Parent Directory | Revision Log
Revision 1742 - (view) (download)
1 : | edgomez | 1382 | ;/**************************************************************************** |
2 : | ; * | ||
3 : | ; * XVID MPEG-4 VIDEO CODEC | ||
4 : | ; * - MMX and XMM YV12->YV12 conversion - | ||
5 : | ; * | ||
6 : | ; * Copyright(C) 2001 Michael Militzer <isibaar@xvid.org> | ||
7 : | ; * | ||
8 : | ; * 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 : | ; * | ||
13 : | ; * 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 : | ; * | ||
18 : | ; * 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 : | ; * | ||
22 : | Skal | 1742 | ; * $Id: colorspace_yuv_mmx.asm,v 1.6 2006-10-30 10:52:00 Skal Exp $ |
23 : | edgomez | 1382 | ; * |
24 : | ; ***************************************************************************/ | ||
25 : | edgomez | 851 | |
26 : | BITS 32 | ||
27 : | |||
28 : | edgomez | 1382 | %macro cglobal 1 |
29 : | edgomez | 1535 | %ifdef PREFIX |
30 : | %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 : | ; Helper macros | ||
48 : | ;============================================================================= | ||
49 : | edgomez | 851 | |
50 : | ;------------------------------------------------------------------------------ | ||
51 : | Skal | 1742 | ; PLANE_COPY ( DST, DST_STRIDE, SRC, SRC_STRIDE, WIDTH, HEIGHT, OPT ) |
52 : | edgomez | 851 | ; DST dst buffer |
53 : | Skal | 1742 | ; DST_STRIDE dst stride |
54 : | edgomez | 851 | ; SRC src destination buffer |
55 : | Skal | 1742 | ; SRC_STRIDE src stride |
56 : | edgomez | 851 | ; WIDTH width |
57 : | ; HEIGHT height | ||
58 : | ; OPT 0=plain mmx, 1=xmm | ||
59 : | ;------------------------------------------------------------------------------ | ||
60 : | edgomez | 1382 | |
61 : | edgomez | 851 | %macro PLANE_COPY 7 |
62 : | Skal | 1742 | %define DST %1 |
63 : | %define DST_STRIDE %2 | ||
64 : | %define SRC %3 | ||
65 : | %define SRC_STRIDE %4 | ||
66 : | edgomez | 851 | %define WIDTH %5 |
67 : | %define HEIGHT %6 | ||
68 : | Skal | 1742 | %define OPT %7 |
69 : | edgomez | 851 | |
70 : | edgomez | 1382 | mov eax, WIDTH |
71 : | mov ebp, HEIGHT ; $ebp$ = height | ||
72 : | mov esi, SRC | ||
73 : | mov edi, DST | ||
74 : | edgomez | 851 | |
75 : | edgomez | 1382 | mov ebx, eax |
76 : | shr eax, 6 ; $eax$ = width / 64 | ||
77 : | and ebx, 63 ; remainder = width % 64 | ||
78 : | mov edx, ebx | ||
79 : | shr ebx, 4 ; $ebx$ = remainder / 16 | ||
80 : | and edx, 15 ; $edx$ = remainder % 16 | ||
81 : | edgomez | 851 | |
82 : | Skal | 1742 | %%loop64_start_pc: |
83 : | push edi | ||
84 : | push esi | ||
85 : | mov ecx, eax ; width64 | ||
86 : | test eax, eax | ||
87 : | jz %%loop16_start_pc | ||
88 : | |||
89 : | %%loop64_pc: | ||
90 : | edgomez | 1382 | %if OPT == 1 ; xmm |
91 : | prefetchnta [esi + 64] ; non temporal prefetch | ||
92 : | prefetchnta [esi + 96] | ||
93 : | edgomez | 851 | %endif |
94 : | Skal | 1742 | movq mm1, [esi ] ; read from src |
95 : | movq mm2, [esi + 8] | ||
96 : | edgomez | 1382 | movq mm3, [esi + 16] |
97 : | movq mm4, [esi + 24] | ||
98 : | movq mm5, [esi + 32] | ||
99 : | movq mm6, [esi + 40] | ||
100 : | movq mm7, [esi + 48] | ||
101 : | movq mm0, [esi + 56] | ||
102 : | edgomez | 851 | |
103 : | edgomez | 1382 | %if OPT == 0 ; plain mmx |
104 : | Skal | 1742 | movq [edi ], mm1 ; write to y_out |
105 : | movq [edi + 8], mm2 | ||
106 : | edgomez | 1382 | movq [edi + 16], mm3 |
107 : | movq [edi + 24], mm4 | ||
108 : | movq [edi + 32], mm5 | ||
109 : | movq [edi + 40], mm6 | ||
110 : | movq [edi + 48], mm7 | ||
111 : | movq [edi + 56], mm0 | ||
112 : | edgomez | 851 | %else |
113 : | Skal | 1742 | movntq [edi ], mm1 ; write to y_out |
114 : | movntq [edi + 8], mm2 | ||
115 : | edgomez | 1382 | movntq [edi + 16], mm3 |
116 : | movntq [edi + 24], mm4 | ||
117 : | movntq [edi + 32], mm5 | ||
118 : | movntq [edi + 40], mm6 | ||
119 : | movntq [edi + 48], mm7 | ||
120 : | movntq [edi + 56], mm0 | ||
121 : | edgomez | 851 | %endif |
122 : | |||
123 : | edgomez | 1382 | add esi, 64 |
124 : | add edi, 64 | ||
125 : | Skal | 1742 | loop %%loop64_pc |
126 : | edgomez | 851 | |
127 : | |||
128 : | Skal | 1742 | %%loop16_start_pc: |
129 : | mov ecx, ebx ; width16 | ||
130 : | test ebx, ebx | ||
131 : | jz %%loop1_start_pc | ||
132 : | |||
133 : | %%loop16_pc: | ||
134 : | edgomez | 1382 | movq mm1, [esi] |
135 : | movq mm2, [esi + 8] | ||
136 : | %if OPT == 0 ; plain mmx | ||
137 : | movq [edi], mm1 | ||
138 : | movq [edi + 8], mm2 | ||
139 : | edgomez | 851 | %else |
140 : | edgomez | 1382 | movntq [edi], mm1 |
141 : | movntq [edi + 8], mm2 | ||
142 : | edgomez | 851 | %endif |
143 : | |||
144 : | edgomez | 1382 | add esi, 16 |
145 : | add edi, 16 | ||
146 : | Skal | 1742 | loop %%loop16_pc |
147 : | edgomez | 851 | |
148 : | |||
149 : | Skal | 1742 | %%loop1_start_pc: |
150 : | edgomez | 1382 | mov ecx, edx |
151 : | rep movsb | ||
152 : | edgomez | 851 | |
153 : | Skal | 1742 | pop esi |
154 : | pop edi | ||
155 : | add esi, SRC_STRIDE | ||
156 : | add edi, DST_STRIDE | ||
157 : | edgomez | 1382 | dec ebp |
158 : | Skal | 1742 | jg near %%loop64_start_pc |
159 : | edgomez | 851 | %endmacro |
160 : | |||
161 : | ;------------------------------------------------------------------------------ | ||
162 : | Skal | 1742 | ; PLANE_FILL ( DST, DST_STRIDE, WIDTH, HEIGHT, OPT ) |
163 : | ; DST dst buffer | ||
164 : | ; DST_STRIDE dst stride | ||
165 : | ; WIDTH width | ||
166 : | ; HEIGHT height | ||
167 : | ; OPT 0=plain mmx, 1=xmm | ||
168 : | ;------------------------------------------------------------------------------ | ||
169 : | |||
170 : | %macro PLANE_FILL 5 | ||
171 : | %define DST %1 | ||
172 : | %define DST_STRIDE %2 | ||
173 : | %define WIDTH %3 | ||
174 : | %define HEIGHT %4 | ||
175 : | %define OPT %5 | ||
176 : | |||
177 : | mov esi, WIDTH | ||
178 : | mov ebp, HEIGHT ; $ebp$ = height | ||
179 : | mov edi, DST | ||
180 : | |||
181 : | mov eax, 0x80808080 | ||
182 : | mov ebx, esi | ||
183 : | shr esi, 6 ; $esi$ = width / 64 | ||
184 : | and ebx, 63 ; ebx = remainder = width % 64 | ||
185 : | movd mm0, eax | ||
186 : | mov edx, ebx | ||
187 : | shr ebx, 4 ; $ebx$ = remainder / 16 | ||
188 : | and edx, 15 ; $edx$ = remainder % 16 | ||
189 : | punpckldq mm0, mm0 | ||
190 : | |||
191 : | %%loop64_start_pf: | ||
192 : | push edi | ||
193 : | mov ecx, esi ; width64 | ||
194 : | test esi, esi | ||
195 : | jz %%loop16_start_pf | ||
196 : | |||
197 : | %%loop64_pf: | ||
198 : | |||
199 : | %if OPT == 0 ; plain mmx | ||
200 : | movq [edi ], mm0 ; write to y_out | ||
201 : | movq [edi + 8], mm0 | ||
202 : | movq [edi + 16], mm0 | ||
203 : | movq [edi + 24], mm0 | ||
204 : | movq [edi + 32], mm0 | ||
205 : | movq [edi + 40], mm0 | ||
206 : | movq [edi + 48], mm0 | ||
207 : | movq [edi + 56], mm0 | ||
208 : | %else | ||
209 : | movntq [edi ], mm0 ; write to y_out | ||
210 : | movntq [edi + 8], mm0 | ||
211 : | movntq [edi + 16], mm0 | ||
212 : | movntq [edi + 24], mm0 | ||
213 : | movntq [edi + 32], mm0 | ||
214 : | movntq [edi + 40], mm0 | ||
215 : | movntq [edi + 48], mm0 | ||
216 : | movntq [edi + 56], mm0 | ||
217 : | %endif | ||
218 : | |||
219 : | add edi, 64 | ||
220 : | loop %%loop64_pf | ||
221 : | |||
222 : | %%loop16_start_pf: | ||
223 : | mov ecx, ebx ; width16 | ||
224 : | test ebx, ebx | ||
225 : | jz %%loop1_start_pf | ||
226 : | |||
227 : | %%loop16_pf: | ||
228 : | %if OPT == 0 ; plain mmx | ||
229 : | movq [edi ], mm0 | ||
230 : | movq [edi + 8], mm0 | ||
231 : | %else | ||
232 : | movntq [edi ], mm0 | ||
233 : | movntq [edi + 8], mm0 | ||
234 : | %endif | ||
235 : | |||
236 : | add edi, 16 | ||
237 : | loop %%loop16_pf | ||
238 : | |||
239 : | %%loop1_start_pf: | ||
240 : | mov ecx, edx | ||
241 : | rep stosb | ||
242 : | |||
243 : | pop edi | ||
244 : | add edi, DST_STRIDE | ||
245 : | dec ebp | ||
246 : | jg near %%loop64_start_pf | ||
247 : | %endmacro | ||
248 : | |||
249 : | ;------------------------------------------------------------------------------ | ||
250 : | edgomez | 851 | ; MAKE_YV12_TO_YV12( NAME, OPT ) |
251 : | ; NAME function name | ||
252 : | ; OPT 0=plain mmx, 1=xmm | ||
253 : | ; | ||
254 : | edgomez | 1382 | ; yv12_to_yv12_mmx(uint8_t * y_dst, uint8_t * u_dst, uint8_t * v_dst, |
255 : | edgomez | 851 | ; int y_dst_stride, int uv_dst_stride, |
256 : | edgomez | 1382 | ; uint8_t * y_src, uint8_t * u_src, uint8_t * v_src, |
257 : | edgomez | 851 | ; int y_src_stride, int uv_src_stride, |
258 : | ; int width, int height, int vflip) | ||
259 : | ;------------------------------------------------------------------------------ | ||
260 : | %macro MAKE_YV12_TO_YV12 2 | ||
261 : | %define NAME %1 | ||
262 : | Skal | 1742 | %define OPT %2 |
263 : | edgomez | 1382 | ALIGN 16 |
264 : | edgomez | 851 | cglobal NAME |
265 : | edgomez | 1382 | NAME: |
266 : | edgomez | 851 | %define pushsize 16 |
267 : | Skal | 1742 | %define localsize 12 |
268 : | edgomez | 851 | |
269 : | Skal | 1742 | %define vflip esp + localsize + pushsize + 52 |
270 : | %define height esp + localsize + pushsize + 48 | ||
271 : | edgomez | 851 | %define width esp + localsize + pushsize + 44 |
272 : | %define uv_src_stride esp + localsize + pushsize + 40 | ||
273 : | %define y_src_stride esp + localsize + pushsize + 36 | ||
274 : | Skal | 1742 | %define v_src esp + localsize + pushsize + 32 |
275 : | %define u_src esp + localsize + pushsize + 28 | ||
276 : | %define y_src esp + localsize + pushsize + 24 | ||
277 : | edgomez | 851 | %define uv_dst_stride esp + localsize + pushsize + 20 |
278 : | %define y_dst_stride esp + localsize + pushsize + 16 | ||
279 : | Skal | 1742 | %define v_dst esp + localsize + pushsize + 12 |
280 : | %define u_dst esp + localsize + pushsize + 8 | ||
281 : | %define y_dst esp + localsize + pushsize + 4 | ||
282 : | %define _ip esp + localsize + pushsize + 0 | ||
283 : | edgomez | 851 | |
284 : | edgomez | 1382 | push ebx ; esp + localsize + 16 |
285 : | push esi ; esp + localsize + 8 | ||
286 : | push edi ; esp + localsize + 4 | ||
287 : | push ebp ; esp + localsize + 0 | ||
288 : | edgomez | 851 | |
289 : | %define width2 esp + localsize - 4 | ||
290 : | %define height2 esp + localsize - 8 | ||
291 : | |||
292 : | edgomez | 1382 | sub esp, localsize |
293 : | edgomez | 851 | |
294 : | edgomez | 1382 | mov eax, [width] |
295 : | mov ebx, [height] | ||
296 : | shr eax, 1 ; calculate widht/2, heigh/2 | ||
297 : | shr ebx, 1 | ||
298 : | mov [width2], eax | ||
299 : | mov [height2], ebx | ||
300 : | edgomez | 851 | |
301 : | Skal | 1742 | mov eax, [vflip] |
302 : | test eax, eax | ||
303 : | jz near .go | ||
304 : | edgomez | 851 | |
305 : | Skal | 1742 | ; flipping support |
306 : | edgomez | 1382 | mov eax, [height] |
307 : | mov esi, [y_src] | ||
308 : | Skal | 1742 | mov ecx, [y_src_stride] |
309 : | sub eax, 1 | ||
310 : | imul eax, ecx | ||
311 : | edgomez | 1382 | add esi, eax ; y_src += (height-1) * y_src_stride |
312 : | Skal | 1742 | neg ecx |
313 : | edgomez | 1382 | mov [y_src], esi |
314 : | Skal | 1742 | mov [y_src_stride], ecx ; y_src_stride = -y_src_stride |
315 : | edgomez | 851 | |
316 : | edgomez | 1382 | mov eax, [height2] |
317 : | mov esi, [u_src] | ||
318 : | mov edi, [v_src] | ||
319 : | Skal | 1742 | mov ecx, [uv_src_stride] |
320 : | test esi, esi | ||
321 : | jz .go | ||
322 : | test edi, edi | ||
323 : | jz .go | ||
324 : | sub eax, 1 ; eax = height2 - 1 | ||
325 : | imul eax, ecx | ||
326 : | edgomez | 1382 | add esi, eax ; u_src += (height2-1) * uv_src_stride |
327 : | add edi, eax ; v_src += (height2-1) * uv_src_stride | ||
328 : | Skal | 1742 | neg ecx |
329 : | edgomez | 1382 | mov [u_src], esi |
330 : | mov [v_src], edi | ||
331 : | Skal | 1742 | mov [uv_src_stride], ecx ; uv_src_stride = -uv_src_stride |
332 : | edgomez | 851 | |
333 : | Skal | 1742 | .go: |
334 : | edgomez | 851 | |
335 : | Skal | 1742 | PLANE_COPY [y_dst], [y_dst_stride], [y_src], [y_src_stride], [width], [height], OPT |
336 : | edgomez | 851 | |
337 : | Skal | 1742 | mov eax, [u_src] |
338 : | or eax, [v_src] | ||
339 : | jz near .UVFill_0x80 | ||
340 : | PLANE_COPY [u_dst], [uv_dst_stride], [u_src], [uv_src_stride], [width2], [height2], OPT | ||
341 : | PLANE_COPY [v_dst], [uv_dst_stride], [v_src], [uv_src_stride], [width2], [height2], OPT | ||
342 : | edgomez | 851 | |
343 : | Skal | 1742 | .Done_UVPlane: |
344 : | edgomez | 1382 | add esp, localsize |
345 : | pop ebp | ||
346 : | pop edi | ||
347 : | pop esi | ||
348 : | pop ebx | ||
349 : | Skal | 1742 | ret |
350 : | edgomez | 851 | |
351 : | Skal | 1742 | .UVFill_0x80: |
352 : | PLANE_FILL [u_dst], [uv_dst_stride], [width2], [height2], OPT | ||
353 : | PLANE_FILL [v_dst], [uv_dst_stride], [width2], [height2], OPT | ||
354 : | jmp near .Done_UVPlane | ||
355 : | edgomez | 1540 | .endfunc |
356 : | edgomez | 851 | %endmacro |
357 : | |||
358 : | edgomez | 1382 | ;============================================================================= |
359 : | ; Code | ||
360 : | ;============================================================================= | ||
361 : | edgomez | 851 | |
362 : | edgomez | 1382 | SECTION .text |
363 : | |||
364 : | edgomez | 851 | MAKE_YV12_TO_YV12 yv12_to_yv12_mmx, 0 |
365 : | edgomez | 1382 | |
366 : | edgomez | 851 | MAKE_YV12_TO_YV12 yv12_to_yv12_xmm, 1 |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |