Parent Directory | Revision Log
Revision 1790 - (view) (download)
1 : | edgomez | 1382 | ;/**************************************************************************** |
2 : | edgomez | 851 | ; * |
3 : | edgomez | 1382 | ; * XVID MPEG-4 VIDEO CODEC |
4 : | ; * - MMX and XMM YUYV<->YV12 conversion - | ||
5 : | edgomez | 851 | ; * |
6 : | edgomez | 1382 | ; * Copyright(C) 2002 Peter Ross <pross@xvid.org> |
7 : | edgomez | 851 | ; * |
8 : | edgomez | 1382 | ; * This program is free software; you can redistribute it and/or modify it |
9 : | ; * under the terms of the GNU General Public License as published by | ||
10 : | ; * the Free Software Foundation; either version 2 of the License, or | ||
11 : | ; * (at your option) any later version. | ||
12 : | edgomez | 851 | ; * |
13 : | edgomez | 1382 | ; * This program is distributed in the hope that it will be useful, |
14 : | ; * but WITHOUT ANY WARRANTY; without even the implied warranty of | ||
15 : | ; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
16 : | ; * GNU General Public License for more details. | ||
17 : | edgomez | 851 | ; * |
18 : | edgomez | 1382 | ; * You should have received a copy of the GNU General Public License |
19 : | ; * along with this program; if not, write to the Free Software | ||
20 : | ; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
21 : | edgomez | 851 | ; * |
22 : | Isibaar | 1790 | ; * $Id: colorspace_yuyv_mmx.asm,v 1.8 2008-08-19 09:06:48 Isibaar Exp $ |
23 : | edgomez | 851 | ; * |
24 : | edgomez | 1382 | ; ***************************************************************************/ |
25 : | edgomez | 851 | |
26 : | edgomez | 1382 | BITS 32 |
27 : | edgomez | 851 | |
28 : | edgomez | 1382 | %macro cglobal 1 |
29 : | edgomez | 851 | %ifdef PREFIX |
30 : | edgomez | 1535 | %ifdef MARK_FUNCS |
31 : | edgomez | 1540 | global _%1:function %1.endfunc-%1 |
32 : | %define %1 _%1:function %1.endfunc-%1 | ||
33 : | edgomez | 1535 | %else |
34 : | global _%1 | ||
35 : | %define %1 _%1 | ||
36 : | %endif | ||
37 : | edgomez | 851 | %else |
38 : | edgomez | 1535 | %ifdef MARK_FUNCS |
39 : | edgomez | 1540 | global %1:function %1.endfunc-%1 |
40 : | edgomez | 1535 | %else |
41 : | global %1 | ||
42 : | %endif | ||
43 : | edgomez | 851 | %endif |
44 : | %endmacro | ||
45 : | |||
46 : | edgomez | 1382 | ;============================================================================= |
47 : | ; Read only data | ||
48 : | ;============================================================================= | ||
49 : | edgomez | 851 | |
50 : | edgomez | 1382 | %ifdef FORMAT_COFF |
51 : | edgomez | 1519 | SECTION .rodata |
52 : | edgomez | 1382 | %else |
53 : | edgomez | 1519 | SECTION .rodata align=16 |
54 : | edgomez | 1382 | %endif |
55 : | edgomez | 851 | |
56 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
57 : | edgomez | 851 | ; yuyv/uyvy mask for extracting yuv components |
58 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
59 : | edgomez | 851 | ; y u y v y u y v |
60 : | |||
61 : | edgomez | 1382 | ALIGN 16 |
62 : | yuyv_mask: db 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0 | ||
63 : | mmx_one: dw 1, 1, 1, 1 | ||
64 : | edgomez | 851 | |
65 : | edgomez | 1382 | ;============================================================================= |
66 : | ; helper macros used with colorspace_mmx.inc | ||
67 : | ;============================================================================= | ||
68 : | edgomez | 851 | |
69 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
70 : | edgomez | 851 | ; YUYV_TO_YV12( TYPE, PAVG ) |
71 : | ; | ||
72 : | ; TYPE 0=yuyv, 1=uyvy | ||
73 : | ; PAVG 0=mmx, pavgusb=3dnow, pavgb=xmm | ||
74 : | ; | ||
75 : | ; bytes=2, pixels = 8, vpixels=2 | ||
76 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
77 : | |||
78 : | edgomez | 851 | %macro YUYV_TO_YV12_INIT 2 |
79 : | edgomez | 1382 | movq mm7, [yuyv_mask] |
80 : | edgomez | 851 | %endmacro |
81 : | |||
82 : | |||
83 : | edgomez | 1382 | %macro YUYV_TO_YV12 2 |
84 : | movq mm0, [edi] ; x_ptr[0] | ||
85 : | movq mm1, [edi + 8] ; x_ptr[8] | ||
86 : | movq mm2, [edi + edx] ; x_ptr[x_stride + 0] | ||
87 : | movq mm3, [edi + edx + 8] ; x_ptr[x_stride + 8] | ||
88 : | edgomez | 851 | |
89 : | edgomez | 1382 | ; average uv-components |
90 : | edgomez | 851 | ;---[ plain mmx ]---------------------------------------------------- |
91 : | edgomez | 1382 | %ifidn %2,0 ; if (%2 eq "0") |
92 : | movq mm4, mm0 | ||
93 : | movq mm5, mm2 | ||
94 : | %if %1 == 0 ; yuyv | ||
95 : | psrlw mm4, 8 | ||
96 : | psrlw mm5, 8 | ||
97 : | edgomez | 851 | %endif |
98 : | edgomez | 1382 | pand mm4, mm7 |
99 : | pand mm5, mm7 | ||
100 : | paddw mm4, mm5 | ||
101 : | edgomez | 851 | |
102 : | edgomez | 1382 | movq mm5, mm1 |
103 : | movq mm6, mm3 | ||
104 : | %if %1 == 0 ; yuyv | ||
105 : | psrlw mm5, 8 | ||
106 : | psrlw mm6, 8 | ||
107 : | edgomez | 851 | %endif |
108 : | edgomez | 1382 | pand mm5, mm7 |
109 : | pand mm6, mm7 | ||
110 : | paddw mm5, mm6 | ||
111 : | paddw mm4, [mmx_one] ; +1 rounding | ||
112 : | paddw mm5, [mmx_one] ; | ||
113 : | psrlw mm4, 1 | ||
114 : | psrlw mm5, 1 | ||
115 : | edgomez | 851 | ;---[ 3dnow/xmm ]---------------------------------------------------- |
116 : | %else | ||
117 : | edgomez | 1382 | movq mm4, mm0 |
118 : | movq mm5, mm1 | ||
119 : | %2 mm4, mm2 ;pavgb/pavgusb mm4, mm2 | ||
120 : | %2 mm5, mm3 ;pavgb/pavgusb mm5, mm3 | ||
121 : | edgomez | 851 | |
122 : | edgomez | 1382 | ;;movq mm6, mm0 ; 0 rounding |
123 : | ;;pxor mm6, mm2 ; | ||
124 : | ;;psubb mm4, mm6 ; | ||
125 : | ;;movq mm6, mm1 ; | ||
126 : | ;;pxor mm6, mm3 ; | ||
127 : | ;;psubb mm5, mm5 ; | ||
128 : | edgomez | 851 | |
129 : | edgomez | 1382 | %if %1 == 0 ; yuyv |
130 : | psrlw mm4, 8 | ||
131 : | psrlw mm5, 8 | ||
132 : | edgomez | 851 | %endif |
133 : | edgomez | 1382 | pand mm4, mm7 |
134 : | pand mm5, mm7 | ||
135 : | edgomez | 851 | %endif |
136 : | ;-------------------------------------------------------------------- | ||
137 : | |||
138 : | edgomez | 1382 | ; write y-component |
139 : | %if %1 == 1 ; uyvy | ||
140 : | psrlw mm0, 8 | ||
141 : | psrlw mm1, 8 | ||
142 : | psrlw mm2, 8 | ||
143 : | psrlw mm3, 8 | ||
144 : | edgomez | 851 | %endif |
145 : | edgomez | 1382 | pand mm0, mm7 |
146 : | pand mm1, mm7 | ||
147 : | pand mm2, mm7 | ||
148 : | pand mm3, mm7 | ||
149 : | packuswb mm0, mm1 | ||
150 : | packuswb mm2, mm3 | ||
151 : | edgomez | 851 | |
152 : | edgomez | 1382 | %ifidn %2,pavgb ; xmm |
153 : | movntq [esi], mm0 | ||
154 : | movntq [esi+eax], mm2 | ||
155 : | %else ; plain mmx,3dnow | ||
156 : | movq [esi], mm0 | ||
157 : | movq [esi+eax], mm2 | ||
158 : | edgomez | 851 | %endif |
159 : | |||
160 : | edgomez | 1382 | ; write uv-components |
161 : | packuswb mm4, mm5 | ||
162 : | movq mm5, mm4 | ||
163 : | psrlq mm4, 8 | ||
164 : | pand mm5, mm7 | ||
165 : | pand mm4, mm7 | ||
166 : | packuswb mm5,mm5 | ||
167 : | packuswb mm4,mm4 | ||
168 : | movd [ebx],mm5 | ||
169 : | movd [ecx],mm4 | ||
170 : | edgomez | 851 | %endmacro |
171 : | |||
172 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
173 : | edgomez | 851 | ; YV12_TO_YUYV( TYPE ) |
174 : | ; | ||
175 : | Skal | 1706 | ; bytes=2, pixels = 16, vpixels=2 |
176 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
177 : | |||
178 : | %macro YV12_TO_YUYV_INIT 2 | ||
179 : | edgomez | 851 | %endmacro |
180 : | |||
181 : | |||
182 : | edgomez | 1382 | %macro YV12_TO_YUYV 2 |
183 : | Skal | 1706 | movq mm6, [ebx] ; [ |uuuu] |
184 : | movq mm2, [ecx] ; [ |vvvv] | ||
185 : | movq mm0, [esi ] ; [yyyy|yyyy] ; y row 0 | ||
186 : | edgomez | 1382 | movq mm1, [esi+eax] ; [yyyy|yyyy] ; y row 1 |
187 : | Skal | 1706 | movq mm7, mm6 |
188 : | punpcklbw mm6, mm2 ; [vuvu|vuvu] ; uv[0..3] | ||
189 : | punpckhbw mm7, mm2 ; [vuvu|vuvu] ; uv[4..7] | ||
190 : | edgomez | 851 | |
191 : | edgomez | 1382 | %if %1 == 0 ; YUYV |
192 : | movq mm2, mm0 | ||
193 : | movq mm3, mm1 | ||
194 : | Skal | 1706 | movq mm4, [esi +8] ; [yyyy|yyyy] ; y[8..15] row 0 |
195 : | movq mm5, [esi+eax+8] ; [yyyy|yyyy] ; y[8..15] row 1 | ||
196 : | punpcklbw mm0, mm6 ; [vyuy|vyuy] ; y row 0 + 0 | ||
197 : | punpckhbw mm2, mm6 ; [vyuy|vyuy] ; y row 0 + 8 | ||
198 : | punpcklbw mm1, mm6 ; [vyuy|vyuy] ; y row 1 + 0 | ||
199 : | punpckhbw mm3, mm6 ; [vyuy|vyuy] ; y row 1 + 8 | ||
200 : | movq [edi ], mm0 | ||
201 : | movq [edi+8 ], mm2 | ||
202 : | movq [edi+edx ], mm1 | ||
203 : | edgomez | 1382 | movq [edi+edx+8], mm3 |
204 : | Skal | 1706 | movq mm0, mm4 |
205 : | movq mm2, mm5 | ||
206 : | punpcklbw mm0, mm7 ; [vyuy|vyuy] ; y row 0 + 16 | ||
207 : | punpckhbw mm4, mm7 ; [vyuy|vyuy] ; y row 0 + 24 | ||
208 : | punpcklbw mm2, mm7 ; [vyuy|vyuy] ; y row 1 + 16 | ||
209 : | punpckhbw mm5, mm7 ; [vyuy|vyuy] ; y row 1 + 24 | ||
210 : | movq [edi +16], mm0 | ||
211 : | movq [edi +24], mm4 | ||
212 : | movq [edi+edx+16], mm2 | ||
213 : | movq [edi+edx+24], mm5 | ||
214 : | edgomez | 1382 | %else ; UYVY |
215 : | Skal | 1706 | movq mm2, mm6 |
216 : | movq mm3, mm6 | ||
217 : | movq mm4, mm6 | ||
218 : | punpcklbw mm2, mm0 ; [yvyu|yvyu] ; y row 0 + 0 | ||
219 : | punpckhbw mm3, mm0 ; [yvyu|yvyu] ; y row 0 + 8 | ||
220 : | movq mm0, [esi +8] ; [yyyy|yyyy] ; y[8..15] row 0 | ||
221 : | movq mm5, [esi+eax+8] ; [yyyy|yyyy] ; y[8..15] row 1 | ||
222 : | punpcklbw mm4, mm1 ; [yvyu|yvyu] ; y row 1 + 0 | ||
223 : | punpckhbw mm6, mm1 ; [yvyu|yvyu] ; y row 1 + 8 | ||
224 : | movq [edi ], mm2 | ||
225 : | movq [edi +8], mm3 | ||
226 : | movq [edi+edx ], mm4 | ||
227 : | movq [edi+edx+8], mm6 | ||
228 : | movq mm2, mm7 | ||
229 : | movq mm3, mm7 | ||
230 : | movq mm6, mm7 | ||
231 : | punpcklbw mm2, mm0 ; [yvyu|yvyu] ; y row 0 + 0 | ||
232 : | punpckhbw mm3, mm0 ; [yvyu|yvyu] ; y row 0 + 8 | ||
233 : | punpcklbw mm6, mm5 ; [yvyu|yvyu] ; y row 1 + 0 | ||
234 : | punpckhbw mm7, mm5 ; [yvyu|yvyu] ; y row 1 + 8 | ||
235 : | movq [edi +16], mm2 | ||
236 : | movq [edi +24], mm3 | ||
237 : | movq [edi+edx+16], mm6 | ||
238 : | movq [edi+edx+24], mm7 | ||
239 : | edgomez | 851 | %endif |
240 : | %endmacro | ||
241 : | |||
242 : | ;------------------------------------------------------------------------------ | ||
243 : | ; YV12_TO_YUYVI( TYPE ) | ||
244 : | ; | ||
245 : | edgomez | 1382 | ; TYPE 0=yuyv, 1=uyvy |
246 : | edgomez | 851 | ; |
247 : | ; bytes=2, pixels = 8, vpixels=4 | ||
248 : | ;------------------------------------------------------------------------------ | ||
249 : | edgomez | 1382 | |
250 : | %macro YV12_TO_YUYVI_INIT 2 | ||
251 : | edgomez | 851 | %endmacro |
252 : | |||
253 : | edgomez | 1382 | %macro YV12_TO_YUYVI 2 |
254 : | xchg ebp, [uv_stride] | ||
255 : | movd mm0, [ebx] ; [ |uuuu] | ||
256 : | movd mm1, [ebx+ebp] ; [ |uuuu] | ||
257 : | punpcklbw mm0, [ecx] ; [vuvu|vuvu] ; uv row 0 | ||
258 : | punpcklbw mm1, [ecx+ebp] ; [vuvu|vuvu] ; uv row 1 | ||
259 : | xchg ebp, [uv_stride] | ||
260 : | edgomez | 851 | |
261 : | edgomez | 1382 | %if %1 == 0 ; YUYV |
262 : | movq mm4, [esi] ; [yyyy|yyyy] ; y row 0 | ||
263 : | movq mm6, [esi+eax] ; [yyyy|yyyy] ; y row 1 | ||
264 : | movq mm5, mm4 | ||
265 : | movq mm7, mm6 | ||
266 : | punpcklbw mm4, mm0 ; [yuyv|yuyv] ; y row 0 + 0 | ||
267 : | punpckhbw mm5, mm0 ; [yuyv|yuyv] ; y row 0 + 8 | ||
268 : | punpcklbw mm6, mm1 ; [yuyv|yuyv] ; y row 1 + 0 | ||
269 : | punpckhbw mm7, mm1 ; [yuyv|yuyv] ; y row 1 + 8 | ||
270 : | movq [edi], mm4 | ||
271 : | movq [edi+8], mm5 | ||
272 : | movq [edi+edx], mm6 | ||
273 : | movq [edi+edx+8], mm7 | ||
274 : | edgomez | 851 | |
275 : | edgomez | 1382 | push esi |
276 : | push edi | ||
277 : | add esi, eax | ||
278 : | add edi, edx | ||
279 : | movq mm4, [esi+eax] ; [yyyy|yyyy] ; y row 2 | ||
280 : | movq mm6, [esi+2*eax] ; [yyyy|yyyy] ; y row 3 | ||
281 : | movq mm5, mm4 | ||
282 : | movq mm7, mm6 | ||
283 : | punpcklbw mm4, mm0 ; [yuyv|yuyv] ; y row 2 + 0 | ||
284 : | punpckhbw mm5, mm0 ; [yuyv|yuyv] ; y row 2 + 8 | ||
285 : | punpcklbw mm6, mm1 ; [yuyv|yuyv] ; y row 3 + 0 | ||
286 : | punpckhbw mm7, mm1 ; [yuyv|yuyv] ; y row 3 + 8 | ||
287 : | movq [edi+edx], mm4 | ||
288 : | movq [edi+edx+8], mm5 | ||
289 : | movq [edi+2*edx], mm6 | ||
290 : | movq [edi+2*edx+8], mm7 | ||
291 : | pop edi | ||
292 : | pop esi | ||
293 : | %else ; UYVY | ||
294 : | movq mm2, [esi] ; [yyyy|yyyy] ; y row 0 | ||
295 : | movq mm3, [esi+eax] ; [yyyy|yyyy] ; y row 1 | ||
296 : | movq mm4, mm0 | ||
297 : | movq mm5, mm0 | ||
298 : | movq mm6, mm1 | ||
299 : | movq mm7, mm1 | ||
300 : | punpcklbw mm4, mm2 ; [uyvy|uyvy] ; y row 0 + 0 | ||
301 : | punpckhbw mm5, mm2 ; [uyvy|uyvy] ; y row 0 + 8 | ||
302 : | punpcklbw mm6, mm3 ; [uyvy|uyvy] ; y row 1 + 0 | ||
303 : | punpckhbw mm7, mm3 ; [uyvy|uyvy] ; y row 1 + 8 | ||
304 : | movq [edi], mm4 | ||
305 : | movq [edi+8], mm5 | ||
306 : | movq [edi+edx], mm6 | ||
307 : | movq [edi+edx+8], mm7 | ||
308 : | edgomez | 851 | |
309 : | edgomez | 1382 | push esi |
310 : | push edi | ||
311 : | add esi, eax | ||
312 : | add edi, edx | ||
313 : | movq mm2, [esi+eax] ; [yyyy|yyyy] ; y row 2 | ||
314 : | movq mm3, [esi+2*eax] ; [yyyy|yyyy] ; y row 3 | ||
315 : | movq mm4, mm0 | ||
316 : | movq mm5, mm0 | ||
317 : | movq mm6, mm1 | ||
318 : | movq mm7, mm1 | ||
319 : | punpcklbw mm4, mm2 ; [uyvy|uyvy] ; y row 2 + 0 | ||
320 : | punpckhbw mm5, mm2 ; [uyvy|uyvy] ; y row 2 + 8 | ||
321 : | punpcklbw mm6, mm3 ; [uyvy|uyvy] ; y row 3 + 0 | ||
322 : | punpckhbw mm7, mm3 ; [uyvy|uyvy] ; y row 3 + 8 | ||
323 : | movq [edi+edx], mm4 | ||
324 : | movq [edi+edx+8], mm5 | ||
325 : | movq [edi+2*edx], mm6 | ||
326 : | movq [edi+2*edx+8], mm7 | ||
327 : | pop edi | ||
328 : | pop esi | ||
329 : | edgomez | 851 | %endif |
330 : | %endmacro | ||
331 : | |||
332 : | edgomez | 1382 | ;============================================================================= |
333 : | ; Code | ||
334 : | ;============================================================================= | ||
335 : | edgomez | 851 | |
336 : | edgomez | 1382 | SECTION .text |
337 : | |||
338 : | %include "colorspace_mmx.inc" | ||
339 : | |||
340 : | edgomez | 851 | ; input |
341 : | |||
342 : | MAKE_COLORSPACE yuyv_to_yv12_mmx,0, 2,8,2, YUYV_TO_YV12, 0, 0 | ||
343 : | MAKE_COLORSPACE yuyv_to_yv12_3dn,0, 2,8,2, YUYV_TO_YV12, 0, pavgusb | ||
344 : | MAKE_COLORSPACE yuyv_to_yv12_xmm,0, 2,8,2, YUYV_TO_YV12, 0, pavgb | ||
345 : | |||
346 : | MAKE_COLORSPACE uyvy_to_yv12_mmx,0, 2,8,2, YUYV_TO_YV12, 1, 0 | ||
347 : | MAKE_COLORSPACE uyvy_to_yv12_3dn,0, 2,8,2, YUYV_TO_YV12, 1, pavgusb | ||
348 : | MAKE_COLORSPACE uyvy_to_yv12_xmm,0, 2,8,2, YUYV_TO_YV12, 1, pavgb | ||
349 : | |||
350 : | ; output | ||
351 : | |||
352 : | Skal | 1706 | MAKE_COLORSPACE yv12_to_yuyv_mmx,0, 2,16,2, YV12_TO_YUYV, 0, -1 |
353 : | MAKE_COLORSPACE yv12_to_uyvy_mmx,0, 2,16,2, YV12_TO_YUYV, 1, -1 | ||
354 : | edgomez | 851 | |
355 : | MAKE_COLORSPACE yv12_to_yuyvi_mmx,0, 2,8,4, YV12_TO_YUYVI, 0, -1 | ||
356 : | MAKE_COLORSPACE yv12_to_uyvyi_mmx,0, 2,8,4, YV12_TO_YUYVI, 1, -1 | ||
357 : | Isibaar | 1790 | |
358 : | %ifidn __OUTPUT_FORMAT__,elf | ||
359 : | section ".note.GNU-stack" noalloc noexec nowrite progbits | ||
360 : | %endif | ||
361 : |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |