Parent Directory | Revision Log
Revision 1002 - (view) (download)
1 : | edgomez | 851 | ;/************************************************************************** |
2 : | Isibaar | 3 | ; * |
3 : | edgomez | 851 | ; * XVID MPEG-4 VIDEO CODEC |
4 : | ; * mmx yuyv/uyvy to yuv planar | ||
5 : | Isibaar | 3 | ; * |
6 : | edgomez | 851 | ; * This program is free software; you can redistribute it and/or modify |
7 : | ; * it under the terms of the GNU General Public License as published by | ||
8 : | ; * the Free Software Foundation; either version 2 of the License, or | ||
9 : | ; * (at your option) any later version. | ||
10 : | Isibaar | 3 | ; * |
11 : | edgomez | 851 | ; * 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 : | Isibaar | 3 | ; * |
16 : | edgomez | 851 | ; * 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 : | Isibaar | 3 | ; * |
20 : | edgomez | 851 | ; *************************************************************************/ |
21 : | |||
22 : | ;/************************************************************************** | ||
23 : | Isibaar | 3 | ; * |
24 : | edgomez | 851 | ; * History: |
25 : | Isibaar | 3 | ; * |
26 : | edgomez | 851 | ; * 30.11.2001 initial version; (c)2001 peter ross <pross@cs.rmit.edu.au> |
27 : | edgomez | 651 | ; * |
28 : | edgomez | 851 | ; *************************************************************************/ |
29 : | Isibaar | 3 | |
30 : | bits 32 | ||
31 : | |||
32 : | |||
33 : | section .data | ||
34 : | |||
35 : | %macro cglobal 1 | ||
36 : | %ifdef PREFIX | ||
37 : | global _%1 | ||
38 : | %define %1 _%1 | ||
39 : | %else | ||
40 : | global %1 | ||
41 : | %endif | ||
42 : | %endmacro | ||
43 : | |||
44 : | align 16 | ||
45 : | |||
46 : | |||
47 : | section .data | ||
48 : | |||
49 : | |||
50 : | ;=========================================================================== | ||
51 : | ; masks for extracting yuv components | ||
52 : | ;=========================================================================== | ||
53 : | ; y u y v y u y v | ||
54 : | |||
55 : | mask1 db 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0 | ||
56 : | mask2 db 0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff | ||
57 : | |||
58 : | |||
59 : | section .text | ||
60 : | |||
61 : | ;=========================================================================== | ||
62 : | ; | ||
63 : | ; void yuyv_to_yv12_mmx(uint8_t * const y_out, | ||
64 : | ; uint8_t * const u_out, | ||
65 : | ; uint8_t * const v_out, | ||
66 : | ; const uint8_t * const src, | ||
67 : | ; const uint32_t width, | ||
68 : | ; const uint32_t height, | ||
69 : | ; const uint32_t stride); | ||
70 : | ; | ||
71 : | ; width must be multiple of 8 | ||
72 : | ; does not flip | ||
73 : | ; ~30% faster than plain c | ||
74 : | ; | ||
75 : | ;=========================================================================== | ||
76 : | |||
77 : | align 16 | ||
78 : | cglobal yuyv_to_yv12_mmx | ||
79 : | yuyv_to_yv12_mmx | ||
80 : | |||
81 : | push ebx | ||
82 : | push ecx | ||
83 : | push esi | ||
84 : | push edi | ||
85 : | push ebp ; STACK BASE = 20 | ||
86 : | |||
87 : | ; some global consants | ||
88 : | |||
89 : | mov ecx, [esp + 20 + 20] ; width | ||
90 : | mov eax, [esp + 20 + 28] ; stride | ||
91 : | sub eax, ecx ; eax = stride - width | ||
92 : | mov edx, eax | ||
93 : | add edx, [esp + 20 + 28] ; edx = y_dif + stride | ||
94 : | push edx ; [esp + 12] = y_dif | ||
95 : | |||
96 : | shr eax, 1 | ||
97 : | push eax ; [esp + 8] = uv_dif | ||
98 : | |||
99 : | shr ecx, 3 | ||
100 : | push ecx ; [esp + 4] = width/8 | ||
101 : | |||
102 : | sub esp, 4 ; [esp + 0] = tmp_height_counter | ||
103 : | ; STACK_BASE = 36 | ||
104 : | |||
105 : | movq mm6, [mask1] | ||
106 : | movq mm7, [mask2] | ||
107 : | |||
108 : | mov edi, [esp + 36 + 4] ; y_out | ||
109 : | mov ebx, [esp + 36 + 8] ; u_out | ||
110 : | mov edx, [esp + 36 + 12] ; v_out | ||
111 : | mov esi, [esp + 36 + 16] ; src | ||
112 : | |||
113 : | mov eax, [esp + 36 + 20] | ||
114 : | mov ebp, [esp + 36 + 24] | ||
115 : | mov ecx, [esp + 36 + 28] ; ecx = stride | ||
116 : | shr ebp, 1 ; ebp = height /= 2 | ||
117 : | add eax, eax ; eax = 2 * width | ||
118 : | |||
119 : | .yloop | ||
120 : | mov [esp], ebp | ||
121 : | mov ebp, [esp + 4] ; width/8 | ||
122 : | .xloop | ||
123 : | movq mm2, [esi] ; y 1st row | ||
124 : | movq mm3, [esi + 8] | ||
125 : | movq mm0, mm2 | ||
126 : | movq mm1, mm3 | ||
127 : | pand mm2, mm6 ; mask1 | ||
128 : | pand mm3, mm6 ; mask1 | ||
129 : | pand mm0, mm7 ; mask2 | ||
130 : | pand mm1, mm7 ; mask2 | ||
131 : | packuswb mm2, mm3 | ||
132 : | psrlq mm0, 8 | ||
133 : | psrlq mm1, 8 | ||
134 : | movq [edi], mm2 | ||
135 : | |||
136 : | movq mm4, [esi + eax] ; y 2nd row | ||
137 : | movq mm5, [esi + eax + 8] | ||
138 : | movq mm2, mm4 | ||
139 : | movq mm3, mm5 | ||
140 : | pand mm4, mm6 ; mask1 | ||
141 : | pand mm5, mm6 ; mask1 | ||
142 : | pand mm2, mm7 ; mask2 | ||
143 : | pand mm3, mm7 ; mask2 | ||
144 : | packuswb mm4, mm5 | ||
145 : | psrlq mm2, 8 | ||
146 : | psrlq mm3, 8 | ||
147 : | movq [edi + ecx], mm4 | ||
148 : | |||
149 : | paddw mm0, mm2 ; uv avg 1st & 2nd | ||
150 : | paddw mm1, mm3 | ||
151 : | psrlw mm0, 1 | ||
152 : | psrlw mm1, 1 | ||
153 : | packuswb mm0, mm1 | ||
154 : | movq mm2, mm0 | ||
155 : | pand mm0, mm6 ; mask1 | ||
156 : | pand mm2, mm7 ; mask2 | ||
157 : | packuswb mm0, mm0 | ||
158 : | psrlq mm2, 8 | ||
159 : | movd [ebx], mm0 | ||
160 : | packuswb mm2, mm2 | ||
161 : | movd [edx], mm2 | ||
162 : | |||
163 : | add esi, 16 | ||
164 : | add edi, 8 | ||
165 : | add ebx, 4 | ||
166 : | add edx, 4 | ||
167 : | dec ebp | ||
168 : | jnz near .xloop | ||
169 : | |||
170 : | mov ebp, [esp] | ||
171 : | |||
172 : | add esi, eax ; += width2 | ||
173 : | add edi, [esp + 12] ; += y_dif + stride | ||
174 : | add ebx, [esp + 8] ; += uv_dif | ||
175 : | add edx, [esp + 8] ; += uv_dif | ||
176 : | |||
177 : | dec ebp | ||
178 : | jnz near .yloop | ||
179 : | |||
180 : | emms | ||
181 : | |||
182 : | add esp, 16 | ||
183 : | pop ebp | ||
184 : | pop edi | ||
185 : | pop esi | ||
186 : | pop ecx | ||
187 : | pop ebx | ||
188 : | |||
189 : | ret | ||
190 : | |||
191 : | |||
192 : | ;=========================================================================== | ||
193 : | ; | ||
194 : | ; void uyvy_to_yv12_mmx(uint8_t * const y_out, | ||
195 : | ; uint8_t * const u_out, | ||
196 : | ; uint8_t * const v_out, | ||
197 : | ; const uint8_t * const src, | ||
198 : | ; const uint32_t width, | ||
199 : | ; const uint32_t height, | ||
200 : | ; const uint32_t stride); | ||
201 : | ; | ||
202 : | ; width must be multiple of 8 | ||
203 : | ; does not flip | ||
204 : | ; ~30% faster than plain c | ||
205 : | ; | ||
206 : | ;=========================================================================== | ||
207 : | |||
208 : | align 16 | ||
209 : | cglobal uyvy_to_yv12_mmx | ||
210 : | uyvy_to_yv12_mmx | ||
211 : | |||
212 : | push ebx | ||
213 : | push ecx | ||
214 : | push esi | ||
215 : | push edi | ||
216 : | push ebp ; STACK BASE = 20 | ||
217 : | |||
218 : | ; some global consants | ||
219 : | |||
220 : | mov ecx, [esp + 20 + 20] ; width | ||
221 : | mov eax, [esp + 20 + 28] ; stride | ||
222 : | sub eax, ecx ; eax = stride - width | ||
223 : | mov edx, eax | ||
224 : | add edx, [esp + 20 + 28] ; edx = y_dif + stride | ||
225 : | push edx ; [esp + 12] = y_dif | ||
226 : | |||
227 : | shr eax, 1 | ||
228 : | push eax ; [esp + 8] = uv_dif | ||
229 : | |||
230 : | shr ecx, 3 | ||
231 : | push ecx ; [esp + 4] = width/8 | ||
232 : | |||
233 : | sub esp, 4 ; [esp + 0] = tmp_height_counter | ||
234 : | ; STACK_BASE = 36 | ||
235 : | |||
236 : | movq mm6, [mask1] | ||
237 : | movq mm7, [mask2] | ||
238 : | |||
239 : | mov edi, [esp + 36 + 4] ; y_out | ||
240 : | mov ebx, [esp + 36 + 8] ; u_out | ||
241 : | mov edx, [esp + 36 + 12] ; v_out | ||
242 : | mov esi, [esp + 36 + 16] ; src | ||
243 : | |||
244 : | mov eax, [esp + 36 + 20] | ||
245 : | mov ebp, [esp + 36 + 24] | ||
246 : | mov ecx, [esp + 36 + 28] ; ecx = stride | ||
247 : | shr ebp, 1 ; ebp = height /= 2 | ||
248 : | add eax, eax ; eax = 2 * width | ||
249 : | |||
250 : | .yloop | ||
251 : | mov [esp], ebp | ||
252 : | mov ebp, [esp + 4] ; width/8 | ||
253 : | .xloop | ||
254 : | movq mm2, [esi] ; y 1st row | ||
255 : | movq mm3, [esi + 8] | ||
256 : | movq mm0, mm2 | ||
257 : | movq mm1, mm3 | ||
258 : | pand mm2, mm7 ; mask2 | ||
259 : | pand mm3, mm7 ; mask2 | ||
260 : | psrlq mm2, 8 | ||
261 : | psrlq mm3, 8 | ||
262 : | pand mm0, mm6 ; mask1 | ||
263 : | pand mm1, mm6 ; mask1 | ||
264 : | packuswb mm2, mm3 | ||
265 : | movq [edi], mm2 | ||
266 : | |||
267 : | movq mm4, [esi + eax] ; y 2nd row | ||
268 : | movq mm5, [esi + eax + 8] | ||
269 : | movq mm2, mm4 | ||
270 : | movq mm3, mm5 | ||
271 : | pand mm4, mm7 ; mask2 | ||
272 : | pand mm5, mm7 ; mask2 | ||
273 : | psrlq mm4, 8 | ||
274 : | psrlq mm5, 8 | ||
275 : | pand mm2, mm6 ; mask1 | ||
276 : | pand mm3, mm6 ; mask1 | ||
277 : | packuswb mm4, mm5 | ||
278 : | movq [edi + ecx], mm4 | ||
279 : | |||
280 : | paddw mm0, mm2 ; uv avg 1st & 2nd | ||
281 : | paddw mm1, mm3 | ||
282 : | psrlw mm0, 1 | ||
283 : | psrlw mm1, 1 | ||
284 : | packuswb mm0, mm1 | ||
285 : | movq mm2, mm0 | ||
286 : | pand mm0, mm6 ; mask1 | ||
287 : | pand mm2, mm7 ; mask2 | ||
288 : | packuswb mm0, mm0 | ||
289 : | psrlq mm2, 8 | ||
290 : | movd [ebx], mm0 | ||
291 : | packuswb mm2, mm2 | ||
292 : | movd [edx], mm2 | ||
293 : | |||
294 : | add esi, 16 | ||
295 : | add edi, 8 | ||
296 : | add ebx, 4 | ||
297 : | add edx, 4 | ||
298 : | dec ebp | ||
299 : | jnz near .xloop | ||
300 : | |||
301 : | mov ebp, [esp] | ||
302 : | |||
303 : | add esi, eax ; += width2 | ||
304 : | add edi, [esp + 12] ; += y_dif + stride | ||
305 : | add ebx, [esp + 8] ; += uv_dif | ||
306 : | add edx, [esp + 8] ; += uv_dif | ||
307 : | |||
308 : | dec ebp | ||
309 : | jnz near .yloop | ||
310 : | |||
311 : | emms | ||
312 : | |||
313 : | add esp, 16 | ||
314 : | pop ebp | ||
315 : | pop edi | ||
316 : | pop esi | ||
317 : | pop ecx | ||
318 : | pop ebx | ||
319 : | |||
320 : | ret |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |