Mercurial > libavcodec.hg
annotate i386/dsputil_mmx_avg.h @ 4116:7a08ccdc8550 libavcodec
Use common define for x86_32 and x86_64.
author | diego |
---|---|
date | Wed, 01 Nov 2006 18:34:40 +0000 |
parents | c8c591fe26f8 |
children | 4dbe6578f811 |
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 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3807
diff
changeset
|
6 * This file is part of FFmpeg. |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3807
diff
changeset
|
7 * |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3807
diff
changeset
|
8 * FFmpeg is free software; you can redistribute it and/or |
429 | 9 * modify it under the terms of the GNU Lesser General Public |
10 * License as published by the Free Software Foundation; either | |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3807
diff
changeset
|
11 * version 2.1 of the License, or (at your option) any later version. |
0 | 12 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3807
diff
changeset
|
13 * FFmpeg is distributed in the hope that it will be useful, |
0 | 14 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
429 | 15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
16 * Lesser General Public License for more details. | |
0 | 17 * |
429 | 18 * You should have received a copy of the GNU Lesser General Public |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3807
diff
changeset
|
19 * License along with FFmpeg; if not, write to the Free Software |
3036
0b546eab515d
Update licensing information: The FSF changed postal address.
diego
parents:
2979
diff
changeset
|
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
0 | 21 * |
22 * MMX optimization by Nick Kurshev <nickols_k@mail.ru> | |
386 | 23 * mostly rewritten by Michael Niedermayer <michaelni@gmx.at> |
441 | 24 * and improved by Zdenek Kabelac <kabi@users.sf.net> |
0 | 25 */ |
2967 | 26 |
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
|
27 /* 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
|
28 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
|
29 */ |
1064 | 30 static void DEF(put_pixels8_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
0 | 31 { |
386 | 32 __asm __volatile( |
2979 | 33 "lea (%3, %3), %%"REG_a" \n\t" |
34 "1: \n\t" | |
35 "movq (%1), %%mm0 \n\t" | |
36 "movq (%1, %3), %%mm1 \n\t" | |
37 PAVGB" 1(%1), %%mm0 \n\t" | |
38 PAVGB" 1(%1, %3), %%mm1 \n\t" | |
39 "movq %%mm0, (%2) \n\t" | |
40 "movq %%mm1, (%2, %3) \n\t" | |
41 "add %%"REG_a", %1 \n\t" | |
42 "add %%"REG_a", %2 \n\t" | |
43 "movq (%1), %%mm0 \n\t" | |
44 "movq (%1, %3), %%mm1 \n\t" | |
45 PAVGB" 1(%1), %%mm0 \n\t" | |
46 PAVGB" 1(%1, %3), %%mm1 \n\t" | |
47 "add %%"REG_a", %1 \n\t" | |
48 "movq %%mm0, (%2) \n\t" | |
49 "movq %%mm1, (%2, %3) \n\t" | |
50 "add %%"REG_a", %2 \n\t" | |
51 "subl $4, %0 \n\t" | |
52 "jnz 1b \n\t" | |
53 :"+g"(h), "+S"(pixels), "+D"(block) | |
54 :"r" ((long)line_size) | |
55 :"%"REG_a, "memory"); | |
441 | 56 } |
651 | 57 |
2209 | 58 static void DEF(put_pixels4_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h) |
59 { | |
60 __asm __volatile( | |
2979 | 61 "testl $1, %0 \n\t" |
62 " jz 1f \n\t" | |
63 "movd (%1), %%mm0 \n\t" | |
64 "movd (%2), %%mm1 \n\t" | |
65 "add %4, %1 \n\t" | |
66 "add $4, %2 \n\t" | |
67 PAVGB" %%mm1, %%mm0 \n\t" | |
68 "movd %%mm0, (%3) \n\t" | |
69 "add %5, %3 \n\t" | |
70 "decl %0 \n\t" | |
71 "1: \n\t" | |
72 "movd (%1), %%mm0 \n\t" | |
73 "add %4, %1 \n\t" | |
74 "movd (%1), %%mm1 \n\t" | |
75 "movd (%2), %%mm2 \n\t" | |
76 "movd 4(%2), %%mm3 \n\t" | |
77 "add %4, %1 \n\t" | |
78 PAVGB" %%mm2, %%mm0 \n\t" | |
79 PAVGB" %%mm3, %%mm1 \n\t" | |
80 "movd %%mm0, (%3) \n\t" | |
81 "add %5, %3 \n\t" | |
82 "movd %%mm1, (%3) \n\t" | |
83 "add %5, %3 \n\t" | |
84 "movd (%1), %%mm0 \n\t" | |
85 "add %4, %1 \n\t" | |
86 "movd (%1), %%mm1 \n\t" | |
87 "movd 8(%2), %%mm2 \n\t" | |
88 "movd 12(%2), %%mm3 \n\t" | |
89 "add %4, %1 \n\t" | |
90 PAVGB" %%mm2, %%mm0 \n\t" | |
91 PAVGB" %%mm3, %%mm1 \n\t" | |
92 "movd %%mm0, (%3) \n\t" | |
93 "add %5, %3 \n\t" | |
94 "movd %%mm1, (%3) \n\t" | |
95 "add %5, %3 \n\t" | |
96 "add $16, %2 \n\t" | |
97 "subl $4, %0 \n\t" | |
98 "jnz 1b \n\t" | |
2209 | 99 #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used |
2979 | 100 :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) |
2209 | 101 #else |
2979 | 102 :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) |
2209 | 103 #endif |
2979 | 104 :"S"((long)src1Stride), "D"((long)dstStride) |
105 :"memory"); | |
2209 | 106 } |
107 | |
108 | |
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
|
109 static void DEF(put_pixels8_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h) |
954 | 110 { |
111 __asm __volatile( | |
2979 | 112 "testl $1, %0 \n\t" |
113 " jz 1f \n\t" | |
114 "movq (%1), %%mm0 \n\t" | |
115 "movq (%2), %%mm1 \n\t" | |
116 "add %4, %1 \n\t" | |
117 "add $8, %2 \n\t" | |
118 PAVGB" %%mm1, %%mm0 \n\t" | |
119 "movq %%mm0, (%3) \n\t" | |
120 "add %5, %3 \n\t" | |
121 "decl %0 \n\t" | |
122 "1: \n\t" | |
123 "movq (%1), %%mm0 \n\t" | |
124 "add %4, %1 \n\t" | |
125 "movq (%1), %%mm1 \n\t" | |
126 "add %4, %1 \n\t" | |
127 PAVGB" (%2), %%mm0 \n\t" | |
128 PAVGB" 8(%2), %%mm1 \n\t" | |
129 "movq %%mm0, (%3) \n\t" | |
130 "add %5, %3 \n\t" | |
131 "movq %%mm1, (%3) \n\t" | |
132 "add %5, %3 \n\t" | |
133 "movq (%1), %%mm0 \n\t" | |
134 "add %4, %1 \n\t" | |
135 "movq (%1), %%mm1 \n\t" | |
136 "add %4, %1 \n\t" | |
137 PAVGB" 16(%2), %%mm0 \n\t" | |
138 PAVGB" 24(%2), %%mm1 \n\t" | |
139 "movq %%mm0, (%3) \n\t" | |
140 "add %5, %3 \n\t" | |
141 "movq %%mm1, (%3) \n\t" | |
142 "add %5, %3 \n\t" | |
143 "add $32, %2 \n\t" | |
144 "subl $4, %0 \n\t" | |
145 "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
|
146 #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used |
2979 | 147 :"+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
|
148 #else |
2979 | 149 :"+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
|
150 #endif |
2979 | 151 :"S"((long)src1Stride), "D"((long)dstStride) |
152 :"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
|
153 //the following should be used, though better not with gcc ... |
2979 | 154 /* :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst) |
155 :"r"(src1Stride), "r"(dstStride) | |
156 :"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
|
157 } |
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 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
|
160 { |
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
|
161 __asm __volatile( |
2979 | 162 "pcmpeqb %%mm6, %%mm6 \n\t" |
163 "testl $1, %0 \n\t" | |
164 " jz 1f \n\t" | |
165 "movq (%1), %%mm0 \n\t" | |
166 "movq (%2), %%mm1 \n\t" | |
167 "add %4, %1 \n\t" | |
168 "add $8, %2 \n\t" | |
169 "pxor %%mm6, %%mm0 \n\t" | |
170 "pxor %%mm6, %%mm1 \n\t" | |
171 PAVGB" %%mm1, %%mm0 \n\t" | |
172 "pxor %%mm6, %%mm0 \n\t" | |
173 "movq %%mm0, (%3) \n\t" | |
174 "add %5, %3 \n\t" | |
175 "decl %0 \n\t" | |
176 "1: \n\t" | |
177 "movq (%1), %%mm0 \n\t" | |
178 "add %4, %1 \n\t" | |
179 "movq (%1), %%mm1 \n\t" | |
180 "add %4, %1 \n\t" | |
181 "movq (%2), %%mm2 \n\t" | |
182 "movq 8(%2), %%mm3 \n\t" | |
183 "pxor %%mm6, %%mm0 \n\t" | |
184 "pxor %%mm6, %%mm1 \n\t" | |
185 "pxor %%mm6, %%mm2 \n\t" | |
186 "pxor %%mm6, %%mm3 \n\t" | |
187 PAVGB" %%mm2, %%mm0 \n\t" | |
188 PAVGB" %%mm3, %%mm1 \n\t" | |
189 "pxor %%mm6, %%mm0 \n\t" | |
190 "pxor %%mm6, %%mm1 \n\t" | |
191 "movq %%mm0, (%3) \n\t" | |
192 "add %5, %3 \n\t" | |
193 "movq %%mm1, (%3) \n\t" | |
194 "add %5, %3 \n\t" | |
195 "movq (%1), %%mm0 \n\t" | |
196 "add %4, %1 \n\t" | |
197 "movq (%1), %%mm1 \n\t" | |
198 "add %4, %1 \n\t" | |
199 "movq 16(%2), %%mm2 \n\t" | |
200 "movq 24(%2), %%mm3 \n\t" | |
201 "pxor %%mm6, %%mm0 \n\t" | |
202 "pxor %%mm6, %%mm1 \n\t" | |
203 "pxor %%mm6, %%mm2 \n\t" | |
204 "pxor %%mm6, %%mm3 \n\t" | |
205 PAVGB" %%mm2, %%mm0 \n\t" | |
206 PAVGB" %%mm3, %%mm1 \n\t" | |
207 "pxor %%mm6, %%mm0 \n\t" | |
208 "pxor %%mm6, %%mm1 \n\t" | |
209 "movq %%mm0, (%3) \n\t" | |
210 "add %5, %3 \n\t" | |
211 "movq %%mm1, (%3) \n\t" | |
212 "add %5, %3 \n\t" | |
213 "add $32, %2 \n\t" | |
214 "subl $4, %0 \n\t" | |
215 "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
|
216 #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used |
2979 | 217 :"+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
|
218 #else |
2979 | 219 :"+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
|
220 #endif |
2979 | 221 :"S"((long)src1Stride), "D"((long)dstStride) |
222 :"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
|
223 //the following should be used, though better not with gcc ... |
2979 | 224 /* :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst) |
225 :"r"(src1Stride), "r"(dstStride) | |
226 :"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
|
227 } |
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
|
228 |
2209 | 229 static void DEF(avg_pixels4_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h) |
230 { | |
231 __asm __volatile( | |
2979 | 232 "testl $1, %0 \n\t" |
233 " jz 1f \n\t" | |
234 "movd (%1), %%mm0 \n\t" | |
235 "movd (%2), %%mm1 \n\t" | |
236 "add %4, %1 \n\t" | |
237 "add $4, %2 \n\t" | |
238 PAVGB" %%mm1, %%mm0 \n\t" | |
239 PAVGB" (%3), %%mm0 \n\t" | |
240 "movd %%mm0, (%3) \n\t" | |
241 "add %5, %3 \n\t" | |
242 "decl %0 \n\t" | |
243 "1: \n\t" | |
244 "movd (%1), %%mm0 \n\t" | |
245 "add %4, %1 \n\t" | |
246 "movd (%1), %%mm1 \n\t" | |
247 "add %4, %1 \n\t" | |
248 PAVGB" (%2), %%mm0 \n\t" | |
249 PAVGB" 4(%2), %%mm1 \n\t" | |
250 PAVGB" (%3), %%mm0 \n\t" | |
251 "movd %%mm0, (%3) \n\t" | |
252 "add %5, %3 \n\t" | |
253 PAVGB" (%3), %%mm1 \n\t" | |
254 "movd %%mm1, (%3) \n\t" | |
255 "add %5, %3 \n\t" | |
256 "movd (%1), %%mm0 \n\t" | |
257 "add %4, %1 \n\t" | |
258 "movd (%1), %%mm1 \n\t" | |
259 "add %4, %1 \n\t" | |
260 PAVGB" 8(%2), %%mm0 \n\t" | |
261 PAVGB" 12(%2), %%mm1 \n\t" | |
262 PAVGB" (%3), %%mm0 \n\t" | |
263 "movd %%mm0, (%3) \n\t" | |
264 "add %5, %3 \n\t" | |
265 PAVGB" (%3), %%mm1 \n\t" | |
266 "movd %%mm1, (%3) \n\t" | |
267 "add %5, %3 \n\t" | |
268 "add $16, %2 \n\t" | |
269 "subl $4, %0 \n\t" | |
270 "jnz 1b \n\t" | |
2209 | 271 #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used |
2979 | 272 :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst) |
2209 | 273 #else |
2979 | 274 :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst) |
2209 | 275 #endif |
2979 | 276 :"S"((long)src1Stride), "D"((long)dstStride) |
277 :"memory"); | |
2209 | 278 } |
279 | |
280 | |
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
|
281 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
|
282 { |
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
|
283 __asm __volatile( |
2979 | 284 "testl $1, %0 \n\t" |
285 " jz 1f \n\t" | |
286 "movq (%1), %%mm0 \n\t" | |
287 "movq (%2), %%mm1 \n\t" | |
288 "add %4, %1 \n\t" | |
289 "add $8, %2 \n\t" | |
290 PAVGB" %%mm1, %%mm0 \n\t" | |
291 PAVGB" (%3), %%mm0 \n\t" | |
292 "movq %%mm0, (%3) \n\t" | |
293 "add %5, %3 \n\t" | |
294 "decl %0 \n\t" | |
295 "1: \n\t" | |
296 "movq (%1), %%mm0 \n\t" | |
297 "add %4, %1 \n\t" | |
298 "movq (%1), %%mm1 \n\t" | |
299 "add %4, %1 \n\t" | |
300 PAVGB" (%2), %%mm0 \n\t" | |
301 PAVGB" 8(%2), %%mm1 \n\t" | |
302 PAVGB" (%3), %%mm0 \n\t" | |
303 "movq %%mm0, (%3) \n\t" | |
304 "add %5, %3 \n\t" | |
305 PAVGB" (%3), %%mm1 \n\t" | |
306 "movq %%mm1, (%3) \n\t" | |
307 "add %5, %3 \n\t" | |
308 "movq (%1), %%mm0 \n\t" | |
309 "add %4, %1 \n\t" | |
310 "movq (%1), %%mm1 \n\t" | |
311 "add %4, %1 \n\t" | |
312 PAVGB" 16(%2), %%mm0 \n\t" | |
313 PAVGB" 24(%2), %%mm1 \n\t" | |
314 PAVGB" (%3), %%mm0 \n\t" | |
315 "movq %%mm0, (%3) \n\t" | |
316 "add %5, %3 \n\t" | |
317 PAVGB" (%3), %%mm1 \n\t" | |
318 "movq %%mm1, (%3) \n\t" | |
319 "add %5, %3 \n\t" | |
320 "add $32, %2 \n\t" | |
321 "subl $4, %0 \n\t" | |
322 "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
|
323 #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used |
2979 | 324 :"+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
|
325 #else |
2979 | 326 :"+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
|
327 #endif |
2979 | 328 :"S"((long)src1Stride), "D"((long)dstStride) |
329 :"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
|
330 //the following should be used, though better not with gcc ... |
2979 | 331 /* :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst) |
332 :"r"(src1Stride), "r"(dstStride) | |
333 :"memory");*/ | |
954 | 334 } |
335 | |
1064 | 336 static void DEF(put_pixels16_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
651 | 337 { |
338 __asm __volatile( | |
2979 | 339 "lea (%3, %3), %%"REG_a" \n\t" |
340 "1: \n\t" | |
341 "movq (%1), %%mm0 \n\t" | |
342 "movq (%1, %3), %%mm1 \n\t" | |
343 "movq 8(%1), %%mm2 \n\t" | |
344 "movq 8(%1, %3), %%mm3 \n\t" | |
345 PAVGB" 1(%1), %%mm0 \n\t" | |
346 PAVGB" 1(%1, %3), %%mm1 \n\t" | |
347 PAVGB" 9(%1), %%mm2 \n\t" | |
348 PAVGB" 9(%1, %3), %%mm3 \n\t" | |
349 "movq %%mm0, (%2) \n\t" | |
350 "movq %%mm1, (%2, %3) \n\t" | |
351 "movq %%mm2, 8(%2) \n\t" | |
352 "movq %%mm3, 8(%2, %3) \n\t" | |
353 "add %%"REG_a", %1 \n\t" | |
354 "add %%"REG_a", %2 \n\t" | |
355 "movq (%1), %%mm0 \n\t" | |
356 "movq (%1, %3), %%mm1 \n\t" | |
357 "movq 8(%1), %%mm2 \n\t" | |
358 "movq 8(%1, %3), %%mm3 \n\t" | |
359 PAVGB" 1(%1), %%mm0 \n\t" | |
360 PAVGB" 1(%1, %3), %%mm1 \n\t" | |
361 PAVGB" 9(%1), %%mm2 \n\t" | |
362 PAVGB" 9(%1, %3), %%mm3 \n\t" | |
363 "add %%"REG_a", %1 \n\t" | |
364 "movq %%mm0, (%2) \n\t" | |
365 "movq %%mm1, (%2, %3) \n\t" | |
366 "movq %%mm2, 8(%2) \n\t" | |
367 "movq %%mm3, 8(%2, %3) \n\t" | |
368 "add %%"REG_a", %2 \n\t" | |
369 "subl $4, %0 \n\t" | |
370 "jnz 1b \n\t" | |
371 :"+g"(h), "+S"(pixels), "+D"(block) | |
372 :"r" ((long)line_size) | |
373 :"%"REG_a, "memory"); | |
651 | 374 } |
954 | 375 |
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
|
376 static void DEF(put_pixels16_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h) |
954 | 377 { |
378 __asm __volatile( | |
2979 | 379 "testl $1, %0 \n\t" |
380 " jz 1f \n\t" | |
381 "movq (%1), %%mm0 \n\t" | |
382 "movq 8(%1), %%mm1 \n\t" | |
383 PAVGB" (%2), %%mm0 \n\t" | |
384 PAVGB" 8(%2), %%mm1 \n\t" | |
385 "add %4, %1 \n\t" | |
386 "add $16, %2 \n\t" | |
387 "movq %%mm0, (%3) \n\t" | |
388 "movq %%mm1, 8(%3) \n\t" | |
389 "add %5, %3 \n\t" | |
390 "decl %0 \n\t" | |
391 "1: \n\t" | |
392 "movq (%1), %%mm0 \n\t" | |
393 "movq 8(%1), %%mm1 \n\t" | |
394 "add %4, %1 \n\t" | |
395 PAVGB" (%2), %%mm0 \n\t" | |
396 PAVGB" 8(%2), %%mm1 \n\t" | |
397 "movq %%mm0, (%3) \n\t" | |
398 "movq %%mm1, 8(%3) \n\t" | |
399 "add %5, %3 \n\t" | |
400 "movq (%1), %%mm0 \n\t" | |
401 "movq 8(%1), %%mm1 \n\t" | |
402 "add %4, %1 \n\t" | |
403 PAVGB" 16(%2), %%mm0 \n\t" | |
404 PAVGB" 24(%2), %%mm1 \n\t" | |
405 "movq %%mm0, (%3) \n\t" | |
406 "movq %%mm1, 8(%3) \n\t" | |
407 "add %5, %3 \n\t" | |
408 "add $32, %2 \n\t" | |
409 "subl $2, %0 \n\t" | |
410 "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
|
411 #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used |
2979 | 412 :"+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
|
413 #else |
2979 | 414 :"+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
|
415 #endif |
2979 | 416 :"S"((long)src1Stride), "D"((long)dstStride) |
417 :"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
|
418 //the following should be used, though better not with gcc ... |
2979 | 419 /* :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst) |
420 :"r"(src1Stride), "r"(dstStride) | |
421 :"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
|
422 } |
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 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
|
425 { |
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
|
426 __asm __volatile( |
2979 | 427 "testl $1, %0 \n\t" |
428 " jz 1f \n\t" | |
429 "movq (%1), %%mm0 \n\t" | |
430 "movq 8(%1), %%mm1 \n\t" | |
431 PAVGB" (%2), %%mm0 \n\t" | |
432 PAVGB" 8(%2), %%mm1 \n\t" | |
433 "add %4, %1 \n\t" | |
434 "add $16, %2 \n\t" | |
435 PAVGB" (%3), %%mm0 \n\t" | |
436 PAVGB" 8(%3), %%mm1 \n\t" | |
437 "movq %%mm0, (%3) \n\t" | |
438 "movq %%mm1, 8(%3) \n\t" | |
439 "add %5, %3 \n\t" | |
440 "decl %0 \n\t" | |
441 "1: \n\t" | |
442 "movq (%1), %%mm0 \n\t" | |
443 "movq 8(%1), %%mm1 \n\t" | |
444 "add %4, %1 \n\t" | |
445 PAVGB" (%2), %%mm0 \n\t" | |
446 PAVGB" 8(%2), %%mm1 \n\t" | |
447 PAVGB" (%3), %%mm0 \n\t" | |
448 PAVGB" 8(%3), %%mm1 \n\t" | |
449 "movq %%mm0, (%3) \n\t" | |
450 "movq %%mm1, 8(%3) \n\t" | |
451 "add %5, %3 \n\t" | |
452 "movq (%1), %%mm0 \n\t" | |
453 "movq 8(%1), %%mm1 \n\t" | |
454 "add %4, %1 \n\t" | |
455 PAVGB" 16(%2), %%mm0 \n\t" | |
456 PAVGB" 24(%2), %%mm1 \n\t" | |
457 PAVGB" (%3), %%mm0 \n\t" | |
458 PAVGB" 8(%3), %%mm1 \n\t" | |
459 "movq %%mm0, (%3) \n\t" | |
460 "movq %%mm1, 8(%3) \n\t" | |
461 "add %5, %3 \n\t" | |
462 "add $32, %2 \n\t" | |
463 "subl $2, %0 \n\t" | |
464 "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
|
465 #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used |
2979 | 466 :"+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
|
467 #else |
2979 | 468 :"+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
|
469 #endif |
2979 | 470 :"S"((long)src1Stride), "D"((long)dstStride) |
471 :"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
|
472 //the following should be used, though better not with gcc ... |
2979 | 473 /* :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst) |
474 :"r"(src1Stride), "r"(dstStride) | |
475 :"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
|
476 } |
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 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
|
479 { |
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
|
480 __asm __volatile( |
2979 | 481 "pcmpeqb %%mm6, %%mm6 \n\t" |
482 "testl $1, %0 \n\t" | |
483 " jz 1f \n\t" | |
484 "movq (%1), %%mm0 \n\t" | |
485 "movq 8(%1), %%mm1 \n\t" | |
486 "movq (%2), %%mm2 \n\t" | |
487 "movq 8(%2), %%mm3 \n\t" | |
488 "pxor %%mm6, %%mm0 \n\t" | |
489 "pxor %%mm6, %%mm1 \n\t" | |
490 "pxor %%mm6, %%mm2 \n\t" | |
491 "pxor %%mm6, %%mm3 \n\t" | |
492 PAVGB" %%mm2, %%mm0 \n\t" | |
493 PAVGB" %%mm3, %%mm1 \n\t" | |
494 "pxor %%mm6, %%mm0 \n\t" | |
495 "pxor %%mm6, %%mm1 \n\t" | |
496 "add %4, %1 \n\t" | |
497 "add $16, %2 \n\t" | |
498 "movq %%mm0, (%3) \n\t" | |
499 "movq %%mm1, 8(%3) \n\t" | |
500 "add %5, %3 \n\t" | |
501 "decl %0 \n\t" | |
502 "1: \n\t" | |
503 "movq (%1), %%mm0 \n\t" | |
504 "movq 8(%1), %%mm1 \n\t" | |
505 "add %4, %1 \n\t" | |
506 "movq (%2), %%mm2 \n\t" | |
507 "movq 8(%2), %%mm3 \n\t" | |
508 "pxor %%mm6, %%mm0 \n\t" | |
509 "pxor %%mm6, %%mm1 \n\t" | |
510 "pxor %%mm6, %%mm2 \n\t" | |
511 "pxor %%mm6, %%mm3 \n\t" | |
512 PAVGB" %%mm2, %%mm0 \n\t" | |
513 PAVGB" %%mm3, %%mm1 \n\t" | |
514 "pxor %%mm6, %%mm0 \n\t" | |
515 "pxor %%mm6, %%mm1 \n\t" | |
516 "movq %%mm0, (%3) \n\t" | |
517 "movq %%mm1, 8(%3) \n\t" | |
518 "add %5, %3 \n\t" | |
519 "movq (%1), %%mm0 \n\t" | |
520 "movq 8(%1), %%mm1 \n\t" | |
521 "add %4, %1 \n\t" | |
522 "movq 16(%2), %%mm2 \n\t" | |
523 "movq 24(%2), %%mm3 \n\t" | |
524 "pxor %%mm6, %%mm0 \n\t" | |
525 "pxor %%mm6, %%mm1 \n\t" | |
526 "pxor %%mm6, %%mm2 \n\t" | |
527 "pxor %%mm6, %%mm3 \n\t" | |
528 PAVGB" %%mm2, %%mm0 \n\t" | |
529 PAVGB" %%mm3, %%mm1 \n\t" | |
530 "pxor %%mm6, %%mm0 \n\t" | |
531 "pxor %%mm6, %%mm1 \n\t" | |
532 "movq %%mm0, (%3) \n\t" | |
533 "movq %%mm1, 8(%3) \n\t" | |
534 "add %5, %3 \n\t" | |
535 "add $32, %2 \n\t" | |
536 "subl $2, %0 \n\t" | |
537 "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
|
538 #ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used |
2979 | 539 :"+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
|
540 #else |
2979 | 541 :"+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
|
542 #endif |
2979 | 543 :"S"((long)src1Stride), "D"((long)dstStride) |
544 :"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
|
545 //the following should be used, though better not with gcc ... |
2979 | 546 /* :"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst) |
547 :"r"(src1Stride), "r"(dstStride) | |
548 :"memory");*/ | |
954 | 549 } |
2967 | 550 |
441 | 551 /* GL: this function does incorrect rounding if overflow */ |
1064 | 552 static void DEF(put_no_rnd_pixels8_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
441 | 553 { |
448 | 554 MOVQ_BONE(mm6); |
441 | 555 __asm __volatile( |
2979 | 556 "lea (%3, %3), %%"REG_a" \n\t" |
557 "1: \n\t" | |
558 "movq (%1), %%mm0 \n\t" | |
559 "movq (%1, %3), %%mm2 \n\t" | |
560 "movq 1(%1), %%mm1 \n\t" | |
561 "movq 1(%1, %3), %%mm3 \n\t" | |
562 "add %%"REG_a", %1 \n\t" | |
563 "psubusb %%mm6, %%mm0 \n\t" | |
564 "psubusb %%mm6, %%mm2 \n\t" | |
565 PAVGB" %%mm1, %%mm0 \n\t" | |
566 PAVGB" %%mm3, %%mm2 \n\t" | |
567 "movq %%mm0, (%2) \n\t" | |
568 "movq %%mm2, (%2, %3) \n\t" | |
569 "movq (%1), %%mm0 \n\t" | |
570 "movq 1(%1), %%mm1 \n\t" | |
571 "movq (%1, %3), %%mm2 \n\t" | |
572 "movq 1(%1, %3), %%mm3 \n\t" | |
573 "add %%"REG_a", %2 \n\t" | |
574 "add %%"REG_a", %1 \n\t" | |
575 "psubusb %%mm6, %%mm0 \n\t" | |
576 "psubusb %%mm6, %%mm2 \n\t" | |
577 PAVGB" %%mm1, %%mm0 \n\t" | |
578 PAVGB" %%mm3, %%mm2 \n\t" | |
579 "movq %%mm0, (%2) \n\t" | |
580 "movq %%mm2, (%2, %3) \n\t" | |
581 "add %%"REG_a", %2 \n\t" | |
582 "subl $4, %0 \n\t" | |
583 "jnz 1b \n\t" | |
584 :"+g"(h), "+S"(pixels), "+D"(block) | |
585 :"r" ((long)line_size) | |
586 :"%"REG_a, "memory"); | |
0 | 587 } |
588 | |
1064 | 589 static void DEF(put_pixels8_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
0 | 590 { |
413
1548abb7bbed
* fix for -fPIC compilation - compiles with 2.95.2 as well - any
kabi
parents:
402
diff
changeset
|
591 __asm __volatile( |
2979 | 592 "lea (%3, %3), %%"REG_a" \n\t" |
593 "movq (%1), %%mm0 \n\t" | |
594 "sub %3, %2 \n\t" | |
595 "1: \n\t" | |
596 "movq (%1, %3), %%mm1 \n\t" | |
597 "movq (%1, %%"REG_a"), %%mm2 \n\t" | |
598 "add %%"REG_a", %1 \n\t" | |
599 PAVGB" %%mm1, %%mm0 \n\t" | |
600 PAVGB" %%mm2, %%mm1 \n\t" | |
601 "movq %%mm0, (%2, %3) \n\t" | |
602 "movq %%mm1, (%2, %%"REG_a") \n\t" | |
603 "movq (%1, %3), %%mm1 \n\t" | |
604 "movq (%1, %%"REG_a"), %%mm0 \n\t" | |
605 "add %%"REG_a", %2 \n\t" | |
606 "add %%"REG_a", %1 \n\t" | |
607 PAVGB" %%mm1, %%mm2 \n\t" | |
608 PAVGB" %%mm0, %%mm1 \n\t" | |
609 "movq %%mm2, (%2, %3) \n\t" | |
610 "movq %%mm1, (%2, %%"REG_a") \n\t" | |
611 "add %%"REG_a", %2 \n\t" | |
612 "subl $4, %0 \n\t" | |
613 "jnz 1b \n\t" | |
614 :"+g"(h), "+S"(pixels), "+D" (block) | |
615 :"r" ((long)line_size) | |
616 :"%"REG_a, "memory"); | |
386 | 617 } |
618 | |
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
|
619 /* GL: this function does incorrect rounding if overflow */ |
1064 | 620 static void DEF(put_no_rnd_pixels8_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
386 | 621 { |
448 | 622 MOVQ_BONE(mm6); |
0 | 623 __asm __volatile( |
2979 | 624 "lea (%3, %3), %%"REG_a" \n\t" |
625 "movq (%1), %%mm0 \n\t" | |
626 "sub %3, %2 \n\t" | |
627 "1: \n\t" | |
628 "movq (%1, %3), %%mm1 \n\t" | |
629 "movq (%1, %%"REG_a"), %%mm2 \n\t" | |
630 "add %%"REG_a", %1 \n\t" | |
631 "psubusb %%mm6, %%mm1 \n\t" | |
632 PAVGB" %%mm1, %%mm0 \n\t" | |
633 PAVGB" %%mm2, %%mm1 \n\t" | |
634 "movq %%mm0, (%2, %3) \n\t" | |
635 "movq %%mm1, (%2, %%"REG_a") \n\t" | |
636 "movq (%1, %3), %%mm1 \n\t" | |
637 "movq (%1, %%"REG_a"), %%mm0 \n\t" | |
638 "add %%"REG_a", %2 \n\t" | |
639 "add %%"REG_a", %1 \n\t" | |
640 "psubusb %%mm6, %%mm1 \n\t" | |
641 PAVGB" %%mm1, %%mm2 \n\t" | |
642 PAVGB" %%mm0, %%mm1 \n\t" | |
643 "movq %%mm2, (%2, %3) \n\t" | |
644 "movq %%mm1, (%2, %%"REG_a") \n\t" | |
645 "add %%"REG_a", %2 \n\t" | |
646 "subl $4, %0 \n\t" | |
647 "jnz 1b \n\t" | |
648 :"+g"(h), "+S"(pixels), "+D" (block) | |
649 :"r" ((long)line_size) | |
650 :"%"REG_a, "memory"); | |
0 | 651 } |
652 | |
1064 | 653 static void DEF(avg_pixels8)(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
0 | 654 { |
655 __asm __volatile( | |
2979 | 656 "lea (%3, %3), %%"REG_a" \n\t" |
657 "1: \n\t" | |
658 "movq (%2), %%mm0 \n\t" | |
659 "movq (%2, %3), %%mm1 \n\t" | |
660 PAVGB" (%1), %%mm0 \n\t" | |
661 PAVGB" (%1, %3), %%mm1 \n\t" | |
662 "movq %%mm0, (%2) \n\t" | |
663 "movq %%mm1, (%2, %3) \n\t" | |
664 "add %%"REG_a", %1 \n\t" | |
665 "add %%"REG_a", %2 \n\t" | |
666 "movq (%2), %%mm0 \n\t" | |
667 "movq (%2, %3), %%mm1 \n\t" | |
668 PAVGB" (%1), %%mm0 \n\t" | |
669 PAVGB" (%1, %3), %%mm1 \n\t" | |
670 "add %%"REG_a", %1 \n\t" | |
671 "movq %%mm0, (%2) \n\t" | |
672 "movq %%mm1, (%2, %3) \n\t" | |
673 "add %%"REG_a", %2 \n\t" | |
674 "subl $4, %0 \n\t" | |
675 "jnz 1b \n\t" | |
676 :"+g"(h), "+S"(pixels), "+D"(block) | |
677 :"r" ((long)line_size) | |
678 :"%"REG_a, "memory"); | |
0 | 679 } |
680 | |
1064 | 681 static void DEF(avg_pixels8_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
0 | 682 { |
683 __asm __volatile( | |
2979 | 684 "lea (%3, %3), %%"REG_a" \n\t" |
685 "1: \n\t" | |
686 "movq (%1), %%mm0 \n\t" | |
687 "movq (%1, %3), %%mm2 \n\t" | |
688 PAVGB" 1(%1), %%mm0 \n\t" | |
689 PAVGB" 1(%1, %3), %%mm2 \n\t" | |
690 PAVGB" (%2), %%mm0 \n\t" | |
691 PAVGB" (%2, %3), %%mm2 \n\t" | |
692 "add %%"REG_a", %1 \n\t" | |
693 "movq %%mm0, (%2) \n\t" | |
694 "movq %%mm2, (%2, %3) \n\t" | |
695 "movq (%1), %%mm0 \n\t" | |
696 "movq (%1, %3), %%mm2 \n\t" | |
697 PAVGB" 1(%1), %%mm0 \n\t" | |
698 PAVGB" 1(%1, %3), %%mm2 \n\t" | |
699 "add %%"REG_a", %2 \n\t" | |
700 "add %%"REG_a", %1 \n\t" | |
701 PAVGB" (%2), %%mm0 \n\t" | |
702 PAVGB" (%2, %3), %%mm2 \n\t" | |
703 "movq %%mm0, (%2) \n\t" | |
704 "movq %%mm2, (%2, %3) \n\t" | |
705 "add %%"REG_a", %2 \n\t" | |
706 "subl $4, %0 \n\t" | |
707 "jnz 1b \n\t" | |
708 :"+g"(h), "+S"(pixels), "+D"(block) | |
709 :"r" ((long)line_size) | |
710 :"%"REG_a, "memory"); | |
0 | 711 } |
712 | |
1064 | 713 static void DEF(avg_pixels8_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
0 | 714 { |
715 __asm __volatile( | |
2979 | 716 "lea (%3, %3), %%"REG_a" \n\t" |
717 "movq (%1), %%mm0 \n\t" | |
718 "sub %3, %2 \n\t" | |
719 "1: \n\t" | |
720 "movq (%1, %3), %%mm1 \n\t" | |
721 "movq (%1, %%"REG_a"), %%mm2 \n\t" | |
722 "add %%"REG_a", %1 \n\t" | |
723 PAVGB" %%mm1, %%mm0 \n\t" | |
724 PAVGB" %%mm2, %%mm1 \n\t" | |
725 "movq (%2, %3), %%mm3 \n\t" | |
726 "movq (%2, %%"REG_a"), %%mm4 \n\t" | |
727 PAVGB" %%mm3, %%mm0 \n\t" | |
728 PAVGB" %%mm4, %%mm1 \n\t" | |
729 "movq %%mm0, (%2, %3) \n\t" | |
730 "movq %%mm1, (%2, %%"REG_a") \n\t" | |
731 "movq (%1, %3), %%mm1 \n\t" | |
732 "movq (%1, %%"REG_a"), %%mm0 \n\t" | |
733 PAVGB" %%mm1, %%mm2 \n\t" | |
734 PAVGB" %%mm0, %%mm1 \n\t" | |
735 "add %%"REG_a", %2 \n\t" | |
736 "add %%"REG_a", %1 \n\t" | |
737 "movq (%2, %3), %%mm3 \n\t" | |
738 "movq (%2, %%"REG_a"), %%mm4 \n\t" | |
739 PAVGB" %%mm3, %%mm2 \n\t" | |
740 PAVGB" %%mm4, %%mm1 \n\t" | |
741 "movq %%mm2, (%2, %3) \n\t" | |
742 "movq %%mm1, (%2, %%"REG_a") \n\t" | |
743 "add %%"REG_a", %2 \n\t" | |
744 "subl $4, %0 \n\t" | |
745 "jnz 1b \n\t" | |
746 :"+g"(h), "+S"(pixels), "+D"(block) | |
747 :"r" ((long)line_size) | |
748 :"%"REG_a, "memory"); | |
0 | 749 } |
750 | |
2967 | 751 // Note this is not correctly rounded, but this function is only used for b frames so it doesnt matter |
1064 | 752 static void DEF(avg_pixels8_xy2)(uint8_t *block, const uint8_t *pixels, int line_size, int h) |
0 | 753 { |
448 | 754 MOVQ_BONE(mm6); |
0 | 755 __asm __volatile( |
2979 | 756 "lea (%3, %3), %%"REG_a" \n\t" |
757 "movq (%1), %%mm0 \n\t" | |
758 PAVGB" 1(%1), %%mm0 \n\t" | |
3576
f7125bf10892
Support for MacIntel, last part: balign directives
gpoirier
parents:
3036
diff
changeset
|
759 ASMALIGN(3) |
2979 | 760 "1: \n\t" |
761 "movq (%1, %%"REG_a"), %%mm2 \n\t" | |
762 "movq (%1, %3), %%mm1 \n\t" | |
763 "psubusb %%mm6, %%mm2 \n\t" | |
764 PAVGB" 1(%1, %3), %%mm1 \n\t" | |
765 PAVGB" 1(%1, %%"REG_a"), %%mm2 \n\t" | |
766 "add %%"REG_a", %1 \n\t" | |
767 PAVGB" %%mm1, %%mm0 \n\t" | |
768 PAVGB" %%mm2, %%mm1 \n\t" | |
769 PAVGB" (%2), %%mm0 \n\t" | |
770 PAVGB" (%2, %3), %%mm1 \n\t" | |
771 "movq %%mm0, (%2) \n\t" | |
772 "movq %%mm1, (%2, %3) \n\t" | |
773 "movq (%1, %3), %%mm1 \n\t" | |
774 "movq (%1, %%"REG_a"), %%mm0 \n\t" | |
775 PAVGB" 1(%1, %3), %%mm1 \n\t" | |
776 PAVGB" 1(%1, %%"REG_a"), %%mm0 \n\t" | |
777 "add %%"REG_a", %2 \n\t" | |
778 "add %%"REG_a", %1 \n\t" | |
779 PAVGB" %%mm1, %%mm2 \n\t" | |
780 PAVGB" %%mm0, %%mm1 \n\t" | |
781 PAVGB" (%2), %%mm2 \n\t" | |
782 PAVGB" (%2, %3), %%mm1 \n\t" | |
783 "movq %%mm2, (%2) \n\t" | |
784 "movq %%mm1, (%2, %3) \n\t" | |
785 "add %%"REG_a", %2 \n\t" | |
786 "subl $4, %0 \n\t" | |
787 "jnz 1b \n\t" | |
788 :"+g"(h), "+S"(pixels), "+D"(block) | |
789 :"r" ((long)line_size) | |
790 :"%"REG_a, "memory"); | |
0 | 791 } |
651 | 792 |
793 //FIXME the following could be optimized too ... | |
1064 | 794 static void DEF(put_no_rnd_pixels16_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){ |
651 | 795 DEF(put_no_rnd_pixels8_x2)(block , pixels , line_size, h); |
796 DEF(put_no_rnd_pixels8_x2)(block+8, pixels+8, line_size, h); | |
797 } | |
1064 | 798 static void DEF(put_pixels16_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){ |
651 | 799 DEF(put_pixels8_y2)(block , pixels , line_size, h); |
800 DEF(put_pixels8_y2)(block+8, pixels+8, line_size, h); | |
801 } | |
1064 | 802 static void DEF(put_no_rnd_pixels16_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){ |
651 | 803 DEF(put_no_rnd_pixels8_y2)(block , pixels , line_size, h); |
804 DEF(put_no_rnd_pixels8_y2)(block+8, pixels+8, line_size, h); | |
805 } | |
1064 | 806 static void DEF(avg_pixels16)(uint8_t *block, const uint8_t *pixels, int line_size, int h){ |
651 | 807 DEF(avg_pixels8)(block , pixels , line_size, h); |
808 DEF(avg_pixels8)(block+8, pixels+8, line_size, h); | |
809 } | |
1064 | 810 static void DEF(avg_pixels16_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){ |
651 | 811 DEF(avg_pixels8_x2)(block , pixels , line_size, h); |
812 DEF(avg_pixels8_x2)(block+8, pixels+8, line_size, h); | |
813 } | |
1064 | 814 static void DEF(avg_pixels16_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){ |
651 | 815 DEF(avg_pixels8_y2)(block , pixels , line_size, h); |
816 DEF(avg_pixels8_y2)(block+8, pixels+8, line_size, h); | |
817 } | |
1064 | 818 static void DEF(avg_pixels16_xy2)(uint8_t *block, const uint8_t *pixels, int line_size, int h){ |
651 | 819 DEF(avg_pixels8_xy2)(block , pixels , line_size, h); |
820 DEF(avg_pixels8_xy2)(block+8, pixels+8, line_size, h); | |
821 } | |
822 | |
3807
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
823 #define QPEL_2TAP_L3(OPNAME) \ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
824 static void DEF(OPNAME ## 2tap_qpel16_l3)(uint8_t *dst, uint8_t *src, int stride, int h, int off1, int off2){\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
825 asm volatile(\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
826 "1: \n\t"\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
827 "movq (%1,%2), %%mm0 \n\t"\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
828 "movq 8(%1,%2), %%mm1 \n\t"\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
829 PAVGB" (%1,%3), %%mm0 \n\t"\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
830 PAVGB" 8(%1,%3), %%mm1 \n\t"\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
831 PAVGB" (%1), %%mm0 \n\t"\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
832 PAVGB" 8(%1), %%mm1 \n\t"\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
833 STORE_OP( (%1,%4),%%mm0)\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
834 STORE_OP(8(%1,%4),%%mm1)\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
835 "movq %%mm0, (%1,%4) \n\t"\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
836 "movq %%mm1, 8(%1,%4) \n\t"\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
837 "add %5, %1 \n\t"\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
838 "decl %0 \n\t"\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
839 "jnz 1b \n\t"\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
840 :"+g"(h), "+r"(src)\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
841 :"r"((long)off1), "r"((long)off2),\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
842 "r"((long)(dst-src)), "r"((long)stride)\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
843 :"memory"\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
844 );\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
845 }\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
846 static void DEF(OPNAME ## 2tap_qpel8_l3)(uint8_t *dst, uint8_t *src, int stride, int h, int off1, int off2){\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
847 asm volatile(\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
848 "1: \n\t"\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
849 "movq (%1,%2), %%mm0 \n\t"\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
850 PAVGB" (%1,%3), %%mm0 \n\t"\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
851 PAVGB" (%1), %%mm0 \n\t"\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
852 STORE_OP((%1,%4),%%mm0)\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
853 "movq %%mm0, (%1,%4) \n\t"\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
854 "add %5, %1 \n\t"\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
855 "decl %0 \n\t"\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
856 "jnz 1b \n\t"\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
857 :"+g"(h), "+r"(src)\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
858 :"r"((long)off1), "r"((long)off2),\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
859 "r"((long)(dst-src)), "r"((long)stride)\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
860 :"memory"\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
861 );\ |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
862 } |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
863 |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
864 #define STORE_OP(a,b) PAVGB" "#a","#b" \n\t" |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
865 QPEL_2TAP_L3(avg_) |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
866 #undef STORE_OP |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
867 #define STORE_OP(a,b) |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
868 QPEL_2TAP_L3(put_) |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
869 #undef STORE_OP |
6a40092eb9e6
approximate qpel functions: sacrifice some quality for some decoding speed. enabled on B-frames with -lavdopts fast.
lorenm
parents:
3576
diff
changeset
|
870 #undef QPEL_2TAP_L3 |