Mercurial > mplayer.hg
annotate postproc/postprocess.c @ 2181:d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
author | michael |
---|---|
date | Sat, 13 Oct 2001 11:58:41 +0000 |
parents | c76bfc70a28c |
children | c5162871cf2b |
rev | line source |
---|---|
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2 Copyright (C) 2001 Michael Niedermayer (michaelni@gmx.at) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3 |
508468a75be0
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 |
508468a75be0
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 |
508468a75be0
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 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
7 (at your option) any later version. |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
8 |
508468a75be0
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, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
10 but WITHOUT ANY WARRANTY; without even the implied warranty of |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
12 GNU General Public License for more details. |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
13 |
508468a75be0
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 |
508468a75be0
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 |
508468a75be0
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 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
17 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
18 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
19 /* |
2159 | 20 C MMX MMX2 3DNow* |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
21 isVertDC Ec Ec |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
22 isVertMinMaxOk Ec Ec |
2159 | 23 doVertLowPass E e e* |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
24 doVertDefFilter Ec Ec Ec |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
25 isHorizDC Ec Ec |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
26 isHorizMinMaxOk a |
2159 | 27 doHorizLowPass E a a* |
28 doHorizDefFilter E ac ac | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
29 deRing |
2179 | 30 Vertical RKAlgo1 E a a* |
31 Vertical X1 a E E* | |
32 Horizontal X1 a E E* | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
33 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
34 |
2159 | 35 * i dont have a 3dnow CPU -> its untested |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
36 E = Exact implementation |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
37 e = allmost exact implementation |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
38 a = alternative / approximate impl |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
39 c = checked against the other implementations (-vo md5) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
40 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
41 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
42 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
43 TODO: |
2179 | 44 verify that everything workes as it should (how?) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
45 reduce the time wasted on the mem transfer |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
46 implement dering |
2159 | 47 implement everything in C at least (done at the moment but ...) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
48 unroll stuff if instructions depend too much on the prior one |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
49 we use 8x8 blocks for the horizontal filters, opendivx seems to use 8x4? |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
50 move YScale thing to the end instead of fixing QP |
2159 | 51 write a faster and higher quality deblocking filter :) |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
52 do something about the speed of the horizontal filters |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
53 make the mainloop more flexible (variable number of blocks at once |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
54 (the if/else stuff per block is slowing things down) |
2179 | 55 compare the quality & speed of all filters |
56 implement a few simple deinterlacing filters | |
57 split this huge file | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
58 ... |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
59 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
60 Notes: |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
61 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
62 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
63 |
2159 | 64 /* |
2179 | 65 Changelog: use the CVS log |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
66 0.1.3 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
67 bugfixes: last 3 lines not brightness/contrast corrected |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
68 brightness statistics messed up with initial black pic |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
69 changed initial values of the brightness statistics |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
70 C++ -> C conversation |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
71 QP range question solved (very likely 1<=QP<=32 according to arpi) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
72 new experimental vertical deblocking filter |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
73 RK filter has 3dNow support now (untested) |
2159 | 74 0.1.2 |
75 fixed a bug in the horizontal default filter | |
76 3dnow version of the Horizontal & Vertical Lowpass filters | |
77 mmx version of the Horizontal Default filter | |
78 mmx2 & C versions of a simple filter described in a paper from ramkishor & karandikar | |
79 added mode flags & quality2mode function | |
80 0.1.1 | |
81 */ | |
82 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
83 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
84 #include <inttypes.h> |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
85 #include <stdio.h> |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
86 #include <stdlib.h> |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
87 #include "../config.h" |
2159 | 88 //#undef HAVE_MMX2 |
89 //#define HAVE_3DNOW | |
90 //#undef HAVE_MMX | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
91 #include "postprocess.h" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
92 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
93 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
94 static uint64_t packedYOffset= 0x0000000000000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
95 static uint64_t packedYScale= 0x0100010001000100LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
96 static uint64_t w05= 0x0005000500050005LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
97 static uint64_t w20= 0x0020002000200020LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
98 static uint64_t w1400= 0x1400140014001400LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
99 static uint64_t bm00000001= 0x00000000000000FFLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
100 static uint64_t bm00010000= 0x000000FF00000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
101 static uint64_t bm00001000= 0x00000000FF000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
102 static uint64_t bm10000000= 0xFF00000000000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
103 static uint64_t bm10000001= 0xFF000000000000FFLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
104 static uint64_t bm11000011= 0xFFFF00000000FFFFLL; |
2159 | 105 static uint64_t bm00000011= 0x000000000000FFFFLL; |
2179 | 106 static uint64_t bm11111110= 0xFFFFFFFFFFFFFF00LL; |
2159 | 107 static uint64_t bm11000000= 0xFFFF000000000000LL; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
108 static uint64_t bm00011000= 0x000000FFFF000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
109 static uint64_t bm00110011= 0x0000FFFF0000FFFFLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
110 static uint64_t bm11001100= 0xFFFF0000FFFF0000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
111 static uint64_t b00= 0x0000000000000000LL; |
2179 | 112 static uint64_t b01= 0x0101010101010101LL; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
113 static uint64_t b02= 0x0202020202020202LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
114 static uint64_t b0F= 0x0F0F0F0F0F0F0F0FLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
115 static uint64_t bFF= 0xFFFFFFFFFFFFFFFFLL; |
2159 | 116 static uint64_t b20= 0x2020202020202020LL; |
117 static uint64_t b80= 0x8080808080808080LL; | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
118 static uint64_t b7E= 0x7E7E7E7E7E7E7E7ELL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
119 static uint64_t b7C= 0x7C7C7C7C7C7C7C7CLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
120 static uint64_t b3F= 0x3F3F3F3F3F3F3F3FLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
121 static uint64_t temp0=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
122 static uint64_t temp1=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
123 static uint64_t temp2=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
124 static uint64_t temp3=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
125 static uint64_t temp4=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
126 static uint64_t temp5=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
127 static uint64_t pQPb=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
128 static uint8_t tempBlock[16*16]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
129 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
130 int hFlatnessThreshold= 56 - 16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
131 int vFlatnessThreshold= 56 - 16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
132 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
133 //amount of "black" u r willing to loose to get a brightness corrected picture |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
134 double maxClippedThreshold= 0.01; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
135 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
136 int maxAllowedY=255; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
137 //FIXME can never make a movie´s black brighter (anyone needs that?) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
138 int minAllowedY=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
139 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
140 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
141 static inline long long rdtsc() |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
142 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
143 long long l; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
144 asm volatile( "rdtsc\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
145 : "=A" (l) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
146 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
147 // printf("%d\n", int(l/1000)); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
148 return l; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
149 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
150 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
151 static inline void prefetchnta(void *p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
152 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
153 asm volatile( "prefetchnta (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
154 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
155 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
156 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
157 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
158 static inline void prefetcht0(void *p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
159 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
160 asm volatile( "prefetcht0 (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
161 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
162 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
163 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
164 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
165 static inline void prefetcht1(void *p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
166 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
167 asm volatile( "prefetcht1 (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
168 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
169 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
170 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
171 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
172 static inline void prefetcht2(void *p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
173 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
174 asm volatile( "prefetcht2 (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
175 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
176 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
177 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
178 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
179 //FIXME? |255-0| = 1 (shouldnt be a problem ...) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
180 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
181 * Check if the middle 8x8 Block in the given 8x10 block is flat |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
182 */ |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
183 static inline int isVertDC(uint8_t src[], int stride){ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
184 // return true; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
185 int numEq= 0; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
186 int y; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
187 src+= stride; // src points to begin of the 8x8 Block |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
188 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
189 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
190 // "int $3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
191 "pushl %1\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
192 "movq b7E, %%mm7 \n\t" // mm7 = 0x7F |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
193 "movq b7C, %%mm6 \n\t" // mm6 = 0x7D |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
194 "movq (%1), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
195 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
196 "movq (%1), %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
197 "psubb %%mm1, %%mm0 \n\t" // mm0 = differnece |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
198 "paddb %%mm7, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
199 "pcmpgtb %%mm6, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
200 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
201 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
202 "movq (%1), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
203 "psubb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
204 "paddb %%mm7, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
205 "pcmpgtb %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
206 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
207 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
208 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
209 "movq (%1), %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
210 "psubb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
211 "paddb %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
212 "pcmpgtb %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
213 "paddb %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
214 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
215 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
216 "movq (%1), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
217 "psubb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
218 "paddb %%mm7, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
219 "pcmpgtb %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
220 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
221 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
222 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
223 "movq (%1), %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
224 "psubb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
225 "paddb %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
226 "pcmpgtb %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
227 "paddb %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
228 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
229 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
230 "movq (%1), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
231 "psubb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
232 "paddb %%mm7, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
233 "pcmpgtb %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
234 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
235 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
236 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
237 "movq (%1), %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
238 "psubb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
239 "paddb %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
240 "pcmpgtb %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
241 "paddb %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
242 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
243 " \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
244 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
245 "psrlw $8, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
246 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
247 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
248 "psrlq $16, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
249 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
250 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
251 "psrlq $32, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
252 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
253 "popl %1\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
254 "movd %%mm0, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
255 : "=r" (numEq) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
256 : "r" (src), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
257 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
258 // printf("%d\n", numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
259 numEq= (256 - (numEq & 0xFF)) &0xFF; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
260 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
261 // int asmEq= numEq; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
262 // numEq=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
263 // uint8_t *temp= src; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
264 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
265 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
266 for(y=0; y<BLOCK_SIZE-1; y++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
267 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
268 if(((src[0] - src[0+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
269 if(((src[1] - src[1+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
270 if(((src[2] - src[2+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
271 if(((src[3] - src[3+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
272 if(((src[4] - src[4+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
273 if(((src[5] - src[5+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
274 if(((src[6] - src[6+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
275 if(((src[7] - src[7+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
276 src+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
277 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
278 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
279 /* if(abs(numEq - asmEq) > 0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
280 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
281 printf("\nasm:%d c:%d\n", asmEq, numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
282 for(int y=0; y<8; y++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
283 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
284 for(int x=0; x<8; x++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
285 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
286 printf("%d ", temp[x + y*stride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
287 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
288 printf("\n"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
289 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
290 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
291 */ |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
292 // for(int i=0; i<numEq/8; i++) src[i]=255; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
293 return (numEq > vFlatnessThreshold) ? 1 : 0; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
294 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
295 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
296 static inline int isVertMinMaxOk(uint8_t src[], int stride, int QP) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
297 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
298 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
299 int isOk; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
300 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
301 // "int $3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
302 "movq (%1, %2), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
303 "movq (%1, %2, 8), %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
304 "movq %%mm0, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
305 "psubusb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
306 "psubusb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
307 "por %%mm1, %%mm0 \n\t" // ABS Diff |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
308 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
309 "movq pQPb, %%mm7 \n\t" // QP,..., QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
310 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
311 "psubusb %%mm7, %%mm0 \n\t" // Diff <= 2QP -> 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
312 "pcmpeqd b00, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
313 "psrlq $16, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
314 "pcmpeqd bFF, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
315 // "movd %%mm0, (%1, %2, 4)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
316 "movd %%mm0, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
317 : "=r" (isOk) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
318 : "r" (src), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
319 ); |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
320 return isOk ? 1 : 0; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
321 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
322 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
323 int isOk2= 1; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
324 int x; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
325 for(x=0; x<BLOCK_SIZE; x++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
326 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
327 if(abs((int)src[x + stride] - (int)src[x + (stride<<3)]) > 2*QP) isOk2=0; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
328 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
329 /* if(isOk && !isOk2 || !isOk && isOk2) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
330 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
331 printf("\nasm:%d c:%d QP:%d\n", isOk, isOk2, QP); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
332 for(int y=0; y<9; y++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
333 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
334 for(int x=0; x<8; x++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
335 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
336 printf("%d ", src[x + y*stride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
337 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
338 printf("\n"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
339 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
340 } */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
341 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
342 return isOk2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
343 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
344 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
345 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
346 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
347 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
348 * Do a vertical low pass filter on the 8x10 block (only write to the 8x8 block in the middle) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
349 * useing the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
350 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
351 static inline void doVertLowPass(uint8_t *src, int stride, int QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
352 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
353 // QP= 64; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
354 |
2159 | 355 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
356 //#ifdef HAVE_MMX2 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
357 asm volatile( //"movv %0 %1 %2\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
358 "pushl %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
359 "movq pQPb, %%mm0 \n\t" // QP,..., QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
360 // "movq bFF , %%mm0 \n\t" // QP,..., QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
361 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
362 "movq (%0), %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
363 "movq (%0, %1), %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
364 "movq %%mm5, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
365 "movq %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
366 "psubusb %%mm6, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
367 "psubusb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
368 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
369 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
370 "pcmpeqb b00, %%mm2 \n\t" // diff <= QP -> FF |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
371 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
372 "pand %%mm2, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
373 "pandn %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
374 "por %%mm2, %%mm6 \n\t"// First Line to Filter |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
375 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
376 "movq (%0, %1, 8), %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
377 "leal (%0, %1, 4), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
378 "leal (%0, %1, 8), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
379 "subl %1, %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
380 "addl %1, %0 \n\t" // %0 points to line 1 not 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
381 "movq (%0, %1, 8), %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
382 "movq %%mm5, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
383 "movq %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
384 "psubusb %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
385 "psubusb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
386 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
387 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
388 "pcmpeqb b00, %%mm2 \n\t" // diff <= QP -> FF |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
389 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
390 "pand %%mm2, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
391 "pandn %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
392 "por %%mm2, %%mm7 \n\t" // First Line to Filter |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
393 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
394 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
395 // 1 2 3 4 5 6 7 8 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
396 // %0 %0+%1 %0+2%1 eax %0+4%1 eax+2%1 ebx eax+4%1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
397 // 6 4 2 2 1 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
398 // 6 4 4 2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
399 // 6 8 2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
400 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
401 "movq %%mm6, %%mm2 \n\t" //1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
402 "movq %%mm6, %%mm3 \n\t" //1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
403 "paddusb b02, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
404 "psrlw $2, %%mm3 \n\t" //1 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
405 "pand b3F, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
406 "psubb %%mm3, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
407 "movq (%0, %1), %%mm0 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
408 "movq %%mm0, %%mm1 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
409 "paddusb b02, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
410 "psrlw $2, %%mm0 \n\t" // 1 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
411 "pand b3F, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
412 "paddusb %%mm2, %%mm0 \n\t" //3 1 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
413 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
414 "movq (%0, %1), %%mm0 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
415 "movq %%mm0, %%mm1 \n\t" // 1 |
2159 | 416 PAVGB(%%mm6, %%mm0) //1 1 /2 |
417 PAVGB(%%mm6, %%mm0) //3 1 /4 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
418 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
419 "movq (%0, %1, 4), %%mm2 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
420 "movq %%mm2, %%mm5 \n\t" // 1 |
2159 | 421 PAVGB((%%eax), %%mm2) // 11 /2 |
422 PAVGB((%0, %1, 2), %%mm2) // 211 /4 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
423 "movq %%mm2, %%mm3 \n\t" // 211 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
424 "movq (%0), %%mm4 \n\t" // 1 |
2159 | 425 PAVGB(%%mm4, %%mm3) // 4 211 /8 |
426 PAVGB(%%mm0, %%mm3) //642211 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
427 "movq %%mm3, (%0) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
428 // mm1=2 mm2=3(211) mm4=1 mm5=5 mm6=0 mm7=9 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
429 "movq %%mm1, %%mm0 \n\t" // 1 |
2159 | 430 PAVGB(%%mm6, %%mm0) //1 1 /2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
431 "movq %%mm4, %%mm3 \n\t" // 1 |
2159 | 432 PAVGB((%0,%1,2), %%mm3) // 1 1 /2 |
433 PAVGB((%%eax,%1,2), %%mm5) // 11 /2 | |
434 PAVGB((%%eax), %%mm5) // 211 /4 | |
435 PAVGB(%%mm5, %%mm3) // 2 2211 /8 | |
436 PAVGB(%%mm0, %%mm3) //4242211 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
437 "movq %%mm3, (%0,%1) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
438 // mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9 |
2159 | 439 PAVGB(%%mm4, %%mm6) //11 /2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
440 "movq (%%ebx), %%mm0 \n\t" // 1 |
2159 | 441 PAVGB((%%eax, %1, 2), %%mm0) // 11/2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
442 "movq %%mm0, %%mm3 \n\t" // 11/2 |
2159 | 443 PAVGB(%%mm1, %%mm0) // 2 11/4 |
444 PAVGB(%%mm6, %%mm0) //222 11/8 | |
445 PAVGB(%%mm2, %%mm0) //22242211/16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
446 "movq (%0, %1, 2), %%mm2 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
447 "movq %%mm0, (%0, %1, 2) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
448 // mm1=2 mm2=3 mm3=6(11) mm4=1 mm5=4(211) mm6=0(11) mm7=9 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
449 "movq (%%eax, %1, 4), %%mm0 \n\t" // 1 |
2159 | 450 PAVGB((%%ebx), %%mm0) // 11 /2 |
451 PAVGB(%%mm0, %%mm6) //11 11 /4 | |
452 PAVGB(%%mm1, %%mm4) // 11 /2 | |
453 PAVGB(%%mm2, %%mm1) // 11 /2 | |
454 PAVGB(%%mm1, %%mm6) //1122 11 /8 | |
455 PAVGB(%%mm5, %%mm6) //112242211 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
456 "movq (%%eax), %%mm5 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
457 "movq %%mm6, (%%eax) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
458 // mm0=7(11) mm1=2(11) mm2=3 mm3=6(11) mm4=1(11) mm5=4 mm7=9 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
459 "movq (%%eax, %1, 4), %%mm6 \n\t" // 1 |
2159 | 460 PAVGB(%%mm7, %%mm6) // 11 /2 |
461 PAVGB(%%mm4, %%mm6) // 11 11 /4 | |
462 PAVGB(%%mm3, %%mm6) // 11 2211 /8 | |
463 PAVGB(%%mm5, %%mm2) // 11 /2 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
464 "movq (%0, %1, 4), %%mm4 \n\t" // 1 |
2159 | 465 PAVGB(%%mm4, %%mm2) // 112 /4 |
466 PAVGB(%%mm2, %%mm6) // 112242211 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
467 "movq %%mm6, (%0, %1, 4) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
468 // mm0=7(11) mm1=2(11) mm2=3(112) mm3=6(11) mm4=5 mm5=4 mm7=9 |
2159 | 469 PAVGB(%%mm7, %%mm1) // 11 2 /4 |
470 PAVGB(%%mm4, %%mm5) // 11 /2 | |
471 PAVGB(%%mm5, %%mm0) // 11 11 /4 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
472 "movq (%%eax, %1, 2), %%mm6 \n\t" // 1 |
2159 | 473 PAVGB(%%mm6, %%mm1) // 11 4 2 /8 |
474 PAVGB(%%mm0, %%mm1) // 11224222 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
475 // "pxor %%mm1, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
476 "movq %%mm1, (%%eax, %1, 2) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
477 // mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9 |
2159 | 478 PAVGB((%%ebx), %%mm2) // 112 4 /8 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
479 "movq (%%eax, %1, 4), %%mm0 \n\t" // 1 |
2159 | 480 PAVGB(%%mm0, %%mm6) // 1 1 /2 |
481 PAVGB(%%mm7, %%mm6) // 1 12 /4 | |
482 PAVGB(%%mm2, %%mm6) // 1122424 /4 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
483 // "pxor %%mm6, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
484 "movq %%mm6, (%%ebx) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
485 // mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9 |
2159 | 486 PAVGB(%%mm7, %%mm5) // 11 2 /4 |
487 PAVGB(%%mm7, %%mm5) // 11 6 /8 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
488 |
2159 | 489 PAVGB(%%mm3, %%mm0) // 112 /4 |
490 PAVGB(%%mm0, %%mm5) // 112246 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
491 // "pxor %%mm5, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
492 // "movq pQPb, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
493 "movq %%mm5, (%%eax, %1, 4) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
494 "popl %0\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
495 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
496 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
497 : "r" (src), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
498 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
499 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
500 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
501 const int l1= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
502 const int l2= stride + l1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
503 const int l3= stride + l2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
504 const int l4= stride + l3; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
505 const int l5= stride + l4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
506 const int l6= stride + l5; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
507 const int l7= stride + l6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
508 const int l8= stride + l7; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
509 const int l9= stride + l8; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
510 int x; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
511 for(x=0; x<BLOCK_SIZE; x++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
512 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
513 const int first= ABS(src[0] - src[l1]) < QP ? src[0] : src[l1]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
514 const int last= ABS(src[l8] - src[l9]) < QP ? src[l9] : src[l8]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
515 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
516 int sums[9]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
517 sums[0] = first + src[l1]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
518 sums[1] = src[l1] + src[l2]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
519 sums[2] = src[l2] + src[l3]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
520 sums[3] = src[l3] + src[l4]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
521 sums[4] = src[l4] + src[l5]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
522 sums[5] = src[l5] + src[l6]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
523 sums[6] = src[l6] + src[l7]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
524 sums[7] = src[l7] + src[l8]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
525 sums[8] = src[l8] + last; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
526 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
527 src[l1]= ((sums[0]<<2) + ((first + sums[2])<<1) + sums[4] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
528 src[l2]= ((src[l2]<<2) + (first + sums[0] + sums[3]<<1) + sums[5] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
529 src[l3]= ((src[l3]<<2) + (first + sums[1] + sums[4]<<1) + sums[6] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
530 src[l4]= ((src[l4]<<2) + (sums[2] + sums[5]<<1) + sums[0] + sums[7] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
531 src[l5]= ((src[l5]<<2) + (sums[3] + sums[6]<<1) + sums[1] + sums[8] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
532 src[l6]= ((src[l6]<<2) + (last + sums[7] + sums[4]<<1) + sums[2] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
533 src[l7]= ((last + src[l7]<<2) + (src[l8] + sums[5]<<1) + sums[3] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
534 src[l8]= ((sums[8]<<2) + (last + sums[6]<<1) + sums[4] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
535 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
536 src++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
537 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
538 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
539 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
540 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
541 |
2159 | 542 /** |
543 * Experimental implementation of the filter (Algorithm 1) described in a paper from Ramkishor & Karandikar | |
544 * values are correctly clipped (MMX2) | |
545 * values are wraparound (C) | |
546 * conclusion: its fast, but introduces ugly horizontal patterns if there is a continious gradient | |
547 0 8 16 24 | |
548 x = 8 | |
549 x/2 = 4 | |
550 x/8 = 1 | |
551 1 12 12 23 | |
552 */ | |
2179 | 553 static inline void vertRK1Filter(uint8_t *src, int stride, int QP) |
2159 | 554 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
555 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2159 | 556 // FIXME rounding |
557 asm volatile( | |
558 "pxor %%mm7, %%mm7 \n\t" // 0 | |
559 "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE | |
560 "leal (%0, %1), %%eax \n\t" | |
561 "leal (%%eax, %1, 4), %%ebx \n\t" | |
562 // 0 1 2 3 4 5 6 7 8 9 | |
563 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 | |
564 "movq pQPb, %%mm0 \n\t" // QP,..., QP | |
565 "movq %%mm0, %%mm1 \n\t" // QP,..., QP | |
566 "paddusb b02, %%mm0 \n\t" | |
567 "psrlw $2, %%mm0 \n\t" | |
568 "pand b3F, %%mm0 \n\t" // QP/4,..., QP/4 | |
569 "paddusb %%mm1, %%mm0 \n\t" // QP*1.25 ... | |
570 "movq (%0, %1, 4), %%mm2 \n\t" // line 4 | |
571 "movq (%%ebx), %%mm3 \n\t" // line 5 | |
572 "movq %%mm2, %%mm4 \n\t" // line 4 | |
573 "pcmpeqb %%mm5, %%mm5 \n\t" // -1 | |
574 "pxor %%mm2, %%mm5 \n\t" // -line 4 - 1 | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
575 PAVGB(%%mm3, %%mm5) |
2159 | 576 "paddb %%mm6, %%mm5 \n\t" // (l5-l4)/2 |
577 "psubusb %%mm3, %%mm4 \n\t" | |
578 "psubusb %%mm2, %%mm3 \n\t" | |
579 "por %%mm3, %%mm4 \n\t" // |l4 - l5| | |
580 "psubusb %%mm0, %%mm4 \n\t" | |
581 "pcmpeqb %%mm7, %%mm4 \n\t" | |
582 "pand %%mm4, %%mm5 \n\t" // d/2 | |
583 | |
584 // "paddb %%mm6, %%mm2 \n\t" // line 4 + 0x80 | |
585 "paddb %%mm5, %%mm2 \n\t" | |
586 // "psubb %%mm6, %%mm2 \n\t" | |
587 "movq %%mm2, (%0,%1, 4) \n\t" | |
588 | |
589 "movq (%%ebx), %%mm2 \n\t" | |
590 // "paddb %%mm6, %%mm2 \n\t" // line 5 + 0x80 | |
591 "psubb %%mm5, %%mm2 \n\t" | |
592 // "psubb %%mm6, %%mm2 \n\t" | |
593 "movq %%mm2, (%%ebx) \n\t" | |
594 | |
595 "paddb %%mm6, %%mm5 \n\t" | |
596 "psrlw $2, %%mm5 \n\t" | |
597 "pand b3F, %%mm5 \n\t" | |
598 "psubb b20, %%mm5 \n\t" // (l5-l4)/8 | |
599 | |
600 "movq (%%eax, %1, 2), %%mm2 \n\t" | |
601 "paddb %%mm6, %%mm2 \n\t" // line 3 + 0x80 | |
602 "paddsb %%mm5, %%mm2 \n\t" | |
603 "psubb %%mm6, %%mm2 \n\t" | |
604 "movq %%mm2, (%%eax, %1, 2) \n\t" | |
605 | |
606 "movq (%%ebx, %1), %%mm2 \n\t" | |
607 "paddb %%mm6, %%mm2 \n\t" // line 6 + 0x80 | |
608 "psubsb %%mm5, %%mm2 \n\t" | |
609 "psubb %%mm6, %%mm2 \n\t" | |
610 "movq %%mm2, (%%ebx, %1) \n\t" | |
611 | |
612 : | |
613 : "r" (src), "r" (stride) | |
614 : "%eax", "%ebx" | |
615 ); | |
616 #else | |
617 const int l1= stride; | |
618 const int l2= stride + l1; | |
619 const int l3= stride + l2; | |
620 const int l4= stride + l3; | |
621 const int l5= stride + l4; | |
622 const int l6= stride + l5; | |
623 const int l7= stride + l6; | |
624 const int l8= stride + l7; | |
625 const int l9= stride + l8; | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
626 int x; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
627 for(x=0; x<BLOCK_SIZE; x++) |
2159 | 628 { |
629 if(ABS(src[l4]-src[l5]) < QP + QP/4) | |
630 { | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
631 int v = (src[l5] - src[l4]); |
2159 | 632 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
633 src[l3] +=v/8; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
634 src[l4] +=v/2; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
635 src[l5] -=v/2; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
636 src[l6] -=v/8; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
637 |
2159 | 638 } |
639 src++; | |
640 } | |
641 | |
642 #endif | |
643 } | |
644 | |
645 /** | |
646 * Experimental Filter 1 | |
2179 | 647 * will not damage linear gradients |
648 * Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
649 * can only smooth blocks at the expected locations (it cant smooth them if they did move) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
650 * MMX2 version does correct clipping C version doesnt |
2159 | 651 */ |
652 static inline void vertX1Filter(uint8_t *src, int stride, int QP) | |
653 { | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
654 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2159 | 655 asm volatile( |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
656 "pxor %%mm7, %%mm7 \n\t" // 0 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
657 // "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
658 "leal (%0, %1), %%eax \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
659 "leal (%%eax, %1, 4), %%ebx \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
660 // 0 1 2 3 4 5 6 7 8 9 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
661 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
662 "movq (%%eax, %1, 2), %%mm0 \n\t" // line 3 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
663 "movq (%0, %1, 4), %%mm1 \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
664 "movq %%mm1, %%mm2 \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
665 "psubusb %%mm0, %%mm1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
666 "psubusb %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
667 "por %%mm1, %%mm0 \n\t" // |l2 - l3| |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
668 "movq (%%ebx), %%mm3 \n\t" // line 5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
669 "movq (%%ebx, %1), %%mm4 \n\t" // line 6 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
670 "movq %%mm3, %%mm5 \n\t" // line 5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
671 "psubusb %%mm4, %%mm3 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
672 "psubusb %%mm5, %%mm4 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
673 "por %%mm4, %%mm3 \n\t" // |l5 - l6| |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
674 PAVGB(%%mm3, %%mm0) // (|l2 - l3| + |l5 - l6|)/2 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
675 "movq %%mm2, %%mm1 \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
676 "psubusb %%mm5, %%mm2 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
677 "movq %%mm2, %%mm4 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
678 "pcmpeqb %%mm7, %%mm2 \n\t" // (l4 - l5) <= 0 ? -1 : 0 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
679 "psubusb %%mm1, %%mm5 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
680 "por %%mm5, %%mm4 \n\t" // |l4 - l5| |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
681 "psubusb %%mm0, %%mm4 \n\t" //d = MAX(0, |l4-l5| - (|l2-l3| + |l5-l6|)/2) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
682 "movq %%mm4, %%mm3 \n\t" // d |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
683 "psubusb pQPb, %%mm4 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
684 "pcmpeqb %%mm7, %%mm4 \n\t" // d <= QP ? -1 : 0 |
2179 | 685 "psubusb b01, %%mm3 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
686 "pand %%mm4, %%mm3 \n\t" // d <= QP ? d : 0 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
687 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
688 PAVGB(%%mm7, %%mm3) // d/2 |
2179 | 689 "movq %%mm3, %%mm1 \n\t" // d/2 |
690 PAVGB(%%mm7, %%mm3) // d/4 | |
691 PAVGB(%%mm1, %%mm3) // 3*d/8 | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
692 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
693 "movq (%0, %1, 4), %%mm0 \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
694 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
695 "psubusb %%mm3, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
696 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
697 "movq %%mm0, (%0, %1, 4) \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
698 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
699 "movq (%%ebx), %%mm0 \n\t" // line 5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
700 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
701 "paddusb %%mm3, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
702 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
703 "movq %%mm0, (%%ebx) \n\t" // line 5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
704 |
2179 | 705 PAVGB(%%mm7, %%mm1) // d/4 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
706 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
707 "movq (%%eax, %1, 2), %%mm0 \n\t" // line 3 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
708 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4 |
2179 | 709 "psubusb %%mm1, %%mm0 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
710 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
711 "movq %%mm0, (%%eax, %1, 2) \n\t" // line 3 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
712 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
713 "movq (%%ebx, %1), %%mm0 \n\t" // line 6 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
714 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5 |
2179 | 715 "paddusb %%mm1, %%mm0 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
716 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
717 "movq %%mm0, (%%ebx, %1) \n\t" // line 6 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
718 |
2179 | 719 PAVGB(%%mm7, %%mm1) // d/8 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
720 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
721 "movq (%%eax, %1), %%mm0 \n\t" // line 2 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
722 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l2-1 : l2 |
2179 | 723 "psubusb %%mm1, %%mm0 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
724 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
725 "movq %%mm0, (%%eax, %1) \n\t" // line 2 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
726 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
727 "movq (%%ebx, %1, 2), %%mm0 \n\t" // line 7 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
728 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l7-1 : l7 |
2179 | 729 "paddusb %%mm1, %%mm0 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
730 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
731 "movq %%mm0, (%%ebx, %1, 2) \n\t" // line 7 |
2159 | 732 |
733 : | |
734 : "r" (src), "r" (stride) | |
735 : "%eax", "%ebx" | |
736 ); | |
737 #else | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
738 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
739 const int l1= stride; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
740 const int l2= stride + l1; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
741 const int l3= stride + l2; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
742 const int l4= stride + l3; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
743 const int l5= stride + l4; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
744 const int l6= stride + l5; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
745 const int l7= stride + l6; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
746 const int l8= stride + l7; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
747 const int l9= stride + l8; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
748 int x; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
749 for(x=0; x<BLOCK_SIZE; x++) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
750 { |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
751 int a= src[l3] - src[l4]; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
752 int b= src[l4] - src[l5]; |
2179 | 753 int c= src[l5] - src[l6]; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
754 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
755 int d= MAX(ABS(b) - (ABS(a) + ABS(c))/2, 0); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
756 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
757 if(d < QP) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
758 { |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
759 int v = d * SIGN(-b); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
760 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
761 src[l2] +=v/8; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
762 src[l3] +=v/4; |
2179 | 763 src[l4] +=3*v/8; |
764 src[l5] -=3*v/8; | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
765 src[l6] -=v/4; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
766 src[l7] -=v/8; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
767 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
768 } |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
769 src++; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
770 } |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
771 /* |
2159 | 772 const int l1= stride; |
773 const int l2= stride + l1; | |
774 const int l3= stride + l2; | |
775 const int l4= stride + l3; | |
776 const int l5= stride + l4; | |
777 const int l6= stride + l5; | |
778 const int l7= stride + l6; | |
779 const int l8= stride + l7; | |
780 const int l9= stride + l8; | |
781 for(int x=0; x<BLOCK_SIZE; x++) | |
782 { | |
783 int v2= src[l2]; | |
784 int v3= src[l3]; | |
785 int v4= src[l4]; | |
786 int v5= src[l5]; | |
787 int v6= src[l6]; | |
788 int v7= src[l7]; | |
789 | |
790 if(ABS(v4-v5)<QP && ABS(v4-v5) - (ABS(v3-v4) + ABS(v5-v6))>0 ) | |
791 { | |
792 src[l3] = (6*v2 + 4*v3 + 3*v4 + 2*v5 + v6 )/16; | |
793 src[l4] = (3*v2 + 3*v3 + 4*v4 + 3*v5 + 2*v6 + v7 )/16; | |
794 src[l5] = (1*v2 + 2*v3 + 3*v4 + 4*v5 + 3*v6 + 3*v7)/16; | |
795 src[l6] = ( 1*v3 + 2*v4 + 3*v5 + 4*v6 + 6*v7)/16; | |
796 } | |
797 src++; | |
798 } | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
799 */ |
2159 | 800 #endif |
801 } | |
802 | |
2179 | 803 /** |
804 * Experimental Filter 1 (Horizontal) | |
805 * will not damage linear gradients | |
806 * Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter | |
807 * can only smooth blocks at the expected locations (it cant smooth them if they did move) | |
808 * MMX2 version does correct clipping C version doesnt | |
809 * not identical with the vertical one | |
810 */ | |
811 static inline void horizX1Filter(uint8_t *src, int stride, int QP) | |
812 { | |
813 int y; | |
814 static uint64_t *lut= NULL; | |
815 if(lut==NULL) | |
816 { | |
817 int i; | |
818 lut= (uint64_t*)memalign(8, 256*8); | |
819 for(i=0; i<256; i++) | |
820 { | |
821 int v= i < 128 ? 2*i : 2*(i-256); | |
822 /* | |
823 //Simulate 112242211 9-Tap filter | |
824 uint64_t a= (v/16) & 0xFF; | |
825 uint64_t b= (v/8) & 0xFF; | |
826 uint64_t c= (v/4) & 0xFF; | |
827 uint64_t d= (3*v/8) & 0xFF; | |
828 */ | |
829 //Simulate piecewise linear interpolation | |
830 uint64_t a= (v/16) & 0xFF; | |
831 uint64_t b= (v*3/16) & 0xFF; | |
832 uint64_t c= (v*5/16) & 0xFF; | |
833 uint64_t d= (7*v/16) & 0xFF; | |
834 uint64_t A= (0x100 - a)&0xFF; | |
835 uint64_t B= (0x100 - b)&0xFF; | |
836 uint64_t C= (0x100 - c)&0xFF; | |
837 uint64_t D= (0x100 - c)&0xFF; | |
838 | |
839 lut[i] = (a<<56) | (b<<48) | (c<<40) | (d<<32) | | |
840 (D<<24) | (C<<16) | (B<<8) | (A); | |
841 //lut[i] = (v<<32) | (v<<24); | |
842 } | |
843 } | |
844 | |
845 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) | |
846 asm volatile( | |
847 "pxor %%mm7, %%mm7 \n\t" // 0 | |
848 // "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE | |
849 "leal (%0, %1), %%eax \n\t" | |
850 "leal (%%eax, %1, 4), %%ebx \n\t" | |
851 | |
852 "movq b80, %%mm6 \n\t" | |
2180 | 853 "movd pQPb, %%mm5 \n\t" // QP |
2179 | 854 "movq %%mm5, %%mm4 \n\t" |
855 "paddusb %%mm5, %%mm5 \n\t" // 2QP | |
856 "paddusb %%mm5, %%mm4 \n\t" // 3QP | |
857 "pxor %%mm5, %%mm5 \n\t" // 0 | |
858 "psubb %%mm4, %%mm5 \n\t" // -3QP | |
859 "por bm11111110, %%mm5 \n\t" // ...,FF,FF,-3QP | |
860 "psllq $24, %%mm5 \n\t" | |
861 | |
862 // 0 1 2 3 4 5 6 7 8 9 | |
863 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 | |
864 | |
865 #define HX1old(a) \ | |
866 "movd " #a ", %%mm0 \n\t"\ | |
867 "movd 4" #a ", %%mm1 \n\t"\ | |
868 "punpckldq %%mm1, %%mm0 \n\t"\ | |
869 "movq %%mm0, %%mm1 \n\t"\ | |
870 "movq %%mm0, %%mm2 \n\t"\ | |
871 "psrlq $8, %%mm1 \n\t"\ | |
872 "psubusb %%mm1, %%mm2 \n\t"\ | |
873 "psubusb %%mm0, %%mm1 \n\t"\ | |
874 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ | |
875 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
876 "pshufw $0x00, %%mm1, %%mm3 \n\t" /* p´5 = |p1 - p2| */\ | |
877 PAVGB(%%mm1, %%mm3) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\ | |
878 "psrlq $16, %%mm3 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\ | |
879 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\ | |
880 "paddb %%mm5, %%mm1 \n\t"\ | |
881 "psubusb %%mm5, %%mm1 \n\t"\ | |
882 PAVGB(%%mm7, %%mm1)\ | |
883 "pxor %%mm2, %%mm1 \n\t"\ | |
884 "psubb %%mm2, %%mm1 \n\t"\ | |
885 "psrlq $24, %%mm1 \n\t"\ | |
886 "movd %%mm1, %%ecx \n\t"\ | |
887 "paddb %%mm6, %%mm0 \n\t"\ | |
888 "paddsb (%3, %%ecx, 8), %%mm0 \n\t"\ | |
889 "paddb %%mm6, %%mm0 \n\t"\ | |
890 "movq %%mm0, " #a " \n\t"\ | |
891 | |
892 /* | |
893 HX1old((%0)) | |
894 HX1old((%%eax)) | |
895 HX1old((%%eax, %1)) | |
896 HX1old((%%eax, %1, 2)) | |
897 HX1old((%0, %1, 4)) | |
898 HX1old((%%ebx)) | |
899 HX1old((%%ebx, %1)) | |
900 HX1old((%%ebx, %1, 2)) | |
901 */ | |
902 | |
903 //FIXME add some comments, its unreadable ... | |
904 #define HX1b(a, c, b, d) \ | |
905 "movd " #a ", %%mm0 \n\t"\ | |
906 "movd 4" #a ", %%mm1 \n\t"\ | |
907 "punpckldq %%mm1, %%mm0 \n\t"\ | |
908 "movd " #b ", %%mm4 \n\t"\ | |
909 "movq %%mm0, %%mm1 \n\t"\ | |
910 "movq %%mm0, %%mm2 \n\t"\ | |
911 "psrlq $8, %%mm1 \n\t"\ | |
912 "movd 4" #b ", %%mm3 \n\t"\ | |
913 "psubusb %%mm1, %%mm2 \n\t"\ | |
914 "psubusb %%mm0, %%mm1 \n\t"\ | |
915 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ | |
916 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
917 "punpckldq %%mm3, %%mm4 \n\t"\ | |
918 "movq %%mm1, %%mm3 \n\t"\ | |
919 "psllq $32, %%mm3 \n\t" /* p´5 = |p1 - p2| */\ | |
920 PAVGB(%%mm1, %%mm3) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\ | |
921 "paddb %%mm6, %%mm0 \n\t"\ | |
922 "psrlq $16, %%mm3 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\ | |
923 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\ | |
924 "movq %%mm4, %%mm3 \n\t"\ | |
925 "paddb %%mm5, %%mm1 \n\t"\ | |
926 "psubusb %%mm5, %%mm1 \n\t"\ | |
927 "psrlq $8, %%mm3 \n\t"\ | |
928 PAVGB(%%mm7, %%mm1)\ | |
929 "pxor %%mm2, %%mm1 \n\t"\ | |
930 "psubb %%mm2, %%mm1 \n\t"\ | |
931 "movq %%mm4, %%mm2 \n\t"\ | |
932 "psrlq $24, %%mm1 \n\t"\ | |
933 "psubusb %%mm3, %%mm2 \n\t"\ | |
934 "movd %%mm1, %%ecx \n\t"\ | |
935 "psubusb %%mm4, %%mm3 \n\t"\ | |
2180 | 936 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\ |
2179 | 937 "por %%mm2, %%mm3 \n\t" /* p´x = |px - p(x+1)| */\ |
938 "paddb %%mm6, %%mm0 \n\t"\ | |
939 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
940 "movq %%mm3, %%mm1 \n\t"\ | |
941 "psllq $32, %%mm1 \n\t" /* p´5 = |p1 - p2| */\ | |
942 "movq %%mm0, " #a " \n\t"\ | |
943 PAVGB(%%mm3, %%mm1) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\ | |
944 "paddb %%mm6, %%mm4 \n\t"\ | |
945 "psrlq $16, %%mm1 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\ | |
946 "psubusb %%mm1, %%mm3 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\ | |
947 "paddb %%mm5, %%mm3 \n\t"\ | |
948 "psubusb %%mm5, %%mm3 \n\t"\ | |
949 PAVGB(%%mm7, %%mm3)\ | |
950 "pxor %%mm2, %%mm3 \n\t"\ | |
951 "psubb %%mm2, %%mm3 \n\t"\ | |
952 "psrlq $24, %%mm3 \n\t"\ | |
953 "movd " #c ", %%mm0 \n\t"\ | |
954 "movd 4" #c ", %%mm1 \n\t"\ | |
955 "punpckldq %%mm1, %%mm0 \n\t"\ | |
956 "paddb %%mm6, %%mm0 \n\t"\ | |
2180 | 957 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\ |
2179 | 958 "paddb %%mm6, %%mm0 \n\t"\ |
959 "movq %%mm0, " #c " \n\t"\ | |
960 "movd %%mm3, %%ecx \n\t"\ | |
961 "movd " #d ", %%mm0 \n\t"\ | |
2180 | 962 "paddsb (%2, %%ecx, 8), %%mm4 \n\t"\ |
2179 | 963 "movd 4" #d ", %%mm1 \n\t"\ |
964 "paddb %%mm6, %%mm4 \n\t"\ | |
965 "punpckldq %%mm1, %%mm0 \n\t"\ | |
966 "movq %%mm4, " #b " \n\t"\ | |
967 "paddb %%mm6, %%mm0 \n\t"\ | |
2180 | 968 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\ |
2179 | 969 "paddb %%mm6, %%mm0 \n\t"\ |
970 "movq %%mm0, " #d " \n\t"\ | |
971 | |
972 HX1b((%0),(%%eax),(%%eax, %1),(%%eax, %1, 2)) | |
973 HX1b((%0, %1, 4),(%%ebx),(%%ebx, %1),(%%ebx, %1, 2)) | |
974 | |
975 | |
976 : | |
2180 | 977 : "r" (src), "r" (stride), "r" (lut) |
2179 | 978 : "%eax", "%ebx", "%ecx" |
979 ); | |
980 #else | |
981 | |
982 //FIXME (has little in common with the mmx2 version) | |
983 for(y=0; y<BLOCK_SIZE; y++) | |
984 { | |
985 int a= src[1] - src[2]; | |
986 int b= src[3] - src[4]; | |
987 int c= src[5] - src[6]; | |
988 | |
989 int d= MAX(ABS(b) - (ABS(a) + ABS(c))/2, 0); | |
990 | |
991 if(d < QP) | |
992 { | |
993 int v = d * SIGN(-b); | |
994 | |
995 src[1] +=v/8; | |
996 src[2] +=v/4; | |
997 src[3] +=3*v/8; | |
998 src[4] -=3*v/8; | |
999 src[5] -=v/4; | |
1000 src[6] -=v/8; | |
1001 | |
1002 } | |
1003 src+=stride; | |
1004 } | |
1005 #endif | |
1006 } | |
1007 | |
2159 | 1008 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1009 static inline void doVertDefFilter(uint8_t src[], int stride, int QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1010 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1011 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1012 src+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1013 //FIXME try pmul for *5 stuff |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1014 // src[0]=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1015 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1016 "pxor %%mm7, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1017 "leal (%0, %1), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1018 "leal (%%eax, %1, 4), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1019 // 0 1 2 3 4 5 6 7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1020 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ebx+%1 ebx+2%1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1021 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1022 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1023 "movq (%0), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1024 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1025 "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1026 "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1027 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1028 "movq (%%eax), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1029 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1030 "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1031 "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1032 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1033 "movq (%%eax, %1), %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1034 "movq %%mm4, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1035 "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1036 "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1037 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1038 "paddw %%mm0, %%mm0 \n\t" // 2L0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1039 "paddw %%mm1, %%mm1 \n\t" // 2H0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1040 "psubw %%mm4, %%mm2 \n\t" // L1 - L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1041 "psubw %%mm5, %%mm3 \n\t" // H1 - H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1042 "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1043 "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1044 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1045 "psllw $2, %%mm2 \n\t" // 4L1 - 4L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1046 "psllw $2, %%mm3 \n\t" // 4H1 - 4H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1047 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1048 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1049 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1050 "movq (%%eax, %1, 2), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1051 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1052 "punpcklbw %%mm7, %%mm2 \n\t" // L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1053 "punpckhbw %%mm7, %%mm3 \n\t" // H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1054 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1055 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1056 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1057 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1058 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1059 "movq %%mm0, temp0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1060 "movq %%mm1, temp1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1061 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1062 "movq (%0, %1, 4), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1063 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1064 "punpcklbw %%mm7, %%mm0 \n\t" // L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1065 "punpckhbw %%mm7, %%mm1 \n\t" // H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1066 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1067 "psubw %%mm0, %%mm2 \n\t" // L3 - L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1068 "psubw %%mm1, %%mm3 \n\t" // H3 - H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1069 "movq %%mm2, temp2 \n\t" // L3 - L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1070 "movq %%mm3, temp3 \n\t" // H3 - H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1071 "paddw %%mm4, %%mm4 \n\t" // 2L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1072 "paddw %%mm5, %%mm5 \n\t" // 2H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1073 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1074 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1075 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1076 "psllw $2, %%mm2 \n\t" // 4L3 - 4L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1077 "psllw $2, %%mm3 \n\t" // 4H3 - 4H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1078 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1079 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1080 //50 opcodes so far |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1081 "movq (%%ebx), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1082 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1083 "punpcklbw %%mm7, %%mm2 \n\t" // L5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1084 "punpckhbw %%mm7, %%mm3 \n\t" // H5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1085 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1086 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1087 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1088 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1089 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1090 "movq (%%ebx, %1), %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1091 "punpcklbw %%mm7, %%mm6 \n\t" // L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1092 "psubw %%mm6, %%mm2 \n\t" // L5 - L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1093 "movq (%%ebx, %1), %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1094 "punpckhbw %%mm7, %%mm6 \n\t" // H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1095 "psubw %%mm6, %%mm3 \n\t" // H5 - H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1096 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1097 "paddw %%mm0, %%mm0 \n\t" // 2L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1098 "paddw %%mm1, %%mm1 \n\t" // 2H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1099 "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1100 "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1101 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1102 "psllw $2, %%mm2 \n\t" // 4L5 - 4L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1103 "psllw $2, %%mm3 \n\t" // 4H5 - 4H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1104 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1105 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1106 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1107 "movq (%%ebx, %1, 2), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1108 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1109 "punpcklbw %%mm7, %%mm2 \n\t" // L7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1110 "punpckhbw %%mm7, %%mm3 \n\t" // H7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1111 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1112 "paddw %%mm2, %%mm2 \n\t" // 2L7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1113 "paddw %%mm3, %%mm3 \n\t" // 2H7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1114 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1115 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1116 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1117 "movq temp0, %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1118 "movq temp1, %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1119 //FIXME pxor, psubw, pmax for abs |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1120 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1121 "pcmpgtw %%mm0, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1122 "pxor %%mm6, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1123 "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1124 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1125 "pcmpgtw %%mm1, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1126 "pxor %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1127 "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1128 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1129 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1130 "pcmpgtw %%mm2, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1131 "pxor %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1132 "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1133 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1134 "pcmpgtw %%mm3, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1135 "pxor %%mm6, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1136 "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1137 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1138 #ifdef HAVE_MMX2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1139 "pminsw %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1140 "pminsw %%mm3, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1141 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1142 "movq %%mm0, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1143 "psubusw %%mm2, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1144 "psubw %%mm6, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1145 "movq %%mm1, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1146 "psubusw %%mm3, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1147 "psubw %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1148 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1149 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1150 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1151 "pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1152 "pxor %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1153 "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1154 "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1155 "pxor %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1156 "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1157 // 100 opcodes |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1158 "movd %2, %%mm2 \n\t" // QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1159 //"pcmpeqb %%mm2, %%mm2\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1160 "punpcklwd %%mm2, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1161 "punpcklwd %%mm2, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1162 "psllw $3, %%mm2 \n\t" // 8QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1163 "movq %%mm2, %%mm3 \n\t" // 8QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1164 "pcmpgtw %%mm4, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1165 "pcmpgtw %%mm5, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1166 "pand %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1167 "pand %%mm3, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1168 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1169 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1170 "psubusw %%mm0, %%mm4 \n\t" // hd |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1171 "psubusw %%mm1, %%mm5 \n\t" // ld |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1172 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1173 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1174 "movq w05, %%mm2 \n\t" // 5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1175 "pmullw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1176 "pmullw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1177 "movq w20, %%mm2 \n\t" // 32 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1178 "paddw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1179 "paddw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1180 "psrlw $6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1181 "psrlw $6, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1182 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1183 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1184 "movq w06, %%mm2 \n\t" // 6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1185 "paddw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1186 "paddw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1187 "movq w1400, %%mm2 \n\t" // 1400h = 5120 = 5/64*2^16 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1188 //FIXME if *5/64 is supposed to be /13 then we should use 5041 instead of 5120 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1189 "pmulhw %%mm2, %%mm4 \n\t" // hd/13 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1190 "pmulhw %%mm2, %%mm5 \n\t" // ld/13 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1191 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1192 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1193 "movq temp2, %%mm0 \n\t" // L3 - L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1194 "movq temp3, %%mm1 \n\t" // H3 - H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1195 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1196 "pxor %%mm2, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1197 "pxor %%mm3, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1198 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1199 // FIXME rounding error |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1200 "psraw $1, %%mm0 \n\t" // (L3 - L4)/2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1201 "psraw $1, %%mm1 \n\t" // (H3 - H4)/2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1202 "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1203 "pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1204 "pxor %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1205 "pxor %%mm3, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1206 "psubw %%mm2, %%mm0 \n\t" // |L3-L4| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1207 "psubw %%mm3, %%mm1 \n\t" // |H3-H4| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1208 // "psrlw $1, %%mm0 \n\t" // |L3 - L4|/2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1209 // "psrlw $1, %%mm1 \n\t" // |H3 - H4|/2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1210 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1211 "pxor %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1212 "pxor %%mm7, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1213 "pand %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1214 "pand %%mm3, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1215 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1216 #ifdef HAVE_MMX2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1217 "pminsw %%mm0, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1218 "pminsw %%mm1, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1219 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1220 "movq %%mm4, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1221 "psubusw %%mm0, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1222 "psubw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1223 "movq %%mm5, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1224 "psubusw %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1225 "psubw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1226 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1227 "pxor %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1228 "pxor %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1229 "psubw %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1230 "psubw %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1231 "packsswb %%mm5, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1232 "movq (%%eax, %1, 2), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1233 "paddb %%mm4, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1234 "movq %%mm0, (%%eax, %1, 2) \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1235 "movq (%0, %1, 4), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1236 "psubb %%mm4, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1237 // "pxor %%mm0, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1238 "movq %%mm0, (%0, %1, 4) \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1239 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1240 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1241 : "r" (src), "r" (stride), "r" (QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1242 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1243 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1244 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1245 const int l1= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1246 const int l2= stride + l1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1247 const int l3= stride + l2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1248 const int l4= stride + l3; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1249 const int l5= stride + l4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1250 const int l6= stride + l5; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1251 const int l7= stride + l6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1252 const int l8= stride + l7; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1253 // const int l9= stride + l8; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1254 int x; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1255 for(x=0; x<BLOCK_SIZE; x++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1256 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1257 const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1258 if(ABS(middleEnergy) < 8*QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1259 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1260 const int q=(src[l4] - src[l5])/2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1261 const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1262 const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1263 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1264 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1265 d= MAX(d, 0); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1266 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1267 d= (5*d + 32) >> 6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1268 d*= SIGN(-middleEnergy); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1269 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1270 if(q>0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1271 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1272 d= d<0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1273 d= d>q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1274 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1275 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1276 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1277 d= d>0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1278 d= d<q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1279 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1280 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1281 src[l4]-= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1282 src[l5]+= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1283 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1284 src++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1285 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1286 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1287 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1288 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1289 //FIXME? |255-0| = 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1290 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1291 * Check if the given 8x8 Block is mostly "flat" and copy the unaliged data into tempBlock. |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1292 */ |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1293 static inline int isHorizDCAndCopy2Temp(uint8_t src[], int stride) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1294 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1295 // src++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1296 int numEq= 0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1297 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1298 asm volatile ( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1299 // "int $3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1300 "pushl %1\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1301 "movq b7E, %%mm7 \n\t" // mm7 = 0x7F |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1302 "movq b7C, %%mm6 \n\t" // mm6 = 0x7D |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1303 "leal tempBlock, %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1304 "pxor %%mm0, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1305 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1306 #define HDC_CHECK_AND_CPY(i) \ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1307 "movq -4(%1), %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1308 "psrlq $32, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1309 "punpckldq 4(%1), %%mm2 \n\t" /* (%1) */\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1310 "movq %%mm2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1311 "psrlq $8, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1312 "psubb %%mm1, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1313 "paddb %%mm7, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1314 "pcmpgtb %%mm6, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1315 "paddb %%mm2, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1316 "movq %%mm1," #i "(%%eax) \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1317 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1318 HDC_CHECK_AND_CPY(0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1319 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1320 HDC_CHECK_AND_CPY(8) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1321 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1322 HDC_CHECK_AND_CPY(16) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1323 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1324 HDC_CHECK_AND_CPY(24) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1325 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1326 HDC_CHECK_AND_CPY(32) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1327 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1328 HDC_CHECK_AND_CPY(40) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1329 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1330 HDC_CHECK_AND_CPY(48) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1331 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1332 HDC_CHECK_AND_CPY(56) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1333 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1334 "psllq $8, %%mm0 \n\t" // remove dummy value |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1335 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1336 "psrlw $8, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1337 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1338 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1339 "psrlq $16, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1340 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1341 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1342 "psrlq $32, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1343 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1344 "popl %1\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1345 "movd %%mm0, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1346 : "=r" (numEq) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1347 : "r" (src), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1348 : "%eax" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1349 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1350 // printf("%d\n", numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1351 numEq= (256 - (numEq & 0xFF)) &0xFF; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1352 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1353 int y; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1354 for(y=0; y<BLOCK_SIZE; y++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1355 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1356 if(((src[0] - src[1] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1357 if(((src[1] - src[2] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1358 if(((src[2] - src[3] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1359 if(((src[3] - src[4] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1360 if(((src[4] - src[5] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1361 if(((src[5] - src[6] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1362 if(((src[6] - src[7] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1363 tempBlock[0 + y*TEMP_STRIDE] = src[0]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1364 tempBlock[1 + y*TEMP_STRIDE] = src[1]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1365 tempBlock[2 + y*TEMP_STRIDE] = src[2]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1366 tempBlock[3 + y*TEMP_STRIDE] = src[3]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1367 tempBlock[4 + y*TEMP_STRIDE] = src[4]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1368 tempBlock[5 + y*TEMP_STRIDE] = src[5]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1369 tempBlock[6 + y*TEMP_STRIDE] = src[6]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1370 tempBlock[7 + y*TEMP_STRIDE] = src[7]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1371 src+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1372 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1373 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1374 /* if(abs(numEq - asmEq) > 0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1375 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1376 // printf("\nasm:%d c:%d\n", asmEq, numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1377 for(int y=0; y<8; y++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1378 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1379 for(int x=0; x<8; x++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1380 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1381 printf("%d ", src[x + y*stride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1382 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1383 printf("\n"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1384 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1385 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1386 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1387 // printf("%d\n", numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1388 return numEq > hFlatnessThreshold; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1389 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1390 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1391 static inline int isHorizMinMaxOk(uint8_t src[], int stride, int QP) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1392 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1393 #ifdef MMX_FIXME |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1394 FIXME |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1395 int isOk; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1396 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1397 // "int $3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1398 "movq (%1, %2), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1399 "movq (%1, %2, 8), %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1400 "movq %%mm0, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1401 "psubusb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1402 "psubusb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1403 "por %%mm1, %%mm0 \n\t" // ABS Diff |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1404 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1405 "movq pQPb, %%mm7 \n\t" // QP,..., QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1406 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1407 "psubusb %%mm7, %%mm0 \n\t" // Diff <= 2QP -> 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1408 "pcmpeqd b00, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1409 "psrlq $16, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1410 "pcmpeqd bFF, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1411 // "movd %%mm0, (%1, %2, 4)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1412 "movd %%mm0, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1413 : "=r" (isOk) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1414 : "r" (src), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1415 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1416 return isOk; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1417 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1418 if(abs(src[0] - src[7]) > 2*QP) return 0; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1419 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1420 return 1; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1421 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1422 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1423 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1424 static inline void doHorizDefFilterAndCopyBack(uint8_t dst[], int stride, int QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1425 { |
2159 | 1426 #ifdef HAVE_MMX |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1427 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1428 "pushl %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1429 "pxor %%mm7, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1430 "movq bm00001000, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1431 "movd %2, %%mm5 \n\t" // QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1432 "movq %%mm5, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1433 "paddusb %%mm5, %%mm5 \n\t" // 2QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1434 "paddusb %%mm5, %%mm4 \n\t" // 3QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1435 "psllq $24, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1436 "pxor %%mm5, %%mm5 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1437 "psubb %%mm4, %%mm5 \n\t" // -QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1438 "leal tempBlock, %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1439 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1440 //FIXME? "unroll by 2" and mix |
2159 | 1441 #ifdef HAVE_MMX2 |
1442 #define HDF(i) \ | |
1443 "movq " #i "(%%eax), %%mm0 \n\t"\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1444 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1445 "movq %%mm0, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1446 "psrlq $8, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1447 "psubusb %%mm1, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1448 "psubusb %%mm0, %%mm1 \n\t"\ |
2159 | 1449 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ |
1450 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
1451 "pshufw $0x00, %%mm1, %%mm3 \n\t" /* p´5 = |p1 - p2| */\ | |
1452 "pminub %%mm1, %%mm3 \n\t" /* p´5 = min(|p2-p1|, |p6-p5|)*/\ | |
1453 "psrlq $16, %%mm3 \n\t" /* p´3 = min(|p2-p1|, |p6-p5|)*/\ | |
1454 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-min(|p1-p2|,|p5-p6|) */\ | |
1455 "paddb %%mm5, %%mm1 \n\t"\ | |
1456 "psubusb %%mm5, %%mm1 \n\t"\ | |
1457 "psrlw $2, %%mm1 \n\t"\ | |
1458 "pxor %%mm2, %%mm1 \n\t"\ | |
1459 "psubb %%mm2, %%mm1 \n\t"\ | |
1460 "pand %%mm6, %%mm1 \n\t"\ | |
1461 "psubb %%mm1, %%mm0 \n\t"\ | |
1462 "psllq $8, %%mm1 \n\t"\ | |
1463 "paddb %%mm1, %%mm0 \n\t"\ | |
1464 "movd %%mm0, (%0) \n\t"\ | |
1465 "psrlq $32, %%mm0 \n\t"\ | |
1466 "movd %%mm0, 4(%0) \n\t" | |
1467 #else | |
1468 #define HDF(i)\ | |
1469 "movq " #i "(%%eax), %%mm0 \n\t"\ | |
1470 "movq %%mm0, %%mm1 \n\t"\ | |
1471 "movq %%mm0, %%mm2 \n\t"\ | |
1472 "psrlq $8, %%mm1 \n\t"\ | |
1473 "psubusb %%mm1, %%mm2 \n\t"\ | |
1474 "psubusb %%mm0, %%mm1 \n\t"\ | |
1475 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ | |
1476 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
1477 "movq %%mm1, %%mm3 \n\t"\ | |
1478 "psllq $32, %%mm3 \n\t"\ | |
1479 "movq %%mm3, %%mm4 \n\t"\ | |
1480 "psubusb %%mm1, %%mm4 \n\t"\ | |
1481 "psubb %%mm4, %%mm3 \n\t"\ | |
1482 "psrlq $16, %%mm3 \n\t" /* p´3 = min(|p2-p1|, |p6-p5|)*/\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1483 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-min(|p1-p2|,|p5,ü6|) */\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1484 "paddb %%mm5, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1485 "psubusb %%mm5, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1486 "psrlw $2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1487 "pxor %%mm2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1488 "psubb %%mm2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1489 "pand %%mm6, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1490 "psubb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1491 "psllq $8, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1492 "paddb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1493 "movd %%mm0, (%0) \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1494 "psrlq $32, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1495 "movd %%mm0, 4(%0) \n\t" |
2159 | 1496 #endif |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1497 HDF(0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1498 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1499 HDF(8) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1500 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1501 HDF(16) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1502 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1503 HDF(24) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1504 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1505 HDF(32) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1506 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1507 HDF(40) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1508 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1509 HDF(48) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1510 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1511 HDF(56) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1512 "popl %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1513 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1514 : "r" (dst), "r" (stride), "r" (QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1515 : "%eax" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1516 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1517 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1518 uint8_t *src= tempBlock; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1519 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1520 int y; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1521 for(y=0; y<BLOCK_SIZE; y++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1522 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1523 dst[0] = src[0]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1524 dst[1] = src[1]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1525 dst[2] = src[2]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1526 dst[3] = src[3]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1527 dst[4] = src[4]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1528 dst[5] = src[5]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1529 dst[6] = src[6]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1530 dst[7] = src[7]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1531 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1532 const int middleEnergy= 5*(src[4] - src[5]) + 2*(src[2] - src[5]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1533 if(ABS(middleEnergy) < 8*QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1534 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1535 const int q=(src[3] - src[4])/2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1536 const int leftEnergy= 5*(src[2] - src[1]) + 2*(src[0] - src[3]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1537 const int rightEnergy= 5*(src[6] - src[5]) + 2*(src[4] - src[7]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1538 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1539 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1540 d= MAX(d, 0); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1541 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1542 d= (5*d + 32) >> 6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1543 d*= SIGN(-middleEnergy); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1544 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1545 if(q>0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1546 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1547 d= d<0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1548 d= d>q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1549 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1550 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1551 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1552 d= d>0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1553 d= d<q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1554 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1555 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1556 dst[3]-= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1557 dst[4]+= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1558 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1559 dst+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1560 src+= TEMP_STRIDE; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1561 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1562 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1563 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1564 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1565 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1566 * Do a horizontal low pass filter on the 8x8 block |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1567 * useing the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 (C version) |
2159 | 1568 * useing approximately the 7-Tap Filter (1,2,3,4,3,2,1)/16 (MMX2/3DNOW version) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1569 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1570 static inline void doHorizLowPassAndCopyBack(uint8_t dst[], int stride, int QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1571 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1572 //return; |
2159 | 1573 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1574 asm volatile( //"movv %0 %1 %2\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1575 "pushl %0\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1576 "pxor %%mm7, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1577 "leal tempBlock, %%eax \n\t" |
2159 | 1578 /* |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1579 #define HLP1 "movq (%0), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1580 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1581 "psllq $8, %%mm0 \n\t"\ |
2159 | 1582 PAVGB(%%mm1, %%mm0)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1583 "psrlw $8, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1584 "pxor %%mm1, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1585 "packuswb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1586 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1587 "movq %%mm0, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1588 "psllq $32, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1589 "paddb %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1590 "psllq $16, %%mm2 \n\t"\ |
2159 | 1591 PAVGB(%%mm2, %%mm0)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1592 "movq %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1593 "pand bm11001100, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1594 "paddusb %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1595 "psrlq $8, %%mm3 \n\t"\ |
2159 | 1596 PAVGB(%%mm1, %%mm4)\ |
1597 PAVGB(%%mm3, %%mm2)\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1598 "psrlq $16, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1599 "punpcklbw %%mm2, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1600 "movq %%mm2, (%0) \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1601 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1602 #define HLP2 "movq (%0), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1603 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1604 "psllq $8, %%mm0 \n\t"\ |
2159 | 1605 PAVGB(%%mm1, %%mm0)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1606 "psrlw $8, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1607 "pxor %%mm1, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1608 "packuswb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1609 "movq %%mm0, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1610 "psllq $32, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1611 "psllq $16, %%mm2 \n\t"\ |
2159 | 1612 PAVGB(%%mm2, %%mm0)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1613 "movq %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1614 "pand bm11001100, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1615 "paddusb %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1616 "psrlq $8, %%mm3 \n\t"\ |
2159 | 1617 PAVGB(%%mm3, %%mm2)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1618 "psrlq $16, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1619 "punpcklbw %%mm2, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1620 "movq %%mm2, (%0) \n\t"\ |
2159 | 1621 */ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1622 // approximately a 7-Tap Filter with Vector (1,2,3,4,3,2,1)/16 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1623 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1624 31 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1625 121 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1626 121 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1627 121 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1628 121 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1629 121 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1630 121 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1631 13 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1632 Implemented Exact 7-Tap |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1633 9421 A321 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1634 36421 64321 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1635 334321 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1636 1234321 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1637 1234321 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1638 123433 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1639 12463 12346 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1640 1249 123A |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1641 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1642 */ |
2159 | 1643 #ifdef HAVE_MMX2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1644 #define HLP3(i) "movq " #i "(%%eax), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1645 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1646 "movq %%mm0, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1647 "movq %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1648 "movq %%mm0, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1649 "psllq $8, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1650 "psrlq $8, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1651 "pand bm00000001, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1652 "pand bm10000000, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1653 "por %%mm3, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1654 "por %%mm4, %%mm2 \n\t"\ |
2159 | 1655 PAVGB(%%mm2, %%mm1)\ |
1656 PAVGB(%%mm1, %%mm0)\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1657 \ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1658 "pshufw $0xF9, %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1659 "pshufw $0x90, %%mm0, %%mm4 \n\t"\ |
2159 | 1660 PAVGB(%%mm3, %%mm4)\ |
1661 PAVGB(%%mm4, %%mm0)\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1662 "movd %%mm0, (%0) \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1663 "psrlq $32, %%mm0 \n\t"\ |
2159 | 1664 "movd %%mm0, 4(%0) \n\t" |
1665 #else | |
1666 #define HLP3(i) "movq " #i "(%%eax), %%mm0 \n\t"\ | |
1667 "movq %%mm0, %%mm1 \n\t"\ | |
1668 "movq %%mm0, %%mm2 \n\t"\ | |
1669 "movq %%mm0, %%mm3 \n\t"\ | |
1670 "movq %%mm0, %%mm4 \n\t"\ | |
1671 "psllq $8, %%mm1 \n\t"\ | |
1672 "psrlq $8, %%mm2 \n\t"\ | |
1673 "pand bm00000001, %%mm3 \n\t"\ | |
1674 "pand bm10000000, %%mm4 \n\t"\ | |
1675 "por %%mm3, %%mm1 \n\t"\ | |
1676 "por %%mm4, %%mm2 \n\t"\ | |
1677 PAVGB(%%mm2, %%mm1)\ | |
1678 PAVGB(%%mm1, %%mm0)\ | |
1679 \ | |
1680 "movq %%mm0, %%mm3 \n\t"\ | |
1681 "movq %%mm0, %%mm4 \n\t"\ | |
1682 "movq %%mm0, %%mm5 \n\t"\ | |
1683 "psrlq $16, %%mm3 \n\t"\ | |
1684 "psllq $16, %%mm4 \n\t"\ | |
1685 "pand bm11000000, %%mm5 \n\t"\ | |
1686 "por %%mm5, %%mm3 \n\t"\ | |
1687 "movq %%mm0, %%mm5 \n\t"\ | |
1688 "pand bm00000011, %%mm5 \n\t"\ | |
1689 "por %%mm5, %%mm4 \n\t"\ | |
1690 PAVGB(%%mm3, %%mm4)\ | |
1691 PAVGB(%%mm4, %%mm0)\ | |
1692 "movd %%mm0, (%0) \n\t"\ | |
1693 "psrlq $32, %%mm0 \n\t"\ | |
1694 "movd %%mm0, 4(%0) \n\t" | |
1695 #endif | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1696 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1697 #define HLP(i) HLP3(i) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1698 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1699 HLP(0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1700 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1701 HLP(8) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1702 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1703 HLP(16) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1704 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1705 HLP(24) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1706 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1707 HLP(32) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1708 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1709 HLP(40) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1710 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1711 HLP(48) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1712 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1713 HLP(56) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1714 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1715 "popl %0\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1716 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1717 : "r" (dst), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1718 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1719 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1720 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1721 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1722 uint8_t *temp= tempBlock; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1723 int y; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1724 for(y=0; y<BLOCK_SIZE; y++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1725 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1726 const int first= ABS(dst[-1] - dst[0]) < QP ? dst[-1] : dst[0]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1727 const int last= ABS(dst[8] - dst[7]) < QP ? dst[8] : dst[7]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1728 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1729 int sums[9]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1730 sums[0] = first + temp[0]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1731 sums[1] = temp[0] + temp[1]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1732 sums[2] = temp[1] + temp[2]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1733 sums[3] = temp[2] + temp[3]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1734 sums[4] = temp[3] + temp[4]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1735 sums[5] = temp[4] + temp[5]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1736 sums[6] = temp[5] + temp[6]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1737 sums[7] = temp[6] + temp[7]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1738 sums[8] = temp[7] + last; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1739 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1740 dst[0]= ((sums[0]<<2) + ((first + sums[2])<<1) + sums[4] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1741 dst[1]= ((dst[1]<<2) + (first + sums[0] + sums[3]<<1) + sums[5] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1742 dst[2]= ((dst[2]<<2) + (first + sums[1] + sums[4]<<1) + sums[6] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1743 dst[3]= ((dst[3]<<2) + (sums[2] + sums[5]<<1) + sums[0] + sums[7] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1744 dst[4]= ((dst[4]<<2) + (sums[3] + sums[6]<<1) + sums[1] + sums[8] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1745 dst[5]= ((dst[5]<<2) + (last + sums[7] + sums[4]<<1) + sums[2] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1746 dst[6]= ((last + dst[6]<<2) + (dst[7] + sums[5]<<1) + sums[3] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1747 dst[7]= ((sums[8]<<2) + (last + sums[6]<<1) + sums[4] + 8)>>4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1748 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1749 dst+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1750 temp+= TEMP_STRIDE; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1751 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1752 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1753 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1754 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1755 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1756 static inline void dering(uint8_t src[], int stride, int QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1757 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1758 //FIXME |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1759 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1760 #ifdef HAVE_MMX2X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1761 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1762 "leal (%0, %1), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1763 "leal (%%eax, %1, 4), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1764 // 0 1 2 3 4 5 6 7 8 9 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1765 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1766 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1767 "pcmpeq %%mm6, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1768 "pxor %%mm7, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1769 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1770 #define FIND_MIN_MAX(addr)\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1771 "movq (" #addr "), %%mm0, \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1772 "pminub %%mm0, %%mm6 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1773 "pmaxub %%mm0, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1774 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1775 FIND_MIN_MAX(%0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1776 FIND_MIN_MAX(%%eax) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1777 FIND_MIN_MAX(%%eax, %1) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1778 FIND_MIN_MAX(%%eax, %1, 2) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1779 FIND_MIN_MAX(%0, %1, 4) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1780 FIND_MIN_MAX(%%ebx) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1781 FIND_MIN_MAX(%%ebx, %1) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1782 FIND_MIN_MAX(%%ebx, %1, 2) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1783 FIND_MIN_MAX(%0, %1, 8) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1784 FIND_MIN_MAX(%%ebx, %1, 2) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1785 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1786 "movq %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1787 "psrlq $32, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1788 "pminub %%mm4, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1789 "movq %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1790 "psrlq $16, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1791 "pminub %%mm4, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1792 "movq %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1793 "psrlq $8, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1794 "pminub %%mm4, %%mm6 \n\t" // min of pixels |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1795 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1796 "movq %%mm7, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1797 "psrlq $32, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1798 "pmaxub %%mm4, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1799 "movq %%mm7, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1800 "psrlq $16, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1801 "pmaxub %%mm4, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1802 "movq %%mm7, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1803 "psrlq $8, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1804 "pmaxub %%mm4, %%mm7 \n\t" // max of pixels |
2159 | 1805 PAVGB(%%mm6, %%mm7) // (max + min)/2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1806 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1807 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1808 : : "r" (src), "r" (stride), "r" (QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1809 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1810 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1811 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1812 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1813 //FIXME |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1814 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1815 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1816 |
2159 | 1817 |
1818 | |
1819 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1820 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1821 * ... |
2159 | 1822 * the mode value is interpreted as a quality value if its negative, its range is then (-1 ... -63) |
1823 * -63 is best quality -1 is worst | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1824 */ |
2169 | 1825 //extern "C"{ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1826 void postprocess(unsigned char * src[], int src_stride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1827 unsigned char * dst[], int dst_stride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1828 int horizontal_size, int vertical_size, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1829 QP_STORE_T *QP_store, int QP_stride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1830 int mode) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1831 { |
2159 | 1832 |
1833 if(mode<0) mode= getModeForQuality(-mode); | |
1834 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1835 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1836 long long T= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1837 for(int y=vertical_size-1; y>=0 ; y--) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1838 memcpy(dst[0] + y*src_stride, src[0] + y*src_stride,src_stride); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1839 // memcpy(dst[0], src[0],src_stride*vertical_size); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1840 printf("%4dk\r", (rdtsc()-T)/1000); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1841 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1842 return; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1843 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1844 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1845 long long T= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1846 while( (rdtsc() - T)/1000 < 4000); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1847 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1848 return; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1849 */ |
2159 | 1850 postProcess(src[0], src_stride, dst[0], dst_stride, |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1851 horizontal_size, vertical_size, QP_store, QP_stride, 0, mode); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1852 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1853 horizontal_size >>= 1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1854 vertical_size >>= 1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1855 src_stride >>= 1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1856 dst_stride >>= 1; |
2179 | 1857 mode= ((mode&0xFF)>>4) | (mode&0xFFFFFF00); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1858 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1859 if(1) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1860 { |
2159 | 1861 postProcess(src[1], src_stride, dst[1], dst_stride, |
2179 | 1862 horizontal_size, vertical_size, QP_store, QP_stride, 1, mode); |
2159 | 1863 postProcess(src[2], src_stride, dst[2], dst_stride, |
2179 | 1864 horizontal_size, vertical_size, QP_store, QP_stride, 1, mode); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1865 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1866 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1867 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1868 memcpy(dst[1], src[1], src_stride*horizontal_size); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1869 memcpy(dst[2], src[2], src_stride*horizontal_size); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1870 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1871 } |
2159 | 1872 /** |
1873 * gets the mode flags for a given quality (larger values mean slower but better postprocessing) | |
1874 * 0 <= quality < 64 | |
1875 */ | |
1876 int getModeForQuality(int quality){ | |
1877 int modes[6]= { | |
1878 LUM_V_DEBLOCK, | |
1879 LUM_V_DEBLOCK | LUM_H_DEBLOCK, | |
1880 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK, | |
1881 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK, | |
1882 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING, | |
1883 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING | CHROM_DERING | |
1884 }; | |
1885 | |
1886 return modes[ (quality*6) >>6 ]; | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1887 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1888 |
2169 | 1889 //} // extern "C" |
2159 | 1890 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1891 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1892 * Copies a block from src to dst and fixes the blacklevel |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1893 * numLines must be a multiple of 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1894 * levelFix == 0 -> dont touch the brighness & contrast |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1895 */ |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1896 static inline void blockCopy(uint8_t dst[], int dstStride, uint8_t src[], int srcStride, |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1897 int numLines, int levelFix) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1898 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1899 int i; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1900 if(levelFix) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1901 { |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1902 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1903 asm volatile( |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1904 "movl %4, %%eax \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1905 "movl %%eax, temp0\n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1906 "pushl %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1907 "pushl %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1908 "leal (%2,%2), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1909 "leal (%3,%3), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1910 "movq packedYOffset, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1911 "movq packedYScale, %%mm3 \n\t" |
2181
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
1912 "pxor %%mm4, %%mm4 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1913 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1914 #define SCALED_CPY \ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1915 "movq (%0), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1916 "movq (%0,%2), %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1917 "psubusb %%mm2, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1918 "psubusb %%mm2, %%mm1 \n\t"\ |
2181
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
1919 "movq %%mm0, %%mm5 \n\t"\ |
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
1920 "punpcklbw %%mm4, %%mm0 \n\t"\ |
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
1921 "punpckhbw %%mm4, %%mm5 \n\t"\ |
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
1922 "psllw $7, %%mm0 \n\t"\ |
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
1923 "psllw $7, %%mm5 \n\t"\ |
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
1924 "pmulhw %%mm3, %%mm0 \n\t"\ |
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
1925 "pmulhw %%mm3, %%mm5 \n\t"\ |
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
1926 "packuswb %%mm5, %%mm0 \n\t"\ |
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
1927 "movq %%mm0, (%1) \n\t"\ |
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
1928 "movq %%mm1, %%mm5 \n\t"\ |
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
1929 "punpcklbw %%mm4, %%mm1 \n\t"\ |
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
1930 "punpckhbw %%mm4, %%mm5 \n\t"\ |
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
1931 "psllw $7, %%mm1 \n\t"\ |
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
1932 "psllw $7, %%mm5 \n\t"\ |
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
1933 "pmulhw %%mm3, %%mm1 \n\t"\ |
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
1934 "pmulhw %%mm3, %%mm5 \n\t"\ |
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
1935 "packuswb %%mm5, %%mm1 \n\t"\ |
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
1936 "movq %%mm1, (%1, %3) \n\t"\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1937 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1938 "1: \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1939 SCALED_CPY |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1940 "addl %%eax, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1941 "addl %%ebx, %1 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1942 SCALED_CPY |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1943 "addl %%eax, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1944 "addl %%ebx, %1 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1945 "decl temp0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1946 "jnz 1b \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1947 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1948 "popl %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1949 "popl %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1950 : : "r" (src), |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1951 "r" (dst), |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1952 "r" (srcStride), |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1953 "r" (dstStride), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1954 "m" (numLines>>2) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1955 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1956 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1957 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1958 for(i=0; i<numLines; i++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1959 memcpy( &(dst[dstStride*i]), |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1960 &(src[srcStride*i]), BLOCK_SIZE); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1961 #endif |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1962 } |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1963 else |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1964 { |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1965 #ifdef HAVE_MMX |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1966 asm volatile( |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1967 "movl %4, %%eax \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1968 "movl %%eax, temp0\n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1969 "pushl %0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1970 "pushl %1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1971 "leal (%2,%2), %%eax \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1972 "leal (%3,%3), %%ebx \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1973 "movq packedYOffset, %%mm2 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1974 "movq packedYScale, %%mm3 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1975 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1976 #define SIMPLE_CPY \ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1977 "movq (%0), %%mm0 \n\t"\ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1978 "movq (%0,%2), %%mm1 \n\t"\ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1979 "movq %%mm0, (%1) \n\t"\ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1980 "movq %%mm1, (%1, %3) \n\t"\ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1981 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1982 "1: \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1983 SIMPLE_CPY |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1984 "addl %%eax, %0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1985 "addl %%ebx, %1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1986 SIMPLE_CPY |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1987 "addl %%eax, %0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1988 "addl %%ebx, %1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1989 "decl temp0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1990 "jnz 1b \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1991 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1992 "popl %1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1993 "popl %0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1994 : : "r" (src), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1995 "r" (dst), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1996 "r" (srcStride), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1997 "r" (dstStride), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1998 "m" (numLines>>2) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1999 : "%eax", "%ebx" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2000 ); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2001 #else |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2002 for(i=0; i<numLines; i++) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2003 memcpy( &(dst[dstStride*i]), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2004 &(src[srcStride*i]), BLOCK_SIZE); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2005 #endif |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2006 } |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2007 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2008 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2009 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2010 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2011 * Filters array of bytes (Y or U or V values) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2012 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2013 void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height, |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2014 QP_STORE_T QPs[], int QPStride, int isColor, int mode) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2015 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2016 int x,y; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2017 /* we need 64bit here otherwise we´ll going to have a problem |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2018 after watching a black picture for 5 hours*/ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2019 static uint64_t *yHistogram= NULL; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2020 int black=0, white=255; // blackest black and whitest white in the picture |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2021 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2022 #ifdef TIMEING |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2023 long long T0, T1, memcpyTime=0, vertTime=0, horizTime=0, sumTime, diffTime=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2024 sumTime= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2025 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2026 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2027 if(!yHistogram) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2028 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2029 int i; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2030 yHistogram= (uint64_t*)malloc(8*256); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2031 for(i=0; i<256; i++) yHistogram[i]= width*height/64*15/256; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2032 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2033 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2034 if(!isColor) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2035 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2036 uint64_t sum= 0; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2037 int i; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2038 static int framenum= -1; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2039 uint64_t maxClipped; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2040 uint64_t clipped; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2041 double scale; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2042 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2043 framenum++; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2044 if(framenum == 1) yHistogram[0]= width*height/64*15/256; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2045 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2046 for(i=0; i<256; i++) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2047 { |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2048 sum+= yHistogram[i]; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2049 // printf("%d ", yHistogram[i]); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2050 } |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2051 // printf("\n\n"); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2052 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2053 /* we allways get a completly black picture first */ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2054 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2055 maxClipped= (uint64_t)(sum * maxClippedThreshold); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2056 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2057 clipped= sum; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2058 for(black=255; black>0; black--) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2059 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2060 if(clipped < maxClipped) break; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2061 clipped-= yHistogram[black]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2062 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2063 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2064 clipped= sum; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2065 for(white=0; white<256; white++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2066 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2067 if(clipped < maxClipped) break; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2068 clipped-= yHistogram[white]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2069 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2070 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2071 // we cant handle negative correctures |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2072 packedYOffset= MAX(black - minAllowedY, 0); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2073 packedYOffset|= packedYOffset<<32; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2074 packedYOffset|= packedYOffset<<16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2075 packedYOffset|= packedYOffset<<8; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2076 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2077 scale= (double)(maxAllowedY - minAllowedY) / (double)(white-black); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2078 |
2181
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
2079 packedYScale= (uint16_t)(scale*512.0 + 0.5); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2080 packedYScale|= packedYScale<<32; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2081 packedYScale|= packedYScale<<16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2082 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2083 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2084 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2085 packedYScale= 0x0100010001000100LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2086 packedYOffset= 0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2087 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2088 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2089 for(x=0; x<width; x+=BLOCK_SIZE) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2090 blockCopy(dst + x, dstStride, src + x, srcStride, 8, mode & LEVEL_FIX); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2091 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2092 for(y=0; y<height; y+=BLOCK_SIZE) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2093 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2094 //1% speedup if these are here instead of the inner loop |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2095 uint8_t *srcBlock= &(src[y*srcStride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2096 uint8_t *dstBlock= &(dst[y*dstStride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2097 uint8_t *vertSrcBlock= &(srcBlock[srcStride*3]); // Blocks are 10x8 -> *3 to start |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2098 uint8_t *vertBlock= &(dstBlock[dstStride*3]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2099 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2100 // finish 1 block before the next otherwise we´ll might have a problem |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2101 // with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2102 for(x=0; x<width; x+=BLOCK_SIZE) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2103 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2104 const int stride= dstStride; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2105 int QP= isColor ? |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2106 QPs[(y>>3)*QPStride + (x>>3)]: |
2180 | 2107 QPs[(y>>4)*QPStride + (x>>4)]; |
2108 if(!isColor && (mode & LEVEL_FIX)) QP= (QP* (packedYScale &0xFFFF))>>8; | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2109 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2110 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2111 "movd %0, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2112 "packuswb %%mm7, %%mm7 \n\t" // 0, 0, 0, QP, 0, 0, 0, QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2113 "packuswb %%mm7, %%mm7 \n\t" // 0,QP, 0, QP, 0,QP, 0, QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2114 "packuswb %%mm7, %%mm7 \n\t" // QP,..., QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2115 "movq %%mm7, pQPb \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2116 : : "r" (QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2117 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2118 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2119 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2120 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2121 if(y + 12 < height) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2122 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2123 #ifdef MORE_TIMEING |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2124 T0= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2125 #endif |
2159 | 2126 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2127 #ifdef HAVE_MMX2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2128 prefetchnta(vertSrcBlock + (((x>>3)&3) + 2)*srcStride + 32); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2129 prefetchnta(vertSrcBlock + (((x>>3)&3) + 6)*srcStride + 32); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2130 prefetcht0(vertBlock + (((x>>3)&3) + 2)*dstStride + 32); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2131 prefetcht0(vertBlock + (((x>>3)&3) + 6)*dstStride + 32); |
2159 | 2132 #elif defined(HAVE_3DNOW) |
2133 //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ... | |
2134 /* prefetch(vertSrcBlock + (((x>>3)&3) + 2)*srcStride + 32); | |
2135 prefetch(vertSrcBlock + (((x>>3)&3) + 6)*srcStride + 32); | |
2136 prefetchw(vertBlock + (((x>>3)&3) + 2)*dstStride + 32); | |
2137 prefetchw(vertBlock + (((x>>3)&3) + 6)*dstStride + 32); | |
2138 */ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2139 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2140 if(!isColor) yHistogram[ srcBlock[0] ]++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2141 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2142 blockCopy(vertBlock + dstStride*2, dstStride, |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2143 vertSrcBlock + srcStride*2, srcStride, 8, mode & LEVEL_FIX); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2144 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2145 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2146 #ifdef MORE_TIMEING |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2147 T1= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2148 memcpyTime+= T1-T0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2149 T0=T1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2150 #endif |
2159 | 2151 if(mode & V_DEBLOCK) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2152 { |
2179 | 2153 if(mode & V_RK1_FILTER) |
2154 vertRK1Filter(vertBlock, stride, QP); | |
2155 else if(mode & V_X1_FILTER) | |
2159 | 2156 vertX1Filter(vertBlock, stride, QP); |
2157 else | |
2158 { | |
2159 if( isVertDC(vertBlock, stride)) | |
2160 { | |
2161 if(isVertMinMaxOk(vertBlock, stride, QP)) | |
2162 doVertLowPass(vertBlock, stride, QP); | |
2163 } | |
2164 else | |
2165 doVertDefFilter(vertBlock, stride, QP); | |
2166 } | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2167 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2168 #ifdef MORE_TIMEING |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2169 T1= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2170 vertTime+= T1-T0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2171 T0=T1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2172 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2173 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2174 else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2175 blockCopy(vertBlock + dstStride*1, dstStride, |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2176 vertSrcBlock + srcStride*1, srcStride, 4, mode & LEVEL_FIX); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2177 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2178 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2179 if(x - 8 >= 0 && x<width) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2180 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2181 #ifdef MORE_TIMEING |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2182 T0= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2183 #endif |
2159 | 2184 if(mode & H_DEBLOCK) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2185 { |
2179 | 2186 if(mode & H_X1_FILTER) |
2187 horizX1Filter(dstBlock-4, stride, QP); | |
2188 else | |
2159 | 2189 { |
2179 | 2190 if( isHorizDCAndCopy2Temp(dstBlock-4, stride)) |
2191 { | |
2192 if(isHorizMinMaxOk(tempBlock, TEMP_STRIDE, QP)) | |
2193 doHorizLowPassAndCopyBack(dstBlock-4, stride, QP); | |
2194 } | |
2195 else | |
2196 doHorizDefFilterAndCopyBack(dstBlock-4, stride, QP); | |
2159 | 2197 } |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2198 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2199 #ifdef MORE_TIMEING |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2200 T1= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2201 horizTime+= T1-T0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2202 T0=T1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2203 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2204 dering(dstBlock - 9 - stride, stride, QP); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2205 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2206 else if(y!=0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2207 dering(dstBlock - stride*9 + width-9, stride, QP); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2208 //FIXME dering filter will not be applied to last block (bottom right) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2209 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2210 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2211 dstBlock+=8; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2212 srcBlock+=8; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2213 vertBlock+=8; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2214 vertSrcBlock+=8; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2215 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2216 } |
2159 | 2217 #ifdef HAVE_3DNOW |
2218 asm volatile("femms"); | |
2219 #elif defined (HAVE_MMX) | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2220 asm volatile("emms"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2221 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2222 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2223 #ifdef TIMEING |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2224 // FIXME diff is mostly the time spent for rdtsc (should subtract that but ...) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2225 sumTime= rdtsc() - sumTime; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2226 if(!isColor) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2227 printf("cpy:%4dk, vert:%4dk, horiz:%4dk, sum:%4dk, diff:%4dk, color: %d/%d \r", |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2228 (int)(memcpyTime/1000), (int)(vertTime/1000), (int)(horizTime/1000), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2229 (int)(sumTime/1000), (int)((sumTime-memcpyTime-vertTime-horizTime)/1000) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2230 , black, white); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2231 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2232 } |
2159 | 2233 |
2234 |