Mercurial > libavcodec.hg
annotate libpostproc/postprocess.c @ 96:29ac11dc53d3 libavcodec
fixed a bug in the horizontal default filter
3dnow version of the Horizontal & Vertical Lowpass filters
mmx version of the Horizontal Default filter
mmx2 & C versions of a simple filter described in a paper from ramkishor & karan
added mode flags & quality2mode function
author | arpi |
---|---|
date | Wed, 10 Oct 2001 22:21:19 +0000 |
parents | 8bce253b537c |
children | e57b1d38d71f |
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 /* |
96 | 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 |
96 | 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 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
26 isHorizMinMaxOk a |
96 | 27 doHorizLowPass E a a* |
28 doHorizDefFilter E ac ac | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
29 deRing |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
30 |
96 | 31 * i dont have a 3dnow CPU -> its untested |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
32 E = Exact implementation |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
33 e = allmost exact implementation |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
34 a = alternative / approximate impl |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
35 c = checked against the other implementations (-vo md5) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
36 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
37 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
38 /* |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
39 TODO: |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
40 verify that everything workes as it should |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
41 reduce the time wasted on the mem transfer |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
42 implement dering |
96 | 43 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
|
44 figure range of QP out (assuming <256 for now) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
45 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
|
46 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
|
47 move YScale thing to the end instead of fixing QP |
96 | 48 write a faster and higher quality deblocking filter :) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
49 ... |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
50 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
51 Notes: |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
52 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
53 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
54 |
96 | 55 /* |
56 Changelog: | |
57 0.1.2 | |
58 fixed a bug in the horizontal default filter | |
59 3dnow version of the Horizontal & Vertical Lowpass filters | |
60 mmx version of the Horizontal Default filter | |
61 mmx2 & C versions of a simple filter described in a paper from ramkishor & karandikar | |
62 added mode flags & quality2mode function | |
63 0.1.1 | |
64 */ | |
65 | |
95
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 #include <inttypes.h> |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
68 #include <stdio.h> |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
69 #include "../config.h" |
96 | 70 //#undef HAVE_MMX2 |
71 //#define HAVE_3DNOW | |
72 //#undef HAVE_MMX | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
73 #include "postprocess.h" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
74 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
75 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
76 static uint64_t packedYOffset= 0x0000000000000000LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
77 static uint64_t packedYScale= 0x0100010001000100LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
78 static uint64_t w05= 0x0005000500050005LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
79 static uint64_t w20= 0x0020002000200020LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
80 static uint64_t w1400= 0x1400140014001400LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
81 static uint64_t bm00000001= 0x00000000000000FFLL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
82 static uint64_t bm00010000= 0x000000FF00000000LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
83 static uint64_t bm00001000= 0x00000000FF000000LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
84 static uint64_t bm10000000= 0xFF00000000000000LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
85 static uint64_t bm10000001= 0xFF000000000000FFLL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
86 static uint64_t bm11000011= 0xFFFF00000000FFFFLL; |
96 | 87 static uint64_t bm00000011= 0x000000000000FFFFLL; |
88 static uint64_t bm11000000= 0xFFFF000000000000LL; | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
89 static uint64_t bm00011000= 0x000000FFFF000000LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
90 static uint64_t bm00110011= 0x0000FFFF0000FFFFLL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
91 static uint64_t bm11001100= 0xFFFF0000FFFF0000LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
92 static uint64_t b00= 0x0000000000000000LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
93 static uint64_t b02= 0x0202020202020202LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
94 static uint64_t b0F= 0x0F0F0F0F0F0F0F0FLL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
95 static uint64_t bFF= 0xFFFFFFFFFFFFFFFFLL; |
96 | 96 static uint64_t b20= 0x2020202020202020LL; |
97 static uint64_t b80= 0x8080808080808080LL; | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
98 static uint64_t b7E= 0x7E7E7E7E7E7E7E7ELL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
99 static uint64_t b7C= 0x7C7C7C7C7C7C7C7CLL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
100 static uint64_t b3F= 0x3F3F3F3F3F3F3F3FLL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
101 static uint64_t temp0=0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
102 static uint64_t temp1=0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
103 static uint64_t temp2=0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
104 static uint64_t temp3=0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
105 static uint64_t temp4=0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
106 static uint64_t temp5=0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
107 static uint64_t pQPb=0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
108 static uint8_t tempBlock[16*16]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
109 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
110 int hFlatnessThreshold= 56 - 16; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
111 int vFlatnessThreshold= 56 - 16; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
112 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
113 //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
|
114 double maxClippedThreshold= 0.01; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
115 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
116 int maxAllowedY=255; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
117 //FIXME can never make a movie´s black brighter (anyone needs that?) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
118 int minAllowedY=0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
119 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
120 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
121 static inline long long rdtsc() |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
122 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
123 long long l; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
124 asm volatile( "rdtsc\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
125 : "=A" (l) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
126 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
127 // printf("%d\n", int(l/1000)); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
128 return l; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
129 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
130 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
131 static inline void prefetchnta(void *p) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
132 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
133 asm volatile( "prefetchnta (%0)\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
134 : : "r" (p) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
135 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
136 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
137 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
138 static inline void prefetcht0(void *p) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
139 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
140 asm volatile( "prefetcht0 (%0)\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
141 : : "r" (p) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
142 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
143 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
144 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
145 static inline void prefetcht1(void *p) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
146 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
147 asm volatile( "prefetcht1 (%0)\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
148 : : "r" (p) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
149 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
150 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
151 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
152 static inline void prefetcht2(void *p) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
153 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
154 asm volatile( "prefetcht2 (%0)\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
155 : : "r" (p) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
156 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
157 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
158 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
159 //FIXME? |255-0| = 1 (shouldnt be a problem ...) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
160 /** |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
161 * Check if the middle 8x8 Block in the given 8x10 block is flat |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
162 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
163 static inline bool isVertDC(uint8_t src[], int stride){ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
164 // return true; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
165 int numEq= 0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
166 src+= stride; // src points to begin of the 8x8 Block |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
167 #ifdef HAVE_MMX |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
168 asm volatile( |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
169 // "int $3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
170 "pushl %1\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
171 "movq b7E, %%mm7 \n\t" // mm7 = 0x7F |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
172 "movq b7C, %%mm6 \n\t" // mm6 = 0x7D |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
173 "movq (%1), %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
174 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
175 "movq (%1), %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
176 "psubb %%mm1, %%mm0 \n\t" // mm0 = differnece |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
177 "paddb %%mm7, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
178 "pcmpgtb %%mm6, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
179 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
180 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
181 "movq (%1), %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
182 "psubb %%mm2, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
183 "paddb %%mm7, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
184 "pcmpgtb %%mm6, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
185 "paddb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
186 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
187 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
188 "movq (%1), %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
189 "psubb %%mm1, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
190 "paddb %%mm7, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
191 "pcmpgtb %%mm6, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
192 "paddb %%mm2, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
193 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
194 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
195 "movq (%1), %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
196 "psubb %%mm2, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
197 "paddb %%mm7, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
198 "pcmpgtb %%mm6, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
199 "paddb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
200 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
201 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
202 "movq (%1), %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
203 "psubb %%mm1, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
204 "paddb %%mm7, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
205 "pcmpgtb %%mm6, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
206 "paddb %%mm2, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
207 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
208 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
209 "movq (%1), %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
210 "psubb %%mm2, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
211 "paddb %%mm7, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
212 "pcmpgtb %%mm6, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
213 "paddb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
214 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
215 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
216 "movq (%1), %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
217 "psubb %%mm1, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
218 "paddb %%mm7, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
219 "pcmpgtb %%mm6, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
220 "paddb %%mm2, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
221 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
222 " \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
223 "movq %%mm0, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
224 "psrlw $8, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
225 "paddb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
226 "movq %%mm0, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
227 "psrlq $16, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
228 "paddb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
229 "movq %%mm0, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
230 "psrlq $32, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
231 "paddb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
232 "popl %1\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
233 "movd %%mm0, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
234 : "=r" (numEq) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
235 : "r" (src), "r" (stride) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
236 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
237 // printf("%d\n", numEq); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
238 numEq= (256 - (numEq & 0xFF)) &0xFF; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
239 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
240 // int asmEq= numEq; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
241 // numEq=0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
242 // uint8_t *temp= src; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
243 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
244 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
245 for(int y=0; y<BLOCK_SIZE-1; y++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
246 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
247 if(((src[0] - src[0+stride] + 1)&0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
248 if(((src[1] - src[1+stride] + 1)&0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
249 if(((src[2] - src[2+stride] + 1)&0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
250 if(((src[3] - src[3+stride] + 1)&0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
251 if(((src[4] - src[4+stride] + 1)&0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
252 if(((src[5] - src[5+stride] + 1)&0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
253 if(((src[6] - src[6+stride] + 1)&0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
254 if(((src[7] - src[7+stride] + 1)&0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
255 src+= stride; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
256 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
257 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
258 /* if(abs(numEq - asmEq) > 0) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
259 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
260 printf("\nasm:%d c:%d\n", asmEq, numEq); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
261 for(int y=0; y<8; y++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
262 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
263 for(int x=0; x<8; x++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
264 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
265 printf("%d ", temp[x + y*stride]); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
266 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
267 printf("\n"); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
268 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
269 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
270 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
271 return numEq > vFlatnessThreshold; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
272 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
273 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
274 static inline bool isVertMinMaxOk(uint8_t src[], int stride, int QP) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
275 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
276 #ifdef HAVE_MMX |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
277 int isOk; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
278 asm volatile( |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
279 // "int $3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
280 "movq (%1, %2), %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
281 "movq (%1, %2, 8), %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
282 "movq %%mm0, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
283 "psubusb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
284 "psubusb %%mm2, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
285 "por %%mm1, %%mm0 \n\t" // ABS Diff |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
286 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
287 "movq pQPb, %%mm7 \n\t" // QP,..., QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
288 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
289 "psubusb %%mm7, %%mm0 \n\t" // Diff <= 2QP -> 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
290 "pcmpeqd b00, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
291 "psrlq $16, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
292 "pcmpeqd bFF, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
293 // "movd %%mm0, (%1, %2, 4)\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
294 "movd %%mm0, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
295 : "=r" (isOk) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
296 : "r" (src), "r" (stride) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
297 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
298 return isOk; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
299 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
300 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
301 int isOk2= true; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
302 for(int x=0; x<BLOCK_SIZE; x++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
303 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
304 if(abs((int)src[x + stride] - (int)src[x + (stride<<3)]) > 2*QP) isOk2=false; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
305 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
306 /* if(isOk && !isOk2 || !isOk && isOk2) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
307 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
308 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
|
309 for(int y=0; y<9; y++) |
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 for(int x=0; x<8; x++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
312 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
313 printf("%d ", src[x + y*stride]); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
314 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
315 printf("\n"); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
316 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
317 } */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
318 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
319 return isOk2; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
320 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
321 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
322 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
323 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
324 /** |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
325 * Do a vertical low pass filter on the 8x10 block (only write to the 8x8 block in the middle) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
326 * useing the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
327 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
328 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
|
329 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
330 // QP= 64; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
331 |
96 | 332 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
333 //#ifdef HAVE_MMX2 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
334 asm volatile( //"movv %0 %1 %2\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
335 "pushl %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
336 "movq pQPb, %%mm0 \n\t" // QP,..., QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
337 // "movq bFF , %%mm0 \n\t" // QP,..., QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
338 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
339 "movq (%0), %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
340 "movq (%0, %1), %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
341 "movq %%mm5, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
342 "movq %%mm6, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
343 "psubusb %%mm6, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
344 "psubusb %%mm1, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
345 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
346 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
347 "pcmpeqb b00, %%mm2 \n\t" // diff <= QP -> FF |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
348 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
349 "pand %%mm2, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
350 "pandn %%mm1, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
351 "por %%mm2, %%mm6 \n\t"// First Line to Filter |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
352 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
353 "movq (%0, %1, 8), %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
354 "leal (%0, %1, 4), %%eax \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
355 "leal (%0, %1, 8), %%ebx \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
356 "subl %1, %%ebx \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
357 "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
|
358 "movq (%0, %1, 8), %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
359 "movq %%mm5, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
360 "movq %%mm7, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
361 "psubusb %%mm7, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
362 "psubusb %%mm1, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
363 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
364 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
365 "pcmpeqb b00, %%mm2 \n\t" // diff <= QP -> FF |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
366 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
367 "pand %%mm2, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
368 "pandn %%mm1, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
369 "por %%mm2, %%mm7 \n\t" // First Line to Filter |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
370 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
371 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
372 // 1 2 3 4 5 6 7 8 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
373 // %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
|
374 // 6 4 2 2 1 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
375 // 6 4 4 2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
376 // 6 8 2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
377 /* |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
378 "movq %%mm6, %%mm2 \n\t" //1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
379 "movq %%mm6, %%mm3 \n\t" //1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
380 "paddusb b02, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
381 "psrlw $2, %%mm3 \n\t" //1 /4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
382 "pand b3F, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
383 "psubb %%mm3, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
384 "movq (%0, %1), %%mm0 \n\t" // 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
385 "movq %%mm0, %%mm1 \n\t" // 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
386 "paddusb b02, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
387 "psrlw $2, %%mm0 \n\t" // 1 /4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
388 "pand b3F, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
389 "paddusb %%mm2, %%mm0 \n\t" //3 1 /4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
390 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
391 "movq (%0, %1), %%mm0 \n\t" // 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
392 "movq %%mm0, %%mm1 \n\t" // 1 |
96 | 393 PAVGB(%%mm6, %%mm0) //1 1 /2 |
394 PAVGB(%%mm6, %%mm0) //3 1 /4 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
395 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
396 "movq (%0, %1, 4), %%mm2 \n\t" // 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
397 "movq %%mm2, %%mm5 \n\t" // 1 |
96 | 398 PAVGB((%%eax), %%mm2) // 11 /2 |
399 PAVGB((%0, %1, 2), %%mm2) // 211 /4 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
400 "movq %%mm2, %%mm3 \n\t" // 211 /4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
401 "movq (%0), %%mm4 \n\t" // 1 |
96 | 402 PAVGB(%%mm4, %%mm3) // 4 211 /8 |
403 PAVGB(%%mm0, %%mm3) //642211 /16 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
404 "movq %%mm3, (%0) \n\t" // X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
405 // 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
|
406 "movq %%mm1, %%mm0 \n\t" // 1 |
96 | 407 PAVGB(%%mm6, %%mm0) //1 1 /2 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
408 "movq %%mm4, %%mm3 \n\t" // 1 |
96 | 409 PAVGB((%0,%1,2), %%mm3) // 1 1 /2 |
410 PAVGB((%%eax,%1,2), %%mm5) // 11 /2 | |
411 PAVGB((%%eax), %%mm5) // 211 /4 | |
412 PAVGB(%%mm5, %%mm3) // 2 2211 /8 | |
413 PAVGB(%%mm0, %%mm3) //4242211 /16 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
414 "movq %%mm3, (%0,%1) \n\t" // X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
415 // mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9 |
96 | 416 PAVGB(%%mm4, %%mm6) //11 /2 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
417 "movq (%%ebx), %%mm0 \n\t" // 1 |
96 | 418 PAVGB((%%eax, %1, 2), %%mm0) // 11/2 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
419 "movq %%mm0, %%mm3 \n\t" // 11/2 |
96 | 420 PAVGB(%%mm1, %%mm0) // 2 11/4 |
421 PAVGB(%%mm6, %%mm0) //222 11/8 | |
422 PAVGB(%%mm2, %%mm0) //22242211/16 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
423 "movq (%0, %1, 2), %%mm2 \n\t" // 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
424 "movq %%mm0, (%0, %1, 2) \n\t" // X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
425 // 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
|
426 "movq (%%eax, %1, 4), %%mm0 \n\t" // 1 |
96 | 427 PAVGB((%%ebx), %%mm0) // 11 /2 |
428 PAVGB(%%mm0, %%mm6) //11 11 /4 | |
429 PAVGB(%%mm1, %%mm4) // 11 /2 | |
430 PAVGB(%%mm2, %%mm1) // 11 /2 | |
431 PAVGB(%%mm1, %%mm6) //1122 11 /8 | |
432 PAVGB(%%mm5, %%mm6) //112242211 /16 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
433 "movq (%%eax), %%mm5 \n\t" // 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
434 "movq %%mm6, (%%eax) \n\t" // X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
435 // 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
|
436 "movq (%%eax, %1, 4), %%mm6 \n\t" // 1 |
96 | 437 PAVGB(%%mm7, %%mm6) // 11 /2 |
438 PAVGB(%%mm4, %%mm6) // 11 11 /4 | |
439 PAVGB(%%mm3, %%mm6) // 11 2211 /8 | |
440 PAVGB(%%mm5, %%mm2) // 11 /2 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
441 "movq (%0, %1, 4), %%mm4 \n\t" // 1 |
96 | 442 PAVGB(%%mm4, %%mm2) // 112 /4 |
443 PAVGB(%%mm2, %%mm6) // 112242211 /16 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
444 "movq %%mm6, (%0, %1, 4) \n\t" // X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
445 // mm0=7(11) mm1=2(11) mm2=3(112) mm3=6(11) mm4=5 mm5=4 mm7=9 |
96 | 446 PAVGB(%%mm7, %%mm1) // 11 2 /4 |
447 PAVGB(%%mm4, %%mm5) // 11 /2 | |
448 PAVGB(%%mm5, %%mm0) // 11 11 /4 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
449 "movq (%%eax, %1, 2), %%mm6 \n\t" // 1 |
96 | 450 PAVGB(%%mm6, %%mm1) // 11 4 2 /8 |
451 PAVGB(%%mm0, %%mm1) // 11224222 /16 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
452 // "pxor %%mm1, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
453 "movq %%mm1, (%%eax, %1, 2) \n\t" // X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
454 // mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9 |
96 | 455 PAVGB((%%ebx), %%mm2) // 112 4 /8 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
456 "movq (%%eax, %1, 4), %%mm0 \n\t" // 1 |
96 | 457 PAVGB(%%mm0, %%mm6) // 1 1 /2 |
458 PAVGB(%%mm7, %%mm6) // 1 12 /4 | |
459 PAVGB(%%mm2, %%mm6) // 1122424 /4 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
460 // "pxor %%mm6, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
461 "movq %%mm6, (%%ebx) \n\t" // X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
462 // mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9 |
96 | 463 PAVGB(%%mm7, %%mm5) // 11 2 /4 |
464 PAVGB(%%mm7, %%mm5) // 11 6 /8 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
465 |
96 | 466 PAVGB(%%mm3, %%mm0) // 112 /4 |
467 PAVGB(%%mm0, %%mm5) // 112246 /16 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
468 // "pxor %%mm5, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
469 // "movq pQPb, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
470 "movq %%mm5, (%%eax, %1, 4) \n\t" // X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
471 "popl %0\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
472 |
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 : "r" (src), "r" (stride) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
475 : "%eax", "%ebx" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
476 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
477 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
478 const int l1= stride; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
479 const int l2= stride + l1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
480 const int l3= stride + l2; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
481 const int l4= stride + l3; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
482 const int l5= stride + l4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
483 const int l6= stride + l5; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
484 const int l7= stride + l6; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
485 const int l8= stride + l7; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
486 const int l9= stride + l8; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
487 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
488 for(int x=0; x<BLOCK_SIZE; x++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
489 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
490 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
|
491 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
|
492 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
493 int sums[9]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
494 sums[0] = first + src[l1]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
495 sums[1] = src[l1] + src[l2]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
496 sums[2] = src[l2] + src[l3]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
497 sums[3] = src[l3] + src[l4]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
498 sums[4] = src[l4] + src[l5]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
499 sums[5] = src[l5] + src[l6]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
500 sums[6] = src[l6] + src[l7]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
501 sums[7] = src[l7] + src[l8]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
502 sums[8] = src[l8] + last; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
503 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
504 src[l1]= ((sums[0]<<2) + ((first + sums[2])<<1) + sums[4] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
505 src[l2]= ((src[l2]<<2) + (first + sums[0] + sums[3]<<1) + sums[5] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
506 src[l3]= ((src[l3]<<2) + (first + sums[1] + sums[4]<<1) + sums[6] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
507 src[l4]= ((src[l4]<<2) + (sums[2] + sums[5]<<1) + sums[0] + sums[7] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
508 src[l5]= ((src[l5]<<2) + (sums[3] + sums[6]<<1) + sums[1] + sums[8] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
509 src[l6]= ((src[l6]<<2) + (last + sums[7] + sums[4]<<1) + sums[2] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
510 src[l7]= ((last + src[l7]<<2) + (src[l8] + sums[5]<<1) + sums[3] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
511 src[l8]= ((sums[8]<<2) + (last + sums[6]<<1) + sums[4] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
512 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
513 src++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
514 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
515 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
516 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
517 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
518 |
96 | 519 /** |
520 * Experimental implementation of the filter (Algorithm 1) described in a paper from Ramkishor & Karandikar | |
521 * values are correctly clipped (MMX2) | |
522 * values are wraparound (C) | |
523 * conclusion: its fast, but introduces ugly horizontal patterns if there is a continious gradient | |
524 0 8 16 24 | |
525 x = 8 | |
526 x/2 = 4 | |
527 x/8 = 1 | |
528 1 12 12 23 | |
529 */ | |
530 static inline void vertRKFilter(uint8_t *src, int stride, int QP) | |
531 { | |
532 #ifdef HAVE_MMX2 | |
533 // FIXME rounding | |
534 asm volatile( | |
535 "pxor %%mm7, %%mm7 \n\t" // 0 | |
536 "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE | |
537 "leal (%0, %1), %%eax \n\t" | |
538 "leal (%%eax, %1, 4), %%ebx \n\t" | |
539 // 0 1 2 3 4 5 6 7 8 9 | |
540 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 | |
541 "movq pQPb, %%mm0 \n\t" // QP,..., QP | |
542 "movq %%mm0, %%mm1 \n\t" // QP,..., QP | |
543 "paddusb b02, %%mm0 \n\t" | |
544 "psrlw $2, %%mm0 \n\t" | |
545 "pand b3F, %%mm0 \n\t" // QP/4,..., QP/4 | |
546 "paddusb %%mm1, %%mm0 \n\t" // QP*1.25 ... | |
547 "movq (%0, %1, 4), %%mm2 \n\t" // line 4 | |
548 "movq (%%ebx), %%mm3 \n\t" // line 5 | |
549 "movq %%mm2, %%mm4 \n\t" // line 4 | |
550 "pcmpeqb %%mm5, %%mm5 \n\t" // -1 | |
551 "pxor %%mm2, %%mm5 \n\t" // -line 4 - 1 | |
552 "pavgb %%mm3, %%mm5 \n\t" | |
553 "paddb %%mm6, %%mm5 \n\t" // (l5-l4)/2 | |
554 "psubusb %%mm3, %%mm4 \n\t" | |
555 "psubusb %%mm2, %%mm3 \n\t" | |
556 "por %%mm3, %%mm4 \n\t" // |l4 - l5| | |
557 "psubusb %%mm0, %%mm4 \n\t" | |
558 "pcmpeqb %%mm7, %%mm4 \n\t" | |
559 "pand %%mm4, %%mm5 \n\t" // d/2 | |
560 | |
561 // "paddb %%mm6, %%mm2 \n\t" // line 4 + 0x80 | |
562 "paddb %%mm5, %%mm2 \n\t" | |
563 // "psubb %%mm6, %%mm2 \n\t" | |
564 "movq %%mm2, (%0,%1, 4) \n\t" | |
565 | |
566 "movq (%%ebx), %%mm2 \n\t" | |
567 // "paddb %%mm6, %%mm2 \n\t" // line 5 + 0x80 | |
568 "psubb %%mm5, %%mm2 \n\t" | |
569 // "psubb %%mm6, %%mm2 \n\t" | |
570 "movq %%mm2, (%%ebx) \n\t" | |
571 | |
572 "paddb %%mm6, %%mm5 \n\t" | |
573 "psrlw $2, %%mm5 \n\t" | |
574 "pand b3F, %%mm5 \n\t" | |
575 "psubb b20, %%mm5 \n\t" // (l5-l4)/8 | |
576 | |
577 "movq (%%eax, %1, 2), %%mm2 \n\t" | |
578 "paddb %%mm6, %%mm2 \n\t" // line 3 + 0x80 | |
579 "paddsb %%mm5, %%mm2 \n\t" | |
580 "psubb %%mm6, %%mm2 \n\t" | |
581 "movq %%mm2, (%%eax, %1, 2) \n\t" | |
582 | |
583 "movq (%%ebx, %1), %%mm2 \n\t" | |
584 "paddb %%mm6, %%mm2 \n\t" // line 6 + 0x80 | |
585 "psubsb %%mm5, %%mm2 \n\t" | |
586 "psubb %%mm6, %%mm2 \n\t" | |
587 "movq %%mm2, (%%ebx, %1) \n\t" | |
588 | |
589 : | |
590 : "r" (src), "r" (stride) | |
591 : "%eax", "%ebx" | |
592 ); | |
593 #else | |
594 const int l1= stride; | |
595 const int l2= stride + l1; | |
596 const int l3= stride + l2; | |
597 const int l4= stride + l3; | |
598 const int l5= stride + l4; | |
599 const int l6= stride + l5; | |
600 const int l7= stride + l6; | |
601 const int l8= stride + l7; | |
602 const int l9= stride + l8; | |
603 for(int x=0; x<BLOCK_SIZE; x++) | |
604 { | |
605 if(ABS(src[l4]-src[l5]) < QP + QP/4) | |
606 { | |
607 int x = src[l5] - src[l4]; | |
608 | |
609 src[l3] +=x/8; | |
610 src[l4] +=x/2; | |
611 src[l5] -=x/2; | |
612 src[l6] -=x/8; | |
613 } | |
614 src++; | |
615 } | |
616 | |
617 #endif | |
618 } | |
619 | |
620 /** | |
621 * Experimental Filter 1 | |
622 */ | |
623 static inline void vertX1Filter(uint8_t *src, int stride, int QP) | |
624 { | |
625 #ifdef HAVE_MMX2X | |
626 // FIXME | |
627 asm volatile( | |
628 | |
629 : | |
630 : "r" (src), "r" (stride) | |
631 : "%eax", "%ebx" | |
632 ); | |
633 #else | |
634 const int l1= stride; | |
635 const int l2= stride + l1; | |
636 const int l3= stride + l2; | |
637 const int l4= stride + l3; | |
638 const int l5= stride + l4; | |
639 const int l6= stride + l5; | |
640 const int l7= stride + l6; | |
641 const int l8= stride + l7; | |
642 const int l9= stride + l8; | |
643 for(int x=0; x<BLOCK_SIZE; x++) | |
644 { | |
645 int v2= src[l2]; | |
646 int v3= src[l3]; | |
647 int v4= src[l4]; | |
648 int v5= src[l5]; | |
649 int v6= src[l6]; | |
650 int v7= src[l7]; | |
651 | |
652 if(ABS(v4-v5)<QP && ABS(v4-v5) - (ABS(v3-v4) + ABS(v5-v6))>0 ) | |
653 { | |
654 src[l3] = (6*v2 + 4*v3 + 3*v4 + 2*v5 + v6 )/16; | |
655 src[l4] = (3*v2 + 3*v3 + 4*v4 + 3*v5 + 2*v6 + v7 )/16; | |
656 src[l5] = (1*v2 + 2*v3 + 3*v4 + 4*v5 + 3*v6 + 3*v7)/16; | |
657 src[l6] = ( 1*v3 + 2*v4 + 3*v5 + 4*v6 + 6*v7)/16; | |
658 } | |
659 src++; | |
660 } | |
661 | |
662 #endif | |
663 } | |
664 | |
665 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
666 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
|
667 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
668 #ifdef HAVE_MMX |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
669 src+= stride; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
670 //FIXME try pmul for *5 stuff |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
671 // src[0]=0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
672 asm volatile( |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
673 "pxor %%mm7, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
674 "leal (%0, %1), %%eax \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
675 "leal (%%eax, %1, 4), %%ebx \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
676 // 0 1 2 3 4 5 6 7 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
677 // %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
|
678 // %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
|
679 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
680 "movq (%0), %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
681 "movq %%mm0, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
682 "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
683 "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
684 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
685 "movq (%%eax), %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
686 "movq %%mm2, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
687 "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
688 "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
689 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
690 "movq (%%eax, %1), %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
691 "movq %%mm4, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
692 "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
693 "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
694 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
695 "paddw %%mm0, %%mm0 \n\t" // 2L0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
696 "paddw %%mm1, %%mm1 \n\t" // 2H0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
697 "psubw %%mm4, %%mm2 \n\t" // L1 - L2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
698 "psubw %%mm5, %%mm3 \n\t" // H1 - H2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
699 "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
700 "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
701 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
702 "psllw $2, %%mm2 \n\t" // 4L1 - 4L2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
703 "psllw $2, %%mm3 \n\t" // 4H1 - 4H2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
704 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
705 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
706 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
707 "movq (%%eax, %1, 2), %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
708 "movq %%mm2, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
709 "punpcklbw %%mm7, %%mm2 \n\t" // L3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
710 "punpckhbw %%mm7, %%mm3 \n\t" // H3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
711 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
712 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
713 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
714 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
715 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
716 "movq %%mm0, temp0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
717 "movq %%mm1, temp1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
718 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
719 "movq (%0, %1, 4), %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
720 "movq %%mm0, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
721 "punpcklbw %%mm7, %%mm0 \n\t" // L4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
722 "punpckhbw %%mm7, %%mm1 \n\t" // H4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
723 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
724 "psubw %%mm0, %%mm2 \n\t" // L3 - L4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
725 "psubw %%mm1, %%mm3 \n\t" // H3 - H4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
726 "movq %%mm2, temp2 \n\t" // L3 - L4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
727 "movq %%mm3, temp3 \n\t" // H3 - H4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
728 "paddw %%mm4, %%mm4 \n\t" // 2L2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
729 "paddw %%mm5, %%mm5 \n\t" // 2H2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
730 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
731 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
732 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
733 "psllw $2, %%mm2 \n\t" // 4L3 - 4L4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
734 "psllw $2, %%mm3 \n\t" // 4H3 - 4H4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
735 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
736 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
737 //50 opcodes so far |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
738 "movq (%%ebx), %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
739 "movq %%mm2, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
740 "punpcklbw %%mm7, %%mm2 \n\t" // L5 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
741 "punpckhbw %%mm7, %%mm3 \n\t" // H5 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
742 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
743 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
744 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
745 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
746 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
747 "movq (%%ebx, %1), %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
748 "punpcklbw %%mm7, %%mm6 \n\t" // L6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
749 "psubw %%mm6, %%mm2 \n\t" // L5 - L6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
750 "movq (%%ebx, %1), %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
751 "punpckhbw %%mm7, %%mm6 \n\t" // H6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
752 "psubw %%mm6, %%mm3 \n\t" // H5 - H6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
753 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
754 "paddw %%mm0, %%mm0 \n\t" // 2L4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
755 "paddw %%mm1, %%mm1 \n\t" // 2H4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
756 "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
757 "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
758 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
759 "psllw $2, %%mm2 \n\t" // 4L5 - 4L6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
760 "psllw $2, %%mm3 \n\t" // 4H5 - 4H6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
761 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
762 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
763 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
764 "movq (%%ebx, %1, 2), %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
765 "movq %%mm2, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
766 "punpcklbw %%mm7, %%mm2 \n\t" // L7 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
767 "punpckhbw %%mm7, %%mm3 \n\t" // H7 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
768 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
769 "paddw %%mm2, %%mm2 \n\t" // 2L7 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
770 "paddw %%mm3, %%mm3 \n\t" // 2H7 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
771 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
772 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
773 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
774 "movq temp0, %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
775 "movq temp1, %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
776 //FIXME pxor, psubw, pmax for abs |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
777 "movq %%mm7, %%mm6 \n\t" // 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
778 "pcmpgtw %%mm0, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
779 "pxor %%mm6, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
780 "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
781 "movq %%mm7, %%mm6 \n\t" // 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
782 "pcmpgtw %%mm1, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
783 "pxor %%mm6, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
784 "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7| |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
785 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
786 "movq %%mm7, %%mm6 \n\t" // 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
787 "pcmpgtw %%mm2, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
788 "pxor %%mm6, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
789 "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3| |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
790 "movq %%mm7, %%mm6 \n\t" // 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
791 "pcmpgtw %%mm3, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
792 "pxor %%mm6, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
793 "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3| |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
794 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
795 #ifdef HAVE_MMX2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
796 "pminsw %%mm2, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
797 "pminsw %%mm3, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
798 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
799 "movq %%mm0, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
800 "psubusw %%mm2, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
801 "psubw %%mm6, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
802 "movq %%mm1, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
803 "psubusw %%mm3, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
804 "psubw %%mm6, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
805 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
806 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
807 "movq %%mm7, %%mm6 \n\t" // 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
808 "pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
809 "pxor %%mm6, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
810 "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5| |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
811 "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
812 "pxor %%mm7, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
813 "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5| |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
814 // 100 opcodes |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
815 "movd %2, %%mm2 \n\t" // QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
816 //"pcmpeqb %%mm2, %%mm2\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
817 "punpcklwd %%mm2, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
818 "punpcklwd %%mm2, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
819 "psllw $3, %%mm2 \n\t" // 8QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
820 "movq %%mm2, %%mm3 \n\t" // 8QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
821 "pcmpgtw %%mm4, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
822 "pcmpgtw %%mm5, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
823 "pand %%mm2, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
824 "pand %%mm3, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
825 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
826 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
827 "psubusw %%mm0, %%mm4 \n\t" // hd |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
828 "psubusw %%mm1, %%mm5 \n\t" // ld |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
829 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
830 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
831 "movq w05, %%mm2 \n\t" // 5 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
832 "pmullw %%mm2, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
833 "pmullw %%mm2, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
834 "movq w20, %%mm2 \n\t" // 32 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
835 "paddw %%mm2, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
836 "paddw %%mm2, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
837 "psrlw $6, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
838 "psrlw $6, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
839 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
840 /* |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
841 "movq w06, %%mm2 \n\t" // 6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
842 "paddw %%mm2, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
843 "paddw %%mm2, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
844 "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
|
845 //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
|
846 "pmulhw %%mm2, %%mm4 \n\t" // hd/13 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
847 "pmulhw %%mm2, %%mm5 \n\t" // ld/13 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
848 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
849 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
850 "movq temp2, %%mm0 \n\t" // L3 - L4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
851 "movq temp3, %%mm1 \n\t" // H3 - H4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
852 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
853 "pxor %%mm2, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
854 "pxor %%mm3, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
855 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
856 // FIXME rounding error |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
857 "psraw $1, %%mm0 \n\t" // (L3 - L4)/2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
858 "psraw $1, %%mm1 \n\t" // (H3 - H4)/2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
859 "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
860 "pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
861 "pxor %%mm2, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
862 "pxor %%mm3, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
863 "psubw %%mm2, %%mm0 \n\t" // |L3-L4| |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
864 "psubw %%mm3, %%mm1 \n\t" // |H3-H4| |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
865 // "psrlw $1, %%mm0 \n\t" // |L3 - L4|/2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
866 // "psrlw $1, %%mm1 \n\t" // |H3 - H4|/2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
867 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
868 "pxor %%mm6, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
869 "pxor %%mm7, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
870 "pand %%mm2, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
871 "pand %%mm3, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
872 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
873 #ifdef HAVE_MMX2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
874 "pminsw %%mm0, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
875 "pminsw %%mm1, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
876 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
877 "movq %%mm4, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
878 "psubusw %%mm0, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
879 "psubw %%mm2, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
880 "movq %%mm5, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
881 "psubusw %%mm1, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
882 "psubw %%mm2, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
883 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
884 "pxor %%mm6, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
885 "pxor %%mm7, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
886 "psubw %%mm6, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
887 "psubw %%mm7, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
888 "packsswb %%mm5, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
889 "movq (%%eax, %1, 2), %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
890 "paddb %%mm4, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
891 "movq %%mm0, (%%eax, %1, 2) \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
892 "movq (%0, %1, 4), %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
893 "psubb %%mm4, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
894 // "pxor %%mm0, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
895 "movq %%mm0, (%0, %1, 4) \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
896 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
897 : |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
898 : "r" (src), "r" (stride), "r" (QP) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
899 : "%eax", "%ebx" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
900 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
901 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
902 const int l1= stride; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
903 const int l2= stride + l1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
904 const int l3= stride + l2; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
905 const int l4= stride + l3; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
906 const int l5= stride + l4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
907 const int l6= stride + l5; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
908 const int l7= stride + l6; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
909 const int l8= stride + l7; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
910 // const int l9= stride + l8; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
911 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
912 for(int x=0; x<BLOCK_SIZE; x++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
913 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
914 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
|
915 if(ABS(middleEnergy) < 8*QP) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
916 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
917 const int q=(src[l4] - src[l5])/2; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
918 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
|
919 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
|
920 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
921 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
922 d= MAX(d, 0); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
923 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
924 d= (5*d + 32) >> 6; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
925 d*= SIGN(-middleEnergy); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
926 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
927 if(q>0) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
928 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
929 d= d<0 ? 0 : d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
930 d= d>q ? q : d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
931 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
932 else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
933 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
934 d= d>0 ? 0 : d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
935 d= d<q ? q : d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
936 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
937 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
938 src[l4]-= d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
939 src[l5]+= d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
940 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
941 src++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
942 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
943 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
944 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
945 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
946 //FIXME? |255-0| = 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
947 /** |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
948 * Check if the given 8x8 Block is mostly "flat" and copy the unaliged data into tempBlock. |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
949 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
950 static inline bool isHorizDCAndCopy2Temp(uint8_t src[], int stride) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
951 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
952 // src++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
953 int numEq= 0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
954 #ifdef HAVE_MMX |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
955 asm volatile ( |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
956 // "int $3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
957 "pushl %1\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
958 "movq b7E, %%mm7 \n\t" // mm7 = 0x7F |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
959 "movq b7C, %%mm6 \n\t" // mm6 = 0x7D |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
960 "leal tempBlock, %%eax \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
961 "pxor %%mm0, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
962 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
963 #define HDC_CHECK_AND_CPY(i) \ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
964 "movq -4(%1), %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
965 "psrlq $32, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
966 "punpckldq 4(%1), %%mm2 \n\t" /* (%1) */\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
967 "movq %%mm2, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
968 "psrlq $8, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
969 "psubb %%mm1, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
970 "paddb %%mm7, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
971 "pcmpgtb %%mm6, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
972 "paddb %%mm2, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
973 "movq %%mm1," #i "(%%eax) \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
974 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
975 HDC_CHECK_AND_CPY(0) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
976 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
977 HDC_CHECK_AND_CPY(8) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
978 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
979 HDC_CHECK_AND_CPY(16) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
980 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
981 HDC_CHECK_AND_CPY(24) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
982 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
983 HDC_CHECK_AND_CPY(32) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
984 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
985 HDC_CHECK_AND_CPY(40) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
986 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
987 HDC_CHECK_AND_CPY(48) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
988 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
989 HDC_CHECK_AND_CPY(56) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
990 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
991 "psllq $8, %%mm0 \n\t" // remove dummy value |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
992 "movq %%mm0, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
993 "psrlw $8, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
994 "paddb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
995 "movq %%mm0, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
996 "psrlq $16, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
997 "paddb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
998 "movq %%mm0, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
999 "psrlq $32, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1000 "paddb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1001 "popl %1\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1002 "movd %%mm0, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1003 : "=r" (numEq) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1004 : "r" (src), "r" (stride) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1005 : "%eax" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1006 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1007 // printf("%d\n", numEq); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1008 numEq= (256 - (numEq & 0xFF)) &0xFF; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1009 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1010 for(int y=0; y<BLOCK_SIZE; y++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1011 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1012 if(((src[0] - src[1] + 1) & 0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1013 if(((src[1] - src[2] + 1) & 0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1014 if(((src[2] - src[3] + 1) & 0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1015 if(((src[3] - src[4] + 1) & 0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1016 if(((src[4] - src[5] + 1) & 0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1017 if(((src[5] - src[6] + 1) & 0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1018 if(((src[6] - src[7] + 1) & 0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1019 tempBlock[0 + y*TEMP_STRIDE] = src[0]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1020 tempBlock[1 + y*TEMP_STRIDE] = src[1]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1021 tempBlock[2 + y*TEMP_STRIDE] = src[2]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1022 tempBlock[3 + y*TEMP_STRIDE] = src[3]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1023 tempBlock[4 + y*TEMP_STRIDE] = src[4]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1024 tempBlock[5 + y*TEMP_STRIDE] = src[5]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1025 tempBlock[6 + y*TEMP_STRIDE] = src[6]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1026 tempBlock[7 + y*TEMP_STRIDE] = src[7]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1027 src+= stride; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1028 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1029 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1030 /* if(abs(numEq - asmEq) > 0) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1031 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1032 // printf("\nasm:%d c:%d\n", asmEq, numEq); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1033 for(int y=0; y<8; y++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1034 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1035 for(int x=0; x<8; x++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1036 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1037 printf("%d ", src[x + y*stride]); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1038 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1039 printf("\n"); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1040 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1041 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1042 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1043 // printf("%d\n", numEq); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1044 return numEq > hFlatnessThreshold; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1045 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1046 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1047 static inline bool isHorizMinMaxOk(uint8_t src[], int stride, int QP) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1048 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1049 #ifdef MMX_FIXME |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1050 FIXME |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1051 int isOk; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1052 asm volatile( |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1053 // "int $3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1054 "movq (%1, %2), %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1055 "movq (%1, %2, 8), %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1056 "movq %%mm0, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1057 "psubusb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1058 "psubusb %%mm2, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1059 "por %%mm1, %%mm0 \n\t" // ABS Diff |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1060 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1061 "movq pQPb, %%mm7 \n\t" // QP,..., QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1062 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1063 "psubusb %%mm7, %%mm0 \n\t" // Diff <= 2QP -> 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1064 "pcmpeqd b00, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1065 "psrlq $16, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1066 "pcmpeqd bFF, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1067 // "movd %%mm0, (%1, %2, 4)\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1068 "movd %%mm0, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1069 : "=r" (isOk) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1070 : "r" (src), "r" (stride) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1071 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1072 return isOk; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1073 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1074 if(abs(src[0] - src[7]) > 2*QP) return false; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1075 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1076 return true; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1077 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1078 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1079 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1080 static inline void doHorizDefFilterAndCopyBack(uint8_t dst[], int stride, int QP) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1081 { |
96 | 1082 #ifdef HAVE_MMX |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1083 asm volatile( |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1084 "pushl %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1085 "pxor %%mm7, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1086 "movq bm00001000, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1087 "movd %2, %%mm5 \n\t" // QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1088 "movq %%mm5, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1089 "paddusb %%mm5, %%mm5 \n\t" // 2QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1090 "paddusb %%mm5, %%mm4 \n\t" // 3QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1091 "psllq $24, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1092 "pxor %%mm5, %%mm5 \n\t" // 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1093 "psubb %%mm4, %%mm5 \n\t" // -QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1094 "leal tempBlock, %%eax \n\t" |
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 //FIXME? "unroll by 2" and mix |
96 | 1097 #ifdef HAVE_MMX2 |
1098 #define HDF(i) \ | |
1099 "movq " #i "(%%eax), %%mm0 \n\t"\ | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1100 "movq %%mm0, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1101 "movq %%mm0, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1102 "psrlq $8, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1103 "psubusb %%mm1, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1104 "psubusb %%mm0, %%mm1 \n\t"\ |
96 | 1105 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ |
1106 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
1107 "pshufw $0x00, %%mm1, %%mm3 \n\t" /* p´5 = |p1 - p2| */\ | |
1108 "pminub %%mm1, %%mm3 \n\t" /* p´5 = min(|p2-p1|, |p6-p5|)*/\ | |
1109 "psrlq $16, %%mm3 \n\t" /* p´3 = min(|p2-p1|, |p6-p5|)*/\ | |
1110 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-min(|p1-p2|,|p5-p6|) */\ | |
1111 "paddb %%mm5, %%mm1 \n\t"\ | |
1112 "psubusb %%mm5, %%mm1 \n\t"\ | |
1113 "psrlw $2, %%mm1 \n\t"\ | |
1114 "pxor %%mm2, %%mm1 \n\t"\ | |
1115 "psubb %%mm2, %%mm1 \n\t"\ | |
1116 "pand %%mm6, %%mm1 \n\t"\ | |
1117 "psubb %%mm1, %%mm0 \n\t"\ | |
1118 "psllq $8, %%mm1 \n\t"\ | |
1119 "paddb %%mm1, %%mm0 \n\t"\ | |
1120 "movd %%mm0, (%0) \n\t"\ | |
1121 "psrlq $32, %%mm0 \n\t"\ | |
1122 "movd %%mm0, 4(%0) \n\t" | |
1123 #else | |
1124 #define HDF(i)\ | |
1125 "movq " #i "(%%eax), %%mm0 \n\t"\ | |
1126 "movq %%mm0, %%mm1 \n\t"\ | |
1127 "movq %%mm0, %%mm2 \n\t"\ | |
1128 "psrlq $8, %%mm1 \n\t"\ | |
1129 "psubusb %%mm1, %%mm2 \n\t"\ | |
1130 "psubusb %%mm0, %%mm1 \n\t"\ | |
1131 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ | |
1132 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
1133 "movq %%mm1, %%mm3 \n\t"\ | |
1134 "psllq $32, %%mm3 \n\t"\ | |
1135 "movq %%mm3, %%mm4 \n\t"\ | |
1136 "psubusb %%mm1, %%mm4 \n\t"\ | |
1137 "psubb %%mm4, %%mm3 \n\t"\ | |
1138 "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
|
1139 "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
|
1140 "paddb %%mm5, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1141 "psubusb %%mm5, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1142 "psrlw $2, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1143 "pxor %%mm2, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1144 "psubb %%mm2, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1145 "pand %%mm6, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1146 "psubb %%mm1, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1147 "psllq $8, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1148 "paddb %%mm1, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1149 "movd %%mm0, (%0) \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1150 "psrlq $32, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1151 "movd %%mm0, 4(%0) \n\t" |
96 | 1152 #endif |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1153 HDF(0) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1154 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1155 HDF(8) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1156 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1157 HDF(16) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1158 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1159 HDF(24) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1160 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1161 HDF(32) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1162 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1163 HDF(40) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1164 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1165 HDF(48) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1166 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1167 HDF(56) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1168 "popl %0 \n\t" |
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 : "r" (dst), "r" (stride), "r" (QP) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1171 : "%eax" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1172 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1173 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1174 uint8_t *src= tempBlock; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1175 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1176 for(int y=0; y<BLOCK_SIZE; y++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1177 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1178 dst[0] = src[0]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1179 dst[1] = src[1]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1180 dst[2] = src[2]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1181 dst[3] = src[3]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1182 dst[4] = src[4]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1183 dst[5] = src[5]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1184 dst[6] = src[6]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1185 dst[7] = src[7]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1186 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1187 const int middleEnergy= 5*(src[4] - src[5]) + 2*(src[2] - src[5]); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1188 if(ABS(middleEnergy) < 8*QP) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1189 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1190 const int q=(src[3] - src[4])/2; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1191 const int leftEnergy= 5*(src[2] - src[1]) + 2*(src[0] - src[3]); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1192 const int rightEnergy= 5*(src[6] - src[5]) + 2*(src[4] - src[7]); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1193 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1194 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1195 d= MAX(d, 0); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1196 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1197 d= (5*d + 32) >> 6; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1198 d*= SIGN(-middleEnergy); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1199 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1200 if(q>0) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1201 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1202 d= d<0 ? 0 : d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1203 d= d>q ? q : d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1204 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1205 else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1206 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1207 d= d>0 ? 0 : d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1208 d= d<q ? q : d; |
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 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1211 dst[3]-= d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1212 dst[4]+= d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1213 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1214 dst+= stride; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1215 src+= TEMP_STRIDE; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1216 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1217 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1218 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1219 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1220 /** |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1221 * Do a horizontal low pass filter on the 8x8 block |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1222 * useing the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 (C version) |
96 | 1223 * useing approximately the 7-Tap Filter (1,2,3,4,3,2,1)/16 (MMX2/3DNOW version) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1224 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1225 static inline void doHorizLowPassAndCopyBack(uint8_t dst[], int stride, int QP) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1226 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1227 //return; |
96 | 1228 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1229 asm volatile( //"movv %0 %1 %2\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1230 "pushl %0\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1231 "pxor %%mm7, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1232 "leal tempBlock, %%eax \n\t" |
96 | 1233 /* |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1234 #define HLP1 "movq (%0), %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1235 "movq %%mm0, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1236 "psllq $8, %%mm0 \n\t"\ |
96 | 1237 PAVGB(%%mm1, %%mm0)\ |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1238 "psrlw $8, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1239 "pxor %%mm1, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1240 "packuswb %%mm1, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1241 "movq %%mm0, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1242 "movq %%mm0, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1243 "psllq $32, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1244 "paddb %%mm0, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1245 "psllq $16, %%mm2 \n\t"\ |
96 | 1246 PAVGB(%%mm2, %%mm0)\ |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1247 "movq %%mm0, %%mm3 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1248 "pand bm11001100, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1249 "paddusb %%mm0, %%mm3 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1250 "psrlq $8, %%mm3 \n\t"\ |
96 | 1251 PAVGB(%%mm1, %%mm4)\ |
1252 PAVGB(%%mm3, %%mm2)\ | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1253 "psrlq $16, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1254 "punpcklbw %%mm2, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1255 "movq %%mm2, (%0) \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1256 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1257 #define HLP2 "movq (%0), %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1258 "movq %%mm0, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1259 "psllq $8, %%mm0 \n\t"\ |
96 | 1260 PAVGB(%%mm1, %%mm0)\ |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1261 "psrlw $8, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1262 "pxor %%mm1, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1263 "packuswb %%mm1, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1264 "movq %%mm0, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1265 "psllq $32, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1266 "psllq $16, %%mm2 \n\t"\ |
96 | 1267 PAVGB(%%mm2, %%mm0)\ |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1268 "movq %%mm0, %%mm3 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1269 "pand bm11001100, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1270 "paddusb %%mm0, %%mm3 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1271 "psrlq $8, %%mm3 \n\t"\ |
96 | 1272 PAVGB(%%mm3, %%mm2)\ |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1273 "psrlq $16, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1274 "punpcklbw %%mm2, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1275 "movq %%mm2, (%0) \n\t"\ |
96 | 1276 */ |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1277 // 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
|
1278 /* |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1279 31 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1280 121 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1281 121 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1282 121 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1283 121 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1284 121 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1285 121 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1286 13 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1287 Implemented Exact 7-Tap |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1288 9421 A321 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1289 36421 64321 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1290 334321 = |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1291 1234321 = |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1292 1234321 = |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1293 123433 = |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1294 12463 12346 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1295 1249 123A |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1296 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1297 */ |
96 | 1298 #ifdef HAVE_MMX2 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1299 #define HLP3(i) "movq " #i "(%%eax), %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1300 "movq %%mm0, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1301 "movq %%mm0, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1302 "movq %%mm0, %%mm3 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1303 "movq %%mm0, %%mm4 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1304 "psllq $8, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1305 "psrlq $8, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1306 "pand bm00000001, %%mm3 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1307 "pand bm10000000, %%mm4 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1308 "por %%mm3, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1309 "por %%mm4, %%mm2 \n\t"\ |
96 | 1310 PAVGB(%%mm2, %%mm1)\ |
1311 PAVGB(%%mm1, %%mm0)\ | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1312 \ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1313 "pshufw $0xF9, %%mm0, %%mm3 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1314 "pshufw $0x90, %%mm0, %%mm4 \n\t"\ |
96 | 1315 PAVGB(%%mm3, %%mm4)\ |
1316 PAVGB(%%mm4, %%mm0)\ | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1317 "movd %%mm0, (%0) \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1318 "psrlq $32, %%mm0 \n\t"\ |
96 | 1319 "movd %%mm0, 4(%0) \n\t" |
1320 #else | |
1321 #define HLP3(i) "movq " #i "(%%eax), %%mm0 \n\t"\ | |
1322 "movq %%mm0, %%mm1 \n\t"\ | |
1323 "movq %%mm0, %%mm2 \n\t"\ | |
1324 "movq %%mm0, %%mm3 \n\t"\ | |
1325 "movq %%mm0, %%mm4 \n\t"\ | |
1326 "psllq $8, %%mm1 \n\t"\ | |
1327 "psrlq $8, %%mm2 \n\t"\ | |
1328 "pand bm00000001, %%mm3 \n\t"\ | |
1329 "pand bm10000000, %%mm4 \n\t"\ | |
1330 "por %%mm3, %%mm1 \n\t"\ | |
1331 "por %%mm4, %%mm2 \n\t"\ | |
1332 PAVGB(%%mm2, %%mm1)\ | |
1333 PAVGB(%%mm1, %%mm0)\ | |
1334 \ | |
1335 "movq %%mm0, %%mm3 \n\t"\ | |
1336 "movq %%mm0, %%mm4 \n\t"\ | |
1337 "movq %%mm0, %%mm5 \n\t"\ | |
1338 "psrlq $16, %%mm3 \n\t"\ | |
1339 "psllq $16, %%mm4 \n\t"\ | |
1340 "pand bm11000000, %%mm5 \n\t"\ | |
1341 "por %%mm5, %%mm3 \n\t"\ | |
1342 "movq %%mm0, %%mm5 \n\t"\ | |
1343 "pand bm00000011, %%mm5 \n\t"\ | |
1344 "por %%mm5, %%mm4 \n\t"\ | |
1345 PAVGB(%%mm3, %%mm4)\ | |
1346 PAVGB(%%mm4, %%mm0)\ | |
1347 "movd %%mm0, (%0) \n\t"\ | |
1348 "psrlq $32, %%mm0 \n\t"\ | |
1349 "movd %%mm0, 4(%0) \n\t" | |
1350 #endif | |
95
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 #define HLP(i) HLP3(i) |
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 HLP(0) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1355 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1356 HLP(8) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1357 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1358 HLP(16) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1359 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1360 HLP(24) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1361 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1362 HLP(32) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1363 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1364 HLP(40) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1365 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1366 HLP(48) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1367 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1368 HLP(56) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1369 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1370 "popl %0\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1371 : |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1372 : "r" (dst), "r" (stride) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1373 : "%eax", "%ebx" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1374 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1375 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1376 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1377 uint8_t *temp= tempBlock; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1378 for(int y=0; y<BLOCK_SIZE; y++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1379 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1380 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
|
1381 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
|
1382 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1383 int sums[9]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1384 sums[0] = first + temp[0]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1385 sums[1] = temp[0] + temp[1]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1386 sums[2] = temp[1] + temp[2]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1387 sums[3] = temp[2] + temp[3]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1388 sums[4] = temp[3] + temp[4]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1389 sums[5] = temp[4] + temp[5]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1390 sums[6] = temp[5] + temp[6]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1391 sums[7] = temp[6] + temp[7]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1392 sums[8] = temp[7] + last; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1393 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1394 dst[0]= ((sums[0]<<2) + ((first + sums[2])<<1) + sums[4] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1395 dst[1]= ((dst[1]<<2) + (first + sums[0] + sums[3]<<1) + sums[5] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1396 dst[2]= ((dst[2]<<2) + (first + sums[1] + sums[4]<<1) + sums[6] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1397 dst[3]= ((dst[3]<<2) + (sums[2] + sums[5]<<1) + sums[0] + sums[7] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1398 dst[4]= ((dst[4]<<2) + (sums[3] + sums[6]<<1) + sums[1] + sums[8] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1399 dst[5]= ((dst[5]<<2) + (last + sums[7] + sums[4]<<1) + sums[2] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1400 dst[6]= ((last + dst[6]<<2) + (dst[7] + sums[5]<<1) + sums[3] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1401 dst[7]= ((sums[8]<<2) + (last + sums[6]<<1) + sums[4] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1402 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1403 dst+= stride; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1404 temp+= TEMP_STRIDE; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1405 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1406 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1407 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1408 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1409 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1410 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
|
1411 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1412 //FIXME |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1413 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1414 #ifdef HAVE_MMX2X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1415 asm volatile( |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1416 "leal (%0, %1), %%eax \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1417 "leal (%%eax, %1, 4), %%ebx \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1418 // 0 1 2 3 4 5 6 7 8 9 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1419 // %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
|
1420 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1421 "pcmpeq %%mm6, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1422 "pxor %%mm7, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1423 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1424 #define FIND_MIN_MAX(addr)\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1425 "movq (" #addr "), %%mm0, \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1426 "pminub %%mm0, %%mm6 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1427 "pmaxub %%mm0, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1428 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1429 FIND_MIN_MAX(%0) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1430 FIND_MIN_MAX(%%eax) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1431 FIND_MIN_MAX(%%eax, %1) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1432 FIND_MIN_MAX(%%eax, %1, 2) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1433 FIND_MIN_MAX(%0, %1, 4) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1434 FIND_MIN_MAX(%%ebx) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1435 FIND_MIN_MAX(%%ebx, %1) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1436 FIND_MIN_MAX(%%ebx, %1, 2) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1437 FIND_MIN_MAX(%0, %1, 8) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1438 FIND_MIN_MAX(%%ebx, %1, 2) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1439 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1440 "movq %%mm6, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1441 "psrlq $32, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1442 "pminub %%mm4, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1443 "movq %%mm6, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1444 "psrlq $16, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1445 "pminub %%mm4, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1446 "movq %%mm6, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1447 "psrlq $8, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1448 "pminub %%mm4, %%mm6 \n\t" // min of pixels |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1449 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1450 "movq %%mm7, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1451 "psrlq $32, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1452 "pmaxub %%mm4, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1453 "movq %%mm7, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1454 "psrlq $16, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1455 "pmaxub %%mm4, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1456 "movq %%mm7, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1457 "psrlq $8, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1458 "pmaxub %%mm4, %%mm7 \n\t" // max of pixels |
96 | 1459 PAVGB(%%mm6, %%mm7) // (max + min)/2 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1460 |
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 : : "r" (src), "r" (stride), "r" (QP) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1463 : "%eax", "%ebx" |
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 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1466 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1467 //FIXME |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1468 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1469 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1470 |
96 | 1471 |
1472 | |
1473 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1474 /** |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1475 * ... |
96 | 1476 * the mode value is interpreted as a quality value if its negative, its range is then (-1 ... -63) |
1477 * -63 is best quality -1 is worst | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1478 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1479 extern "C"{ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1480 void postprocess(unsigned char * src[], int src_stride, |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1481 unsigned char * dst[], int dst_stride, |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1482 int horizontal_size, int vertical_size, |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1483 QP_STORE_T *QP_store, int QP_stride, |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1484 int mode) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1485 { |
96 | 1486 |
1487 if(mode<0) mode= getModeForQuality(-mode); | |
1488 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1489 /* |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1490 long long T= rdtsc(); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1491 for(int y=vertical_size-1; y>=0 ; y--) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1492 memcpy(dst[0] + y*src_stride, src[0] + y*src_stride,src_stride); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1493 // memcpy(dst[0], src[0],src_stride*vertical_size); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1494 printf("%4dk\r", (rdtsc()-T)/1000); |
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 return; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1497 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1498 /* |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1499 long long T= rdtsc(); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1500 while( (rdtsc() - T)/1000 < 4000); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1501 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1502 return; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1503 */ |
96 | 1504 postProcess(src[0], src_stride, dst[0], dst_stride, |
1505 horizontal_size, vertical_size, QP_store, QP_stride, false, mode); | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1506 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1507 horizontal_size >>= 1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1508 vertical_size >>= 1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1509 src_stride >>= 1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1510 dst_stride >>= 1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1511 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1512 if(1) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1513 { |
96 | 1514 postProcess(src[1], src_stride, dst[1], dst_stride, |
1515 horizontal_size, vertical_size, QP_store, QP_stride, true, mode >>4); | |
1516 postProcess(src[2], src_stride, dst[2], dst_stride, | |
1517 horizontal_size, vertical_size, QP_store, QP_stride, true, mode >>4); | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1518 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1519 else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1520 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1521 memcpy(dst[1], src[1], src_stride*horizontal_size); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1522 memcpy(dst[2], src[2], src_stride*horizontal_size); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1523 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1524 } |
96 | 1525 /** |
1526 * gets the mode flags for a given quality (larger values mean slower but better postprocessing) | |
1527 * 0 <= quality < 64 | |
1528 */ | |
1529 int getModeForQuality(int quality){ | |
1530 int modes[6]= { | |
1531 LUM_V_DEBLOCK, | |
1532 LUM_V_DEBLOCK | LUM_H_DEBLOCK, | |
1533 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK, | |
1534 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK, | |
1535 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING, | |
1536 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING | CHROM_DERING | |
1537 }; | |
1538 | |
1539 return modes[ (quality*6) >>6 ]; | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1540 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1541 |
96 | 1542 } // extern "C" |
1543 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1544 /** |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1545 * Copies a block from src to dst and fixes the blacklevel |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1546 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1547 static inline void blockCopy(uint8_t dst[], int dstStride, uint8_t src[], int srcStride) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1548 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1549 #ifdef HAVE_MMX |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1550 asm volatile( |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1551 "pushl %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1552 "pushl %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1553 "leal (%2,%2), %%eax \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1554 "leal (%3,%3), %%ebx \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1555 "movq packedYOffset, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1556 "movq packedYScale, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1557 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1558 #define SIMPLE_CPY \ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1559 "movq (%0), %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1560 "movq (%0,%2), %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1561 "psubusb %%mm2, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1562 "psubusb %%mm2, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1563 "movq %%mm0, (%1) \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1564 "movq %%mm1, (%1, %3) \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1565 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1566 #define SCALED_CPY \ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1567 "movq (%0), %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1568 "movq (%0,%2), %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1569 "psubusb %%mm2, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1570 "psubusb %%mm2, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1571 "pxor %%mm4, %%mm4 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1572 "pxor %%mm5, %%mm5 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1573 "punpcklbw %%mm0, %%mm4 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1574 "punpckhbw %%mm0, %%mm5 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1575 "pmulhuw %%mm3, %%mm4 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1576 "pmulhuw %%mm3, %%mm5 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1577 "packuswb %%mm5, %%mm4 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1578 "movq %%mm4, (%1) \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1579 "pxor %%mm4, %%mm4 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1580 "pxor %%mm5, %%mm5 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1581 "punpcklbw %%mm1, %%mm4 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1582 "punpckhbw %%mm1, %%mm5 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1583 "pmulhuw %%mm3, %%mm4 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1584 "pmulhuw %%mm3, %%mm5 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1585 "packuswb %%mm5, %%mm4 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1586 "movq %%mm4, (%1, %3) \n\t"\ |
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 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1589 #define CPY SCALED_CPY |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1590 //#define CPY SIMPLE_CPY |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1591 // "prefetchnta 8(%0)\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1592 CPY |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1593 "addl %%eax, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1594 "addl %%ebx, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1595 CPY |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1596 "addl %%eax, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1597 "addl %%ebx, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1598 CPY |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1599 "addl %%eax, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1600 "addl %%ebx, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1601 CPY |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1602 "popl %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1603 "popl %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1604 : : "r" (src), |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1605 "r" (dst), |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1606 "r" (srcStride), |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1607 "r" (dstStride) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1608 : "%eax", "%ebx" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1609 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1610 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1611 for(int i=0; i<BLOCK_SIZE; i++) // last 10x8 Block is copied allready so +2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1612 memcpy( &(dst[dstStride*i]), |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1613 &(src[srcStride*i]), BLOCK_SIZE); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1614 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1615 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1616 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1617 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1618 /** |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1619 * Filters array of bytes (Y or U or V values) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1620 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1621 void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height, |
96 | 1622 QP_STORE_T QPs[], int QPStride, bool isColor, int mode) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1623 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1624 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1625 #ifdef TIMEING |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1626 long long T0, T1, memcpyTime=0, vertTime=0, horizTime=0, sumTime, diffTime=0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1627 sumTime= rdtsc(); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1628 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1629 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1630 /* we need 64bit here otherwise we´ll going to have a problem |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1631 after watching a black picture for 5 hours*/ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1632 static uint64_t *yHistogram= NULL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1633 if(!yHistogram) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1634 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1635 yHistogram= new uint64_t[256]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1636 for(int i=0; i<256; i++) yHistogram[i]= width*height/64/256; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1637 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1638 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1639 int black=0, white=255; // blackest black and whitest white in the picture |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1640 if(!isColor) |
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 uint64_t sum= 0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1643 for(int i=0; i<256; i++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1644 sum+= yHistogram[i]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1645 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1646 uint64_t maxClipped= (uint64_t)(sum * maxClippedThreshold); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1647 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1648 uint64_t clipped= sum; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1649 for(black=255; black>0; black--) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1650 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1651 if(clipped < maxClipped) break; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1652 clipped-= yHistogram[black]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1653 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1654 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1655 clipped= sum; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1656 for(white=0; white<256; white++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1657 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1658 if(clipped < maxClipped) break; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1659 clipped-= yHistogram[white]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1660 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1661 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1662 // we cant handle negative correctures |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1663 packedYOffset= MAX(black - minAllowedY, 0); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1664 packedYOffset|= packedYOffset<<32; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1665 packedYOffset|= packedYOffset<<16; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1666 packedYOffset|= packedYOffset<<8; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1667 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1668 double scale= (double)(maxAllowedY - minAllowedY) / (double)(white-black); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1669 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1670 packedYScale= uint16_t(scale*256.0 + 0.5); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1671 packedYScale|= packedYScale<<32; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1672 packedYScale|= packedYScale<<16; |
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 else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1675 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1676 packedYScale= 0x0100010001000100LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1677 packedYOffset= 0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1678 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1679 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1680 for(int x=0; x<width; x+=BLOCK_SIZE) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1681 blockCopy(dst + x, dstStride, src + x, srcStride); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1682 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1683 for(int y=0; y<height; y+=BLOCK_SIZE) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1684 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1685 //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
|
1686 uint8_t *srcBlock= &(src[y*srcStride]); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1687 uint8_t *dstBlock= &(dst[y*dstStride]); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1688 uint8_t *vertSrcBlock= &(srcBlock[srcStride*3]); // Blocks are 10x8 -> *3 to start |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1689 uint8_t *vertBlock= &(dstBlock[dstStride*3]); |
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 // 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
|
1692 // with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1693 for(int x=0; x<width; x+=BLOCK_SIZE) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1694 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1695 int QP= isColor ? |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1696 QPs[(y>>3)*QPStride + (x>>3)]: |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1697 (QPs[(y>>4)*QPStride + (x>>4)] * (packedYScale &0xFFFF))>>8; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1698 #ifdef HAVE_MMX |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1699 asm volatile( |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1700 "movd %0, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1701 "packuswb %%mm7, %%mm7 \n\t" // 0, 0, 0, QP, 0, 0, 0, QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1702 "packuswb %%mm7, %%mm7 \n\t" // 0,QP, 0, QP, 0,QP, 0, QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1703 "packuswb %%mm7, %%mm7 \n\t" // QP,..., QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1704 "movq %%mm7, pQPb \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1705 : : "r" (QP) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1706 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1707 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1708 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1709 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1710 const int stride= dstStride; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1711 if(y + 12 < height) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1712 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1713 #ifdef MORE_TIMEING |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1714 T0= rdtsc(); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1715 #endif |
96 | 1716 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1717 #ifdef HAVE_MMX2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1718 prefetchnta(vertSrcBlock + (((x>>3)&3) + 2)*srcStride + 32); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1719 prefetchnta(vertSrcBlock + (((x>>3)&3) + 6)*srcStride + 32); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1720 prefetcht0(vertBlock + (((x>>3)&3) + 2)*dstStride + 32); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1721 prefetcht0(vertBlock + (((x>>3)&3) + 6)*dstStride + 32); |
96 | 1722 #elif defined(HAVE_3DNOW) |
1723 //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ... | |
1724 /* prefetch(vertSrcBlock + (((x>>3)&3) + 2)*srcStride + 32); | |
1725 prefetch(vertSrcBlock + (((x>>3)&3) + 6)*srcStride + 32); | |
1726 prefetchw(vertBlock + (((x>>3)&3) + 2)*dstStride + 32); | |
1727 prefetchw(vertBlock + (((x>>3)&3) + 6)*dstStride + 32); | |
1728 */ | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1729 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1730 if(!isColor) yHistogram[ srcBlock[0] ]++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1731 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1732 blockCopy(vertBlock + dstStride*2, dstStride, |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1733 vertSrcBlock + srcStride*2, srcStride); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1734 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1735 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1736 #ifdef MORE_TIMEING |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1737 T1= rdtsc(); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1738 memcpyTime+= T1-T0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1739 T0=T1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1740 #endif |
96 | 1741 if(mode & V_DEBLOCK) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1742 { |
96 | 1743 if(mode & RK_FILTER) |
1744 vertRKFilter(vertBlock, stride, QP); | |
1745 else if(0) | |
1746 vertX1Filter(vertBlock, stride, QP); | |
1747 else | |
1748 { | |
1749 if( isVertDC(vertBlock, stride)) | |
1750 { | |
1751 if(isVertMinMaxOk(vertBlock, stride, QP)) | |
1752 doVertLowPass(vertBlock, stride, QP); | |
1753 } | |
1754 else | |
1755 doVertDefFilter(vertBlock, stride, QP); | |
1756 } | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1757 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1758 #ifdef MORE_TIMEING |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1759 T1= rdtsc(); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1760 vertTime+= T1-T0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1761 T0=T1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1762 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1763 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1764 else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1765 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1766 for(int i=2; i<BLOCK_SIZE/2+1; i++) // last 10x8 Block is copied allready so +2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1767 memcpy( &(vertBlock[dstStride*i]), |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1768 &(vertSrcBlock[srcStride*i]), BLOCK_SIZE); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1769 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1770 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1771 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1772 if(x - 8 >= 0 && x<width) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1773 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1774 #ifdef MORE_TIMEING |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1775 T0= rdtsc(); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1776 #endif |
96 | 1777 if(mode & H_DEBLOCK) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1778 { |
96 | 1779 if( isHorizDCAndCopy2Temp(dstBlock-4, stride)) |
1780 { | |
1781 if(isHorizMinMaxOk(tempBlock, TEMP_STRIDE, QP)) | |
1782 doHorizLowPassAndCopyBack(dstBlock-4, stride, QP); | |
1783 } | |
1784 else | |
1785 doHorizDefFilterAndCopyBack(dstBlock-4, stride, QP); | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1786 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1787 #ifdef MORE_TIMEING |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1788 T1= rdtsc(); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1789 horizTime+= T1-T0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1790 T0=T1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1791 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1792 dering(dstBlock - 9 - stride, stride, QP); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1793 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1794 else if(y!=0) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1795 dering(dstBlock - stride*9 + width-9, stride, QP); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1796 //FIXME dering filter will not be applied to last block (bottom right) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1797 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1798 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1799 dstBlock+=8; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1800 srcBlock+=8; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1801 vertBlock+=8; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1802 vertSrcBlock+=8; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1803 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1804 } |
96 | 1805 #ifdef HAVE_3DNOW |
1806 asm volatile("femms"); | |
1807 #elif defined (HAVE_MMX) | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1808 asm volatile("emms"); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1809 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1810 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1811 #ifdef TIMEING |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1812 // 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
|
1813 sumTime= rdtsc() - sumTime; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1814 if(!isColor) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1815 printf("cpy:%4dk, vert:%4dk, horiz:%4dk, sum:%4dk, diff:%4dk, color: %d/%d \r", |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1816 int(memcpyTime/1000), int(vertTime/1000), int(horizTime/1000), |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1817 int(sumTime/1000), int((sumTime-memcpyTime-vertTime-horizTime)/1000) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1818 , black, white); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1819 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1820 } |
96 | 1821 |
1822 |