Mercurial > mplayer.hg
annotate libmpcodecs/vf_tfields.c @ 25549:16c0e23971f3
consistency cosmetics: Do not #define multiple inclusion guards to 1.
author | diego |
---|---|
date | Tue, 01 Jan 2008 20:11:08 +0000 |
parents | 00fff9a3b735 |
children | 08d18fe9da52 |
rev | line source |
---|---|
9514 | 1 #include <stdio.h> |
2 #include <stdlib.h> | |
3 #include <string.h> | |
4 | |
17012 | 5 #include "config.h" |
6 #include "mp_msg.h" | |
7 #include "cpudetect.h" | |
9514 | 8 |
9 #include "img_format.h" | |
10 #include "mp_image.h" | |
11 #include "vf.h" | |
12 | |
17012 | 13 #include "libvo/fastmemcpy.h" |
9514 | 14 |
15 struct vf_priv_s { | |
16 int mode; | |
14888
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
17 int parity; |
18917
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
18 int buffered_i; |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
19 mp_image_t *buffered_mpi; |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
20 double buffered_pts; |
9514 | 21 }; |
22 | |
23 static void deint(unsigned char *dest, int ds, unsigned char *src, int ss, int w, int h, int field) | |
24 { | |
25 int x, y; | |
26 src += ss; | |
27 dest += ds; | |
28 if (field) { | |
29 src += ss; | |
30 dest += ds; | |
31 h -= 2; | |
32 } | |
33 for (y=h/2; y; y--) { | |
34 for (x=0; x<w; x++) { | |
35 if (((src[x-ss] < src[x]) && (src[x+ss] < src[x])) || | |
36 ((src[x-ss] > src[x]) && (src[x+ss] > src[x]))) { | |
37 //dest[x] = (src[x+ss] + src[x-ss])>>1; | |
38 dest[x] = ((src[x+ss]<<1) + (src[x-ss]<<1) | |
39 + src[x+ss+1] + src[x-ss+1] | |
40 + src[x+ss-1] + src[x-ss-1])>>3; | |
41 } | |
42 else dest[x] = src[x]; | |
43 } | |
44 dest += ds<<1; | |
45 src += ss<<1; | |
46 } | |
47 } | |
48 | |
10020 | 49 #ifdef HAVE_3DNOW |
10049
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
50 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
|
51 { |
10020 | 52 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
|
53 long crap1, crap2; |
10020 | 54 if (up) { |
55 ssd = -ss; | |
23457
a124f3abc1ec
Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents:
18986
diff
changeset
|
56 fast_memcpy(d, s, w); |
10020 | 57 d += ds; |
58 s += ss; | |
59 } | |
10009
69f10d08c3be
new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents:
9593
diff
changeset
|
60 for (i=h-1; i; i--) { |
10049
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
61 asm volatile( |
10020 | 62 "1: \n\t" |
13720
821f464b4d90
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents:
10078
diff
changeset
|
63 "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
|
64 "movq (%%"REG_S",%%"REG_a"), %%mm1 \n\t" |
10020 | 65 "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
|
66 "add $8, %%"REG_S" \n\t" |
10020 | 67 "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
|
68 "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
|
69 "add $8, %%"REG_D" \n\t" |
10020 | 70 "decl %%ecx \n\t" |
71 "jnz 1b \n\t" | |
72 : "=S"(crap1), "=D"(crap2) | |
13720
821f464b4d90
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents:
10078
diff
changeset
|
73 : "c"(w>>3), "S"(s), "D"(d), "a"((long)ssd) |
10020 | 74 ); |
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 for (j=w-(w&7); j<w; j++) |
10020 | 76 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
|
77 d += ds; |
69f10d08c3be
new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents:
9593
diff
changeset
|
78 s += ss; |
69f10d08c3be
new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents:
9593
diff
changeset
|
79 } |
23457
a124f3abc1ec
Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents:
18986
diff
changeset
|
80 if (!up) fast_memcpy(d, s, w); |
10020 | 81 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
|
82 } |
10020 | 83 #endif |
10009
69f10d08c3be
new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents:
9593
diff
changeset
|
84 |
10020 | 85 #ifdef 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
|
86 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
|
87 { |
10020 | 88 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
|
89 long crap1, crap2; |
10020 | 90 if (up) { |
91 ssd = -ss; | |
23457
a124f3abc1ec
Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents:
18986
diff
changeset
|
92 fast_memcpy(d, s, w); |
10020 | 93 d += ds; |
94 s += ss; | |
95 } | |
10009
69f10d08c3be
new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents:
9593
diff
changeset
|
96 for (i=h-1; i; i--) { |
10049
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
97 asm volatile( |
10020 | 98 "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
|
99 "2: \n\t" |
13720
821f464b4d90
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents:
10078
diff
changeset
|
100 "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
|
101 "movq (%%"REG_S",%%"REG_a"), %%mm1 \n\t" |
10020 | 102 "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
|
103 "add $8, %%"REG_S" \n\t" |
10020 | 104 "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
|
105 "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
|
106 "add $8, %%"REG_D" \n\t" |
10020 | 107 "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
|
108 "jnz 2b \n\t" |
10020 | 109 : "=S"(crap1), "=D"(crap2) |
13720
821f464b4d90
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents:
10078
diff
changeset
|
110 : "c"(w>>3), "S"(s), "D"(d), "a"((long)ssd) |
10020 | 111 ); |
10049
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
112 for (j=w-(w&7); j<w; j++) |
10020 | 113 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
|
114 d += ds; |
69f10d08c3be
new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents:
9593
diff
changeset
|
115 s += ss; |
69f10d08c3be
new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents:
9593
diff
changeset
|
116 } |
23457
a124f3abc1ec
Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents:
18986
diff
changeset
|
117 if (!up) fast_memcpy(d, s, w); |
10020 | 118 asm volatile("emms \n\t" : : : "memory"); |
119 } | |
120 #endif | |
121 | |
122 #ifdef 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
|
123 static void qpel_li_MMX(unsigned char *d, unsigned char *s, int w, int h, int ds, int ss, int up) |
10020 | 124 { |
125 int i, j, ssd=ss; | |
126 int crap1, crap2; | |
127 if (up) { | |
128 ssd = -ss; | |
23457
a124f3abc1ec
Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents:
18986
diff
changeset
|
129 fast_memcpy(d, s, w); |
10020 | 130 d += ds; |
131 s += ss; | |
132 } | |
133 for (i=h-1; i; i--) { | |
10049
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
134 asm volatile( |
10020 | 135 "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
|
136 "3: \n\t" |
13720
821f464b4d90
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents:
10078
diff
changeset
|
137 "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
|
138 "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
|
139 "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
|
140 "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
|
141 "add $8, %%"REG_S" \n\t" |
10020 | 142 "punpcklbw %%mm7, %%mm0 \n\t" |
143 "punpckhbw %%mm7, %%mm1 \n\t" | |
144 "punpcklbw %%mm7, %%mm2 \n\t" | |
145 "punpckhbw %%mm7, %%mm3 \n\t" | |
146 "paddw %%mm0, %%mm2 \n\t" | |
147 "paddw %%mm1, %%mm3 \n\t" | |
148 "paddw %%mm0, %%mm2 \n\t" | |
149 "paddw %%mm1, %%mm3 \n\t" | |
150 "paddw %%mm0, %%mm2 \n\t" | |
151 "paddw %%mm1, %%mm3 \n\t" | |
152 "psrlw $2, %%mm2 \n\t" | |
153 "psrlw $2, %%mm3 \n\t" | |
154 "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
|
155 "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
|
156 "add $8, %%"REG_D" \n\t" |
10020 | 157 "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
|
158 "jnz 3b \n\t" |
10020 | 159 : "=S"(crap1), "=D"(crap2) |
13720
821f464b4d90
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents:
10078
diff
changeset
|
160 : "c"(w>>3), "S"(s), "D"(d), "a"((long)ssd) |
10020 | 161 ); |
10049
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
162 for (j=w-(w&7); j<w; j++) |
10020 | 163 d[j] = (s[j+ssd] + 3*s[j])>>2; |
164 d += ds; | |
165 s += ss; | |
166 } | |
23457
a124f3abc1ec
Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents:
18986
diff
changeset
|
167 if (!up) fast_memcpy(d, s, w); |
10020 | 168 asm volatile("emms \n\t" : : : "memory"); |
169 } | |
10049
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
170 |
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
171 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
|
172 { |
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
173 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
|
174 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
|
175 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
|
176 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
|
177 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
|
178 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
|
179 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
|
180 ssd = -ss; |
23457
a124f3abc1ec
Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents:
18986
diff
changeset
|
181 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
|
182 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
|
183 } |
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
184 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
|
185 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
|
186 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
|
187 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
|
188 asm volatile( |
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
189 "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
|
190 "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
|
191 "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
|
192 "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
|
193 "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
|
194 "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
|
195 |
13720
821f464b4d90
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents:
10078
diff
changeset
|
196 "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
|
197 "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
|
198 "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
|
199 "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
|
200 "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
|
201 "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
|
202 "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
|
203 "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
|
204 "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
|
205 "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
|
206 "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
|
207 "psubusw %%mm3, %%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
|
208 "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
|
209 "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
|
210 "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
|
211 "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
|
212 |
13720
821f464b4d90
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents:
10078
diff
changeset
|
213 "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
|
214 "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
|
215 "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
|
216 "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
|
217 "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
|
218 "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
|
219 "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
|
220 "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
|
221 "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
|
222 "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
|
223 "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
|
224 "movq (%%"REG_S",%%"REG_a",2), %%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 "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
|
226 "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
|
227 "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
|
228 "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
|
229 "psrlw $7, %%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 |
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
231 "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
|
232 "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
|
233 "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
|
234 "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
|
235 "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
|
236 : "=S"(crap1), "=D"(crap2) |
13720
821f464b4d90
adapting existing mmx/mmx2/sse/3dnow optimizations so they work on x86_64
aurel
parents:
10078
diff
changeset
|
237 : "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
|
238 ); |
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
239 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
|
240 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
|
241 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
|
242 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
|
243 } |
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
244 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
|
245 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
|
246 d += ds; s += ss; |
23457
a124f3abc1ec
Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents:
18986
diff
changeset
|
247 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
|
248 asm volatile("emms \n\t" : : : "memory"); |
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
249 } |
10020 | 250 #endif |
251 | |
10049
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
252 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
|
253 { |
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
254 // 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
|
255 // 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
|
256 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
|
257 } |
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
258 |
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
259 static void qpel_li_C(unsigned char *d, unsigned char *s, int w, int h, int ds, int ss, int up) |
10020 | 260 { |
261 int i, j, ssd=ss; | |
262 if (up) { | |
263 ssd = -ss; | |
23457
a124f3abc1ec
Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents:
18986
diff
changeset
|
264 fast_memcpy(d, s, w); |
10020 | 265 d += ds; |
266 s += ss; | |
267 } | |
268 for (i=h-1; i; i--) { | |
269 for (j=0; j<w; j++) | |
270 d[j] = (s[j+ssd] + 3*s[j])>>2; | |
271 d += ds; | |
272 s += ss; | |
273 } | |
23457
a124f3abc1ec
Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents:
18986
diff
changeset
|
274 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
|
275 } |
9514 | 276 |
10049
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
277 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
|
278 { |
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
279 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
|
280 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
|
281 ssd = -ss; |
23457
a124f3abc1ec
Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents:
18986
diff
changeset
|
282 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
|
283 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
|
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 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
|
286 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
|
287 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
|
288 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
|
289 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
|
290 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
|
291 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
|
292 } |
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
293 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
|
294 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
|
295 d += ds; s += ss; |
23457
a124f3abc1ec
Replace implicit use of fast_memcpy via macro by explicit use to allow
reimar
parents:
18986
diff
changeset
|
296 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
|
297 } |
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
298 |
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
299 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
|
300 static void (*qpel_4tap)(unsigned char *d, unsigned char *s, int w, int h, int ds, int ss, int up); |
9514 | 301 |
18917
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
302 static int continue_buffered_image(struct vf_instance_s *); |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
303 extern int correct_pts; |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
304 |
17906
20aca9baf5d8
passing pts through the filter layer (lets see if pts or cola comes out at the end)
michael
parents:
17012
diff
changeset
|
305 static int put_image(struct vf_instance_s* vf, mp_image_t *mpi, double pts) |
9514 | 306 { |
18917
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
307 vf->priv->buffered_mpi = mpi; |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
308 vf->priv->buffered_pts = pts; |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
309 vf->priv->buffered_i = 0; |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
310 return continue_buffered_image(vf); |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
311 } |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
312 |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
313 static int continue_buffered_image(struct vf_instance_s *vf) |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
314 { |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
315 int i=vf->priv->buffered_i; |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
316 double pts = vf->priv->buffered_pts; |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
317 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
|
318 int ret=0; |
9514 | 319 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
|
320 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
|
321 int bpp=1; |
14888
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
322 int tff; |
10078
379f48cace77
support more image formats. hopefully this bpp handling is correct...
rfelker
parents:
10052
diff
changeset
|
323 |
18986
d743c48823cc
c++ decls, 100000000000l to whoever broke my code like this..
rfelker
parents:
18917
diff
changeset
|
324 if (i == 0) |
d743c48823cc
c++ decls, 100000000000l to whoever broke my code like this..
rfelker
parents:
18917
diff
changeset
|
325 vf_queue_frame(vf, continue_buffered_image); |
d743c48823cc
c++ decls, 100000000000l to whoever broke my code like this..
rfelker
parents:
18917
diff
changeset
|
326 pts += i * .02; // XXX not right |
d743c48823cc
c++ decls, 100000000000l to whoever broke my code like this..
rfelker
parents:
18917
diff
changeset
|
327 |
10078
379f48cace77
support more image formats. hopefully this bpp handling is correct...
rfelker
parents:
10052
diff
changeset
|
328 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
|
329 if (vf->priv->parity < 0) { |
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
330 if (mpi->fields & MP_IMGFIELD_ORDERED) |
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
331 tff = mpi->fields & MP_IMGFIELD_TOP_FIRST; |
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
332 else |
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
333 tff = 1; |
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
334 } |
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
335 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
|
336 |
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
337 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
|
338 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
|
339 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
|
340 break; |
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
341 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
|
342 // 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
|
343 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
|
344 break; |
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
345 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
|
346 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
|
347 break; |
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
348 } |
9514 | 349 |
350 switch (vf->priv->mode) { | |
351 case 0: | |
18917
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
352 for (; i<2; i++) { |
14888
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
353 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
|
354 MP_IMGTYPE_EXPORT, MP_IMGFLAG_ACCEPT_STRIDE, |
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
355 mpi->width, mpi->height/2); |
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
356 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
|
357 dmpi->stride[0] = 2*mpi->stride[0]; |
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
358 if (mpi->flags & MP_IMGFLAG_PLANAR) { |
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
359 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
|
360 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
|
361 dmpi->stride[1] = 2*mpi->stride[1]; |
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
362 dmpi->stride[2] = 2*mpi->stride[2]; |
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
363 } |
18917
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
364 ret |= vf_next_put_image(vf, dmpi, pts); |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
365 if (correct_pts) |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
366 break; |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
367 else |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
368 if (!i) vf_next_control(vf, VFCTRL_FLIP_PAGE, NULL); |
9514 | 369 } |
14888
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
370 break; |
9514 | 371 case 1: |
18917
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
372 for (; i<2; i++) { |
14888
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
373 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
|
374 MP_IMGTYPE_TEMP, MP_IMGFLAG_ACCEPT_STRIDE, |
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
375 mpi->width, mpi->height); |
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
376 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
|
377 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
|
378 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
|
379 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
|
380 if (mpi->flags & MP_IMGFLAG_PLANAR) { |
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
381 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
|
382 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
|
383 mpi->chroma_width, mpi->chroma_height/2, |
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
384 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
|
385 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
|
386 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
|
387 mpi->chroma_width, mpi->chroma_height/2, |
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
388 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
|
389 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
|
390 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
|
391 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
|
392 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
|
393 } |
18917
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
394 ret |= vf_next_put_image(vf, dmpi, pts); |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
395 if (correct_pts) |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
396 break; |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
397 else |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
398 if (!i) vf_next_control(vf, VFCTRL_FLIP_PAGE, NULL); |
9514 | 399 } |
14888
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
400 break; |
10009
69f10d08c3be
new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents:
9593
diff
changeset
|
401 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
|
402 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
|
403 case 4: |
18917
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
404 for (; i<2; i++) { |
14888
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
405 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
|
406 MP_IMGTYPE_TEMP, MP_IMGFLAG_ACCEPT_STRIDE, |
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
407 mpi->width, mpi->height/2); |
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
408 qpel(dmpi->planes[0], mpi->planes[0] + (i^!tff)*mpi->stride[0], |
15012 | 409 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
|
410 if (mpi->flags & MP_IMGFLAG_PLANAR) { |
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
411 qpel(dmpi->planes[1], |
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
412 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
|
413 mpi->chroma_width, mpi->chroma_height/2, |
15012 | 414 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
|
415 qpel(dmpi->planes[2], |
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
416 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
|
417 mpi->chroma_width, mpi->chroma_height/2, |
15012 | 418 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
|
419 } |
18917
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
420 ret |= vf_next_put_image(vf, dmpi, pts); |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
421 if (correct_pts) |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
422 break; |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
423 else |
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
424 if (!i) vf_next_control(vf, VFCTRL_FLIP_PAGE, NULL); |
10009
69f10d08c3be
new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents:
9593
diff
changeset
|
425 } |
14888
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
426 break; |
9514 | 427 } |
18917
d9a75b26da6c
Add a new video pts tracking mode, enabled by option -correct-pts.
uau
parents:
17906
diff
changeset
|
428 vf->priv->buffered_i = 1; |
14888
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
429 return ret; |
9514 | 430 } |
431 | |
24605
8a6f80593529
Disable unused query_format functions for now until they are
diego
parents:
23666
diff
changeset
|
432 #if 0 |
9514 | 433 static int query_format(struct vf_instance_s* vf, unsigned int fmt) |
434 { | |
435 /* FIXME - figure out which other formats work */ | |
436 switch (fmt) { | |
437 case IMGFMT_YV12: | |
438 case IMGFMT_IYUV: | |
439 case IMGFMT_I420: | |
440 return vf_next_query_format(vf, fmt); | |
441 } | |
442 return 0; | |
443 } | |
24605
8a6f80593529
Disable unused query_format functions for now until they are
diego
parents:
23666
diff
changeset
|
444 #endif |
9514 | 445 |
446 static int config(struct vf_instance_s* vf, | |
447 int width, int height, int d_width, int d_height, | |
448 unsigned int flags, unsigned int outfmt) | |
449 { | |
450 switch (vf->priv->mode) { | |
451 case 0: | |
10009
69f10d08c3be
new mode for tfields filter -- shifts fields by a quarter-pixel so the
rfelker
parents:
9593
diff
changeset
|
452 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
|
453 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
|
454 case 4: |
9514 | 455 return vf_next_config(vf,width,height/2,d_width,d_height,flags,outfmt); |
456 case 1: | |
457 return vf_next_config(vf,width,height,d_width,d_height,flags,outfmt); | |
458 } | |
459 return 0; | |
460 } | |
461 | |
462 static void uninit(struct vf_instance_s* vf) | |
463 { | |
464 free(vf->priv); | |
465 } | |
466 | |
467 static int open(vf_instance_t *vf, char* args) | |
468 { | |
469 struct vf_priv_s *p; | |
470 vf->config = config; | |
471 vf->put_image = put_image; | |
10078
379f48cace77
support more image formats. hopefully this bpp handling is correct...
rfelker
parents:
10052
diff
changeset
|
472 //vf->query_format = query_format; |
9514 | 473 vf->uninit = uninit; |
474 vf->default_reqs = VFCAP_ACCEPT_STRIDE; | |
475 vf->priv = p = calloc(1, sizeof(struct vf_priv_s)); | |
15013 | 476 vf->priv->mode = 4; |
14888
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
477 vf->priv->parity = -1; |
32dcf8672086
configurable field parity (default from source); bugfixes; speed up mode 0
rfelker
parents:
13720
diff
changeset
|
478 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
|
479 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
|
480 qpel_4tap = qpel_4tap_C; |
10020 | 481 #ifdef 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
|
482 if(gCpuCaps.hasMMX) qpel_li = qpel_li_MMX; |
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
483 if(gCpuCaps.hasMMX) qpel_4tap = qpel_4tap_MMX; |
10020 | 484 #endif |
485 #ifdef 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
|
486 if(gCpuCaps.hasMMX2) qpel_li = qpel_li_MMX2; |
10020 | 487 #endif |
488 #ifdef HAVE_3DNOW | |
10049
765c2276aa0c
more 10l's -- fortunately part of the bug was that the buggy code didn't get called...
rfelker
parents:
10020
diff
changeset
|
489 if(gCpuCaps.has3DNow) qpel_li = qpel_li_3DNOW; |
10020 | 490 #endif |
9514 | 491 return 1; |
492 } | |
493 | |
25221 | 494 const vf_info_t vf_info_tfields = { |
9514 | 495 "temporal field separation", |
496 "tfields", | |
497 "Rich Felker", | |
498 "", | |
9593
e9a2af584986
Add the new -vf option wich is the same as vop in reverse order.
albeu
parents:
9514
diff
changeset
|
499 open, |
e9a2af584986
Add the new -vf option wich is the same as vop in reverse order.
albeu
parents:
9514
diff
changeset
|
500 NULL |
9514 | 501 }; |
502 | |
503 |