annotate libmpcodecs/vf_tfields.c @ 33444:fac64fecdb03

Automatically add "-fno-tree-vectorize" in CFLAGS. Automatic tree vectorization is enabled for high optimization levels (-O3 and above). However it have long history of breaking code. Latest been with gcc-4.6.0 and "-march=corei7-avx -O3". FFmpeg also disables it.
author iive
date Tue, 31 May 2011 20:45:39 +0000
parents 3525dc14e3dc
children 064be070d8c8
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
30421
bbb6ebec87a0 Add missing license headers to all files in the libmpcodecs directory.
diego
parents: 29639
diff changeset
1 /*
bbb6ebec87a0 Add missing license headers to all files in the libmpcodecs directory.
diego
parents: 29639
diff changeset
2 * This file is part of MPlayer.
bbb6ebec87a0 Add missing license headers to all files in the libmpcodecs directory.
diego
parents: 29639
diff changeset
3 *
bbb6ebec87a0 Add missing license headers to all files in the libmpcodecs directory.
diego
parents: 29639
diff changeset
4 * MPlayer is free software; you can redistribute it and/or modify
bbb6ebec87a0 Add missing license headers to all files in the libmpcodecs directory.
diego
parents: 29639
diff changeset
5 * it under the terms of the GNU General Public License as published by
bbb6ebec87a0 Add missing license headers to all files in the libmpcodecs directory.
diego
parents: 29639
diff changeset
6 * the Free Software Foundation; either version 2 of the License, or
bbb6ebec87a0 Add missing license headers to all files in the libmpcodecs directory.
diego
parents: 29639
diff changeset
7 * (at your option) any later version.
bbb6ebec87a0 Add missing license headers to all files in the libmpcodecs directory.
diego
parents: 29639
diff changeset
8 *
bbb6ebec87a0 Add missing license headers to all files in the libmpcodecs directory.
diego
parents: 29639
diff changeset
9 * MPlayer is distributed in the hope that it will be useful,
bbb6ebec87a0 Add missing license headers to all files in the libmpcodecs directory.
diego
parents: 29639
diff changeset
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
bbb6ebec87a0 Add missing license headers to all files in the libmpcodecs directory.
diego
parents: 29639
diff changeset
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
bbb6ebec87a0 Add missing license headers to all files in the libmpcodecs directory.
diego
parents: 29639
diff changeset
12 * GNU General Public License for more details.
bbb6ebec87a0 Add missing license headers to all files in the libmpcodecs directory.
diego
parents: 29639
diff changeset
13 *
bbb6ebec87a0 Add missing license headers to all files in the libmpcodecs directory.
diego
parents: 29639
diff changeset
14 * You should have received a copy of the GNU General Public License along
bbb6ebec87a0 Add missing license headers to all files in the libmpcodecs directory.
diego
parents: 29639
diff changeset
15 * with MPlayer; if not, write to the Free Software Foundation, Inc.,
bbb6ebec87a0 Add missing license headers to all files in the libmpcodecs directory.
diego
parents: 29639
diff changeset
16 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
bbb6ebec87a0 Add missing license headers to all files in the libmpcodecs directory.
diego
parents: 29639
diff changeset
17 */
bbb6ebec87a0 Add missing license headers to all files in the libmpcodecs directory.
diego
parents: 29639
diff changeset
18
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
19 #include <stdio.h>
08264c647f46 new filter
rfelker
parents:
diff changeset
20 #include <stdlib.h>
08264c647f46 new filter
rfelker
parents:
diff changeset
21 #include <string.h>
08264c647f46 new filter
rfelker
parents:
diff changeset
22
17012
6ff3379a0862 Unify include path handling, -I.. is in CFLAGS.
diego
parents: 15013
diff changeset
23 #include "config.h"
6ff3379a0862 Unify include path handling, -I.. is in CFLAGS.
diego
parents: 15013
diff changeset
24 #include "mp_msg.h"
6ff3379a0862 Unify include path handling, -I.. is in CFLAGS.
diego
parents: 15013
diff changeset
25 #include "cpudetect.h"
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
26
08264c647f46 new filter
rfelker
parents:
diff changeset
27 #include "img_format.h"
08264c647f46 new filter
rfelker
parents:
diff changeset
28 #include "mp_image.h"
08264c647f46 new filter
rfelker
parents:
diff changeset
29 #include "vf.h"
08264c647f46 new filter
rfelker
parents:
diff changeset
30
31978
3525dc14e3dc Add the proper include instead of declaring the correct_pts variable extern.
diego
parents: 30642
diff changeset
31 #include "libmpdemux/demuxer.h"
17012
6ff3379a0862 Unify include path handling, -I.. is in CFLAGS.
diego
parents: 15013
diff changeset
32 #include "libvo/fastmemcpy.h"
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
33
08264c647f46 new filter
rfelker
parents:
diff changeset
34 struct vf_priv_s {
08264c647f46 new filter
rfelker
parents:
diff changeset
35 int mode;
14888
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
36 int parity;
18917
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
37 int buffered_i;
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
38 mp_image_t *buffered_mpi;
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
39 double buffered_pts;
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
40 };
08264c647f46 new filter
rfelker
parents:
diff changeset
41
08264c647f46 new filter
rfelker
parents:
diff changeset
42 static void deint(unsigned char *dest, int ds, unsigned char *src, int ss, int w, int h, int field)
08264c647f46 new filter
rfelker
parents:
diff changeset
43 {
08264c647f46 new filter
rfelker
parents:
diff changeset
44 int x, y;
08264c647f46 new filter
rfelker
parents:
diff changeset
45 src += ss;
08264c647f46 new filter
rfelker
parents:
diff changeset
46 dest += ds;
30441
9c15719d8769 Fix -vf tfields=1 so it does not read out of bounds or leave parts of the
reimar
parents: 30421
diff changeset
47 h--;
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
48 if (field) {
30441
9c15719d8769 Fix -vf tfields=1 so it does not read out of bounds or leave parts of the
reimar
parents: 30421
diff changeset
49 fast_memcpy(dest - ds, src - ss, w);
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
50 src += ss;
08264c647f46 new filter
rfelker
parents:
diff changeset
51 dest += ds;
30441
9c15719d8769 Fix -vf tfields=1 so it does not read out of bounds or leave parts of the
reimar
parents: 30421
diff changeset
52 h--;
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
53 }
30441
9c15719d8769 Fix -vf tfields=1 so it does not read out of bounds or leave parts of the
reimar
parents: 30421
diff changeset
54 for (y=h/2; y > 0; y--) {
9c15719d8769 Fix -vf tfields=1 so it does not read out of bounds or leave parts of the
reimar
parents: 30421
diff changeset
55 dest[0] = src[0];
9c15719d8769 Fix -vf tfields=1 so it does not read out of bounds or leave parts of the
reimar
parents: 30421
diff changeset
56 for (x=1; x<w-1; x++) {
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
57 if (((src[x-ss] < src[x]) && (src[x+ss] < src[x])) ||
08264c647f46 new filter
rfelker
parents:
diff changeset
58 ((src[x-ss] > src[x]) && (src[x+ss] > src[x]))) {
08264c647f46 new filter
rfelker
parents:
diff changeset
59 //dest[x] = (src[x+ss] + src[x-ss])>>1;
08264c647f46 new filter
rfelker
parents:
diff changeset
60 dest[x] = ((src[x+ss]<<1) + (src[x-ss]<<1)
08264c647f46 new filter
rfelker
parents:
diff changeset
61 + src[x+ss+1] + src[x-ss+1]
08264c647f46 new filter
rfelker
parents:
diff changeset
62 + src[x+ss-1] + src[x-ss-1])>>3;
08264c647f46 new filter
rfelker
parents:
diff changeset
63 }
08264c647f46 new filter
rfelker
parents:
diff changeset
64 else dest[x] = src[x];
08264c647f46 new filter
rfelker
parents:
diff changeset
65 }
30441
9c15719d8769 Fix -vf tfields=1 so it does not read out of bounds or leave parts of the
reimar
parents: 30421
diff changeset
66 dest[w-1] = src[w-1];
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
67 dest += ds<<1;
08264c647f46 new filter
rfelker
parents:
diff changeset
68 src += ss<<1;
08264c647f46 new filter
rfelker
parents:
diff changeset
69 }
30441
9c15719d8769 Fix -vf tfields=1 so it does not read out of bounds or leave parts of the
reimar
parents: 30421
diff changeset
70 if (h & 1)
9c15719d8769 Fix -vf tfields=1 so it does not read out of bounds or leave parts of the
reimar
parents: 30421
diff changeset
71 fast_memcpy(dest, src, w);
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
72 }
08264c647f46 new filter
rfelker
parents:
diff changeset
73
28335
31287e75b5d8 HAVE_3DNOW --> HAVE_AMD3DNOW
diego
parents: 28290
diff changeset
74 #if HAVE_AMD3DNOW
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
75 static void qpel_li_3DNOW(unsigned char *d, unsigned char *s, int w, int h, int ds, int ss, int up)
10009
69f10d08c3be new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents: 9593
diff changeset
76 {
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
77 int i, j, ssd=ss;
13720
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
78 long crap1, crap2;
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
79 if (up) {
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
80 ssd = -ss;
23457
a124f3abc1ec Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents: 18986
diff changeset
81 fast_memcpy(d, s, w);
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
82 d += ds;
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
83 s += ss;
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
84 }
10009
69f10d08c3be new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents: 9593
diff changeset
85 for (i=h-1; i; i--) {
27754
08d18fe9da52 Change all occurrences of asm and __asm to __asm__, same as was done for FFmpeg.
diego
parents: 25221
diff changeset
86 __asm__ volatile(
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
87 "1: \n\t"
13720
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
88 "movq (%%"REG_S"), %%mm0 \n\t"
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
89 "movq (%%"REG_S",%%"REG_a"), %%mm1 \n\t"
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
90 "pavgusb %%mm0, %%mm1 \n\t"
13720
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
91 "add $8, %%"REG_S" \n\t"
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
92 "pavgusb %%mm0, %%mm1 \n\t"
13720
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
93 "movq %%mm1, (%%"REG_D") \n\t"
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
94 "add $8, %%"REG_D" \n\t"
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
95 "decl %%ecx \n\t"
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
96 "jnz 1b \n\t"
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
97 : "=S"(crap1), "=D"(crap2)
13720
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
98 : "c"(w>>3), "S"(s), "D"(d), "a"((long)ssd)
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
99 );
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
100 for (j=w-(w&7); j<w; j++)
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
101 d[j] = (s[j+ssd] + 3*s[j])>>2;
10009
69f10d08c3be new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents: 9593
diff changeset
102 d += ds;
69f10d08c3be new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents: 9593
diff changeset
103 s += ss;
69f10d08c3be new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents: 9593
diff changeset
104 }
23457
a124f3abc1ec Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents: 18986
diff changeset
105 if (!up) fast_memcpy(d, s, w);
27754
08d18fe9da52 Change all occurrences of asm and __asm to __asm__, same as was done for FFmpeg.
diego
parents: 25221
diff changeset
106 __asm__ volatile("emms \n\t" : : : "memory");
10009
69f10d08c3be new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents: 9593
diff changeset
107 }
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
108 #endif
10009
69f10d08c3be new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents: 9593
diff changeset
109
28290
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_...
reimar
parents: 28175
diff changeset
110 #if HAVE_MMX2
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
111 static void qpel_li_MMX2(unsigned char *d, unsigned char *s, int w, int h, int ds, int ss, int up)
10009
69f10d08c3be new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents: 9593
diff changeset
112 {
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
113 int i, j, ssd=ss;
13720
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
114 long crap1, crap2;
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
115 if (up) {
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
116 ssd = -ss;
23457
a124f3abc1ec Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents: 18986
diff changeset
117 fast_memcpy(d, s, w);
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
118 d += ds;
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
119 s += ss;
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
120 }
10009
69f10d08c3be new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents: 9593
diff changeset
121 for (i=h-1; i; i--) {
27754
08d18fe9da52 Change all occurrences of asm and __asm to __asm__, same as was done for FFmpeg.
diego
parents: 25221
diff changeset
122 __asm__ volatile(
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
123 "pxor %%mm7, %%mm7 \n\t"
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
124 "2: \n\t"
13720
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
125 "movq (%%"REG_S"), %%mm0 \n\t"
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
126 "movq (%%"REG_S",%%"REG_a"), %%mm1 \n\t"
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
127 "pavgb %%mm0, %%mm1 \n\t"
13720
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
128 "add $8, %%"REG_S" \n\t"
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
129 "pavgb %%mm0, %%mm1 \n\t"
13720
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
130 "movq %%mm1, (%%"REG_D") \n\t"
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
131 "add $8, %%"REG_D" \n\t"
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
132 "decl %%ecx \n\t"
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
133 "jnz 2b \n\t"
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
134 : "=S"(crap1), "=D"(crap2)
13720
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
135 : "c"(w>>3), "S"(s), "D"(d), "a"((long)ssd)
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
136 );
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
137 for (j=w-(w&7); j<w; j++)
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
138 d[j] = (s[j+ssd] + 3*s[j])>>2;
10009
69f10d08c3be new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents: 9593
diff changeset
139 d += ds;
69f10d08c3be new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents: 9593
diff changeset
140 s += ss;
69f10d08c3be new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents: 9593
diff changeset
141 }
23457
a124f3abc1ec Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents: 18986
diff changeset
142 if (!up) fast_memcpy(d, s, w);
27754
08d18fe9da52 Change all occurrences of asm and __asm to __asm__, same as was done for FFmpeg.
diego
parents: 25221
diff changeset
143 __asm__ volatile("emms \n\t" : : : "memory");
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
144 }
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
145 #endif
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
146
28290
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_...
reimar
parents: 28175
diff changeset
147 #if HAVE_MMX
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
148 static void qpel_li_MMX(unsigned char *d, unsigned char *s, int w, int h, int ds, int ss, int up)
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
149 {
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
150 int i, j, ssd=ss;
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
151 int crap1, crap2;
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
152 if (up) {
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
153 ssd = -ss;
23457
a124f3abc1ec Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents: 18986
diff changeset
154 fast_memcpy(d, s, w);
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
155 d += ds;
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
156 s += ss;
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
157 }
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
158 for (i=h-1; i; i--) {
27754
08d18fe9da52 Change all occurrences of asm and __asm to __asm__, same as was done for FFmpeg.
diego
parents: 25221
diff changeset
159 __asm__ volatile(
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
160 "pxor %%mm7, %%mm7 \n\t"
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
161 "3: \n\t"
13720
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
162 "movq (%%"REG_S"), %%mm0 \n\t"
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
163 "movq (%%"REG_S"), %%mm1 \n\t"
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
164 "movq (%%"REG_S",%%"REG_a"), %%mm2 \n\t"
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
165 "movq (%%"REG_S",%%"REG_a"), %%mm3 \n\t"
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
166 "add $8, %%"REG_S" \n\t"
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
167 "punpcklbw %%mm7, %%mm0 \n\t"
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
168 "punpckhbw %%mm7, %%mm1 \n\t"
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
169 "punpcklbw %%mm7, %%mm2 \n\t"
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
170 "punpckhbw %%mm7, %%mm3 \n\t"
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
171 "paddw %%mm0, %%mm2 \n\t"
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
172 "paddw %%mm1, %%mm3 \n\t"
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
173 "paddw %%mm0, %%mm2 \n\t"
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
174 "paddw %%mm1, %%mm3 \n\t"
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
175 "paddw %%mm0, %%mm2 \n\t"
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
176 "paddw %%mm1, %%mm3 \n\t"
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
177 "psrlw $2, %%mm2 \n\t"
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
178 "psrlw $2, %%mm3 \n\t"
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
179 "packsswb %%mm3, %%mm2 \n\t"
13720
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
180 "movq %%mm2, (%%"REG_D") \n\t"
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
181 "add $8, %%"REG_D" \n\t"
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
182 "decl %%ecx \n\t"
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
183 "jnz 3b \n\t"
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
184 : "=S"(crap1), "=D"(crap2)
13720
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
185 : "c"(w>>3), "S"(s), "D"(d), "a"((long)ssd)
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
186 );
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
187 for (j=w-(w&7); j<w; j++)
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
188 d[j] = (s[j+ssd] + 3*s[j])>>2;
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
189 d += ds;
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
190 s += ss;
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
191 }
23457
a124f3abc1ec Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents: 18986
diff changeset
192 if (!up) fast_memcpy(d, s, w);
27754
08d18fe9da52 Change all occurrences of asm and __asm to __asm__, same as was done for FFmpeg.
diego
parents: 25221
diff changeset
193 __asm__ volatile("emms \n\t" : : : "memory");
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
194 }
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
195
29639
26a355ffe458 Add several HAVE_EBX_AVAILABLE conditions where necessary
reimar
parents: 29263
diff changeset
196 #if HAVE_EBX_AVAILABLE
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
197 static void qpel_4tap_MMX(unsigned char *d, unsigned char *s, int w, int h, int ds, int ss, int up)
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
198 {
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
199 int i, j, ssd=ss;
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
200 static const short filter[] = {
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
201 29, 29, 29, 29, 110, 110, 110, 110,
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
202 9, 9, 9, 9, 3, 3, 3, 3,
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
203 64, 64, 64, 64 };
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
204 int crap1, crap2;
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
205 if (up) {
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
206 ssd = -ss;
23457
a124f3abc1ec Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents: 18986
diff changeset
207 fast_memcpy(d, s, w);
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
208 d += ds; s += ss;
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
209 }
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
210 for (j=0; j<w; j++)
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
211 d[j] = (s[j+ssd] + 3*s[j])>>2;
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
212 d += ds; s += ss;
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
213 for (i=h-3; i; i--) {
27754
08d18fe9da52 Change all occurrences of asm and __asm to __asm__, same as was done for FFmpeg.
diego
parents: 25221
diff changeset
214 __asm__ volatile(
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
215 "pxor %%mm0, %%mm0 \n\t"
13720
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
216 "movq (%%"REG_d"), %%mm4 \n\t"
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
217 "movq 8(%%"REG_d"), %%mm5 \n\t"
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
218 "movq 16(%%"REG_d"), %%mm6 \n\t"
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
219 "movq 24(%%"REG_d"), %%mm7 \n\t"
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
220 "4: \n\t"
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
221
13720
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
222 "movq (%%"REG_S",%%"REG_a"), %%mm1 \n\t"
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
223 "movq (%%"REG_S"), %%mm2 \n\t"
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
224 "movq (%%"REG_S",%%"REG_b"), %%mm3 \n\t"
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
225 "punpcklbw %%mm0, %%mm1 \n\t"
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
226 "punpcklbw %%mm0, %%mm2 \n\t"
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
227 "pmullw %%mm4, %%mm1 \n\t"
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
228 "punpcklbw %%mm0, %%mm3 \n\t"
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
229 "pmullw %%mm5, %%mm2 \n\t"
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
230 "paddusw %%mm2, %%mm1 \n\t"
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
231 "pmullw %%mm6, %%mm3 \n\t"
13720
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
232 "movq (%%"REG_S",%%"REG_a",2), %%mm2 \n\t"
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
233 "psubusw %%mm3, %%mm1 \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 28335
diff changeset
234 "punpcklbw %%mm0, %%mm2 \n\t"
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
235 "pmullw %%mm7, %%mm2 \n\t"
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
236 "psubusw %%mm2, %%mm1 \n\t"
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
237 "psrlw $7, %%mm1 \n\t"
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
238
13720
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
239 "movq (%%"REG_S",%%"REG_a"), %%mm2 \n\t"
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
240 "movq (%%"REG_S"), %%mm3 \n\t"
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
241 "punpckhbw %%mm0, %%mm2 \n\t"
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
242 "punpckhbw %%mm0, %%mm3 \n\t"
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
243 "pmullw %%mm4, %%mm2 \n\t"
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
244 "pmullw %%mm5, %%mm3 \n\t"
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
245 "paddusw %%mm3, %%mm2 \n\t"
13720
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
246 "movq (%%"REG_S",%%"REG_b"), %%mm3 \n\t"
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
247 "punpckhbw %%mm0, %%mm3 \n\t"
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
248 "pmullw %%mm6, %%mm3 \n\t"
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
249 "psubusw %%mm3, %%mm2 \n\t"
13720
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
250 "movq (%%"REG_S",%%"REG_a",2), %%mm3 \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 28335
diff changeset
251 "punpckhbw %%mm0, %%mm3 \n\t"
13720
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
252 "add $8, %%"REG_S" \n\t"
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
253 "pmullw %%mm7, %%mm3 \n\t"
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
254 "psubusw %%mm3, %%mm2 \n\t"
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
255 "psrlw $7, %%mm2 \n\t"
29263
0f1b5b68af32 whitespace cosmetics: Remove all trailing whitespace.
diego
parents: 28335
diff changeset
256
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
257 "packuswb %%mm2, %%mm1 \n\t"
13720
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
258 "movq %%mm1, (%%"REG_D") \n\t"
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
259 "add $8, %%"REG_D" \n\t"
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
260 "decl %%ecx \n\t"
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
261 "jnz 4b \n\t"
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
262 : "=S"(crap1), "=D"(crap2)
13720
821f464b4d90 adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents: 10078
diff changeset
263 : "c"(w>>3), "S"(s), "D"(d), "a"((long)ssd), "b"((long)-ssd), "d"(filter)
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
264 );
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
265 for (j=w-(w&7); j<w; j++)
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
266 d[j] = (-9*s[j-ssd] + 111*s[j] + 29*s[j+ssd] - 3*s[j+ssd+ssd])>>7;
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
267 d += ds;
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
268 s += ss;
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
269 }
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
270 for (j=0; j<w; j++)
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
271 d[j] = (s[j+ssd] + 3*s[j])>>2;
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
272 d += ds; s += ss;
23457
a124f3abc1ec Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents: 18986
diff changeset
273 if (!up) fast_memcpy(d, s, w);
27754
08d18fe9da52 Change all occurrences of asm and __asm to __asm__, same as was done for FFmpeg.
diego
parents: 25221
diff changeset
274 __asm__ volatile("emms \n\t" : : : "memory");
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
275 }
29639
26a355ffe458 Add several HAVE_EBX_AVAILABLE conditions where necessary
reimar
parents: 29263
diff changeset
276 #endif /* HAVE_EBX_AVAILABLE */
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
277 #endif
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
278
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
279 static inline int clamp(int a)
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
280 {
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
281 // If a<512, this is equivalent to:
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
282 // return (a<0) ? 0 : ( (a>255) ? 255 : a);
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
283 return (~(a>>31)) & (a | ((a<<23)>>31));
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
284 }
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
285
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
286 static void qpel_li_C(unsigned char *d, unsigned char *s, int w, int h, int ds, int ss, int up)
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
287 {
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
288 int i, j, ssd=ss;
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
289 if (up) {
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
290 ssd = -ss;
23457
a124f3abc1ec Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents: 18986
diff changeset
291 fast_memcpy(d, s, w);
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
292 d += ds;
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
293 s += ss;
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
294 }
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
295 for (i=h-1; i; i--) {
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
296 for (j=0; j<w; j++)
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
297 d[j] = (s[j+ssd] + 3*s[j])>>2;
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
298 d += ds;
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
299 s += ss;
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
300 }
23457
a124f3abc1ec Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents: 18986
diff changeset
301 if (!up) fast_memcpy(d, s, w);
10009
69f10d08c3be new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents: 9593
diff changeset
302 }
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
303
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
304 static void qpel_4tap_C(unsigned char *d, unsigned char *s, int w, int h, int ds, int ss, int up)
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
305 {
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
306 int i, j, ssd=ss;
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
307 if (up) {
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
308 ssd = -ss;
23457
a124f3abc1ec Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents: 18986
diff changeset
309 fast_memcpy(d, s, w);
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
310 d += ds; s += ss;
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
311 }
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
312 for (j=0; j<w; j++)
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
313 d[j] = (s[j+ssd] + 3*s[j] + 2)>>2;
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
314 d += ds; s += ss;
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
315 for (i=h-3; i; i--) {
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
316 for (j=0; j<w; j++)
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
317 d[j] = clamp((-9*s[j-ssd] + 111*s[j] + 29*s[j+ssd] - 3*s[j+ssd+ssd] + 64)>>7);
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
318 d += ds; s += ss;
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
319 }
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
320 for (j=0; j<w; j++)
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
321 d[j] = (s[j+ssd] + 3*s[j] + 2)>>2;
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
322 d += ds; s += ss;
23457
a124f3abc1ec Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents: 18986
diff changeset
323 if (!up) fast_memcpy(d, s, w);
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
324 }
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
325
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
326 static void (*qpel_li)(unsigned char *d, unsigned char *s, int w, int h, int ds, int ss, int up);
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
327 static void (*qpel_4tap)(unsigned char *d, unsigned char *s, int w, int h, int ds, int ss, int up);
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
328
30642
a972c1a4a012 cosmetics: Rename struct vf_instance_s --> vf_instance.
diego
parents: 30638
diff changeset
329 static int continue_buffered_image(struct vf_instance *vf);
18917
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
330
30642
a972c1a4a012 cosmetics: Rename struct vf_instance_s --> vf_instance.
diego
parents: 30638
diff changeset
331 static int put_image(struct vf_instance *vf, mp_image_t *mpi, double pts)
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
332 {
18917
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
333 vf->priv->buffered_mpi = mpi;
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
334 vf->priv->buffered_pts = pts;
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
335 vf->priv->buffered_i = 0;
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
336 return continue_buffered_image(vf);
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
337 }
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
338
30443
f60a1db3aee4 Estimate pts of additional frame generated by tfields also if correct-pts is
reimar
parents: 30442
diff changeset
339 static double calc_pts(double base_pts, int field)
f60a1db3aee4 Estimate pts of additional frame generated by tfields also if correct-pts is
reimar
parents: 30442
diff changeset
340 {
f60a1db3aee4 Estimate pts of additional frame generated by tfields also if correct-pts is
reimar
parents: 30442
diff changeset
341 // FIXME this assumes 25 fps / 50 fields per second
f60a1db3aee4 Estimate pts of additional frame generated by tfields also if correct-pts is
reimar
parents: 30442
diff changeset
342 return base_pts + 0.02 * field;
f60a1db3aee4 Estimate pts of additional frame generated by tfields also if correct-pts is
reimar
parents: 30442
diff changeset
343 }
f60a1db3aee4 Estimate pts of additional frame generated by tfields also if correct-pts is
reimar
parents: 30442
diff changeset
344
30642
a972c1a4a012 cosmetics: Rename struct vf_instance_s --> vf_instance.
diego
parents: 30638
diff changeset
345 static int continue_buffered_image(struct vf_instance *vf)
18917
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
346 {
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
347 int i=vf->priv->buffered_i;
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
348 double pts = vf->priv->buffered_pts;
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
349 mp_image_t *mpi = vf->priv->buffered_mpi;
14888
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
350 int ret=0;
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
351 mp_image_t *dmpi;
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
352 void (*qpel)(unsigned char *, unsigned char *, int, int, int, int, int);
10078
379f48cace77 support more image formats. hopefully this bpp handling is correct...
rfelker
parents: 10052
diff changeset
353 int bpp=1;
14888
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
354 int tff;
10078
379f48cace77 support more image formats. hopefully this bpp handling is correct...
rfelker
parents: 10052
diff changeset
355
18986
d743c48823cc c++ decls, 100000000000l to whoever broke my code like this..
rfelker
parents: 18917
diff changeset
356 if (i == 0)
d743c48823cc c++ decls, 100000000000l to whoever broke my code like this..
rfelker
parents: 18917
diff changeset
357 vf_queue_frame(vf, continue_buffered_image);
d743c48823cc c++ decls, 100000000000l to whoever broke my code like this..
rfelker
parents: 18917
diff changeset
358
10078
379f48cace77 support more image formats. hopefully this bpp handling is correct...
rfelker
parents: 10052
diff changeset
359 if (!(mpi->flags & MP_IMGFLAG_PLANAR)) bpp = mpi->bpp/8;
14888
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
360 if (vf->priv->parity < 0) {
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
361 if (mpi->fields & MP_IMGFIELD_ORDERED)
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
362 tff = mpi->fields & MP_IMGFIELD_TOP_FIRST;
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
363 else
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
364 tff = 1;
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
365 }
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
366 else tff = (vf->priv->parity&1)^1;
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
367
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
368 switch (vf->priv->mode) {
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
369 case 2:
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
370 qpel = qpel_li;
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
371 break;
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
372 case 3:
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
373 // TODO: add 3tap filter
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
374 qpel = qpel_4tap;
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
375 break;
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
376 case 4:
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
377 qpel = qpel_4tap;
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
378 break;
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
379 }
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
380
08264c647f46 new filter
rfelker
parents:
diff changeset
381 switch (vf->priv->mode) {
08264c647f46 new filter
rfelker
parents:
diff changeset
382 case 0:
18917
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
383 for (; i<2; i++) {
14888
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
384 dmpi = vf_get_image(vf->next, mpi->imgfmt,
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
385 MP_IMGTYPE_EXPORT, MP_IMGFLAG_ACCEPT_STRIDE,
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
386 mpi->width, mpi->height/2);
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
387 dmpi->planes[0] = mpi->planes[0] + (i^!tff)*mpi->stride[0];
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
388 dmpi->stride[0] = 2*mpi->stride[0];
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
389 if (mpi->flags & MP_IMGFLAG_PLANAR) {
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
390 dmpi->planes[1] = mpi->planes[1] + (i^!tff)*mpi->stride[1];
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
391 dmpi->planes[2] = mpi->planes[2] + (i^!tff)*mpi->stride[2];
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
392 dmpi->stride[1] = 2*mpi->stride[1];
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
393 dmpi->stride[2] = 2*mpi->stride[2];
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
394 }
30443
f60a1db3aee4 Estimate pts of additional frame generated by tfields also if correct-pts is
reimar
parents: 30442
diff changeset
395 ret |= vf_next_put_image(vf, dmpi, calc_pts(pts, i));
18917
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
396 if (correct_pts)
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
397 break;
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
398 else
28175
60402016152c Fix OSD flicker with tfields as well.
reimar
parents: 27754
diff changeset
399 if (!i) vf_extra_flip(vf);
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
400 }
14888
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
401 break;
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
402 case 1:
18917
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
403 for (; i<2; i++) {
14888
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
404 dmpi = vf_get_image(vf->next, mpi->imgfmt,
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
405 MP_IMGTYPE_TEMP, MP_IMGFLAG_ACCEPT_STRIDE,
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
406 mpi->width, mpi->height);
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
407 my_memcpy_pic(dmpi->planes[0] + (i^!tff)*dmpi->stride[0],
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
408 mpi->planes[0] + (i^!tff)*mpi->stride[0],
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
409 mpi->w*bpp, mpi->h/2, dmpi->stride[0]*2, mpi->stride[0]*2);
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
410 deint(dmpi->planes[0], dmpi->stride[0], mpi->planes[0], mpi->stride[0], mpi->w, mpi->h, (i^!tff));
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
411 if (mpi->flags & MP_IMGFLAG_PLANAR) {
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
412 my_memcpy_pic(dmpi->planes[1] + (i^!tff)*dmpi->stride[1],
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
413 mpi->planes[1] + (i^!tff)*mpi->stride[1],
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
414 mpi->chroma_width, mpi->chroma_height/2,
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
415 dmpi->stride[1]*2, mpi->stride[1]*2);
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
416 my_memcpy_pic(dmpi->planes[2] + (i^!tff)*dmpi->stride[2],
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
417 mpi->planes[2] + (i^!tff)*mpi->stride[2],
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
418 mpi->chroma_width, mpi->chroma_height/2,
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
419 dmpi->stride[2]*2, mpi->stride[2]*2);
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
420 deint(dmpi->planes[1], dmpi->stride[1], mpi->planes[1], mpi->stride[1],
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
421 mpi->chroma_width, mpi->chroma_height, (i^!tff));
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
422 deint(dmpi->planes[2], dmpi->stride[2], mpi->planes[2], mpi->stride[2],
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
423 mpi->chroma_width, mpi->chroma_height, (i^!tff));
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
424 }
30443
f60a1db3aee4 Estimate pts of additional frame generated by tfields also if correct-pts is
reimar
parents: 30442
diff changeset
425 ret |= vf_next_put_image(vf, dmpi, calc_pts(pts, i));
18917
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
426 if (correct_pts)
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
427 break;
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
428 else
28175
60402016152c Fix OSD flicker with tfields as well.
reimar
parents: 27754
diff changeset
429 if (!i) vf_extra_flip(vf);
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
430 }
14888
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
431 break;
10009
69f10d08c3be new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents: 9593
diff changeset
432 case 2:
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
433 case 3:
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
434 case 4:
18917
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
435 for (; i<2; i++) {
14888
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
436 dmpi = vf_get_image(vf->next, mpi->imgfmt,
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
437 MP_IMGTYPE_TEMP, MP_IMGFLAG_ACCEPT_STRIDE,
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
438 mpi->width, mpi->height/2);
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
439 qpel(dmpi->planes[0], mpi->planes[0] + (i^!tff)*mpi->stride[0],
15012
ad8815f740d3 1000l, last commit broke qpel interp entirely
rfelker
parents: 14888
diff changeset
440 mpi->w*bpp, mpi->h/2, dmpi->stride[0], mpi->stride[0]*2, (i^!tff));
14888
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
441 if (mpi->flags & MP_IMGFLAG_PLANAR) {
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
442 qpel(dmpi->planes[1],
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
443 mpi->planes[1] + (i^!tff)*mpi->stride[1],
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
444 mpi->chroma_width, mpi->chroma_height/2,
15012
ad8815f740d3 1000l, last commit broke qpel interp entirely
rfelker
parents: 14888
diff changeset
445 dmpi->stride[1], mpi->stride[1]*2, (i^!tff));
14888
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
446 qpel(dmpi->planes[2],
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
447 mpi->planes[2] + (i^!tff)*mpi->stride[2],
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
448 mpi->chroma_width, mpi->chroma_height/2,
15012
ad8815f740d3 1000l, last commit broke qpel interp entirely
rfelker
parents: 14888
diff changeset
449 dmpi->stride[2], mpi->stride[2]*2, (i^!tff));
14888
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
450 }
30443
f60a1db3aee4 Estimate pts of additional frame generated by tfields also if correct-pts is
reimar
parents: 30442
diff changeset
451 ret |= vf_next_put_image(vf, dmpi, calc_pts(pts, i));
18917
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
452 if (correct_pts)
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
453 break;
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
454 else
28175
60402016152c Fix OSD flicker with tfields as well.
reimar
parents: 27754
diff changeset
455 if (!i) vf_extra_flip(vf);
10009
69f10d08c3be new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents: 9593
diff changeset
456 }
14888
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
457 break;
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
458 }
18917
d9a75b26da6c Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents: 17906
diff changeset
459 vf->priv->buffered_i = 1;
14888
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
460 return ret;
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
461 }
08264c647f46 new filter
rfelker
parents:
diff changeset
462
30642
a972c1a4a012 cosmetics: Rename struct vf_instance_s --> vf_instance.
diego
parents: 30638
diff changeset
463 static int query_format(struct vf_instance *vf, unsigned int fmt)
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
464 {
30442
8d7a061d427b Enable query_format for vf_tfields to make sure it isn't used for formats
reimar
parents: 30441
diff changeset
465 /* FIXME - figure out which formats exactly work */
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
466 switch (fmt) {
30442
8d7a061d427b Enable query_format for vf_tfields to make sure it isn't used for formats
reimar
parents: 30441
diff changeset
467 default:
8d7a061d427b Enable query_format for vf_tfields to make sure it isn't used for formats
reimar
parents: 30441
diff changeset
468 if (vf->priv->mode == 1)
8d7a061d427b Enable query_format for vf_tfields to make sure it isn't used for formats
reimar
parents: 30441
diff changeset
469 return 0;
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
470 case IMGFMT_YV12:
08264c647f46 new filter
rfelker
parents:
diff changeset
471 case IMGFMT_IYUV:
08264c647f46 new filter
rfelker
parents:
diff changeset
472 case IMGFMT_I420:
08264c647f46 new filter
rfelker
parents:
diff changeset
473 return vf_next_query_format(vf, fmt);
08264c647f46 new filter
rfelker
parents:
diff changeset
474 }
08264c647f46 new filter
rfelker
parents:
diff changeset
475 return 0;
08264c647f46 new filter
rfelker
parents:
diff changeset
476 }
08264c647f46 new filter
rfelker
parents:
diff changeset
477
30642
a972c1a4a012 cosmetics: Rename struct vf_instance_s --> vf_instance.
diego
parents: 30638
diff changeset
478 static int config(struct vf_instance *vf,
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
479 int width, int height, int d_width, int d_height,
08264c647f46 new filter
rfelker
parents:
diff changeset
480 unsigned int flags, unsigned int outfmt)
08264c647f46 new filter
rfelker
parents:
diff changeset
481 {
08264c647f46 new filter
rfelker
parents:
diff changeset
482 switch (vf->priv->mode) {
08264c647f46 new filter
rfelker
parents:
diff changeset
483 case 0:
10009
69f10d08c3be new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents: 9593
diff changeset
484 case 2:
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
485 case 3:
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
486 case 4:
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
487 return vf_next_config(vf,width,height/2,d_width,d_height,flags,outfmt);
08264c647f46 new filter
rfelker
parents:
diff changeset
488 case 1:
08264c647f46 new filter
rfelker
parents:
diff changeset
489 return vf_next_config(vf,width,height,d_width,d_height,flags,outfmt);
08264c647f46 new filter
rfelker
parents:
diff changeset
490 }
08264c647f46 new filter
rfelker
parents:
diff changeset
491 return 0;
08264c647f46 new filter
rfelker
parents:
diff changeset
492 }
08264c647f46 new filter
rfelker
parents:
diff changeset
493
30642
a972c1a4a012 cosmetics: Rename struct vf_instance_s --> vf_instance.
diego
parents: 30638
diff changeset
494 static void uninit(struct vf_instance *vf)
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
495 {
08264c647f46 new filter
rfelker
parents:
diff changeset
496 free(vf->priv);
08264c647f46 new filter
rfelker
parents:
diff changeset
497 }
08264c647f46 new filter
rfelker
parents:
diff changeset
498
30638
a7b908875c14 Rename open() vf initialization function to vf_open().
diego
parents: 30633
diff changeset
499 static int vf_open(vf_instance_t *vf, char *args)
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
500 {
08264c647f46 new filter
rfelker
parents:
diff changeset
501 struct vf_priv_s *p;
08264c647f46 new filter
rfelker
parents:
diff changeset
502 vf->config = config;
08264c647f46 new filter
rfelker
parents:
diff changeset
503 vf->put_image = put_image;
30442
8d7a061d427b Enable query_format for vf_tfields to make sure it isn't used for formats
reimar
parents: 30441
diff changeset
504 vf->query_format = query_format;
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
505 vf->uninit = uninit;
08264c647f46 new filter
rfelker
parents:
diff changeset
506 vf->default_reqs = VFCAP_ACCEPT_STRIDE;
08264c647f46 new filter
rfelker
parents:
diff changeset
507 vf->priv = p = calloc(1, sizeof(struct vf_priv_s));
15013
0bb95cd581b6 sane default mode
rfelker
parents: 15012
diff changeset
508 vf->priv->mode = 4;
14888
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
509 vf->priv->parity = -1;
32dcf8672086 configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents: 13720
diff changeset
510 if (args) sscanf(args, "%d:%d", &vf->priv->mode, &vf->priv->parity);
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
511 qpel_li = qpel_li_C;
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
512 qpel_4tap = qpel_4tap_C;
28290
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_...
reimar
parents: 28175
diff changeset
513 #if HAVE_MMX
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
514 if(gCpuCaps.hasMMX) qpel_li = qpel_li_MMX;
29639
26a355ffe458 Add several HAVE_EBX_AVAILABLE conditions where necessary
reimar
parents: 29263
diff changeset
515 #if HAVE_EBX_AVAILABLE
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
516 if(gCpuCaps.hasMMX) qpel_4tap = qpel_4tap_MMX;
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
517 #endif
29639
26a355ffe458 Add several HAVE_EBX_AVAILABLE conditions where necessary
reimar
parents: 29263
diff changeset
518 #endif
28290
25337a2147e7 Lots and lots of #ifdef ARCH_... -> #if ARCH_...
reimar
parents: 28175
diff changeset
519 #if HAVE_MMX2
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
520 if(gCpuCaps.hasMMX2) qpel_li = qpel_li_MMX2;
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
521 #endif
28335
31287e75b5d8 HAVE_3DNOW --> HAVE_AMD3DNOW
diego
parents: 28290
diff changeset
522 #if HAVE_AMD3DNOW
10049
765c2276aa0c more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents: 10020
diff changeset
523 if(gCpuCaps.has3DNow) qpel_li = qpel_li_3DNOW;
10020
9829b7e61b55 new mmx/mmx2/3dnow code for improved performance
rfelker
parents: 10009
diff changeset
524 #endif
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
525 return 1;
08264c647f46 new filter
rfelker
parents:
diff changeset
526 }
08264c647f46 new filter
rfelker
parents:
diff changeset
527
25221
00fff9a3b735 Make all vf_info_t structs const
reimar
parents: 24605
diff changeset
528 const vf_info_t vf_info_tfields = {
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
529 "temporal field separation",
08264c647f46 new filter
rfelker
parents:
diff changeset
530 "tfields",
08264c647f46 new filter
rfelker
parents:
diff changeset
531 "Rich Felker",
08264c647f46 new filter
rfelker
parents:
diff changeset
532 "",
30638
a7b908875c14 Rename open() vf initialization function to vf_open().
diego
parents: 30633
diff changeset
533 vf_open,
9593
e9a2af584986 Add the new -vf option wich is the same as vop in reverse order.
albeu
parents: 9514
diff changeset
534 NULL
9514
08264c647f46 new filter
rfelker
parents:
diff changeset
535 };