annotate libpostproc/postprocess.c @ 157:bc12fd7e6153 libavcodec

temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
author michael
date Wed, 14 Nov 2001 02:46:58 +0000
parents c09459686be3
children d1a4f4ca7178
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1 /*
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
2 Copyright (C) 2001 Michael Niedermayer (michaelni@gmx.at)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
4 This program is free software; you can redistribute it and/or modify
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
5 it under the terms of the GNU General Public License as published by
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
6 the Free Software Foundation; either version 2 of the License, or
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
7 (at your option) any later version.
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
8
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
9 This program is distributed in the hope that it will be useful,
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
10 but WITHOUT ANY WARRANTY; without even the implied warranty of
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
12 GNU General Public License for more details.
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
13
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
14 You should have received a copy of the GNU General Public License
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
15 along with this program; if not, write to the Free Software
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
16 Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
17 */
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
18
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
19 /*
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
20 C MMX MMX2 3DNow
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
21 isVertDC Ec Ec
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
22 isVertMinMaxOk Ec Ec
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
23 doVertLowPass E e e
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
24 doVertDefFilter Ec Ec Ec
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
25 isHorizDC Ec Ec
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
26 isHorizMinMaxOk a E
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
27 doHorizLowPass E e e
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
28 doHorizDefFilter Ec Ec Ec
134
2c469e390117 dering in c
michael
parents: 133
diff changeset
29 deRing E e e*
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
30 Vertical RKAlgo1 E a a
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
31 Horizontal RKAlgo1 a a
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
32 Vertical X1# a E E
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
33 Horizontal X1# a E E
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
34 LinIpolDeinterlace e E E*
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
35 CubicIpolDeinterlace a e e*
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
36 LinBlendDeinterlace e E E*
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
37 MedianDeinterlace# Ec Ec
157
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
38 TempDeNoiser# E e e
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
39
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
40 * i dont have a 3dnow CPU -> its untested, but noone said it doesnt work so it seems to work
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
41 # more or less selfinvented filters so the exactness isnt too meaningfull
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
42 E = Exact implementation
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
43 e = allmost exact implementation (slightly different rounding,...)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
44 a = alternative / approximate impl
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
45 c = checked against the other implementations (-vo md5)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
46 */
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
47
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
48 /*
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
49 TODO:
99
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
50 verify that everything workes as it should (how?)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
51 reduce the time wasted on the mem transfer
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
52 implement everything in C at least (done at the moment but ...)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
53 unroll stuff if instructions depend too much on the prior one
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
54 we use 8x8 blocks for the horizontal filters, opendivx seems to use 8x4?
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
55 move YScale thing to the end instead of fixing QP
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
56 write a faster and higher quality deblocking filter :)
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
57 make the mainloop more flexible (variable number of blocks at once
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
58 (the if/else stuff per block is slowing things down)
99
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
59 compare the quality & speed of all filters
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
60 split this huge file
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
61 border remover
140
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
62 optimize c versions
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
63 try to unroll inner for(x=0 ... loop to avoid these damn if(x ... checks
157
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
64 smart blur
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
65 ...
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
66
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
67 Notes:
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
68 */
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
69
107
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
70 //Changelog: use the CVS log
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
71
150
668dda5ef46d added missing #include <malloc.h>
pl
parents: 149
diff changeset
72 #include "../config.h"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
73 #include <inttypes.h>
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
74 #include <stdio.h>
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
75 #include <stdlib.h>
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
76 #include <string.h>
133
e0bf2e9ce24c a few warning fixes (missing #include's)
pl
parents: 132
diff changeset
77 #ifdef HAVE_MALLOC_H
e0bf2e9ce24c a few warning fixes (missing #include's)
pl
parents: 132
diff changeset
78 #include <malloc.h>
e0bf2e9ce24c a few warning fixes (missing #include's)
pl
parents: 132
diff changeset
79 #endif
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
80 //#undef HAVE_MMX2
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
81 //#define HAVE_3DNOW
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
82 //#undef HAVE_MMX
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
83 #include "postprocess.h"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
84
104
9607b48e2c2d Cleanup:
arpi
parents: 102
diff changeset
85 #define MIN(a,b) ((a) > (b) ? (b) : (a))
9607b48e2c2d Cleanup:
arpi
parents: 102
diff changeset
86 #define MAX(a,b) ((a) < (b) ? (b) : (a))
9607b48e2c2d Cleanup:
arpi
parents: 102
diff changeset
87 #define ABS(a) ((a) > 0 ? (a) : (-(a)))
9607b48e2c2d Cleanup:
arpi
parents: 102
diff changeset
88 #define SIGN(a) ((a) > 0 ? 1 : -1)
9607b48e2c2d Cleanup:
arpi
parents: 102
diff changeset
89
9607b48e2c2d Cleanup:
arpi
parents: 102
diff changeset
90 #ifdef HAVE_MMX2
9607b48e2c2d Cleanup:
arpi
parents: 102
diff changeset
91 #define PAVGB(a,b) "pavgb " #a ", " #b " \n\t"
9607b48e2c2d Cleanup:
arpi
parents: 102
diff changeset
92 #elif defined (HAVE_3DNOW)
9607b48e2c2d Cleanup:
arpi
parents: 102
diff changeset
93 #define PAVGB(a,b) "pavgusb " #a ", " #b " \n\t"
9607b48e2c2d Cleanup:
arpi
parents: 102
diff changeset
94 #endif
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
95
134
2c469e390117 dering in c
michael
parents: 133
diff changeset
96 #ifdef HAVE_MMX2
2c469e390117 dering in c
michael
parents: 133
diff changeset
97 #define PMINUB(a,b,t) "pminub " #a ", " #b " \n\t"
2c469e390117 dering in c
michael
parents: 133
diff changeset
98 #elif defined (HAVE_MMX)
2c469e390117 dering in c
michael
parents: 133
diff changeset
99 #define PMINUB(b,a,t) \
2c469e390117 dering in c
michael
parents: 133
diff changeset
100 "movq " #a ", " #t " \n\t"\
2c469e390117 dering in c
michael
parents: 133
diff changeset
101 "psubusb " #b ", " #t " \n\t"\
2c469e390117 dering in c
michael
parents: 133
diff changeset
102 "psubb " #t ", " #a " \n\t"
2c469e390117 dering in c
michael
parents: 133
diff changeset
103 #endif
2c469e390117 dering in c
michael
parents: 133
diff changeset
104
2c469e390117 dering in c
michael
parents: 133
diff changeset
105 #ifdef HAVE_MMX2
2c469e390117 dering in c
michael
parents: 133
diff changeset
106 #define PMAXUB(a,b) "pmaxub " #a ", " #b " \n\t"
2c469e390117 dering in c
michael
parents: 133
diff changeset
107 #elif defined (HAVE_MMX)
2c469e390117 dering in c
michael
parents: 133
diff changeset
108 #define PMAXUB(a,b) \
2c469e390117 dering in c
michael
parents: 133
diff changeset
109 "psubusb " #a ", " #b " \n\t"\
2c469e390117 dering in c
michael
parents: 133
diff changeset
110 "paddb " #a ", " #b " \n\t"
2c469e390117 dering in c
michael
parents: 133
diff changeset
111 #endif
2c469e390117 dering in c
michael
parents: 133
diff changeset
112
2c469e390117 dering in c
michael
parents: 133
diff changeset
113
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
114 #define GET_MODE_BUFFER_SIZE 500
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
115 #define OPTIONS_ARRAY_SIZE 10
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
116
148
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
117 #ifdef HAVE_MMX
149
3fc494e92f0d gcc does optimize writes to non volatile variables away if it didnt know that they were read in between
michael
parents: 148
diff changeset
118 static volatile uint64_t __attribute__((aligned(8))) packedYOffset= 0x0000000000000000LL;
3fc494e92f0d gcc does optimize writes to non volatile variables away if it didnt know that they were read in between
michael
parents: 148
diff changeset
119 static volatile uint64_t __attribute__((aligned(8))) packedYScale= 0x0100010001000100LL;
148
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
120 static uint64_t __attribute__((aligned(8))) w05= 0x0005000500050005LL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
121 static uint64_t __attribute__((aligned(8))) w20= 0x0020002000200020LL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
122 static uint64_t __attribute__((aligned(8))) w1400= 0x1400140014001400LL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
123 static uint64_t __attribute__((aligned(8))) bm00000001= 0x00000000000000FFLL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
124 static uint64_t __attribute__((aligned(8))) bm00010000= 0x000000FF00000000LL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
125 static uint64_t __attribute__((aligned(8))) bm00001000= 0x00000000FF000000LL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
126 static uint64_t __attribute__((aligned(8))) bm10000000= 0xFF00000000000000LL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
127 static uint64_t __attribute__((aligned(8))) bm10000001= 0xFF000000000000FFLL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
128 static uint64_t __attribute__((aligned(8))) bm11000011= 0xFFFF00000000FFFFLL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
129 static uint64_t __attribute__((aligned(8))) bm00000011= 0x000000000000FFFFLL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
130 static uint64_t __attribute__((aligned(8))) bm11111110= 0xFFFFFFFFFFFFFF00LL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
131 static uint64_t __attribute__((aligned(8))) bm11000000= 0xFFFF000000000000LL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
132 static uint64_t __attribute__((aligned(8))) bm00011000= 0x000000FFFF000000LL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
133 static uint64_t __attribute__((aligned(8))) bm00110011= 0x0000FFFF0000FFFFLL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
134 static uint64_t __attribute__((aligned(8))) bm11001100= 0xFFFF0000FFFF0000LL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
135 static uint64_t __attribute__((aligned(8))) b00= 0x0000000000000000LL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
136 static uint64_t __attribute__((aligned(8))) b01= 0x0101010101010101LL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
137 static uint64_t __attribute__((aligned(8))) b02= 0x0202020202020202LL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
138 static uint64_t __attribute__((aligned(8))) b0F= 0x0F0F0F0F0F0F0F0FLL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
139 static uint64_t __attribute__((aligned(8))) b04= 0x0404040404040404LL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
140 static uint64_t __attribute__((aligned(8))) b08= 0x0808080808080808LL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
141 static uint64_t __attribute__((aligned(8))) bFF= 0xFFFFFFFFFFFFFFFFLL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
142 static uint64_t __attribute__((aligned(8))) b20= 0x2020202020202020LL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
143 static uint64_t __attribute__((aligned(8))) b80= 0x8080808080808080LL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
144 static uint64_t __attribute__((aligned(8))) b7E= 0x7E7E7E7E7E7E7E7ELL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
145 static uint64_t __attribute__((aligned(8))) b7C= 0x7C7C7C7C7C7C7C7CLL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
146 static uint64_t __attribute__((aligned(8))) b3F= 0x3F3F3F3F3F3F3F3FLL;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
147 static uint64_t __attribute__((aligned(8))) temp0=0;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
148 static uint64_t __attribute__((aligned(8))) temp1=0;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
149 static uint64_t __attribute__((aligned(8))) temp2=0;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
150 static uint64_t __attribute__((aligned(8))) temp3=0;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
151 static uint64_t __attribute__((aligned(8))) temp4=0;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
152 static uint64_t __attribute__((aligned(8))) temp5=0;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
153 static uint64_t __attribute__((aligned(8))) pQPb=0;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
154 static uint64_t __attribute__((aligned(8))) pQPb2=0;
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
155 static uint8_t __attribute__((aligned(8))) tempBlocks[8*16*2]; //used for the horizontal code
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
156 #else
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
157 static uint64_t packedYOffset= 0x0000000000000000LL;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
158 static uint64_t packedYScale= 0x0100010001000100LL;
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
159 static uint8_t tempBlocks[8*16*2]; //used for the horizontal code
148
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
160 #endif
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
161
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
162 int hFlatnessThreshold= 56 - 16;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
163 int vFlatnessThreshold= 56 - 16;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
164
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
165 //amount of "black" u r willing to loose to get a brightness corrected picture
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
166 double maxClippedThreshold= 0.01;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
167
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
168 int maxAllowedY=234;
112
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
169 int minAllowedY=16;
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
170
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
171 static struct PPFilter filters[]=
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
172 {
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
173 {"hb", "hdeblock", 1, 1, 3, H_DEBLOCK},
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
174 {"vb", "vdeblock", 1, 2, 4, V_DEBLOCK},
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
175 {"vr", "rkvdeblock", 1, 2, 4, H_RK1_FILTER},
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
176 {"h1", "x1hdeblock", 1, 1, 3, H_X1_FILTER},
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
177 {"v1", "x1vdeblock", 1, 2, 4, V_X1_FILTER},
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
178 {"dr", "dering", 1, 5, 6, DERING},
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
179 {"al", "autolevels", 0, 1, 2, LEVEL_FIX},
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
180 {"lb", "linblenddeint", 0, 1, 6, LINEAR_BLEND_DEINT_FILTER},
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
181 {"li", "linipoldeint", 0, 1, 6, LINEAR_IPOL_DEINT_FILTER},
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
182 {"ci", "cubicipoldeint", 0, 1, 6, CUBIC_IPOL_DEINT_FILTER},
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
183 {"md", "mediandeint", 0, 1, 6, MEDIAN_DEINT_FILTER},
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
184 {"tn", "tmpnoise", 1, 7, 8, TEMP_NOISE_FILTER},
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
185 {NULL, NULL,0,0,0,0} //End Marker
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
186 };
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
187
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
188 static char *replaceTable[]=
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
189 {
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
190 "default", "hdeblock:a,vdeblock:a,dering:a,autolevels,tmpnoise:a:150:200:400",
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
191 "de", "hdeblock:a,vdeblock:a,dering:a,autolevels,tmpnoise:a:150:200:400",
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
192 "fast", "x1hdeblock:a,x1vdeblock:a,dering:a,autolevels,tmpnoise:a:150:200:400",
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
193 "fa", "x1hdeblock:a,x1vdeblock:a,dering:a,autolevels,tmpnoise:a:150:200:400",
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
194 NULL //End Marker
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
195 };
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
196
148
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
197 #ifdef HAVE_MMX
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
198 static inline void unusedVariableWarningFixer()
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
199 {
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
200 if(
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
201 packedYOffset + packedYScale + w05 + w20 + w1400 + bm00000001 + bm00010000
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
202 + bm00001000 + bm10000000 + bm10000001 + bm11000011 + bm00000011 + bm11111110
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
203 + bm11000000 + bm00011000 + bm00110011 + bm11001100 + b00 + b01 + b02 + b0F
148
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
204 + bFF + b20 + b04+ b08 + pQPb2 + b80 + b7E + b7C + b3F + temp0 + temp1 + temp2 + temp3 + temp4
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
205 + temp5 + pQPb== 0) b00=0;
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
206 }
148
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
207 #endif
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
208
107
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
209 #ifdef TIMING
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
210 static inline long long rdtsc()
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
211 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
212 long long l;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
213 asm volatile( "rdtsc\n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
214 : "=A" (l)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
215 );
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
216 // printf("%d\n", int(l/1000));
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
217 return l;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
218 }
102
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
219 #endif
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
220
102
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
221 #ifdef HAVE_MMX2
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
222 static inline void prefetchnta(void *p)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
223 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
224 asm volatile( "prefetchnta (%0)\n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
225 : : "r" (p)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
226 );
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
227 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
228
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
229 static inline void prefetcht0(void *p)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
230 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
231 asm volatile( "prefetcht0 (%0)\n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
232 : : "r" (p)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
233 );
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
234 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
235
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
236 static inline void prefetcht1(void *p)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
237 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
238 asm volatile( "prefetcht1 (%0)\n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
239 : : "r" (p)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
240 );
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
241 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
242
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
243 static inline void prefetcht2(void *p)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
244 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
245 asm volatile( "prefetcht2 (%0)\n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
246 : : "r" (p)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
247 );
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
248 }
102
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
249 #endif
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
250
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
251 //FIXME? |255-0| = 1 (shouldnt be a problem ...)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
252 /**
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
253 * Check if the middle 8x8 Block in the given 8x16 block is flat
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
254 */
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
255 static inline int isVertDC(uint8_t src[], int stride){
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
256 int numEq= 0;
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
257 #ifndef HAVE_MMX
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
258 int y;
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
259 #endif
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
260 src+= stride*4; // src points to begin of the 8x8 Block
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
261 #ifdef HAVE_MMX
119
b2f0e40866b1 optimizations (+2% speedup)
michael
parents: 118
diff changeset
262 asm volatile(
b2f0e40866b1 optimizations (+2% speedup)
michael
parents: 118
diff changeset
263 "leal (%1, %2), %%eax \n\t"
b2f0e40866b1 optimizations (+2% speedup)
michael
parents: 118
diff changeset
264 "leal (%%eax, %2, 4), %%ebx \n\t"
b2f0e40866b1 optimizations (+2% speedup)
michael
parents: 118
diff changeset
265 // 0 1 2 3 4 5 6 7 8 9
b2f0e40866b1 optimizations (+2% speedup)
michael
parents: 118
diff changeset
266 // %1 eax eax+%2 eax+2%2 %1+4%2 ebx ebx+%2 ebx+2%2 %1+8%2 ebx+4%2
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
267 "movq b7E, %%mm7 \n\t" // mm7 = 0x7F
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
268 "movq b7C, %%mm6 \n\t" // mm6 = 0x7D
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
269 "movq (%1), %%mm0 \n\t"
119
b2f0e40866b1 optimizations (+2% speedup)
michael
parents: 118
diff changeset
270 "movq (%%eax), %%mm1 \n\t"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
271 "psubb %%mm1, %%mm0 \n\t" // mm0 = differnece
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
272 "paddb %%mm7, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
273 "pcmpgtb %%mm6, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
274
119
b2f0e40866b1 optimizations (+2% speedup)
michael
parents: 118
diff changeset
275 "movq (%%eax,%2), %%mm2 \n\t"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
276 "psubb %%mm2, %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
277 "paddb %%mm7, %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
278 "pcmpgtb %%mm6, %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
279 "paddb %%mm1, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
280
119
b2f0e40866b1 optimizations (+2% speedup)
michael
parents: 118
diff changeset
281 "movq (%%eax, %2, 2), %%mm1 \n\t"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
282 "psubb %%mm1, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
283 "paddb %%mm7, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
284 "pcmpgtb %%mm6, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
285 "paddb %%mm2, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
286
119
b2f0e40866b1 optimizations (+2% speedup)
michael
parents: 118
diff changeset
287 "movq (%1, %2, 4), %%mm2 \n\t"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
288 "psubb %%mm2, %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
289 "paddb %%mm7, %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
290 "pcmpgtb %%mm6, %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
291 "paddb %%mm1, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
292
119
b2f0e40866b1 optimizations (+2% speedup)
michael
parents: 118
diff changeset
293 "movq (%%ebx), %%mm1 \n\t"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
294 "psubb %%mm1, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
295 "paddb %%mm7, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
296 "pcmpgtb %%mm6, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
297 "paddb %%mm2, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
298
119
b2f0e40866b1 optimizations (+2% speedup)
michael
parents: 118
diff changeset
299 "movq (%%ebx, %2), %%mm2 \n\t"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
300 "psubb %%mm2, %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
301 "paddb %%mm7, %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
302 "pcmpgtb %%mm6, %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
303 "paddb %%mm1, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
304
119
b2f0e40866b1 optimizations (+2% speedup)
michael
parents: 118
diff changeset
305 "movq (%%ebx, %2, 2), %%mm1 \n\t"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
306 "psubb %%mm1, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
307 "paddb %%mm7, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
308 "pcmpgtb %%mm6, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
309 "paddb %%mm2, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
310
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
311 " \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
312 "movq %%mm0, %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
313 "psrlw $8, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
314 "paddb %%mm1, %%mm0 \n\t"
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
315 #ifdef HAVE_MMX2
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
316 "pshufw $0xF9, %%mm0, %%mm1 \n\t"
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
317 "paddb %%mm1, %%mm0 \n\t"
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
318 "pshufw $0xFE, %%mm0, %%mm1 \n\t"
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
319 #else
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
320 "movq %%mm0, %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
321 "psrlq $16, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
322 "paddb %%mm1, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
323 "movq %%mm0, %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
324 "psrlq $32, %%mm0 \n\t"
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
325 #endif
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
326 "paddb %%mm1, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
327 "movd %%mm0, %0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
328 : "=r" (numEq)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
329 : "r" (src), "r" (stride)
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
330 : "%eax", "%ebx"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
331 );
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
332
119
b2f0e40866b1 optimizations (+2% speedup)
michael
parents: 118
diff changeset
333 numEq= (256 - numEq) &0xFF;
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
334
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
335 #else
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
336 for(y=0; y<BLOCK_SIZE-1; y++)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
337 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
338 if(((src[0] - src[0+stride] + 1)&0xFFFF) < 3) numEq++;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
339 if(((src[1] - src[1+stride] + 1)&0xFFFF) < 3) numEq++;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
340 if(((src[2] - src[2+stride] + 1)&0xFFFF) < 3) numEq++;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
341 if(((src[3] - src[3+stride] + 1)&0xFFFF) < 3) numEq++;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
342 if(((src[4] - src[4+stride] + 1)&0xFFFF) < 3) numEq++;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
343 if(((src[5] - src[5+stride] + 1)&0xFFFF) < 3) numEq++;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
344 if(((src[6] - src[6+stride] + 1)&0xFFFF) < 3) numEq++;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
345 if(((src[7] - src[7+stride] + 1)&0xFFFF) < 3) numEq++;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
346 src+= stride;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
347 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
348 #endif
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
349 /* if(abs(numEq - asmEq) > 0)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
350 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
351 printf("\nasm:%d c:%d\n", asmEq, numEq);
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
352 for(int y=0; y<8; y++)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
353 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
354 for(int x=0; x<8; x++)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
355 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
356 printf("%d ", temp[x + y*stride]);
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
357 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
358 printf("\n");
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
359 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
360 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
361 */
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
362 // for(int i=0; i<numEq/8; i++) src[i]=255;
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
363 return (numEq > vFlatnessThreshold) ? 1 : 0;
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
364 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
365
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
366 static inline int isVertMinMaxOk(uint8_t src[], int stride, int QP)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
367 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
368 #ifdef HAVE_MMX
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
369 int isOk;
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
370 src+= stride*3;
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
371 asm volatile(
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
372 // "int $3 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
373 "movq (%1, %2), %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
374 "movq (%1, %2, 8), %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
375 "movq %%mm0, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
376 "psubusb %%mm1, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
377 "psubusb %%mm2, %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
378 "por %%mm1, %%mm0 \n\t" // ABS Diff
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
379
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
380 "movq pQPb, %%mm7 \n\t" // QP,..., QP
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
381 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
382 "psubusb %%mm7, %%mm0 \n\t" // Diff <= 2QP -> 0
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
383 "pcmpeqd b00, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
384 "psrlq $16, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
385 "pcmpeqd bFF, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
386 // "movd %%mm0, (%1, %2, 4)\n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
387 "movd %%mm0, %0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
388 : "=r" (isOk)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
389 : "r" (src), "r" (stride)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
390 );
120
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
391 return isOk;
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
392 #else
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
393
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
394 int isOk2= 1;
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
395 int x;
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
396 src+= stride*3;
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
397 for(x=0; x<BLOCK_SIZE; x++)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
398 {
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
399 if(abs((int)src[x + stride] - (int)src[x + (stride<<3)]) > 2*QP) isOk2=0;
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
400 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
401 /* if(isOk && !isOk2 || !isOk && isOk2)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
402 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
403 printf("\nasm:%d c:%d QP:%d\n", isOk, isOk2, QP);
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
404 for(int y=0; y<9; y++)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
405 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
406 for(int x=0; x<8; x++)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
407 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
408 printf("%d ", src[x + y*stride]);
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
409 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
410 printf("\n");
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
411 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
412 } */
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
413
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
414 return isOk2;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
415 #endif
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
416
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
417 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
418
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
419 /**
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
420 * Do a vertical low pass filter on the 8x16 block (only write to the 8x8 block in the middle)
107
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
421 * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
422 */
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
423 static inline void doVertLowPass(uint8_t *src, int stride, int QP)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
424 {
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
425 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
426 src+= stride*3;
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
427 asm volatile( //"movv %0 %1 %2\n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
428 "movq pQPb, %%mm0 \n\t" // QP,..., QP
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
429
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
430 "movq (%0), %%mm6 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
431 "movq (%0, %1), %%mm5 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
432 "movq %%mm5, %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
433 "movq %%mm6, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
434 "psubusb %%mm6, %%mm5 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
435 "psubusb %%mm1, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
436 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
437 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
438 "pcmpeqb b00, %%mm2 \n\t" // diff <= QP -> FF
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
439
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
440 "pand %%mm2, %%mm6 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
441 "pandn %%mm1, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
442 "por %%mm2, %%mm6 \n\t"// First Line to Filter
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
443
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
444 "movq (%0, %1, 8), %%mm5 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
445 "leal (%0, %1, 4), %%eax \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
446 "leal (%0, %1, 8), %%ebx \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
447 "subl %1, %%ebx \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
448 "addl %1, %0 \n\t" // %0 points to line 1 not 0
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
449 "movq (%0, %1, 8), %%mm7 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
450 "movq %%mm5, %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
451 "movq %%mm7, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
452 "psubusb %%mm7, %%mm5 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
453 "psubusb %%mm1, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
454 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
455 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
456 "pcmpeqb b00, %%mm2 \n\t" // diff <= QP -> FF
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
457
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
458 "pand %%mm2, %%mm7 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
459 "pandn %%mm1, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
460 "por %%mm2, %%mm7 \n\t" // First Line to Filter
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
461
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
462
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
463 // 1 2 3 4 5 6 7 8
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
464 // %0 %0+%1 %0+2%1 eax %0+4%1 eax+2%1 ebx eax+4%1
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
465 // 6 4 2 2 1 1
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
466 // 6 4 4 2
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
467 // 6 8 2
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
468
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
469 "movq (%0, %1), %%mm0 \n\t" // 1
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
470 "movq %%mm0, %%mm1 \n\t" // 1
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
471 PAVGB(%%mm6, %%mm0) //1 1 /2
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
472 PAVGB(%%mm6, %%mm0) //3 1 /4
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
473
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
474 "movq (%0, %1, 4), %%mm2 \n\t" // 1
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
475 "movq %%mm2, %%mm5 \n\t" // 1
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
476 PAVGB((%%eax), %%mm2) // 11 /2
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
477 PAVGB((%0, %1, 2), %%mm2) // 211 /4
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
478 "movq %%mm2, %%mm3 \n\t" // 211 /4
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
479 "movq (%0), %%mm4 \n\t" // 1
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
480 PAVGB(%%mm4, %%mm3) // 4 211 /8
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
481 PAVGB(%%mm0, %%mm3) //642211 /16
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
482 "movq %%mm3, (%0) \n\t" // X
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
483 // mm1=2 mm2=3(211) mm4=1 mm5=5 mm6=0 mm7=9
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
484 "movq %%mm1, %%mm0 \n\t" // 1
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
485 PAVGB(%%mm6, %%mm0) //1 1 /2
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
486 "movq %%mm4, %%mm3 \n\t" // 1
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
487 PAVGB((%0,%1,2), %%mm3) // 1 1 /2
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
488 PAVGB((%%eax,%1,2), %%mm5) // 11 /2
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
489 PAVGB((%%eax), %%mm5) // 211 /4
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
490 PAVGB(%%mm5, %%mm3) // 2 2211 /8
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
491 PAVGB(%%mm0, %%mm3) //4242211 /16
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
492 "movq %%mm3, (%0,%1) \n\t" // X
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
493 // mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
494 PAVGB(%%mm4, %%mm6) //11 /2
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
495 "movq (%%ebx), %%mm0 \n\t" // 1
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
496 PAVGB((%%eax, %1, 2), %%mm0) // 11/2
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
497 "movq %%mm0, %%mm3 \n\t" // 11/2
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
498 PAVGB(%%mm1, %%mm0) // 2 11/4
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
499 PAVGB(%%mm6, %%mm0) //222 11/8
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
500 PAVGB(%%mm2, %%mm0) //22242211/16
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
501 "movq (%0, %1, 2), %%mm2 \n\t" // 1
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
502 "movq %%mm0, (%0, %1, 2) \n\t" // X
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
503 // mm1=2 mm2=3 mm3=6(11) mm4=1 mm5=4(211) mm6=0(11) mm7=9
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
504 "movq (%%eax, %1, 4), %%mm0 \n\t" // 1
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
505 PAVGB((%%ebx), %%mm0) // 11 /2
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
506 PAVGB(%%mm0, %%mm6) //11 11 /4
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
507 PAVGB(%%mm1, %%mm4) // 11 /2
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
508 PAVGB(%%mm2, %%mm1) // 11 /2
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
509 PAVGB(%%mm1, %%mm6) //1122 11 /8
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
510 PAVGB(%%mm5, %%mm6) //112242211 /16
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
511 "movq (%%eax), %%mm5 \n\t" // 1
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
512 "movq %%mm6, (%%eax) \n\t" // X
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
513 // mm0=7(11) mm1=2(11) mm2=3 mm3=6(11) mm4=1(11) mm5=4 mm7=9
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
514 "movq (%%eax, %1, 4), %%mm6 \n\t" // 1
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
515 PAVGB(%%mm7, %%mm6) // 11 /2
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
516 PAVGB(%%mm4, %%mm6) // 11 11 /4
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
517 PAVGB(%%mm3, %%mm6) // 11 2211 /8
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
518 PAVGB(%%mm5, %%mm2) // 11 /2
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
519 "movq (%0, %1, 4), %%mm4 \n\t" // 1
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
520 PAVGB(%%mm4, %%mm2) // 112 /4
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
521 PAVGB(%%mm2, %%mm6) // 112242211 /16
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
522 "movq %%mm6, (%0, %1, 4) \n\t" // X
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
523 // mm0=7(11) mm1=2(11) mm2=3(112) mm3=6(11) mm4=5 mm5=4 mm7=9
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
524 PAVGB(%%mm7, %%mm1) // 11 2 /4
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
525 PAVGB(%%mm4, %%mm5) // 11 /2
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
526 PAVGB(%%mm5, %%mm0) // 11 11 /4
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
527 "movq (%%eax, %1, 2), %%mm6 \n\t" // 1
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
528 PAVGB(%%mm6, %%mm1) // 11 4 2 /8
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
529 PAVGB(%%mm0, %%mm1) // 11224222 /16
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
530 "movq %%mm1, (%%eax, %1, 2) \n\t" // X
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
531 // mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
532 PAVGB((%%ebx), %%mm2) // 112 4 /8
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
533 "movq (%%eax, %1, 4), %%mm0 \n\t" // 1
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
534 PAVGB(%%mm0, %%mm6) // 1 1 /2
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
535 PAVGB(%%mm7, %%mm6) // 1 12 /4
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
536 PAVGB(%%mm2, %%mm6) // 1122424 /4
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
537 "movq %%mm6, (%%ebx) \n\t" // X
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
538 // mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
539 PAVGB(%%mm7, %%mm5) // 11 2 /4
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
540 PAVGB(%%mm7, %%mm5) // 11 6 /8
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
541
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
542 PAVGB(%%mm3, %%mm0) // 112 /4
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
543 PAVGB(%%mm0, %%mm5) // 112246 /16
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
544 "movq %%mm5, (%%eax, %1, 4) \n\t" // X
140
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
545 "subl %1, %0 \n\t"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
546
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
547 :
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
548 : "r" (src), "r" (stride)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
549 : "%eax", "%ebx"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
550 );
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
551 #else
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
552 const int l1= stride;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
553 const int l2= stride + l1;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
554 const int l3= stride + l2;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
555 const int l4= stride + l3;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
556 const int l5= stride + l4;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
557 const int l6= stride + l5;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
558 const int l7= stride + l6;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
559 const int l8= stride + l7;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
560 const int l9= stride + l8;
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
561 int x;
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
562 src+= stride*3;
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
563 for(x=0; x<BLOCK_SIZE; x++)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
564 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
565 const int first= ABS(src[0] - src[l1]) < QP ? src[0] : src[l1];
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
566 const int last= ABS(src[l8] - src[l9]) < QP ? src[l9] : src[l8];
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
567
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
568 int sums[9];
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
569 sums[0] = first + src[l1];
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
570 sums[1] = src[l1] + src[l2];
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
571 sums[2] = src[l2] + src[l3];
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
572 sums[3] = src[l3] + src[l4];
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
573 sums[4] = src[l4] + src[l5];
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
574 sums[5] = src[l5] + src[l6];
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
575 sums[6] = src[l6] + src[l7];
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
576 sums[7] = src[l7] + src[l8];
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
577 sums[8] = src[l8] + last;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
578
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
579 src[l1]= ((sums[0]<<2) + ((first + sums[2])<<1) + sums[4] + 8)>>4;
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
580 src[l2]= ((src[l2]<<2) + ((first + sums[0] + sums[3])<<1) + sums[5] + 8)>>4;
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
581 src[l3]= ((src[l3]<<2) + ((first + sums[1] + sums[4])<<1) + sums[6] + 8)>>4;
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
582 src[l4]= ((src[l4]<<2) + ((sums[2] + sums[5])<<1) + sums[0] + sums[7] + 8)>>4;
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
583 src[l5]= ((src[l5]<<2) + ((sums[3] + sums[6])<<1) + sums[1] + sums[8] + 8)>>4;
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
584 src[l6]= ((src[l6]<<2) + ((last + sums[7] + sums[4])<<1) + sums[2] + 8)>>4;
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
585 src[l7]= (((last + src[l7])<<2) + ((src[l8] + sums[5])<<1) + sums[3] + 8)>>4;
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
586 src[l8]= ((sums[8]<<2) + ((last + sums[6])<<1) + sums[4] + 8)>>4;
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
587
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
588 src++;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
589 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
590
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
591 #endif
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
592 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
593
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
594 /**
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
595 * Experimental implementation of the filter (Algorithm 1) described in a paper from Ramkishor & Karandikar
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
596 * values are correctly clipped (MMX2)
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
597 * values are wraparound (C)
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
598 * conclusion: its fast, but introduces ugly horizontal patterns if there is a continious gradient
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
599 0 8 16 24
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
600 x = 8
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
601 x/2 = 4
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
602 x/8 = 1
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
603 1 12 12 23
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
604 */
99
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
605 static inline void vertRK1Filter(uint8_t *src, int stride, int QP)
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
606 {
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
607 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
608 src+= stride*3;
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
609 // FIXME rounding
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
610 asm volatile(
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
611 "pxor %%mm7, %%mm7 \n\t" // 0
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
612 "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
613 "leal (%0, %1), %%eax \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
614 "leal (%%eax, %1, 4), %%ebx \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
615 // 0 1 2 3 4 5 6 7 8 9
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
616 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
617 "movq pQPb, %%mm0 \n\t" // QP,..., QP
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
618 "movq %%mm0, %%mm1 \n\t" // QP,..., QP
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
619 "paddusb b02, %%mm0 \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
620 "psrlw $2, %%mm0 \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
621 "pand b3F, %%mm0 \n\t" // QP/4,..., QP/4
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
622 "paddusb %%mm1, %%mm0 \n\t" // QP*1.25 ...
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
623 "movq (%0, %1, 4), %%mm2 \n\t" // line 4
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
624 "movq (%%ebx), %%mm3 \n\t" // line 5
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
625 "movq %%mm2, %%mm4 \n\t" // line 4
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
626 "pcmpeqb %%mm5, %%mm5 \n\t" // -1
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
627 "pxor %%mm2, %%mm5 \n\t" // -line 4 - 1
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
628 PAVGB(%%mm3, %%mm5)
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
629 "paddb %%mm6, %%mm5 \n\t" // (l5-l4)/2
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
630 "psubusb %%mm3, %%mm4 \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
631 "psubusb %%mm2, %%mm3 \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
632 "por %%mm3, %%mm4 \n\t" // |l4 - l5|
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
633 "psubusb %%mm0, %%mm4 \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
634 "pcmpeqb %%mm7, %%mm4 \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
635 "pand %%mm4, %%mm5 \n\t" // d/2
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
636
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
637 // "paddb %%mm6, %%mm2 \n\t" // line 4 + 0x80
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
638 "paddb %%mm5, %%mm2 \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
639 // "psubb %%mm6, %%mm2 \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
640 "movq %%mm2, (%0,%1, 4) \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
641
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
642 "movq (%%ebx), %%mm2 \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
643 // "paddb %%mm6, %%mm2 \n\t" // line 5 + 0x80
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
644 "psubb %%mm5, %%mm2 \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
645 // "psubb %%mm6, %%mm2 \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
646 "movq %%mm2, (%%ebx) \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
647
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
648 "paddb %%mm6, %%mm5 \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
649 "psrlw $2, %%mm5 \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
650 "pand b3F, %%mm5 \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
651 "psubb b20, %%mm5 \n\t" // (l5-l4)/8
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
652
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
653 "movq (%%eax, %1, 2), %%mm2 \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
654 "paddb %%mm6, %%mm2 \n\t" // line 3 + 0x80
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
655 "paddsb %%mm5, %%mm2 \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
656 "psubb %%mm6, %%mm2 \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
657 "movq %%mm2, (%%eax, %1, 2) \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
658
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
659 "movq (%%ebx, %1), %%mm2 \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
660 "paddb %%mm6, %%mm2 \n\t" // line 6 + 0x80
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
661 "psubsb %%mm5, %%mm2 \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
662 "psubb %%mm6, %%mm2 \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
663 "movq %%mm2, (%%ebx, %1) \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
664
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
665 :
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
666 : "r" (src), "r" (stride)
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
667 : "%eax", "%ebx"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
668 );
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
669 #else
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
670 const int l1= stride;
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
671 const int l2= stride + l1;
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
672 const int l3= stride + l2;
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
673 const int l4= stride + l3;
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
674 const int l5= stride + l4;
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
675 const int l6= stride + l5;
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
676 // const int l7= stride + l6;
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
677 // const int l8= stride + l7;
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
678 // const int l9= stride + l8;
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
679 int x;
141
626bfabff1f5 c speedup (x1, rk1 filters)
michael
parents: 140
diff changeset
680 const int QP15= QP + (QP>>2);
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
681 src+= stride*3;
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
682 for(x=0; x<BLOCK_SIZE; x++)
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
683 {
141
626bfabff1f5 c speedup (x1, rk1 filters)
michael
parents: 140
diff changeset
684 const int v = (src[x+l5] - src[x+l4]);
626bfabff1f5 c speedup (x1, rk1 filters)
michael
parents: 140
diff changeset
685 if(ABS(v) < QP15)
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
686 {
141
626bfabff1f5 c speedup (x1, rk1 filters)
michael
parents: 140
diff changeset
687 src[x+l3] +=v>>3;
626bfabff1f5 c speedup (x1, rk1 filters)
michael
parents: 140
diff changeset
688 src[x+l4] +=v>>1;
626bfabff1f5 c speedup (x1, rk1 filters)
michael
parents: 140
diff changeset
689 src[x+l5] -=v>>1;
626bfabff1f5 c speedup (x1, rk1 filters)
michael
parents: 140
diff changeset
690 src[x+l6] -=v>>3;
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
691
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
692 }
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
693 }
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
694
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
695 #endif
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
696 }
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
697
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
698 /**
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
699 * Experimental Filter 1
99
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
700 * will not damage linear gradients
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
701 * Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
702 * can only smooth blocks at the expected locations (it cant smooth them if they did move)
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
703 * MMX2 version does correct clipping C version doesnt
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
704 */
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
705 static inline void vertX1Filter(uint8_t *src, int stride, int QP)
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
706 {
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
707 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
708 src+= stride*3;
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
709
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
710 asm volatile(
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
711 "pxor %%mm7, %%mm7 \n\t" // 0
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
712 // "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
713 "leal (%0, %1), %%eax \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
714 "leal (%%eax, %1, 4), %%ebx \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
715 // 0 1 2 3 4 5 6 7 8 9
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
716 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
717 "movq (%%eax, %1, 2), %%mm0 \n\t" // line 3
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
718 "movq (%0, %1, 4), %%mm1 \n\t" // line 4
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
719 "movq %%mm1, %%mm2 \n\t" // line 4
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
720 "psubusb %%mm0, %%mm1 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
721 "psubusb %%mm2, %%mm0 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
722 "por %%mm1, %%mm0 \n\t" // |l2 - l3|
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
723 "movq (%%ebx), %%mm3 \n\t" // line 5
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
724 "movq (%%ebx, %1), %%mm4 \n\t" // line 6
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
725 "movq %%mm3, %%mm5 \n\t" // line 5
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
726 "psubusb %%mm4, %%mm3 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
727 "psubusb %%mm5, %%mm4 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
728 "por %%mm4, %%mm3 \n\t" // |l5 - l6|
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
729 PAVGB(%%mm3, %%mm0) // (|l2 - l3| + |l5 - l6|)/2
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
730 "movq %%mm2, %%mm1 \n\t" // line 4
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
731 "psubusb %%mm5, %%mm2 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
732 "movq %%mm2, %%mm4 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
733 "pcmpeqb %%mm7, %%mm2 \n\t" // (l4 - l5) <= 0 ? -1 : 0
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
734 "psubusb %%mm1, %%mm5 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
735 "por %%mm5, %%mm4 \n\t" // |l4 - l5|
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
736 "psubusb %%mm0, %%mm4 \n\t" //d = MAX(0, |l4-l5| - (|l2-l3| + |l5-l6|)/2)
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
737 "movq %%mm4, %%mm3 \n\t" // d
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
738 "psubusb pQPb, %%mm4 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
739 "pcmpeqb %%mm7, %%mm4 \n\t" // d <= QP ? -1 : 0
99
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
740 "psubusb b01, %%mm3 \n\t"
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
741 "pand %%mm4, %%mm3 \n\t" // d <= QP ? d : 0
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
742
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
743 PAVGB(%%mm7, %%mm3) // d/2
99
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
744 "movq %%mm3, %%mm1 \n\t" // d/2
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
745 PAVGB(%%mm7, %%mm3) // d/4
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
746 PAVGB(%%mm1, %%mm3) // 3*d/8
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
747
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
748 "movq (%0, %1, 4), %%mm0 \n\t" // line 4
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
749 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
750 "psubusb %%mm3, %%mm0 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
751 "pxor %%mm2, %%mm0 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
752 "movq %%mm0, (%0, %1, 4) \n\t" // line 4
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
753
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
754 "movq (%%ebx), %%mm0 \n\t" // line 5
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
755 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
756 "paddusb %%mm3, %%mm0 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
757 "pxor %%mm2, %%mm0 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
758 "movq %%mm0, (%%ebx) \n\t" // line 5
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
759
99
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
760 PAVGB(%%mm7, %%mm1) // d/4
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
761
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
762 "movq (%%eax, %1, 2), %%mm0 \n\t" // line 3
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
763 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
99
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
764 "psubusb %%mm1, %%mm0 \n\t"
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
765 "pxor %%mm2, %%mm0 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
766 "movq %%mm0, (%%eax, %1, 2) \n\t" // line 3
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
767
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
768 "movq (%%ebx, %1), %%mm0 \n\t" // line 6
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
769 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
99
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
770 "paddusb %%mm1, %%mm0 \n\t"
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
771 "pxor %%mm2, %%mm0 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
772 "movq %%mm0, (%%ebx, %1) \n\t" // line 6
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
773
99
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
774 PAVGB(%%mm7, %%mm1) // d/8
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
775
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
776 "movq (%%eax, %1), %%mm0 \n\t" // line 2
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
777 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l2-1 : l2
99
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
778 "psubusb %%mm1, %%mm0 \n\t"
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
779 "pxor %%mm2, %%mm0 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
780 "movq %%mm0, (%%eax, %1) \n\t" // line 2
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
781
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
782 "movq (%%ebx, %1, 2), %%mm0 \n\t" // line 7
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
783 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l7-1 : l7
99
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
784 "paddusb %%mm1, %%mm0 \n\t"
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
785 "pxor %%mm2, %%mm0 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
786 "movq %%mm0, (%%ebx, %1, 2) \n\t" // line 7
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
787
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
788 :
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
789 : "r" (src), "r" (stride)
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
790 : "%eax", "%ebx"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
791 );
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
792 #else
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
793
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
794 const int l1= stride;
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
795 const int l2= stride + l1;
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
796 const int l3= stride + l2;
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
797 const int l4= stride + l3;
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
798 const int l5= stride + l4;
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
799 const int l6= stride + l5;
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
800 const int l7= stride + l6;
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
801 // const int l8= stride + l7;
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
802 // const int l9= stride + l8;
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
803 int x;
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
804
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
805 src+= stride*3;
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
806 for(x=0; x<BLOCK_SIZE; x++)
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
807 {
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
808 int a= src[l3] - src[l4];
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
809 int b= src[l4] - src[l5];
99
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
810 int c= src[l5] - src[l6];
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
811
141
626bfabff1f5 c speedup (x1, rk1 filters)
michael
parents: 140
diff changeset
812 int d= ABS(b) - ((ABS(a) + ABS(c))>>1);
626bfabff1f5 c speedup (x1, rk1 filters)
michael
parents: 140
diff changeset
813 d= MAX(d, 0);
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
814
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
815 if(d < QP)
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
816 {
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
817 int v = d * SIGN(-b);
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
818
141
626bfabff1f5 c speedup (x1, rk1 filters)
michael
parents: 140
diff changeset
819 src[l2] +=v>>3;
626bfabff1f5 c speedup (x1, rk1 filters)
michael
parents: 140
diff changeset
820 src[l3] +=v>>2;
626bfabff1f5 c speedup (x1, rk1 filters)
michael
parents: 140
diff changeset
821 src[l4] +=(3*v)>>3;
626bfabff1f5 c speedup (x1, rk1 filters)
michael
parents: 140
diff changeset
822 src[l5] -=(3*v)>>3;
626bfabff1f5 c speedup (x1, rk1 filters)
michael
parents: 140
diff changeset
823 src[l6] -=v>>2;
626bfabff1f5 c speedup (x1, rk1 filters)
michael
parents: 140
diff changeset
824 src[l7] -=v>>3;
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
825
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
826 }
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
827 src++;
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
828 }
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
829 /*
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
830 const int l1= stride;
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
831 const int l2= stride + l1;
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
832 const int l3= stride + l2;
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
833 const int l4= stride + l3;
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
834 const int l5= stride + l4;
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
835 const int l6= stride + l5;
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
836 const int l7= stride + l6;
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
837 const int l8= stride + l7;
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
838 const int l9= stride + l8;
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
839 for(int x=0; x<BLOCK_SIZE; x++)
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
840 {
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
841 int v2= src[l2];
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
842 int v3= src[l3];
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
843 int v4= src[l4];
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
844 int v5= src[l5];
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
845 int v6= src[l6];
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
846 int v7= src[l7];
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
847
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
848 if(ABS(v4-v5)<QP && ABS(v4-v5) - (ABS(v3-v4) + ABS(v5-v6))>0 )
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
849 {
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
850 src[l3] = (6*v2 + 4*v3 + 3*v4 + 2*v5 + v6 )/16;
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
851 src[l4] = (3*v2 + 3*v3 + 4*v4 + 3*v5 + 2*v6 + v7 )/16;
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
852 src[l5] = (1*v2 + 2*v3 + 3*v4 + 4*v5 + 3*v6 + 3*v7)/16;
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
853 src[l6] = ( 1*v3 + 2*v4 + 3*v5 + 4*v6 + 6*v7)/16;
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
854 }
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
855 src++;
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
856 }
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
857 */
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
858 #endif
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
859 }
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
860
99
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
861 /**
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
862 * Experimental Filter 1 (Horizontal)
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
863 * will not damage linear gradients
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
864 * Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
865 * can only smooth blocks at the expected locations (it cant smooth them if they did move)
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
866 * MMX2 version does correct clipping C version doesnt
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
867 * not identical with the vertical one
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
868 */
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
869 static inline void horizX1Filter(uint8_t *src, int stride, int QP)
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
870 {
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
871 int y;
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
872 static uint64_t *lut= NULL;
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
873 if(lut==NULL)
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
874 {
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
875 int i;
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
876 lut= (uint64_t*)memalign(8, 256*8);
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
877 for(i=0; i<256; i++)
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
878 {
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
879 int v= i < 128 ? 2*i : 2*(i-256);
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
880 /*
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
881 //Simulate 112242211 9-Tap filter
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
882 uint64_t a= (v/16) & 0xFF;
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
883 uint64_t b= (v/8) & 0xFF;
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
884 uint64_t c= (v/4) & 0xFF;
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
885 uint64_t d= (3*v/8) & 0xFF;
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
886 */
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
887 //Simulate piecewise linear interpolation
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
888 uint64_t a= (v/16) & 0xFF;
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
889 uint64_t b= (v*3/16) & 0xFF;
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
890 uint64_t c= (v*5/16) & 0xFF;
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
891 uint64_t d= (7*v/16) & 0xFF;
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
892 uint64_t A= (0x100 - a)&0xFF;
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
893 uint64_t B= (0x100 - b)&0xFF;
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
894 uint64_t C= (0x100 - c)&0xFF;
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
895 uint64_t D= (0x100 - c)&0xFF;
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
896
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
897 lut[i] = (a<<56) | (b<<48) | (c<<40) | (d<<32) |
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
898 (D<<24) | (C<<16) | (B<<8) | (A);
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
899 //lut[i] = (v<<32) | (v<<24);
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
900 }
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
901 }
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
902
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
903 #if 0
99
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
904 asm volatile(
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
905 "pxor %%mm7, %%mm7 \n\t" // 0
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
906 // "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
907 "leal (%0, %1), %%eax \n\t"
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
908 "leal (%%eax, %1, 4), %%ebx \n\t"
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
909
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
910 "movq b80, %%mm6 \n\t"
100
1d1182345591 minor QP scaling & auto-contrast bugfix
michael
parents: 99
diff changeset
911 "movd pQPb, %%mm5 \n\t" // QP
99
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
912 "movq %%mm5, %%mm4 \n\t"
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
913 "paddusb %%mm5, %%mm5 \n\t" // 2QP
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
914 "paddusb %%mm5, %%mm4 \n\t" // 3QP
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
915 "pxor %%mm5, %%mm5 \n\t" // 0
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
916 "psubb %%mm4, %%mm5 \n\t" // -3QP
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
917 "por bm11111110, %%mm5 \n\t" // ...,FF,FF,-3QP
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
918 "psllq $24, %%mm5 \n\t"
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
919
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
920 // 0 1 2 3 4 5 6 7 8 9
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
921 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
922
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
923 #define HX1old(a) \
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
924 "movd " #a ", %%mm0 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
925 "movd 4" #a ", %%mm1 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
926 "punpckldq %%mm1, %%mm0 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
927 "movq %%mm0, %%mm1 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
928 "movq %%mm0, %%mm2 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
929 "psrlq $8, %%mm1 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
930 "psubusb %%mm1, %%mm2 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
931 "psubusb %%mm0, %%mm1 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
932 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
933 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
934 "pshufw $0x00, %%mm1, %%mm3 \n\t" /* p´5 = |p1 - p2| */\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
935 PAVGB(%%mm1, %%mm3) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
936 "psrlq $16, %%mm3 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
937 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
938 "paddb %%mm5, %%mm1 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
939 "psubusb %%mm5, %%mm1 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
940 PAVGB(%%mm7, %%mm1)\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
941 "pxor %%mm2, %%mm1 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
942 "psubb %%mm2, %%mm1 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
943 "psrlq $24, %%mm1 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
944 "movd %%mm1, %%ecx \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
945 "paddb %%mm6, %%mm0 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
946 "paddsb (%3, %%ecx, 8), %%mm0 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
947 "paddb %%mm6, %%mm0 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
948 "movq %%mm0, " #a " \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
949
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
950 /*
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
951 HX1old((%0))
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
952 HX1old((%%eax))
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
953 HX1old((%%eax, %1))
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
954 HX1old((%%eax, %1, 2))
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
955 HX1old((%0, %1, 4))
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
956 HX1old((%%ebx))
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
957 HX1old((%%ebx, %1))
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
958 HX1old((%%ebx, %1, 2))
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
959 */
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
960
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
961 //FIXME add some comments, its unreadable ...
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
962 #define HX1b(a, c, b, d) \
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
963 "movd " #a ", %%mm0 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
964 "movd 4" #a ", %%mm1 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
965 "punpckldq %%mm1, %%mm0 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
966 "movd " #b ", %%mm4 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
967 "movq %%mm0, %%mm1 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
968 "movq %%mm0, %%mm2 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
969 "psrlq $8, %%mm1 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
970 "movd 4" #b ", %%mm3 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
971 "psubusb %%mm1, %%mm2 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
972 "psubusb %%mm0, %%mm1 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
973 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
974 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
975 "punpckldq %%mm3, %%mm4 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
976 "movq %%mm1, %%mm3 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
977 "psllq $32, %%mm3 \n\t" /* p´5 = |p1 - p2| */\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
978 PAVGB(%%mm1, %%mm3) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
979 "paddb %%mm6, %%mm0 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
980 "psrlq $16, %%mm3 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
981 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
982 "movq %%mm4, %%mm3 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
983 "paddb %%mm5, %%mm1 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
984 "psubusb %%mm5, %%mm1 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
985 "psrlq $8, %%mm3 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
986 PAVGB(%%mm7, %%mm1)\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
987 "pxor %%mm2, %%mm1 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
988 "psubb %%mm2, %%mm1 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
989 "movq %%mm4, %%mm2 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
990 "psrlq $24, %%mm1 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
991 "psubusb %%mm3, %%mm2 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
992 "movd %%mm1, %%ecx \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
993 "psubusb %%mm4, %%mm3 \n\t"\
100
1d1182345591 minor QP scaling & auto-contrast bugfix
michael
parents: 99
diff changeset
994 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\
99
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
995 "por %%mm2, %%mm3 \n\t" /* p´x = |px - p(x+1)| */\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
996 "paddb %%mm6, %%mm0 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
997 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
998 "movq %%mm3, %%mm1 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
999 "psllq $32, %%mm1 \n\t" /* p´5 = |p1 - p2| */\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1000 "movq %%mm0, " #a " \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1001 PAVGB(%%mm3, %%mm1) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1002 "paddb %%mm6, %%mm4 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1003 "psrlq $16, %%mm1 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1004 "psubusb %%mm1, %%mm3 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1005 "paddb %%mm5, %%mm3 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1006 "psubusb %%mm5, %%mm3 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1007 PAVGB(%%mm7, %%mm3)\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1008 "pxor %%mm2, %%mm3 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1009 "psubb %%mm2, %%mm3 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1010 "psrlq $24, %%mm3 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1011 "movd " #c ", %%mm0 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1012 "movd 4" #c ", %%mm1 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1013 "punpckldq %%mm1, %%mm0 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1014 "paddb %%mm6, %%mm0 \n\t"\
100
1d1182345591 minor QP scaling & auto-contrast bugfix
michael
parents: 99
diff changeset
1015 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\
99
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1016 "paddb %%mm6, %%mm0 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1017 "movq %%mm0, " #c " \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1018 "movd %%mm3, %%ecx \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1019 "movd " #d ", %%mm0 \n\t"\
100
1d1182345591 minor QP scaling & auto-contrast bugfix
michael
parents: 99
diff changeset
1020 "paddsb (%2, %%ecx, 8), %%mm4 \n\t"\
99
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1021 "movd 4" #d ", %%mm1 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1022 "paddb %%mm6, %%mm4 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1023 "punpckldq %%mm1, %%mm0 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1024 "movq %%mm4, " #b " \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1025 "paddb %%mm6, %%mm0 \n\t"\
100
1d1182345591 minor QP scaling & auto-contrast bugfix
michael
parents: 99
diff changeset
1026 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\
99
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1027 "paddb %%mm6, %%mm0 \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1028 "movq %%mm0, " #d " \n\t"\
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1029
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1030 HX1b((%0),(%%eax),(%%eax, %1),(%%eax, %1, 2))
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1031 HX1b((%0, %1, 4),(%%ebx),(%%ebx, %1),(%%ebx, %1, 2))
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1032
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1033
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1034 :
100
1d1182345591 minor QP scaling & auto-contrast bugfix
michael
parents: 99
diff changeset
1035 : "r" (src), "r" (stride), "r" (lut)
99
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1036 : "%eax", "%ebx", "%ecx"
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1037 );
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1038 #else
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1039
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1040 //FIXME (has little in common with the mmx2 version)
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1041 for(y=0; y<BLOCK_SIZE; y++)
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1042 {
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1043 int a= src[1] - src[2];
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1044 int b= src[3] - src[4];
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1045 int c= src[5] - src[6];
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1046
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1047 int d= MAX(ABS(b) - (ABS(a) + ABS(c))/2, 0);
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1048
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1049 if(d < QP)
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1050 {
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1051 int v = d * SIGN(-b);
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1052
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1053 src[1] +=v/8;
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1054 src[2] +=v/4;
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1055 src[3] +=3*v/8;
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1056 src[4] -=3*v/8;
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1057 src[5] -=v/4;
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1058 src[6] -=v/8;
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1059
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1060 }
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1061 src+=stride;
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1062 }
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1063 #endif
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1064 }
4f072fa99ccf fixed a rounding bug thing in the X1 Filter
michael
parents: 98
diff changeset
1065
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1066
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1067 static inline void doVertDefFilter(uint8_t src[], int stride, int QP)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1068 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1069 #ifdef HAVE_MMX
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
1070 src+= stride*4;
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1071 //FIXME try pmul for *5 stuff
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1072 // src[0]=0;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1073 asm volatile(
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1074 "pxor %%mm7, %%mm7 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1075 "leal (%0, %1), %%eax \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1076 "leal (%%eax, %1, 4), %%ebx \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1077 // 0 1 2 3 4 5 6 7
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1078 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ebx+%1 ebx+2%1
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1079 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1080
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1081 "movq (%0), %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1082 "movq %%mm0, %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1083 "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1084 "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1085
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1086 "movq (%%eax), %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1087 "movq %%mm2, %%mm3 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1088 "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1089 "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1090
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1091 "movq (%%eax, %1), %%mm4 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1092 "movq %%mm4, %%mm5 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1093 "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1094 "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1095
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1096 "paddw %%mm0, %%mm0 \n\t" // 2L0
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1097 "paddw %%mm1, %%mm1 \n\t" // 2H0
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1098 "psubw %%mm4, %%mm2 \n\t" // L1 - L2
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1099 "psubw %%mm5, %%mm3 \n\t" // H1 - H2
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1100 "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1101 "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1102
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1103 "psllw $2, %%mm2 \n\t" // 4L1 - 4L2
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1104 "psllw $2, %%mm3 \n\t" // 4H1 - 4H2
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1105 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1106 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1107
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1108 "movq (%%eax, %1, 2), %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1109 "movq %%mm2, %%mm3 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1110 "punpcklbw %%mm7, %%mm2 \n\t" // L3
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1111 "punpckhbw %%mm7, %%mm3 \n\t" // H3
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1112
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1113 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1114 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1115 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1116 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1117 "movq %%mm0, temp0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1118 "movq %%mm1, temp1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1119
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1120 "movq (%0, %1, 4), %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1121 "movq %%mm0, %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1122 "punpcklbw %%mm7, %%mm0 \n\t" // L4
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1123 "punpckhbw %%mm7, %%mm1 \n\t" // H4
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1124
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1125 "psubw %%mm0, %%mm2 \n\t" // L3 - L4
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1126 "psubw %%mm1, %%mm3 \n\t" // H3 - H4
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1127 "movq %%mm2, temp2 \n\t" // L3 - L4
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1128 "movq %%mm3, temp3 \n\t" // H3 - H4
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1129 "paddw %%mm4, %%mm4 \n\t" // 2L2
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1130 "paddw %%mm5, %%mm5 \n\t" // 2H2
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1131 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1132 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1133
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1134 "psllw $2, %%mm2 \n\t" // 4L3 - 4L4
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1135 "psllw $2, %%mm3 \n\t" // 4H3 - 4H4
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1136 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1137 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1138 //50 opcodes so far
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1139 "movq (%%ebx), %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1140 "movq %%mm2, %%mm3 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1141 "punpcklbw %%mm7, %%mm2 \n\t" // L5
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1142 "punpckhbw %%mm7, %%mm3 \n\t" // H5
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1143 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1144 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1145 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1146 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1147
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1148 "movq (%%ebx, %1), %%mm6 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1149 "punpcklbw %%mm7, %%mm6 \n\t" // L6
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1150 "psubw %%mm6, %%mm2 \n\t" // L5 - L6
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1151 "movq (%%ebx, %1), %%mm6 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1152 "punpckhbw %%mm7, %%mm6 \n\t" // H6
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1153 "psubw %%mm6, %%mm3 \n\t" // H5 - H6
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1154
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1155 "paddw %%mm0, %%mm0 \n\t" // 2L4
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1156 "paddw %%mm1, %%mm1 \n\t" // 2H4
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1157 "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1158 "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1159
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1160 "psllw $2, %%mm2 \n\t" // 4L5 - 4L6
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1161 "psllw $2, %%mm3 \n\t" // 4H5 - 4H6
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1162 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1163 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1164
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1165 "movq (%%ebx, %1, 2), %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1166 "movq %%mm2, %%mm3 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1167 "punpcklbw %%mm7, %%mm2 \n\t" // L7
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1168 "punpckhbw %%mm7, %%mm3 \n\t" // H7
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1169
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1170 "paddw %%mm2, %%mm2 \n\t" // 2L7
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1171 "paddw %%mm3, %%mm3 \n\t" // 2H7
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1172 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1173 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1174
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1175 "movq temp0, %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1176 "movq temp1, %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
140
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
1177
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
1178 #ifdef HAVE_MMX2
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
1179 "movq %%mm7, %%mm6 \n\t" // 0
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
1180 "psubw %%mm0, %%mm6 \n\t"
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
1181 "pmaxsw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
1182 "movq %%mm7, %%mm6 \n\t" // 0
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
1183 "psubw %%mm1, %%mm6 \n\t"
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
1184 "pmaxsw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
1185 "movq %%mm7, %%mm6 \n\t" // 0
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
1186 "psubw %%mm2, %%mm6 \n\t"
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
1187 "pmaxsw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
1188 "movq %%mm7, %%mm6 \n\t" // 0
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
1189 "psubw %%mm3, %%mm6 \n\t"
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
1190 "pmaxsw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
1191 #else
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1192 "movq %%mm7, %%mm6 \n\t" // 0
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1193 "pcmpgtw %%mm0, %%mm6 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1194 "pxor %%mm6, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1195 "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1196 "movq %%mm7, %%mm6 \n\t" // 0
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1197 "pcmpgtw %%mm1, %%mm6 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1198 "pxor %%mm6, %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1199 "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1200 "movq %%mm7, %%mm6 \n\t" // 0
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1201 "pcmpgtw %%mm2, %%mm6 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1202 "pxor %%mm6, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1203 "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1204 "movq %%mm7, %%mm6 \n\t" // 0
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1205 "pcmpgtw %%mm3, %%mm6 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1206 "pxor %%mm6, %%mm3 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1207 "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
140
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
1208 #endif
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1209
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1210 #ifdef HAVE_MMX2
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1211 "pminsw %%mm2, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1212 "pminsw %%mm3, %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1213 #else
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1214 "movq %%mm0, %%mm6 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1215 "psubusw %%mm2, %%mm6 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1216 "psubw %%mm6, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1217 "movq %%mm1, %%mm6 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1218 "psubusw %%mm3, %%mm6 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1219 "psubw %%mm6, %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1220 #endif
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1221
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1222 "movq %%mm7, %%mm6 \n\t" // 0
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1223 "pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1224 "pxor %%mm6, %%mm4 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1225 "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5|
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1226 "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1227 "pxor %%mm7, %%mm5 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1228 "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5|
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1229 // 100 opcodes
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1230 "movd %2, %%mm2 \n\t" // QP
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1231 "punpcklwd %%mm2, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1232 "punpcklwd %%mm2, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1233 "psllw $3, %%mm2 \n\t" // 8QP
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1234 "movq %%mm2, %%mm3 \n\t" // 8QP
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1235 "pcmpgtw %%mm4, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1236 "pcmpgtw %%mm5, %%mm3 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1237 "pand %%mm2, %%mm4 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1238 "pand %%mm3, %%mm5 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1239
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1240
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1241 "psubusw %%mm0, %%mm4 \n\t" // hd
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1242 "psubusw %%mm1, %%mm5 \n\t" // ld
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1243
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1244
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1245 "movq w05, %%mm2 \n\t" // 5
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1246 "pmullw %%mm2, %%mm4 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1247 "pmullw %%mm2, %%mm5 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1248 "movq w20, %%mm2 \n\t" // 32
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1249 "paddw %%mm2, %%mm4 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1250 "paddw %%mm2, %%mm5 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1251 "psrlw $6, %%mm4 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1252 "psrlw $6, %%mm5 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1253
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1254 /*
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1255 "movq w06, %%mm2 \n\t" // 6
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1256 "paddw %%mm2, %%mm4 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1257 "paddw %%mm2, %%mm5 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1258 "movq w1400, %%mm2 \n\t" // 1400h = 5120 = 5/64*2^16
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1259 //FIXME if *5/64 is supposed to be /13 then we should use 5041 instead of 5120
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1260 "pmulhw %%mm2, %%mm4 \n\t" // hd/13
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1261 "pmulhw %%mm2, %%mm5 \n\t" // ld/13
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1262 */
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1263
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1264 "movq temp2, %%mm0 \n\t" // L3 - L4
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1265 "movq temp3, %%mm1 \n\t" // H3 - H4
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1266
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1267 "pxor %%mm2, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1268 "pxor %%mm3, %%mm3 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1269
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1270 "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1271 "pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1272 "pxor %%mm2, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1273 "pxor %%mm3, %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1274 "psubw %%mm2, %%mm0 \n\t" // |L3-L4|
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1275 "psubw %%mm3, %%mm1 \n\t" // |H3-H4|
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1276 "psrlw $1, %%mm0 \n\t" // |L3 - L4|/2
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1277 "psrlw $1, %%mm1 \n\t" // |H3 - H4|/2
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1278
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1279 "pxor %%mm6, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1280 "pxor %%mm7, %%mm3 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1281 "pand %%mm2, %%mm4 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1282 "pand %%mm3, %%mm5 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1283
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1284 #ifdef HAVE_MMX2
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1285 "pminsw %%mm0, %%mm4 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1286 "pminsw %%mm1, %%mm5 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1287 #else
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1288 "movq %%mm4, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1289 "psubusw %%mm0, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1290 "psubw %%mm2, %%mm4 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1291 "movq %%mm5, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1292 "psubusw %%mm1, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1293 "psubw %%mm2, %%mm5 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1294 #endif
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1295 "pxor %%mm6, %%mm4 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1296 "pxor %%mm7, %%mm5 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1297 "psubw %%mm6, %%mm4 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1298 "psubw %%mm7, %%mm5 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1299 "packsswb %%mm5, %%mm4 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1300 "movq (%%eax, %1, 2), %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1301 "paddb %%mm4, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1302 "movq %%mm0, (%%eax, %1, 2) \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1303 "movq (%0, %1, 4), %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1304 "psubb %%mm4, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1305 "movq %%mm0, (%0, %1, 4) \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1306
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1307 :
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1308 : "r" (src), "r" (stride), "r" (QP)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1309 : "%eax", "%ebx"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1310 );
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1311 #else
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1312 const int l1= stride;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1313 const int l2= stride + l1;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1314 const int l3= stride + l2;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1315 const int l4= stride + l3;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1316 const int l5= stride + l4;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1317 const int l6= stride + l5;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1318 const int l7= stride + l6;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1319 const int l8= stride + l7;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1320 // const int l9= stride + l8;
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
1321 int x;
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
1322 src+= stride*3;
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
1323 for(x=0; x<BLOCK_SIZE; x++)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1324 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1325 const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]);
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1326 if(ABS(middleEnergy) < 8*QP)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1327 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1328 const int q=(src[l4] - src[l5])/2;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1329 const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]);
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1330 const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]);
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1331
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1332 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) );
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1333 d= MAX(d, 0);
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1334
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1335 d= (5*d + 32) >> 6;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1336 d*= SIGN(-middleEnergy);
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1337
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1338 if(q>0)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1339 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1340 d= d<0 ? 0 : d;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1341 d= d>q ? q : d;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1342 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1343 else
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1344 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1345 d= d>0 ? 0 : d;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1346 d= d<q ? q : d;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1347 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1348
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1349 src[l4]-= d;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1350 src[l5]+= d;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1351 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1352 src++;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1353 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1354 #endif
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1355 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1356
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1357 //FIXME? |255-0| = 1
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1358 /**
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
1359 * Check if the given 8x8 Block is mostly "flat"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1360 */
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
1361 static inline int isHorizDC(uint8_t src[], int stride)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1362 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1363 // src++;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1364 int numEq= 0;
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
1365 #if 0
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1366 asm volatile (
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1367 // "int $3 \n\t"
120
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1368 "leal (%1, %2), %%ecx \n\t"
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1369 "leal (%%ecx, %2, 4), %%ebx \n\t"
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1370 // 0 1 2 3 4 5 6 7 8 9
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1371 // %1 ecx ecx+%2 ecx+2%2 %1+4%2 ebx ebx+%2 ebx+2%2 %1+8%2 ebx+4%2
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1372 "movq b7E, %%mm7 \n\t" // mm7 = 0x7F
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1373 "movq b7C, %%mm6 \n\t" // mm6 = 0x7D
120
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1374 "pxor %%mm0, %%mm0 \n\t"
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1375 "movl %1, %%eax \n\t"
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1376 "andl $0x1F, %%eax \n\t"
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1377 "cmpl $24, %%eax \n\t"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1378 "leal tempBlock, %%eax \n\t"
120
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1379 "jb 1f \n\t"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1380
120
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1381 #define HDC_CHECK_AND_CPY(src, dst) \
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1382 "movd " #src ", %%mm2 \n\t"\
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1383 "punpckldq 4" #src ", %%mm2 \n\t" /* (%1) */\
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1384 "movq %%mm2, %%mm1 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1385 "psrlq $8, %%mm2 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1386 "psubb %%mm1, %%mm2 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1387 "paddb %%mm7, %%mm2 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1388 "pcmpgtb %%mm6, %%mm2 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1389 "paddb %%mm2, %%mm0 \n\t"\
120
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1390 "movq %%mm1," #dst "(%%eax) \n\t"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1391
120
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1392 HDC_CHECK_AND_CPY((%1),0)
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1393 HDC_CHECK_AND_CPY((%%ecx),8)
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1394 HDC_CHECK_AND_CPY((%%ecx, %2),16)
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1395 HDC_CHECK_AND_CPY((%%ecx, %2, 2),24)
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1396 HDC_CHECK_AND_CPY((%1, %2, 4),32)
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1397 HDC_CHECK_AND_CPY((%%ebx),40)
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1398 HDC_CHECK_AND_CPY((%%ebx, %2),48)
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1399 HDC_CHECK_AND_CPY((%%ebx, %2, 2),56)
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1400 "jmp 2f \n\t"
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1401 "1: \n\t"
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1402 // src does not cross a 32 byte cache line so dont waste time with alignment
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1403 #define HDC_CHECK_AND_CPY2(src, dst) \
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1404 "movq " #src ", %%mm2 \n\t"\
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1405 "movq " #src ", %%mm1 \n\t"\
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1406 "psrlq $8, %%mm2 \n\t"\
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1407 "psubb %%mm1, %%mm2 \n\t"\
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1408 "paddb %%mm7, %%mm2 \n\t"\
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1409 "pcmpgtb %%mm6, %%mm2 \n\t"\
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1410 "paddb %%mm2, %%mm0 \n\t"\
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1411 "movq %%mm1," #dst "(%%eax) \n\t"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1412
120
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1413 HDC_CHECK_AND_CPY2((%1),0)
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1414 HDC_CHECK_AND_CPY2((%%ecx),8)
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1415 HDC_CHECK_AND_CPY2((%%ecx, %2),16)
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1416 HDC_CHECK_AND_CPY2((%%ecx, %2, 2),24)
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1417 HDC_CHECK_AND_CPY2((%1, %2, 4),32)
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1418 HDC_CHECK_AND_CPY2((%%ebx),40)
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1419 HDC_CHECK_AND_CPY2((%%ebx, %2),48)
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1420 HDC_CHECK_AND_CPY2((%%ebx, %2, 2),56)
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1421 "2: \n\t"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1422 "psllq $8, %%mm0 \n\t" // remove dummy value
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1423 "movq %%mm0, %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1424 "psrlw $8, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1425 "paddb %%mm1, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1426 "movq %%mm0, %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1427 "psrlq $16, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1428 "paddb %%mm1, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1429 "movq %%mm0, %%mm1 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1430 "psrlq $32, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1431 "paddb %%mm1, %%mm0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1432 "movd %%mm0, %0 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1433 : "=r" (numEq)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1434 : "r" (src), "r" (stride)
120
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1435 : "%eax", "%ebx", "%ecx"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1436 );
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1437 // printf("%d\n", numEq);
120
b0b89f5d0288 and another +2% speedup
michael
parents: 119
diff changeset
1438 numEq= (256 - numEq) &0xFF;
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1439 #else
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
1440 int y;
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
1441 for(y=0; y<BLOCK_SIZE; y++)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1442 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1443 if(((src[0] - src[1] + 1) & 0xFFFF) < 3) numEq++;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1444 if(((src[1] - src[2] + 1) & 0xFFFF) < 3) numEq++;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1445 if(((src[2] - src[3] + 1) & 0xFFFF) < 3) numEq++;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1446 if(((src[3] - src[4] + 1) & 0xFFFF) < 3) numEq++;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1447 if(((src[4] - src[5] + 1) & 0xFFFF) < 3) numEq++;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1448 if(((src[5] - src[6] + 1) & 0xFFFF) < 3) numEq++;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1449 if(((src[6] - src[7] + 1) & 0xFFFF) < 3) numEq++;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1450 src+= stride;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1451 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1452 #endif
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1453 /* if(abs(numEq - asmEq) > 0)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1454 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1455 // printf("\nasm:%d c:%d\n", asmEq, numEq);
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1456 for(int y=0; y<8; y++)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1457 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1458 for(int x=0; x<8; x++)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1459 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1460 printf("%d ", src[x + y*stride]);
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1461 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1462 printf("\n");
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1463 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1464 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1465 */
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1466 // printf("%d\n", numEq);
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1467 return numEq > hFlatnessThreshold;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1468 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1469
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
1470 static inline int isHorizMinMaxOk(uint8_t src[], int stride, int QP)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1471 {
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
1472 if(abs(src[0] - src[7]) > 2*QP) return 0;
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1473
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
1474 return 1;
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1475 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1476
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
1477 static inline void doHorizDefFilter(uint8_t dst[], int stride, int QP)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1478 {
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
1479 #if 0
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1480 asm volatile(
121
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1481 "leal (%0, %1), %%ecx \n\t"
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1482 "leal (%%ecx, %1, 4), %%ebx \n\t"
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1483 // 0 1 2 3 4 5 6 7 8 9
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1484 // %0 ecx ecx+%1 ecx+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1485 "pxor %%mm7, %%mm7 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1486 "movq bm00001000, %%mm6 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1487 "movd %2, %%mm5 \n\t" // QP
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1488 "movq %%mm5, %%mm4 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1489 "paddusb %%mm5, %%mm5 \n\t" // 2QP
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1490 "paddusb %%mm5, %%mm4 \n\t" // 3QP
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1491 "psllq $24, %%mm4 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1492 "pxor %%mm5, %%mm5 \n\t" // 0
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1493 "psubb %%mm4, %%mm5 \n\t" // -QP
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1494 "leal tempBlock, %%eax \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1495
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1496 //FIXME? "unroll by 2" and mix
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1497 #ifdef HAVE_MMX2
121
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1498 #define HDF(src, dst) \
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1499 "movq " #src "(%%eax), %%mm0 \n\t"\
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1500 "movq " #src "(%%eax), %%mm1 \n\t"\
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1501 "movq " #src "(%%eax), %%mm2 \n\t"\
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1502 "psrlq $8, %%mm1 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1503 "psubusb %%mm1, %%mm2 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1504 "psubusb %%mm0, %%mm1 \n\t"\
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1505 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1506 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1507 "pshufw $0x00, %%mm1, %%mm3 \n\t" /* p´5 = |p1 - p2| */\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1508 "pminub %%mm1, %%mm3 \n\t" /* p´5 = min(|p2-p1|, |p6-p5|)*/\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1509 "psrlq $16, %%mm3 \n\t" /* p´3 = min(|p2-p1|, |p6-p5|)*/\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1510 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-min(|p1-p2|,|p5-p6|) */\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1511 "paddb %%mm5, %%mm1 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1512 "psubusb %%mm5, %%mm1 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1513 "psrlw $2, %%mm1 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1514 "pxor %%mm2, %%mm1 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1515 "psubb %%mm2, %%mm1 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1516 "pand %%mm6, %%mm1 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1517 "psubb %%mm1, %%mm0 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1518 "psllq $8, %%mm1 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1519 "paddb %%mm1, %%mm0 \n\t"\
121
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1520 "movd %%mm0, " #dst" \n\t"\
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1521 "psrlq $32, %%mm0 \n\t"\
121
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1522 "movd %%mm0, 4" #dst" \n\t"
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1523 #else
121
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1524 #define HDF(src, dst)\
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1525 "movq " #src "(%%eax), %%mm0 \n\t"\
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1526 "movq %%mm0, %%mm1 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1527 "movq %%mm0, %%mm2 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1528 "psrlq $8, %%mm1 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1529 "psubusb %%mm1, %%mm2 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1530 "psubusb %%mm0, %%mm1 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1531 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1532 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1533 "movq %%mm1, %%mm3 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1534 "psllq $32, %%mm3 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1535 "movq %%mm3, %%mm4 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1536 "psubusb %%mm1, %%mm4 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1537 "psubb %%mm4, %%mm3 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1538 "psrlq $16, %%mm3 \n\t" /* p´3 = min(|p2-p1|, |p6-p5|)*/\
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1539 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-min(|p1-p2|,|p5,ü6|) */\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1540 "paddb %%mm5, %%mm1 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1541 "psubusb %%mm5, %%mm1 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1542 "psrlw $2, %%mm1 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1543 "pxor %%mm2, %%mm1 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1544 "psubb %%mm2, %%mm1 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1545 "pand %%mm6, %%mm1 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1546 "psubb %%mm1, %%mm0 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1547 "psllq $8, %%mm1 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1548 "paddb %%mm1, %%mm0 \n\t"\
121
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1549 "movd %%mm0, " #dst " \n\t"\
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1550 "psrlq $32, %%mm0 \n\t"\
121
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1551 "movd %%mm0, 4" #dst " \n\t"
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1552 #endif
121
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1553 HDF(0,(%0))
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1554 HDF(8,(%%ecx))
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1555 HDF(16,(%%ecx, %1))
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1556 HDF(24,(%%ecx, %1, 2))
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1557 HDF(32,(%0, %1, 4))
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1558 HDF(40,(%%ebx))
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1559 HDF(48,(%%ebx, %1))
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1560 HDF(56,(%%ebx, %1, 2))
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1561 :
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1562 : "r" (dst), "r" (stride), "r" (QP)
121
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1563 : "%eax", "%ebx", "%ecx"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1564 );
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1565 #else
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
1566 int y;
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
1567 for(y=0; y<BLOCK_SIZE; y++)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1568 {
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
1569 const int middleEnergy= 5*(dst[4] - dst[5]) + 2*(dst[2] - dst[5]);
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1570
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1571 if(ABS(middleEnergy) < 8*QP)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1572 {
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
1573 const int q=(dst[3] - dst[4])/2;
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
1574 const int leftEnergy= 5*(dst[2] - dst[1]) + 2*(dst[0] - dst[3]);
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
1575 const int rightEnergy= 5*(dst[6] - dst[5]) + 2*(dst[4] - dst[7]);
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1576
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1577 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) );
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1578 d= MAX(d, 0);
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1579
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1580 d= (5*d + 32) >> 6;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1581 d*= SIGN(-middleEnergy);
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1582
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1583 if(q>0)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1584 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1585 d= d<0 ? 0 : d;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1586 d= d>q ? q : d;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1587 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1588 else
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1589 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1590 d= d>0 ? 0 : d;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1591 d= d<q ? q : d;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1592 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1593
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1594 dst[3]-= d;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1595 dst[4]+= d;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1596 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1597 dst+= stride;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1598 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1599 #endif
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1600 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1601
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1602 /**
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1603 * Do a horizontal low pass filter on the 10x8 block (dst points to middle 8x8 Block)
107
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
1604 * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 (C version)
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
1605 * using the 7-Tap Filter (2,2,2,4,2,2,2)/16 (MMX2/3DNOW version)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1606 */
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
1607 static inline void doHorizLowPass(uint8_t dst[], int stride, int QP)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1608 {
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
1609
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
1610 #if 0
121
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1611 asm volatile(
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1612 "leal (%0, %1), %%ecx \n\t"
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1613 "leal (%%ecx, %1, 4), %%ebx \n\t"
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1614 // 0 1 2 3 4 5 6 7 8 9
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1615 // %0 ecx ecx+%1 ecx+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1616 "pxor %%mm7, %%mm7 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1617 "leal tempBlock, %%eax \n\t"
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1618 /*
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1619 #define HLP1 "movq (%0), %%mm0 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1620 "movq %%mm0, %%mm1 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1621 "psllq $8, %%mm0 \n\t"\
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1622 PAVGB(%%mm1, %%mm0)\
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1623 "psrlw $8, %%mm0 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1624 "pxor %%mm1, %%mm1 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1625 "packuswb %%mm1, %%mm0 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1626 "movq %%mm0, %%mm1 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1627 "movq %%mm0, %%mm2 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1628 "psllq $32, %%mm0 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1629 "paddb %%mm0, %%mm1 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1630 "psllq $16, %%mm2 \n\t"\
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1631 PAVGB(%%mm2, %%mm0)\
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1632 "movq %%mm0, %%mm3 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1633 "pand bm11001100, %%mm0 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1634 "paddusb %%mm0, %%mm3 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1635 "psrlq $8, %%mm3 \n\t"\
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1636 PAVGB(%%mm1, %%mm4)\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1637 PAVGB(%%mm3, %%mm2)\
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1638 "psrlq $16, %%mm2 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1639 "punpcklbw %%mm2, %%mm2 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1640 "movq %%mm2, (%0) \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1641
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1642 #define HLP2 "movq (%0), %%mm0 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1643 "movq %%mm0, %%mm1 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1644 "psllq $8, %%mm0 \n\t"\
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1645 PAVGB(%%mm1, %%mm0)\
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1646 "psrlw $8, %%mm0 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1647 "pxor %%mm1, %%mm1 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1648 "packuswb %%mm1, %%mm0 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1649 "movq %%mm0, %%mm2 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1650 "psllq $32, %%mm0 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1651 "psllq $16, %%mm2 \n\t"\
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1652 PAVGB(%%mm2, %%mm0)\
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1653 "movq %%mm0, %%mm3 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1654 "pand bm11001100, %%mm0 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1655 "paddusb %%mm0, %%mm3 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1656 "psrlq $8, %%mm3 \n\t"\
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1657 PAVGB(%%mm3, %%mm2)\
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1658 "psrlq $16, %%mm2 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1659 "punpcklbw %%mm2, %%mm2 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1660 "movq %%mm2, (%0) \n\t"\
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1661 */
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1662 // approximately a 7-Tap Filter with Vector (1,2,3,4,3,2,1)/16
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1663 /*
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1664 Implemented Exact 7-Tap
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1665 9421 A321
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1666 36421 64321
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1667 334321 =
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1668 1234321 =
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1669 1234321 =
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1670 123433 =
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1671 12463 12346
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1672 1249 123A
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1673
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1674 */
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1675
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1676 #ifdef HAVE_MMX2
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1677 #define HLP3(i) "movq " #i "(%%eax), %%mm0 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1678 "movq %%mm0, %%mm1 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1679 "movq %%mm0, %%mm2 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1680 "movq %%mm0, %%mm3 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1681 "movq %%mm0, %%mm4 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1682 "psllq $8, %%mm1 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1683 "psrlq $8, %%mm2 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1684 "pand bm00000001, %%mm3 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1685 "pand bm10000000, %%mm4 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1686 "por %%mm3, %%mm1 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1687 "por %%mm4, %%mm2 \n\t"\
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1688 PAVGB(%%mm2, %%mm1)\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1689 PAVGB(%%mm1, %%mm0)\
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1690 \
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1691 "pshufw $0xF9, %%mm0, %%mm3 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1692 "pshufw $0x90, %%mm0, %%mm4 \n\t"\
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1693 PAVGB(%%mm3, %%mm4)\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1694 PAVGB(%%mm4, %%mm0)\
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1695 "movd %%mm0, (%0) \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1696 "psrlq $32, %%mm0 \n\t"\
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1697 "movd %%mm0, 4(%0) \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1698 #else
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1699 #define HLP3(i) "movq " #i "(%%eax), %%mm0 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1700 "movq %%mm0, %%mm1 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1701 "movq %%mm0, %%mm2 \n\t"\
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1702 "movd -4(%0), %%mm3 \n\t" /*0001000*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1703 "movd 8(%0), %%mm4 \n\t" /*0001000*/\
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1704 "psllq $8, %%mm1 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1705 "psrlq $8, %%mm2 \n\t"\
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1706 "psrlq $24, %%mm3 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1707 "psllq $56, %%mm4 \n\t"\
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1708 "por %%mm3, %%mm1 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1709 "por %%mm4, %%mm2 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1710 PAVGB(%%mm2, %%mm1)\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1711 PAVGB(%%mm1, %%mm0)\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1712 \
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1713 "movq %%mm0, %%mm3 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1714 "movq %%mm0, %%mm4 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1715 "movq %%mm0, %%mm5 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1716 "psrlq $16, %%mm3 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1717 "psllq $16, %%mm4 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1718 "pand bm11000000, %%mm5 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1719 "por %%mm5, %%mm3 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1720 "movq %%mm0, %%mm5 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1721 "pand bm00000011, %%mm5 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1722 "por %%mm5, %%mm4 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1723 PAVGB(%%mm3, %%mm4)\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1724 PAVGB(%%mm4, %%mm0)\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1725 "movd %%mm0, (%0) \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1726 "psrlq $32, %%mm0 \n\t"\
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1727 "movd %%mm0, 4(%0) \n\t"
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
1728 #endif
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1729
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1730 /* uses the 7-Tap Filter: 1112111 */
121
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1731 #define NEW_HLP(src, dst)\
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1732 "movq " #src "(%%eax), %%mm1 \n\t"\
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1733 "movq " #src "(%%eax), %%mm2 \n\t"\
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1734 "psllq $8, %%mm1 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1735 "psrlq $8, %%mm2 \n\t"\
121
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1736 "movd -4" #dst ", %%mm3 \n\t" /*0001000*/\
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1737 "movd 8" #dst ", %%mm4 \n\t" /*0001000*/\
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1738 "psrlq $24, %%mm3 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1739 "psllq $56, %%mm4 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1740 "por %%mm3, %%mm1 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1741 "por %%mm4, %%mm2 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1742 "movq %%mm1, %%mm5 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1743 PAVGB(%%mm2, %%mm1)\
121
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1744 "movq " #src "(%%eax), %%mm0 \n\t"\
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1745 PAVGB(%%mm1, %%mm0)\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1746 "psllq $8, %%mm5 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1747 "psrlq $8, %%mm2 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1748 "por %%mm3, %%mm5 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1749 "por %%mm4, %%mm2 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1750 "movq %%mm5, %%mm1 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1751 PAVGB(%%mm2, %%mm5)\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1752 "psllq $8, %%mm1 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1753 "psrlq $8, %%mm2 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1754 "por %%mm3, %%mm1 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1755 "por %%mm4, %%mm2 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1756 PAVGB(%%mm2, %%mm1)\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1757 PAVGB(%%mm1, %%mm5)\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1758 PAVGB(%%mm5, %%mm0)\
121
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1759 "movd %%mm0, " #dst " \n\t"\
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1760 "psrlq $32, %%mm0 \n\t"\
121
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1761 "movd %%mm0, 4" #dst " \n\t"
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1762
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1763 /* uses the 9-Tap Filter: 112242211 */
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1764 #define NEW_HLP2(i)\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1765 "movq " #i "(%%eax), %%mm0 \n\t" /*0001000*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1766 "movq %%mm0, %%mm1 \n\t" /*0001000*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1767 "movq %%mm0, %%mm2 \n\t" /*0001000*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1768 "movd -4(%0), %%mm3 \n\t" /*0001000*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1769 "movd 8(%0), %%mm4 \n\t" /*0001000*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1770 "psllq $8, %%mm1 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1771 "psrlq $8, %%mm2 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1772 "psrlq $24, %%mm3 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1773 "psllq $56, %%mm4 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1774 "por %%mm3, %%mm1 \n\t" /*0010000*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1775 "por %%mm4, %%mm2 \n\t" /*0000100*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1776 "movq %%mm1, %%mm5 \n\t" /*0010000*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1777 PAVGB(%%mm2, %%mm1) /*0010100*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1778 PAVGB(%%mm1, %%mm0) /*0012100*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1779 "psllq $8, %%mm5 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1780 "psrlq $8, %%mm2 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1781 "por %%mm3, %%mm5 \n\t" /*0100000*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1782 "por %%mm4, %%mm2 \n\t" /*0000010*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1783 "movq %%mm5, %%mm1 \n\t" /*0100000*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1784 PAVGB(%%mm2, %%mm5) /*0100010*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1785 "psllq $8, %%mm1 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1786 "psrlq $8, %%mm2 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1787 "por %%mm3, %%mm1 \n\t" /*1000000*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1788 "por %%mm4, %%mm2 \n\t" /*0000001*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1789 "movq %%mm1, %%mm6 \n\t" /*1000000*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1790 PAVGB(%%mm2, %%mm1) /*1000001*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1791 "psllq $8, %%mm6 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1792 "psrlq $8, %%mm2 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1793 "por %%mm3, %%mm6 \n\t"/*100000000*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1794 "por %%mm4, %%mm2 \n\t"/*000000001*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1795 PAVGB(%%mm2, %%mm6) /*100000001*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1796 PAVGB(%%mm6, %%mm1) /*110000011*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1797 PAVGB(%%mm1, %%mm5) /*112000211*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1798 PAVGB(%%mm5, %%mm0) /*112242211*/\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1799 "movd %%mm0, (%0) \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1800 "psrlq $32, %%mm0 \n\t"\
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1801 "movd %%mm0, 4(%0) \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
1802
121
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1803 #define HLP(src, dst) NEW_HLP(src, dst)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1804
121
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1805 HLP(0, (%0))
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1806 HLP(8, (%%ecx))
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1807 HLP(16, (%%ecx, %1))
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1808 HLP(24, (%%ecx, %1, 2))
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1809 HLP(32, (%0, %1, 4))
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1810 HLP(40, (%%ebx))
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1811 HLP(48, (%%ebx, %1))
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1812 HLP(56, (%%ebx, %1, 2))
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1813
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1814 :
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1815 : "r" (dst), "r" (stride)
121
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
1816 : "%eax", "%ebx", "%ecx"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1817 );
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1818
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1819 #else
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
1820 int y;
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
1821 for(y=0; y<BLOCK_SIZE; y++)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1822 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1823 const int first= ABS(dst[-1] - dst[0]) < QP ? dst[-1] : dst[0];
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1824 const int last= ABS(dst[8] - dst[7]) < QP ? dst[8] : dst[7];
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1825
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1826 int sums[9];
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
1827 sums[0] = first + dst[0];
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
1828 sums[1] = dst[0] + dst[1];
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
1829 sums[2] = dst[1] + dst[2];
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
1830 sums[3] = dst[2] + dst[3];
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
1831 sums[4] = dst[3] + dst[4];
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
1832 sums[5] = dst[4] + dst[5];
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
1833 sums[6] = dst[5] + dst[6];
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
1834 sums[7] = dst[6] + dst[7];
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
1835 sums[8] = dst[7] + last;
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1836
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1837 dst[0]= ((sums[0]<<2) + ((first + sums[2])<<1) + sums[4] + 8)>>4;
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1838 dst[1]= ((dst[1]<<2) + ((first + sums[0] + sums[3])<<1) + sums[5] + 8)>>4;
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1839 dst[2]= ((dst[2]<<2) + ((first + sums[1] + sums[4])<<1) + sums[6] + 8)>>4;
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1840 dst[3]= ((dst[3]<<2) + ((sums[2] + sums[5])<<1) + sums[0] + sums[7] + 8)>>4;
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1841 dst[4]= ((dst[4]<<2) + ((sums[3] + sums[6])<<1) + sums[1] + sums[8] + 8)>>4;
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1842 dst[5]= ((dst[5]<<2) + ((last + sums[7] + sums[4])<<1) + sums[2] + 8)>>4;
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1843 dst[6]= (((last + dst[6])<<2) + ((dst[7] + sums[5])<<1) + sums[3] + 8)>>4;
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1844 dst[7]= ((sums[8]<<2) + ((last + sums[6])<<1) + sums[4] + 8)>>4;
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1845
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1846 dst+= stride;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1847 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1848 #endif
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1849 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1850
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1851 static inline void dering(uint8_t src[], int stride, int QP)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1852 {
132
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1853 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1854 asm volatile(
130
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1855 "movq pQPb, %%mm0 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1856 "paddusb %%mm0, %%mm0 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1857 "movq %%mm0, pQPb2 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1858
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1859 "leal (%0, %1), %%eax \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1860 "leal (%%eax, %1, 4), %%ebx \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1861 // 0 1 2 3 4 5 6 7 8 9
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1862 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1863
130
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1864 "pcmpeqb %%mm6, %%mm6 \n\t"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1865 "pxor %%mm7, %%mm7 \n\t"
132
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1866 #ifdef HAVE_MMX2
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1867 #define FIND_MIN_MAX(addr)\
130
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1868 "movq " #addr ", %%mm0 \n\t"\
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1869 "pminub %%mm0, %%mm6 \n\t"\
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1870 "pmaxub %%mm0, %%mm7 \n\t"
132
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1871 #else
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1872 #define FIND_MIN_MAX(addr)\
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1873 "movq " #addr ", %%mm0 \n\t"\
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1874 "movq %%mm6, %%mm1 \n\t"\
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1875 "psubusb %%mm0, %%mm7 \n\t"\
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1876 "paddb %%mm0, %%mm7 \n\t"\
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1877 "psubusb %%mm0, %%mm1 \n\t"\
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1878 "psubb %%mm1, %%mm6 \n\t"
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1879 #endif
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1880
130
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1881 FIND_MIN_MAX((%%eax))
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1882 FIND_MIN_MAX((%%eax, %1))
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1883 FIND_MIN_MAX((%%eax, %1, 2))
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1884 FIND_MIN_MAX((%0, %1, 4))
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1885 FIND_MIN_MAX((%%ebx))
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1886 FIND_MIN_MAX((%%ebx, %1))
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1887 FIND_MIN_MAX((%%ebx, %1, 2))
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1888 FIND_MIN_MAX((%0, %1, 8))
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1889
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1890 "movq %%mm6, %%mm4 \n\t"
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1891 "psrlq $8, %%mm6 \n\t"
132
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1892 #ifdef HAVE_MMX2
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1893 "pminub %%mm4, %%mm6 \n\t" // min of pixels
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1894 "pshufw $0xF9, %%mm6, %%mm4 \n\t"
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1895 "pminub %%mm4, %%mm6 \n\t" // min of pixels
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1896 "pshufw $0xFE, %%mm6, %%mm4 \n\t"
132
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1897 "pminub %%mm4, %%mm6 \n\t"
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1898 #else
132
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1899 "movq %%mm6, %%mm1 \n\t"
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1900 "psubusb %%mm4, %%mm1 \n\t"
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1901 "psubb %%mm1, %%mm6 \n\t"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1902 "movq %%mm6, %%mm4 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1903 "psrlq $16, %%mm6 \n\t"
132
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1904 "movq %%mm6, %%mm1 \n\t"
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1905 "psubusb %%mm4, %%mm1 \n\t"
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1906 "psubb %%mm1, %%mm6 \n\t"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1907 "movq %%mm6, %%mm4 \n\t"
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1908 "psrlq $32, %%mm6 \n\t"
132
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1909 "movq %%mm6, %%mm1 \n\t"
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1910 "psubusb %%mm4, %%mm1 \n\t"
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1911 "psubb %%mm1, %%mm6 \n\t"
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1912 #endif
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1913
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1914
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1915 "movq %%mm7, %%mm4 \n\t"
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1916 "psrlq $8, %%mm7 \n\t"
132
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1917 #ifdef HAVE_MMX2
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1918 "pmaxub %%mm4, %%mm7 \n\t" // max of pixels
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1919 "pshufw $0xF9, %%mm7, %%mm4 \n\t"
132
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1920 "pmaxub %%mm4, %%mm7 \n\t"
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1921 "pshufw $0xFE, %%mm7, %%mm4 \n\t"
132
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1922 "pmaxub %%mm4, %%mm7 \n\t"
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1923 #else
132
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1924 "psubusb %%mm4, %%mm7 \n\t"
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1925 "paddb %%mm4, %%mm7 \n\t"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1926 "movq %%mm7, %%mm4 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1927 "psrlq $16, %%mm7 \n\t"
132
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1928 "psubusb %%mm4, %%mm7 \n\t"
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1929 "paddb %%mm4, %%mm7 \n\t"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
1930 "movq %%mm7, %%mm4 \n\t"
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1931 "psrlq $32, %%mm7 \n\t"
132
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1932 "psubusb %%mm4, %%mm7 \n\t"
c4caf29acc1a 3dnow dering
michael
parents: 130
diff changeset
1933 "paddb %%mm4, %%mm7 \n\t"
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1934 #endif
130
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1935 PAVGB(%%mm6, %%mm7) // a=(max + min)/2
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1936 "punpcklbw %%mm7, %%mm7 \n\t"
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1937 "punpcklbw %%mm7, %%mm7 \n\t"
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
1938 "punpcklbw %%mm7, %%mm7 \n\t"
130
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1939 "movq %%mm7, temp0 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1940
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1941 "movq (%0), %%mm0 \n\t" // L10
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1942 "movq %%mm0, %%mm1 \n\t" // L10
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1943 "movq %%mm0, %%mm2 \n\t" // L10
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1944 "psllq $8, %%mm1 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1945 "psrlq $8, %%mm2 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1946 "movd -4(%0), %%mm3 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1947 "movd 8(%0), %%mm4 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1948 "psrlq $24, %%mm3 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1949 "psllq $56, %%mm4 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1950 "por %%mm3, %%mm1 \n\t" // L00
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1951 "por %%mm4, %%mm2 \n\t" // L20
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1952 "movq %%mm1, %%mm3 \n\t" // L00
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1953 PAVGB(%%mm2, %%mm1) // (L20 + L00)/2
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1954 PAVGB(%%mm0, %%mm1) // (L20 + L00 + 2L10)/4
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1955 "psubusb %%mm7, %%mm0 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1956 "psubusb %%mm7, %%mm2 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1957 "psubusb %%mm7, %%mm3 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1958 "pcmpeqb b00, %%mm0 \n\t" // L10 > a ? 0 : -1
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1959 "pcmpeqb b00, %%mm2 \n\t" // L20 > a ? 0 : -1
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1960 "pcmpeqb b00, %%mm3 \n\t" // L00 > a ? 0 : -1
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1961 "paddb %%mm2, %%mm0 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1962 "paddb %%mm3, %%mm0 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1963
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1964 "movq (%%eax), %%mm2 \n\t" // L11
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1965 "movq %%mm2, %%mm3 \n\t" // L11
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1966 "movq %%mm2, %%mm4 \n\t" // L11
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1967 "psllq $8, %%mm3 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1968 "psrlq $8, %%mm4 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1969 "movd -4(%%eax), %%mm5 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1970 "movd 8(%%eax), %%mm6 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1971 "psrlq $24, %%mm5 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1972 "psllq $56, %%mm6 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1973 "por %%mm5, %%mm3 \n\t" // L01
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1974 "por %%mm6, %%mm4 \n\t" // L21
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1975 "movq %%mm3, %%mm5 \n\t" // L01
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1976 PAVGB(%%mm4, %%mm3) // (L21 + L01)/2
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1977 PAVGB(%%mm2, %%mm3) // (L21 + L01 + 2L11)/4
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1978 "psubusb %%mm7, %%mm2 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1979 "psubusb %%mm7, %%mm4 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1980 "psubusb %%mm7, %%mm5 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1981 "pcmpeqb b00, %%mm2 \n\t" // L11 > a ? 0 : -1
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1982 "pcmpeqb b00, %%mm4 \n\t" // L21 > a ? 0 : -1
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1983 "pcmpeqb b00, %%mm5 \n\t" // L01 > a ? 0 : -1
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1984 "paddb %%mm4, %%mm2 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1985 "paddb %%mm5, %%mm2 \n\t"
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1986 // 0, 2, 3, 1
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1987 #define DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1988 "movq " #src ", " #sx " \n\t" /* src[0] */\
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1989 "movq " #sx ", " #lx " \n\t" /* src[0] */\
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1990 "movq " #sx ", " #t0 " \n\t" /* src[0] */\
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1991 "psllq $8, " #lx " \n\t"\
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1992 "psrlq $8, " #t0 " \n\t"\
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1993 "movd -4" #src ", " #t1 " \n\t"\
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1994 "psrlq $24, " #t1 " \n\t"\
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1995 "por " #t1 ", " #lx " \n\t" /* src[-1] */\
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1996 "movd 8" #src ", " #t1 " \n\t"\
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1997 "psllq $56, " #t1 " \n\t"\
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1998 "por " #t1 ", " #t0 " \n\t" /* src[+1] */\
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
1999 "movq " #lx ", " #t1 " \n\t" /* src[-1] */\
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2000 PAVGB(t0, lx) /* (src[-1] + src[+1])/2 */\
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2001 PAVGB(sx, lx) /* (src[-1] + 2src[0] + src[+1])/4 */\
135
5083d662ff85 faster dering
michael
parents: 134
diff changeset
2002 PAVGB(lx, pplx) \
5083d662ff85 faster dering
michael
parents: 134
diff changeset
2003 "movq " #lx ", temp1 \n\t"\
5083d662ff85 faster dering
michael
parents: 134
diff changeset
2004 "movq temp0, " #lx " \n\t"\
140
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
2005 "psubusb " #lx ", " #t1 " \n\t"\
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
2006 "psubusb " #lx ", " #t0 " \n\t"\
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
2007 "psubusb " #lx ", " #sx " \n\t"\
135
5083d662ff85 faster dering
michael
parents: 134
diff changeset
2008 "movq b00, " #lx " \n\t"\
140
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
2009 "pcmpeqb " #lx ", " #t1 " \n\t" /* src[-1] > a ? 0 : -1*/\
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
2010 "pcmpeqb " #lx ", " #t0 " \n\t" /* src[+1] > a ? 0 : -1*/\
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
2011 "pcmpeqb " #lx ", " #sx " \n\t" /* src[0] > a ? 0 : -1*/\
130
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2012 "paddb " #t1 ", " #t0 " \n\t"\
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2013 "paddb " #t0 ", " #sx " \n\t"\
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2014 \
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2015 PAVGB(plx, pplx) /* filtered */\
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2016 "movq " #dst ", " #t0 " \n\t" /* dst */\
134
2c469e390117 dering in c
michael
parents: 133
diff changeset
2017 "movq " #t0 ", " #t1 " \n\t" /* dst */\
2c469e390117 dering in c
michael
parents: 133
diff changeset
2018 "psubusb pQPb2, " #t0 " \n\t"\
2c469e390117 dering in c
michael
parents: 133
diff changeset
2019 "paddusb pQPb2, " #t1 " \n\t"\
2c469e390117 dering in c
michael
parents: 133
diff changeset
2020 PMAXUB(t0, pplx)\
2c469e390117 dering in c
michael
parents: 133
diff changeset
2021 PMINUB(t1, pplx, t0)\
130
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2022 "paddb " #sx ", " #ppsx " \n\t"\
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2023 "paddb " #psx ", " #ppsx " \n\t"\
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2024 "#paddb b02, " #ppsx " \n\t"\
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2025 "pand b08, " #ppsx " \n\t"\
140
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
2026 "pcmpeqb " #lx ", " #ppsx " \n\t"\
134
2c469e390117 dering in c
michael
parents: 133
diff changeset
2027 "pand " #ppsx ", " #pplx " \n\t"\
130
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2028 "pandn " #dst ", " #ppsx " \n\t"\
140
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
2029 "por " #pplx ", " #ppsx " \n\t"\
135
5083d662ff85 faster dering
michael
parents: 134
diff changeset
2030 "movq " #ppsx ", " #dst " \n\t"\
5083d662ff85 faster dering
michael
parents: 134
diff changeset
2031 "movq temp1, " #lx " \n\t"
134
2c469e390117 dering in c
michael
parents: 133
diff changeset
2032
130
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2033 /*
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2034 0000000
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2035 1111111
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2036
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2037 1111110
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2038 1111101
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2039 1111100
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2040 1111011
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2041 1111010
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2042 1111001
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2043
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2044 1111000
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2045 1110111
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2046
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2047 */
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2048 //DERING_CORE(dst,src ,ppsx ,psx ,sx ,pplx ,plx ,lx ,t0 ,t1)
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2049 DERING_CORE((%%eax),(%%eax, %1) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2050 DERING_CORE((%%eax, %1),(%%eax, %1, 2) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2051 DERING_CORE((%%eax, %1, 2),(%0, %1, 4) ,%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2052 DERING_CORE((%0, %1, 4),(%%ebx) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2053 DERING_CORE((%%ebx),(%%ebx, %1) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2054 DERING_CORE((%%ebx, %1), (%%ebx, %1, 2),%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2055 DERING_CORE((%%ebx, %1, 2),(%0, %1, 8) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
2056 DERING_CORE((%0, %1, 8),(%%ebx, %1, 4) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
2057
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
2058
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
2059 : : "r" (src), "r" (stride), "r" (QP)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
2060 : "%eax", "%ebx"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
2061 );
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
2062 #else
134
2c469e390117 dering in c
michael
parents: 133
diff changeset
2063 int y;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2064 int min=255;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2065 int max=0;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2066 int avg;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2067 uint8_t *p;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2068 int s[10];
2c469e390117 dering in c
michael
parents: 133
diff changeset
2069
2c469e390117 dering in c
michael
parents: 133
diff changeset
2070 for(y=1; y<9; y++)
2c469e390117 dering in c
michael
parents: 133
diff changeset
2071 {
2c469e390117 dering in c
michael
parents: 133
diff changeset
2072 int x;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2073 p= src + stride*y;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2074 for(x=1; x<9; x++)
2c469e390117 dering in c
michael
parents: 133
diff changeset
2075 {
2c469e390117 dering in c
michael
parents: 133
diff changeset
2076 p++;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2077 if(*p > max) max= *p;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2078 if(*p < min) min= *p;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2079 }
2c469e390117 dering in c
michael
parents: 133
diff changeset
2080 }
2c469e390117 dering in c
michael
parents: 133
diff changeset
2081 avg= (min + max + 1)/2;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2082
2c469e390117 dering in c
michael
parents: 133
diff changeset
2083 for(y=0; y<10; y++)
2c469e390117 dering in c
michael
parents: 133
diff changeset
2084 {
2c469e390117 dering in c
michael
parents: 133
diff changeset
2085 int x;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2086 int t = 0;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2087 p= src + stride*y;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2088 for(x=0; x<10; x++)
2c469e390117 dering in c
michael
parents: 133
diff changeset
2089 {
2c469e390117 dering in c
michael
parents: 133
diff changeset
2090 if(*p > avg) t |= (1<<x);
2c469e390117 dering in c
michael
parents: 133
diff changeset
2091 p++;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2092 }
2c469e390117 dering in c
michael
parents: 133
diff changeset
2093 t |= (~t)<<16;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2094 t &= (t<<1) & (t>>1);
2c469e390117 dering in c
michael
parents: 133
diff changeset
2095 s[y] = t;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2096 }
2c469e390117 dering in c
michael
parents: 133
diff changeset
2097
2c469e390117 dering in c
michael
parents: 133
diff changeset
2098 for(y=1; y<9; y++)
2c469e390117 dering in c
michael
parents: 133
diff changeset
2099 {
2c469e390117 dering in c
michael
parents: 133
diff changeset
2100 int x;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2101 int t = s[y-1] & s[y] & s[y+1];
2c469e390117 dering in c
michael
parents: 133
diff changeset
2102 t|= t>>16;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2103
2c469e390117 dering in c
michael
parents: 133
diff changeset
2104 p= src + stride*y;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2105 for(x=1; x<9; x++)
2c469e390117 dering in c
michael
parents: 133
diff changeset
2106 {
2c469e390117 dering in c
michael
parents: 133
diff changeset
2107 p++;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2108 if(t & (1<<x))
2c469e390117 dering in c
michael
parents: 133
diff changeset
2109 {
2c469e390117 dering in c
michael
parents: 133
diff changeset
2110 int f= (*(p-stride-1)) + 2*(*(p-stride)) + (*(p-stride+1))
2c469e390117 dering in c
michael
parents: 133
diff changeset
2111 +2*(*(p -1)) + 4*(*p ) + 2*(*(p +1))
2c469e390117 dering in c
michael
parents: 133
diff changeset
2112 +(*(p+stride-1)) + 2*(*(p+stride)) + (*(p+stride+1));
2c469e390117 dering in c
michael
parents: 133
diff changeset
2113 f= (f + 8)>>4;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2114
2c469e390117 dering in c
michael
parents: 133
diff changeset
2115 if (*p + 2*QP < f) *p= *p + 2*QP;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2116 else if(*p - 2*QP > f) *p= *p - 2*QP;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2117 else *p=f;
2c469e390117 dering in c
michael
parents: 133
diff changeset
2118 }
2c469e390117 dering in c
michael
parents: 133
diff changeset
2119 }
2c469e390117 dering in c
michael
parents: 133
diff changeset
2120 }
2c469e390117 dering in c
michael
parents: 133
diff changeset
2121
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
2122 #endif
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
2123 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
2124
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2125 /**
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2126 * Deinterlaces the given block
142
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
2127 * will be called for every 8x8 block and can read & write from line 4-15
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
2128 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
2129 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2130 */
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2131 static inline void deInterlaceInterpolateLinear(uint8_t src[], int stride)
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2132 {
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2133 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
142
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
2134 src+= 4*stride;
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2135 asm volatile(
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2136 "leal (%0, %1), %%eax \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2137 "leal (%%eax, %1, 4), %%ebx \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2138 // 0 1 2 3 4 5 6 7 8 9
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2139 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2140
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2141 "movq (%0), %%mm0 \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2142 "movq (%%eax, %1), %%mm1 \n\t"
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2143 PAVGB(%%mm1, %%mm0)
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2144 "movq %%mm0, (%%eax) \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2145 "movq (%0, %1, 4), %%mm0 \n\t"
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2146 PAVGB(%%mm0, %%mm1)
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2147 "movq %%mm1, (%%eax, %1, 2) \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2148 "movq (%%ebx, %1), %%mm1 \n\t"
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2149 PAVGB(%%mm1, %%mm0)
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2150 "movq %%mm0, (%%ebx) \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2151 "movq (%0, %1, 8), %%mm0 \n\t"
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2152 PAVGB(%%mm0, %%mm1)
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2153 "movq %%mm1, (%%ebx, %1, 2) \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2154
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2155 : : "r" (src), "r" (stride)
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2156 : "%eax", "%ebx"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2157 );
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2158 #else
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2159 int x;
142
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
2160 src+= 4*stride;
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2161 for(x=0; x<8; x++)
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2162 {
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2163 src[stride] = (src[0] + src[stride*2])>>1;
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2164 src[stride*3] = (src[stride*2] + src[stride*4])>>1;
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2165 src[stride*5] = (src[stride*4] + src[stride*6])>>1;
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2166 src[stride*7] = (src[stride*6] + src[stride*8])>>1;
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2167 src++;
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2168 }
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2169 #endif
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2170 }
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2171
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2172 /**
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2173 * Deinterlaces the given block
142
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
2174 * will be called for every 8x8 block and can read & write from line 4-15
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
2175 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
2176 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
2177 * this filter will read lines 3-15 and write 7-13
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2178 * no cliping in C version
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2179 */
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2180 static inline void deInterlaceInterpolateCubic(uint8_t src[], int stride)
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2181 {
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2182 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
142
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
2183 src+= stride*3;
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2184 asm volatile(
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2185 "leal (%0, %1), %%eax \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2186 "leal (%%eax, %1, 4), %%ebx \n\t"
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2187 "leal (%%ebx, %1, 4), %%ecx \n\t"
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2188 "addl %1, %%ecx \n\t"
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2189 "pxor %%mm7, %%mm7 \n\t"
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2190 // 0 1 2 3 4 5 6 7 8 9 10
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2191 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 ecx
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2192
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2193 #define DEINT_CUBIC(a,b,c,d,e)\
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2194 "movq " #a ", %%mm0 \n\t"\
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2195 "movq " #b ", %%mm1 \n\t"\
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2196 "movq " #d ", %%mm2 \n\t"\
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2197 "movq " #e ", %%mm3 \n\t"\
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2198 PAVGB(%%mm2, %%mm1) /* (b+d) /2 */\
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2199 PAVGB(%%mm3, %%mm0) /* a(a+e) /2 */\
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2200 "movq %%mm0, %%mm2 \n\t"\
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2201 "punpcklbw %%mm7, %%mm0 \n\t"\
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2202 "punpckhbw %%mm7, %%mm2 \n\t"\
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2203 "movq %%mm1, %%mm3 \n\t"\
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2204 "punpcklbw %%mm7, %%mm1 \n\t"\
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2205 "punpckhbw %%mm7, %%mm3 \n\t"\
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2206 "psubw %%mm1, %%mm0 \n\t" /* L(a+e - (b+d))/2 */\
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2207 "psubw %%mm3, %%mm2 \n\t" /* H(a+e - (b+d))/2 */\
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2208 "psraw $3, %%mm0 \n\t" /* L(a+e - (b+d))/16 */\
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2209 "psraw $3, %%mm2 \n\t" /* H(a+e - (b+d))/16 */\
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2210 "psubw %%mm0, %%mm1 \n\t" /* L(9b + 9d - a - e)/16 */\
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2211 "psubw %%mm2, %%mm3 \n\t" /* H(9b + 9d - a - e)/16 */\
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2212 "packuswb %%mm3, %%mm1 \n\t"\
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2213 "movq %%mm1, " #c " \n\t"
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2214
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2215 DEINT_CUBIC((%0), (%%eax, %1), (%%eax, %1, 2), (%0, %1, 4), (%%ebx, %1))
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2216 DEINT_CUBIC((%%eax, %1), (%0, %1, 4), (%%ebx), (%%ebx, %1), (%0, %1, 8))
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2217 DEINT_CUBIC((%0, %1, 4), (%%ebx, %1), (%%ebx, %1, 2), (%0, %1, 8), (%%ecx))
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2218 DEINT_CUBIC((%%ebx, %1), (%0, %1, 8), (%%ebx, %1, 4), (%%ecx), (%%ecx, %1, 2))
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2219
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2220 : : "r" (src), "r" (stride)
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2221 : "%eax", "%ebx", "ecx"
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2222 );
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2223 #else
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2224 int x;
142
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
2225 src+= stride*3;
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2226 for(x=0; x<8; x++)
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2227 {
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2228 src[stride*3] = (-src[0] + 9*src[stride*2] + 9*src[stride*4] - src[stride*6])>>4;
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2229 src[stride*5] = (-src[stride*2] + 9*src[stride*4] + 9*src[stride*6] - src[stride*8])>>4;
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2230 src[stride*7] = (-src[stride*4] + 9*src[stride*6] + 9*src[stride*8] - src[stride*10])>>4;
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
2231 src[stride*9] = (-src[stride*6] + 9*src[stride*8] + 9*src[stride*10] - src[stride*12])>>4;
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2232 src++;
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2233 }
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2234 #endif
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2235 }
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2236
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2237 /**
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2238 * Deinterlaces the given block
142
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
2239 * will be called for every 8x8 block and can read & write from line 4-15
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
2240 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
2241 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2242 * will shift the image up by 1 line (FIXME if this is a problem)
142
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
2243 * this filter will read lines 4-13 and write 4-11
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2244 */
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2245 static inline void deInterlaceBlendLinear(uint8_t src[], int stride)
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2246 {
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2247 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
142
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
2248 src+= 4*stride;
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2249 asm volatile(
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2250 "leal (%0, %1), %%eax \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2251 "leal (%%eax, %1, 4), %%ebx \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2252 // 0 1 2 3 4 5 6 7 8 9
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2253 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2254
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2255 "movq (%0), %%mm0 \n\t" // L0
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2256 "movq (%%eax, %1), %%mm1 \n\t" // L2
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2257 PAVGB(%%mm1, %%mm0) // L0+L2
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2258 "movq (%%eax), %%mm2 \n\t" // L1
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2259 PAVGB(%%mm2, %%mm0)
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2260 "movq %%mm0, (%0) \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2261 "movq (%%eax, %1, 2), %%mm0 \n\t" // L3
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2262 PAVGB(%%mm0, %%mm2) // L1+L3
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2263 PAVGB(%%mm1, %%mm2) // 2L2 + L1 + L3
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2264 "movq %%mm2, (%%eax) \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2265 "movq (%0, %1, 4), %%mm2 \n\t" // L4
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2266 PAVGB(%%mm2, %%mm1) // L2+L4
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2267 PAVGB(%%mm0, %%mm1) // 2L3 + L2 + L4
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2268 "movq %%mm1, (%%eax, %1) \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2269 "movq (%%ebx), %%mm1 \n\t" // L5
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2270 PAVGB(%%mm1, %%mm0) // L3+L5
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2271 PAVGB(%%mm2, %%mm0) // 2L4 + L3 + L5
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2272 "movq %%mm0, (%%eax, %1, 2) \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2273 "movq (%%ebx, %1), %%mm0 \n\t" // L6
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2274 PAVGB(%%mm0, %%mm2) // L4+L6
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2275 PAVGB(%%mm1, %%mm2) // 2L5 + L4 + L6
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2276 "movq %%mm2, (%0, %1, 4) \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2277 "movq (%%ebx, %1, 2), %%mm2 \n\t" // L7
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2278 PAVGB(%%mm2, %%mm1) // L5+L7
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2279 PAVGB(%%mm0, %%mm1) // 2L6 + L5 + L7
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2280 "movq %%mm1, (%%ebx) \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2281 "movq (%0, %1, 8), %%mm1 \n\t" // L8
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2282 PAVGB(%%mm1, %%mm0) // L6+L8
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2283 PAVGB(%%mm2, %%mm0) // 2L7 + L6 + L8
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2284 "movq %%mm0, (%%ebx, %1) \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2285 "movq (%%ebx, %1, 4), %%mm0 \n\t" // L9
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2286 PAVGB(%%mm0, %%mm2) // L7+L9
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2287 PAVGB(%%mm1, %%mm2) // 2L8 + L7 + L9
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2288 "movq %%mm2, (%%ebx, %1, 2) \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2289
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2290
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2291 : : "r" (src), "r" (stride)
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2292 : "%eax", "%ebx"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2293 );
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2294 #else
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2295 int x;
142
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
2296 src+= 4*stride;
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2297 for(x=0; x<8; x++)
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2298 {
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2299 src[0 ] = (src[0 ] + 2*src[stride ] + src[stride*2])>>2;
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2300 src[stride ] = (src[stride ] + 2*src[stride*2] + src[stride*3])>>2;
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2301 src[stride*2] = (src[stride*2] + 2*src[stride*3] + src[stride*4])>>2;
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2302 src[stride*3] = (src[stride*3] + 2*src[stride*4] + src[stride*5])>>2;
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2303 src[stride*4] = (src[stride*4] + 2*src[stride*5] + src[stride*6])>>2;
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2304 src[stride*5] = (src[stride*5] + 2*src[stride*6] + src[stride*7])>>2;
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2305 src[stride*6] = (src[stride*6] + 2*src[stride*7] + src[stride*8])>>2;
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2306 src[stride*7] = (src[stride*7] + 2*src[stride*8] + src[stride*9])>>2;
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2307 src++;
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2308 }
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2309 #endif
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2310 }
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2311
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2312 /**
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2313 * Deinterlaces the given block
142
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
2314 * will be called for every 8x8 block and can read & write from line 4-15,
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
2315 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
2316 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2317 */
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2318 static inline void deInterlaceMedian(uint8_t src[], int stride)
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2319 {
107
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2320 #ifdef HAVE_MMX
142
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
2321 src+= 4*stride;
107
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2322 #ifdef HAVE_MMX2
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2323 asm volatile(
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2324 "leal (%0, %1), %%eax \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2325 "leal (%%eax, %1, 4), %%ebx \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2326 // 0 1 2 3 4 5 6 7 8 9
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2327 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2328
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2329 "movq (%0), %%mm0 \n\t" //
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2330 "movq (%%eax, %1), %%mm2 \n\t" //
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2331 "movq (%%eax), %%mm1 \n\t" //
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2332 "movq %%mm0, %%mm3 \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2333 "pmaxub %%mm1, %%mm0 \n\t" //
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2334 "pminub %%mm3, %%mm1 \n\t" //
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2335 "pmaxub %%mm2, %%mm1 \n\t" //
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2336 "pminub %%mm1, %%mm0 \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2337 "movq %%mm0, (%%eax) \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2338
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2339 "movq (%0, %1, 4), %%mm0 \n\t" //
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2340 "movq (%%eax, %1, 2), %%mm1 \n\t" //
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2341 "movq %%mm2, %%mm3 \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2342 "pmaxub %%mm1, %%mm2 \n\t" //
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2343 "pminub %%mm3, %%mm1 \n\t" //
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2344 "pmaxub %%mm0, %%mm1 \n\t" //
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2345 "pminub %%mm1, %%mm2 \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2346 "movq %%mm2, (%%eax, %1, 2) \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2347
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2348 "movq (%%ebx), %%mm2 \n\t" //
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2349 "movq (%%ebx, %1), %%mm1 \n\t" //
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2350 "movq %%mm2, %%mm3 \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2351 "pmaxub %%mm0, %%mm2 \n\t" //
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2352 "pminub %%mm3, %%mm0 \n\t" //
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2353 "pmaxub %%mm1, %%mm0 \n\t" //
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2354 "pminub %%mm0, %%mm2 \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2355 "movq %%mm2, (%%ebx) \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2356
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2357 "movq (%%ebx, %1, 2), %%mm2 \n\t" //
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2358 "movq (%0, %1, 8), %%mm0 \n\t" //
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2359 "movq %%mm2, %%mm3 \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2360 "pmaxub %%mm0, %%mm2 \n\t" //
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2361 "pminub %%mm3, %%mm0 \n\t" //
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2362 "pmaxub %%mm1, %%mm0 \n\t" //
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2363 "pminub %%mm0, %%mm2 \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2364 "movq %%mm2, (%%ebx, %1, 2) \n\t"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2365
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2366
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2367 : : "r" (src), "r" (stride)
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2368 : "%eax", "%ebx"
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2369 );
107
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2370
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2371 #else // MMX without MMX2
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2372 asm volatile(
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2373 "leal (%0, %1), %%eax \n\t"
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2374 "leal (%%eax, %1, 4), %%ebx \n\t"
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2375 // 0 1 2 3 4 5 6 7 8 9
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2376 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2377 "pxor %%mm7, %%mm7 \n\t"
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2378
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2379 #define MEDIAN(a,b,c)\
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2380 "movq " #a ", %%mm0 \n\t"\
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2381 "movq " #b ", %%mm2 \n\t"\
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2382 "movq " #c ", %%mm1 \n\t"\
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2383 "movq %%mm0, %%mm3 \n\t"\
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2384 "movq %%mm1, %%mm4 \n\t"\
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2385 "movq %%mm2, %%mm5 \n\t"\
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2386 "psubusb %%mm1, %%mm3 \n\t"\
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2387 "psubusb %%mm2, %%mm4 \n\t"\
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2388 "psubusb %%mm0, %%mm5 \n\t"\
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2389 "pcmpeqb %%mm7, %%mm3 \n\t"\
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2390 "pcmpeqb %%mm7, %%mm4 \n\t"\
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2391 "pcmpeqb %%mm7, %%mm5 \n\t"\
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2392 "movq %%mm3, %%mm6 \n\t"\
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2393 "pxor %%mm4, %%mm3 \n\t"\
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2394 "pxor %%mm5, %%mm4 \n\t"\
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2395 "pxor %%mm6, %%mm5 \n\t"\
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2396 "por %%mm3, %%mm1 \n\t"\
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2397 "por %%mm4, %%mm2 \n\t"\
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2398 "por %%mm5, %%mm0 \n\t"\
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2399 "pand %%mm2, %%mm0 \n\t"\
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2400 "pand %%mm1, %%mm0 \n\t"\
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2401 "movq %%mm0, " #b " \n\t"
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2402
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2403 MEDIAN((%0), (%%eax), (%%eax, %1))
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2404 MEDIAN((%%eax, %1), (%%eax, %1, 2), (%0, %1, 4))
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2405 MEDIAN((%0, %1, 4), (%%ebx), (%%ebx, %1))
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2406 MEDIAN((%%ebx, %1), (%%ebx, %1, 2), (%0, %1, 8))
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2407
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2408 : : "r" (src), "r" (stride)
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2409 : "%eax", "%ebx"
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2410 );
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
2411 #endif // MMX
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2412 #else
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2413 //FIXME
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2414 int x;
142
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
2415 src+= 4*stride;
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2416 for(x=0; x<8; x++)
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2417 {
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2418 src[0 ] = (src[0 ] + 2*src[stride ] + src[stride*2])>>2;
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2419 src[stride ] = (src[stride ] + 2*src[stride*2] + src[stride*3])>>2;
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2420 src[stride*2] = (src[stride*2] + 2*src[stride*3] + src[stride*4])>>2;
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2421 src[stride*3] = (src[stride*3] + 2*src[stride*4] + src[stride*5])>>2;
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2422 src[stride*4] = (src[stride*4] + 2*src[stride*5] + src[stride*6])>>2;
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2423 src[stride*5] = (src[stride*5] + 2*src[stride*6] + src[stride*7])>>2;
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2424 src[stride*6] = (src[stride*6] + 2*src[stride*7] + src[stride*8])>>2;
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2425 src[stride*7] = (src[stride*7] + 2*src[stride*8] + src[stride*9])>>2;
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2426 src++;
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2427 }
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2428 #endif
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2429 }
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
2430
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
2431 #ifdef HAVE_MMX
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2432 /**
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2433 * transposes and shift the given 8x8 Block into dst1 and dst2
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2434 */
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2435 static inline void transpose1(uint8_t *dst1, uint8_t *dst2, uint8_t *src, int srcStride)
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2436 {
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2437 asm(
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2438 "leal (%0, %1), %%eax \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2439 "leal (%%eax, %1, 4), %%ebx \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2440 // 0 1 2 3 4 5 6 7 8 9
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2441 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2442 "movq (%0), %%mm0 \n\t" // 12345678
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2443 "movq (%%eax), %%mm1 \n\t" // abcdefgh
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2444 "movq %%mm0, %%mm2 \n\t" // 12345678
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2445 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2446 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2447
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2448 "movq (%%eax, %1), %%mm1 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2449 "movq (%%eax, %1, 2), %%mm3 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2450 "movq %%mm1, %%mm4 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2451 "punpcklbw %%mm3, %%mm1 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2452 "punpckhbw %%mm3, %%mm4 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2453
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2454 "movq %%mm0, %%mm3 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2455 "punpcklwd %%mm1, %%mm0 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2456 "punpckhwd %%mm1, %%mm3 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2457 "movq %%mm2, %%mm1 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2458 "punpcklwd %%mm4, %%mm2 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2459 "punpckhwd %%mm4, %%mm1 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2460
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2461 "movd %%mm0, 128(%2) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2462 "psrlq $32, %%mm0 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2463 "movd %%mm0, 144(%2) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2464 "movd %%mm3, 160(%2) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2465 "psrlq $32, %%mm3 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2466 "movd %%mm3, 176(%2) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2467 "movd %%mm3, 48(%3) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2468 "movd %%mm2, 192(%2) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2469 "movd %%mm2, 64(%3) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2470 "psrlq $32, %%mm2 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2471 "movd %%mm2, 80(%3) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2472 "movd %%mm1, 96(%3) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2473 "psrlq $32, %%mm1 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2474 "movd %%mm1, 112(%3) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2475
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2476 "movq (%0, %1, 4), %%mm0 \n\t" // 12345678
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2477 "movq (%%ebx), %%mm1 \n\t" // abcdefgh
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2478 "movq %%mm0, %%mm2 \n\t" // 12345678
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2479 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2480 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2481
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2482 "movq (%%ebx, %1), %%mm1 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2483 "movq (%%ebx, %1, 2), %%mm3 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2484 "movq %%mm1, %%mm4 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2485 "punpcklbw %%mm3, %%mm1 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2486 "punpckhbw %%mm3, %%mm4 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2487
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2488 "movq %%mm0, %%mm3 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2489 "punpcklwd %%mm1, %%mm0 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2490 "punpckhwd %%mm1, %%mm3 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2491 "movq %%mm2, %%mm1 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2492 "punpcklwd %%mm4, %%mm2 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2493 "punpckhwd %%mm4, %%mm1 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2494
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2495 "movd %%mm0, 132(%2) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2496 "psrlq $32, %%mm0 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2497 "movd %%mm0, 148(%2) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2498 "movd %%mm3, 164(%2) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2499 "psrlq $32, %%mm3 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2500 "movd %%mm3, 180(%2) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2501 "movd %%mm3, 52(%3) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2502 "movd %%mm2, 196(%2) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2503 "movd %%mm2, 68(%3) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2504 "psrlq $32, %%mm2 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2505 "movd %%mm2, 84(%3) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2506 "movd %%mm1, 100(%3) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2507 "psrlq $32, %%mm1 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2508 "movd %%mm1, 116(%3) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2509
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2510
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2511 :: "r" (src), "r" (srcStride), "r" (dst1), "r" (dst2)
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2512 : "%eax", "%ebx"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2513 );
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2514 }
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2515
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2516 /**
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2517 * transposes the given 8x8 block
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2518 */
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2519 static inline void transpose2(uint8_t *dst, int dstStride, uint8_t *src)
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2520 {
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2521 asm(
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2522 "leal (%0, %1), %%eax \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2523 "leal (%%eax, %1, 4), %%ebx \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2524 // 0 1 2 3 4 5 6 7 8 9
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2525 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2526 "movq (%2), %%mm0 \n\t" // 12345678
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2527 "movq 16(%2), %%mm1 \n\t" // abcdefgh
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2528 "movq %%mm0, %%mm2 \n\t" // 12345678
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2529 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2530 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2531
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2532 "movq 32(%2), %%mm1 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2533 "movq 48(%2), %%mm3 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2534 "movq %%mm1, %%mm4 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2535 "punpcklbw %%mm3, %%mm1 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2536 "punpckhbw %%mm3, %%mm4 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2537
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2538 "movq %%mm0, %%mm3 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2539 "punpcklwd %%mm1, %%mm0 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2540 "punpckhwd %%mm1, %%mm3 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2541 "movq %%mm2, %%mm1 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2542 "punpcklwd %%mm4, %%mm2 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2543 "punpckhwd %%mm4, %%mm1 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2544
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2545 "movd %%mm0, (%0) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2546 "psrlq $32, %%mm0 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2547 "movd %%mm0, (%%eax) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2548 "movd %%mm3, (%%eax, %1) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2549 "psrlq $32, %%mm3 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2550 "movd %%mm3, (%%eax, %1, 2) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2551 "movd %%mm2, (%0, %1, 4) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2552 "psrlq $32, %%mm2 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2553 "movd %%mm2, (%%ebx) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2554 "movd %%mm1, (%%ebx, %1) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2555 "psrlq $32, %%mm1 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2556 "movd %%mm1, (%%ebx, %1, 2) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2557
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2558
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2559 "movq 64(%2), %%mm0 \n\t" // 12345678
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2560 "movq 80(%2), %%mm1 \n\t" // abcdefgh
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2561 "movq %%mm0, %%mm2 \n\t" // 12345678
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2562 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2563 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2564
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2565 "movq 96(%2), %%mm1 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2566 "movq 112(%2), %%mm3 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2567 "movq %%mm1, %%mm4 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2568 "punpcklbw %%mm3, %%mm1 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2569 "punpckhbw %%mm3, %%mm4 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2570
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2571 "movq %%mm0, %%mm3 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2572 "punpcklwd %%mm1, %%mm0 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2573 "punpckhwd %%mm1, %%mm3 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2574 "movq %%mm2, %%mm1 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2575 "punpcklwd %%mm4, %%mm2 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2576 "punpckhwd %%mm4, %%mm1 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2577
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2578 "movd %%mm0, 4(%0) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2579 "psrlq $32, %%mm0 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2580 "movd %%mm0, 4(%%eax) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2581 "movd %%mm3, 4(%%eax, %1) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2582 "psrlq $32, %%mm3 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2583 "movd %%mm3, 4(%%eax, %1, 2) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2584 "movd %%mm2, 4(%0, %1, 4) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2585 "psrlq $32, %%mm2 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2586 "movd %%mm2, 4(%%ebx) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2587 "movd %%mm1, 4(%%ebx, %1) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2588 "psrlq $32, %%mm1 \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2589 "movd %%mm1, 4(%%ebx, %1, 2) \n\t"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2590
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2591 :: "r" (dst), "r" (dstStride), "r" (src)
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2592 : "%eax", "%ebx"
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2593 );
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2594 }
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
2595 #endif
157
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2596 //static int test=0;
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
2597
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2598 static void inline tempNoiseReducer(uint8_t *src, int stride,
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2599 uint8_t *tempBlured, int *maxNoise)
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2600 {
157
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2601 #define FAST_L2_DIFF
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2602 //#define L1_DIFF //u should change the thresholds too if u try that one
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2603 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2604 asm volatile(
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2605 "leal (%2, %2, 2), %%eax \n\t" // 3*stride
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2606 "leal (%2, %2, 4), %%ebx \n\t" // 5*stride
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2607 "leal (%%ebx, %2, 2), %%ecx \n\t" // 7*stride
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2608 // 0 1 2 3 4 5 6 7 8 9
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2609 // %x %x+%2 %x+2%2 %x+eax %x+4%2 %x+ebx %x+2eax %x+ecx %x+8%2
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2610 //FIXME reorder?
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2611 #ifdef L1_DIFF //needs mmx2
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2612 "movq (%0), %%mm0 \n\t" // L0
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2613 "psadbw (%1), %%mm0 \n\t" // |L0-R0|
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2614 "movq (%0, %2), %%mm1 \n\t" // L1
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2615 "psadbw (%1, %2), %%mm1 \n\t" // |L1-R1|
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2616 "movq (%0, %2, 2), %%mm2 \n\t" // L2
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2617 "psadbw (%1, %2, 2), %%mm2 \n\t" // |L2-R2|
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2618 "movq (%0, %%eax), %%mm3 \n\t" // L3
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2619 "psadbw (%1, %%eax), %%mm3 \n\t" // |L3-R3|
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2620
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2621 "movq (%0, %2, 4), %%mm4 \n\t" // L4
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2622 "paddw %%mm1, %%mm0 \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2623 "psadbw (%1, %2, 4), %%mm4 \n\t" // |L4-R4|
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2624 "movq (%0, %%ebx), %%mm5 \n\t" // L5
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2625 "paddw %%mm2, %%mm0 \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2626 "psadbw (%1, %%ebx), %%mm5 \n\t" // |L5-R5|
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2627 "movq (%0, %%eax, 2), %%mm6 \n\t" // L6
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2628 "paddw %%mm3, %%mm0 \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2629 "psadbw (%1, %%eax, 2), %%mm6 \n\t" // |L6-R6|
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2630 "movq (%0, %%ecx), %%mm7 \n\t" // L7
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2631 "paddw %%mm4, %%mm0 \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2632 "psadbw (%1, %%ecx), %%mm7 \n\t" // |L7-R7|
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2633 "paddw %%mm5, %%mm6 \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2634 "paddw %%mm7, %%mm6 \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2635 "paddw %%mm6, %%mm0 \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2636 #elif defined (FAST_L2_DIFF)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2637 "pcmpeqb %%mm7, %%mm7 \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2638 "movq b80, %%mm6 \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2639 "pxor %%mm0, %%mm0 \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2640 #define L2_DIFF_CORE(a, b)\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2641 "movq " #a ", %%mm5 \n\t"\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2642 "movq " #b ", %%mm2 \n\t"\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2643 "pxor %%mm7, %%mm2 \n\t"\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2644 PAVGB(%%mm2, %%mm5)\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2645 "paddb %%mm6, %%mm5 \n\t"\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2646 "movq %%mm5, %%mm2 \n\t"\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2647 "psllw $8, %%mm5 \n\t"\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2648 "pmaddwd %%mm5, %%mm5 \n\t"\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2649 "pmaddwd %%mm2, %%mm2 \n\t"\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2650 "paddd %%mm2, %%mm5 \n\t"\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2651 "psrld $14, %%mm5 \n\t"\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2652 "paddd %%mm5, %%mm0 \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2653
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2654 L2_DIFF_CORE((%0), (%1))
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2655 L2_DIFF_CORE((%0, %2), (%1, %2))
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2656 L2_DIFF_CORE((%0, %2, 2), (%1, %2, 2))
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2657 L2_DIFF_CORE((%0, %%eax), (%1, %%eax))
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2658 L2_DIFF_CORE((%0, %2, 4), (%1, %2, 4))
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2659 L2_DIFF_CORE((%0, %%ebx), (%1, %%ebx))
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2660 L2_DIFF_CORE((%0, %%eax,2), (%1, %%eax,2))
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2661 L2_DIFF_CORE((%0, %%ecx), (%1, %%ecx))
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2662
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2663 #else
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2664 "pxor %%mm7, %%mm7 \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2665 "pxor %%mm0, %%mm0 \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2666 #define L2_DIFF_CORE(a, b)\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2667 "movq " #a ", %%mm5 \n\t"\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2668 "movq " #b ", %%mm2 \n\t"\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2669 "movq %%mm5, %%mm1 \n\t"\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2670 "movq %%mm2, %%mm3 \n\t"\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2671 "punpcklbw %%mm7, %%mm5 \n\t"\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2672 "punpckhbw %%mm7, %%mm1 \n\t"\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2673 "punpcklbw %%mm7, %%mm2 \n\t"\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2674 "punpckhbw %%mm7, %%mm3 \n\t"\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2675 "psubw %%mm2, %%mm5 \n\t"\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2676 "psubw %%mm3, %%mm1 \n\t"\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2677 "pmaddwd %%mm5, %%mm5 \n\t"\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2678 "pmaddwd %%mm1, %%mm1 \n\t"\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2679 "paddd %%mm1, %%mm5 \n\t"\
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2680 "paddd %%mm5, %%mm0 \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2681
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2682 L2_DIFF_CORE((%0), (%1))
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2683 L2_DIFF_CORE((%0, %2), (%1, %2))
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2684 L2_DIFF_CORE((%0, %2, 2), (%1, %2, 2))
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2685 L2_DIFF_CORE((%0, %%eax), (%1, %%eax))
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2686 L2_DIFF_CORE((%0, %2, 4), (%1, %2, 4))
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2687 L2_DIFF_CORE((%0, %%ebx), (%1, %%ebx))
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2688 L2_DIFF_CORE((%0, %%eax,2), (%1, %%eax,2))
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2689 L2_DIFF_CORE((%0, %%ecx), (%1, %%ecx))
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2690
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2691 #endif
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2692
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2693 "movq %%mm0, %%mm4 \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2694 "psrlq $32, %%mm0 \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2695 "paddd %%mm0, %%mm4 \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2696 "movd %%mm4, %%ecx \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2697 // "movl %3, %%ecx \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2698 // "movl %%ecx, test \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2699 // "jmp 4f \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2700 "cmpl %4, %%ecx \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2701 " jb 2f \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2702 "cmpl %5, %%ecx \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2703 " jb 1f \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2704
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2705 "leal (%%ebx, %2, 2), %%ecx \n\t" // 7*stride
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2706 "movq (%0), %%mm0 \n\t" // L0
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2707 "movq (%0, %2), %%mm1 \n\t" // L1
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2708 "movq (%0, %2, 2), %%mm2 \n\t" // L2
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2709 "movq (%0, %%eax), %%mm3 \n\t" // L3
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2710 "movq (%0, %2, 4), %%mm4 \n\t" // L4
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2711 "movq (%0, %%ebx), %%mm5 \n\t" // L5
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2712 "movq (%0, %%eax, 2), %%mm6 \n\t" // L6
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2713 "movq (%0, %%ecx), %%mm7 \n\t" // L7
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2714 "movq %%mm0, (%1) \n\t" // L0
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2715 "movq %%mm1, (%1, %2) \n\t" // L1
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2716 "movq %%mm2, (%1, %2, 2) \n\t" // L2
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2717 "movq %%mm3, (%1, %%eax) \n\t" // L3
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2718 "movq %%mm4, (%1, %2, 4) \n\t" // L4
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2719 "movq %%mm5, (%1, %%ebx) \n\t" // L5
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2720 "movq %%mm6, (%1, %%eax, 2) \n\t" // L6
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2721 "movq %%mm7, (%1, %%ecx) \n\t" // L7
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2722 "jmp 4f \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2723
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2724 "1: \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2725 "leal (%%ebx, %2, 2), %%ecx \n\t" // 7*stride
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2726 "movq (%0), %%mm0 \n\t" // L0
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2727 "pavgb (%1), %%mm0 \n\t" // L0
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2728 "movq (%0, %2), %%mm1 \n\t" // L1
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2729 "pavgb (%1, %2), %%mm1 \n\t" // L1
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2730 "movq (%0, %2, 2), %%mm2 \n\t" // L2
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2731 "pavgb (%1, %2, 2), %%mm2 \n\t" // L2
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2732 "movq (%0, %%eax), %%mm3 \n\t" // L3
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2733 "pavgb (%1, %%eax), %%mm3 \n\t" // L3
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2734 "movq (%0, %2, 4), %%mm4 \n\t" // L4
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2735 "pavgb (%1, %2, 4), %%mm4 \n\t" // L4
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2736 "movq (%0, %%ebx), %%mm5 \n\t" // L5
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2737 "pavgb (%1, %%ebx), %%mm5 \n\t" // L5
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2738 "movq (%0, %%eax, 2), %%mm6 \n\t" // L6
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2739 "pavgb (%1, %%eax, 2), %%mm6 \n\t" // L6
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2740 "movq (%0, %%ecx), %%mm7 \n\t" // L7
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2741 "pavgb (%1, %%ecx), %%mm7 \n\t" // L7
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2742 "movq %%mm0, (%1) \n\t" // R0
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2743 "movq %%mm1, (%1, %2) \n\t" // R1
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2744 "movq %%mm2, (%1, %2, 2) \n\t" // R2
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2745 "movq %%mm3, (%1, %%eax) \n\t" // R3
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2746 "movq %%mm4, (%1, %2, 4) \n\t" // R4
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2747 "movq %%mm5, (%1, %%ebx) \n\t" // R5
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2748 "movq %%mm6, (%1, %%eax, 2) \n\t" // R6
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2749 "movq %%mm7, (%1, %%ecx) \n\t" // R7
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2750 "movq %%mm0, (%0) \n\t" // L0
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2751 "movq %%mm1, (%0, %2) \n\t" // L1
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2752 "movq %%mm2, (%0, %2, 2) \n\t" // L2
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2753 "movq %%mm3, (%0, %%eax) \n\t" // L3
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2754 "movq %%mm4, (%0, %2, 4) \n\t" // L4
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2755 "movq %%mm5, (%0, %%ebx) \n\t" // L5
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2756 "movq %%mm6, (%0, %%eax, 2) \n\t" // L6
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2757 "movq %%mm7, (%0, %%ecx) \n\t" // L7
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2758 "jmp 4f \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2759
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2760 "2: \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2761 "cmpl %3, %%ecx \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2762 " jb 3f \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2763
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2764 "leal (%%ebx, %2, 2), %%ecx \n\t" // 7*stride
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2765 "movq (%0), %%mm0 \n\t" // L0
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2766 "movq (%0, %2), %%mm1 \n\t" // L1
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2767 "movq (%0, %2, 2), %%mm2 \n\t" // L2
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2768 "movq (%0, %%eax), %%mm3 \n\t" // L3
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2769 "movq (%1), %%mm4 \n\t" // R0
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2770 "movq (%1, %2), %%mm5 \n\t" // R1
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2771 "movq (%1, %2, 2), %%mm6 \n\t" // R2
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2772 "movq (%1, %%eax), %%mm7 \n\t" // R3
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2773 PAVGB(%%mm4, %%mm0)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2774 PAVGB(%%mm5, %%mm1)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2775 PAVGB(%%mm6, %%mm2)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2776 PAVGB(%%mm7, %%mm3)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2777 PAVGB(%%mm4, %%mm0)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2778 PAVGB(%%mm5, %%mm1)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2779 PAVGB(%%mm6, %%mm2)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2780 PAVGB(%%mm7, %%mm3)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2781 "movq %%mm0, (%1) \n\t" // R0
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2782 "movq %%mm1, (%1, %2) \n\t" // R1
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2783 "movq %%mm2, (%1, %2, 2) \n\t" // R2
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2784 "movq %%mm3, (%1, %%eax) \n\t" // R3
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2785 "movq %%mm0, (%0) \n\t" // L0
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2786 "movq %%mm1, (%0, %2) \n\t" // L1
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2787 "movq %%mm2, (%0, %2, 2) \n\t" // L2
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2788 "movq %%mm3, (%0, %%eax) \n\t" // L3
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2789
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2790 "movq (%0, %2, 4), %%mm0 \n\t" // L4
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2791 "movq (%0, %%ebx), %%mm1 \n\t" // L5
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2792 "movq (%0, %%eax, 2), %%mm2 \n\t" // L6
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2793 "movq (%0, %%ecx), %%mm3 \n\t" // L7
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2794 "movq (%1, %2, 4), %%mm4 \n\t" // R4
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2795 "movq (%1, %%ebx), %%mm5 \n\t" // R5
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2796 "movq (%1, %%eax, 2), %%mm6 \n\t" // R6
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2797 "movq (%1, %%ecx), %%mm7 \n\t" // R7
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2798 PAVGB(%%mm4, %%mm0)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2799 PAVGB(%%mm5, %%mm1)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2800 PAVGB(%%mm6, %%mm2)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2801 PAVGB(%%mm7, %%mm3)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2802 PAVGB(%%mm4, %%mm0)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2803 PAVGB(%%mm5, %%mm1)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2804 PAVGB(%%mm6, %%mm2)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2805 PAVGB(%%mm7, %%mm3)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2806 "movq %%mm0, (%1, %2, 4) \n\t" // R4
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2807 "movq %%mm1, (%1, %%ebx) \n\t" // R5
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2808 "movq %%mm2, (%1, %%eax, 2) \n\t" // R6
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2809 "movq %%mm3, (%1, %%ecx) \n\t" // R7
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2810 "movq %%mm0, (%0, %2, 4) \n\t" // L4
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2811 "movq %%mm1, (%0, %%ebx) \n\t" // L5
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2812 "movq %%mm2, (%0, %%eax, 2) \n\t" // L6
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2813 "movq %%mm3, (%0, %%ecx) \n\t" // L7
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2814 "jmp 4f \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2815
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2816 "3: \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2817 "leal (%%ebx, %2, 2), %%ecx \n\t" // 7*stride
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2818 "movq (%0), %%mm0 \n\t" // L0
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2819 "movq (%0, %2), %%mm1 \n\t" // L1
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2820 "movq (%0, %2, 2), %%mm2 \n\t" // L2
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2821 "movq (%0, %%eax), %%mm3 \n\t" // L3
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2822 "movq (%1), %%mm4 \n\t" // R0
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2823 "movq (%1, %2), %%mm5 \n\t" // R1
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2824 "movq (%1, %2, 2), %%mm6 \n\t" // R2
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2825 "movq (%1, %%eax), %%mm7 \n\t" // R3
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2826 PAVGB(%%mm4, %%mm0)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2827 PAVGB(%%mm5, %%mm1)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2828 PAVGB(%%mm6, %%mm2)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2829 PAVGB(%%mm7, %%mm3)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2830 PAVGB(%%mm4, %%mm0)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2831 PAVGB(%%mm5, %%mm1)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2832 PAVGB(%%mm6, %%mm2)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2833 PAVGB(%%mm7, %%mm3)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2834 PAVGB(%%mm4, %%mm0)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2835 PAVGB(%%mm5, %%mm1)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2836 PAVGB(%%mm6, %%mm2)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2837 PAVGB(%%mm7, %%mm3)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2838 "movq %%mm0, (%1) \n\t" // R0
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2839 "movq %%mm1, (%1, %2) \n\t" // R1
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2840 "movq %%mm2, (%1, %2, 2) \n\t" // R2
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2841 "movq %%mm3, (%1, %%eax) \n\t" // R3
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2842 "movq %%mm0, (%0) \n\t" // L0
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2843 "movq %%mm1, (%0, %2) \n\t" // L1
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2844 "movq %%mm2, (%0, %2, 2) \n\t" // L2
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2845 "movq %%mm3, (%0, %%eax) \n\t" // L3
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2846
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2847 "movq (%0, %2, 4), %%mm0 \n\t" // L4
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2848 "movq (%0, %%ebx), %%mm1 \n\t" // L5
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2849 "movq (%0, %%eax, 2), %%mm2 \n\t" // L6
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2850 "movq (%0, %%ecx), %%mm3 \n\t" // L7
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2851 "movq (%1, %2, 4), %%mm4 \n\t" // R4
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2852 "movq (%1, %%ebx), %%mm5 \n\t" // R5
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2853 "movq (%1, %%eax, 2), %%mm6 \n\t" // R6
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2854 "movq (%1, %%ecx), %%mm7 \n\t" // R7
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2855 PAVGB(%%mm4, %%mm0)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2856 PAVGB(%%mm5, %%mm1)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2857 PAVGB(%%mm6, %%mm2)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2858 PAVGB(%%mm7, %%mm3)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2859 PAVGB(%%mm4, %%mm0)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2860 PAVGB(%%mm5, %%mm1)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2861 PAVGB(%%mm6, %%mm2)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2862 PAVGB(%%mm7, %%mm3)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2863 PAVGB(%%mm4, %%mm0)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2864 PAVGB(%%mm5, %%mm1)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2865 PAVGB(%%mm6, %%mm2)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2866 PAVGB(%%mm7, %%mm3)
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2867 "movq %%mm0, (%1, %2, 4) \n\t" // R4
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2868 "movq %%mm1, (%1, %%ebx) \n\t" // R5
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2869 "movq %%mm2, (%1, %%eax, 2) \n\t" // R6
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2870 "movq %%mm3, (%1, %%ecx) \n\t" // R7
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2871 "movq %%mm0, (%0, %2, 4) \n\t" // L4
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2872 "movq %%mm1, (%0, %%ebx) \n\t" // L5
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2873 "movq %%mm2, (%0, %%eax, 2) \n\t" // L6
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2874 "movq %%mm3, (%0, %%ecx) \n\t" // L7
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2875
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2876 "4: \n\t"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2877
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2878 :: "r" (src), "r" (tempBlured), "r"(stride),
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2879 "m"(maxNoise[0]), "m"(maxNoise[1]), "m"(maxNoise[2])
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2880 : "%eax", "%ebx", "%ecx", "memory"
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2881 );
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2882 //printf("%d\n", test);
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2883 #else
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2884 int y;
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2885 int d=0;
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2886 int sysd=0;
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2887
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2888 for(y=0; y<8; y++)
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2889 {
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2890 int x;
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2891 for(x=0; x<8; x++)
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2892 {
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2893 int ref= tempBlured[ x + y*stride ];
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2894 int cur= src[ x + y*stride ];
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2895 int d1=ref - cur;
157
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2896 // if(x==0 || x==7) d1+= d1>>1;
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2897 // if(y==0 || y==7) d1+= d1>>1;
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2898 // d+= ABS(d1);
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2899 d+= d1*d1;
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2900 sysd+= d1;
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2901 }
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2902 }
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2903 //printf("%d %d %d\n", maxNoise[0], maxNoise[1], maxNoise[2]);
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2904 /*
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2905 Switch between
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2906 1 0 0 0 0 0 0 (0)
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2907 64 32 16 8 4 2 1 (1)
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2908 64 48 36 27 20 15 11 (33) (approx)
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2909 64 56 49 43 37 33 29 (200) (approx)
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2910 */
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2911 if(d > maxNoise[1])
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2912 {
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2913 if(d < maxNoise[2])
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2914 {
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2915 for(y=0; y<8; y++)
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2916 {
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2917 int x;
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2918 for(x=0; x<8; x++)
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2919 {
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2920 int ref= tempBlured[ x + y*stride ];
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2921 int cur= src[ x + y*stride ];
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2922 tempBlured[ x + y*stride ]=
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2923 src[ x + y*stride ]=
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2924 (ref + cur + 1)>>1;
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2925 }
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2926 }
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2927 }
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2928 else
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2929 {
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2930 for(y=0; y<8; y++)
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2931 {
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2932 int x;
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2933 for(x=0; x<8; x++)
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2934 {
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2935 tempBlured[ x + y*stride ]= src[ x + y*stride ];
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2936 }
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2937 }
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2938 }
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2939 }
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2940 else
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2941 {
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2942 if(d < maxNoise[0])
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2943 {
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2944 for(y=0; y<8; y++)
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2945 {
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2946 int x;
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2947 for(x=0; x<8; x++)
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2948 {
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2949 int ref= tempBlured[ x + y*stride ];
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2950 int cur= src[ x + y*stride ];
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2951 tempBlured[ x + y*stride ]=
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2952 src[ x + y*stride ]=
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2953 (ref*7 + cur + 4)>>3;
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2954 }
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2955 }
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2956 }
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2957 else
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2958 {
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2959 for(y=0; y<8; y++)
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2960 {
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2961 int x;
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2962 for(x=0; x<8; x++)
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2963 {
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2964 int ref= tempBlured[ x + y*stride ];
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2965 int cur= src[ x + y*stride ];
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2966 tempBlured[ x + y*stride ]=
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2967 src[ x + y*stride ]=
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2968 (ref*3 + cur + 2)>>2;
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2969 }
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2970 }
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2971 }
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2972 }
157
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
2973 #endif
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2974 }
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2975
102
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
2976 #ifdef HAVE_ODIVX_POSTPROCESS
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
2977 #include "../opendivx/postprocess.h"
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
2978 int use_old_pp=0;
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
2979 #endif
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
2980
102
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
2981 static void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2982 QP_STORE_T QPs[], int QPStride, int isColor, struct PPMode *ppMode);
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
2983
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
2984 /* -pp Command line Help
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
2985 NOTE/FIXME: put this at an appropriate place (--help, html docs, man mplayer)?
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
2986
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
2987 -pp <filterName>[:<option>[:<option>...]][,[-]<filterName>[:<option>...]]...
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
2988
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
2989 long form example:
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
2990 -pp vdeblock:autoq,hdeblock:autoq,linblenddeint -pp default,-vdeblock
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
2991 short form example:
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
2992 -pp vb:a,hb:a,lb -pp de,-vb
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2993 more examples:
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
2994 -pp tn:64:128:256
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
2995
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
2996 Filters Options
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
2997 short long name short long option Description
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
2998 * * a autoq cpu power dependant enabler
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
2999 c chrom chrominance filtring enabled
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3000 y nochrom chrominance filtring disabled
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3001 hb hdeblock horizontal deblocking filter
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3002 vb vdeblock vertical deblocking filter
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3003 vr rkvdeblock
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3004 h1 x1hdeblock Experimental horizontal deblock filter 1
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3005 v1 x1vdeblock Experimental vertical deblock filter 1
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3006 dr dering not implemented yet
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3007 al autolevels automatic brightness / contrast fixer
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3008 f fullyrange stretch luminance range to (0..255)
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3009 lb linblenddeint linear blend deinterlacer
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3010 li linipoldeint linear interpolating deinterlacer
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3011 ci cubicipoldeint cubic interpolating deinterlacer
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3012 md mediandeint median deinterlacer
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3013 de default hdeblock:a,vdeblock:a,dering:a,autolevels
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3014 fa fast x1hdeblock:a,x1vdeblock:a,dering:a,autolevels
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3015 tn tmpnoise (3 Thresholds) Temporal Noise Reducer
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3016 */
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3017
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3018 /**
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3019 * returns a PPMode struct which will have a non 0 error variable if an error occured
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3020 * name is the string after "-pp" on the command line
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3021 * quality is a number from 0 to GET_PP_QUALITY_MAX
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3022 */
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3023 struct PPMode getPPModeByNameAndQuality(char *name, int quality)
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3024 {
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3025 char temp[GET_MODE_BUFFER_SIZE];
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3026 char *p= temp;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3027 char *filterDelimiters= ",";
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3028 char *optionDelimiters= ":";
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3029 struct PPMode ppMode= {0,0,0,0,0,0,{150,200,400}};
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3030 char *filterToken;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3031
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3032 strncpy(temp, name, GET_MODE_BUFFER_SIZE);
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3033
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3034 printf("%s\n", name);
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3035
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3036 for(;;){
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3037 char *filterName;
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3038 int q= 1000000; //GET_PP_QUALITY_MAX;
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3039 int chrom=-1;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3040 char *option;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3041 char *options[OPTIONS_ARRAY_SIZE];
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3042 int i;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3043 int filterNameOk=0;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3044 int numOfUnknownOptions=0;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3045 int enable=1; //does the user want us to enabled or disabled the filter
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3046
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3047 filterToken= strtok(p, filterDelimiters);
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3048 if(filterToken == NULL) break;
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3049 p+= strlen(filterToken) + 1; // p points to next filterToken
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3050 filterName= strtok(filterToken, optionDelimiters);
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3051 printf("%s::%s\n", filterToken, filterName);
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3052
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3053 if(*filterName == '-')
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3054 {
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3055 enable=0;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3056 filterName++;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3057 }
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3058
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3059 for(;;){ //for all options
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3060 option= strtok(NULL, optionDelimiters);
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3061 if(option == NULL) break;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3062
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3063 printf("%s\n", option);
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3064 if(!strcmp("autoq", option) || !strcmp("a", option)) q= quality;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3065 else if(!strcmp("nochrom", option) || !strcmp("y", option)) chrom=0;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3066 else if(!strcmp("chrom", option) || !strcmp("c", option)) chrom=1;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3067 else
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3068 {
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3069 options[numOfUnknownOptions] = option;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3070 numOfUnknownOptions++;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3071 }
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3072 if(numOfUnknownOptions >= OPTIONS_ARRAY_SIZE-1) break;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3073 }
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3074 options[numOfUnknownOptions] = NULL;
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3075
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3076 /* replace stuff from the replace Table */
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3077 for(i=0; replaceTable[2*i]!=NULL; i++)
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3078 {
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3079 if(!strcmp(replaceTable[2*i], filterName))
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3080 {
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3081 int newlen= strlen(replaceTable[2*i + 1]);
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3082 int plen;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3083 int spaceLeft;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3084
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3085 if(p==NULL) p= temp, *p=0; //last filter
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3086 else p--, *p=','; //not last filter
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3087
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3088 plen= strlen(p);
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3089 spaceLeft= (int)p - (int)temp + plen;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3090 if(spaceLeft + newlen >= GET_MODE_BUFFER_SIZE)
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3091 {
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3092 ppMode.error++;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3093 break;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3094 }
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3095 memmove(p + newlen, p, plen+1);
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3096 memcpy(p, replaceTable[2*i + 1], newlen);
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3097 filterNameOk=1;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3098 }
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3099 }
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3100
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3101 for(i=0; filters[i].shortName!=NULL; i++)
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3102 {
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3103 // printf("Compareing %s, %s, %s\n", filters[i].shortName,filters[i].longName, filterName);
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3104 if( !strcmp(filters[i].longName, filterName)
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3105 || !strcmp(filters[i].shortName, filterName))
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3106 {
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3107 ppMode.lumMode &= ~filters[i].mask;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3108 ppMode.chromMode &= ~filters[i].mask;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3109
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3110 filterNameOk=1;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3111 if(!enable) break; // user wants to disable it
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3112
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3113 if(q >= filters[i].minLumQuality)
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3114 ppMode.lumMode|= filters[i].mask;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3115 if(chrom==1 || (chrom==-1 && filters[i].chromDefault))
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3116 if(q >= filters[i].minChromQuality)
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3117 ppMode.chromMode|= filters[i].mask;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3118
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3119 if(filters[i].mask == LEVEL_FIX)
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3120 {
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3121 int o;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3122 ppMode.minAllowedY= 16;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3123 ppMode.maxAllowedY= 234;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3124 for(o=0; options[o]!=NULL; o++)
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3125 if( !strcmp(options[o],"fullyrange")
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3126 ||!strcmp(options[o],"f"))
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3127 {
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3128 ppMode.minAllowedY= 0;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3129 ppMode.maxAllowedY= 255;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3130 numOfUnknownOptions--;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3131 }
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3132 }
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3133 else if(filters[i].mask == TEMP_NOISE_FILTER)
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3134 {
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3135 int o;
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3136 int numOfNoises=0;
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3137 ppMode.maxTmpNoise[0]= 150;
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3138 ppMode.maxTmpNoise[1]= 200;
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3139 ppMode.maxTmpNoise[2]= 400;
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3140
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3141 for(o=0; options[o]!=NULL; o++)
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3142 {
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3143 char *tail;
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3144 ppMode.maxTmpNoise[numOfNoises]=
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3145 strtol(options[o], &tail, 0);
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3146 if(tail!=options[o])
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3147 {
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3148 numOfNoises++;
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3149 numOfUnknownOptions--;
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3150 if(numOfNoises >= 3) break;
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3151 }
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3152 }
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3153 }
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3154 }
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3155 }
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3156 if(!filterNameOk) ppMode.error++;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3157 ppMode.error += numOfUnknownOptions;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3158 }
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3159
152
561ec01a1a78 QP_store==null bugfix and no opendivx bugfix
michael
parents: 150
diff changeset
3160 #ifdef HAVE_ODIVX_POSTPROCESS
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3161 if(ppMode.lumMode & H_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_Y_H;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3162 if(ppMode.lumMode & V_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_Y_V;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3163 if(ppMode.chromMode & H_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_C_H;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3164 if(ppMode.chromMode & V_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_C_V;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3165 if(ppMode.lumMode & DERING) ppMode.oldMode |= PP_DERING_Y;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3166 if(ppMode.chromMode & DERING) ppMode.oldMode |= PP_DERING_C;
152
561ec01a1a78 QP_store==null bugfix and no opendivx bugfix
michael
parents: 150
diff changeset
3167 #endif
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3168
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3169 return ppMode;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3170 }
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3171
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3172 /**
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3173 * Obsolete, dont use it, use postprocess2() instead
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3174 */
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3175 void postprocess(unsigned char * src[], int src_stride,
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3176 unsigned char * dst[], int dst_stride,
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3177 int horizontal_size, int vertical_size,
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3178 QP_STORE_T *QP_store, int QP_stride,
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3179 int mode)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3180 {
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3181 struct PPMode ppMode;
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3182 static QP_STORE_T zeroArray[2048/8];
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3183 /*
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3184 static int qual=0;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3185
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3186 ppMode= getPPModeByNameAndQuality("fast,default,-hdeblock,-vdeblock,tmpnoise:150:200:300", qual);
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3187 printf("OK\n");
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3188 qual++;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3189 qual%=7;
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3190 printf("\n%X %X %X %X :%d: %d %d %d\n", ppMode.lumMode, ppMode.chromMode, ppMode.oldMode, ppMode.error,
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3191 qual, ppMode.maxTmpNoise[0], ppMode.maxTmpNoise[1], ppMode.maxTmpNoise[2]);
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3192 postprocess2(src, src_stride, dst, dst_stride,
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3193 horizontal_size, vertical_size, QP_store, QP_stride, &ppMode);
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3194
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3195 return;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3196 */
152
561ec01a1a78 QP_store==null bugfix and no opendivx bugfix
michael
parents: 150
diff changeset
3197 if(QP_store==NULL)
561ec01a1a78 QP_store==null bugfix and no opendivx bugfix
michael
parents: 150
diff changeset
3198 {
561ec01a1a78 QP_store==null bugfix and no opendivx bugfix
michael
parents: 150
diff changeset
3199 QP_store= zeroArray;
561ec01a1a78 QP_store==null bugfix and no opendivx bugfix
michael
parents: 150
diff changeset
3200 QP_stride= 0;
561ec01a1a78 QP_store==null bugfix and no opendivx bugfix
michael
parents: 150
diff changeset
3201 }
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3202
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3203 ppMode.lumMode= mode;
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3204 mode= ((mode&0xFF)>>4) | (mode&0xFFFFFF00);
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3205 ppMode.chromMode= mode;
157
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
3206 ppMode.maxTmpNoise[0]= 700;
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
3207 ppMode.maxTmpNoise[1]= 1500;
bc12fd7e6153 temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents: 156
diff changeset
3208 ppMode.maxTmpNoise[2]= 3000;
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3209
102
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3210 #ifdef HAVE_ODIVX_POSTPROCESS
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3211 // Note: I could make this shit outside of this file, but it would mean one
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3212 // more function call...
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3213 if(use_old_pp){
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3214 odivx_postprocess(src,src_stride,dst,dst_stride,horizontal_size,vertical_size,QP_store,QP_stride,mode);
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3215 return;
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3216 }
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3217 #endif
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3218
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3219 postProcess(src[0], src_stride, dst[0], dst_stride,
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3220 horizontal_size, vertical_size, QP_store, QP_stride, 0, &ppMode);
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3221
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3222 horizontal_size >>= 1;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3223 vertical_size >>= 1;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3224 src_stride >>= 1;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3225 dst_stride >>= 1;
142
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3226 // mode&= ~(LINEAR_IPOL_DEINT_FILTER | LINEAR_BLEND_DEINT_FILTER |
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3227 // MEDIAN_DEINT_FILTER | CUBIC_IPOL_DEINT_FILTER);
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3228
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3229 if(1)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3230 {
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3231 postProcess(src[1], src_stride, dst[1], dst_stride,
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3232 horizontal_size, vertical_size, QP_store, QP_stride, 1, &ppMode);
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3233 postProcess(src[2], src_stride, dst[2], dst_stride,
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3234 horizontal_size, vertical_size, QP_store, QP_stride, 2, &ppMode);
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3235 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3236 else
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3237 {
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3238 memset(dst[1], 128, dst_stride*vertical_size);
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3239 memset(dst[2], 128, dst_stride*vertical_size);
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3240 // memcpy(dst[1], src[1], src_stride*horizontal_size);
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3241 // memcpy(dst[2], src[2], src_stride*horizontal_size);
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3242 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3243 }
102
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3244
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3245 void postprocess2(unsigned char * src[], int src_stride,
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3246 unsigned char * dst[], int dst_stride,
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3247 int horizontal_size, int vertical_size,
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3248 QP_STORE_T *QP_store, int QP_stride,
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3249 struct PPMode *mode)
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3250 {
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3251
152
561ec01a1a78 QP_store==null bugfix and no opendivx bugfix
michael
parents: 150
diff changeset
3252 static QP_STORE_T zeroArray[2048/8];
561ec01a1a78 QP_store==null bugfix and no opendivx bugfix
michael
parents: 150
diff changeset
3253 if(QP_store==NULL)
561ec01a1a78 QP_store==null bugfix and no opendivx bugfix
michael
parents: 150
diff changeset
3254 {
561ec01a1a78 QP_store==null bugfix and no opendivx bugfix
michael
parents: 150
diff changeset
3255 QP_store= zeroArray;
561ec01a1a78 QP_store==null bugfix and no opendivx bugfix
michael
parents: 150
diff changeset
3256 QP_stride= 0;
561ec01a1a78 QP_store==null bugfix and no opendivx bugfix
michael
parents: 150
diff changeset
3257 }
561ec01a1a78 QP_store==null bugfix and no opendivx bugfix
michael
parents: 150
diff changeset
3258
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3259 #ifdef HAVE_ODIVX_POSTPROCESS
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3260 // Note: I could make this shit outside of this file, but it would mean one
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3261 // more function call...
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3262 if(use_old_pp){
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3263 odivx_postprocess(src,src_stride,dst,dst_stride,horizontal_size,vertical_size,QP_store,QP_stride,
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3264 mode->oldMode);
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3265 return;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3266 }
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3267 #endif
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3268
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3269 postProcess(src[0], src_stride, dst[0], dst_stride,
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3270 horizontal_size, vertical_size, QP_store, QP_stride, 0, mode);
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3271
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3272 horizontal_size >>= 1;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3273 vertical_size >>= 1;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3274 src_stride >>= 1;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3275 dst_stride >>= 1;
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3276
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3277 postProcess(src[1], src_stride, dst[1], dst_stride,
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3278 horizontal_size, vertical_size, QP_store, QP_stride, 1, mode);
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3279 postProcess(src[2], src_stride, dst[2], dst_stride,
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3280 horizontal_size, vertical_size, QP_store, QP_stride, 2, mode);
116
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3281 }
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3282
1895a8fa81ea auto brightness/ contrast bugfix
michael
parents: 115
diff changeset
3283
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3284 /**
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3285 * gets the mode flags for a given quality (larger values mean slower but better postprocessing)
102
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3286 * 0 <= quality <= 6
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3287 */
102
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3288 int getPpModeForQuality(int quality){
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3289 int modes[1+GET_PP_QUALITY_MAX]= {
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3290 0,
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3291 #if 1
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3292 // horizontal filters first
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3293 LUM_H_DEBLOCK,
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3294 LUM_H_DEBLOCK | LUM_V_DEBLOCK,
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3295 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK,
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3296 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK,
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3297 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK | LUM_DERING,
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3298 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK | LUM_DERING | CHROM_DERING
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3299 #else
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3300 // vertical filters first
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3301 LUM_V_DEBLOCK,
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3302 LUM_V_DEBLOCK | LUM_H_DEBLOCK,
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3303 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK,
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3304 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK,
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3305 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING,
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3306 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING | CHROM_DERING
102
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3307 #endif
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3308 };
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3309
102
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3310 #ifdef HAVE_ODIVX_POSTPROCESS
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3311 int odivx_modes[1+GET_PP_QUALITY_MAX]= {
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3312 0,
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3313 PP_DEBLOCK_Y_H,
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3314 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V,
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3315 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H,
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3316 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V,
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3317 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V|PP_DERING_Y,
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3318 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V|PP_DERING_Y|PP_DERING_C
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3319 };
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3320 if(use_old_pp) return odivx_modes[quality];
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3321 #endif
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3322 return modes[quality];
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3323 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3324
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3325 /**
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3326 * Copies a block from src to dst and fixes the blacklevel
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3327 * numLines must be a multiple of 4
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3328 * levelFix == 0 -> dont touch the brighness & contrast
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3329 */
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3330 static inline void blockCopy(uint8_t dst[], int dstStride, uint8_t src[], int srcStride,
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3331 int numLines, int levelFix)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3332 {
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
3333 #ifndef HAVE_MMX
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3334 int i;
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
3335 #endif
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3336 if(levelFix)
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3337 {
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3338 #ifdef HAVE_MMX
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3339 asm volatile(
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3340 "leal (%2,%2), %%eax \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3341 "leal (%3,%3), %%ebx \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3342 "movq packedYOffset, %%mm2 \n\t"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3343 "movq packedYScale, %%mm3 \n\t"
101
fcf4e8fcb34b fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents: 100
diff changeset
3344 "pxor %%mm4, %%mm4 \n\t"
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3345
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3346 #define SCALED_CPY \
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3347 "movq (%0), %%mm0 \n\t"\
118
3dd1950ac98d brightness / contrast fix/copy optimizations +2% speedup
michael
parents: 117
diff changeset
3348 "movq (%0), %%mm5 \n\t"\
101
fcf4e8fcb34b fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents: 100
diff changeset
3349 "punpcklbw %%mm4, %%mm0 \n\t"\
fcf4e8fcb34b fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents: 100
diff changeset
3350 "punpckhbw %%mm4, %%mm5 \n\t"\
117
a02f3088b0cf negative black bugfix
michael
parents: 116
diff changeset
3351 "psubw %%mm2, %%mm0 \n\t"\
a02f3088b0cf negative black bugfix
michael
parents: 116
diff changeset
3352 "psubw %%mm2, %%mm5 \n\t"\
118
3dd1950ac98d brightness / contrast fix/copy optimizations +2% speedup
michael
parents: 117
diff changeset
3353 "movq (%0,%2), %%mm1 \n\t"\
117
a02f3088b0cf negative black bugfix
michael
parents: 116
diff changeset
3354 "psllw $6, %%mm0 \n\t"\
a02f3088b0cf negative black bugfix
michael
parents: 116
diff changeset
3355 "psllw $6, %%mm5 \n\t"\
101
fcf4e8fcb34b fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents: 100
diff changeset
3356 "pmulhw %%mm3, %%mm0 \n\t"\
118
3dd1950ac98d brightness / contrast fix/copy optimizations +2% speedup
michael
parents: 117
diff changeset
3357 "movq (%0,%2), %%mm6 \n\t"\
101
fcf4e8fcb34b fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents: 100
diff changeset
3358 "pmulhw %%mm3, %%mm5 \n\t"\
fcf4e8fcb34b fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents: 100
diff changeset
3359 "punpcklbw %%mm4, %%mm1 \n\t"\
118
3dd1950ac98d brightness / contrast fix/copy optimizations +2% speedup
michael
parents: 117
diff changeset
3360 "punpckhbw %%mm4, %%mm6 \n\t"\
117
a02f3088b0cf negative black bugfix
michael
parents: 116
diff changeset
3361 "psubw %%mm2, %%mm1 \n\t"\
118
3dd1950ac98d brightness / contrast fix/copy optimizations +2% speedup
michael
parents: 117
diff changeset
3362 "psubw %%mm2, %%mm6 \n\t"\
117
a02f3088b0cf negative black bugfix
michael
parents: 116
diff changeset
3363 "psllw $6, %%mm1 \n\t"\
118
3dd1950ac98d brightness / contrast fix/copy optimizations +2% speedup
michael
parents: 117
diff changeset
3364 "psllw $6, %%mm6 \n\t"\
101
fcf4e8fcb34b fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents: 100
diff changeset
3365 "pmulhw %%mm3, %%mm1 \n\t"\
118
3dd1950ac98d brightness / contrast fix/copy optimizations +2% speedup
michael
parents: 117
diff changeset
3366 "pmulhw %%mm3, %%mm6 \n\t"\
3dd1950ac98d brightness / contrast fix/copy optimizations +2% speedup
michael
parents: 117
diff changeset
3367 "addl %%eax, %0 \n\t"\
3dd1950ac98d brightness / contrast fix/copy optimizations +2% speedup
michael
parents: 117
diff changeset
3368 "packuswb %%mm5, %%mm0 \n\t"\
3dd1950ac98d brightness / contrast fix/copy optimizations +2% speedup
michael
parents: 117
diff changeset
3369 "packuswb %%mm6, %%mm1 \n\t"\
3dd1950ac98d brightness / contrast fix/copy optimizations +2% speedup
michael
parents: 117
diff changeset
3370 "movq %%mm0, (%1) \n\t"\
101
fcf4e8fcb34b fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents: 100
diff changeset
3371 "movq %%mm1, (%1, %3) \n\t"\
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3372
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3373 SCALED_CPY
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3374 "addl %%ebx, %1 \n\t"
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3375 SCALED_CPY
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3376 "addl %%ebx, %1 \n\t"
118
3dd1950ac98d brightness / contrast fix/copy optimizations +2% speedup
michael
parents: 117
diff changeset
3377 SCALED_CPY
3dd1950ac98d brightness / contrast fix/copy optimizations +2% speedup
michael
parents: 117
diff changeset
3378 "addl %%ebx, %1 \n\t"
3dd1950ac98d brightness / contrast fix/copy optimizations +2% speedup
michael
parents: 117
diff changeset
3379 SCALED_CPY
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3380
119
b2f0e40866b1 optimizations (+2% speedup)
michael
parents: 118
diff changeset
3381 : "+r"(src),
b2f0e40866b1 optimizations (+2% speedup)
michael
parents: 118
diff changeset
3382 "+r"(dst)
b2f0e40866b1 optimizations (+2% speedup)
michael
parents: 118
diff changeset
3383 :"r" (srcStride),
118
3dd1950ac98d brightness / contrast fix/copy optimizations +2% speedup
michael
parents: 117
diff changeset
3384 "r" (dstStride)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3385 : "%eax", "%ebx"
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3386 );
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3387 #else
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3388 for(i=0; i<numLines; i++)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3389 memcpy( &(dst[dstStride*i]),
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3390 &(src[srcStride*i]), BLOCK_SIZE);
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3391 #endif
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3392 }
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3393 else
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3394 {
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3395 #ifdef HAVE_MMX
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3396 asm volatile(
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3397 "movl %4, %%eax \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3398 "movl %%eax, temp0\n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3399 "pushl %0 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3400 "pushl %1 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3401 "leal (%2,%2), %%eax \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3402 "leal (%3,%3), %%ebx \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3403 "movq packedYOffset, %%mm2 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3404 "movq packedYScale, %%mm3 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3405
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3406 #define SIMPLE_CPY \
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3407 "movq (%0), %%mm0 \n\t"\
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3408 "movq (%0,%2), %%mm1 \n\t"\
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3409 "movq %%mm0, (%1) \n\t"\
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3410 "movq %%mm1, (%1, %3) \n\t"\
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3411
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3412 "1: \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3413 SIMPLE_CPY
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3414 "addl %%eax, %0 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3415 "addl %%ebx, %1 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3416 SIMPLE_CPY
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3417 "addl %%eax, %0 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3418 "addl %%ebx, %1 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3419 "decl temp0 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3420 "jnz 1b \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3421
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3422 "popl %1 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3423 "popl %0 \n\t"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3424 : : "r" (src),
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3425 "r" (dst),
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3426 "r" (srcStride),
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3427 "r" (dstStride),
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3428 "m" (numLines>>2)
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3429 : "%eax", "%ebx"
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3430 );
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3431 #else
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3432 for(i=0; i<numLines; i++)
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3433 memcpy( &(dst[dstStride*i]),
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3434 &(src[srcStride*i]), BLOCK_SIZE);
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3435 #endif
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3436 }
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3437 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3438
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3439
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3440 /**
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3441 * Filters array of bytes (Y or U or V values)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3442 */
102
2cb2660cbd7e small changes for mplayer integration:
arpi
parents: 101
diff changeset
3443 static void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3444 QP_STORE_T QPs[], int QPStride, int isColor, struct PPMode *ppMode)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3445 {
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3446 int x,y;
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3447 const int mode= isColor ? ppMode->chromMode : ppMode->lumMode;
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3448
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3449 /* we need 64bit here otherwise we´ll going to have a problem
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3450 after watching a black picture for 5 hours*/
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3451 static uint64_t *yHistogram= NULL;
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3452 int black=0, white=255; // blackest black and whitest white in the picture
126
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3453 int QPCorrecture= 256;
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3454
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3455 /* Temporary buffers for handling the last row(s) */
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3456 static uint8_t *tempDst= NULL;
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3457 static uint8_t *tempSrc= NULL;
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3458
112
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3459 /* Temporary buffers for handling the last block */
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3460 static uint8_t *tempDstBlock= NULL;
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3461 static uint8_t *tempSrcBlock= NULL;
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3462
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3463 /* Temporal noise reducing buffers */
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3464 static uint8_t *tempBlured[3]= {NULL,NULL,NULL};
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3465
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
3466 #ifdef PP_FUNNY_STRIDE
112
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3467 uint8_t *dstBlockPtrBackup;
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3468 uint8_t *srcBlockPtrBackup;
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
3469 #endif
112
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3470
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
3471 #ifdef MORE_TIMING
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
3472 long long T0, T1, diffTime=0;
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
3473 #endif
107
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
3474 #ifdef TIMING
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
3475 long long memcpyTime=0, vertTime=0, horizTime=0, sumTime;
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3476 sumTime= rdtsc();
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3477 #endif
140
52ed0baddd56 minor speedup
michael
parents: 135
diff changeset
3478 //mode= 0x7F;
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3479
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3480 if(tempDst==NULL)
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3481 {
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3482 tempDst= (uint8_t*)memalign(8, 1024*24);
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3483 tempSrc= (uint8_t*)memalign(8, 1024*24);
112
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3484 tempDstBlock= (uint8_t*)memalign(8, 1024*24);
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3485 tempSrcBlock= (uint8_t*)memalign(8, 1024*24);
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3486 }
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3487
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3488 if(tempBlured[isColor]==NULL && (mode & TEMP_NOISE_FILTER))
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3489 {
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3490 // printf("%d %d %d\n", isColor, dstStride, height);
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3491 //FIXME works only as long as the size doesnt increase
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3492 //Note:the +17*1024 is just there so i dont have to worry about r/w over te end
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3493 tempBlured[isColor]= (uint8_t*)memalign(8, dstStride*((height+7)&(~7)) + 17*1024);
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3494
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3495 memset(tempBlured[isColor], 0, dstStride*((height+7)&(~7)) + 17*1024);
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3496 }
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3497
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3498 if(!yHistogram)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3499 {
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3500 int i;
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3501 yHistogram= (uint64_t*)malloc(8*256);
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3502 for(i=0; i<256; i++) yHistogram[i]= width*height/64*15/256;
112
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3503
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3504 if(mode & FULL_Y_RANGE)
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3505 {
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3506 maxAllowedY=255;
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3507 minAllowedY=0;
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3508 }
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3509 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3510
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3511 if(!isColor)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3512 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3513 uint64_t sum= 0;
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3514 int i;
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3515 static int framenum= -1;
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3516 uint64_t maxClipped;
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3517 uint64_t clipped;
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3518 double scale;
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3519
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3520 framenum++;
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3521 if(framenum == 1) yHistogram[0]= width*height/64*15/256;
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3522
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3523 for(i=0; i<256; i++)
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3524 {
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3525 sum+= yHistogram[i];
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3526 // printf("%d ", yHistogram[i]);
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3527 }
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3528 // printf("\n\n");
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3529
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3530 /* we allways get a completly black picture first */
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3531 maxClipped= (uint64_t)(sum * maxClippedThreshold);
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3532
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3533 clipped= sum;
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3534 for(black=255; black>0; black--)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3535 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3536 if(clipped < maxClipped) break;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3537 clipped-= yHistogram[black];
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3538 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3539
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3540 clipped= sum;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3541 for(white=0; white<256; white++)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3542 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3543 if(clipped < maxClipped) break;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3544 clipped-= yHistogram[white];
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3545 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3546
117
a02f3088b0cf negative black bugfix
michael
parents: 116
diff changeset
3547 packedYOffset= (black - minAllowedY) & 0xFFFF;
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3548 packedYOffset|= packedYOffset<<32;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3549 packedYOffset|= packedYOffset<<16;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3550
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3551 scale= (double)(maxAllowedY - minAllowedY) / (double)(white-black);
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3552
117
a02f3088b0cf negative black bugfix
michael
parents: 116
diff changeset
3553 packedYScale= (uint16_t)(scale*1024.0 + 0.5);
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3554 packedYScale|= packedYScale<<32;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3555 packedYScale|= packedYScale<<16;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3556 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3557 else
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3558 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3559 packedYScale= 0x0100010001000100LL;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3560 packedYOffset= 0;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3561 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3562
126
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3563 if(mode & LEVEL_FIX) QPCorrecture= packedYScale &0xFFFF;
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3564 else QPCorrecture= 256;
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3565
148
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
3566 /* copy & deinterlace first row of blocks */
142
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3567 y=-BLOCK_SIZE;
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3568 {
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3569 //1% speedup if these are here instead of the inner loop
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3570 uint8_t *srcBlock= &(src[y*srcStride]);
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3571 uint8_t *dstBlock= &(dst[y*dstStride]);
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3572
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3573 dstBlock= tempDst + dstStride;
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3574
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3575 // From this point on it is guranteed that we can read and write 16 lines downward
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3576 // finish 1 block before the next otherwise we´ll might have a problem
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3577 // with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3578 for(x=0; x<width; x+=BLOCK_SIZE)
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3579 {
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3580
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3581 #ifdef HAVE_MMX2
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3582 /*
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3583 prefetchnta(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3584 prefetchnta(srcBlock + (((x>>3)&3) + 9)*srcStride + 32);
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3585 prefetcht0(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3586 prefetcht0(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3587 */
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3588 /*
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3589 prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32);
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3590 prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32);
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3591 prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32);
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3592 prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32);
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3593 */
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3594
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3595 asm(
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3596 "movl %4, %%eax \n\t"
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3597 "shrl $2, %%eax \n\t"
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3598 "andl $6, %%eax \n\t"
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3599 "addl $8, %%eax \n\t"
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3600 "movl %%eax, %%ebx \n\t"
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3601 "imul %1, %%eax \n\t"
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3602 "imul %3, %%ebx \n\t"
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3603 "prefetchnta 32(%%eax, %0) \n\t"
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3604 "prefetcht0 32(%%ebx, %2) \n\t"
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3605 "addl %1, %%eax \n\t"
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3606 "addl %3, %%ebx \n\t"
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3607 "prefetchnta 32(%%eax, %0) \n\t"
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3608 "prefetcht0 32(%%ebx, %2) \n\t"
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3609 :: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride),
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3610 "m" (x)
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3611 : "%eax", "%ebx"
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3612 );
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3613
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3614 #elif defined(HAVE_3DNOW)
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3615 //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ...
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3616 /* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3617 prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32);
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3618 prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3619 prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3620 */
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3621 #endif
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3622
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3623 blockCopy(dstBlock + dstStride*8, dstStride,
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3624 srcBlock + srcStride*8, srcStride, 8, mode & LEVEL_FIX);
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3625
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3626 if(mode & LINEAR_IPOL_DEINT_FILTER)
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3627 deInterlaceInterpolateLinear(dstBlock, dstStride);
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3628 else if(mode & LINEAR_BLEND_DEINT_FILTER)
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3629 deInterlaceBlendLinear(dstBlock, dstStride);
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3630 else if(mode & MEDIAN_DEINT_FILTER)
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3631 deInterlaceMedian(dstBlock, dstStride);
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3632 else if(mode & CUBIC_IPOL_DEINT_FILTER)
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3633 deInterlaceInterpolateCubic(dstBlock, dstStride);
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3634 /* else if(mode & CUBIC_BLEND_DEINT_FILTER)
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3635 deInterlaceBlendCubic(dstBlock, dstStride);
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3636 */
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3637 dstBlock+=8;
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3638 srcBlock+=8;
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3639 }
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3640 memcpy(&(dst[y*dstStride]) + 8*dstStride, tempDst + 9*dstStride, 8*dstStride );
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3641 }
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3642
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3643 for(y=0; y<height; y+=BLOCK_SIZE)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3644 {
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3645 //1% speedup if these are here instead of the inner loop
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3646 uint8_t *srcBlock= &(src[y*srcStride]);
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3647 uint8_t *dstBlock= &(dst[y*dstStride]);
126
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3648 #ifdef ARCH_X86
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3649 int *QPptr= isColor ? &QPs[(y>>3)*QPStride] :&QPs[(y>>4)*QPStride];
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3650 int QPDelta= isColor ? 1<<(32-3) : 1<<(32-4);
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3651 int QPFrac= QPDelta;
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3652 uint8_t *tempBlock1= tempBlocks;
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3653 uint8_t *tempBlock2= tempBlocks + 8;
126
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3654 #endif
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3655 int QP=0;
130
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
3656 /* can we mess with a 8x16 block from srcBlock/dstBlock downwards and 1 line upwards
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
3657 if not than use a temporary buffer */
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3658 if(y+15 >= height)
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3659 {
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3660 int i;
142
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3661 /* copy from line 8 to 15 of src, these will be copied with
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3662 blockcopy to dst later */
142
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3663 memcpy(tempSrc + srcStride*8, srcBlock + srcStride*8,
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3664 srcStride*MAX(height-y-8, 0) );
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3665
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3666 /* duplicate last line of src to fill the void upto line 15 */
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3667 for(i=MAX(height-y, 8); i<=15; i++)
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3668 memcpy(tempSrc + srcStride*i, src + srcStride*(height-1), srcStride);
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3669
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3670 /* copy up to 9 lines of dst (line -1 to 7)*/
142
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3671 memcpy(tempDst, dstBlock - dstStride, dstStride*MIN(height-y+1, 9) );
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3672
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3673 /* duplicate last line of dst to fill the void upto line 8 */
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3674 for(i=height-y+1; i<=8; i++)
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3675 memcpy(tempDst + dstStride*i, dst + dstStride*(height-1), dstStride);
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3676
130
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
3677 dstBlock= tempDst + dstStride;
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3678 srcBlock= tempSrc;
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3679 }
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3680
112
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3681 // From this point on it is guranteed that we can read and write 16 lines downward
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3682 // finish 1 block before the next otherwise we´ll might have a problem
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3683 // with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3684 for(x=0; x<width; x+=BLOCK_SIZE)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3685 {
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3686 const int stride= dstStride;
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3687 uint8_t *tmpXchg;
126
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3688 #ifdef ARCH_X86
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3689 QP= *QPptr;
126
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3690 asm volatile(
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3691 "addl %2, %1 \n\t"
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3692 "sbbl %%eax, %%eax \n\t"
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3693 "shll $2, %%eax \n\t"
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3694 "subl %%eax, %0 \n\t"
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3695 : "+r" (QPptr), "+m" (QPFrac)
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3696 : "r" (QPDelta)
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3697 : "%eax"
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3698 );
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3699 #else
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3700 QP= isColor ?
126
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3701 QPs[(y>>3)*QPStride + (x>>3)]:
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3702 QPs[(y>>4)*QPStride + (x>>4)];
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3703 #endif
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3704 if(!isColor)
121
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
3705 {
126
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3706 QP= (QP* QPCorrecture)>>8;
148
1cfc4d567c0a minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents: 142
diff changeset
3707 yHistogram[ srcBlock[srcStride*12 + 4] ]++;
121
3ecf2a90c65e more speed
michael
parents: 120
diff changeset
3708 }
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3709 #ifdef HAVE_MMX
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3710 asm volatile(
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3711 "movd %0, %%mm7 \n\t"
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3712 "packuswb %%mm7, %%mm7 \n\t" // 0, 0, 0, QP, 0, 0, 0, QP
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3713 "packuswb %%mm7, %%mm7 \n\t" // 0,QP, 0, QP, 0,QP, 0, QP
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3714 "packuswb %%mm7, %%mm7 \n\t" // QP,..., QP
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3715 "movq %%mm7, pQPb \n\t"
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3716 : : "r" (QP)
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3717 );
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3718 #endif
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3719
107
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
3720 #ifdef MORE_TIMING
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3721 T0= rdtsc();
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3722 #endif
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3723
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3724 #ifdef HAVE_MMX2
126
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3725 /*
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3726 prefetchnta(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3727 prefetchnta(srcBlock + (((x>>3)&3) + 9)*srcStride + 32);
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3728 prefetcht0(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3729 prefetcht0(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
126
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3730 */
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3731 /*
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3732 prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32);
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3733 prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32);
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3734 prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32);
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3735 prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32);
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3736 */
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3737
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3738 asm(
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3739 "movl %4, %%eax \n\t"
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3740 "shrl $2, %%eax \n\t"
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3741 "andl $6, %%eax \n\t"
142
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3742 "addl $8, %%eax \n\t"
126
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3743 "movl %%eax, %%ebx \n\t"
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3744 "imul %1, %%eax \n\t"
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3745 "imul %3, %%ebx \n\t"
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3746 "prefetchnta 32(%%eax, %0) \n\t"
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3747 "prefetcht0 32(%%ebx, %2) \n\t"
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3748 "addl %1, %%eax \n\t"
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3749 "addl %3, %%ebx \n\t"
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3750 "prefetchnta 32(%%eax, %0) \n\t"
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3751 "prefetcht0 32(%%ebx, %2) \n\t"
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3752 :: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride),
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3753 "m" (x)
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3754 : "%eax", "%ebx"
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3755 );
55f57883bbf5 more speed
michael
parents: 121
diff changeset
3756
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3757 #elif defined(HAVE_3DNOW)
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3758 //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ...
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3759 /* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3760 prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32);
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3761 prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3762 prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3763 */
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3764 #endif
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3765
113
3e0dcdb6b340 added compiletime option to turn width%8==0 on
michael
parents: 112
diff changeset
3766 #ifdef PP_FUNNY_STRIDE
112
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3767 //can we mess with a 8x16 block, if not use a temp buffer, yes again
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3768 if(x+7 >= width)
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3769 {
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3770 int i;
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3771 dstBlockPtrBackup= dstBlock;
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3772 srcBlockPtrBackup= srcBlock;
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3773
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3774 for(i=0;i<BLOCK_SIZE*2; i++)
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3775 {
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3776 memcpy(tempSrcBlock+i*srcStride, srcBlock+i*srcStride, width-x);
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3777 memcpy(tempDstBlock+i*dstStride, dstBlock+i*dstStride, width-x);
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3778 }
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3779
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3780 dstBlock= tempDstBlock;
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3781 srcBlock= tempSrcBlock;
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3782 }
113
3e0dcdb6b340 added compiletime option to turn width%8==0 on
michael
parents: 112
diff changeset
3783 #endif
112
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3784
142
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3785 blockCopy(dstBlock + dstStride*8, dstStride,
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3786 srcBlock + srcStride*8, srcStride, 8, mode & LEVEL_FIX);
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3787
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3788 if(mode & LINEAR_IPOL_DEINT_FILTER)
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3789 deInterlaceInterpolateLinear(dstBlock, dstStride);
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3790 else if(mode & LINEAR_BLEND_DEINT_FILTER)
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3791 deInterlaceBlendLinear(dstBlock, dstStride);
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3792 else if(mode & MEDIAN_DEINT_FILTER)
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3793 deInterlaceMedian(dstBlock, dstStride);
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3794 else if(mode & CUBIC_IPOL_DEINT_FILTER)
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3795 deInterlaceInterpolateCubic(dstBlock, dstStride);
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3796 /* else if(mode & CUBIC_BLEND_DEINT_FILTER)
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3797 deInterlaceBlendCubic(dstBlock, dstStride);
106
389391a6d0bf rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents: 105
diff changeset
3798 */
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3799
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3800 /* only deblock if we have 2 blocks */
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3801 if(y + 8 < height)
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3802 {
107
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
3803 #ifdef MORE_TIMING
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3804 T1= rdtsc();
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3805 memcpyTime+= T1-T0;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3806 T0=T1;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3807 #endif
115
4514b8e7f0f1 more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents: 113
diff changeset
3808 if(mode & V_RK1_FILTER)
4514b8e7f0f1 more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents: 113
diff changeset
3809 vertRK1Filter(dstBlock, stride, QP);
4514b8e7f0f1 more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents: 113
diff changeset
3810 else if(mode & V_X1_FILTER)
4514b8e7f0f1 more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents: 113
diff changeset
3811 vertX1Filter(dstBlock, stride, QP);
4514b8e7f0f1 more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents: 113
diff changeset
3812 else if(mode & V_DEBLOCK)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3813 {
115
4514b8e7f0f1 more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents: 113
diff changeset
3814 if( isVertDC(dstBlock, stride))
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3815 {
115
4514b8e7f0f1 more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents: 113
diff changeset
3816 if(isVertMinMaxOk(dstBlock, stride, QP))
4514b8e7f0f1 more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents: 113
diff changeset
3817 doVertLowPass(dstBlock, stride, QP);
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3818 }
115
4514b8e7f0f1 more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents: 113
diff changeset
3819 else
4514b8e7f0f1 more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents: 113
diff changeset
3820 doVertDefFilter(dstBlock, stride, QP);
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3821 }
107
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
3822 #ifdef MORE_TIMING
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3823 T1= rdtsc();
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3824 vertTime+= T1-T0;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3825 T0=T1;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3826 #endif
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3827 }
130
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
3828
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3829 #ifdef HAVE_MMX
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3830 transpose1(tempBlock1, tempBlock2, dstBlock, dstStride);
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3831 #endif
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3832 /* check if we have a previous block to deblock it with dstBlock */
112
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3833 if(x - 8 >= 0)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3834 {
107
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
3835 #ifdef MORE_TIMING
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3836 T0= rdtsc();
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3837 #endif
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3838 #ifdef HAVE_MMX
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3839 if(mode & H_RK1_FILTER)
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3840 vertRK1Filter(tempBlock1, 16, QP);
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3841 else if(mode & H_X1_FILTER)
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3842 vertX1Filter(tempBlock1, 16, QP);
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3843 else if(mode & H_DEBLOCK)
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3844 {
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3845 if( isVertDC(tempBlock1, 16))
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3846 {
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3847 if(isVertMinMaxOk(tempBlock1, 16, QP))
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3848 doVertLowPass(tempBlock1, 16, QP);
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3849 }
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3850 else
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3851 doVertDefFilter(tempBlock1, 16, QP);
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3852 }
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3853
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3854 transpose2(dstBlock-4, dstStride, tempBlock1 + 4*16);
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3855
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3856 #else
115
4514b8e7f0f1 more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents: 113
diff changeset
3857 if(mode & H_X1_FILTER)
4514b8e7f0f1 more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents: 113
diff changeset
3858 horizX1Filter(dstBlock-4, stride, QP);
4514b8e7f0f1 more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents: 113
diff changeset
3859 else if(mode & H_DEBLOCK)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3860 {
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3861 if( isHorizDC(dstBlock-4, stride))
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3862 {
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3863 if(isHorizMinMaxOk(dstBlock-4, stride, QP))
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3864 doHorizLowPass(dstBlock-4, stride, QP);
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3865 }
115
4514b8e7f0f1 more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents: 113
diff changeset
3866 else
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3867 doHorizDefFilter(dstBlock-4, stride, QP);
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3868 }
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3869 #endif
107
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
3870 #ifdef MORE_TIMING
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3871 T1= rdtsc();
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3872 horizTime+= T1-T0;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3873 T0=T1;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3874 #endif
130
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
3875 if(mode & DERING)
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
3876 {
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
3877 //FIXME filter first line
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
3878 if(y>0) dering(dstBlock - stride - 8, stride, QP);
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
3879 }
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3880
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3881 if(mode & TEMP_NOISE_FILTER)
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3882 {
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3883 tempNoiseReducer(dstBlock-8, stride,
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3884 tempBlured[isColor] + y*dstStride + x,
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3885 ppMode->maxTmpNoise);
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3886 }
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3887 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3888
113
3e0dcdb6b340 added compiletime option to turn width%8==0 on
michael
parents: 112
diff changeset
3889 #ifdef PP_FUNNY_STRIDE
112
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3890 /* did we use a tmp-block buffer */
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3891 if(x+7 >= width)
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3892 {
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3893 int i;
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3894 dstBlock= dstBlockPtrBackup;
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3895 srcBlock= srcBlockPtrBackup;
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3896
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3897 for(i=0;i<BLOCK_SIZE*2; i++)
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3898 {
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3899 memcpy(dstBlock+i*dstStride, tempDstBlock+i*dstStride, width-x);
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3900 }
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3901 }
113
3e0dcdb6b340 added compiletime option to turn width%8==0 on
michael
parents: 112
diff changeset
3902 #endif
112
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3903
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3904 dstBlock+=8;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3905 srcBlock+=8;
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3906
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
3907 #ifdef HAVE_MMX
128
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3908 tmpXchg= tempBlock1;
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3909 tempBlock1= tempBlock2;
e5266b8e79be much better horizontal filters (transpose & use the vertical ones) :)
michael
parents: 126
diff changeset
3910 tempBlock2 = tmpXchg;
129
be35346e27c1 fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents: 128
diff changeset
3911 #endif
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3912 }
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3913
156
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3914 if(mode & DERING)
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3915 {
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3916 if(y > 0) dering(dstBlock - dstStride - 8, dstStride, QP);
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3917 }
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3918
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3919 if((mode & TEMP_NOISE_FILTER))
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3920 {
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3921 tempNoiseReducer(dstBlock-8, dstStride,
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3922 tempBlured[isColor] + y*dstStride + x,
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3923 ppMode->maxTmpNoise);
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3924 }
c09459686be3 temporal noise reducer in C (-pp 0x100000)
michael
parents: 152
diff changeset
3925
142
da4c751fc151 deinterlace bugfix
michael
parents: 141
diff changeset
3926 /* did we use a tmp buffer for the last lines*/
112
a2c063b6ecf9 fixed a bug in the tmp buffer
michael
parents: 111
diff changeset
3927 if(y+15 >= height)
111
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3928 {
8e4c5a16c9fc fixed the height%8!=0 bug
michael
parents: 109
diff changeset
3929 uint8_t *dstBlock= &(dst[y*dstStride]);
130
0cce5d30d1d8 dering in mmx2
michael
parents: 129
diff changeset
3930 memcpy(dstBlock, tempDst + dstStride, dstStride*(height-y) );
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3931 }
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3932 }
96
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3933 #ifdef HAVE_3DNOW
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3934 asm volatile("femms");
29ac11dc53d3 fixed a bug in the horizontal default filter
arpi
parents: 95
diff changeset
3935 #elif defined (HAVE_MMX)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3936 asm volatile("emms");
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3937 #endif
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3938
107
bd163e13a0fb minor cleanups
michael
parents: 106
diff changeset
3939 #ifdef TIMING
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3940 // FIXME diff is mostly the time spent for rdtsc (should subtract that but ...)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3941 sumTime= rdtsc() - sumTime;
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3942 if(!isColor)
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3943 printf("cpy:%4dk, vert:%4dk, horiz:%4dk, sum:%4dk, diff:%4dk, color: %d/%d \r",
97
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3944 (int)(memcpyTime/1000), (int)(vertTime/1000), (int)(horizTime/1000),
e57b1d38d71f bugfixes: last 3 lines not brightness/contrast corrected
michael
parents: 96
diff changeset
3945 (int)(sumTime/1000), (int)((sumTime-memcpyTime-vertTime-horizTime)/1000)
95
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3946 , black, white);
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3947 #endif
8bce253b537c new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff changeset
3948 }