1 |
/* |
2 |
* PearPC |
3 |
* vaccel.S |
4 |
* |
5 |
* Copyright (C) 2004-2006 Sebastian Biallas (sb@biallas.net) |
6 |
* |
7 |
* This program is free software; you can redistribute it and/or modify |
8 |
* it under the terms of the GNU General Public License version 2 as |
9 |
* published by the Free Software Foundation. |
10 |
* |
11 |
* This program is distributed in the hope that it will be useful, |
12 |
* but WITHOUT ANY WARRANTY; without even the implied warranty of |
13 |
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
14 |
* GNU General Public License for more details. |
15 |
* |
16 |
* You should have received a copy of the GNU General Public License |
17 |
* along with this program; if not, write to the Free Software |
18 |
* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
19 |
*/ |
20 |
|
21 |
#ifndef PREFIX |
22 |
#define PREFIX |
23 |
#endif |
24 |
|
25 |
#define EXPORT(sym) EXPORT2(PREFIX, sym) |
26 |
#define EXPORT2(p, sym) EXPORT3(p, sym) |
27 |
#define EXPORT3(p, sym) .globl p##sym; p##sym |
28 |
|
29 |
#define EXTERN(sym) EXTERN2(PREFIX, sym) |
30 |
#define EXTERN2(p, sym) EXTERN3(p, sym) |
31 |
#define EXTERN3(p, sym) p##sym |
32 |
|
33 |
.intel_syntax |
34 |
|
35 |
.text |
36 |
|
37 |
.balign 16 |
38 |
d1: .long 0x00ff00ff |
39 |
.long 0x00ff00ff |
40 |
.long 0x00ff00ff |
41 |
.long 0x00ff00ff |
42 |
d2: .long 0xff00ff00 |
43 |
.long 0xff00ff00 |
44 |
.long 0xff00ff00 |
45 |
.long 0xff00ff00 |
46 |
|
47 |
_2be555_mask_r: .long 0x7c007c00 |
48 |
.long 0x7c007c00 |
49 |
.long 0x7c007c00 |
50 |
.long 0x7c007c00 |
51 |
_2be555_mask_g: .long 0x03e003e0 |
52 |
.long 0x03e003e0 |
53 |
.long 0x03e003e0 |
54 |
.long 0x03e003e0 |
55 |
_2be555_mask_b: .long 0x001f001f |
56 |
.long 0x001f001f |
57 |
.long 0x001f001f |
58 |
.long 0x001f001f |
59 |
|
60 |
.balign 16 |
61 |
################################################################################# |
62 |
## |
63 |
## IN: eax -- number of pixels to convert |
64 |
## %edx -- input |
65 |
## %ecx -- output |
66 |
|
67 |
EXPORT(x86_mmx_convert_2be555_to_2le555): |
68 |
add %eax, 7 |
69 |
shr %eax, 3 # we can convert 8 pixels at a time |
70 |
movq %mm5, [d1] |
71 |
movq %mm6, [d2] |
72 |
jz 2f |
73 |
1: |
74 |
movq %mm1, [%edx] |
75 |
movq %mm3, [%edx+8] |
76 |
|
77 |
## convert big to little endian |
78 |
movq %mm2, %mm1 |
79 |
movq %mm4, %mm3 |
80 |
pand %mm1, %mm5 |
81 |
pand %mm2, %mm6 |
82 |
pand %mm3, %mm5 |
83 |
pand %mm4, %mm6 |
84 |
psllw %mm1, 8 |
85 |
psrlw %mm2, 8 |
86 |
psllw %mm3, 8 |
87 |
psrlw %mm4, 8 |
88 |
por %mm1, %mm2 |
89 |
por %mm3, %mm4 |
90 |
|
91 |
movq [%ecx], %mm1 |
92 |
movq [%ecx+8], %mm3 |
93 |
add %edx, 16 |
94 |
add %ecx, 16 |
95 |
dec %eax |
96 |
jnz 1b |
97 |
|
98 |
emms |
99 |
2: |
100 |
ret |
101 |
|
102 |
.balign 16 |
103 |
################################################################################# |
104 |
## |
105 |
## IN: %eax -- number of pixels to convert |
106 |
## %edx -- input |
107 |
## %ecx -- output |
108 |
|
109 |
EXPORT(x86_mmx_convert_2be555_to_2le565): |
110 |
add %eax, 7 |
111 |
shr %eax, 3 # we can convert 8 pixels at a time |
112 |
movq %mm0, [d1] |
113 |
movq %mm7, [d2] |
114 |
jz 2f |
115 |
1: |
116 |
movq %mm1, [%edx] |
117 |
movq %mm3, [%edx+8] |
118 |
## convert big to little endian |
119 |
movq %mm2, %mm1 |
120 |
movq %mm4, %mm3 |
121 |
pand %mm1, %mm0 |
122 |
pand %mm2, %mm7 |
123 |
pand %mm3, %mm0 |
124 |
pand %mm4, %mm7 |
125 |
psllw %mm1, 8 |
126 |
psrlw %mm2, 8 |
127 |
psllw %mm3, 8 |
128 |
psrlw %mm4, 8 |
129 |
por %mm1, %mm2 |
130 |
por %mm4, %mm3 |
131 |
|
132 |
movq %mm2, %mm1 |
133 |
movq %mm3, %mm1 |
134 |
movq %mm5, %mm4 |
135 |
movq %mm6, %mm4 |
136 |
pand %mm1, [_2be555_mask_r] |
137 |
pand %mm2, [_2be555_mask_g] |
138 |
pand %mm3, [_2be555_mask_b] |
139 |
pand %mm4, [_2be555_mask_r] |
140 |
pand %mm5, [_2be555_mask_g] |
141 |
pand %mm6, [_2be555_mask_b] |
142 |
psllw %mm1, 1 # red |
143 |
psllw %mm2, 1 # green |
144 |
# psllw %mm3, 0 # blue |
145 |
psllw %mm4, 1 # red |
146 |
psllw %mm5, 1 # green |
147 |
# psllw %mm6, 0 # blue |
148 |
por %mm1, %mm2 |
149 |
por %mm4, %mm5 |
150 |
por %mm1, %mm3 |
151 |
por %mm4, %mm6 |
152 |
movq [%ecx], %mm1 |
153 |
movq [%ecx+8], %mm4 |
154 |
add %edx, 16 |
155 |
add %ecx, 16 |
156 |
dec %eax |
157 |
jnz 1b |
158 |
|
159 |
emms |
160 |
2: |
161 |
ret |
162 |
|
163 |
.balign 16 |
164 |
################################################################################# |
165 |
## |
166 |
## IN: %eax -- number of pixels to convert |
167 |
## %edx -- input |
168 |
## %ecx -- output |
169 |
|
170 |
EXPORT(x86_mmx_convert_2be555_to_4le888): |
171 |
add %eax, 3 |
172 |
shr %eax, 2 # we can convert 4 pixels at a time |
173 |
movq %mm7, [d1] |
174 |
jz 2f |
175 |
|
176 |
pxor %mm0, %mm0 |
177 |
1: |
178 |
movq %mm1, [%edx] |
179 |
|
180 |
## convert big to little endian |
181 |
movq %mm3, %mm1 |
182 |
pand %mm1, %mm7 |
183 |
pand %mm3, [d2] |
184 |
psllw %mm1, 8 |
185 |
psrlw %mm3, 8 |
186 |
por %mm1, %mm3 |
187 |
|
188 |
movq %mm2, %mm1 |
189 |
movq %mm3, %mm1 |
190 |
pand %mm1, [_2be555_mask_r] |
191 |
pand %mm2, [_2be555_mask_g] |
192 |
pand %mm3, [_2be555_mask_b] |
193 |
movq %mm4, %mm1 |
194 |
movq %mm5, %mm2 |
195 |
movq %mm6, %mm3 |
196 |
punpcklwd %mm1, %mm0 |
197 |
punpcklwd %mm2, %mm0 |
198 |
punpcklwd %mm3, %mm0 |
199 |
punpckhwd %mm4, %mm0 |
200 |
punpckhwd %mm5, %mm0 |
201 |
punpckhwd %mm6, %mm0 |
202 |
pslld %mm1, 16-10+3 # red |
203 |
pslld %mm2, 8-5+3 # green |
204 |
pslld %mm3, 0+3 # blue |
205 |
pslld %mm4, 16-10+3 # red |
206 |
pslld %mm5, 8-5+3 # green |
207 |
pslld %mm6, 0+3 # blue |
208 |
por %mm1, %mm2 |
209 |
por %mm1, %mm3 |
210 |
por %mm4, %mm5 |
211 |
por %mm4, %mm6 |
212 |
movq [%ecx], %mm1 |
213 |
movq [%ecx+8], %mm4 |
214 |
add %edx, 8 |
215 |
add %ecx, 16 |
216 |
dec %eax |
217 |
jnz 1b |
218 |
|
219 |
2: |
220 |
emms |
221 |
ret |
222 |
|
223 |
|
224 |
.balign 16 |
225 |
################################################################################# |
226 |
## |
227 |
## IN: %eax -- number of pixels to convert |
228 |
## %edx -- input |
229 |
## %ecx -- output |
230 |
|
231 |
EXPORT(x86_convert_4be888_to_4le888): |
232 |
add %eax, 3 |
233 |
shr %eax, 2 # we can convert 4 pixels at a time |
234 |
jz 2f |
235 |
|
236 |
push %ebx |
237 |
push %ebp |
238 |
push %esi |
239 |
push %edi |
240 |
1: |
241 |
mov %ebx, [%edx] |
242 |
mov %ebp, [%edx+4] |
243 |
mov %esi, [%edx+8] |
244 |
mov %edi, [%edx+12] |
245 |
## convert big to little endian |
246 |
bswap %ebx |
247 |
bswap %ebp |
248 |
bswap %esi |
249 |
bswap %edi |
250 |
add %edx, 16 |
251 |
mov [%ecx], %ebx |
252 |
mov [%ecx+4], %ebp |
253 |
mov [%ecx+8], %esi |
254 |
mov [%ecx+12], %edi |
255 |
add %ecx, 16 |
256 |
dec %eax |
257 |
jnz 1b |
258 |
|
259 |
pop %edi |
260 |
pop %esi |
261 |
pop %ebp |
262 |
pop %ebx |
263 |
2: |
264 |
ret |
265 |
|
266 |
.balign 16 |
267 |
################################################################################# |
268 |
## |
269 |
## IN: eax -- number of pixels to convert |
270 |
## %edx -- input |
271 |
## %ecx -- output |
272 |
|
273 |
EXPORT(x86_sse2_convert_2be555_to_2le555): |
274 |
add %eax, 15 |
275 |
shr %eax, 4 # we can convert 16 pixels at a time |
276 |
movdqa %xmm5, [d1] |
277 |
movdqa %xmm6, [d2] |
278 |
jz 2f |
279 |
1: |
280 |
movdqa %xmm1, [%edx] |
281 |
movdqa %xmm3, [%edx+16] |
282 |
|
283 |
## convert big to little endian |
284 |
movdqa %xmm2, %xmm1 |
285 |
movdqa %xmm4, %xmm3 |
286 |
pand %xmm1, %xmm5 |
287 |
pand %xmm2, %xmm6 |
288 |
pand %xmm3, %xmm5 |
289 |
pand %xmm4, %xmm6 |
290 |
psllw %xmm1, 8 |
291 |
psrlw %xmm2, 8 |
292 |
psllw %xmm3, 8 |
293 |
psrlw %xmm4, 8 |
294 |
por %xmm1, %xmm2 |
295 |
por %xmm3, %xmm4 |
296 |
|
297 |
movdqa [%ecx], %xmm1 |
298 |
movdqa [%ecx+16], %xmm3 |
299 |
add %edx, 32 |
300 |
add %ecx, 32 |
301 |
dec %eax |
302 |
jnz 1b |
303 |
|
304 |
2: |
305 |
ret |
306 |
|
307 |
.balign 16 |
308 |
################################################################################# |
309 |
## |
310 |
## IN: %eax -- number of pixels to convert |
311 |
## %edx -- input |
312 |
## %ecx -- output |
313 |
|
314 |
EXPORT(x86_sse2_convert_2be555_to_2le565): |
315 |
add %eax, 15 |
316 |
shr %eax, 4 # we can convert 16 pixels at a time |
317 |
movdqa %xmm0, [d1] |
318 |
movdqa %xmm7, [d2] |
319 |
jz 2f |
320 |
1: |
321 |
movdqa %xmm1, [%edx] |
322 |
movdqa %xmm3, [%edx+16] |
323 |
## convert big to little endian |
324 |
movdqa %xmm2, %xmm1 |
325 |
movdqa %xmm4, %xmm3 |
326 |
pand %xmm1, %xmm0 |
327 |
pand %xmm2, %xmm7 |
328 |
pand %xmm3, %xmm0 |
329 |
pand %xmm4, %xmm7 |
330 |
psllw %xmm1, 8 |
331 |
psrlw %xmm2, 8 |
332 |
psllw %xmm3, 8 |
333 |
psrlw %xmm4, 8 |
334 |
por %xmm1, %xmm2 |
335 |
por %xmm4, %xmm3 |
336 |
|
337 |
movdqa %xmm2, %xmm1 |
338 |
movdqa %xmm3, %xmm1 |
339 |
movdqa %xmm5, %xmm4 |
340 |
movdqa %xmm6, %xmm4 |
341 |
pand %xmm1, [_2be555_mask_r] |
342 |
pand %xmm2, [_2be555_mask_g] |
343 |
pand %xmm3, [_2be555_mask_b] |
344 |
pand %xmm4, [_2be555_mask_r] |
345 |
pand %xmm5, [_2be555_mask_g] |
346 |
pand %xmm6, [_2be555_mask_b] |
347 |
psllw %xmm1, 1 # red |
348 |
psllw %xmm2, 1 # green |
349 |
# psllw %xmm3, 0 # blue |
350 |
psllw %xmm4, 1 # red |
351 |
psllw %xmm5, 1 # green |
352 |
# psllw %xmm6, 0 # blue |
353 |
por %xmm1, %xmm2 |
354 |
por %xmm4, %xmm5 |
355 |
por %xmm1, %xmm3 |
356 |
por %xmm4, %xmm6 |
357 |
movdqa [%ecx], %xmm1 |
358 |
movdqa [%ecx+16], %xmm4 |
359 |
add %edx, 32 |
360 |
add %ecx, 32 |
361 |
dec %eax |
362 |
jnz 1b |
363 |
2: |
364 |
ret |
365 |
|
366 |
.balign 16 |
367 |
################################################################################# |
368 |
## |
369 |
## IN: %eax -- number of pixels to convert |
370 |
## %edx -- input |
371 |
## %ecx -- output |
372 |
|
373 |
EXPORT(x86_sse2_convert_2be555_to_4le888): |
374 |
add %eax, 3 |
375 |
shr %eax, 3 # we can convert 8 pixels at a time |
376 |
movdqa %xmm7, [d1] |
377 |
jz 2f |
378 |
|
379 |
pxor %xmm0, %xmm0 |
380 |
1: |
381 |
movdqa %xmm1, [%edx] |
382 |
|
383 |
## convert big to little endian |
384 |
movdqa %xmm3, %xmm1 |
385 |
pand %xmm1, %xmm7 |
386 |
pand %xmm3, [d2] |
387 |
psllw %xmm1, 8 |
388 |
psrlw %xmm3, 8 |
389 |
por %xmm1, %xmm3 |
390 |
|
391 |
movdqa %xmm2, %xmm1 |
392 |
movdqa %xmm3, %xmm1 |
393 |
pand %xmm1, [_2be555_mask_r] |
394 |
pand %xmm2, [_2be555_mask_g] |
395 |
pand %xmm3, [_2be555_mask_b] |
396 |
movdqa %xmm4, %xmm1 |
397 |
movdqa %xmm5, %xmm2 |
398 |
movdqa %xmm6, %xmm3 |
399 |
punpcklwd %xmm1, %xmm0 |
400 |
punpcklwd %xmm2, %xmm0 |
401 |
punpcklwd %xmm3, %xmm0 |
402 |
punpckhwd %xmm4, %xmm0 |
403 |
punpckhwd %xmm5, %xmm0 |
404 |
punpckhwd %xmm6, %xmm0 |
405 |
pslld %xmm1, 16-10+3 # red |
406 |
pslld %xmm2, 8-5+3 # green |
407 |
pslld %xmm3, 0+3 # blue |
408 |
pslld %xmm4, 16-10+3 # red |
409 |
pslld %xmm5, 8-5+3 # green |
410 |
pslld %xmm6, 0+3 # blue |
411 |
por %xmm1, %xmm2 |
412 |
por %xmm1, %xmm3 |
413 |
por %xmm4, %xmm5 |
414 |
por %xmm4, %xmm6 |
415 |
movdqa [%ecx], %xmm1 |
416 |
movdqa [%ecx+16], %xmm4 |
417 |
add %edx, 16 |
418 |
add %ecx, 32 |
419 |
dec %eax |
420 |
jnz 1b |
421 |
2: |
422 |
ret |
423 |
|