Mercurial > mplayer.hg
annotate postproc/postprocess_template.c @ 2207:726b7877f1ac
-dumpmpsub added
-subcp added (correct me if wrong/missing sth)
author | gabucino |
---|---|
date | Mon, 15 Oct 2001 10:45:02 +0000 |
parents | f90b6e259dc8 |
children | 9fd911c931cd |
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 /* |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
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 |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
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 |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
27 doHorizLowPass E a a |
2159 | 28 doHorizDefFilter E ac ac |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
29 deRing |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
30 Vertical RKAlgo1 E a a |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
31 Vertical X1 a E E |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
32 Horizontal X1 a E E |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
33 LinIpolDeinterlace a E E* |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
34 LinBlendDeinterlace a E E* |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
35 MedianDeinterlace a E |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
36 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
37 |
2159 | 38 * i dont have a 3dnow CPU -> its untested |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
39 E = Exact implementation |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
40 e = allmost exact implementation |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
41 a = alternative / approximate impl |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
42 c = checked against the other implementations (-vo md5) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
43 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
44 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
45 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
46 TODO: |
2179 | 47 verify that everything workes as it should (how?) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
48 reduce the time wasted on the mem transfer |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
49 implement dering |
2159 | 50 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
|
51 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
|
52 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
|
53 move YScale thing to the end instead of fixing QP |
2159 | 54 write a faster and higher quality deblocking filter :) |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
55 do something about the speed of the horizontal filters |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
56 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
|
57 (the if/else stuff per block is slowing things down) |
2179 | 58 compare the quality & speed of all filters |
59 implement a few simple deinterlacing filters | |
60 split this huge file | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
61 fix warnings (unused vars, ...) |
2158
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 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
64 Notes: |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
65 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
66 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
67 |
2159 | 68 /* |
2179 | 69 Changelog: use the CVS log |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
70 rewrote the horizontal lowpass filter to fix a bug which caused a blocky look |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
71 added deinterlace filters (linear interpolate, linear blend, median) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
72 minor cleanups (removed some outcommented stuff) |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
73 0.1.3 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
74 bugfixes: last 3 lines not brightness/contrast corrected |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
75 brightness statistics messed up with initial black pic |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
76 changed initial values of the brightness statistics |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
77 C++ -> C conversation |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
78 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
|
79 new experimental vertical deblocking filter |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
80 RK filter has 3dNow support now (untested) |
2159 | 81 0.1.2 |
82 fixed a bug in the horizontal default filter | |
83 3dnow version of the Horizontal & Vertical Lowpass filters | |
84 mmx version of the Horizontal Default filter | |
85 mmx2 & C versions of a simple filter described in a paper from ramkishor & karandikar | |
86 added mode flags & quality2mode function | |
87 0.1.1 | |
88 */ | |
89 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
90 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
91 #include <inttypes.h> |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
92 #include <stdio.h> |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
93 #include <stdlib.h> |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
94 #include "../config.h" |
2159 | 95 //#undef HAVE_MMX2 |
96 //#define HAVE_3DNOW | |
97 //#undef HAVE_MMX | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
98 #include "postprocess.h" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
99 |
2189 | 100 #define MIN(a,b) ((a) > (b) ? (b) : (a)) |
101 #define MAX(a,b) ((a) < (b) ? (b) : (a)) | |
102 #define ABS(a) ((a) > 0 ? (a) : (-(a))) | |
103 #define SIGN(a) ((a) > 0 ? 1 : -1) | |
104 | |
105 #ifdef HAVE_MMX2 | |
106 #define PAVGB(a,b) "pavgb " #a ", " #b " \n\t" | |
107 #elif defined (HAVE_3DNOW) | |
108 #define PAVGB(a,b) "pavgusb " #a ", " #b " \n\t" | |
109 #endif | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
110 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
111 static uint64_t packedYOffset= 0x0000000000000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
112 static uint64_t packedYScale= 0x0100010001000100LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
113 static uint64_t w05= 0x0005000500050005LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
114 static uint64_t w20= 0x0020002000200020LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
115 static uint64_t w1400= 0x1400140014001400LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
116 static uint64_t bm00000001= 0x00000000000000FFLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
117 static uint64_t bm00010000= 0x000000FF00000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
118 static uint64_t bm00001000= 0x00000000FF000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
119 static uint64_t bm10000000= 0xFF00000000000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
120 static uint64_t bm10000001= 0xFF000000000000FFLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
121 static uint64_t bm11000011= 0xFFFF00000000FFFFLL; |
2159 | 122 static uint64_t bm00000011= 0x000000000000FFFFLL; |
2179 | 123 static uint64_t bm11111110= 0xFFFFFFFFFFFFFF00LL; |
2159 | 124 static uint64_t bm11000000= 0xFFFF000000000000LL; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
125 static uint64_t bm00011000= 0x000000FFFF000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
126 static uint64_t bm00110011= 0x0000FFFF0000FFFFLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
127 static uint64_t bm11001100= 0xFFFF0000FFFF0000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
128 static uint64_t b00= 0x0000000000000000LL; |
2179 | 129 static uint64_t b01= 0x0101010101010101LL; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
130 static uint64_t b02= 0x0202020202020202LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
131 static uint64_t b0F= 0x0F0F0F0F0F0F0F0FLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
132 static uint64_t bFF= 0xFFFFFFFFFFFFFFFFLL; |
2159 | 133 static uint64_t b20= 0x2020202020202020LL; |
134 static uint64_t b80= 0x8080808080808080LL; | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
135 static uint64_t b7E= 0x7E7E7E7E7E7E7E7ELL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
136 static uint64_t b7C= 0x7C7C7C7C7C7C7C7CLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
137 static uint64_t b3F= 0x3F3F3F3F3F3F3F3FLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
138 static uint64_t temp0=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
139 static uint64_t temp1=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
140 static uint64_t temp2=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
141 static uint64_t temp3=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
142 static uint64_t temp4=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
143 static uint64_t temp5=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
144 static uint64_t pQPb=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
145 static uint8_t tempBlock[16*16]; |
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 int hFlatnessThreshold= 56 - 16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
148 int vFlatnessThreshold= 56 - 16; |
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 //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
|
151 double maxClippedThreshold= 0.01; |
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 int maxAllowedY=255; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
154 //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
|
155 int minAllowedY=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
156 |
2185 | 157 #ifdef TIMEING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
158 static inline long long rdtsc() |
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 long long l; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
161 asm volatile( "rdtsc\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
162 : "=A" (l) |
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 // printf("%d\n", int(l/1000)); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
165 return l; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
166 } |
2185 | 167 #endif |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
168 |
2185 | 169 #ifdef HAVE_MMX2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
170 static inline void prefetchnta(void *p) |
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 asm volatile( "prefetchnta (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
173 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
174 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
175 } |
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 static inline void prefetcht0(void *p) |
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 asm volatile( "prefetcht0 (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
180 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
181 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
182 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
183 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
184 static inline void prefetcht1(void *p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
185 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
186 asm volatile( "prefetcht1 (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
187 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
188 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
189 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
190 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
191 static inline void prefetcht2(void *p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
192 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
193 asm volatile( "prefetcht2 (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
194 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
195 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
196 } |
2185 | 197 #endif |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
198 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
199 //FIXME? |255-0| = 1 (shouldnt be a problem ...) |
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 * 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
|
202 */ |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
203 static inline int isVertDC(uint8_t src[], int stride){ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
204 int numEq= 0; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
205 int y; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
206 src+= stride; // src points to begin of the 8x8 Block |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
207 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
208 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
209 "pushl %1\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
210 "movq b7E, %%mm7 \n\t" // mm7 = 0x7F |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
211 "movq b7C, %%mm6 \n\t" // mm6 = 0x7D |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
212 "movq (%1), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
213 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
214 "movq (%1), %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
215 "psubb %%mm1, %%mm0 \n\t" // mm0 = differnece |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
216 "paddb %%mm7, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
217 "pcmpgtb %%mm6, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
218 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
219 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
220 "movq (%1), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
221 "psubb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
222 "paddb %%mm7, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
223 "pcmpgtb %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
224 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
225 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
226 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
227 "movq (%1), %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
228 "psubb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
229 "paddb %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
230 "pcmpgtb %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
231 "paddb %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
232 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
233 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
234 "movq (%1), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
235 "psubb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
236 "paddb %%mm7, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
237 "pcmpgtb %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
238 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
239 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
240 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
241 "movq (%1), %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
242 "psubb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
243 "paddb %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
244 "pcmpgtb %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
245 "paddb %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
246 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
247 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
248 "movq (%1), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
249 "psubb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
250 "paddb %%mm7, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
251 "pcmpgtb %%mm6, %%mm1 \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 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
254 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
255 "movq (%1), %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
256 "psubb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
257 "paddb %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
258 "pcmpgtb %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
259 "paddb %%mm2, %%mm0 \n\t" |
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 " \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
262 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
263 "psrlw $8, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
264 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
265 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
266 "psrlq $16, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
267 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
268 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
269 "psrlq $32, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
270 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
271 "popl %1\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
272 "movd %%mm0, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
273 : "=r" (numEq) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
274 : "r" (src), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
275 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
276 // printf("%d\n", numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
277 numEq= (256 - (numEq & 0xFF)) &0xFF; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
278 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
279 // int asmEq= numEq; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
280 // numEq=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
281 // uint8_t *temp= src; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
282 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
283 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
284 for(y=0; y<BLOCK_SIZE-1; y++) |
2158
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 if(((src[0] - src[0+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
287 if(((src[1] - src[1+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
288 if(((src[2] - src[2+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
289 if(((src[3] - src[3+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
290 if(((src[4] - src[4+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
291 if(((src[5] - src[5+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
292 if(((src[6] - src[6+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
293 if(((src[7] - src[7+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
294 src+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
295 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
296 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
297 /* if(abs(numEq - asmEq) > 0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
298 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
299 printf("\nasm:%d c:%d\n", asmEq, numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
300 for(int y=0; y<8; y++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
301 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
302 for(int x=0; x<8; x++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
303 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
304 printf("%d ", temp[x + y*stride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
305 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
306 printf("\n"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
307 } |
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 */ |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
310 // 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
|
311 return (numEq > vFlatnessThreshold) ? 1 : 0; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
312 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
313 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
314 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
|
315 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
316 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
317 int isOk; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
318 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
319 // "int $3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
320 "movq (%1, %2), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
321 "movq (%1, %2, 8), %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
322 "movq %%mm0, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
323 "psubusb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
324 "psubusb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
325 "por %%mm1, %%mm0 \n\t" // ABS Diff |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
326 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
327 "movq pQPb, %%mm7 \n\t" // QP,..., QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
328 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
329 "psubusb %%mm7, %%mm0 \n\t" // Diff <= 2QP -> 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
330 "pcmpeqd b00, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
331 "psrlq $16, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
332 "pcmpeqd bFF, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
333 // "movd %%mm0, (%1, %2, 4)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
334 "movd %%mm0, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
335 : "=r" (isOk) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
336 : "r" (src), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
337 ); |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
338 return isOk ? 1 : 0; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
339 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
340 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
341 int isOk2= 1; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
342 int x; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
343 for(x=0; x<BLOCK_SIZE; x++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
344 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
345 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
|
346 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
347 /* if(isOk && !isOk2 || !isOk && isOk2) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
348 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
349 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
|
350 for(int y=0; y<9; y++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
351 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
352 for(int x=0; x<8; x++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
353 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
354 printf("%d ", src[x + y*stride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
355 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
356 printf("\n"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
357 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
358 } */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
359 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
360 return isOk2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
361 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
362 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
363 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
364 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
365 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
366 * 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
|
367 * 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
|
368 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
369 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
|
370 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
371 // QP= 64; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
372 |
2159 | 373 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
374 //#ifdef HAVE_MMX2 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
375 asm volatile( //"movv %0 %1 %2\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
376 "pushl %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
377 "movq pQPb, %%mm0 \n\t" // QP,..., QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
378 // "movq bFF , %%mm0 \n\t" // QP,..., QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
379 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
380 "movq (%0), %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
381 "movq (%0, %1), %%mm5 \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 %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
384 "psubusb %%mm6, %%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, %%mm6 \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, %%mm6 \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 "movq (%0, %1, 8), %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
395 "leal (%0, %1, 4), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
396 "leal (%0, %1, 8), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
397 "subl %1, %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
398 "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
|
399 "movq (%0, %1, 8), %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
400 "movq %%mm5, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
401 "movq %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
402 "psubusb %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
403 "psubusb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
404 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
405 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
406 "pcmpeqb b00, %%mm2 \n\t" // diff <= QP -> FF |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
407 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
408 "pand %%mm2, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
409 "pandn %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
410 "por %%mm2, %%mm7 \n\t" // First Line to Filter |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
411 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
412 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
413 // 1 2 3 4 5 6 7 8 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
414 // %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
|
415 // 6 4 2 2 1 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
416 // 6 4 4 2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
417 // 6 8 2 |
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 %%mm6, %%mm2 \n\t" //1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
420 "movq %%mm6, %%mm3 \n\t" //1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
421 "paddusb b02, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
422 "psrlw $2, %%mm3 \n\t" //1 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
423 "pand b3F, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
424 "psubb %%mm3, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
425 "movq (%0, %1), %%mm0 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
426 "movq %%mm0, %%mm1 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
427 "paddusb b02, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
428 "psrlw $2, %%mm0 \n\t" // 1 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
429 "pand b3F, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
430 "paddusb %%mm2, %%mm0 \n\t" //3 1 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
431 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
432 "movq (%0, %1), %%mm0 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
433 "movq %%mm0, %%mm1 \n\t" // 1 |
2159 | 434 PAVGB(%%mm6, %%mm0) //1 1 /2 |
435 PAVGB(%%mm6, %%mm0) //3 1 /4 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
436 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
437 "movq (%0, %1, 4), %%mm2 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
438 "movq %%mm2, %%mm5 \n\t" // 1 |
2159 | 439 PAVGB((%%eax), %%mm2) // 11 /2 |
440 PAVGB((%0, %1, 2), %%mm2) // 211 /4 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
441 "movq %%mm2, %%mm3 \n\t" // 211 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
442 "movq (%0), %%mm4 \n\t" // 1 |
2159 | 443 PAVGB(%%mm4, %%mm3) // 4 211 /8 |
444 PAVGB(%%mm0, %%mm3) //642211 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
445 "movq %%mm3, (%0) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
446 // 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
|
447 "movq %%mm1, %%mm0 \n\t" // 1 |
2159 | 448 PAVGB(%%mm6, %%mm0) //1 1 /2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
449 "movq %%mm4, %%mm3 \n\t" // 1 |
2159 | 450 PAVGB((%0,%1,2), %%mm3) // 1 1 /2 |
451 PAVGB((%%eax,%1,2), %%mm5) // 11 /2 | |
452 PAVGB((%%eax), %%mm5) // 211 /4 | |
453 PAVGB(%%mm5, %%mm3) // 2 2211 /8 | |
454 PAVGB(%%mm0, %%mm3) //4242211 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
455 "movq %%mm3, (%0,%1) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
456 // mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9 |
2159 | 457 PAVGB(%%mm4, %%mm6) //11 /2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
458 "movq (%%ebx), %%mm0 \n\t" // 1 |
2159 | 459 PAVGB((%%eax, %1, 2), %%mm0) // 11/2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
460 "movq %%mm0, %%mm3 \n\t" // 11/2 |
2159 | 461 PAVGB(%%mm1, %%mm0) // 2 11/4 |
462 PAVGB(%%mm6, %%mm0) //222 11/8 | |
463 PAVGB(%%mm2, %%mm0) //22242211/16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
464 "movq (%0, %1, 2), %%mm2 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
465 "movq %%mm0, (%0, %1, 2) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
466 // 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
|
467 "movq (%%eax, %1, 4), %%mm0 \n\t" // 1 |
2159 | 468 PAVGB((%%ebx), %%mm0) // 11 /2 |
469 PAVGB(%%mm0, %%mm6) //11 11 /4 | |
470 PAVGB(%%mm1, %%mm4) // 11 /2 | |
471 PAVGB(%%mm2, %%mm1) // 11 /2 | |
472 PAVGB(%%mm1, %%mm6) //1122 11 /8 | |
473 PAVGB(%%mm5, %%mm6) //112242211 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
474 "movq (%%eax), %%mm5 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
475 "movq %%mm6, (%%eax) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
476 // 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
|
477 "movq (%%eax, %1, 4), %%mm6 \n\t" // 1 |
2159 | 478 PAVGB(%%mm7, %%mm6) // 11 /2 |
479 PAVGB(%%mm4, %%mm6) // 11 11 /4 | |
480 PAVGB(%%mm3, %%mm6) // 11 2211 /8 | |
481 PAVGB(%%mm5, %%mm2) // 11 /2 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
482 "movq (%0, %1, 4), %%mm4 \n\t" // 1 |
2159 | 483 PAVGB(%%mm4, %%mm2) // 112 /4 |
484 PAVGB(%%mm2, %%mm6) // 112242211 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
485 "movq %%mm6, (%0, %1, 4) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
486 // mm0=7(11) mm1=2(11) mm2=3(112) mm3=6(11) mm4=5 mm5=4 mm7=9 |
2159 | 487 PAVGB(%%mm7, %%mm1) // 11 2 /4 |
488 PAVGB(%%mm4, %%mm5) // 11 /2 | |
489 PAVGB(%%mm5, %%mm0) // 11 11 /4 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
490 "movq (%%eax, %1, 2), %%mm6 \n\t" // 1 |
2159 | 491 PAVGB(%%mm6, %%mm1) // 11 4 2 /8 |
492 PAVGB(%%mm0, %%mm1) // 11224222 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
493 // "pxor %%mm1, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
494 "movq %%mm1, (%%eax, %1, 2) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
495 // mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9 |
2159 | 496 PAVGB((%%ebx), %%mm2) // 112 4 /8 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
497 "movq (%%eax, %1, 4), %%mm0 \n\t" // 1 |
2159 | 498 PAVGB(%%mm0, %%mm6) // 1 1 /2 |
499 PAVGB(%%mm7, %%mm6) // 1 12 /4 | |
500 PAVGB(%%mm2, %%mm6) // 1122424 /4 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
501 // "pxor %%mm6, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
502 "movq %%mm6, (%%ebx) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
503 // mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9 |
2159 | 504 PAVGB(%%mm7, %%mm5) // 11 2 /4 |
505 PAVGB(%%mm7, %%mm5) // 11 6 /8 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
506 |
2159 | 507 PAVGB(%%mm3, %%mm0) // 112 /4 |
508 PAVGB(%%mm0, %%mm5) // 112246 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
509 // "pxor %%mm5, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
510 // "movq pQPb, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
511 "movq %%mm5, (%%eax, %1, 4) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
512 "popl %0\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
513 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
514 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
515 : "r" (src), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
516 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
517 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
518 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
519 const int l1= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
520 const int l2= stride + l1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
521 const int l3= stride + l2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
522 const int l4= stride + l3; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
523 const int l5= stride + l4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
524 const int l6= stride + l5; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
525 const int l7= stride + l6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
526 const int l8= stride + l7; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
527 const int l9= stride + l8; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
528 int x; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
529 for(x=0; x<BLOCK_SIZE; x++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
530 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
531 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
|
532 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
|
533 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
534 int sums[9]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
535 sums[0] = first + src[l1]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
536 sums[1] = src[l1] + src[l2]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
537 sums[2] = src[l2] + src[l3]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
538 sums[3] = src[l3] + src[l4]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
539 sums[4] = src[l4] + src[l5]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
540 sums[5] = src[l5] + src[l6]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
541 sums[6] = src[l6] + src[l7]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
542 sums[7] = src[l7] + src[l8]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
543 sums[8] = src[l8] + last; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
544 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
545 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
|
546 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
|
547 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
|
548 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
|
549 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
|
550 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
|
551 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
|
552 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
|
553 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
554 src++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
555 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
556 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
557 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
558 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
559 |
2159 | 560 /** |
561 * Experimental implementation of the filter (Algorithm 1) described in a paper from Ramkishor & Karandikar | |
562 * values are correctly clipped (MMX2) | |
563 * values are wraparound (C) | |
564 * conclusion: its fast, but introduces ugly horizontal patterns if there is a continious gradient | |
565 0 8 16 24 | |
566 x = 8 | |
567 x/2 = 4 | |
568 x/8 = 1 | |
569 1 12 12 23 | |
570 */ | |
2179 | 571 static inline void vertRK1Filter(uint8_t *src, int stride, int QP) |
2159 | 572 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
573 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2159 | 574 // FIXME rounding |
575 asm volatile( | |
576 "pxor %%mm7, %%mm7 \n\t" // 0 | |
577 "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE | |
578 "leal (%0, %1), %%eax \n\t" | |
579 "leal (%%eax, %1, 4), %%ebx \n\t" | |
580 // 0 1 2 3 4 5 6 7 8 9 | |
581 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 | |
582 "movq pQPb, %%mm0 \n\t" // QP,..., QP | |
583 "movq %%mm0, %%mm1 \n\t" // QP,..., QP | |
584 "paddusb b02, %%mm0 \n\t" | |
585 "psrlw $2, %%mm0 \n\t" | |
586 "pand b3F, %%mm0 \n\t" // QP/4,..., QP/4 | |
587 "paddusb %%mm1, %%mm0 \n\t" // QP*1.25 ... | |
588 "movq (%0, %1, 4), %%mm2 \n\t" // line 4 | |
589 "movq (%%ebx), %%mm3 \n\t" // line 5 | |
590 "movq %%mm2, %%mm4 \n\t" // line 4 | |
591 "pcmpeqb %%mm5, %%mm5 \n\t" // -1 | |
592 "pxor %%mm2, %%mm5 \n\t" // -line 4 - 1 | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
593 PAVGB(%%mm3, %%mm5) |
2159 | 594 "paddb %%mm6, %%mm5 \n\t" // (l5-l4)/2 |
595 "psubusb %%mm3, %%mm4 \n\t" | |
596 "psubusb %%mm2, %%mm3 \n\t" | |
597 "por %%mm3, %%mm4 \n\t" // |l4 - l5| | |
598 "psubusb %%mm0, %%mm4 \n\t" | |
599 "pcmpeqb %%mm7, %%mm4 \n\t" | |
600 "pand %%mm4, %%mm5 \n\t" // d/2 | |
601 | |
602 // "paddb %%mm6, %%mm2 \n\t" // line 4 + 0x80 | |
603 "paddb %%mm5, %%mm2 \n\t" | |
604 // "psubb %%mm6, %%mm2 \n\t" | |
605 "movq %%mm2, (%0,%1, 4) \n\t" | |
606 | |
607 "movq (%%ebx), %%mm2 \n\t" | |
608 // "paddb %%mm6, %%mm2 \n\t" // line 5 + 0x80 | |
609 "psubb %%mm5, %%mm2 \n\t" | |
610 // "psubb %%mm6, %%mm2 \n\t" | |
611 "movq %%mm2, (%%ebx) \n\t" | |
612 | |
613 "paddb %%mm6, %%mm5 \n\t" | |
614 "psrlw $2, %%mm5 \n\t" | |
615 "pand b3F, %%mm5 \n\t" | |
616 "psubb b20, %%mm5 \n\t" // (l5-l4)/8 | |
617 | |
618 "movq (%%eax, %1, 2), %%mm2 \n\t" | |
619 "paddb %%mm6, %%mm2 \n\t" // line 3 + 0x80 | |
620 "paddsb %%mm5, %%mm2 \n\t" | |
621 "psubb %%mm6, %%mm2 \n\t" | |
622 "movq %%mm2, (%%eax, %1, 2) \n\t" | |
623 | |
624 "movq (%%ebx, %1), %%mm2 \n\t" | |
625 "paddb %%mm6, %%mm2 \n\t" // line 6 + 0x80 | |
626 "psubsb %%mm5, %%mm2 \n\t" | |
627 "psubb %%mm6, %%mm2 \n\t" | |
628 "movq %%mm2, (%%ebx, %1) \n\t" | |
629 | |
630 : | |
631 : "r" (src), "r" (stride) | |
632 : "%eax", "%ebx" | |
633 ); | |
634 #else | |
635 const int l1= stride; | |
636 const int l2= stride + l1; | |
637 const int l3= stride + l2; | |
638 const int l4= stride + l3; | |
639 const int l5= stride + l4; | |
640 const int l6= stride + l5; | |
641 const int l7= stride + l6; | |
642 const int l8= stride + l7; | |
643 const int l9= stride + l8; | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
644 int x; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
645 for(x=0; x<BLOCK_SIZE; x++) |
2159 | 646 { |
647 if(ABS(src[l4]-src[l5]) < QP + QP/4) | |
648 { | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
649 int v = (src[l5] - src[l4]); |
2159 | 650 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
651 src[l3] +=v/8; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
652 src[l4] +=v/2; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
653 src[l5] -=v/2; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
654 src[l6] -=v/8; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
655 |
2159 | 656 } |
657 src++; | |
658 } | |
659 | |
660 #endif | |
661 } | |
662 | |
663 /** | |
664 * Experimental Filter 1 | |
2179 | 665 * will not damage linear gradients |
666 * 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
|
667 * 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
|
668 * MMX2 version does correct clipping C version doesnt |
2159 | 669 */ |
670 static inline void vertX1Filter(uint8_t *src, int stride, int QP) | |
671 { | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
672 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2159 | 673 asm volatile( |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
674 "pxor %%mm7, %%mm7 \n\t" // 0 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
675 // "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
676 "leal (%0, %1), %%eax \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
677 "leal (%%eax, %1, 4), %%ebx \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
678 // 0 1 2 3 4 5 6 7 8 9 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
679 // %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
|
680 "movq (%%eax, %1, 2), %%mm0 \n\t" // line 3 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
681 "movq (%0, %1, 4), %%mm1 \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
682 "movq %%mm1, %%mm2 \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
683 "psubusb %%mm0, %%mm1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
684 "psubusb %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
685 "por %%mm1, %%mm0 \n\t" // |l2 - l3| |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
686 "movq (%%ebx), %%mm3 \n\t" // line 5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
687 "movq (%%ebx, %1), %%mm4 \n\t" // line 6 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
688 "movq %%mm3, %%mm5 \n\t" // line 5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
689 "psubusb %%mm4, %%mm3 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
690 "psubusb %%mm5, %%mm4 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
691 "por %%mm4, %%mm3 \n\t" // |l5 - l6| |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
692 PAVGB(%%mm3, %%mm0) // (|l2 - l3| + |l5 - l6|)/2 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
693 "movq %%mm2, %%mm1 \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
694 "psubusb %%mm5, %%mm2 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
695 "movq %%mm2, %%mm4 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
696 "pcmpeqb %%mm7, %%mm2 \n\t" // (l4 - l5) <= 0 ? -1 : 0 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
697 "psubusb %%mm1, %%mm5 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
698 "por %%mm5, %%mm4 \n\t" // |l4 - l5| |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
699 "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
|
700 "movq %%mm4, %%mm3 \n\t" // d |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
701 "psubusb pQPb, %%mm4 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
702 "pcmpeqb %%mm7, %%mm4 \n\t" // d <= QP ? -1 : 0 |
2179 | 703 "psubusb b01, %%mm3 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
704 "pand %%mm4, %%mm3 \n\t" // d <= QP ? d : 0 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
705 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
706 PAVGB(%%mm7, %%mm3) // d/2 |
2179 | 707 "movq %%mm3, %%mm1 \n\t" // d/2 |
708 PAVGB(%%mm7, %%mm3) // d/4 | |
709 PAVGB(%%mm1, %%mm3) // 3*d/8 | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
710 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
711 "movq (%0, %1, 4), %%mm0 \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
712 "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
|
713 "psubusb %%mm3, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
714 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
715 "movq %%mm0, (%0, %1, 4) \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
716 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
717 "movq (%%ebx), %%mm0 \n\t" // line 5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
718 "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
|
719 "paddusb %%mm3, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
720 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
721 "movq %%mm0, (%%ebx) \n\t" // line 5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
722 |
2179 | 723 PAVGB(%%mm7, %%mm1) // d/4 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
724 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
725 "movq (%%eax, %1, 2), %%mm0 \n\t" // line 3 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
726 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4 |
2179 | 727 "psubusb %%mm1, %%mm0 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
728 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
729 "movq %%mm0, (%%eax, %1, 2) \n\t" // line 3 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
730 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
731 "movq (%%ebx, %1), %%mm0 \n\t" // line 6 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
732 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5 |
2179 | 733 "paddusb %%mm1, %%mm0 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
734 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
735 "movq %%mm0, (%%ebx, %1) \n\t" // line 6 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
736 |
2179 | 737 PAVGB(%%mm7, %%mm1) // d/8 |
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 "movq (%%eax, %1), %%mm0 \n\t" // line 2 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
740 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l2-1 : l2 |
2179 | 741 "psubusb %%mm1, %%mm0 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
742 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
743 "movq %%mm0, (%%eax, %1) \n\t" // line 2 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
744 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
745 "movq (%%ebx, %1, 2), %%mm0 \n\t" // line 7 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
746 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l7-1 : l7 |
2179 | 747 "paddusb %%mm1, %%mm0 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
748 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
749 "movq %%mm0, (%%ebx, %1, 2) \n\t" // line 7 |
2159 | 750 |
751 : | |
752 : "r" (src), "r" (stride) | |
753 : "%eax", "%ebx" | |
754 ); | |
755 #else | |
2168
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 const int l1= stride; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
758 const int l2= stride + l1; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
759 const int l3= stride + l2; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
760 const int l4= stride + l3; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
761 const int l5= stride + l4; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
762 const int l6= stride + l5; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
763 const int l7= stride + l6; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
764 const int l8= stride + l7; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
765 const int l9= stride + l8; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
766 int x; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
767 for(x=0; x<BLOCK_SIZE; x++) |
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 int a= src[l3] - src[l4]; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
770 int b= src[l4] - src[l5]; |
2179 | 771 int c= src[l5] - src[l6]; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
772 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
773 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
|
774 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
775 if(d < QP) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
776 { |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
777 int v = d * SIGN(-b); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
778 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
779 src[l2] +=v/8; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
780 src[l3] +=v/4; |
2179 | 781 src[l4] +=3*v/8; |
782 src[l5] -=3*v/8; | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
783 src[l6] -=v/4; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
784 src[l7] -=v/8; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
785 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
786 } |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
787 src++; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
788 } |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
789 /* |
2159 | 790 const int l1= stride; |
791 const int l2= stride + l1; | |
792 const int l3= stride + l2; | |
793 const int l4= stride + l3; | |
794 const int l5= stride + l4; | |
795 const int l6= stride + l5; | |
796 const int l7= stride + l6; | |
797 const int l8= stride + l7; | |
798 const int l9= stride + l8; | |
799 for(int x=0; x<BLOCK_SIZE; x++) | |
800 { | |
801 int v2= src[l2]; | |
802 int v3= src[l3]; | |
803 int v4= src[l4]; | |
804 int v5= src[l5]; | |
805 int v6= src[l6]; | |
806 int v7= src[l7]; | |
807 | |
808 if(ABS(v4-v5)<QP && ABS(v4-v5) - (ABS(v3-v4) + ABS(v5-v6))>0 ) | |
809 { | |
810 src[l3] = (6*v2 + 4*v3 + 3*v4 + 2*v5 + v6 )/16; | |
811 src[l4] = (3*v2 + 3*v3 + 4*v4 + 3*v5 + 2*v6 + v7 )/16; | |
812 src[l5] = (1*v2 + 2*v3 + 3*v4 + 4*v5 + 3*v6 + 3*v7)/16; | |
813 src[l6] = ( 1*v3 + 2*v4 + 3*v5 + 4*v6 + 6*v7)/16; | |
814 } | |
815 src++; | |
816 } | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
817 */ |
2159 | 818 #endif |
819 } | |
820 | |
2179 | 821 /** |
822 * Experimental Filter 1 (Horizontal) | |
823 * will not damage linear gradients | |
824 * Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter | |
825 * can only smooth blocks at the expected locations (it cant smooth them if they did move) | |
826 * MMX2 version does correct clipping C version doesnt | |
827 * not identical with the vertical one | |
828 */ | |
829 static inline void horizX1Filter(uint8_t *src, int stride, int QP) | |
830 { | |
831 int y; | |
832 static uint64_t *lut= NULL; | |
833 if(lut==NULL) | |
834 { | |
835 int i; | |
836 lut= (uint64_t*)memalign(8, 256*8); | |
837 for(i=0; i<256; i++) | |
838 { | |
839 int v= i < 128 ? 2*i : 2*(i-256); | |
840 /* | |
841 //Simulate 112242211 9-Tap filter | |
842 uint64_t a= (v/16) & 0xFF; | |
843 uint64_t b= (v/8) & 0xFF; | |
844 uint64_t c= (v/4) & 0xFF; | |
845 uint64_t d= (3*v/8) & 0xFF; | |
846 */ | |
847 //Simulate piecewise linear interpolation | |
848 uint64_t a= (v/16) & 0xFF; | |
849 uint64_t b= (v*3/16) & 0xFF; | |
850 uint64_t c= (v*5/16) & 0xFF; | |
851 uint64_t d= (7*v/16) & 0xFF; | |
852 uint64_t A= (0x100 - a)&0xFF; | |
853 uint64_t B= (0x100 - b)&0xFF; | |
854 uint64_t C= (0x100 - c)&0xFF; | |
855 uint64_t D= (0x100 - c)&0xFF; | |
856 | |
857 lut[i] = (a<<56) | (b<<48) | (c<<40) | (d<<32) | | |
858 (D<<24) | (C<<16) | (B<<8) | (A); | |
859 //lut[i] = (v<<32) | (v<<24); | |
860 } | |
861 } | |
862 | |
863 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) | |
864 asm volatile( | |
865 "pxor %%mm7, %%mm7 \n\t" // 0 | |
866 // "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE | |
867 "leal (%0, %1), %%eax \n\t" | |
868 "leal (%%eax, %1, 4), %%ebx \n\t" | |
869 | |
870 "movq b80, %%mm6 \n\t" | |
2180 | 871 "movd pQPb, %%mm5 \n\t" // QP |
2179 | 872 "movq %%mm5, %%mm4 \n\t" |
873 "paddusb %%mm5, %%mm5 \n\t" // 2QP | |
874 "paddusb %%mm5, %%mm4 \n\t" // 3QP | |
875 "pxor %%mm5, %%mm5 \n\t" // 0 | |
876 "psubb %%mm4, %%mm5 \n\t" // -3QP | |
877 "por bm11111110, %%mm5 \n\t" // ...,FF,FF,-3QP | |
878 "psllq $24, %%mm5 \n\t" | |
879 | |
880 // 0 1 2 3 4 5 6 7 8 9 | |
881 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 | |
882 | |
883 #define HX1old(a) \ | |
884 "movd " #a ", %%mm0 \n\t"\ | |
885 "movd 4" #a ", %%mm1 \n\t"\ | |
886 "punpckldq %%mm1, %%mm0 \n\t"\ | |
887 "movq %%mm0, %%mm1 \n\t"\ | |
888 "movq %%mm0, %%mm2 \n\t"\ | |
889 "psrlq $8, %%mm1 \n\t"\ | |
890 "psubusb %%mm1, %%mm2 \n\t"\ | |
891 "psubusb %%mm0, %%mm1 \n\t"\ | |
892 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ | |
893 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
894 "pshufw $0x00, %%mm1, %%mm3 \n\t" /* p´5 = |p1 - p2| */\ | |
895 PAVGB(%%mm1, %%mm3) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\ | |
896 "psrlq $16, %%mm3 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\ | |
897 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\ | |
898 "paddb %%mm5, %%mm1 \n\t"\ | |
899 "psubusb %%mm5, %%mm1 \n\t"\ | |
900 PAVGB(%%mm7, %%mm1)\ | |
901 "pxor %%mm2, %%mm1 \n\t"\ | |
902 "psubb %%mm2, %%mm1 \n\t"\ | |
903 "psrlq $24, %%mm1 \n\t"\ | |
904 "movd %%mm1, %%ecx \n\t"\ | |
905 "paddb %%mm6, %%mm0 \n\t"\ | |
906 "paddsb (%3, %%ecx, 8), %%mm0 \n\t"\ | |
907 "paddb %%mm6, %%mm0 \n\t"\ | |
908 "movq %%mm0, " #a " \n\t"\ | |
909 | |
910 /* | |
911 HX1old((%0)) | |
912 HX1old((%%eax)) | |
913 HX1old((%%eax, %1)) | |
914 HX1old((%%eax, %1, 2)) | |
915 HX1old((%0, %1, 4)) | |
916 HX1old((%%ebx)) | |
917 HX1old((%%ebx, %1)) | |
918 HX1old((%%ebx, %1, 2)) | |
919 */ | |
920 | |
921 //FIXME add some comments, its unreadable ... | |
922 #define HX1b(a, c, b, d) \ | |
923 "movd " #a ", %%mm0 \n\t"\ | |
924 "movd 4" #a ", %%mm1 \n\t"\ | |
925 "punpckldq %%mm1, %%mm0 \n\t"\ | |
926 "movd " #b ", %%mm4 \n\t"\ | |
927 "movq %%mm0, %%mm1 \n\t"\ | |
928 "movq %%mm0, %%mm2 \n\t"\ | |
929 "psrlq $8, %%mm1 \n\t"\ | |
930 "movd 4" #b ", %%mm3 \n\t"\ | |
931 "psubusb %%mm1, %%mm2 \n\t"\ | |
932 "psubusb %%mm0, %%mm1 \n\t"\ | |
933 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ | |
934 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
935 "punpckldq %%mm3, %%mm4 \n\t"\ | |
936 "movq %%mm1, %%mm3 \n\t"\ | |
937 "psllq $32, %%mm3 \n\t" /* p´5 = |p1 - p2| */\ | |
938 PAVGB(%%mm1, %%mm3) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\ | |
939 "paddb %%mm6, %%mm0 \n\t"\ | |
940 "psrlq $16, %%mm3 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\ | |
941 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\ | |
942 "movq %%mm4, %%mm3 \n\t"\ | |
943 "paddb %%mm5, %%mm1 \n\t"\ | |
944 "psubusb %%mm5, %%mm1 \n\t"\ | |
945 "psrlq $8, %%mm3 \n\t"\ | |
946 PAVGB(%%mm7, %%mm1)\ | |
947 "pxor %%mm2, %%mm1 \n\t"\ | |
948 "psubb %%mm2, %%mm1 \n\t"\ | |
949 "movq %%mm4, %%mm2 \n\t"\ | |
950 "psrlq $24, %%mm1 \n\t"\ | |
951 "psubusb %%mm3, %%mm2 \n\t"\ | |
952 "movd %%mm1, %%ecx \n\t"\ | |
953 "psubusb %%mm4, %%mm3 \n\t"\ | |
2180 | 954 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\ |
2179 | 955 "por %%mm2, %%mm3 \n\t" /* p´x = |px - p(x+1)| */\ |
956 "paddb %%mm6, %%mm0 \n\t"\ | |
957 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
958 "movq %%mm3, %%mm1 \n\t"\ | |
959 "psllq $32, %%mm1 \n\t" /* p´5 = |p1 - p2| */\ | |
960 "movq %%mm0, " #a " \n\t"\ | |
961 PAVGB(%%mm3, %%mm1) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\ | |
962 "paddb %%mm6, %%mm4 \n\t"\ | |
963 "psrlq $16, %%mm1 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\ | |
964 "psubusb %%mm1, %%mm3 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\ | |
965 "paddb %%mm5, %%mm3 \n\t"\ | |
966 "psubusb %%mm5, %%mm3 \n\t"\ | |
967 PAVGB(%%mm7, %%mm3)\ | |
968 "pxor %%mm2, %%mm3 \n\t"\ | |
969 "psubb %%mm2, %%mm3 \n\t"\ | |
970 "psrlq $24, %%mm3 \n\t"\ | |
971 "movd " #c ", %%mm0 \n\t"\ | |
972 "movd 4" #c ", %%mm1 \n\t"\ | |
973 "punpckldq %%mm1, %%mm0 \n\t"\ | |
974 "paddb %%mm6, %%mm0 \n\t"\ | |
2180 | 975 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\ |
2179 | 976 "paddb %%mm6, %%mm0 \n\t"\ |
977 "movq %%mm0, " #c " \n\t"\ | |
978 "movd %%mm3, %%ecx \n\t"\ | |
979 "movd " #d ", %%mm0 \n\t"\ | |
2180 | 980 "paddsb (%2, %%ecx, 8), %%mm4 \n\t"\ |
2179 | 981 "movd 4" #d ", %%mm1 \n\t"\ |
982 "paddb %%mm6, %%mm4 \n\t"\ | |
983 "punpckldq %%mm1, %%mm0 \n\t"\ | |
984 "movq %%mm4, " #b " \n\t"\ | |
985 "paddb %%mm6, %%mm0 \n\t"\ | |
2180 | 986 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\ |
2179 | 987 "paddb %%mm6, %%mm0 \n\t"\ |
988 "movq %%mm0, " #d " \n\t"\ | |
989 | |
990 HX1b((%0),(%%eax),(%%eax, %1),(%%eax, %1, 2)) | |
991 HX1b((%0, %1, 4),(%%ebx),(%%ebx, %1),(%%ebx, %1, 2)) | |
992 | |
993 | |
994 : | |
2180 | 995 : "r" (src), "r" (stride), "r" (lut) |
2179 | 996 : "%eax", "%ebx", "%ecx" |
997 ); | |
998 #else | |
999 | |
1000 //FIXME (has little in common with the mmx2 version) | |
1001 for(y=0; y<BLOCK_SIZE; y++) | |
1002 { | |
1003 int a= src[1] - src[2]; | |
1004 int b= src[3] - src[4]; | |
1005 int c= src[5] - src[6]; | |
1006 | |
1007 int d= MAX(ABS(b) - (ABS(a) + ABS(c))/2, 0); | |
1008 | |
1009 if(d < QP) | |
1010 { | |
1011 int v = d * SIGN(-b); | |
1012 | |
1013 src[1] +=v/8; | |
1014 src[2] +=v/4; | |
1015 src[3] +=3*v/8; | |
1016 src[4] -=3*v/8; | |
1017 src[5] -=v/4; | |
1018 src[6] -=v/8; | |
1019 | |
1020 } | |
1021 src+=stride; | |
1022 } | |
1023 #endif | |
1024 } | |
1025 | |
2159 | 1026 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1027 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
|
1028 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1029 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1030 src+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1031 //FIXME try pmul for *5 stuff |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1032 // src[0]=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1033 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1034 "pxor %%mm7, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1035 "leal (%0, %1), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1036 "leal (%%eax, %1, 4), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1037 // 0 1 2 3 4 5 6 7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1038 // %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
|
1039 // %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
|
1040 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1041 "movq (%0), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1042 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1043 "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1044 "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1045 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1046 "movq (%%eax), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1047 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1048 "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1049 "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1050 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1051 "movq (%%eax, %1), %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1052 "movq %%mm4, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1053 "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1054 "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1055 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1056 "paddw %%mm0, %%mm0 \n\t" // 2L0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1057 "paddw %%mm1, %%mm1 \n\t" // 2H0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1058 "psubw %%mm4, %%mm2 \n\t" // L1 - L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1059 "psubw %%mm5, %%mm3 \n\t" // H1 - H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1060 "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1061 "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1062 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1063 "psllw $2, %%mm2 \n\t" // 4L1 - 4L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1064 "psllw $2, %%mm3 \n\t" // 4H1 - 4H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1065 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1066 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1067 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1068 "movq (%%eax, %1, 2), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1069 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1070 "punpcklbw %%mm7, %%mm2 \n\t" // L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1071 "punpckhbw %%mm7, %%mm3 \n\t" // H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1072 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1073 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1074 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1075 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1076 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1077 "movq %%mm0, temp0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1078 "movq %%mm1, temp1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1079 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1080 "movq (%0, %1, 4), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1081 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1082 "punpcklbw %%mm7, %%mm0 \n\t" // L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1083 "punpckhbw %%mm7, %%mm1 \n\t" // H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1084 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1085 "psubw %%mm0, %%mm2 \n\t" // L3 - L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1086 "psubw %%mm1, %%mm3 \n\t" // H3 - H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1087 "movq %%mm2, temp2 \n\t" // L3 - L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1088 "movq %%mm3, temp3 \n\t" // H3 - H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1089 "paddw %%mm4, %%mm4 \n\t" // 2L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1090 "paddw %%mm5, %%mm5 \n\t" // 2H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1091 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1092 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1093 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1094 "psllw $2, %%mm2 \n\t" // 4L3 - 4L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1095 "psllw $2, %%mm3 \n\t" // 4H3 - 4H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1096 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1097 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1098 //50 opcodes so far |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1099 "movq (%%ebx), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1100 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1101 "punpcklbw %%mm7, %%mm2 \n\t" // L5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1102 "punpckhbw %%mm7, %%mm3 \n\t" // H5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1103 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1104 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1105 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1106 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1107 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1108 "movq (%%ebx, %1), %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1109 "punpcklbw %%mm7, %%mm6 \n\t" // L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1110 "psubw %%mm6, %%mm2 \n\t" // L5 - L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1111 "movq (%%ebx, %1), %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1112 "punpckhbw %%mm7, %%mm6 \n\t" // H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1113 "psubw %%mm6, %%mm3 \n\t" // H5 - H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1114 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1115 "paddw %%mm0, %%mm0 \n\t" // 2L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1116 "paddw %%mm1, %%mm1 \n\t" // 2H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1117 "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1118 "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1119 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1120 "psllw $2, %%mm2 \n\t" // 4L5 - 4L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1121 "psllw $2, %%mm3 \n\t" // 4H5 - 4H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1122 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1123 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1124 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1125 "movq (%%ebx, %1, 2), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1126 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1127 "punpcklbw %%mm7, %%mm2 \n\t" // L7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1128 "punpckhbw %%mm7, %%mm3 \n\t" // H7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1129 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1130 "paddw %%mm2, %%mm2 \n\t" // 2L7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1131 "paddw %%mm3, %%mm3 \n\t" // 2H7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1132 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1133 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1134 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1135 "movq temp0, %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1136 "movq temp1, %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1137 //FIXME pxor, psubw, pmax for abs |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1138 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1139 "pcmpgtw %%mm0, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1140 "pxor %%mm6, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1141 "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1142 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1143 "pcmpgtw %%mm1, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1144 "pxor %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1145 "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1146 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1147 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1148 "pcmpgtw %%mm2, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1149 "pxor %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1150 "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1151 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1152 "pcmpgtw %%mm3, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1153 "pxor %%mm6, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1154 "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1155 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1156 #ifdef HAVE_MMX2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1157 "pminsw %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1158 "pminsw %%mm3, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1159 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1160 "movq %%mm0, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1161 "psubusw %%mm2, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1162 "psubw %%mm6, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1163 "movq %%mm1, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1164 "psubusw %%mm3, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1165 "psubw %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1166 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1167 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1168 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1169 "pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1170 "pxor %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1171 "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1172 "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1173 "pxor %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1174 "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1175 // 100 opcodes |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1176 "movd %2, %%mm2 \n\t" // QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1177 //"pcmpeqb %%mm2, %%mm2\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1178 "punpcklwd %%mm2, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1179 "punpcklwd %%mm2, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1180 "psllw $3, %%mm2 \n\t" // 8QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1181 "movq %%mm2, %%mm3 \n\t" // 8QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1182 "pcmpgtw %%mm4, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1183 "pcmpgtw %%mm5, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1184 "pand %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1185 "pand %%mm3, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1186 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1187 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1188 "psubusw %%mm0, %%mm4 \n\t" // hd |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1189 "psubusw %%mm1, %%mm5 \n\t" // ld |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1190 |
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 "movq w05, %%mm2 \n\t" // 5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1193 "pmullw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1194 "pmullw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1195 "movq w20, %%mm2 \n\t" // 32 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1196 "paddw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1197 "paddw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1198 "psrlw $6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1199 "psrlw $6, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1200 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1201 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1202 "movq w06, %%mm2 \n\t" // 6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1203 "paddw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1204 "paddw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1205 "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
|
1206 //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
|
1207 "pmulhw %%mm2, %%mm4 \n\t" // hd/13 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1208 "pmulhw %%mm2, %%mm5 \n\t" // ld/13 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1209 */ |
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 "movq temp2, %%mm0 \n\t" // L3 - L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1212 "movq temp3, %%mm1 \n\t" // H3 - H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1213 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1214 "pxor %%mm2, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1215 "pxor %%mm3, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1216 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1217 // FIXME rounding error |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1218 "psraw $1, %%mm0 \n\t" // (L3 - L4)/2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1219 "psraw $1, %%mm1 \n\t" // (H3 - H4)/2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1220 "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1221 "pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1222 "pxor %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1223 "pxor %%mm3, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1224 "psubw %%mm2, %%mm0 \n\t" // |L3-L4| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1225 "psubw %%mm3, %%mm1 \n\t" // |H3-H4| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1226 // "psrlw $1, %%mm0 \n\t" // |L3 - L4|/2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1227 // "psrlw $1, %%mm1 \n\t" // |H3 - H4|/2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1228 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1229 "pxor %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1230 "pxor %%mm7, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1231 "pand %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1232 "pand %%mm3, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1233 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1234 #ifdef HAVE_MMX2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1235 "pminsw %%mm0, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1236 "pminsw %%mm1, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1237 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1238 "movq %%mm4, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1239 "psubusw %%mm0, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1240 "psubw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1241 "movq %%mm5, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1242 "psubusw %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1243 "psubw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1244 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1245 "pxor %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1246 "pxor %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1247 "psubw %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1248 "psubw %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1249 "packsswb %%mm5, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1250 "movq (%%eax, %1, 2), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1251 "paddb %%mm4, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1252 "movq %%mm0, (%%eax, %1, 2) \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1253 "movq (%0, %1, 4), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1254 "psubb %%mm4, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1255 // "pxor %%mm0, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1256 "movq %%mm0, (%0, %1, 4) \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1257 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1258 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1259 : "r" (src), "r" (stride), "r" (QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1260 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1261 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1262 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1263 const int l1= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1264 const int l2= stride + l1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1265 const int l3= stride + l2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1266 const int l4= stride + l3; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1267 const int l5= stride + l4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1268 const int l6= stride + l5; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1269 const int l7= stride + l6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1270 const int l8= stride + l7; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1271 // const int l9= stride + l8; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1272 int x; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1273 for(x=0; x<BLOCK_SIZE; x++) |
2158
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 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
|
1276 if(ABS(middleEnergy) < 8*QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1277 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1278 const int q=(src[l4] - src[l5])/2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1279 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
|
1280 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
|
1281 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1282 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1283 d= MAX(d, 0); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1284 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1285 d= (5*d + 32) >> 6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1286 d*= SIGN(-middleEnergy); |
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 if(q>0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1289 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1290 d= d<0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1291 d= d>q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1292 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1293 else |
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 d= d>0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1296 d= d<q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1297 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1298 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1299 src[l4]-= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1300 src[l5]+= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1301 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1302 src++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1303 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1304 #endif |
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 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1307 //FIXME? |255-0| = 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1308 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1309 * 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
|
1310 */ |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1311 static inline int isHorizDCAndCopy2Temp(uint8_t src[], int stride) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1312 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1313 // src++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1314 int numEq= 0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1315 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1316 asm volatile ( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1317 // "int $3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1318 "pushl %1\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1319 "movq b7E, %%mm7 \n\t" // mm7 = 0x7F |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1320 "movq b7C, %%mm6 \n\t" // mm6 = 0x7D |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1321 "leal tempBlock, %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1322 "pxor %%mm0, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1323 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1324 #define HDC_CHECK_AND_CPY(i) \ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1325 "movq -4(%1), %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1326 "psrlq $32, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1327 "punpckldq 4(%1), %%mm2 \n\t" /* (%1) */\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1328 "movq %%mm2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1329 "psrlq $8, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1330 "psubb %%mm1, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1331 "paddb %%mm7, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1332 "pcmpgtb %%mm6, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1333 "paddb %%mm2, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1334 "movq %%mm1," #i "(%%eax) \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1335 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1336 HDC_CHECK_AND_CPY(0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1337 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1338 HDC_CHECK_AND_CPY(8) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1339 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1340 HDC_CHECK_AND_CPY(16) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1341 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1342 HDC_CHECK_AND_CPY(24) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1343 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1344 HDC_CHECK_AND_CPY(32) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1345 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1346 HDC_CHECK_AND_CPY(40) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1347 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1348 HDC_CHECK_AND_CPY(48) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1349 "addl %2, %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1350 HDC_CHECK_AND_CPY(56) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1351 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1352 "psllq $8, %%mm0 \n\t" // remove dummy value |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1353 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1354 "psrlw $8, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1355 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1356 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1357 "psrlq $16, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1358 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1359 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1360 "psrlq $32, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1361 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1362 "popl %1\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1363 "movd %%mm0, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1364 : "=r" (numEq) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1365 : "r" (src), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1366 : "%eax" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1367 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1368 // printf("%d\n", numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1369 numEq= (256 - (numEq & 0xFF)) &0xFF; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1370 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1371 int y; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1372 for(y=0; y<BLOCK_SIZE; y++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1373 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1374 if(((src[0] - src[1] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1375 if(((src[1] - src[2] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1376 if(((src[2] - src[3] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1377 if(((src[3] - src[4] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1378 if(((src[4] - src[5] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1379 if(((src[5] - src[6] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1380 if(((src[6] - src[7] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1381 tempBlock[0 + y*TEMP_STRIDE] = src[0]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1382 tempBlock[1 + y*TEMP_STRIDE] = src[1]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1383 tempBlock[2 + y*TEMP_STRIDE] = src[2]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1384 tempBlock[3 + y*TEMP_STRIDE] = src[3]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1385 tempBlock[4 + y*TEMP_STRIDE] = src[4]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1386 tempBlock[5 + y*TEMP_STRIDE] = src[5]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1387 tempBlock[6 + y*TEMP_STRIDE] = src[6]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1388 tempBlock[7 + y*TEMP_STRIDE] = src[7]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1389 src+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1390 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1391 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1392 /* if(abs(numEq - asmEq) > 0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1393 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1394 // printf("\nasm:%d c:%d\n", asmEq, numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1395 for(int y=0; y<8; y++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1396 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1397 for(int x=0; x<8; x++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1398 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1399 printf("%d ", src[x + y*stride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1400 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1401 printf("\n"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1402 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1403 } |
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 // printf("%d\n", numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1406 return numEq > hFlatnessThreshold; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1407 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1408 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1409 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
|
1410 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1411 #ifdef MMX_FIXME |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1412 FIXME |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1413 int isOk; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1414 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1415 // "int $3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1416 "movq (%1, %2), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1417 "movq (%1, %2, 8), %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1418 "movq %%mm0, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1419 "psubusb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1420 "psubusb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1421 "por %%mm1, %%mm0 \n\t" // ABS Diff |
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 "movq pQPb, %%mm7 \n\t" // QP,..., QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1424 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1425 "psubusb %%mm7, %%mm0 \n\t" // Diff <= 2QP -> 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1426 "pcmpeqd b00, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1427 "psrlq $16, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1428 "pcmpeqd bFF, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1429 // "movd %%mm0, (%1, %2, 4)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1430 "movd %%mm0, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1431 : "=r" (isOk) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1432 : "r" (src), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1433 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1434 return isOk; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1435 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1436 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
|
1437 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1438 return 1; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1439 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1440 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1441 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1442 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
|
1443 { |
2159 | 1444 #ifdef HAVE_MMX |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1445 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1446 "pushl %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1447 "pxor %%mm7, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1448 "movq bm00001000, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1449 "movd %2, %%mm5 \n\t" // QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1450 "movq %%mm5, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1451 "paddusb %%mm5, %%mm5 \n\t" // 2QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1452 "paddusb %%mm5, %%mm4 \n\t" // 3QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1453 "psllq $24, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1454 "pxor %%mm5, %%mm5 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1455 "psubb %%mm4, %%mm5 \n\t" // -QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1456 "leal tempBlock, %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1457 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1458 //FIXME? "unroll by 2" and mix |
2159 | 1459 #ifdef HAVE_MMX2 |
1460 #define HDF(i) \ | |
1461 "movq " #i "(%%eax), %%mm0 \n\t"\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1462 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1463 "movq %%mm0, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1464 "psrlq $8, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1465 "psubusb %%mm1, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1466 "psubusb %%mm0, %%mm1 \n\t"\ |
2159 | 1467 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ |
1468 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
1469 "pshufw $0x00, %%mm1, %%mm3 \n\t" /* p´5 = |p1 - p2| */\ | |
1470 "pminub %%mm1, %%mm3 \n\t" /* p´5 = min(|p2-p1|, |p6-p5|)*/\ | |
1471 "psrlq $16, %%mm3 \n\t" /* p´3 = min(|p2-p1|, |p6-p5|)*/\ | |
1472 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-min(|p1-p2|,|p5-p6|) */\ | |
1473 "paddb %%mm5, %%mm1 \n\t"\ | |
1474 "psubusb %%mm5, %%mm1 \n\t"\ | |
1475 "psrlw $2, %%mm1 \n\t"\ | |
1476 "pxor %%mm2, %%mm1 \n\t"\ | |
1477 "psubb %%mm2, %%mm1 \n\t"\ | |
1478 "pand %%mm6, %%mm1 \n\t"\ | |
1479 "psubb %%mm1, %%mm0 \n\t"\ | |
1480 "psllq $8, %%mm1 \n\t"\ | |
1481 "paddb %%mm1, %%mm0 \n\t"\ | |
1482 "movd %%mm0, (%0) \n\t"\ | |
1483 "psrlq $32, %%mm0 \n\t"\ | |
1484 "movd %%mm0, 4(%0) \n\t" | |
1485 #else | |
1486 #define HDF(i)\ | |
1487 "movq " #i "(%%eax), %%mm0 \n\t"\ | |
1488 "movq %%mm0, %%mm1 \n\t"\ | |
1489 "movq %%mm0, %%mm2 \n\t"\ | |
1490 "psrlq $8, %%mm1 \n\t"\ | |
1491 "psubusb %%mm1, %%mm2 \n\t"\ | |
1492 "psubusb %%mm0, %%mm1 \n\t"\ | |
1493 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ | |
1494 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
1495 "movq %%mm1, %%mm3 \n\t"\ | |
1496 "psllq $32, %%mm3 \n\t"\ | |
1497 "movq %%mm3, %%mm4 \n\t"\ | |
1498 "psubusb %%mm1, %%mm4 \n\t"\ | |
1499 "psubb %%mm4, %%mm3 \n\t"\ | |
1500 "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
|
1501 "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
|
1502 "paddb %%mm5, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1503 "psubusb %%mm5, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1504 "psrlw $2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1505 "pxor %%mm2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1506 "psubb %%mm2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1507 "pand %%mm6, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1508 "psubb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1509 "psllq $8, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1510 "paddb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1511 "movd %%mm0, (%0) \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1512 "psrlq $32, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1513 "movd %%mm0, 4(%0) \n\t" |
2159 | 1514 #endif |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1515 HDF(0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1516 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1517 HDF(8) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1518 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1519 HDF(16) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1520 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1521 HDF(24) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1522 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1523 HDF(32) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1524 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1525 HDF(40) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1526 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1527 HDF(48) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1528 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1529 HDF(56) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1530 "popl %0 \n\t" |
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 : "r" (dst), "r" (stride), "r" (QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1533 : "%eax" |
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 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1536 uint8_t *src= tempBlock; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1537 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1538 int y; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1539 for(y=0; y<BLOCK_SIZE; y++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1540 { |
2195 | 1541 const int middleEnergy= 5*(src[4] - src[5]) + 2*(src[2] - src[5]); |
1542 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1543 dst[0] = src[0]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1544 dst[1] = src[1]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1545 dst[2] = src[2]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1546 dst[3] = src[3]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1547 dst[4] = src[4]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1548 dst[5] = src[5]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1549 dst[6] = src[6]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1550 dst[7] = src[7]; |
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 if(ABS(middleEnergy) < 8*QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1553 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1554 const int q=(src[3] - src[4])/2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1555 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
|
1556 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
|
1557 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1558 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1559 d= MAX(d, 0); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1560 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1561 d= (5*d + 32) >> 6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1562 d*= SIGN(-middleEnergy); |
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 if(q>0) |
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 d= d<0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1567 d= d>q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1568 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1569 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1570 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1571 d= d>0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1572 d= d<q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1573 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1574 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1575 dst[3]-= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1576 dst[4]+= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1577 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1578 dst+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1579 src+= TEMP_STRIDE; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1580 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1581 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1582 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1583 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1584 /** |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1585 * Do a horizontal low pass filter on the 10x8 block (dst points to middle 8x8 Block) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1586 * useing the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 (C version) |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1587 * useing the 7-Tap Filter (2,2,2,4,2,2,2)/16 (MMX2/3DNOW version) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1588 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1589 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
|
1590 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1591 //return; |
2159 | 1592 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1593 asm volatile( //"movv %0 %1 %2\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1594 "pushl %0\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1595 "pxor %%mm7, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1596 "leal tempBlock, %%eax \n\t" |
2159 | 1597 /* |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1598 #define HLP1 "movq (%0), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1599 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1600 "psllq $8, %%mm0 \n\t"\ |
2159 | 1601 PAVGB(%%mm1, %%mm0)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1602 "psrlw $8, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1603 "pxor %%mm1, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1604 "packuswb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1605 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1606 "movq %%mm0, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1607 "psllq $32, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1608 "paddb %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1609 "psllq $16, %%mm2 \n\t"\ |
2159 | 1610 PAVGB(%%mm2, %%mm0)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1611 "movq %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1612 "pand bm11001100, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1613 "paddusb %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1614 "psrlq $8, %%mm3 \n\t"\ |
2159 | 1615 PAVGB(%%mm1, %%mm4)\ |
1616 PAVGB(%%mm3, %%mm2)\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1617 "psrlq $16, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1618 "punpcklbw %%mm2, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1619 "movq %%mm2, (%0) \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1620 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1621 #define HLP2 "movq (%0), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1622 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1623 "psllq $8, %%mm0 \n\t"\ |
2159 | 1624 PAVGB(%%mm1, %%mm0)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1625 "psrlw $8, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1626 "pxor %%mm1, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1627 "packuswb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1628 "movq %%mm0, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1629 "psllq $32, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1630 "psllq $16, %%mm2 \n\t"\ |
2159 | 1631 PAVGB(%%mm2, %%mm0)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1632 "movq %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1633 "pand bm11001100, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1634 "paddusb %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1635 "psrlq $8, %%mm3 \n\t"\ |
2159 | 1636 PAVGB(%%mm3, %%mm2)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1637 "psrlq $16, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1638 "punpcklbw %%mm2, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1639 "movq %%mm2, (%0) \n\t"\ |
2159 | 1640 */ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1641 // 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
|
1642 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1643 Implemented Exact 7-Tap |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1644 9421 A321 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1645 36421 64321 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1646 334321 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1647 1234321 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1648 1234321 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1649 123433 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1650 12463 12346 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1651 1249 123A |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1652 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1653 */ |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1654 |
2159 | 1655 #ifdef HAVE_MMX2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1656 #define HLP3(i) "movq " #i "(%%eax), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1657 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1658 "movq %%mm0, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1659 "movq %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1660 "movq %%mm0, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1661 "psllq $8, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1662 "psrlq $8, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1663 "pand bm00000001, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1664 "pand bm10000000, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1665 "por %%mm3, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1666 "por %%mm4, %%mm2 \n\t"\ |
2159 | 1667 PAVGB(%%mm2, %%mm1)\ |
1668 PAVGB(%%mm1, %%mm0)\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1669 \ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1670 "pshufw $0xF9, %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1671 "pshufw $0x90, %%mm0, %%mm4 \n\t"\ |
2159 | 1672 PAVGB(%%mm3, %%mm4)\ |
1673 PAVGB(%%mm4, %%mm0)\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1674 "movd %%mm0, (%0) \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1675 "psrlq $32, %%mm0 \n\t"\ |
2159 | 1676 "movd %%mm0, 4(%0) \n\t" |
1677 #else | |
1678 #define HLP3(i) "movq " #i "(%%eax), %%mm0 \n\t"\ | |
1679 "movq %%mm0, %%mm1 \n\t"\ | |
1680 "movq %%mm0, %%mm2 \n\t"\ | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1681 "movd -4(%0), %%mm3 \n\t" /*0001000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1682 "movd 8(%0), %%mm4 \n\t" /*0001000*/\ |
2159 | 1683 "psllq $8, %%mm1 \n\t"\ |
1684 "psrlq $8, %%mm2 \n\t"\ | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1685 "psrlq $24, %%mm3 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1686 "psllq $56, %%mm4 \n\t"\ |
2159 | 1687 "por %%mm3, %%mm1 \n\t"\ |
1688 "por %%mm4, %%mm2 \n\t"\ | |
1689 PAVGB(%%mm2, %%mm1)\ | |
1690 PAVGB(%%mm1, %%mm0)\ | |
1691 \ | |
1692 "movq %%mm0, %%mm3 \n\t"\ | |
1693 "movq %%mm0, %%mm4 \n\t"\ | |
1694 "movq %%mm0, %%mm5 \n\t"\ | |
1695 "psrlq $16, %%mm3 \n\t"\ | |
1696 "psllq $16, %%mm4 \n\t"\ | |
1697 "pand bm11000000, %%mm5 \n\t"\ | |
1698 "por %%mm5, %%mm3 \n\t"\ | |
1699 "movq %%mm0, %%mm5 \n\t"\ | |
1700 "pand bm00000011, %%mm5 \n\t"\ | |
1701 "por %%mm5, %%mm4 \n\t"\ | |
1702 PAVGB(%%mm3, %%mm4)\ | |
1703 PAVGB(%%mm4, %%mm0)\ | |
1704 "movd %%mm0, (%0) \n\t"\ | |
1705 "psrlq $32, %%mm0 \n\t"\ | |
1706 "movd %%mm0, 4(%0) \n\t" | |
1707 #endif | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1708 |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1709 /* uses the 7-Tap Filter: 1112111 */ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1710 #define NEW_HLP(i)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1711 "movq " #i "(%%eax), %%mm0 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1712 "movq %%mm0, %%mm1 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1713 "movq %%mm0, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1714 "movd -4(%0), %%mm3 \n\t" /*0001000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1715 "movd 8(%0), %%mm4 \n\t" /*0001000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1716 "psllq $8, %%mm1 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1717 "psrlq $8, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1718 "psrlq $24, %%mm3 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1719 "psllq $56, %%mm4 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1720 "por %%mm3, %%mm1 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1721 "por %%mm4, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1722 "movq %%mm1, %%mm5 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1723 PAVGB(%%mm2, %%mm1)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1724 PAVGB(%%mm1, %%mm0)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1725 "psllq $8, %%mm5 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1726 "psrlq $8, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1727 "por %%mm3, %%mm5 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1728 "por %%mm4, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1729 "movq %%mm5, %%mm1 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1730 PAVGB(%%mm2, %%mm5)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1731 "psllq $8, %%mm1 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1732 "psrlq $8, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1733 "por %%mm3, %%mm1 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1734 "por %%mm4, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1735 PAVGB(%%mm2, %%mm1)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1736 PAVGB(%%mm1, %%mm5)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1737 PAVGB(%%mm5, %%mm0)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1738 "movd %%mm0, (%0) \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1739 "psrlq $32, %%mm0 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1740 "movd %%mm0, 4(%0) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1741 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1742 /* uses the 9-Tap Filter: 112242211 */ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1743 #define NEW_HLP2(i)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1744 "movq " #i "(%%eax), %%mm0 \n\t" /*0001000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1745 "movq %%mm0, %%mm1 \n\t" /*0001000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1746 "movq %%mm0, %%mm2 \n\t" /*0001000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1747 "movd -4(%0), %%mm3 \n\t" /*0001000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1748 "movd 8(%0), %%mm4 \n\t" /*0001000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1749 "psllq $8, %%mm1 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1750 "psrlq $8, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1751 "psrlq $24, %%mm3 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1752 "psllq $56, %%mm4 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1753 "por %%mm3, %%mm1 \n\t" /*0010000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1754 "por %%mm4, %%mm2 \n\t" /*0000100*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1755 "movq %%mm1, %%mm5 \n\t" /*0010000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1756 PAVGB(%%mm2, %%mm1) /*0010100*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1757 PAVGB(%%mm1, %%mm0) /*0012100*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1758 "psllq $8, %%mm5 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1759 "psrlq $8, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1760 "por %%mm3, %%mm5 \n\t" /*0100000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1761 "por %%mm4, %%mm2 \n\t" /*0000010*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1762 "movq %%mm5, %%mm1 \n\t" /*0100000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1763 PAVGB(%%mm2, %%mm5) /*0100010*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1764 "psllq $8, %%mm1 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1765 "psrlq $8, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1766 "por %%mm3, %%mm1 \n\t" /*1000000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1767 "por %%mm4, %%mm2 \n\t" /*0000001*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1768 "movq %%mm1, %%mm6 \n\t" /*1000000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1769 PAVGB(%%mm2, %%mm1) /*1000001*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1770 "psllq $8, %%mm6 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1771 "psrlq $8, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1772 "por %%mm3, %%mm6 \n\t"/*100000000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1773 "por %%mm4, %%mm2 \n\t"/*000000001*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1774 PAVGB(%%mm2, %%mm6) /*100000001*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1775 PAVGB(%%mm6, %%mm1) /*110000011*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1776 PAVGB(%%mm1, %%mm5) /*112000211*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1777 PAVGB(%%mm5, %%mm0) /*112242211*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1778 "movd %%mm0, (%0) \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1779 "psrlq $32, %%mm0 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1780 "movd %%mm0, 4(%0) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1781 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1782 #define HLP(i) NEW_HLP(i) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1783 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1784 HLP(0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1785 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1786 HLP(8) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1787 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1788 HLP(16) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1789 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1790 HLP(24) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1791 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1792 HLP(32) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1793 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1794 HLP(40) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1795 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1796 HLP(48) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1797 "addl %1, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1798 HLP(56) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1799 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1800 "popl %0\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1801 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1802 : "r" (dst), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1803 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1804 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1805 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1806 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1807 uint8_t *temp= tempBlock; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1808 int y; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1809 for(y=0; y<BLOCK_SIZE; y++) |
2158
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 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
|
1812 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
|
1813 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1814 int sums[9]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1815 sums[0] = first + temp[0]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1816 sums[1] = temp[0] + temp[1]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1817 sums[2] = temp[1] + temp[2]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1818 sums[3] = temp[2] + temp[3]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1819 sums[4] = temp[3] + temp[4]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1820 sums[5] = temp[4] + temp[5]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1821 sums[6] = temp[5] + temp[6]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1822 sums[7] = temp[6] + temp[7]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1823 sums[8] = temp[7] + last; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1824 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1825 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
|
1826 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
|
1827 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
|
1828 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
|
1829 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
|
1830 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
|
1831 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
|
1832 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
|
1833 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1834 dst+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1835 temp+= TEMP_STRIDE; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1836 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1837 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1838 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1839 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1840 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1841 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
|
1842 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1843 //FIXME |
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 #ifdef HAVE_MMX2X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1846 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1847 "leal (%0, %1), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1848 "leal (%%eax, %1, 4), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1849 // 0 1 2 3 4 5 6 7 8 9 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1850 // %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
|
1851 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1852 "pcmpeq %%mm6, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1853 "pxor %%mm7, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1854 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1855 #define FIND_MIN_MAX(addr)\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1856 "movq (" #addr "), %%mm0, \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1857 "pminub %%mm0, %%mm6 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1858 "pmaxub %%mm0, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1859 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1860 FIND_MIN_MAX(%0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1861 FIND_MIN_MAX(%%eax) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1862 FIND_MIN_MAX(%%eax, %1) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1863 FIND_MIN_MAX(%%eax, %1, 2) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1864 FIND_MIN_MAX(%0, %1, 4) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1865 FIND_MIN_MAX(%%ebx) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1866 FIND_MIN_MAX(%%ebx, %1) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1867 FIND_MIN_MAX(%%ebx, %1, 2) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1868 FIND_MIN_MAX(%0, %1, 8) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1869 FIND_MIN_MAX(%%ebx, %1, 2) |
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 "movq %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1872 "psrlq $32, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1873 "pminub %%mm4, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1874 "movq %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1875 "psrlq $16, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1876 "pminub %%mm4, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1877 "movq %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1878 "psrlq $8, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1879 "pminub %%mm4, %%mm6 \n\t" // min of pixels |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1880 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1881 "movq %%mm7, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1882 "psrlq $32, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1883 "pmaxub %%mm4, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1884 "movq %%mm7, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1885 "psrlq $16, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1886 "pmaxub %%mm4, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1887 "movq %%mm7, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1888 "psrlq $8, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1889 "pmaxub %%mm4, %%mm7 \n\t" // max of pixels |
2159 | 1890 PAVGB(%%mm6, %%mm7) // (max + min)/2 |
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 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1893 : : "r" (src), "r" (stride), "r" (QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1894 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1895 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1896 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1897 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1898 //FIXME |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1899 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1900 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1901 |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1902 /** |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1903 * Deinterlaces the given block |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1904 * will be called for every 8x8 block, except the last row, and can read & write into an 8x16 block |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1905 */ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1906 static inline void deInterlaceInterpolateLinear(uint8_t src[], int stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1907 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1908 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1909 asm volatile( |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1910 "leal (%0, %1), %%eax \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1911 "leal (%%eax, %1, 4), %%ebx \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1912 // 0 1 2 3 4 5 6 7 8 9 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1913 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1914 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1915 "movq (%0), %%mm0 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1916 "movq (%%eax, %1), %%mm1 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1917 PAVGB(%%mm1, %%mm0)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1918 "movq %%mm0, (%%eax) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1919 "movq (%0, %1, 4), %%mm0 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1920 PAVGB(%%mm0, %%mm1)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1921 "movq %%mm1, (%%eax, %1, 2) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1922 "movq (%%ebx, %1), %%mm1 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1923 PAVGB(%%mm1, %%mm0)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1924 "movq %%mm0, (%%ebx) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1925 "movq (%0, %1, 8), %%mm0 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1926 PAVGB(%%mm0, %%mm1)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1927 "movq %%mm1, (%%ebx, %1, 2) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1928 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1929 : : "r" (src), "r" (stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1930 : "%eax", "%ebx" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1931 ); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1932 #else |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1933 int x; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1934 for(x=0; x<8; x++) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1935 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1936 src[stride] = (src[0] + src[stride*2])>>1; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1937 src[stride*3] = (src[stride*2] + src[stride*4])>>1; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1938 src[stride*5] = (src[stride*4] + src[stride*6])>>1; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1939 src[stride*7] = (src[stride*6] + src[stride*8])>>1; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1940 src++; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1941 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1942 #endif |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1943 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1944 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1945 /** |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1946 * Deinterlaces the given block |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1947 * will be called for every 8x8 block, in the last row, and can read & write into an 8x8 block |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1948 */ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1949 static inline void deInterlaceInterpolateLinearLastRow(uint8_t src[], int stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1950 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1951 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1952 asm volatile( |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1953 "leal (%0, %1), %%eax \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1954 "leal (%%eax, %1, 4), %%ebx \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1955 // 0 1 2 3 4 5 6 7 8 9 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1956 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1957 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1958 "movq (%0), %%mm0 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1959 "movq (%%eax, %1), %%mm1 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1960 PAVGB(%%mm1, %%mm0)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1961 "movq %%mm0, (%%eax) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1962 "movq (%0, %1, 4), %%mm0 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1963 PAVGB(%%mm0, %%mm1)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1964 "movq %%mm1, (%%eax, %1, 2) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1965 "movq (%%ebx, %1), %%mm1 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1966 PAVGB(%%mm1, %%mm0)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1967 "movq %%mm0, (%%ebx) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1968 "movq %%mm1, (%%ebx, %1, 2) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1969 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1970 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1971 : : "r" (src), "r" (stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1972 : "%eax", "%ebx" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1973 ); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1974 #else |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1975 int x; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1976 for(x=0; x<8; x++) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1977 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1978 src[stride] = (src[0] + src[stride*2])>>1; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1979 src[stride*3] = (src[stride*2] + src[stride*4])>>1; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1980 src[stride*5] = (src[stride*4] + src[stride*6])>>1; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1981 src[stride*7] = src[stride*6]; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1982 src++; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1983 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1984 #endif |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1985 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1986 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1987 /** |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1988 * Deinterlaces the given block |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1989 * will be called for every 8x8 block, except the last row, and can read & write into an 8x16 block |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1990 * will shift the image up by 1 line (FIXME if this is a problem) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1991 */ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1992 static inline void deInterlaceBlendLinear(uint8_t src[], int stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1993 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1994 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1995 asm volatile( |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1996 "leal (%0, %1), %%eax \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1997 "leal (%%eax, %1, 4), %%ebx \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1998 // 0 1 2 3 4 5 6 7 8 9 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1999 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2000 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2001 "movq (%0), %%mm0 \n\t" // L0 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2002 "movq (%%eax, %1), %%mm1 \n\t" // L2 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2003 PAVGB(%%mm1, %%mm0) // L0+L2 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2004 "movq (%%eax), %%mm2 \n\t" // L1 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2005 PAVGB(%%mm2, %%mm0) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2006 "movq %%mm0, (%0) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2007 "movq (%%eax, %1, 2), %%mm0 \n\t" // L3 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2008 PAVGB(%%mm0, %%mm2) // L1+L3 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2009 PAVGB(%%mm1, %%mm2) // 2L2 + L1 + L3 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2010 "movq %%mm2, (%%eax) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2011 "movq (%0, %1, 4), %%mm2 \n\t" // L4 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2012 PAVGB(%%mm2, %%mm1) // L2+L4 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2013 PAVGB(%%mm0, %%mm1) // 2L3 + L2 + L4 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2014 "movq %%mm1, (%%eax, %1) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2015 "movq (%%ebx), %%mm1 \n\t" // L5 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2016 PAVGB(%%mm1, %%mm0) // L3+L5 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2017 PAVGB(%%mm2, %%mm0) // 2L4 + L3 + L5 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2018 "movq %%mm0, (%%eax, %1, 2) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2019 "movq (%%ebx, %1), %%mm0 \n\t" // L6 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2020 PAVGB(%%mm0, %%mm2) // L4+L6 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2021 PAVGB(%%mm1, %%mm2) // 2L5 + L4 + L6 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2022 "movq %%mm2, (%0, %1, 4) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2023 "movq (%%ebx, %1, 2), %%mm2 \n\t" // L7 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2024 PAVGB(%%mm2, %%mm1) // L5+L7 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2025 PAVGB(%%mm0, %%mm1) // 2L6 + L5 + L7 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2026 "movq %%mm1, (%%ebx) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2027 "movq (%0, %1, 8), %%mm1 \n\t" // L8 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2028 PAVGB(%%mm1, %%mm0) // L6+L8 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2029 PAVGB(%%mm2, %%mm0) // 2L7 + L6 + L8 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2030 "movq %%mm0, (%%ebx, %1) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2031 "movq (%%ebx, %1, 4), %%mm0 \n\t" // L9 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2032 PAVGB(%%mm0, %%mm2) // L7+L9 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2033 PAVGB(%%mm1, %%mm2) // 2L8 + L7 + L9 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2034 "movq %%mm2, (%%ebx, %1, 2) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2035 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2036 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2037 : : "r" (src), "r" (stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2038 : "%eax", "%ebx" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2039 ); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2040 #else |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2041 int x; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2042 for(x=0; x<8; x++) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2043 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2044 src[0 ] = (src[0 ] + 2*src[stride ] + src[stride*2])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2045 src[stride ] = (src[stride ] + 2*src[stride*2] + src[stride*3])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2046 src[stride*2] = (src[stride*2] + 2*src[stride*3] + src[stride*4])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2047 src[stride*3] = (src[stride*3] + 2*src[stride*4] + src[stride*5])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2048 src[stride*4] = (src[stride*4] + 2*src[stride*5] + src[stride*6])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2049 src[stride*5] = (src[stride*5] + 2*src[stride*6] + src[stride*7])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2050 src[stride*6] = (src[stride*6] + 2*src[stride*7] + src[stride*8])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2051 src[stride*7] = (src[stride*7] + 2*src[stride*8] + src[stride*9])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2052 src++; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2053 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2054 #endif |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2055 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2056 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2057 /** |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2058 * Deinterlaces the given block |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2059 * will be called for every 8x8 block, in the last row, and can read & write into an 8x8 block |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2060 * will shift the image up by 1 line (FIXME if this is a problem) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2061 */ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2062 static inline void deInterlaceBlendLinearLastRow(uint8_t src[], int stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2063 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2064 #if defined (HAVE_MMSX2) || defined (HAVE_3DNOW) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2065 asm volatile( |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2066 "leal (%0, %1), %%eax \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2067 "leal (%%eax, %1, 4), %%ebx \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2068 // 0 1 2 3 4 5 6 7 8 9 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2069 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2070 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2071 "movq (%0), %%mm0 \n\t" // L0 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2072 "movq (%%eax, %1), %%mm1 \n\t" // L2 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2073 PAVGB(%%mm1, %%mm0) // L0+L2 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2074 "movq (%%eax), %%mm2 \n\t" // L1 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2075 PAVGB(%%mm2, %%mm0) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2076 "movq %%mm0, (%0) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2077 "movq (%%eax, %1, 2), %%mm0 \n\t" // L3 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2078 PAVGB(%%mm0, %%mm2) // L1+L3 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2079 PAVGB(%%mm1, %%mm2) // 2L2 + L1 + L3 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2080 "movq %%mm2, (%%eax) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2081 "movq (%0, %1, 4), %%mm2 \n\t" // L4 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2082 PAVGB(%%mm2, %%mm1) // L2+L4 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2083 PAVGB(%%mm0, %%mm1) // 2L3 + L2 + L4 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2084 "movq %%mm1, (%%eax, %1) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2085 "movq (%%ebx), %%mm1 \n\t" // L5 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2086 PAVGB(%%mm1, %%mm0) // L3+L5 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2087 PAVGB(%%mm2, %%mm0) // 2L4 + L3 + L5 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2088 "movq %%mm0, (%%eax, %1, 2) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2089 "movq (%%ebx, %1), %%mm0 \n\t" // L6 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2090 PAVGB(%%mm0, %%mm2) // L4+L6 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2091 PAVGB(%%mm1, %%mm2) // 2L5 + L4 + L6 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2092 "movq %%mm2, (%0, %1, 4) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2093 "movq (%%ebx, %1, 2), %%mm2 \n\t" // L7 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2094 PAVGB(%%mm2, %%mm1) // L5+L7 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2095 PAVGB(%%mm0, %%mm1) // 2L6 + L5 + L7 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2096 "movq %%mm1, (%%ebx) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2097 PAVGB(%%mm2, %%mm0) // L7 + L8 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2098 "movq %%mm0, (%%ebx, %1) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2099 "movq %%mm0, (%%ebx, %1, 2) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2100 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2101 : : "r" (src), "r" (stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2102 : "%eax", "%ebx" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2103 ); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2104 #else |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2105 int x; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2106 for(x=0; x<8; x++) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2107 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2108 src[0 ] = (src[0 ] + 2*src[stride ] + src[stride*2])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2109 src[stride ] = (src[stride ] + 2*src[stride*2] + src[stride*3])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2110 src[stride*2] = (src[stride*2] + 2*src[stride*3] + src[stride*4])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2111 src[stride*3] = (src[stride*3] + 2*src[stride*4] + src[stride*5])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2112 src[stride*4] = (src[stride*4] + 2*src[stride*5] + src[stride*6])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2113 src[stride*5] = (src[stride*5] + 2*src[stride*6] + src[stride*7])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2114 src[stride*6] = (src[stride*6] + src[stride*7])>>1; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2115 src[stride*7] = src[stride*6]; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2116 src++; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2117 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2118 #endif |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2119 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2120 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2121 /** |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2122 * Deinterlaces the given block |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2123 * will be called for every 8x8 block, except the last row, and can read & write into an 8x16 block |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2124 */ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2125 static inline void deInterlaceMedian(uint8_t src[], int stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2126 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2127 #if defined (HAVE_MMX2) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2128 asm volatile( |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2129 "leal (%0, %1), %%eax \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2130 "leal (%%eax, %1, 4), %%ebx \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2131 // 0 1 2 3 4 5 6 7 8 9 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2132 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2133 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2134 "movq (%0), %%mm0 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2135 "movq (%%eax, %1), %%mm2 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2136 "movq (%%eax), %%mm1 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2137 "movq %%mm0, %%mm3 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2138 "pmaxub %%mm1, %%mm0 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2139 "pminub %%mm3, %%mm1 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2140 "pmaxub %%mm2, %%mm1 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2141 "pminub %%mm1, %%mm0 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2142 "movq %%mm0, (%%eax) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2143 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2144 "movq (%0, %1, 4), %%mm0 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2145 "movq (%%eax, %1, 2), %%mm1 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2146 "movq %%mm2, %%mm3 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2147 "pmaxub %%mm1, %%mm2 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2148 "pminub %%mm3, %%mm1 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2149 "pmaxub %%mm0, %%mm1 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2150 "pminub %%mm1, %%mm2 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2151 "movq %%mm2, (%%eax, %1, 2) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2152 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2153 "movq (%%ebx), %%mm2 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2154 "movq (%%ebx, %1), %%mm1 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2155 "movq %%mm2, %%mm3 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2156 "pmaxub %%mm0, %%mm2 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2157 "pminub %%mm3, %%mm0 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2158 "pmaxub %%mm1, %%mm0 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2159 "pminub %%mm0, %%mm2 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2160 "movq %%mm2, (%%ebx) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2161 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2162 "movq (%%ebx, %1, 2), %%mm2 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2163 "movq (%0, %1, 8), %%mm0 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2164 "movq %%mm2, %%mm3 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2165 "pmaxub %%mm0, %%mm2 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2166 "pminub %%mm3, %%mm0 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2167 "pmaxub %%mm1, %%mm0 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2168 "pminub %%mm0, %%mm2 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2169 "movq %%mm2, (%%ebx, %1, 2) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2170 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2171 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2172 : : "r" (src), "r" (stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2173 : "%eax", "%ebx" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2174 ); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2175 #else |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2176 //FIXME |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2177 int x; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2178 for(x=0; x<8; x++) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2179 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2180 src[0 ] = (src[0 ] + 2*src[stride ] + src[stride*2])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2181 src[stride ] = (src[stride ] + 2*src[stride*2] + src[stride*3])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2182 src[stride*2] = (src[stride*2] + 2*src[stride*3] + src[stride*4])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2183 src[stride*3] = (src[stride*3] + 2*src[stride*4] + src[stride*5])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2184 src[stride*4] = (src[stride*4] + 2*src[stride*5] + src[stride*6])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2185 src[stride*5] = (src[stride*5] + 2*src[stride*6] + src[stride*7])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2186 src[stride*6] = (src[stride*6] + 2*src[stride*7] + src[stride*8])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2187 src[stride*7] = (src[stride*7] + 2*src[stride*8] + src[stride*9])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2188 src++; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2189 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2190 #endif |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2191 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2192 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2193 /** |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2194 * Deinterlaces the given block |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2195 * will be called for every 8x8 block, in the last row, and can read & write into an 8x8 block |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2196 * will shift the image up by 1 line (FIXME if this is a problem) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2197 */ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2198 static inline void deInterlaceMedianLastRow(uint8_t src[], int stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2199 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2200 #if defined (HAVE_MMX2) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2201 asm volatile( |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2202 "leal (%0, %1), %%eax \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2203 "leal (%%eax, %1, 4), %%ebx \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2204 // 0 1 2 3 4 5 6 7 8 9 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2205 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2206 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2207 "movq (%0), %%mm0 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2208 "movq (%%eax, %1), %%mm2 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2209 "movq (%%eax), %%mm1 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2210 "movq %%mm0, %%mm3 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2211 "pmaxub %%mm1, %%mm0 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2212 "pminub %%mm3, %%mm1 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2213 "pmaxub %%mm2, %%mm1 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2214 "pminub %%mm1, %%mm0 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2215 "movq %%mm0, (%%eax) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2216 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2217 "movq (%0, %1, 4), %%mm0 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2218 "movq (%%eax, %1, 2), %%mm1 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2219 "movq %%mm2, %%mm3 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2220 "pmaxub %%mm1, %%mm2 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2221 "pminub %%mm3, %%mm1 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2222 "pmaxub %%mm0, %%mm1 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2223 "pminub %%mm1, %%mm2 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2224 "movq %%mm2, (%%eax, %1, 2) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2225 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2226 "movq (%%ebx), %%mm2 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2227 "movq (%%ebx, %1), %%mm1 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2228 "movq %%mm2, %%mm3 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2229 "pmaxub %%mm0, %%mm2 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2230 "pminub %%mm3, %%mm0 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2231 "pmaxub %%mm1, %%mm0 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2232 "pminub %%mm0, %%mm2 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2233 "movq %%mm2, (%%ebx) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2234 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2235 "movq %%mm1, (%%ebx, %1, 2) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2236 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2237 : : "r" (src), "r" (stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2238 : "%eax", "%ebx" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2239 ); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2240 #else |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2241 //FIXME |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2242 int x; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2243 for(x=0; x<8; x++) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2244 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2245 src[0 ] = (src[0 ] + 2*src[stride ] + src[stride*2])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2246 src[stride ] = (src[stride ] + 2*src[stride*2] + src[stride*3])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2247 src[stride*2] = (src[stride*2] + 2*src[stride*3] + src[stride*4])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2248 src[stride*3] = (src[stride*3] + 2*src[stride*4] + src[stride*5])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2249 src[stride*4] = (src[stride*4] + 2*src[stride*5] + src[stride*6])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2250 src[stride*5] = (src[stride*5] + 2*src[stride*6] + src[stride*7])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2251 src[stride*6] = (src[stride*6] + src[stride*7])>>1; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2252 src[stride*7] = src[stride*6]; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2253 src++; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2254 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2255 #endif |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2256 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2257 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2258 |
2185 | 2259 #ifdef HAVE_ODIVX_POSTPROCESS |
2260 #include "../opendivx/postprocess.h" | |
2261 int use_old_pp=0; | |
2262 #endif | |
2159 | 2263 |
2185 | 2264 static void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height, |
2265 QP_STORE_T QPs[], int QPStride, int isColor, int mode); | |
2159 | 2266 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2267 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2268 * ... |
2159 | 2269 * the mode value is interpreted as a quality value if its negative, its range is then (-1 ... -63) |
2270 * -63 is best quality -1 is worst | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2271 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2272 void postprocess(unsigned char * src[], int src_stride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2273 unsigned char * dst[], int dst_stride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2274 int horizontal_size, int vertical_size, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2275 QP_STORE_T *QP_store, int QP_stride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2276 int mode) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2277 { |
2159 | 2278 |
2185 | 2279 #ifdef HAVE_ODIVX_POSTPROCESS |
2280 // Note: I could make this shit outside of this file, but it would mean one | |
2281 // more function call... | |
2282 if(use_old_pp){ | |
2283 odivx_postprocess(src,src_stride,dst,dst_stride,horizontal_size,vertical_size,QP_store,QP_stride,mode); | |
2284 return; | |
2285 } | |
2286 #endif | |
2287 | |
2288 // I'm calling this from dec_video.c:video_set_postprocess() | |
2289 // if(mode<0) mode= getModeForQuality(-mode); | |
2159 | 2290 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2291 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2292 long long T= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2293 for(int y=vertical_size-1; y>=0 ; y--) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2294 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
|
2295 // memcpy(dst[0], src[0],src_stride*vertical_size); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2296 printf("%4dk\r", (rdtsc()-T)/1000); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2297 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2298 return; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2299 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2300 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2301 long long T= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2302 while( (rdtsc() - T)/1000 < 4000); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2303 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2304 return; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2305 */ |
2159 | 2306 postProcess(src[0], src_stride, dst[0], dst_stride, |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2307 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
|
2308 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2309 horizontal_size >>= 1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2310 vertical_size >>= 1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2311 src_stride >>= 1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2312 dst_stride >>= 1; |
2179 | 2313 mode= ((mode&0xFF)>>4) | (mode&0xFFFFFF00); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2314 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2315 if(1) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2316 { |
2159 | 2317 postProcess(src[1], src_stride, dst[1], dst_stride, |
2179 | 2318 horizontal_size, vertical_size, QP_store, QP_stride, 1, mode); |
2159 | 2319 postProcess(src[2], src_stride, dst[2], dst_stride, |
2179 | 2320 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
|
2321 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2322 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2323 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2324 memcpy(dst[1], src[1], src_stride*horizontal_size); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2325 memcpy(dst[2], src[2], src_stride*horizontal_size); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2326 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2327 } |
2185 | 2328 |
2159 | 2329 /** |
2330 * gets the mode flags for a given quality (larger values mean slower but better postprocessing) | |
2185 | 2331 * 0 <= quality <= 6 |
2159 | 2332 */ |
2185 | 2333 int getPpModeForQuality(int quality){ |
2334 int modes[1+GET_PP_QUALITY_MAX]= { | |
2335 0, | |
2336 #if 1 | |
2337 // horizontal filters first | |
2338 LUM_H_DEBLOCK, | |
2339 LUM_H_DEBLOCK | LUM_V_DEBLOCK, | |
2340 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK, | |
2341 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK, | |
2342 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK | LUM_DERING, | |
2343 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK | LUM_DERING | CHROM_DERING | |
2344 #else | |
2345 // vertical filters first | |
2159 | 2346 LUM_V_DEBLOCK, |
2347 LUM_V_DEBLOCK | LUM_H_DEBLOCK, | |
2348 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK, | |
2349 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK, | |
2350 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING, | |
2351 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING | CHROM_DERING | |
2185 | 2352 #endif |
2353 }; | |
2159 | 2354 |
2185 | 2355 #ifdef HAVE_ODIVX_POSTPROCESS |
2356 int odivx_modes[1+GET_PP_QUALITY_MAX]= { | |
2357 0, | |
2358 PP_DEBLOCK_Y_H, | |
2359 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V, | |
2360 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H, | |
2361 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V, | |
2362 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V|PP_DERING_Y, | |
2363 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V|PP_DERING_Y|PP_DERING_C | |
2364 }; | |
2365 if(use_old_pp) return odivx_modes[quality]; | |
2366 #endif | |
2367 return modes[quality]; | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2368 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2369 |
2169 | 2370 //} // extern "C" |
2159 | 2371 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2372 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2373 * 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
|
2374 * numLines must be a multiple of 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2375 * levelFix == 0 -> dont touch the brighness & contrast |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2376 */ |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2377 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
|
2378 int numLines, int levelFix) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2379 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2380 int i; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2381 if(levelFix) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2382 { |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2383 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2384 asm volatile( |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2385 "movl %4, %%eax \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2386 "movl %%eax, temp0\n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2387 "pushl %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2388 "pushl %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2389 "leal (%2,%2), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2390 "leal (%3,%3), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2391 "movq packedYOffset, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2392 "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
|
2393 "pxor %%mm4, %%mm4 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2394 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2395 #define SCALED_CPY \ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2396 "movq (%0), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2397 "movq (%0,%2), %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2398 "psubusb %%mm2, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2399 "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
|
2400 "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
|
2401 "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
|
2402 "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
|
2403 "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
|
2404 "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
|
2405 "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
|
2406 "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
|
2407 "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
|
2408 "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
|
2409 "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
|
2410 "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
|
2411 "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
|
2412 "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
|
2413 "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
|
2414 "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
|
2415 "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
|
2416 "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
|
2417 "movq %%mm1, (%1, %3) \n\t"\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2418 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2419 "1: \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2420 SCALED_CPY |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2421 "addl %%eax, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2422 "addl %%ebx, %1 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2423 SCALED_CPY |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2424 "addl %%eax, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2425 "addl %%ebx, %1 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2426 "decl temp0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2427 "jnz 1b \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2428 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2429 "popl %1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2430 "popl %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2431 : : "r" (src), |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2432 "r" (dst), |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2433 "r" (srcStride), |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2434 "r" (dstStride), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2435 "m" (numLines>>2) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2436 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2437 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2438 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2439 for(i=0; i<numLines; i++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2440 memcpy( &(dst[dstStride*i]), |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2441 &(src[srcStride*i]), BLOCK_SIZE); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2442 #endif |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2443 } |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2444 else |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2445 { |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2446 #ifdef HAVE_MMX |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2447 asm volatile( |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2448 "movl %4, %%eax \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2449 "movl %%eax, temp0\n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2450 "pushl %0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2451 "pushl %1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2452 "leal (%2,%2), %%eax \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2453 "leal (%3,%3), %%ebx \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2454 "movq packedYOffset, %%mm2 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2455 "movq packedYScale, %%mm3 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2456 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2457 #define SIMPLE_CPY \ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2458 "movq (%0), %%mm0 \n\t"\ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2459 "movq (%0,%2), %%mm1 \n\t"\ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2460 "movq %%mm0, (%1) \n\t"\ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2461 "movq %%mm1, (%1, %3) \n\t"\ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2462 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2463 "1: \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2464 SIMPLE_CPY |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2465 "addl %%eax, %0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2466 "addl %%ebx, %1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2467 SIMPLE_CPY |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2468 "addl %%eax, %0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2469 "addl %%ebx, %1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2470 "decl temp0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2471 "jnz 1b \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2472 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2473 "popl %1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2474 "popl %0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2475 : : "r" (src), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2476 "r" (dst), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2477 "r" (srcStride), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2478 "r" (dstStride), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2479 "m" (numLines>>2) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2480 : "%eax", "%ebx" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2481 ); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2482 #else |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2483 for(i=0; i<numLines; i++) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2484 memcpy( &(dst[dstStride*i]), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2485 &(src[srcStride*i]), BLOCK_SIZE); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2486 #endif |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2487 } |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2488 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2489 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2490 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2491 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2492 * Filters array of bytes (Y or U or V values) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2493 */ |
2185 | 2494 static 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
|
2495 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
|
2496 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2497 int x,y; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2498 /* 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
|
2499 after watching a black picture for 5 hours*/ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2500 static uint64_t *yHistogram= NULL; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2501 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
|
2502 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2503 #ifdef TIMEING |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2504 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
|
2505 sumTime= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2506 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2507 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2508 if(!yHistogram) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2509 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2510 int i; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2511 yHistogram= (uint64_t*)malloc(8*256); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2512 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
|
2513 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2514 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2515 if(!isColor) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2516 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2517 uint64_t sum= 0; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2518 int i; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2519 static int framenum= -1; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2520 uint64_t maxClipped; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2521 uint64_t clipped; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2522 double scale; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2523 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2524 framenum++; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2525 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
|
2526 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2527 for(i=0; i<256; i++) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2528 { |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2529 sum+= yHistogram[i]; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2530 // printf("%d ", yHistogram[i]); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2531 } |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2532 // printf("\n\n"); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2533 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2534 /* we allways get a completly black picture first */ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2535 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2536 maxClipped= (uint64_t)(sum * maxClippedThreshold); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2537 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2538 clipped= sum; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2539 for(black=255; black>0; black--) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2540 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2541 if(clipped < maxClipped) break; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2542 clipped-= yHistogram[black]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2543 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2544 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2545 clipped= sum; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2546 for(white=0; white<256; white++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2547 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2548 if(clipped < maxClipped) break; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2549 clipped-= yHistogram[white]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2550 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2551 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2552 // we cant handle negative correctures |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2553 packedYOffset= MAX(black - minAllowedY, 0); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2554 packedYOffset|= packedYOffset<<32; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2555 packedYOffset|= packedYOffset<<16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2556 packedYOffset|= packedYOffset<<8; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2557 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2558 scale= (double)(maxAllowedY - minAllowedY) / (double)(white-black); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2559 |
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
|
2560 packedYScale= (uint16_t)(scale*512.0 + 0.5); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2561 packedYScale|= packedYScale<<32; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2562 packedYScale|= packedYScale<<16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2563 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2564 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2565 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2566 packedYScale= 0x0100010001000100LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2567 packedYOffset= 0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2568 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2569 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2570 for(x=0; x<width; x+=BLOCK_SIZE) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2571 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
|
2572 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2573 for(y=0; y<height; y+=BLOCK_SIZE) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2574 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2575 //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
|
2576 uint8_t *srcBlock= &(src[y*srcStride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2577 uint8_t *dstBlock= &(dst[y*dstStride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2578 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
|
2579 uint8_t *vertBlock= &(dstBlock[dstStride*3]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2580 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2581 // 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
|
2582 // 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
|
2583 for(x=0; x<width; x+=BLOCK_SIZE) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2584 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2585 const int stride= dstStride; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2586 int QP= isColor ? |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2587 QPs[(y>>3)*QPStride + (x>>3)]: |
2180 | 2588 QPs[(y>>4)*QPStride + (x>>4)]; |
2589 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
|
2590 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2591 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2592 "movd %0, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2593 "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
|
2594 "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
|
2595 "packuswb %%mm7, %%mm7 \n\t" // QP,..., QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2596 "movq %%mm7, pQPb \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2597 : : "r" (QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2598 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2599 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2600 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2601 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2602 if(y + 12 < height) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2603 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2604 #ifdef MORE_TIMEING |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2605 T0= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2606 #endif |
2159 | 2607 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2608 #ifdef HAVE_MMX2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2609 prefetchnta(vertSrcBlock + (((x>>3)&3) + 2)*srcStride + 32); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2610 prefetchnta(vertSrcBlock + (((x>>3)&3) + 6)*srcStride + 32); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2611 prefetcht0(vertBlock + (((x>>3)&3) + 2)*dstStride + 32); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2612 prefetcht0(vertBlock + (((x>>3)&3) + 6)*dstStride + 32); |
2159 | 2613 #elif defined(HAVE_3DNOW) |
2614 //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ... | |
2615 /* prefetch(vertSrcBlock + (((x>>3)&3) + 2)*srcStride + 32); | |
2616 prefetch(vertSrcBlock + (((x>>3)&3) + 6)*srcStride + 32); | |
2617 prefetchw(vertBlock + (((x>>3)&3) + 2)*dstStride + 32); | |
2618 prefetchw(vertBlock + (((x>>3)&3) + 6)*dstStride + 32); | |
2619 */ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2620 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2621 if(!isColor) yHistogram[ srcBlock[0] ]++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2622 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2623 blockCopy(vertBlock + dstStride*2, dstStride, |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2624 vertSrcBlock + srcStride*2, srcStride, 8, mode & LEVEL_FIX); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2625 |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2626 if(mode & LINEAR_IPOL_DEINT_FILTER) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2627 deInterlaceInterpolateLinear(dstBlock, dstStride); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2628 else if(mode & LINEAR_BLEND_DEINT_FILTER) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2629 deInterlaceBlendLinear(dstBlock, dstStride); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2630 else if(mode & MEDIAN_DEINT_FILTER) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2631 deInterlaceMedian(dstBlock, dstStride); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2632 /* else if(mode & CUBIC_IPOL_DEINT_FILTER) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2633 deInterlaceInterpolateCubic(dstBlock, dstStride); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2634 else if(mode & CUBIC_BLEND_DEINT_FILTER) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2635 deInterlaceBlendCubic(dstBlock, dstStride); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2636 */ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2637 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2638 #ifdef MORE_TIMEING |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2639 T1= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2640 memcpyTime+= T1-T0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2641 T0=T1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2642 #endif |
2159 | 2643 if(mode & V_DEBLOCK) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2644 { |
2179 | 2645 if(mode & V_RK1_FILTER) |
2646 vertRK1Filter(vertBlock, stride, QP); | |
2647 else if(mode & V_X1_FILTER) | |
2159 | 2648 vertX1Filter(vertBlock, stride, QP); |
2649 else | |
2650 { | |
2651 if( isVertDC(vertBlock, stride)) | |
2652 { | |
2653 if(isVertMinMaxOk(vertBlock, stride, QP)) | |
2654 doVertLowPass(vertBlock, stride, QP); | |
2655 } | |
2656 else | |
2657 doVertDefFilter(vertBlock, stride, QP); | |
2658 } | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2659 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2660 #ifdef MORE_TIMEING |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2661 T1= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2662 vertTime+= T1-T0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2663 T0=T1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2664 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2665 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2666 else |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2667 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2668 blockCopy(vertBlock + dstStride*1, dstStride, |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2669 vertSrcBlock + srcStride*1, srcStride, 4, mode & LEVEL_FIX); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2670 |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2671 if(mode & LINEAR_IPOL_DEINT_FILTER) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2672 deInterlaceInterpolateLinearLastRow(dstBlock, dstStride); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2673 else if(mode & LINEAR_BLEND_DEINT_FILTER) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2674 deInterlaceBlendLinearLastRow(dstBlock, dstStride); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2675 else if(mode & MEDIAN_DEINT_FILTER) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2676 deInterlaceMedianLastRow(dstBlock, dstStride); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2677 /* else if(mode & CUBIC_IPOL_DEINT_FILTER) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2678 deInterlaceInterpolateCubicLastRow(dstBlock, dstStride); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2679 else if(mode & CUBIC_BLEND_DEINT_FILTER) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2680 deInterlaceBlendCubicLastRow(dstBlock, dstStride); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2681 */ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2682 } |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2683 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2684 if(x - 8 >= 0 && x<width) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2685 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2686 #ifdef MORE_TIMEING |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2687 T0= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2688 #endif |
2159 | 2689 if(mode & H_DEBLOCK) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2690 { |
2179 | 2691 if(mode & H_X1_FILTER) |
2692 horizX1Filter(dstBlock-4, stride, QP); | |
2693 else | |
2159 | 2694 { |
2179 | 2695 if( isHorizDCAndCopy2Temp(dstBlock-4, stride)) |
2696 { | |
2697 if(isHorizMinMaxOk(tempBlock, TEMP_STRIDE, QP)) | |
2698 doHorizLowPassAndCopyBack(dstBlock-4, stride, QP); | |
2699 } | |
2700 else | |
2701 doHorizDefFilterAndCopyBack(dstBlock-4, stride, QP); | |
2159 | 2702 } |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2703 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2704 #ifdef MORE_TIMEING |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2705 T1= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2706 horizTime+= T1-T0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2707 T0=T1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2708 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2709 dering(dstBlock - 9 - stride, stride, QP); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2710 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2711 else if(y!=0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2712 dering(dstBlock - stride*9 + width-9, stride, QP); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2713 //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
|
2714 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2715 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2716 dstBlock+=8; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2717 srcBlock+=8; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2718 vertBlock+=8; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2719 vertSrcBlock+=8; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2720 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2721 } |
2159 | 2722 #ifdef HAVE_3DNOW |
2723 asm volatile("femms"); | |
2724 #elif defined (HAVE_MMX) | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2725 asm volatile("emms"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2726 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2727 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2728 #ifdef TIMEING |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2729 // 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
|
2730 sumTime= rdtsc() - sumTime; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2731 if(!isColor) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2732 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
|
2733 (int)(memcpyTime/1000), (int)(vertTime/1000), (int)(horizTime/1000), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2734 (int)(sumTime/1000), (int)((sumTime-memcpyTime-vertTime-horizTime)/1000) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2735 , black, white); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2736 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2737 } |
2159 | 2738 |
2739 |