Parent Directory | Revision Log
Revision 1793 - (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 : | Isibaar | 1793 | ; * $Id: colorspace_yuv_mmx.asm,v 1.8 2008-11-11 20:46:24 Isibaar 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 : | Isibaar | 1793 | %define ENDFUNC .endfunc |
34 : | edgomez | 1535 | %else |
35 : | global _%1 | ||
36 : | %define %1 _%1 | ||
37 : | Isibaar | 1793 | %define ENDFUNC |
38 : | edgomez | 1535 | %endif |
39 : | edgomez | 851 | %else |
40 : | edgomez | 1535 | %ifdef MARK_FUNCS |
41 : | edgomez | 1540 | global %1:function %1.endfunc-%1 |
42 : | Isibaar | 1793 | %define ENDFUNC .endfunc |
43 : | edgomez | 1535 | %else |
44 : | global %1 | ||
45 : | Isibaar | 1793 | %define ENDFUNC |
46 : | edgomez | 1535 | %endif |
47 : | edgomez | 851 | %endif |
48 : | %endmacro | ||
49 : | |||
50 : | edgomez | 1382 | ;============================================================================= |
51 : | ; Helper macros | ||
52 : | ;============================================================================= | ||
53 : | edgomez | 851 | |
54 : | ;------------------------------------------------------------------------------ | ||
55 : | Skal | 1742 | ; PLANE_COPY ( DST, DST_STRIDE, SRC, SRC_STRIDE, WIDTH, HEIGHT, OPT ) |
56 : | edgomez | 851 | ; DST dst buffer |
57 : | Skal | 1742 | ; DST_STRIDE dst stride |
58 : | edgomez | 851 | ; SRC src destination buffer |
59 : | Skal | 1742 | ; SRC_STRIDE src stride |
60 : | edgomez | 851 | ; WIDTH width |
61 : | ; HEIGHT height | ||
62 : | ; OPT 0=plain mmx, 1=xmm | ||
63 : | ;------------------------------------------------------------------------------ | ||
64 : | edgomez | 1382 | |
65 : | edgomez | 851 | %macro PLANE_COPY 7 |
66 : | Skal | 1742 | %define DST %1 |
67 : | %define DST_STRIDE %2 | ||
68 : | %define SRC %3 | ||
69 : | %define SRC_STRIDE %4 | ||
70 : | edgomez | 851 | %define WIDTH %5 |
71 : | %define HEIGHT %6 | ||
72 : | Skal | 1742 | %define OPT %7 |
73 : | edgomez | 851 | |
74 : | edgomez | 1382 | mov eax, WIDTH |
75 : | mov ebp, HEIGHT ; $ebp$ = height | ||
76 : | mov esi, SRC | ||
77 : | mov edi, DST | ||
78 : | edgomez | 851 | |
79 : | edgomez | 1382 | mov ebx, eax |
80 : | shr eax, 6 ; $eax$ = width / 64 | ||
81 : | and ebx, 63 ; remainder = width % 64 | ||
82 : | mov edx, ebx | ||
83 : | shr ebx, 4 ; $ebx$ = remainder / 16 | ||
84 : | and edx, 15 ; $edx$ = remainder % 16 | ||
85 : | edgomez | 851 | |
86 : | Skal | 1742 | %%loop64_start_pc: |
87 : | push edi | ||
88 : | push esi | ||
89 : | mov ecx, eax ; width64 | ||
90 : | test eax, eax | ||
91 : | jz %%loop16_start_pc | ||
92 : | |||
93 : | %%loop64_pc: | ||
94 : | edgomez | 1382 | %if OPT == 1 ; xmm |
95 : | prefetchnta [esi + 64] ; non temporal prefetch | ||
96 : | prefetchnta [esi + 96] | ||
97 : | edgomez | 851 | %endif |
98 : | Skal | 1742 | movq mm1, [esi ] ; read from src |
99 : | movq mm2, [esi + 8] | ||
100 : | edgomez | 1382 | movq mm3, [esi + 16] |
101 : | movq mm4, [esi + 24] | ||
102 : | movq mm5, [esi + 32] | ||
103 : | movq mm6, [esi + 40] | ||
104 : | movq mm7, [esi + 48] | ||
105 : | movq mm0, [esi + 56] | ||
106 : | edgomez | 851 | |
107 : | edgomez | 1382 | %if OPT == 0 ; plain mmx |
108 : | Skal | 1742 | movq [edi ], mm1 ; write to y_out |
109 : | movq [edi + 8], mm2 | ||
110 : | edgomez | 1382 | movq [edi + 16], mm3 |
111 : | movq [edi + 24], mm4 | ||
112 : | movq [edi + 32], mm5 | ||
113 : | movq [edi + 40], mm6 | ||
114 : | movq [edi + 48], mm7 | ||
115 : | movq [edi + 56], mm0 | ||
116 : | edgomez | 851 | %else |
117 : | Skal | 1742 | movntq [edi ], mm1 ; write to y_out |
118 : | movntq [edi + 8], mm2 | ||
119 : | edgomez | 1382 | movntq [edi + 16], mm3 |
120 : | movntq [edi + 24], mm4 | ||
121 : | movntq [edi + 32], mm5 | ||
122 : | movntq [edi + 40], mm6 | ||
123 : | movntq [edi + 48], mm7 | ||
124 : | movntq [edi + 56], mm0 | ||
125 : | edgomez | 851 | %endif |
126 : | |||
127 : | edgomez | 1382 | add esi, 64 |
128 : | add edi, 64 | ||
129 : | Skal | 1742 | loop %%loop64_pc |
130 : | edgomez | 851 | |
131 : | |||
132 : | Skal | 1742 | %%loop16_start_pc: |
133 : | mov ecx, ebx ; width16 | ||
134 : | test ebx, ebx | ||
135 : | jz %%loop1_start_pc | ||
136 : | |||
137 : | %%loop16_pc: | ||
138 : | edgomez | 1382 | movq mm1, [esi] |
139 : | movq mm2, [esi + 8] | ||
140 : | %if OPT == 0 ; plain mmx | ||
141 : | movq [edi], mm1 | ||
142 : | movq [edi + 8], mm2 | ||
143 : | edgomez | 851 | %else |
144 : | edgomez | 1382 | movntq [edi], mm1 |
145 : | movntq [edi + 8], mm2 | ||
146 : | edgomez | 851 | %endif |
147 : | |||
148 : | edgomez | 1382 | add esi, 16 |
149 : | add edi, 16 | ||
150 : | Skal | 1742 | loop %%loop16_pc |
151 : | edgomez | 851 | |
152 : | |||
153 : | Skal | 1742 | %%loop1_start_pc: |
154 : | edgomez | 1382 | mov ecx, edx |
155 : | rep movsb | ||
156 : | edgomez | 851 | |
157 : | Skal | 1742 | pop esi |
158 : | pop edi | ||
159 : | add esi, SRC_STRIDE | ||
160 : | add edi, DST_STRIDE | ||
161 : | edgomez | 1382 | dec ebp |
162 : | Skal | 1742 | jg near %%loop64_start_pc |
163 : | edgomez | 851 | %endmacro |
164 : | |||
165 : | ;------------------------------------------------------------------------------ | ||
166 : | Skal | 1742 | ; PLANE_FILL ( DST, DST_STRIDE, WIDTH, HEIGHT, OPT ) |
167 : | ; DST dst buffer | ||
168 : | ; DST_STRIDE dst stride | ||
169 : | ; WIDTH width | ||
170 : | ; HEIGHT height | ||
171 : | ; OPT 0=plain mmx, 1=xmm | ||
172 : | ;------------------------------------------------------------------------------ | ||
173 : | |||
174 : | %macro PLANE_FILL 5 | ||
175 : | %define DST %1 | ||
176 : | %define DST_STRIDE %2 | ||
177 : | %define WIDTH %3 | ||
178 : | %define HEIGHT %4 | ||
179 : | %define OPT %5 | ||
180 : | |||
181 : | mov esi, WIDTH | ||
182 : | mov ebp, HEIGHT ; $ebp$ = height | ||
183 : | mov edi, DST | ||
184 : | |||
185 : | mov eax, 0x80808080 | ||
186 : | mov ebx, esi | ||
187 : | shr esi, 6 ; $esi$ = width / 64 | ||
188 : | and ebx, 63 ; ebx = remainder = width % 64 | ||
189 : | movd mm0, eax | ||
190 : | mov edx, ebx | ||
191 : | shr ebx, 4 ; $ebx$ = remainder / 16 | ||
192 : | and edx, 15 ; $edx$ = remainder % 16 | ||
193 : | punpckldq mm0, mm0 | ||
194 : | |||
195 : | %%loop64_start_pf: | ||
196 : | push edi | ||
197 : | mov ecx, esi ; width64 | ||
198 : | test esi, esi | ||
199 : | jz %%loop16_start_pf | ||
200 : | |||
201 : | %%loop64_pf: | ||
202 : | |||
203 : | %if OPT == 0 ; plain mmx | ||
204 : | movq [edi ], mm0 ; write to y_out | ||
205 : | movq [edi + 8], mm0 | ||
206 : | movq [edi + 16], mm0 | ||
207 : | movq [edi + 24], mm0 | ||
208 : | movq [edi + 32], mm0 | ||
209 : | movq [edi + 40], mm0 | ||
210 : | movq [edi + 48], mm0 | ||
211 : | movq [edi + 56], mm0 | ||
212 : | %else | ||
213 : | movntq [edi ], mm0 ; write to y_out | ||
214 : | movntq [edi + 8], mm0 | ||
215 : | movntq [edi + 16], mm0 | ||
216 : | movntq [edi + 24], mm0 | ||
217 : | movntq [edi + 32], mm0 | ||
218 : | movntq [edi + 40], mm0 | ||
219 : | movntq [edi + 48], mm0 | ||
220 : | movntq [edi + 56], mm0 | ||
221 : | %endif | ||
222 : | |||
223 : | add edi, 64 | ||
224 : | loop %%loop64_pf | ||
225 : | |||
226 : | %%loop16_start_pf: | ||
227 : | mov ecx, ebx ; width16 | ||
228 : | test ebx, ebx | ||
229 : | jz %%loop1_start_pf | ||
230 : | |||
231 : | %%loop16_pf: | ||
232 : | %if OPT == 0 ; plain mmx | ||
233 : | movq [edi ], mm0 | ||
234 : | movq [edi + 8], mm0 | ||
235 : | %else | ||
236 : | movntq [edi ], mm0 | ||
237 : | movntq [edi + 8], mm0 | ||
238 : | %endif | ||
239 : | |||
240 : | add edi, 16 | ||
241 : | loop %%loop16_pf | ||
242 : | |||
243 : | %%loop1_start_pf: | ||
244 : | mov ecx, edx | ||
245 : | rep stosb | ||
246 : | |||
247 : | pop edi | ||
248 : | add edi, DST_STRIDE | ||
249 : | dec ebp | ||
250 : | jg near %%loop64_start_pf | ||
251 : | %endmacro | ||
252 : | |||
253 : | ;------------------------------------------------------------------------------ | ||
254 : | edgomez | 851 | ; MAKE_YV12_TO_YV12( NAME, OPT ) |
255 : | ; NAME function name | ||
256 : | ; OPT 0=plain mmx, 1=xmm | ||
257 : | ; | ||
258 : | edgomez | 1382 | ; yv12_to_yv12_mmx(uint8_t * y_dst, uint8_t * u_dst, uint8_t * v_dst, |
259 : | edgomez | 851 | ; int y_dst_stride, int uv_dst_stride, |
260 : | edgomez | 1382 | ; uint8_t * y_src, uint8_t * u_src, uint8_t * v_src, |
261 : | edgomez | 851 | ; int y_src_stride, int uv_src_stride, |
262 : | ; int width, int height, int vflip) | ||
263 : | ;------------------------------------------------------------------------------ | ||
264 : | %macro MAKE_YV12_TO_YV12 2 | ||
265 : | %define NAME %1 | ||
266 : | Skal | 1742 | %define OPT %2 |
267 : | edgomez | 1382 | ALIGN 16 |
268 : | edgomez | 851 | cglobal NAME |
269 : | edgomez | 1382 | NAME: |
270 : | edgomez | 851 | %define pushsize 16 |
271 : | Skal | 1742 | %define localsize 12 |
272 : | edgomez | 851 | |
273 : | Skal | 1742 | %define vflip esp + localsize + pushsize + 52 |
274 : | %define height esp + localsize + pushsize + 48 | ||
275 : | edgomez | 851 | %define width esp + localsize + pushsize + 44 |
276 : | %define uv_src_stride esp + localsize + pushsize + 40 | ||
277 : | %define y_src_stride esp + localsize + pushsize + 36 | ||
278 : | Skal | 1742 | %define v_src esp + localsize + pushsize + 32 |
279 : | %define u_src esp + localsize + pushsize + 28 | ||
280 : | %define y_src esp + localsize + pushsize + 24 | ||
281 : | edgomez | 851 | %define uv_dst_stride esp + localsize + pushsize + 20 |
282 : | %define y_dst_stride esp + localsize + pushsize + 16 | ||
283 : | Skal | 1742 | %define v_dst esp + localsize + pushsize + 12 |
284 : | %define u_dst esp + localsize + pushsize + 8 | ||
285 : | %define y_dst esp + localsize + pushsize + 4 | ||
286 : | %define _ip esp + localsize + pushsize + 0 | ||
287 : | edgomez | 851 | |
288 : | edgomez | 1382 | push ebx ; esp + localsize + 16 |
289 : | push esi ; esp + localsize + 8 | ||
290 : | push edi ; esp + localsize + 4 | ||
291 : | push ebp ; esp + localsize + 0 | ||
292 : | edgomez | 851 | |
293 : | %define width2 esp + localsize - 4 | ||
294 : | %define height2 esp + localsize - 8 | ||
295 : | |||
296 : | edgomez | 1382 | sub esp, localsize |
297 : | edgomez | 851 | |
298 : | edgomez | 1382 | mov eax, [width] |
299 : | mov ebx, [height] | ||
300 : | shr eax, 1 ; calculate widht/2, heigh/2 | ||
301 : | shr ebx, 1 | ||
302 : | mov [width2], eax | ||
303 : | mov [height2], ebx | ||
304 : | edgomez | 851 | |
305 : | Skal | 1742 | mov eax, [vflip] |
306 : | test eax, eax | ||
307 : | jz near .go | ||
308 : | edgomez | 851 | |
309 : | Skal | 1742 | ; flipping support |
310 : | edgomez | 1382 | mov eax, [height] |
311 : | mov esi, [y_src] | ||
312 : | Skal | 1742 | mov ecx, [y_src_stride] |
313 : | sub eax, 1 | ||
314 : | imul eax, ecx | ||
315 : | edgomez | 1382 | add esi, eax ; y_src += (height-1) * y_src_stride |
316 : | Skal | 1742 | neg ecx |
317 : | edgomez | 1382 | mov [y_src], esi |
318 : | Skal | 1742 | mov [y_src_stride], ecx ; y_src_stride = -y_src_stride |
319 : | edgomez | 851 | |
320 : | edgomez | 1382 | mov eax, [height2] |
321 : | mov esi, [u_src] | ||
322 : | mov edi, [v_src] | ||
323 : | Skal | 1742 | mov ecx, [uv_src_stride] |
324 : | test esi, esi | ||
325 : | jz .go | ||
326 : | test edi, edi | ||
327 : | jz .go | ||
328 : | sub eax, 1 ; eax = height2 - 1 | ||
329 : | imul eax, ecx | ||
330 : | edgomez | 1382 | add esi, eax ; u_src += (height2-1) * uv_src_stride |
331 : | add edi, eax ; v_src += (height2-1) * uv_src_stride | ||
332 : | Skal | 1742 | neg ecx |
333 : | edgomez | 1382 | mov [u_src], esi |
334 : | mov [v_src], edi | ||
335 : | Skal | 1742 | mov [uv_src_stride], ecx ; uv_src_stride = -uv_src_stride |
336 : | edgomez | 851 | |
337 : | Skal | 1742 | .go: |
338 : | edgomez | 851 | |
339 : | Skal | 1742 | PLANE_COPY [y_dst], [y_dst_stride], [y_src], [y_src_stride], [width], [height], OPT |
340 : | edgomez | 851 | |
341 : | Skal | 1742 | mov eax, [u_src] |
342 : | or eax, [v_src] | ||
343 : | jz near .UVFill_0x80 | ||
344 : | PLANE_COPY [u_dst], [uv_dst_stride], [u_src], [uv_src_stride], [width2], [height2], OPT | ||
345 : | PLANE_COPY [v_dst], [uv_dst_stride], [v_src], [uv_src_stride], [width2], [height2], OPT | ||
346 : | edgomez | 851 | |
347 : | Skal | 1742 | .Done_UVPlane: |
348 : | edgomez | 1382 | add esp, localsize |
349 : | pop ebp | ||
350 : | pop edi | ||
351 : | pop esi | ||
352 : | pop ebx | ||
353 : | Skal | 1742 | ret |
354 : | edgomez | 851 | |
355 : | Skal | 1742 | .UVFill_0x80: |
356 : | PLANE_FILL [u_dst], [uv_dst_stride], [width2], [height2], OPT | ||
357 : | PLANE_FILL [v_dst], [uv_dst_stride], [width2], [height2], OPT | ||
358 : | jmp near .Done_UVPlane | ||
359 : | Isibaar | 1793 | ENDFUNC |
360 : | edgomez | 851 | %endmacro |
361 : | |||
362 : | edgomez | 1382 | ;============================================================================= |
363 : | ; Code | ||
364 : | ;============================================================================= | ||
365 : | edgomez | 851 | |
366 : | edgomez | 1382 | SECTION .text |
367 : | |||
368 : | edgomez | 851 | MAKE_YV12_TO_YV12 yv12_to_yv12_mmx, 0 |
369 : | edgomez | 1382 | |
370 : | edgomez | 851 | MAKE_YV12_TO_YV12 yv12_to_yv12_xmm, 1 |
371 : | Isibaar | 1790 | |
372 : | %ifidn __OUTPUT_FORMAT__,elf | ||
373 : | section ".note.GNU-stack" noalloc noexec nowrite progbits | ||
374 : | %endif | ||
375 : |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |