Parent Directory | Revision Log
Revision 1192 - (view) (download)
1 : | edgomez | 1192 | ;/**************************************************************************** |
2 : | edgomez | 851 | ; * |
3 : | edgomez | 1192 | ; * XVID MPEG-4 VIDEO CODEC |
4 : | ; * - MMX and XMM YUYV<->YV12 conversion - | ||
5 : | edgomez | 851 | ; * |
6 : | edgomez | 1192 | ; * Copyright(C) 2002 Peter Ross <pross@xvid.org> |
7 : | edgomez | 851 | ; * |
8 : | edgomez | 1192 | ; * 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 | 1192 | ; * 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 | 1192 | ; * 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 : | edgomez | 1192 | ; * $Id: colorspace_yuyv_mmx.asm,v 1.2.2.1 2003-10-28 22:23:03 edgomez Exp $ |
23 : | edgomez | 851 | ; * |
24 : | edgomez | 1192 | ; ***************************************************************************/ |
25 : | edgomez | 851 | |
26 : | edgomez | 1192 | BITS 32 |
27 : | edgomez | 851 | |
28 : | edgomez | 1192 | %macro cglobal 1 |
29 : | edgomez | 851 | %ifdef PREFIX |
30 : | edgomez | 1192 | global _%1 |
31 : | edgomez | 851 | %define %1 _%1 |
32 : | %else | ||
33 : | global %1 | ||
34 : | %endif | ||
35 : | %endmacro | ||
36 : | |||
37 : | edgomez | 1192 | ;============================================================================= |
38 : | ; Read only data | ||
39 : | ;============================================================================= | ||
40 : | edgomez | 851 | |
41 : | edgomez | 1192 | SECTION .rodata |
42 : | edgomez | 851 | |
43 : | edgomez | 1192 | ;----------------------------------------------------------------------------- |
44 : | edgomez | 851 | ; yuyv/uyvy mask for extracting yuv components |
45 : | edgomez | 1192 | ;----------------------------------------------------------------------------- |
46 : | edgomez | 851 | ; y u y v y u y v |
47 : | |||
48 : | edgomez | 1192 | ALIGN 16 |
49 : | yuyv_mask: db 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0 | ||
50 : | mmx_one: dw 1, 1, 1, 1 | ||
51 : | edgomez | 851 | |
52 : | edgomez | 1192 | ;============================================================================= |
53 : | ; helper macros used with colorspace_mmx.inc | ||
54 : | ;============================================================================= | ||
55 : | edgomez | 851 | |
56 : | edgomez | 1192 | ;----------------------------------------------------------------------------- |
57 : | edgomez | 851 | ; YUYV_TO_YV12( TYPE, PAVG ) |
58 : | ; | ||
59 : | ; TYPE 0=yuyv, 1=uyvy | ||
60 : | ; PAVG 0=mmx, pavgusb=3dnow, pavgb=xmm | ||
61 : | ; | ||
62 : | ; bytes=2, pixels = 8, vpixels=2 | ||
63 : | edgomez | 1192 | ;----------------------------------------------------------------------------- |
64 : | |||
65 : | edgomez | 851 | %macro YUYV_TO_YV12_INIT 2 |
66 : | edgomez | 1192 | movq mm7, [yuyv_mask] |
67 : | edgomez | 851 | %endmacro |
68 : | |||
69 : | |||
70 : | edgomez | 1192 | %macro YUYV_TO_YV12 2 |
71 : | movq mm0, [edi] ; x_ptr[0] | ||
72 : | movq mm1, [edi + 8] ; x_ptr[8] | ||
73 : | movq mm2, [edi + edx] ; x_ptr[x_stride + 0] | ||
74 : | movq mm3, [edi + edx + 8] ; x_ptr[x_stride + 8] | ||
75 : | edgomez | 851 | |
76 : | edgomez | 1192 | ; average uv-components |
77 : | edgomez | 851 | ;---[ plain mmx ]---------------------------------------------------- |
78 : | edgomez | 1192 | %ifidn %2,0 ; if (%2 eq "0") |
79 : | movq mm4, mm0 | ||
80 : | movq mm5, mm2 | ||
81 : | %if %1 == 0 ; yuyv | ||
82 : | psrlw mm4, 8 | ||
83 : | psrlw mm5, 8 | ||
84 : | edgomez | 851 | %endif |
85 : | edgomez | 1192 | pand mm4, mm7 |
86 : | pand mm5, mm7 | ||
87 : | paddw mm4, mm5 | ||
88 : | edgomez | 851 | |
89 : | edgomez | 1192 | movq mm5, mm1 |
90 : | movq mm6, mm3 | ||
91 : | %if %1 == 0 ; yuyv | ||
92 : | psrlw mm5, 8 | ||
93 : | psrlw mm6, 8 | ||
94 : | edgomez | 851 | %endif |
95 : | edgomez | 1192 | pand mm5, mm7 |
96 : | pand mm6, mm7 | ||
97 : | paddw mm5, mm6 | ||
98 : | paddw mm4, [mmx_one] ; +1 rounding | ||
99 : | paddw mm5, [mmx_one] ; | ||
100 : | psrlw mm4, 1 | ||
101 : | psrlw mm5, 1 | ||
102 : | edgomez | 851 | ;---[ 3dnow/xmm ]---------------------------------------------------- |
103 : | %else | ||
104 : | edgomez | 1192 | movq mm4, mm0 |
105 : | movq mm5, mm1 | ||
106 : | %2 mm4, mm2 ;pavgb/pavgusb mm4, mm2 | ||
107 : | %2 mm5, mm3 ;pavgb/pavgusb mm5, mm3 | ||
108 : | edgomez | 851 | |
109 : | edgomez | 1192 | ;;movq mm6, mm0 ; 0 rounding |
110 : | ;;pxor mm6, mm2 ; | ||
111 : | ;;psubb mm4, mm6 ; | ||
112 : | ;;movq mm6, mm1 ; | ||
113 : | ;;pxor mm6, mm3 ; | ||
114 : | ;;psubb mm5, mm5 ; | ||
115 : | edgomez | 851 | |
116 : | edgomez | 1192 | %if %1 == 0 ; yuyv |
117 : | psrlw mm4, 8 | ||
118 : | psrlw mm5, 8 | ||
119 : | edgomez | 851 | %endif |
120 : | edgomez | 1192 | pand mm4, mm7 |
121 : | pand mm5, mm7 | ||
122 : | edgomez | 851 | %endif |
123 : | ;-------------------------------------------------------------------- | ||
124 : | |||
125 : | edgomez | 1192 | ; write y-component |
126 : | %if %1 == 1 ; uyvy | ||
127 : | psrlw mm0, 8 | ||
128 : | psrlw mm1, 8 | ||
129 : | psrlw mm2, 8 | ||
130 : | psrlw mm3, 8 | ||
131 : | edgomez | 851 | %endif |
132 : | edgomez | 1192 | pand mm0, mm7 |
133 : | pand mm1, mm7 | ||
134 : | pand mm2, mm7 | ||
135 : | pand mm3, mm7 | ||
136 : | packuswb mm0, mm1 | ||
137 : | packuswb mm2, mm3 | ||
138 : | edgomez | 851 | |
139 : | edgomez | 1192 | %ifidn %2,pavgb ; xmm |
140 : | movntq [esi], mm0 | ||
141 : | movntq [esi+eax], mm2 | ||
142 : | %else ; plain mmx,3dnow | ||
143 : | movq [esi], mm0 | ||
144 : | movq [esi+eax], mm2 | ||
145 : | edgomez | 851 | %endif |
146 : | |||
147 : | edgomez | 1192 | ; write uv-components |
148 : | packuswb mm4, mm5 | ||
149 : | movq mm5, mm4 | ||
150 : | psrlq mm4, 8 | ||
151 : | pand mm5, mm7 | ||
152 : | pand mm4, mm7 | ||
153 : | packuswb mm5,mm5 | ||
154 : | packuswb mm4,mm4 | ||
155 : | movd [ebx],mm5 | ||
156 : | movd [ecx],mm4 | ||
157 : | edgomez | 851 | %endmacro |
158 : | |||
159 : | edgomez | 1192 | ;----------------------------------------------------------------------------- |
160 : | edgomez | 851 | ; YV12_TO_YUYV( TYPE ) |
161 : | ; | ||
162 : | edgomez | 1192 | ; TYPE 0=yuyv, 1=uyvy |
163 : | edgomez | 851 | ; |
164 : | ; bytes=2, pixels = 8, vpixels=2 | ||
165 : | edgomez | 1192 | ;----------------------------------------------------------------------------- |
166 : | |||
167 : | %macro YV12_TO_YUYV_INIT 2 | ||
168 : | edgomez | 851 | %endmacro |
169 : | |||
170 : | |||
171 : | edgomez | 1192 | %macro YV12_TO_YUYV 2 |
172 : | movd mm4, [ebx] ; [ |uuuu] | ||
173 : | movd mm5, [ecx] ; [ |vvvv] | ||
174 : | movq mm0, [esi] ; [yyyy|yyyy] ; y row 0 | ||
175 : | movq mm1, [esi+eax] ; [yyyy|yyyy] ; y row 1 | ||
176 : | punpcklbw mm4, mm5 ; [vuvu|vuvu] ; uv row 0 | ||
177 : | edgomez | 851 | |
178 : | edgomez | 1192 | %if %1 == 0 ; YUYV |
179 : | movq mm2, mm0 | ||
180 : | movq mm3, mm1 | ||
181 : | punpcklbw mm0, mm4 ; [vyuy|vyuy] ; y row 0 + 0 | ||
182 : | punpckhbw mm2, mm4 ; [vyuy|vyuy] ; y row 0 + 8 | ||
183 : | punpcklbw mm1, mm4 ; [vyuy|vyuy] ; y row 1 + 0 | ||
184 : | punpckhbw mm3, mm4 ; [vyuy|vyuy] ; y row 1 + 8 | ||
185 : | movq [edi], mm0 | ||
186 : | movq [edi+8], mm2 | ||
187 : | movq [edi+edx], mm1 | ||
188 : | movq [edi+edx+8], mm3 | ||
189 : | %else ; UYVY | ||
190 : | movq mm5, mm4 | ||
191 : | movq mm6, mm4 | ||
192 : | movq mm7, mm4 | ||
193 : | punpcklbw mm4, mm0 ; [yvyu|yvyu] ; y row 0 + 0 | ||
194 : | punpckhbw mm5, mm0 ; [yvyu|yvyu] ; y row 0 + 8 | ||
195 : | punpcklbw mm6, mm1 ; [yvyu|yvyu] ; y row 1 + 0 | ||
196 : | punpckhbw mm7, mm1 ; [yvyu|yvyu] ; y row 1 + 8 | ||
197 : | movq [edi], mm4 | ||
198 : | movq [edi+8], mm5 | ||
199 : | movq [edi+edx], mm6 | ||
200 : | movq [edi+edx+8], mm7 | ||
201 : | edgomez | 851 | %endif |
202 : | %endmacro | ||
203 : | |||
204 : | ;------------------------------------------------------------------------------ | ||
205 : | ; YV12_TO_YUYVI( TYPE ) | ||
206 : | ; | ||
207 : | edgomez | 1192 | ; TYPE 0=yuyv, 1=uyvy |
208 : | edgomez | 851 | ; |
209 : | ; bytes=2, pixels = 8, vpixels=4 | ||
210 : | ;------------------------------------------------------------------------------ | ||
211 : | edgomez | 1192 | |
212 : | %macro YV12_TO_YUYVI_INIT 2 | ||
213 : | edgomez | 851 | %endmacro |
214 : | |||
215 : | edgomez | 1192 | %macro YV12_TO_YUYVI 2 |
216 : | xchg ebp, [uv_stride] | ||
217 : | movd mm0, [ebx] ; [ |uuuu] | ||
218 : | movd mm1, [ebx+ebp] ; [ |uuuu] | ||
219 : | punpcklbw mm0, [ecx] ; [vuvu|vuvu] ; uv row 0 | ||
220 : | punpcklbw mm1, [ecx+ebp] ; [vuvu|vuvu] ; uv row 1 | ||
221 : | xchg ebp, [uv_stride] | ||
222 : | edgomez | 851 | |
223 : | edgomez | 1192 | %if %1 == 0 ; YUYV |
224 : | movq mm4, [esi] ; [yyyy|yyyy] ; y row 0 | ||
225 : | movq mm6, [esi+eax] ; [yyyy|yyyy] ; y row 1 | ||
226 : | movq mm5, mm4 | ||
227 : | movq mm7, mm6 | ||
228 : | punpcklbw mm4, mm0 ; [yuyv|yuyv] ; y row 0 + 0 | ||
229 : | punpckhbw mm5, mm0 ; [yuyv|yuyv] ; y row 0 + 8 | ||
230 : | punpcklbw mm6, mm1 ; [yuyv|yuyv] ; y row 1 + 0 | ||
231 : | punpckhbw mm7, mm1 ; [yuyv|yuyv] ; y row 1 + 8 | ||
232 : | movq [edi], mm4 | ||
233 : | movq [edi+8], mm5 | ||
234 : | movq [edi+edx], mm6 | ||
235 : | movq [edi+edx+8], mm7 | ||
236 : | edgomez | 851 | |
237 : | edgomez | 1192 | push esi |
238 : | push edi | ||
239 : | add esi, eax | ||
240 : | add edi, edx | ||
241 : | movq mm4, [esi+eax] ; [yyyy|yyyy] ; y row 2 | ||
242 : | movq mm6, [esi+2*eax] ; [yyyy|yyyy] ; y row 3 | ||
243 : | movq mm5, mm4 | ||
244 : | movq mm7, mm6 | ||
245 : | punpcklbw mm4, mm0 ; [yuyv|yuyv] ; y row 2 + 0 | ||
246 : | punpckhbw mm5, mm0 ; [yuyv|yuyv] ; y row 2 + 8 | ||
247 : | punpcklbw mm6, mm1 ; [yuyv|yuyv] ; y row 3 + 0 | ||
248 : | punpckhbw mm7, mm1 ; [yuyv|yuyv] ; y row 3 + 8 | ||
249 : | movq [edi+edx], mm4 | ||
250 : | movq [edi+edx+8], mm5 | ||
251 : | movq [edi+2*edx], mm6 | ||
252 : | movq [edi+2*edx+8], mm7 | ||
253 : | pop edi | ||
254 : | pop esi | ||
255 : | %else ; UYVY | ||
256 : | movq mm2, [esi] ; [yyyy|yyyy] ; y row 0 | ||
257 : | movq mm3, [esi+eax] ; [yyyy|yyyy] ; y row 1 | ||
258 : | movq mm4, mm0 | ||
259 : | movq mm5, mm0 | ||
260 : | movq mm6, mm1 | ||
261 : | movq mm7, mm1 | ||
262 : | punpcklbw mm4, mm2 ; [uyvy|uyvy] ; y row 0 + 0 | ||
263 : | punpckhbw mm5, mm2 ; [uyvy|uyvy] ; y row 0 + 8 | ||
264 : | punpcklbw mm6, mm3 ; [uyvy|uyvy] ; y row 1 + 0 | ||
265 : | punpckhbw mm7, mm3 ; [uyvy|uyvy] ; y row 1 + 8 | ||
266 : | movq [edi], mm4 | ||
267 : | movq [edi+8], mm5 | ||
268 : | movq [edi+edx], mm6 | ||
269 : | movq [edi+edx+8], mm7 | ||
270 : | edgomez | 851 | |
271 : | edgomez | 1192 | push esi |
272 : | push edi | ||
273 : | add esi, eax | ||
274 : | add edi, edx | ||
275 : | movq mm2, [esi+eax] ; [yyyy|yyyy] ; y row 2 | ||
276 : | movq mm3, [esi+2*eax] ; [yyyy|yyyy] ; y row 3 | ||
277 : | movq mm4, mm0 | ||
278 : | movq mm5, mm0 | ||
279 : | movq mm6, mm1 | ||
280 : | movq mm7, mm1 | ||
281 : | punpcklbw mm4, mm2 ; [uyvy|uyvy] ; y row 2 + 0 | ||
282 : | punpckhbw mm5, mm2 ; [uyvy|uyvy] ; y row 2 + 8 | ||
283 : | punpcklbw mm6, mm3 ; [uyvy|uyvy] ; y row 3 + 0 | ||
284 : | punpckhbw mm7, mm3 ; [uyvy|uyvy] ; y row 3 + 8 | ||
285 : | movq [edi+edx], mm4 | ||
286 : | movq [edi+edx+8], mm5 | ||
287 : | movq [edi+2*edx], mm6 | ||
288 : | movq [edi+2*edx+8], mm7 | ||
289 : | pop edi | ||
290 : | pop esi | ||
291 : | edgomez | 851 | %endif |
292 : | %endmacro | ||
293 : | |||
294 : | edgomez | 1192 | ;============================================================================= |
295 : | ; Code | ||
296 : | ;============================================================================= | ||
297 : | edgomez | 851 | |
298 : | edgomez | 1192 | SECTION .text |
299 : | |||
300 : | %include "colorspace_mmx.inc" | ||
301 : | |||
302 : | edgomez | 851 | ; input |
303 : | |||
304 : | MAKE_COLORSPACE yuyv_to_yv12_mmx,0, 2,8,2, YUYV_TO_YV12, 0, 0 | ||
305 : | MAKE_COLORSPACE yuyv_to_yv12_3dn,0, 2,8,2, YUYV_TO_YV12, 0, pavgusb | ||
306 : | MAKE_COLORSPACE yuyv_to_yv12_xmm,0, 2,8,2, YUYV_TO_YV12, 0, pavgb | ||
307 : | |||
308 : | MAKE_COLORSPACE uyvy_to_yv12_mmx,0, 2,8,2, YUYV_TO_YV12, 1, 0 | ||
309 : | MAKE_COLORSPACE uyvy_to_yv12_3dn,0, 2,8,2, YUYV_TO_YV12, 1, pavgusb | ||
310 : | MAKE_COLORSPACE uyvy_to_yv12_xmm,0, 2,8,2, YUYV_TO_YV12, 1, pavgb | ||
311 : | |||
312 : | ; output | ||
313 : | |||
314 : | MAKE_COLORSPACE yv12_to_yuyv_mmx,0, 2,8,2, YV12_TO_YUYV, 0, -1 | ||
315 : | MAKE_COLORSPACE yv12_to_uyvy_mmx,0, 2,8,2, YV12_TO_YUYV, 1, -1 | ||
316 : | |||
317 : | MAKE_COLORSPACE yv12_to_yuyvi_mmx,0, 2,8,4, YV12_TO_YUYVI, 0, -1 | ||
318 : | MAKE_COLORSPACE yv12_to_uyvyi_mmx,0, 2,8,4, YV12_TO_YUYVI, 1, -1 |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |