix2_180 reg\r |
284 | mode2_4pix_getpix3 r3, \reg\r |
285 | mode2_4pix_getpix2 r4, \reg\r |
286 | mode2_4pix_getpix1 r5, \reg\r |
287 | mode2_4pix_getpix0 r6, \reg\r |
288 | .endm\r |
289 | \r |
290 | @ takes byte-sized pixels from reg, fetches from pal and stores to r3-r5\r |
291 | @ r11=0xFF, r2=pal, r10=0xfcfcfc, r6=tmp\r |
292 | .macro mode2_4pix_to3 reg is180\r |
293 | .if \is180\r |
294 | mode2_4pix_getpix3 r3, \reg\r |
295 | mode2_4pix_getpix2 r4, \reg\r |
296 | .else\r |
297 | mode2_4pix_getpix0 r3, \reg @ gathering loads cause a weird-hang\r |
298 | mode2_4pix_getpix1 r4, \reg\r |
299 | .endif\r |
300 | \r |
301 | sub r3, r3, r3, lsr #2 @ r3 *= 0.75\r |
302 | add r3, r3, r4, lsr #2 @ r3 += r4 * 0.25\r |
303 | and r3, r3, r10\r |
304 | \r |
305 | .if \is180\r |
306 | mode2_4pix_getpix1 r5, \reg\r |
307 | mode2_4pix_getpix0 r6, \reg\r |
308 | .else\r |
309 | mode2_4pix_getpix2 r5, \reg\r |
310 | mode2_4pix_getpix3 r6, \reg\r |
311 | .endif\r |
312 | \r |
313 | mov r4, r4, lsr #1\r |
314 | add r4, r4, r5, lsr #1 @ r4 = (r4 + r5) / 2;\r |
315 | @ and r4, r4, r10\r |
316 | sub r6, r6, r6, lsr #2 @ r6 *= 0.75\r |
317 | add r5, r6, r5, lsr #2 @ r5 = r6 + r5 * 0.25\r |
318 | and r5, r5, r10\r |
319 | .endm\r |
320 | \r |
321 | \r |
322 | @ void *to, void *from, void *pal, int width\r |
323 | .macro vidConvCpyM2_landscape is270\r |
324 | stmfd sp!, {r4-r11,lr}\r |
325 | \r |
326 | mov r11, #0xff\r |
327 | \r |
328 | mov r12, #(224/4-1)<<16 @ row counter\r |
329 | orr r12, r12, r3, lsl #1 @ we do 4 pixel wide copies (right to left)\r |
330 | \r |
331 | .if \is270\r |
332 | add r1, r1, #324\r |
333 | .else\r |
334 | add r1, r1, #0x11c00\r |
335 | add r1, r1, #0x00308 @ 328*224+8\r |
336 | .endif\r |
337 | mov r9, r1\r |
338 | \r |
339 | mov r3, #0 @ fill top border\r |
340 | mov r4, #0\r |
341 | mov r5, #0\r |
342 | mov r6, #0\r |
343 | stmia r0!, {r3-r6}\r |
344 | stmia r0!, {r3-r6}\r |
345 | add r7, r0, #256*4-8*4\r |
346 | stmia r7!, {r3-r6}\r |
347 | stmia r7!, {r3-r6}\r |
348 | add r7, r7, #256*4-8*4\r |
349 | stmia r7!, {r3-r6}\r |
350 | stmia r7!, {r3-r6}\r |
351 | add r7, r7, #256*4-8*4\r |
352 | stmia r7!, {r3-r6}\r |
353 | stmia r7!, {r3-r6}\r |
354 | \r |
355 | 0: @ .loopM2RGB32_270:\r |
356 | subs r12, r12, #1<<16\r |
357 | \r |
358 | .if \is270\r |
359 | ldr r3, [r1], #328\r |
360 | ldr r4, [r1], #328\r |
361 | ldr r5, [r1], #328\r |
362 | ldr r6, [r1], #328\r |
363 | .else\r |
364 | ldr r3, [r1, #-328]!\r |
365 | ldr r4, [r1, #-328]!\r |
366 | ldr r5, [r1, #-328]!\r |
367 | ldr r6, [r1, #-328]!\r |
368 | .endif\r |
369 | \r |
370 | .if \is270\r |
371 | mode2_4pix 24\r |
372 | .else\r |
373 | mode2_4pix 0\r |
374 | .endif\r |
375 | stmia r0, {r7,r8,r10,lr}\r |
376 | add r0, r0, #256*4\r |
377 | \r |
378 | .if \is270\r |
379 | mode2_4pix 16\r |
380 | .else\r |
381 | mode2_4pix 8\r |
382 | .endif\r |
383 | stmia r0, {r7,r8,r10,lr}\r |
384 | add r0, r0, #256*4\r |
385 | \r |
386 | .if \is270\r |
387 | mode2_4pix 8\r |
388 | .else\r |
389 | mode2_4pix 16\r |
390 | .endif\r |
391 | stmia r0, {r7,r8,r10,lr}\r |
392 | add r0, r0, #256*4\r |
393 | \r |
394 | .if \is270\r |
395 | mode2_4pix 0\r |
396 | .else\r |
397 | mode2_4pix 24\r |
398 | .endif\r |
399 | stmia r0!,{r7,r8,r10,lr}\r |
400 | sub r0, r0, #256*4*3\r |
401 | \r |
402 | bpl 0b @ .loopM2RGB32_270\r |
403 | \r |
404 | mov r3, #0 @ bottom border\r |
405 | mov r4, #0\r |
406 | mov r5, #0\r |
407 | mov r6, #0\r |
408 | stmia r0!, {r3-r6}\r |
409 | stmia r0!, {r3-r6}\r |
410 | add r0, r0, #256*4-8*4\r |
411 | stmia r0!, {r3-r6}\r |
412 | stmia r0!, {r3-r6}\r |
413 | add r0, r0, #256*4-8*4\r |
414 | stmia r0!, {r3-r6}\r |
415 | stmia r0!, {r3-r6}\r |
416 | add r0, r0, #256*4-8*4\r |
417 | stmia r0!, {r3-r6}\r |
418 | nop @ phone crashes if this is commented out. Do I stress it too much?\r |
419 | stmia r0!, {r3-r6}\r |
420 | \r |
421 | add r12, r12, #1<<16\r |
422 | subs r12, r12, #1\r |
423 | ldmeqfd sp!, {r4-r11,pc} @ return\r |
424 | \r |
425 | add r0, r0, #16*4\r |
426 | .if \is270\r |
427 | sub r9, r9, #4 @ fix src pointer\r |
428 | .else\r |
429 | add r9, r9, #4\r |
430 | .endif\r |
431 | mov r1, r9\r |
432 | \r |
433 | stmia r0!, {r3-r6} @ top border\r |
434 | stmia r0!, {r3-r6}\r |
435 | add r7, r0, #256*4-8*4\r |
436 | stmia r7!, {r3-r6}\r |
437 | stmia r7!, {r3-r6}\r |
438 | add r7, r7, #256*4-8*4\r |
439 | stmia r7!, {r3-r6}\r |
440 | stmia r7!, {r3-r6}\r |
441 | add r7, r7, #256*4-8*4\r |
442 | stmia r7!, {r3-r6}\r |
443 | stmia r7!, {r3-r6}\r |
444 | \r |
445 | orr r12, r12, #(224/4-1)<<16 @ restore row counter\r |
446 | b 0b @ .loopM2RGB32_270\r |
447 | .endm\r |
448 | \r |
449 | \r |
450 | .global vidConvCpy_90 @ void *to, void *from, void *pal, int width\r |
451 | \r |
452 | vidConvCpy_90:\r |
453 | vidConvCpyM2_landscape 0\r |
454 | \r |
455 | \r |
456 | .global vidConvCpy_270 @ void *to, void *from, void *pal, int width\r |
457 | \r |
458 | vidConvCpy_270:\r |
459 | vidConvCpyM2_landscape 1\r |
460 | \r |
461 | \r |
462 | .global vidConvCpy_center_0 @ void *to, void *from, void *pal\r |
463 | \r |
464 | vidConvCpy_center_0:\r |
465 | stmfd sp!, {r4-r6,r11,lr}\r |
466 | \r |
467 | mov r11, #0xff\r |
468 | add r1, r1, #8 @ not border (centering 32col here)\r |
469 | \r |
470 | mov r12, #(240/4-1)<<16\r |
471 | orr r12, r12, #224\r |
472 | \r |
473 | .loopRGB32_c0:\r |
474 | ldr lr, [r1], #4\r |
475 | subs r12, r12, #1<<16\r |
476 | \r |
477 | mode2_4pix2_0 lr\r |
478 | stmia r0!, {r3-r6}\r |
479 | bpl .loopRGB32_c0\r |
480 | \r |
481 | sub r12, r12, #1\r |
482 | adds r12, r12, #1<<16\r |
483 | ldmeqfd sp!, {r4-r6,r11,pc} @ return\r |
484 | add r0, r0, #16*4\r |
485 | add r1, r1, #88\r |
486 | orr r12, #(240/4-1)<<16\r |
487 | b .loopRGB32_c0\r |
488 | \r |
489 | \r |
490 | .global vidConvCpy_center_180 @ void *to, void *from, void *pal\r |
491 | \r |
492 | vidConvCpy_center_180:\r |
493 | stmfd sp!, {r4-r6,r11,lr}\r |
494 | \r |
495 | mov r11, #0xff\r |
496 | add r1, r1, #0x11c00\r |
497 | add r1, r1, #0x002B8 @ #328*224-72\r |
498 | \r |
499 | mov r12, #(240/4-1)<<16\r |
500 | orr r12, r12, #224\r |
501 | \r |
502 | .loopRGB32_c180:\r |
503 | ldr lr, [r1, #-4]!\r |
504 | subs r12, r12, #1<<16\r |
505 | \r |
506 | mode2_4pix2_180 lr\r |
507 | stmia r0!, {r3-r6}\r |
508 | bpl .loopRGB32_c180\r |
509 | \r |
510 | sub r12, r12, #1\r |
511 | adds r12, r12, #1<<16\r |
512 | ldmeqfd sp!, {r4-r6,r11,pc} @ return\r |
513 | add r0, r0, #16*4\r |
514 | sub r1, r1, #88\r |
515 | orr r12, #(240/4-1)<<16\r |
516 | b .loopRGB32_c180\r |
517 | \r |
518 | \r |
519 | @ note: the following code assumes that (pal[x] & 0x030303) == 0\r |
520 | \r |
521 | .global vidConvCpy_center2_40c_0 @ void *to, void *from, void *pal, int lines\r |
522 | \r |
523 | vidConvCpy_center2_40c_0:\r |
524 | stmfd sp!, {r4-r6,r10,r11,lr}\r |
525 | \r |
526 | mov r11, #0xff\r |
527 | mov r10, #0xfc\r |
528 | orr r10, r10, lsl #8\r |
529 | orr r10, r10, lsl #8\r |
530 | add r1, r1, #8 @ border\r |
531 | \r |
532 | mov r12, #(240/3-1)<<16\r |
533 | orr r12, r12, r3\r |
534 | \r |
535 | .loopRGB32_c2_40c_0:\r |
536 | ldr lr, [r1], #4\r |
537 | subs r12, r12, #1<<16\r |
538 | \r |
539 | mode2_4pix_to3 lr, 0\r |
540 | \r |
541 | stmia r0!, {r3-r5}\r |
542 | bpl .loopRGB32_c2_40c_0\r |
543 | \r |
544 | sub r12, r12, #1\r |
545 | adds r12, r12, #1<<16\r |
546 | ldmeqfd sp!, {r4-r6,r10,r11,pc} @ return\r |
547 | add r0, r0, #16*4\r |
548 | add r1, r1, #8\r |
549 | orr r12, #(240/3-1)<<16\r |
550 | b .loopRGB32_c2_40c_0\r |
551 | \r |
552 | \r |
553 | .global vidConvCpy_center2_40c_180 @ void *to, void *from, void *pal, int lines\r |
554 | \r |
555 | vidConvCpy_center2_40c_180:\r |
556 | stmfd sp!, {r4-r6,r10,r11,lr}\r |
557 | \r |
558 | mov r11, #0xff\r |
559 | mov r10, #0xfc\r |
560 | orr r10, r10, lsl #8\r |
561 | orr r10, r10, lsl #8\r |
562 | \r |
563 | mov r4, #328\r |
564 | mla r1, r3, r4, r1\r |
565 | @ add r1, r1, #0x11000\r |
566 | @ add r1, r1, #0x00f00 @ #328*224\r |
567 | \r |
568 | mov r12, #(240/3-1)<<16\r |
569 | orr r12, r12, r3\r |
570 | \r |
571 | .loop_c2_40c_180:\r |
572 | ldr lr, [r1, #-4]!\r |
573 | subs r12, r12, #1<<16\r |
574 | \r |
575 | mode2_4pix_to3 lr, 1\r |
576 | \r |
577 | stmia r0!, {r3-r5}\r |
578 | bpl .loop_c2_40c_180\r |
579 | \r |
580 | sub r12, r12, #1\r |
581 | adds r12, r12, #1<<16\r |
582 | ldmeqfd sp!, {r4-r6,r10,r11,pc} @ return\r |
583 | add r0, r0, #16*4\r |
584 | sub r1, r1, #8\r |
585 | orr r12, #(240/3-1)<<16\r |
586 | b .loop_c2_40c_180\r |
587 | \r |
588 | \r |
589 | .global vidConvCpy_center2_32c_0 @ void *to, void *from, void *pal, int lines\r |
590 | \r |
591 | vidConvCpy_center2_32c_0:\r |
592 | stmfd sp!, {r4-r11,lr}\r |
593 | \r |
594 | mov r10, #0xfc\r |
595 | orr r10, r10, lsl #8\r |
596 | orr r10, r10, lsl #8\r |
597 | mov r11, #0xff\r |
598 | add r1, r1, #8 @ border\r |
599 | \r |
600 | mov r12, #(240/15-1)<<16\r |
601 | orr r12, r12, r3\r |
602 | \r |
603 | .loop_c2_32c_0:\r |
604 | ldmia r1!, {r7-r9,lr}\r |
605 | subs r12, r12, #1<<16\r |
606 | \r |
607 | mode2_4pix2_0 r7\r |
608 | stmia r0!, {r3-r6}\r |
609 | mode2_4pix2_0 r8\r |
610 | stmia r0!, {r3-r6}\r |
611 | mode2_4pix2_0 r9\r |
612 | stmia r0!, {r3-r6}\r |
613 | mode2_4pix_to3 lr, 0\r |
614 | stmia r0!, {r3-r5}\r |
615 | bpl .loop_c2_32c_0\r |
616 | \r |
617 | sub r12, r12, #1\r |
618 | adds r12, r12, #1<<16\r |
619 | ldmeqfd sp!, {r4-r11,pc} @ return\r |
620 | add r0, r0, #16*4\r |
621 | add r1, r1, #64+8\r |
622 | orr r12, #(240/15-1)<<16\r |
623 | b .loop_c2_32c_0\r |
624 | \r |
625 | \r |
626 | .global vidConvCpy_center2_32c_180 @ void *to, void *from, void *pal, int lines\r |
627 | \r |
628 | vidConvCpy_center2_32c_180:\r |
629 | stmfd sp!, {r4-r11,lr}\r |
630 | \r |
631 | mov r10, #0xfc\r |
632 | orr r10, r10, lsl #8\r |
633 | orr r10, r10, lsl #8\r |
634 | mov r11, #0xff\r |
635 | \r |
636 | mov r4, #328\r |
637 | mla r1, r3, r4, r1\r |
638 | @ add r1, r1, #0x11000\r |
639 | @ add r1, r1, #0x00f00 @ #328*224\r |
640 | \r |
641 | mov r12, #(240/15-1)<<16\r |
642 | orr r12, r12, r3\r |
643 | \r |
644 | .loop_c2_32c_180:\r |
645 | ldmdb r1!, {r7-r9,lr}\r |
646 | subs r12, r12, #1<<16\r |
647 | \r |
648 | mode2_4pix2_180 lr\r |
649 | stmia r0!, {r3-r6}\r |
650 | mode2_4pix2_180 r9\r |
651 | stmia r0!, {r3-r6}\r |
652 | mode2_4pix2_180 r8\r |
653 | stmia r0!, {r3-r6}\r |
654 | mode2_4pix_to3 r7, 1\r |
655 | stmia r0!, {r3-r5}\r |
656 | bpl .loop_c2_32c_180\r |
657 | \r |
658 | sub r12, r12, #1\r |
659 | adds r12, r12, #1<<16\r |
660 | ldmeqfd sp!, {r4-r11,pc} @ return\r |
661 | add r0, r0, #16*4\r |
662 | sub r1, r1, #64+8\r |
663 | orr r12, #(240/15-1)<<16\r |
664 | b .loop_c2_32c_180\r |
665 | \r |
666 | \r |
667 | @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@\r |
668 | \r |
669 | \r |
670 | .global vidClear @ void *to, int lines\r |
671 | \r |
672 | vidClear:\r |
673 | stmfd sp!, {lr}\r |
674 | mov r12, #240/16-1\r |
675 | orr r12, r1, r12, lsl #16\r |
676 | mov r1, #0\r |
677 | mov r2, #0\r |
678 | mov r3, #0\r |
679 | mov lr, #0\r |
680 | \r |
681 | .loopVidClear:\r |
682 | subs r12, r12, #1<<16\r |
683 | \r |
684 | stmia r0!, {r1-r3,lr}\r |
685 | stmia r0!, {r1-r3,lr}\r |
686 | stmia r0!, {r1-r3,lr}\r |
687 | stmia r0!, {r1-r3,lr}\r |
688 | bpl .loopVidClear\r |
689 | \r |
690 | sub r12, r12, #1\r |
691 | adds r12, r12, #1<<16\r |
692 | ldmeqfd sp!, {pc} @ return\r |
693 | add r0, r0, #16*4\r |
694 | orr r12, #(240/16-1)<<16\r |
695 | b .loopVidClear\r |
696 | \r |
ca482e5d |
697 | @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@\r |
698 | \r |
699 | .equ EExecSetExceptionHandler, (90)\r |
700 | \r |
701 | .global my_SetExceptionHandler\r |
702 | \r |
703 | my_SetExceptionHandler:\r |
704 | mov ip, lr\r |
705 | swi EExecSetExceptionHandler\r |
706 | \r |