Mercurial > libavcodec.hg
annotate i386/dsputil_mmx_avg.h @ 3029:7a92269a3ccd libavcodec
tweak h264_biweight
author | lorenm |
---|---|
date | Mon, 09 Jan 2006 03:38:37 +0000 |
parents | bfabfdf9ce55 |
children | 0b546eab515d |
rev | line source |
---|---|
0 | 1 /* |
2 * DSP utils : average functions are compiled twice for 3dnow/mmx2 | |
429 | 3 * Copyright (c) 2000, 2001 Fabrice Bellard. |
1739
07a484280a82
copyright year update of the files i touched and remembered, things look annoyingly unmaintained otherwise
michael
parents:
1064
diff
changeset
|
4 * Copyright (c) 2002-2004 Michael Niedermayer |
0 | 5 * |
429 | 6 * This library is free software; you can redistribute it and/or |
7 * modify it under the terms of the GNU Lesser General Public | |
8 * License as published by the Free Software Foundation; either | |
9 * version 2 of the License, or (at your option) any later version. | |
0 | 10 * |
429 | 11 * This library is distributed in the hope that it will be useful, |
0 | 12 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
429 | 13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
14 * Lesser General Public License for more details. | |
0 | 15 * |
429 | 16 * You should have received a copy of the GNU Lesser General Public |
17 * License along with this library; if not, write to the Free Software | |
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | |
0 | 19 * |
20 * MMX optimization by Nick Kurshev <nickols_k@mail.ru> | |
386 | 21 * mostly rewritten by Michael Niedermayer <michaelni@gmx.at> |
441 | 22 * and improved by Zdenek Kabelac <kabi@users.sf.net> |
0 | 23 */ |
2967 | 24 |
389
f874493a1970
tried to avoid gcc 2.95.2 bug by puting explicit register constraints - added comment about rounding bug in some functions (need to correct or suppress them for regression tests)
glantau
parents:
387
diff
changeset
|
25 /* XXX: we use explicit registers to avoid a gcc 2.95.2 register asm |
413
1548abb7bbed
* fix for -fPIC compilation - compiles with 2.95.2 as well - any
kabi
parents:
402
diff
changeset
|
26 clobber bug - now it will work with 2.95.2 and also with -fPIC |
1548abb7bbed
* fix for -fPIC compilation - compiles with 2.95.2 as well - any
kabi
parents:
402
diff
changeset
|
27 */ |
1064 | 28 static void DEF(put_pixels8_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
0 | 29 { |
386 | 30 __asm __volatile( |
2979 | 31 "lea (%3, %3), %%"REG_a" \n\t" |
32 "1: \n\t" | |
33 "movq (%1), %%mm0 \n\t" | |
34 "movq (%1, %3), %%mm1 \n\t" | |
35 PAVGB" 1(%1), %%mm0 \n\t" | |
36 PAVGB" 1(%1, %3), %%mm1 \n\t" | |
37 "movq %%mm0, (%2) \n\t" | |
38 "movq %%mm1, (%2, %3) \n\t" | |
39 "add %%"REG_a", %1 \n\t" | |
40 "add %%"REG_a", %2 \n\t" | |
41 "movq (%1), %%mm0 \n\t" | |
42 "movq (%1, %3), %%mm1 \n\t" | |
43 PAVGB" 1(%1), %%mm0 \n\t" | |
44 PAVGB" 1(%1, %3), %%mm1 \n\t" | |
45 "add %%"REG_a", %1 \n\t" | |
46 "movq %%mm0, (%2) \n\t" | |
47 "movq %%mm1, (%2, %3) \n\t" | |
48 "add %%"REG_a", %2 \n\t" | |
49 "subl $4, %0 \n\t" | |
50 "jnz 1b \n\t" | |
51 :"+g"(h), "+S"(pixels), "+D"(block) | |
52 :"r" ((long)line_size) | |
53 :"%"REG_a, "memory"); | |
441 | 54 } |
651 | 55 |
2209 | 56 static void DEF(put_pixels4_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h) |
57 { | |
58 __asm __volatile( | |
2979 | 59 "testl $1, %0 \n\t" |
60 " jz 1f \n\t" | |
61 "movd (%1), %%mm0 \n\t" | |
62 "movd (%2), %%mm1 \n\t" | |
63 "add %4, %1 \n\t" | |
64 "add $4, %2 \n\t" | |
65 PAVGB" %%mm1, %%mm0 \n\t" | |
66 "movd %%mm0, (%3) \n\t" | |
67 "add %5, %3 \n\t" | |
68 "decl %0 \n\t" | |
69 "1: \n\t" | |
70 "movd (%1), %%mm0 \n\t" | |
71 "add %4, %1 \n\t" | |
72 "movd (%1), %%mm1 \n\t" | |
73 "movd (%2), %%mm2 \n\t" | |
74 "movd 4(%2), %%mm3 \n\t" | |
75 "add %4, %1 \n\t" | |
76 PAVGB" %%mm2, %%mm0 \n\t" | |
77 PAVGB" %%mm3, %%mm1 \n\t" | |
78 "movd %%mm0, (%3) \n\t" | |
79 "add %5, %3 \n\t" | |
80 "movd %%mm1, (%3) \n\t" | |
81 "add %5, %3 \n\t" | |
82 "movd (%1), %%mm0 \n\t" | |
83 "add %4, %1 \n\t" | |
84 "movd (%1), %%mm1 \n\t" | |
85 "movd 8(%2), %%mm2 \n\t" | |
86 "movd 12(%2), %%mm3 \n\t" | |
87 "add %4, %1 \n\t" | |
88 PAVGB" %%mm2, %%mm0 \n\t" | |
89 PAVGB" %%mm3, %%mm1 \n\t" | |
90 "movd %%mm0, (%3) \n\t" | |
91 "add %5, %3 \n\t" | |
92 "movd %%mm1, (%3) \n\t" | |
93 "add %5, %3 \n\t" | |
94 "add $16, %2 \n\t" | |
95 "subl $4, %0 \n\t" | |
96 "jnz 1b \n\t" | |
2209 | 97 #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used |
2979 | 98 :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) |
2209 | 99 #else |
2979 | 100 :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) |
2209 | 101 #endif |
2979 | 102 :"S"((long)src1Stride), "D"((long)dstStride) |
103 :"memory"); | |
2209 | 104 } |
105 | |
106 | |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
107 static void DEF(put_pixels8_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h) |
954 | 108 { |
109 __asm __volatile( | |
2979 | 110 "testl $1, %0 \n\t" |
111 " jz 1f \n\t" | |
112 "movq (%1), %%mm0 \n\t" | |
113 "movq (%2), %%mm1 \n\t" | |
114 "add %4, %1 \n\t" | |
115 "add $8, %2 \n\t" | |
116 PAVGB" %%mm1, %%mm0 \n\t" | |
117 "movq %%mm0, (%3) \n\t" | |
118 "add %5, %3 \n\t" | |
119 "decl %0 \n\t" | |
120 "1: \n\t" | |
121 "movq (%1), %%mm0 \n\t" | |
122 "add %4, %1 \n\t" | |
123 "movq (%1), %%mm1 \n\t" | |
124 "add %4, %1 \n\t" | |
125 PAVGB" (%2), %%mm0 \n\t" | |
126 PAVGB" 8(%2), %%mm1 \n\t" | |
127 "movq %%mm0, (%3) \n\t" | |
128 "add %5, %3 \n\t" | |
129 "movq %%mm1, (%3) \n\t" | |
130 "add %5, %3 \n\t" | |
131 "movq (%1), %%mm0 \n\t" | |
132 "add %4, %1 \n\t" | |
133 "movq (%1), %%mm1 \n\t" | |
134 "add %4, %1 \n\t" | |
135 PAVGB" 16(%2), %%mm0 \n\t" | |
136 PAVGB" 24(%2), %%mm1 \n\t" | |
137 "movq %%mm0, (%3) \n\t" | |
138 "add %5, %3 \n\t" | |
139 "movq %%mm1, (%3) \n\t" | |
140 "add %5, %3 \n\t" | |
141 "add $32, %2 \n\t" | |
142 "subl $4, %0 \n\t" | |
143 "jnz 1b \n\t" | |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
144 #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used |
2979 | 145 :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
146 #else |
2979 | 147 :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
148 #endif |
2979 | 149 :"S"((long)src1Stride), "D"((long)dstStride) |
150 :"memory"); | |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
151 //the following should be used, though better not with gcc ... |
2979 | 152 /* :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst) |
153 :"r"(src1Stride), "r"(dstStride) | |
154 :"memory");*/ | |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
155 } |
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
156 |
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
157 static void DEF(put_no_rnd_pixels8_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h) |
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
158 { |
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
159 __asm __volatile( |
2979 | 160 "pcmpeqb %%mm6, %%mm6 \n\t" |
161 "testl $1, %0 \n\t" | |
162 " jz 1f \n\t" | |
163 "movq (%1), %%mm0 \n\t" | |
164 "movq (%2), %%mm1 \n\t" | |
165 "add %4, %1 \n\t" | |
166 "add $8, %2 \n\t" | |
167 "pxor %%mm6, %%mm0 \n\t" | |
168 "pxor %%mm6, %%mm1 \n\t" | |
169 PAVGB" %%mm1, %%mm0 \n\t" | |
170 "pxor %%mm6, %%mm0 \n\t" | |
171 "movq %%mm0, (%3) \n\t" | |
172 "add %5, %3 \n\t" | |
173 "decl %0 \n\t" | |
174 "1: \n\t" | |
175 "movq (%1), %%mm0 \n\t" | |
176 "add %4, %1 \n\t" | |
177 "movq (%1), %%mm1 \n\t" | |
178 "add %4, %1 \n\t" | |
179 "movq (%2), %%mm2 \n\t" | |
180 "movq 8(%2), %%mm3 \n\t" | |
181 "pxor %%mm6, %%mm0 \n\t" | |
182 "pxor %%mm6, %%mm1 \n\t" | |
183 "pxor %%mm6, %%mm2 \n\t" | |
184 "pxor %%mm6, %%mm3 \n\t" | |
185 PAVGB" %%mm2, %%mm0 \n\t" | |
186 PAVGB" %%mm3, %%mm1 \n\t" | |
187 "pxor %%mm6, %%mm0 \n\t" | |
188 "pxor %%mm6, %%mm1 \n\t" | |
189 "movq %%mm0, (%3) \n\t" | |
190 "add %5, %3 \n\t" | |
191 "movq %%mm1, (%3) \n\t" | |
192 "add %5, %3 \n\t" | |
193 "movq (%1), %%mm0 \n\t" | |
194 "add %4, %1 \n\t" | |
195 "movq (%1), %%mm1 \n\t" | |
196 "add %4, %1 \n\t" | |
197 "movq 16(%2), %%mm2 \n\t" | |
198 "movq 24(%2), %%mm3 \n\t" | |
199 "pxor %%mm6, %%mm0 \n\t" | |
200 "pxor %%mm6, %%mm1 \n\t" | |
201 "pxor %%mm6, %%mm2 \n\t" | |
202 "pxor %%mm6, %%mm3 \n\t" | |
203 PAVGB" %%mm2, %%mm0 \n\t" | |
204 PAVGB" %%mm3, %%mm1 \n\t" | |
205 "pxor %%mm6, %%mm0 \n\t" | |
206 "pxor %%mm6, %%mm1 \n\t" | |
207 "movq %%mm0, (%3) \n\t" | |
208 "add %5, %3 \n\t" | |
209 "movq %%mm1, (%3) \n\t" | |
210 "add %5, %3 \n\t" | |
211 "add $32, %2 \n\t" | |
212 "subl $4, %0 \n\t" | |
213 "jnz 1b \n\t" | |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
214 #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used |
2979 | 215 :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
216 #else |
2979 | 217 :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
218 #endif |
2979 | 219 :"S"((long)src1Stride), "D"((long)dstStride) |
220 :"memory"); | |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
221 //the following should be used, though better not with gcc ... |
2979 | 222 /* :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst) |
223 :"r"(src1Stride), "r"(dstStride) | |
224 :"memory");*/ | |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
225 } |
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
226 |
2209 | 227 static void DEF(avg_pixels4_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h) |
228 { | |
229 __asm __volatile( | |
2979 | 230 "testl $1, %0 \n\t" |
231 " jz 1f \n\t" | |
232 "movd (%1), %%mm0 \n\t" | |
233 "movd (%2), %%mm1 \n\t" | |
234 "add %4, %1 \n\t" | |
235 "add $4, %2 \n\t" | |
236 PAVGB" %%mm1, %%mm0 \n\t" | |
237 PAVGB" (%3), %%mm0 \n\t" | |
238 "movd %%mm0, (%3) \n\t" | |
239 "add %5, %3 \n\t" | |
240 "decl %0 \n\t" | |
241 "1: \n\t" | |
242 "movd (%1), %%mm0 \n\t" | |
243 "add %4, %1 \n\t" | |
244 "movd (%1), %%mm1 \n\t" | |
245 "add %4, %1 \n\t" | |
246 PAVGB" (%2), %%mm0 \n\t" | |
247 PAVGB" 4(%2), %%mm1 \n\t" | |
248 PAVGB" (%3), %%mm0 \n\t" | |
249 "movd %%mm0, (%3) \n\t" | |
250 "add %5, %3 \n\t" | |
251 PAVGB" (%3), %%mm1 \n\t" | |
252 "movd %%mm1, (%3) \n\t" | |
253 "add %5, %3 \n\t" | |
254 "movd (%1), %%mm0 \n\t" | |
255 "add %4, %1 \n\t" | |
256 "movd (%1), %%mm1 \n\t" | |
257 "add %4, %1 \n\t" | |
258 PAVGB" 8(%2), %%mm0 \n\t" | |
259 PAVGB" 12(%2), %%mm1 \n\t" | |
260 PAVGB" (%3), %%mm0 \n\t" | |
261 "movd %%mm0, (%3) \n\t" | |
262 "add %5, %3 \n\t" | |
263 PAVGB" (%3), %%mm1 \n\t" | |
264 "movd %%mm1, (%3) \n\t" | |
265 "add %5, %3 \n\t" | |
266 "add $16, %2 \n\t" | |
267 "subl $4, %0 \n\t" | |
268 "jnz 1b \n\t" | |
2209 | 269 #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used |
2979 | 270 :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) |
2209 | 271 #else |
2979 | 272 :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) |
2209 | 273 #endif |
2979 | 274 :"S"((long)src1Stride), "D"((long)dstStride) |
275 :"memory"); | |
2209 | 276 } |
277 | |
278 | |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
279 static void DEF(avg_pixels8_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h) |
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
280 { |
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
281 __asm __volatile( |
2979 | 282 "testl $1, %0 \n\t" |
283 " jz 1f \n\t" | |
284 "movq (%1), %%mm0 \n\t" | |
285 "movq (%2), %%mm1 \n\t" | |
286 "add %4, %1 \n\t" | |
287 "add $8, %2 \n\t" | |
288 PAVGB" %%mm1, %%mm0 \n\t" | |
289 PAVGB" (%3), %%mm0 \n\t" | |
290 "movq %%mm0, (%3) \n\t" | |
291 "add %5, %3 \n\t" | |
292 "decl %0 \n\t" | |
293 "1: \n\t" | |
294 "movq (%1), %%mm0 \n\t" | |
295 "add %4, %1 \n\t" | |
296 "movq (%1), %%mm1 \n\t" | |
297 "add %4, %1 \n\t" | |
298 PAVGB" (%2), %%mm0 \n\t" | |
299 PAVGB" 8(%2), %%mm1 \n\t" | |
300 PAVGB" (%3), %%mm0 \n\t" | |
301 "movq %%mm0, (%3) \n\t" | |
302 "add %5, %3 \n\t" | |
303 PAVGB" (%3), %%mm1 \n\t" | |
304 "movq %%mm1, (%3) \n\t" | |
305 "add %5, %3 \n\t" | |
306 "movq (%1), %%mm0 \n\t" | |
307 "add %4, %1 \n\t" | |
308 "movq (%1), %%mm1 \n\t" | |
309 "add %4, %1 \n\t" | |
310 PAVGB" 16(%2), %%mm0 \n\t" | |
311 PAVGB" 24(%2), %%mm1 \n\t" | |
312 PAVGB" (%3), %%mm0 \n\t" | |
313 "movq %%mm0, (%3) \n\t" | |
314 "add %5, %3 \n\t" | |
315 PAVGB" (%3), %%mm1 \n\t" | |
316 "movq %%mm1, (%3) \n\t" | |
317 "add %5, %3 \n\t" | |
318 "add $32, %2 \n\t" | |
319 "subl $4, %0 \n\t" | |
320 "jnz 1b \n\t" | |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
321 #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used |
2979 | 322 :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
323 #else |
2979 | 324 :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
325 #endif |
2979 | 326 :"S"((long)src1Stride), "D"((long)dstStride) |
327 :"memory"); | |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
328 //the following should be used, though better not with gcc ... |
2979 | 329 /* :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst) |
330 :"r"(src1Stride), "r"(dstStride) | |
331 :"memory");*/ | |
954 | 332 } |
333 | |
1064 | 334 static void DEF(put_pixels16_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
651 | 335 { |
336 __asm __volatile( | |
2979 | 337 "lea (%3, %3), %%"REG_a" \n\t" |
338 "1: \n\t" | |
339 "movq (%1), %%mm0 \n\t" | |
340 "movq (%1, %3), %%mm1 \n\t" | |
341 "movq 8(%1), %%mm2 \n\t" | |
342 "movq 8(%1, %3), %%mm3 \n\t" | |
343 PAVGB" 1(%1), %%mm0 \n\t" | |
344 PAVGB" 1(%1, %3), %%mm1 \n\t" | |
345 PAVGB" 9(%1), %%mm2 \n\t" | |
346 PAVGB" 9(%1, %3), %%mm3 \n\t" | |
347 "movq %%mm0, (%2) \n\t" | |
348 "movq %%mm1, (%2, %3) \n\t" | |
349 "movq %%mm2, 8(%2) \n\t" | |
350 "movq %%mm3, 8(%2, %3) \n\t" | |
351 "add %%"REG_a", %1 \n\t" | |
352 "add %%"REG_a", %2 \n\t" | |
353 "movq (%1), %%mm0 \n\t" | |
354 "movq (%1, %3), %%mm1 \n\t" | |
355 "movq 8(%1), %%mm2 \n\t" | |
356 "movq 8(%1, %3), %%mm3 \n\t" | |
357 PAVGB" 1(%1), %%mm0 \n\t" | |
358 PAVGB" 1(%1, %3), %%mm1 \n\t" | |
359 PAVGB" 9(%1), %%mm2 \n\t" | |
360 PAVGB" 9(%1, %3), %%mm3 \n\t" | |
361 "add %%"REG_a", %1 \n\t" | |
362 "movq %%mm0, (%2) \n\t" | |
363 "movq %%mm1, (%2, %3) \n\t" | |
364 "movq %%mm2, 8(%2) \n\t" | |
365 "movq %%mm3, 8(%2, %3) \n\t" | |
366 "add %%"REG_a", %2 \n\t" | |
367 "subl $4, %0 \n\t" | |
368 "jnz 1b \n\t" | |
369 :"+g"(h), "+S"(pixels), "+D"(block) | |
370 :"r" ((long)line_size) | |
371 :"%"REG_a, "memory"); | |
651 | 372 } |
954 | 373 |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
374 static void DEF(put_pixels16_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h) |
954 | 375 { |
376 __asm __volatile( | |
2979 | 377 "testl $1, %0 \n\t" |
378 " jz 1f \n\t" | |
379 "movq (%1), %%mm0 \n\t" | |
380 "movq 8(%1), %%mm1 \n\t" | |
381 PAVGB" (%2), %%mm0 \n\t" | |
382 PAVGB" 8(%2), %%mm1 \n\t" | |
383 "add %4, %1 \n\t" | |
384 "add $16, %2 \n\t" | |
385 "movq %%mm0, (%3) \n\t" | |
386 "movq %%mm1, 8(%3) \n\t" | |
387 "add %5, %3 \n\t" | |
388 "decl %0 \n\t" | |
389 "1: \n\t" | |
390 "movq (%1), %%mm0 \n\t" | |
391 "movq 8(%1), %%mm1 \n\t" | |
392 "add %4, %1 \n\t" | |
393 PAVGB" (%2), %%mm0 \n\t" | |
394 PAVGB" 8(%2), %%mm1 \n\t" | |
395 "movq %%mm0, (%3) \n\t" | |
396 "movq %%mm1, 8(%3) \n\t" | |
397 "add %5, %3 \n\t" | |
398 "movq (%1), %%mm0 \n\t" | |
399 "movq 8(%1), %%mm1 \n\t" | |
400 "add %4, %1 \n\t" | |
401 PAVGB" 16(%2), %%mm0 \n\t" | |
402 PAVGB" 24(%2), %%mm1 \n\t" | |
403 "movq %%mm0, (%3) \n\t" | |
404 "movq %%mm1, 8(%3) \n\t" | |
405 "add %5, %3 \n\t" | |
406 "add $32, %2 \n\t" | |
407 "subl $2, %0 \n\t" | |
408 "jnz 1b \n\t" | |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
409 #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used |
2979 | 410 :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
411 #else |
2979 | 412 :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
413 #endif |
2979 | 414 :"S"((long)src1Stride), "D"((long)dstStride) |
415 :"memory"); | |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
416 //the following should be used, though better not with gcc ... |
2979 | 417 /* :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst) |
418 :"r"(src1Stride), "r"(dstStride) | |
419 :"memory");*/ | |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
420 } |
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
421 |
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
422 static void DEF(avg_pixels16_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h) |
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
423 { |
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
424 __asm __volatile( |
2979 | 425 "testl $1, %0 \n\t" |
426 " jz 1f \n\t" | |
427 "movq (%1), %%mm0 \n\t" | |
428 "movq 8(%1), %%mm1 \n\t" | |
429 PAVGB" (%2), %%mm0 \n\t" | |
430 PAVGB" 8(%2), %%mm1 \n\t" | |
431 "add %4, %1 \n\t" | |
432 "add $16, %2 \n\t" | |
433 PAVGB" (%3), %%mm0 \n\t" | |
434 PAVGB" 8(%3), %%mm1 \n\t" | |
435 "movq %%mm0, (%3) \n\t" | |
436 "movq %%mm1, 8(%3) \n\t" | |
437 "add %5, %3 \n\t" | |
438 "decl %0 \n\t" | |
439 "1: \n\t" | |
440 "movq (%1), %%mm0 \n\t" | |
441 "movq 8(%1), %%mm1 \n\t" | |
442 "add %4, %1 \n\t" | |
443 PAVGB" (%2), %%mm0 \n\t" | |
444 PAVGB" 8(%2), %%mm1 \n\t" | |
445 PAVGB" (%3), %%mm0 \n\t" | |
446 PAVGB" 8(%3), %%mm1 \n\t" | |
447 "movq %%mm0, (%3) \n\t" | |
448 "movq %%mm1, 8(%3) \n\t" | |
449 "add %5, %3 \n\t" | |
450 "movq (%1), %%mm0 \n\t" | |
451 "movq 8(%1), %%mm1 \n\t" | |
452 "add %4, %1 \n\t" | |
453 PAVGB" 16(%2), %%mm0 \n\t" | |
454 PAVGB" 24(%2), %%mm1 \n\t" | |
455 PAVGB" (%3), %%mm0 \n\t" | |
456 PAVGB" 8(%3), %%mm1 \n\t" | |
457 "movq %%mm0, (%3) \n\t" | |
458 "movq %%mm1, 8(%3) \n\t" | |
459 "add %5, %3 \n\t" | |
460 "add $32, %2 \n\t" | |
461 "subl $2, %0 \n\t" | |
462 "jnz 1b \n\t" | |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
463 #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used |
2979 | 464 :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
465 #else |
2979 | 466 :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
467 #endif |
2979 | 468 :"S"((long)src1Stride), "D"((long)dstStride) |
469 :"memory"); | |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
470 //the following should be used, though better not with gcc ... |
2979 | 471 /* :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst) |
472 :"r"(src1Stride), "r"(dstStride) | |
473 :"memory");*/ | |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
474 } |
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
475 |
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
476 static void DEF(put_no_rnd_pixels16_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h) |
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
477 { |
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
478 __asm __volatile( |
2979 | 479 "pcmpeqb %%mm6, %%mm6 \n\t" |
480 "testl $1, %0 \n\t" | |
481 " jz 1f \n\t" | |
482 "movq (%1), %%mm0 \n\t" | |
483 "movq 8(%1), %%mm1 \n\t" | |
484 "movq (%2), %%mm2 \n\t" | |
485 "movq 8(%2), %%mm3 \n\t" | |
486 "pxor %%mm6, %%mm0 \n\t" | |
487 "pxor %%mm6, %%mm1 \n\t" | |
488 "pxor %%mm6, %%mm2 \n\t" | |
489 "pxor %%mm6, %%mm3 \n\t" | |
490 PAVGB" %%mm2, %%mm0 \n\t" | |
491 PAVGB" %%mm3, %%mm1 \n\t" | |
492 "pxor %%mm6, %%mm0 \n\t" | |
493 "pxor %%mm6, %%mm1 \n\t" | |
494 "add %4, %1 \n\t" | |
495 "add $16, %2 \n\t" | |
496 "movq %%mm0, (%3) \n\t" | |
497 "movq %%mm1, 8(%3) \n\t" | |
498 "add %5, %3 \n\t" | |
499 "decl %0 \n\t" | |
500 "1: \n\t" | |
501 "movq (%1), %%mm0 \n\t" | |
502 "movq 8(%1), %%mm1 \n\t" | |
503 "add %4, %1 \n\t" | |
504 "movq (%2), %%mm2 \n\t" | |
505 "movq 8(%2), %%mm3 \n\t" | |
506 "pxor %%mm6, %%mm0 \n\t" | |
507 "pxor %%mm6, %%mm1 \n\t" | |
508 "pxor %%mm6, %%mm2 \n\t" | |
509 "pxor %%mm6, %%mm3 \n\t" | |
510 PAVGB" %%mm2, %%mm0 \n\t" | |
511 PAVGB" %%mm3, %%mm1 \n\t" | |
512 "pxor %%mm6, %%mm0 \n\t" | |
513 "pxor %%mm6, %%mm1 \n\t" | |
514 "movq %%mm0, (%3) \n\t" | |
515 "movq %%mm1, 8(%3) \n\t" | |
516 "add %5, %3 \n\t" | |
517 "movq (%1), %%mm0 \n\t" | |
518 "movq 8(%1), %%mm1 \n\t" | |
519 "add %4, %1 \n\t" | |
520 "movq 16(%2), %%mm2 \n\t" | |
521 "movq 24(%2), %%mm3 \n\t" | |
522 "pxor %%mm6, %%mm0 \n\t" | |
523 "pxor %%mm6, %%mm1 \n\t" | |
524 "pxor %%mm6, %%mm2 \n\t" | |
525 "pxor %%mm6, %%mm3 \n\t" | |
526 PAVGB" %%mm2, %%mm0 \n\t" | |
527 PAVGB" %%mm3, %%mm1 \n\t" | |
528 "pxor %%mm6, %%mm0 \n\t" | |
529 "pxor %%mm6, %%mm1 \n\t" | |
530 "movq %%mm0, (%3) \n\t" | |
531 "movq %%mm1, 8(%3) \n\t" | |
532 "add %5, %3 \n\t" | |
533 "add $32, %2 \n\t" | |
534 "subl $2, %0 \n\t" | |
535 "jnz 1b \n\t" | |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
536 #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used |
2979 | 537 :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
538 #else |
2979 | 539 :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
540 #endif |
2979 | 541 :"S"((long)src1Stride), "D"((long)dstStride) |
542 :"memory"); | |
2207
22b768f1261a
10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster
michael
parents:
2024
diff
changeset
|
543 //the following should be used, though better not with gcc ... |
2979 | 544 /* :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst) |
545 :"r"(src1Stride), "r"(dstStride) | |
546 :"memory");*/ | |
954 | 547 } |
2967 | 548 |
441 | 549 /* GL: this function does incorrect rounding if overflow */ |
1064 | 550 static void DEF(put_no_rnd_pixels8_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
441 | 551 { |
448 | 552 MOVQ_BONE(mm6); |
441 | 553 __asm __volatile( |
2979 | 554 "lea (%3, %3), %%"REG_a" \n\t" |
555 "1: \n\t" | |
556 "movq (%1), %%mm0 \n\t" | |
557 "movq (%1, %3), %%mm2 \n\t" | |
558 "movq 1(%1), %%mm1 \n\t" | |
559 "movq 1(%1, %3), %%mm3 \n\t" | |
560 "add %%"REG_a", %1 \n\t" | |
561 "psubusb %%mm6, %%mm0 \n\t" | |
562 "psubusb %%mm6, %%mm2 \n\t" | |
563 PAVGB" %%mm1, %%mm0 \n\t" | |
564 PAVGB" %%mm3, %%mm2 \n\t" | |
565 "movq %%mm0, (%2) \n\t" | |
566 "movq %%mm2, (%2, %3) \n\t" | |
567 "movq (%1), %%mm0 \n\t" | |
568 "movq 1(%1), %%mm1 \n\t" | |
569 "movq (%1, %3), %%mm2 \n\t" | |
570 "movq 1(%1, %3), %%mm3 \n\t" | |
571 "add %%"REG_a", %2 \n\t" | |
572 "add %%"REG_a", %1 \n\t" | |
573 "psubusb %%mm6, %%mm0 \n\t" | |
574 "psubusb %%mm6, %%mm2 \n\t" | |
575 PAVGB" %%mm1, %%mm0 \n\t" | |
576 PAVGB" %%mm3, %%mm2 \n\t" | |
577 "movq %%mm0, (%2) \n\t" | |
578 "movq %%mm2, (%2, %3) \n\t" | |
579 "add %%"REG_a", %2 \n\t" | |
580 "subl $4, %0 \n\t" | |
581 "jnz 1b \n\t" | |
582 :"+g"(h), "+S"(pixels), "+D"(block) | |
583 :"r" ((long)line_size) | |
584 :"%"REG_a, "memory"); | |
0 | 585 } |
586 | |
1064 | 587 static void DEF(put_pixels8_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
0 | 588 { |
413
1548abb7bbed
* fix for -fPIC compilation - compiles with 2.95.2 as well - any
kabi
parents:
402
diff
changeset
|
589 __asm __volatile( |
2979 | 590 "lea (%3, %3), %%"REG_a" \n\t" |
591 "movq (%1), %%mm0 \n\t" | |
592 "sub %3, %2 \n\t" | |
593 "1: \n\t" | |
594 "movq (%1, %3), %%mm1 \n\t" | |
595 "movq (%1, %%"REG_a"), %%mm2 \n\t" | |
596 "add %%"REG_a", %1 \n\t" | |
597 PAVGB" %%mm1, %%mm0 \n\t" | |
598 PAVGB" %%mm2, %%mm1 \n\t" | |
599 "movq %%mm0, (%2, %3) \n\t" | |
600 "movq %%mm1, (%2, %%"REG_a") \n\t" | |
601 "movq (%1, %3), %%mm1 \n\t" | |
602 "movq (%1, %%"REG_a"), %%mm0 \n\t" | |
603 "add %%"REG_a", %2 \n\t" | |
604 "add %%"REG_a", %1 \n\t" | |
605 PAVGB" %%mm1, %%mm2 \n\t" | |
606 PAVGB" %%mm0, %%mm1 \n\t" | |
607 "movq %%mm2, (%2, %3) \n\t" | |
608 "movq %%mm1, (%2, %%"REG_a") \n\t" | |
609 "add %%"REG_a", %2 \n\t" | |
610 "subl $4, %0 \n\t" | |
611 "jnz 1b \n\t" | |
612 :"+g"(h), "+S"(pixels), "+D" (block) | |
613 :"r" ((long)line_size) | |
614 :"%"REG_a, "memory"); | |
386 | 615 } |
616 | |
389
f874493a1970
tried to avoid gcc 2.95.2 bug by puting explicit register constraints - added comment about rounding bug in some functions (need to correct or suppress them for regression tests)
glantau
parents:
387
diff
changeset
|
617 /* GL: this function does incorrect rounding if overflow */ |
1064 | 618 static void DEF(put_no_rnd_pixels8_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
386 | 619 { |
448 | 620 MOVQ_BONE(mm6); |
0 | 621 __asm __volatile( |
2979 | 622 "lea (%3, %3), %%"REG_a" \n\t" |
623 "movq (%1), %%mm0 \n\t" | |
624 "sub %3, %2 \n\t" | |
625 "1: \n\t" | |
626 "movq (%1, %3), %%mm1 \n\t" | |
627 "movq (%1, %%"REG_a"), %%mm2 \n\t" | |
628 "add %%"REG_a", %1 \n\t" | |
629 "psubusb %%mm6, %%mm1 \n\t" | |
630 PAVGB" %%mm1, %%mm0 \n\t" | |
631 PAVGB" %%mm2, %%mm1 \n\t" | |
632 "movq %%mm0, (%2, %3) \n\t" | |
633 "movq %%mm1, (%2, %%"REG_a") \n\t" | |
634 "movq (%1, %3), %%mm1 \n\t" | |
635 "movq (%1, %%"REG_a"), %%mm0 \n\t" | |
636 "add %%"REG_a", %2 \n\t" | |
637 "add %%"REG_a", %1 \n\t" | |
638 "psubusb %%mm6, %%mm1 \n\t" | |
639 PAVGB" %%mm1, %%mm2 \n\t" | |
640 PAVGB" %%mm0, %%mm1 \n\t" | |
641 "movq %%mm2, (%2, %3) \n\t" | |
642 "movq %%mm1, (%2, %%"REG_a") \n\t" | |
643 "add %%"REG_a", %2 \n\t" | |
644 "subl $4, %0 \n\t" | |
645 "jnz 1b \n\t" | |
646 :"+g"(h), "+S"(pixels), "+D" (block) | |
647 :"r" ((long)line_size) | |
648 :"%"REG_a, "memory"); | |
0 | 649 } |
650 | |
1064 | 651 static void DEF(avg_pixels8)(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
0 | 652 { |
653 __asm __volatile( | |
2979 | 654 "lea (%3, %3), %%"REG_a" \n\t" |
655 "1: \n\t" | |
656 "movq (%2), %%mm0 \n\t" | |
657 "movq (%2, %3), %%mm1 \n\t" | |
658 PAVGB" (%1), %%mm0 \n\t" | |
659 PAVGB" (%1, %3), %%mm1 \n\t" | |
660 "movq %%mm0, (%2) \n\t" | |
661 "movq %%mm1, (%2, %3) \n\t" | |
662 "add %%"REG_a", %1 \n\t" | |
663 "add %%"REG_a", %2 \n\t" | |
664 "movq (%2), %%mm0 \n\t" | |
665 "movq (%2, %3), %%mm1 \n\t" | |
666 PAVGB" (%1), %%mm0 \n\t" | |
667 PAVGB" (%1, %3), %%mm1 \n\t" | |
668 "add %%"REG_a", %1 \n\t" | |
669 "movq %%mm0, (%2) \n\t" | |
670 "movq %%mm1, (%2, %3) \n\t" | |
671 "add %%"REG_a", %2 \n\t" | |
672 "subl $4, %0 \n\t" | |
673 "jnz 1b \n\t" | |
674 :"+g"(h), "+S"(pixels), "+D"(block) | |
675 :"r" ((long)line_size) | |
676 :"%"REG_a, "memory"); | |
0 | 677 } |
678 | |
1064 | 679 static void DEF(avg_pixels8_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
0 | 680 { |
681 __asm __volatile( | |
2979 | 682 "lea (%3, %3), %%"REG_a" \n\t" |
683 "1: \n\t" | |
684 "movq (%1), %%mm0 \n\t" | |
685 "movq (%1, %3), %%mm2 \n\t" | |
686 PAVGB" 1(%1), %%mm0 \n\t" | |
687 PAVGB" 1(%1, %3), %%mm2 \n\t" | |
688 PAVGB" (%2), %%mm0 \n\t" | |
689 PAVGB" (%2, %3), %%mm2 \n\t" | |
690 "add %%"REG_a", %1 \n\t" | |
691 "movq %%mm0, (%2) \n\t" | |
692 "movq %%mm2, (%2, %3) \n\t" | |
693 "movq (%1), %%mm0 \n\t" | |
694 "movq (%1, %3), %%mm2 \n\t" | |
695 PAVGB" 1(%1), %%mm0 \n\t" | |
696 PAVGB" 1(%1, %3), %%mm2 \n\t" | |
697 "add %%"REG_a", %2 \n\t" | |
698 "add %%"REG_a", %1 \n\t" | |
699 PAVGB" (%2), %%mm0 \n\t" | |
700 PAVGB" (%2, %3), %%mm2 \n\t" | |
701 "movq %%mm0, (%2) \n\t" | |
702 "movq %%mm2, (%2, %3) \n\t" | |
703 "add %%"REG_a", %2 \n\t" | |
704 "subl $4, %0 \n\t" | |
705 "jnz 1b \n\t" | |
706 :"+g"(h), "+S"(pixels), "+D"(block) | |
707 :"r" ((long)line_size) | |
708 :"%"REG_a, "memory"); | |
0 | 709 } |
710 | |
1064 | 711 static void DEF(avg_pixels8_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
0 | 712 { |
713 __asm __volatile( | |
2979 | 714 "lea (%3, %3), %%"REG_a" \n\t" |
715 "movq (%1), %%mm0 \n\t" | |
716 "sub %3, %2 \n\t" | |
717 "1: \n\t" | |
718 "movq (%1, %3), %%mm1 \n\t" | |
719 "movq (%1, %%"REG_a"), %%mm2 \n\t" | |
720 "add %%"REG_a", %1 \n\t" | |
721 PAVGB" %%mm1, %%mm0 \n\t" | |
722 PAVGB" %%mm2, %%mm1 \n\t" | |
723 "movq (%2, %3), %%mm3 \n\t" | |
724 "movq (%2, %%"REG_a"), %%mm4 \n\t" | |
725 PAVGB" %%mm3, %%mm0 \n\t" | |
726 PAVGB" %%mm4, %%mm1 \n\t" | |
727 "movq %%mm0, (%2, %3) \n\t" | |
728 "movq %%mm1, (%2, %%"REG_a") \n\t" | |
729 "movq (%1, %3), %%mm1 \n\t" | |
730 "movq (%1, %%"REG_a"), %%mm0 \n\t" | |
731 PAVGB" %%mm1, %%mm2 \n\t" | |
732 PAVGB" %%mm0, %%mm1 \n\t" | |
733 "add %%"REG_a", %2 \n\t" | |
734 "add %%"REG_a", %1 \n\t" | |
735 "movq (%2, %3), %%mm3 \n\t" | |
736 "movq (%2, %%"REG_a"), %%mm4 \n\t" | |
737 PAVGB" %%mm3, %%mm2 \n\t" | |
738 PAVGB" %%mm4, %%mm1 \n\t" | |
739 "movq %%mm2, (%2, %3) \n\t" | |
740 "movq %%mm1, (%2, %%"REG_a") \n\t" | |
741 "add %%"REG_a", %2 \n\t" | |
742 "subl $4, %0 \n\t" | |
743 "jnz 1b \n\t" | |
744 :"+g"(h), "+S"(pixels), "+D"(block) | |
745 :"r" ((long)line_size) | |
746 :"%"REG_a, "memory"); | |
0 | 747 } |
748 | |
2967 | 749 // Note this is not correctly rounded, but this function is only used for b frames so it doesnt matter |
1064 | 750 static void DEF(avg_pixels8_xy2)(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
0 | 751 { |
448 | 752 MOVQ_BONE(mm6); |
0 | 753 __asm __volatile( |
2979 | 754 "lea (%3, %3), %%"REG_a" \n\t" |
755 "movq (%1), %%mm0 \n\t" | |
756 PAVGB" 1(%1), %%mm0 \n\t" | |
757 ".balign 8 \n\t" | |
758 "1: \n\t" | |
759 "movq (%1, %%"REG_a"), %%mm2 \n\t" | |
760 "movq (%1, %3), %%mm1 \n\t" | |
761 "psubusb %%mm6, %%mm2 \n\t" | |
762 PAVGB" 1(%1, %3), %%mm1 \n\t" | |
763 PAVGB" 1(%1, %%"REG_a"), %%mm2 \n\t" | |
764 "add %%"REG_a", %1 \n\t" | |
765 PAVGB" %%mm1, %%mm0 \n\t" | |
766 PAVGB" %%mm2, %%mm1 \n\t" | |
767 PAVGB" (%2), %%mm0 \n\t" | |
768 PAVGB" (%2, %3), %%mm1 \n\t" | |
769 "movq %%mm0, (%2) \n\t" | |
770 "movq %%mm1, (%2, %3) \n\t" | |
771 "movq (%1, %3), %%mm1 \n\t" | |
772 "movq (%1, %%"REG_a"), %%mm0 \n\t" | |
773 PAVGB" 1(%1, %3), %%mm1 \n\t" | |
774 PAVGB" 1(%1, %%"REG_a"), %%mm0 \n\t" | |
775 "add %%"REG_a", %2 \n\t" | |
776 "add %%"REG_a", %1 \n\t" | |
777 PAVGB" %%mm1, %%mm2 \n\t" | |
778 PAVGB" %%mm0, %%mm1 \n\t" | |
779 PAVGB" (%2), %%mm2 \n\t" | |
780 PAVGB" (%2, %3), %%mm1 \n\t" | |
781 "movq %%mm2, (%2) \n\t" | |
782 "movq %%mm1, (%2, %3) \n\t" | |
783 "add %%"REG_a", %2 \n\t" | |
784 "subl $4, %0 \n\t" | |
785 "jnz 1b \n\t" | |
786 :"+g"(h), "+S"(pixels), "+D"(block) | |
787 :"r" ((long)line_size) | |
788 :"%"REG_a, "memory"); | |
0 | 789 } |
651 | 790 |
791 //FIXME the following could be optimized too ... | |
1064 | 792 static void DEF(put_no_rnd_pixels16_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){ |
651 | 793 DEF(put_no_rnd_pixels8_x2)(block , pixels , line_size, h); |
794 DEF(put_no_rnd_pixels8_x2)(block+8, pixels+8, line_size, h); | |
795 } | |
1064 | 796 static void DEF(put_pixels16_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){ |
651 | 797 DEF(put_pixels8_y2)(block , pixels , line_size, h); |
798 DEF(put_pixels8_y2)(block+8, pixels+8, line_size, h); | |
799 } | |
1064 | 800 static void DEF(put_no_rnd_pixels16_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){ |
651 | 801 DEF(put_no_rnd_pixels8_y2)(block , pixels , line_size, h); |
802 DEF(put_no_rnd_pixels8_y2)(block+8, pixels+8, line_size, h); | |
803 } | |
1064 | 804 static void DEF(avg_pixels16)(uint8_t *block, const uint8_t *pixels, int line_size, int h){ |
651 | 805 DEF(avg_pixels8)(block , pixels , line_size, h); |
806 DEF(avg_pixels8)(block+8, pixels+8, line_size, h); | |
807 } | |
1064 | 808 static void DEF(avg_pixels16_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){ |
651 | 809 DEF(avg_pixels8_x2)(block , pixels , line_size, h); |
810 DEF(avg_pixels8_x2)(block+8, pixels+8, line_size, h); | |
811 } | |
1064 | 812 static void DEF(avg_pixels16_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){ |
651 | 813 DEF(avg_pixels8_y2)(block , pixels , line_size, h); |
814 DEF(avg_pixels8_y2)(block+8, pixels+8, line_size, h); | |
815 } | |
1064 | 816 static void DEF(avg_pixels16_xy2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){ |
651 | 817 DEF(avg_pixels8_xy2)(block , pixels , line_size, h); |
818 DEF(avg_pixels8_xy2)(block+8, pixels+8, line_size, h); | |
819 } | |
820 |