Mercurial > libavcodec.hg
annotate libpostproc/postprocess_template.c @ 116:1895a8fa81ea libavcodec
auto brightness/ contrast bugfix
getPPModeByNameAndQuality
author | michael |
---|---|
date | Mon, 22 Oct 2001 23:36:35 +0000 |
parents | 4514b8e7f0f1 |
children | a02f3088b0cf |
rev | line source |
---|---|
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1 /* |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2 Copyright (C) 2001 Michael Niedermayer (michaelni@gmx.at) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4 This program is free software; you can redistribute it and/or modify |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
5 it under the terms of the GNU General Public License as published by |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
6 the Free Software Foundation; either version 2 of the License, or |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
7 (at your option) any later version. |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
8 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
9 This program is distributed in the hope that it will be useful, |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
10 but WITHOUT ANY WARRANTY; without even the implied warranty of |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
12 GNU General Public License for more details. |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
13 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
14 You should have received a copy of the GNU General Public License |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
15 along with this program; if not, write to the Free Software |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
16 Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
17 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
18 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
19 /* |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
20 C MMX MMX2 3DNow |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
21 isVertDC Ec Ec |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
22 isVertMinMaxOk Ec Ec |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
23 doVertLowPass E e e |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
24 doVertDefFilter Ec Ec Ec |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
25 isHorizDC Ec Ec |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
26 isHorizMinMaxOk a |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
27 doHorizLowPass E a a |
96 | 28 doHorizDefFilter E ac ac |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
29 deRing |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
30 Vertical RKAlgo1 E a a |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
31 Vertical X1 a E E |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
32 Horizontal X1 a E E |
111 | 33 LinIpolDeinterlace e E E* |
34 CubicIpolDeinterlace a e e* | |
35 LinBlendDeinterlace e E E* | |
107 | 36 MedianDeinterlace Ec Ec |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
37 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
38 |
96 | 39 * i dont have a 3dnow CPU -> its untested |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
40 E = Exact implementation |
111 | 41 e = allmost exact implementation (slightly different rounding,...) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
42 a = alternative / approximate impl |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
43 c = checked against the other implementations (-vo md5) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
44 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
45 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
46 /* |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
47 TODO: |
99 | 48 verify that everything workes as it should (how?) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
49 reduce the time wasted on the mem transfer |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
50 implement dering |
96 | 51 implement everything in C at least (done at the moment but ...) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
52 unroll stuff if instructions depend too much on the prior one |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
53 we use 8x8 blocks for the horizontal filters, opendivx seems to use 8x4? |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
54 move YScale thing to the end instead of fixing QP |
96 | 55 write a faster and higher quality deblocking filter :) |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
56 do something about the speed of the horizontal filters |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
57 make the mainloop more flexible (variable number of blocks at once |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
58 (the if/else stuff per block is slowing things down) |
99 | 59 compare the quality & speed of all filters |
60 split this huge file | |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
61 fix warnings (unused vars, ...) |
107 | 62 noise reduction filters |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
63 ... |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
64 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
65 Notes: |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
66 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
67 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
68 |
107 | 69 //Changelog: use the CVS log |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
70 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
71 #include <inttypes.h> |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
72 #include <stdio.h> |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
73 #include <stdlib.h> |
116 | 74 #include <string.h> |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
75 #include "../config.h" |
96 | 76 //#undef HAVE_MMX2 |
77 //#define HAVE_3DNOW | |
78 //#undef HAVE_MMX | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
79 #include "postprocess.h" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
80 |
104 | 81 #define MIN(a,b) ((a) > (b) ? (b) : (a)) |
82 #define MAX(a,b) ((a) < (b) ? (b) : (a)) | |
83 #define ABS(a) ((a) > 0 ? (a) : (-(a))) | |
84 #define SIGN(a) ((a) > 0 ? 1 : -1) | |
85 | |
86 #ifdef HAVE_MMX2 | |
87 #define PAVGB(a,b) "pavgb " #a ", " #b " \n\t" | |
88 #elif defined (HAVE_3DNOW) | |
89 #define PAVGB(a,b) "pavgusb " #a ", " #b " \n\t" | |
90 #endif | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
91 |
116 | 92 #define GET_MODE_BUFFER_SIZE 500 |
93 #define OPTIONS_ARRAY_SIZE 10 | |
94 | |
95 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
96 static uint64_t packedYOffset= 0x0000000000000000LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
97 static uint64_t packedYScale= 0x0100010001000100LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
98 static uint64_t w05= 0x0005000500050005LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
99 static uint64_t w20= 0x0020002000200020LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
100 static uint64_t w1400= 0x1400140014001400LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
101 static uint64_t bm00000001= 0x00000000000000FFLL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
102 static uint64_t bm00010000= 0x000000FF00000000LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
103 static uint64_t bm00001000= 0x00000000FF000000LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
104 static uint64_t bm10000000= 0xFF00000000000000LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
105 static uint64_t bm10000001= 0xFF000000000000FFLL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
106 static uint64_t bm11000011= 0xFFFF00000000FFFFLL; |
96 | 107 static uint64_t bm00000011= 0x000000000000FFFFLL; |
99 | 108 static uint64_t bm11111110= 0xFFFFFFFFFFFFFF00LL; |
96 | 109 static uint64_t bm11000000= 0xFFFF000000000000LL; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
110 static uint64_t bm00011000= 0x000000FFFF000000LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
111 static uint64_t bm00110011= 0x0000FFFF0000FFFFLL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
112 static uint64_t bm11001100= 0xFFFF0000FFFF0000LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
113 static uint64_t b00= 0x0000000000000000LL; |
99 | 114 static uint64_t b01= 0x0101010101010101LL; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
115 static uint64_t b02= 0x0202020202020202LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
116 static uint64_t b0F= 0x0F0F0F0F0F0F0F0FLL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
117 static uint64_t bFF= 0xFFFFFFFFFFFFFFFFLL; |
96 | 118 static uint64_t b20= 0x2020202020202020LL; |
119 static uint64_t b80= 0x8080808080808080LL; | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
120 static uint64_t b7E= 0x7E7E7E7E7E7E7E7ELL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
121 static uint64_t b7C= 0x7C7C7C7C7C7C7C7CLL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
122 static uint64_t b3F= 0x3F3F3F3F3F3F3F3FLL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
123 static uint64_t temp0=0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
124 static uint64_t temp1=0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
125 static uint64_t temp2=0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
126 static uint64_t temp3=0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
127 static uint64_t temp4=0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
128 static uint64_t temp5=0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
129 static uint64_t pQPb=0; |
112 | 130 static uint8_t tempBlock[16*16]; //used so the horizontal code gets aligned data |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
131 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
132 int hFlatnessThreshold= 56 - 16; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
133 int vFlatnessThreshold= 56 - 16; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
134 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
135 //amount of "black" u r willing to loose to get a brightness corrected picture |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
136 double maxClippedThreshold= 0.01; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
137 |
116 | 138 int maxAllowedY=234; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
139 //FIXME can never make a movie´s black brighter (anyone needs that?) |
112 | 140 int minAllowedY=16; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
141 |
116 | 142 static struct PPFilter filters[]= |
143 { | |
144 {"hb", "hdeblock", 1, 1, 3, H_DEBLOCK}, | |
145 {"vb", "vdeblock", 1, 2, 4, V_DEBLOCK}, | |
146 {"vr", "rkvdeblock", 1, 2, 4, H_RK1_FILTER}, | |
147 {"h1", "x1hdeblock", 1, 1, 3, H_X1_FILTER}, | |
148 {"v1", "x1vdeblock", 1, 2, 4, V_X1_FILTER}, | |
149 {"dr", "dering", 1, 5, 6, DERING}, | |
150 {"al", "autolevels", 0, 1, 2, LEVEL_FIX}, | |
151 {"lb", "linblenddeint", 0, 1, 6, LINEAR_BLEND_DEINT_FILTER}, | |
152 {"li", "linipoldeint", 0, 1, 6, LINEAR_IPOL_DEINT_FILTER}, | |
153 {"ci", "cubicipoldeint", 0, 1, 6, CUBIC_IPOL_DEINT_FILTER}, | |
154 {"md", "mediandeint", 0, 1, 6, MEDIAN_DEINT_FILTER}, | |
155 {NULL, NULL,0,0,0,0} //End Marker | |
156 }; | |
157 | |
158 static char *replaceTable[]= | |
159 { | |
160 "default", "hdeblock:a,vdeblock:a,dering:a,autolevels", | |
161 "de", "hdeblock:a,vdeblock:a,dering:a,autolevels", | |
162 "fast", "x1hdeblock:a,x1vdeblock:a,dering:a,autolevels", | |
163 "fa", "x1hdeblock:a,x1vdeblock:a,dering:a,autolevels", | |
164 NULL //End Marker | |
165 }; | |
166 | |
107 | 167 #ifdef TIMING |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
168 static inline long long rdtsc() |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
169 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
170 long long l; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
171 asm volatile( "rdtsc\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
172 : "=A" (l) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
173 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
174 // printf("%d\n", int(l/1000)); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
175 return l; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
176 } |
102 | 177 #endif |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
178 |
102 | 179 #ifdef HAVE_MMX2 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
180 static inline void prefetchnta(void *p) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
181 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
182 asm volatile( "prefetchnta (%0)\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
183 : : "r" (p) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
184 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
185 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
186 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
187 static inline void prefetcht0(void *p) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
188 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
189 asm volatile( "prefetcht0 (%0)\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
190 : : "r" (p) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
191 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
192 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
193 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
194 static inline void prefetcht1(void *p) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
195 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
196 asm volatile( "prefetcht1 (%0)\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
197 : : "r" (p) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
198 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
199 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
200 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
201 static inline void prefetcht2(void *p) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
202 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
203 asm volatile( "prefetcht2 (%0)\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
204 : : "r" (p) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
205 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
206 } |
102 | 207 #endif |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
208 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
209 //FIXME? |255-0| = 1 (shouldnt be a problem ...) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
210 /** |
111 | 211 * Check if the middle 8x8 Block in the given 8x16 block is flat |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
212 */ |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
213 static inline int isVertDC(uint8_t src[], int stride){ |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
214 int numEq= 0; |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
215 int y; |
111 | 216 src+= stride*4; // src points to begin of the 8x8 Block |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
217 #ifdef HAVE_MMX |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
218 asm volatile( |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
219 "pushl %1\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
220 "movq b7E, %%mm7 \n\t" // mm7 = 0x7F |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
221 "movq b7C, %%mm6 \n\t" // mm6 = 0x7D |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
222 "movq (%1), %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
223 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
224 "movq (%1), %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
225 "psubb %%mm1, %%mm0 \n\t" // mm0 = differnece |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
226 "paddb %%mm7, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
227 "pcmpgtb %%mm6, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
228 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
229 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
230 "movq (%1), %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
231 "psubb %%mm2, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
232 "paddb %%mm7, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
233 "pcmpgtb %%mm6, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
234 "paddb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
235 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
236 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
237 "movq (%1), %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
238 "psubb %%mm1, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
239 "paddb %%mm7, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
240 "pcmpgtb %%mm6, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
241 "paddb %%mm2, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
242 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
243 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
244 "movq (%1), %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
245 "psubb %%mm2, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
246 "paddb %%mm7, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
247 "pcmpgtb %%mm6, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
248 "paddb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
249 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
250 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
251 "movq (%1), %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
252 "psubb %%mm1, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
253 "paddb %%mm7, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
254 "pcmpgtb %%mm6, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
255 "paddb %%mm2, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
256 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
257 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
258 "movq (%1), %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
259 "psubb %%mm2, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
260 "paddb %%mm7, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
261 "pcmpgtb %%mm6, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
262 "paddb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
263 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
264 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
265 "movq (%1), %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
266 "psubb %%mm1, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
267 "paddb %%mm7, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
268 "pcmpgtb %%mm6, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
269 "paddb %%mm2, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
270 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
271 " \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
272 "movq %%mm0, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
273 "psrlw $8, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
274 "paddb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
275 "movq %%mm0, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
276 "psrlq $16, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
277 "paddb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
278 "movq %%mm0, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
279 "psrlq $32, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
280 "paddb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
281 "popl %1\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
282 "movd %%mm0, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
283 : "=r" (numEq) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
284 : "r" (src), "r" (stride) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
285 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
286 // printf("%d\n", numEq); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
287 numEq= (256 - (numEq & 0xFF)) &0xFF; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
288 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
289 // int asmEq= numEq; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
290 // numEq=0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
291 // uint8_t *temp= src; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
292 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
293 #else |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
294 for(y=0; y<BLOCK_SIZE-1; y++) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
295 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
296 if(((src[0] - src[0+stride] + 1)&0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
297 if(((src[1] - src[1+stride] + 1)&0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
298 if(((src[2] - src[2+stride] + 1)&0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
299 if(((src[3] - src[3+stride] + 1)&0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
300 if(((src[4] - src[4+stride] + 1)&0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
301 if(((src[5] - src[5+stride] + 1)&0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
302 if(((src[6] - src[6+stride] + 1)&0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
303 if(((src[7] - src[7+stride] + 1)&0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
304 src+= stride; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
305 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
306 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
307 /* if(abs(numEq - asmEq) > 0) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
308 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
309 printf("\nasm:%d c:%d\n", asmEq, numEq); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
310 for(int y=0; y<8; y++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
311 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
312 for(int x=0; x<8; x++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
313 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
314 printf("%d ", temp[x + y*stride]); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
315 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
316 printf("\n"); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
317 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
318 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
319 */ |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
320 // for(int i=0; i<numEq/8; i++) src[i]=255; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
321 return (numEq > vFlatnessThreshold) ? 1 : 0; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
322 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
323 |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
324 static inline int isVertMinMaxOk(uint8_t src[], int stride, int QP) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
325 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
326 #ifdef HAVE_MMX |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
327 int isOk; |
111 | 328 src+= stride*3; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
329 asm volatile( |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
330 // "int $3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
331 "movq (%1, %2), %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
332 "movq (%1, %2, 8), %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
333 "movq %%mm0, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
334 "psubusb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
335 "psubusb %%mm2, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
336 "por %%mm1, %%mm0 \n\t" // ABS Diff |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
337 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
338 "movq pQPb, %%mm7 \n\t" // QP,..., QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
339 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
340 "psubusb %%mm7, %%mm0 \n\t" // Diff <= 2QP -> 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
341 "pcmpeqd b00, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
342 "psrlq $16, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
343 "pcmpeqd bFF, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
344 // "movd %%mm0, (%1, %2, 4)\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
345 "movd %%mm0, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
346 : "=r" (isOk) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
347 : "r" (src), "r" (stride) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
348 ); |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
349 return isOk ? 1 : 0; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
350 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
351 |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
352 int isOk2= 1; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
353 int x; |
111 | 354 src+= stride*3; |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
355 for(x=0; x<BLOCK_SIZE; x++) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
356 { |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
357 if(abs((int)src[x + stride] - (int)src[x + (stride<<3)]) > 2*QP) isOk2=0; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
358 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
359 /* if(isOk && !isOk2 || !isOk && isOk2) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
360 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
361 printf("\nasm:%d c:%d QP:%d\n", isOk, isOk2, QP); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
362 for(int y=0; y<9; y++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
363 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
364 for(int x=0; x<8; x++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
365 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
366 printf("%d ", src[x + y*stride]); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
367 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
368 printf("\n"); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
369 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
370 } */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
371 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
372 return isOk2; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
373 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
374 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
375 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
376 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
377 /** |
111 | 378 * Do a vertical low pass filter on the 8x16 block (only write to the 8x8 block in the middle) |
107 | 379 * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
380 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
381 static inline void doVertLowPass(uint8_t *src, int stride, int QP) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
382 { |
96 | 383 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
111 | 384 src+= stride*3; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
385 asm volatile( //"movv %0 %1 %2\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
386 "pushl %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
387 "movq pQPb, %%mm0 \n\t" // QP,..., QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
388 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
389 "movq (%0), %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
390 "movq (%0, %1), %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
391 "movq %%mm5, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
392 "movq %%mm6, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
393 "psubusb %%mm6, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
394 "psubusb %%mm1, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
395 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
396 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
397 "pcmpeqb b00, %%mm2 \n\t" // diff <= QP -> FF |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
398 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
399 "pand %%mm2, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
400 "pandn %%mm1, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
401 "por %%mm2, %%mm6 \n\t"// First Line to Filter |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
402 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
403 "movq (%0, %1, 8), %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
404 "leal (%0, %1, 4), %%eax \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
405 "leal (%0, %1, 8), %%ebx \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
406 "subl %1, %%ebx \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
407 "addl %1, %0 \n\t" // %0 points to line 1 not 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
408 "movq (%0, %1, 8), %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
409 "movq %%mm5, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
410 "movq %%mm7, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
411 "psubusb %%mm7, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
412 "psubusb %%mm1, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
413 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
414 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
415 "pcmpeqb b00, %%mm2 \n\t" // diff <= QP -> FF |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
416 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
417 "pand %%mm2, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
418 "pandn %%mm1, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
419 "por %%mm2, %%mm7 \n\t" // First Line to Filter |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
420 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
421 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
422 // 1 2 3 4 5 6 7 8 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
423 // %0 %0+%1 %0+2%1 eax %0+4%1 eax+2%1 ebx eax+4%1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
424 // 6 4 2 2 1 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
425 // 6 4 4 2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
426 // 6 8 2 |
111 | 427 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
428 "movq (%0, %1), %%mm0 \n\t" // 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
429 "movq %%mm0, %%mm1 \n\t" // 1 |
96 | 430 PAVGB(%%mm6, %%mm0) //1 1 /2 |
431 PAVGB(%%mm6, %%mm0) //3 1 /4 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
432 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
433 "movq (%0, %1, 4), %%mm2 \n\t" // 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
434 "movq %%mm2, %%mm5 \n\t" // 1 |
96 | 435 PAVGB((%%eax), %%mm2) // 11 /2 |
436 PAVGB((%0, %1, 2), %%mm2) // 211 /4 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
437 "movq %%mm2, %%mm3 \n\t" // 211 /4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
438 "movq (%0), %%mm4 \n\t" // 1 |
96 | 439 PAVGB(%%mm4, %%mm3) // 4 211 /8 |
440 PAVGB(%%mm0, %%mm3) //642211 /16 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
441 "movq %%mm3, (%0) \n\t" // X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
442 // mm1=2 mm2=3(211) mm4=1 mm5=5 mm6=0 mm7=9 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
443 "movq %%mm1, %%mm0 \n\t" // 1 |
96 | 444 PAVGB(%%mm6, %%mm0) //1 1 /2 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
445 "movq %%mm4, %%mm3 \n\t" // 1 |
96 | 446 PAVGB((%0,%1,2), %%mm3) // 1 1 /2 |
447 PAVGB((%%eax,%1,2), %%mm5) // 11 /2 | |
448 PAVGB((%%eax), %%mm5) // 211 /4 | |
449 PAVGB(%%mm5, %%mm3) // 2 2211 /8 | |
450 PAVGB(%%mm0, %%mm3) //4242211 /16 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
451 "movq %%mm3, (%0,%1) \n\t" // X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
452 // mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9 |
96 | 453 PAVGB(%%mm4, %%mm6) //11 /2 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
454 "movq (%%ebx), %%mm0 \n\t" // 1 |
96 | 455 PAVGB((%%eax, %1, 2), %%mm0) // 11/2 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
456 "movq %%mm0, %%mm3 \n\t" // 11/2 |
96 | 457 PAVGB(%%mm1, %%mm0) // 2 11/4 |
458 PAVGB(%%mm6, %%mm0) //222 11/8 | |
459 PAVGB(%%mm2, %%mm0) //22242211/16 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
460 "movq (%0, %1, 2), %%mm2 \n\t" // 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
461 "movq %%mm0, (%0, %1, 2) \n\t" // X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
462 // mm1=2 mm2=3 mm3=6(11) mm4=1 mm5=4(211) mm6=0(11) mm7=9 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
463 "movq (%%eax, %1, 4), %%mm0 \n\t" // 1 |
96 | 464 PAVGB((%%ebx), %%mm0) // 11 /2 |
465 PAVGB(%%mm0, %%mm6) //11 11 /4 | |
466 PAVGB(%%mm1, %%mm4) // 11 /2 | |
467 PAVGB(%%mm2, %%mm1) // 11 /2 | |
468 PAVGB(%%mm1, %%mm6) //1122 11 /8 | |
469 PAVGB(%%mm5, %%mm6) //112242211 /16 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
470 "movq (%%eax), %%mm5 \n\t" // 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
471 "movq %%mm6, (%%eax) \n\t" // X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
472 // mm0=7(11) mm1=2(11) mm2=3 mm3=6(11) mm4=1(11) mm5=4 mm7=9 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
473 "movq (%%eax, %1, 4), %%mm6 \n\t" // 1 |
96 | 474 PAVGB(%%mm7, %%mm6) // 11 /2 |
475 PAVGB(%%mm4, %%mm6) // 11 11 /4 | |
476 PAVGB(%%mm3, %%mm6) // 11 2211 /8 | |
477 PAVGB(%%mm5, %%mm2) // 11 /2 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
478 "movq (%0, %1, 4), %%mm4 \n\t" // 1 |
96 | 479 PAVGB(%%mm4, %%mm2) // 112 /4 |
480 PAVGB(%%mm2, %%mm6) // 112242211 /16 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
481 "movq %%mm6, (%0, %1, 4) \n\t" // X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
482 // mm0=7(11) mm1=2(11) mm2=3(112) mm3=6(11) mm4=5 mm5=4 mm7=9 |
96 | 483 PAVGB(%%mm7, %%mm1) // 11 2 /4 |
484 PAVGB(%%mm4, %%mm5) // 11 /2 | |
485 PAVGB(%%mm5, %%mm0) // 11 11 /4 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
486 "movq (%%eax, %1, 2), %%mm6 \n\t" // 1 |
96 | 487 PAVGB(%%mm6, %%mm1) // 11 4 2 /8 |
488 PAVGB(%%mm0, %%mm1) // 11224222 /16 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
489 "movq %%mm1, (%%eax, %1, 2) \n\t" // X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
490 // mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9 |
96 | 491 PAVGB((%%ebx), %%mm2) // 112 4 /8 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
492 "movq (%%eax, %1, 4), %%mm0 \n\t" // 1 |
96 | 493 PAVGB(%%mm0, %%mm6) // 1 1 /2 |
494 PAVGB(%%mm7, %%mm6) // 1 12 /4 | |
495 PAVGB(%%mm2, %%mm6) // 1122424 /4 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
496 "movq %%mm6, (%%ebx) \n\t" // X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
497 // mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9 |
96 | 498 PAVGB(%%mm7, %%mm5) // 11 2 /4 |
499 PAVGB(%%mm7, %%mm5) // 11 6 /8 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
500 |
96 | 501 PAVGB(%%mm3, %%mm0) // 112 /4 |
502 PAVGB(%%mm0, %%mm5) // 112246 /16 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
503 "movq %%mm5, (%%eax, %1, 4) \n\t" // X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
504 "popl %0\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
505 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
506 : |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
507 : "r" (src), "r" (stride) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
508 : "%eax", "%ebx" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
509 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
510 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
511 const int l1= stride; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
512 const int l2= stride + l1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
513 const int l3= stride + l2; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
514 const int l4= stride + l3; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
515 const int l5= stride + l4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
516 const int l6= stride + l5; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
517 const int l7= stride + l6; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
518 const int l8= stride + l7; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
519 const int l9= stride + l8; |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
520 int x; |
111 | 521 src+= stride*3; |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
522 for(x=0; x<BLOCK_SIZE; x++) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
523 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
524 const int first= ABS(src[0] - src[l1]) < QP ? src[0] : src[l1]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
525 const int last= ABS(src[l8] - src[l9]) < QP ? src[l9] : src[l8]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
526 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
527 int sums[9]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
528 sums[0] = first + src[l1]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
529 sums[1] = src[l1] + src[l2]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
530 sums[2] = src[l2] + src[l3]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
531 sums[3] = src[l3] + src[l4]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
532 sums[4] = src[l4] + src[l5]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
533 sums[5] = src[l5] + src[l6]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
534 sums[6] = src[l6] + src[l7]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
535 sums[7] = src[l7] + src[l8]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
536 sums[8] = src[l8] + last; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
537 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
538 src[l1]= ((sums[0]<<2) + ((first + sums[2])<<1) + sums[4] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
539 src[l2]= ((src[l2]<<2) + (first + sums[0] + sums[3]<<1) + sums[5] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
540 src[l3]= ((src[l3]<<2) + (first + sums[1] + sums[4]<<1) + sums[6] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
541 src[l4]= ((src[l4]<<2) + (sums[2] + sums[5]<<1) + sums[0] + sums[7] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
542 src[l5]= ((src[l5]<<2) + (sums[3] + sums[6]<<1) + sums[1] + sums[8] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
543 src[l6]= ((src[l6]<<2) + (last + sums[7] + sums[4]<<1) + sums[2] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
544 src[l7]= ((last + src[l7]<<2) + (src[l8] + sums[5]<<1) + sums[3] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
545 src[l8]= ((sums[8]<<2) + (last + sums[6]<<1) + sums[4] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
546 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
547 src++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
548 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
549 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
550 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
551 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
552 |
96 | 553 /** |
554 * Experimental implementation of the filter (Algorithm 1) described in a paper from Ramkishor & Karandikar | |
555 * values are correctly clipped (MMX2) | |
556 * values are wraparound (C) | |
557 * conclusion: its fast, but introduces ugly horizontal patterns if there is a continious gradient | |
558 0 8 16 24 | |
559 x = 8 | |
560 x/2 = 4 | |
561 x/8 = 1 | |
562 1 12 12 23 | |
563 */ | |
99 | 564 static inline void vertRK1Filter(uint8_t *src, int stride, int QP) |
96 | 565 { |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
566 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
111 | 567 src+= stride*3; |
96 | 568 // FIXME rounding |
569 asm volatile( | |
570 "pxor %%mm7, %%mm7 \n\t" // 0 | |
571 "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE | |
572 "leal (%0, %1), %%eax \n\t" | |
573 "leal (%%eax, %1, 4), %%ebx \n\t" | |
574 // 0 1 2 3 4 5 6 7 8 9 | |
575 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 | |
576 "movq pQPb, %%mm0 \n\t" // QP,..., QP | |
577 "movq %%mm0, %%mm1 \n\t" // QP,..., QP | |
578 "paddusb b02, %%mm0 \n\t" | |
579 "psrlw $2, %%mm0 \n\t" | |
580 "pand b3F, %%mm0 \n\t" // QP/4,..., QP/4 | |
581 "paddusb %%mm1, %%mm0 \n\t" // QP*1.25 ... | |
582 "movq (%0, %1, 4), %%mm2 \n\t" // line 4 | |
583 "movq (%%ebx), %%mm3 \n\t" // line 5 | |
584 "movq %%mm2, %%mm4 \n\t" // line 4 | |
585 "pcmpeqb %%mm5, %%mm5 \n\t" // -1 | |
586 "pxor %%mm2, %%mm5 \n\t" // -line 4 - 1 | |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
587 PAVGB(%%mm3, %%mm5) |
96 | 588 "paddb %%mm6, %%mm5 \n\t" // (l5-l4)/2 |
589 "psubusb %%mm3, %%mm4 \n\t" | |
590 "psubusb %%mm2, %%mm3 \n\t" | |
591 "por %%mm3, %%mm4 \n\t" // |l4 - l5| | |
592 "psubusb %%mm0, %%mm4 \n\t" | |
593 "pcmpeqb %%mm7, %%mm4 \n\t" | |
594 "pand %%mm4, %%mm5 \n\t" // d/2 | |
595 | |
596 // "paddb %%mm6, %%mm2 \n\t" // line 4 + 0x80 | |
597 "paddb %%mm5, %%mm2 \n\t" | |
598 // "psubb %%mm6, %%mm2 \n\t" | |
599 "movq %%mm2, (%0,%1, 4) \n\t" | |
600 | |
601 "movq (%%ebx), %%mm2 \n\t" | |
602 // "paddb %%mm6, %%mm2 \n\t" // line 5 + 0x80 | |
603 "psubb %%mm5, %%mm2 \n\t" | |
604 // "psubb %%mm6, %%mm2 \n\t" | |
605 "movq %%mm2, (%%ebx) \n\t" | |
606 | |
607 "paddb %%mm6, %%mm5 \n\t" | |
608 "psrlw $2, %%mm5 \n\t" | |
609 "pand b3F, %%mm5 \n\t" | |
610 "psubb b20, %%mm5 \n\t" // (l5-l4)/8 | |
611 | |
612 "movq (%%eax, %1, 2), %%mm2 \n\t" | |
613 "paddb %%mm6, %%mm2 \n\t" // line 3 + 0x80 | |
614 "paddsb %%mm5, %%mm2 \n\t" | |
615 "psubb %%mm6, %%mm2 \n\t" | |
616 "movq %%mm2, (%%eax, %1, 2) \n\t" | |
617 | |
618 "movq (%%ebx, %1), %%mm2 \n\t" | |
619 "paddb %%mm6, %%mm2 \n\t" // line 6 + 0x80 | |
620 "psubsb %%mm5, %%mm2 \n\t" | |
621 "psubb %%mm6, %%mm2 \n\t" | |
622 "movq %%mm2, (%%ebx, %1) \n\t" | |
623 | |
624 : | |
625 : "r" (src), "r" (stride) | |
626 : "%eax", "%ebx" | |
627 ); | |
628 #else | |
629 const int l1= stride; | |
630 const int l2= stride + l1; | |
631 const int l3= stride + l2; | |
632 const int l4= stride + l3; | |
633 const int l5= stride + l4; | |
634 const int l6= stride + l5; | |
635 const int l7= stride + l6; | |
636 const int l8= stride + l7; | |
637 const int l9= stride + l8; | |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
638 int x; |
111 | 639 src+= stride*3; |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
640 for(x=0; x<BLOCK_SIZE; x++) |
96 | 641 { |
642 if(ABS(src[l4]-src[l5]) < QP + QP/4) | |
643 { | |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
644 int v = (src[l5] - src[l4]); |
96 | 645 |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
646 src[l3] +=v/8; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
647 src[l4] +=v/2; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
648 src[l5] -=v/2; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
649 src[l6] -=v/8; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
650 |
96 | 651 } |
652 src++; | |
653 } | |
654 | |
655 #endif | |
656 } | |
657 | |
658 /** | |
659 * Experimental Filter 1 | |
99 | 660 * will not damage linear gradients |
661 * Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter | |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
662 * can only smooth blocks at the expected locations (it cant smooth them if they did move) |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
663 * MMX2 version does correct clipping C version doesnt |
96 | 664 */ |
665 static inline void vertX1Filter(uint8_t *src, int stride, int QP) | |
666 { | |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
667 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
111 | 668 src+= stride*3; |
669 | |
96 | 670 asm volatile( |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
671 "pxor %%mm7, %%mm7 \n\t" // 0 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
672 // "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
673 "leal (%0, %1), %%eax \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
674 "leal (%%eax, %1, 4), %%ebx \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
675 // 0 1 2 3 4 5 6 7 8 9 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
676 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
677 "movq (%%eax, %1, 2), %%mm0 \n\t" // line 3 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
678 "movq (%0, %1, 4), %%mm1 \n\t" // line 4 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
679 "movq %%mm1, %%mm2 \n\t" // line 4 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
680 "psubusb %%mm0, %%mm1 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
681 "psubusb %%mm2, %%mm0 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
682 "por %%mm1, %%mm0 \n\t" // |l2 - l3| |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
683 "movq (%%ebx), %%mm3 \n\t" // line 5 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
684 "movq (%%ebx, %1), %%mm4 \n\t" // line 6 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
685 "movq %%mm3, %%mm5 \n\t" // line 5 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
686 "psubusb %%mm4, %%mm3 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
687 "psubusb %%mm5, %%mm4 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
688 "por %%mm4, %%mm3 \n\t" // |l5 - l6| |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
689 PAVGB(%%mm3, %%mm0) // (|l2 - l3| + |l5 - l6|)/2 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
690 "movq %%mm2, %%mm1 \n\t" // line 4 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
691 "psubusb %%mm5, %%mm2 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
692 "movq %%mm2, %%mm4 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
693 "pcmpeqb %%mm7, %%mm2 \n\t" // (l4 - l5) <= 0 ? -1 : 0 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
694 "psubusb %%mm1, %%mm5 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
695 "por %%mm5, %%mm4 \n\t" // |l4 - l5| |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
696 "psubusb %%mm0, %%mm4 \n\t" //d = MAX(0, |l4-l5| - (|l2-l3| + |l5-l6|)/2) |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
697 "movq %%mm4, %%mm3 \n\t" // d |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
698 "psubusb pQPb, %%mm4 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
699 "pcmpeqb %%mm7, %%mm4 \n\t" // d <= QP ? -1 : 0 |
99 | 700 "psubusb b01, %%mm3 \n\t" |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
701 "pand %%mm4, %%mm3 \n\t" // d <= QP ? d : 0 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
702 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
703 PAVGB(%%mm7, %%mm3) // d/2 |
99 | 704 "movq %%mm3, %%mm1 \n\t" // d/2 |
705 PAVGB(%%mm7, %%mm3) // d/4 | |
706 PAVGB(%%mm1, %%mm3) // 3*d/8 | |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
707 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
708 "movq (%0, %1, 4), %%mm0 \n\t" // line 4 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
709 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
710 "psubusb %%mm3, %%mm0 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
711 "pxor %%mm2, %%mm0 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
712 "movq %%mm0, (%0, %1, 4) \n\t" // line 4 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
713 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
714 "movq (%%ebx), %%mm0 \n\t" // line 5 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
715 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
716 "paddusb %%mm3, %%mm0 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
717 "pxor %%mm2, %%mm0 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
718 "movq %%mm0, (%%ebx) \n\t" // line 5 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
719 |
99 | 720 PAVGB(%%mm7, %%mm1) // d/4 |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
721 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
722 "movq (%%eax, %1, 2), %%mm0 \n\t" // line 3 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
723 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4 |
99 | 724 "psubusb %%mm1, %%mm0 \n\t" |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
725 "pxor %%mm2, %%mm0 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
726 "movq %%mm0, (%%eax, %1, 2) \n\t" // line 3 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
727 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
728 "movq (%%ebx, %1), %%mm0 \n\t" // line 6 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
729 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5 |
99 | 730 "paddusb %%mm1, %%mm0 \n\t" |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
731 "pxor %%mm2, %%mm0 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
732 "movq %%mm0, (%%ebx, %1) \n\t" // line 6 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
733 |
99 | 734 PAVGB(%%mm7, %%mm1) // d/8 |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
735 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
736 "movq (%%eax, %1), %%mm0 \n\t" // line 2 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
737 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l2-1 : l2 |
99 | 738 "psubusb %%mm1, %%mm0 \n\t" |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
739 "pxor %%mm2, %%mm0 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
740 "movq %%mm0, (%%eax, %1) \n\t" // line 2 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
741 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
742 "movq (%%ebx, %1, 2), %%mm0 \n\t" // line 7 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
743 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l7-1 : l7 |
99 | 744 "paddusb %%mm1, %%mm0 \n\t" |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
745 "pxor %%mm2, %%mm0 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
746 "movq %%mm0, (%%ebx, %1, 2) \n\t" // line 7 |
96 | 747 |
748 : | |
749 : "r" (src), "r" (stride) | |
750 : "%eax", "%ebx" | |
751 ); | |
752 #else | |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
753 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
754 const int l1= stride; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
755 const int l2= stride + l1; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
756 const int l3= stride + l2; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
757 const int l4= stride + l3; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
758 const int l5= stride + l4; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
759 const int l6= stride + l5; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
760 const int l7= stride + l6; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
761 const int l8= stride + l7; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
762 const int l9= stride + l8; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
763 int x; |
111 | 764 |
765 src+= stride*3; | |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
766 for(x=0; x<BLOCK_SIZE; x++) |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
767 { |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
768 int a= src[l3] - src[l4]; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
769 int b= src[l4] - src[l5]; |
99 | 770 int c= src[l5] - src[l6]; |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
771 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
772 int d= MAX(ABS(b) - (ABS(a) + ABS(c))/2, 0); |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
773 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
774 if(d < QP) |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
775 { |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
776 int v = d * SIGN(-b); |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
777 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
778 src[l2] +=v/8; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
779 src[l3] +=v/4; |
99 | 780 src[l4] +=3*v/8; |
781 src[l5] -=3*v/8; | |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
782 src[l6] -=v/4; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
783 src[l7] -=v/8; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
784 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
785 } |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
786 src++; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
787 } |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
788 /* |
96 | 789 const int l1= stride; |
790 const int l2= stride + l1; | |
791 const int l3= stride + l2; | |
792 const int l4= stride + l3; | |
793 const int l5= stride + l4; | |
794 const int l6= stride + l5; | |
795 const int l7= stride + l6; | |
796 const int l8= stride + l7; | |
797 const int l9= stride + l8; | |
798 for(int x=0; x<BLOCK_SIZE; x++) | |
799 { | |
800 int v2= src[l2]; | |
801 int v3= src[l3]; | |
802 int v4= src[l4]; | |
803 int v5= src[l5]; | |
804 int v6= src[l6]; | |
805 int v7= src[l7]; | |
806 | |
807 if(ABS(v4-v5)<QP && ABS(v4-v5) - (ABS(v3-v4) + ABS(v5-v6))>0 ) | |
808 { | |
809 src[l3] = (6*v2 + 4*v3 + 3*v4 + 2*v5 + v6 )/16; | |
810 src[l4] = (3*v2 + 3*v3 + 4*v4 + 3*v5 + 2*v6 + v7 )/16; | |
811 src[l5] = (1*v2 + 2*v3 + 3*v4 + 4*v5 + 3*v6 + 3*v7)/16; | |
812 src[l6] = ( 1*v3 + 2*v4 + 3*v5 + 4*v6 + 6*v7)/16; | |
813 } | |
814 src++; | |
815 } | |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
816 */ |
96 | 817 #endif |
818 } | |
819 | |
99 | 820 /** |
821 * Experimental Filter 1 (Horizontal) | |
822 * will not damage linear gradients | |
823 * Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter | |
824 * can only smooth blocks at the expected locations (it cant smooth them if they did move) | |
825 * MMX2 version does correct clipping C version doesnt | |
826 * not identical with the vertical one | |
827 */ | |
828 static inline void horizX1Filter(uint8_t *src, int stride, int QP) | |
829 { | |
830 int y; | |
831 static uint64_t *lut= NULL; | |
832 if(lut==NULL) | |
833 { | |
834 int i; | |
835 lut= (uint64_t*)memalign(8, 256*8); | |
836 for(i=0; i<256; i++) | |
837 { | |
838 int v= i < 128 ? 2*i : 2*(i-256); | |
839 /* | |
840 //Simulate 112242211 9-Tap filter | |
841 uint64_t a= (v/16) & 0xFF; | |
842 uint64_t b= (v/8) & 0xFF; | |
843 uint64_t c= (v/4) & 0xFF; | |
844 uint64_t d= (3*v/8) & 0xFF; | |
845 */ | |
846 //Simulate piecewise linear interpolation | |
847 uint64_t a= (v/16) & 0xFF; | |
848 uint64_t b= (v*3/16) & 0xFF; | |
849 uint64_t c= (v*5/16) & 0xFF; | |
850 uint64_t d= (7*v/16) & 0xFF; | |
851 uint64_t A= (0x100 - a)&0xFF; | |
852 uint64_t B= (0x100 - b)&0xFF; | |
853 uint64_t C= (0x100 - c)&0xFF; | |
854 uint64_t D= (0x100 - c)&0xFF; | |
855 | |
856 lut[i] = (a<<56) | (b<<48) | (c<<40) | (d<<32) | | |
857 (D<<24) | (C<<16) | (B<<8) | (A); | |
858 //lut[i] = (v<<32) | (v<<24); | |
859 } | |
860 } | |
861 | |
862 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) | |
863 asm volatile( | |
864 "pxor %%mm7, %%mm7 \n\t" // 0 | |
865 // "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE | |
866 "leal (%0, %1), %%eax \n\t" | |
867 "leal (%%eax, %1, 4), %%ebx \n\t" | |
868 | |
869 "movq b80, %%mm6 \n\t" | |
100 | 870 "movd pQPb, %%mm5 \n\t" // QP |
99 | 871 "movq %%mm5, %%mm4 \n\t" |
872 "paddusb %%mm5, %%mm5 \n\t" // 2QP | |
873 "paddusb %%mm5, %%mm4 \n\t" // 3QP | |
874 "pxor %%mm5, %%mm5 \n\t" // 0 | |
875 "psubb %%mm4, %%mm5 \n\t" // -3QP | |
876 "por bm11111110, %%mm5 \n\t" // ...,FF,FF,-3QP | |
877 "psllq $24, %%mm5 \n\t" | |
878 | |
879 // 0 1 2 3 4 5 6 7 8 9 | |
880 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 | |
881 | |
882 #define HX1old(a) \ | |
883 "movd " #a ", %%mm0 \n\t"\ | |
884 "movd 4" #a ", %%mm1 \n\t"\ | |
885 "punpckldq %%mm1, %%mm0 \n\t"\ | |
886 "movq %%mm0, %%mm1 \n\t"\ | |
887 "movq %%mm0, %%mm2 \n\t"\ | |
888 "psrlq $8, %%mm1 \n\t"\ | |
889 "psubusb %%mm1, %%mm2 \n\t"\ | |
890 "psubusb %%mm0, %%mm1 \n\t"\ | |
891 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ | |
892 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
893 "pshufw $0x00, %%mm1, %%mm3 \n\t" /* p´5 = |p1 - p2| */\ | |
894 PAVGB(%%mm1, %%mm3) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\ | |
895 "psrlq $16, %%mm3 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\ | |
896 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\ | |
897 "paddb %%mm5, %%mm1 \n\t"\ | |
898 "psubusb %%mm5, %%mm1 \n\t"\ | |
899 PAVGB(%%mm7, %%mm1)\ | |
900 "pxor %%mm2, %%mm1 \n\t"\ | |
901 "psubb %%mm2, %%mm1 \n\t"\ | |
902 "psrlq $24, %%mm1 \n\t"\ | |
903 "movd %%mm1, %%ecx \n\t"\ | |
904 "paddb %%mm6, %%mm0 \n\t"\ | |
905 "paddsb (%3, %%ecx, 8), %%mm0 \n\t"\ | |
906 "paddb %%mm6, %%mm0 \n\t"\ | |
907 "movq %%mm0, " #a " \n\t"\ | |
908 | |
909 /* | |
910 HX1old((%0)) | |
911 HX1old((%%eax)) | |
912 HX1old((%%eax, %1)) | |
913 HX1old((%%eax, %1, 2)) | |
914 HX1old((%0, %1, 4)) | |
915 HX1old((%%ebx)) | |
916 HX1old((%%ebx, %1)) | |
917 HX1old((%%ebx, %1, 2)) | |
918 */ | |
919 | |
920 //FIXME add some comments, its unreadable ... | |
921 #define HX1b(a, c, b, d) \ | |
922 "movd " #a ", %%mm0 \n\t"\ | |
923 "movd 4" #a ", %%mm1 \n\t"\ | |
924 "punpckldq %%mm1, %%mm0 \n\t"\ | |
925 "movd " #b ", %%mm4 \n\t"\ | |
926 "movq %%mm0, %%mm1 \n\t"\ | |
927 "movq %%mm0, %%mm2 \n\t"\ | |
928 "psrlq $8, %%mm1 \n\t"\ | |
929 "movd 4" #b ", %%mm3 \n\t"\ | |
930 "psubusb %%mm1, %%mm2 \n\t"\ | |
931 "psubusb %%mm0, %%mm1 \n\t"\ | |
932 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ | |
933 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
934 "punpckldq %%mm3, %%mm4 \n\t"\ | |
935 "movq %%mm1, %%mm3 \n\t"\ | |
936 "psllq $32, %%mm3 \n\t" /* p´5 = |p1 - p2| */\ | |
937 PAVGB(%%mm1, %%mm3) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\ | |
938 "paddb %%mm6, %%mm0 \n\t"\ | |
939 "psrlq $16, %%mm3 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\ | |
940 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\ | |
941 "movq %%mm4, %%mm3 \n\t"\ | |
942 "paddb %%mm5, %%mm1 \n\t"\ | |
943 "psubusb %%mm5, %%mm1 \n\t"\ | |
944 "psrlq $8, %%mm3 \n\t"\ | |
945 PAVGB(%%mm7, %%mm1)\ | |
946 "pxor %%mm2, %%mm1 \n\t"\ | |
947 "psubb %%mm2, %%mm1 \n\t"\ | |
948 "movq %%mm4, %%mm2 \n\t"\ | |
949 "psrlq $24, %%mm1 \n\t"\ | |
950 "psubusb %%mm3, %%mm2 \n\t"\ | |
951 "movd %%mm1, %%ecx \n\t"\ | |
952 "psubusb %%mm4, %%mm3 \n\t"\ | |
100 | 953 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\ |
99 | 954 "por %%mm2, %%mm3 \n\t" /* p´x = |px - p(x+1)| */\ |
955 "paddb %%mm6, %%mm0 \n\t"\ | |
956 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
957 "movq %%mm3, %%mm1 \n\t"\ | |
958 "psllq $32, %%mm1 \n\t" /* p´5 = |p1 - p2| */\ | |
959 "movq %%mm0, " #a " \n\t"\ | |
960 PAVGB(%%mm3, %%mm1) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\ | |
961 "paddb %%mm6, %%mm4 \n\t"\ | |
962 "psrlq $16, %%mm1 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\ | |
963 "psubusb %%mm1, %%mm3 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\ | |
964 "paddb %%mm5, %%mm3 \n\t"\ | |
965 "psubusb %%mm5, %%mm3 \n\t"\ | |
966 PAVGB(%%mm7, %%mm3)\ | |
967 "pxor %%mm2, %%mm3 \n\t"\ | |
968 "psubb %%mm2, %%mm3 \n\t"\ | |
969 "psrlq $24, %%mm3 \n\t"\ | |
970 "movd " #c ", %%mm0 \n\t"\ | |
971 "movd 4" #c ", %%mm1 \n\t"\ | |
972 "punpckldq %%mm1, %%mm0 \n\t"\ | |
973 "paddb %%mm6, %%mm0 \n\t"\ | |
100 | 974 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\ |
99 | 975 "paddb %%mm6, %%mm0 \n\t"\ |
976 "movq %%mm0, " #c " \n\t"\ | |
977 "movd %%mm3, %%ecx \n\t"\ | |
978 "movd " #d ", %%mm0 \n\t"\ | |
100 | 979 "paddsb (%2, %%ecx, 8), %%mm4 \n\t"\ |
99 | 980 "movd 4" #d ", %%mm1 \n\t"\ |
981 "paddb %%mm6, %%mm4 \n\t"\ | |
982 "punpckldq %%mm1, %%mm0 \n\t"\ | |
983 "movq %%mm4, " #b " \n\t"\ | |
984 "paddb %%mm6, %%mm0 \n\t"\ | |
100 | 985 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\ |
99 | 986 "paddb %%mm6, %%mm0 \n\t"\ |
987 "movq %%mm0, " #d " \n\t"\ | |
988 | |
989 HX1b((%0),(%%eax),(%%eax, %1),(%%eax, %1, 2)) | |
990 HX1b((%0, %1, 4),(%%ebx),(%%ebx, %1),(%%ebx, %1, 2)) | |
991 | |
992 | |
993 : | |
100 | 994 : "r" (src), "r" (stride), "r" (lut) |
99 | 995 : "%eax", "%ebx", "%ecx" |
996 ); | |
997 #else | |
998 | |
999 //FIXME (has little in common with the mmx2 version) | |
1000 for(y=0; y<BLOCK_SIZE; y++) | |
1001 { | |
1002 int a= src[1] - src[2]; | |
1003 int b= src[3] - src[4]; | |
1004 int c= src[5] - src[6]; | |
1005 | |
1006 int d= MAX(ABS(b) - (ABS(a) + ABS(c))/2, 0); | |
1007 | |
1008 if(d < QP) | |
1009 { | |
1010 int v = d * SIGN(-b); | |
1011 | |
1012 src[1] +=v/8; | |
1013 src[2] +=v/4; | |
1014 src[3] +=3*v/8; | |
1015 src[4] -=3*v/8; | |
1016 src[5] -=v/4; | |
1017 src[6] -=v/8; | |
1018 | |
1019 } | |
1020 src+=stride; | |
1021 } | |
1022 #endif | |
1023 } | |
1024 | |
96 | 1025 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1026 static inline void doVertDefFilter(uint8_t src[], int stride, int QP) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1027 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1028 #ifdef HAVE_MMX |
111 | 1029 src+= stride*4; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1030 //FIXME try pmul for *5 stuff |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1031 // src[0]=0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1032 asm volatile( |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1033 "pxor %%mm7, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1034 "leal (%0, %1), %%eax \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1035 "leal (%%eax, %1, 4), %%ebx \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1036 // 0 1 2 3 4 5 6 7 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1037 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ebx+%1 ebx+2%1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1038 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1039 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1040 "movq (%0), %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1041 "movq %%mm0, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1042 "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1043 "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1044 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1045 "movq (%%eax), %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1046 "movq %%mm2, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1047 "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1048 "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1049 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1050 "movq (%%eax, %1), %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1051 "movq %%mm4, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1052 "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1053 "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1054 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1055 "paddw %%mm0, %%mm0 \n\t" // 2L0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1056 "paddw %%mm1, %%mm1 \n\t" // 2H0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1057 "psubw %%mm4, %%mm2 \n\t" // L1 - L2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1058 "psubw %%mm5, %%mm3 \n\t" // H1 - H2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1059 "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1060 "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1061 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1062 "psllw $2, %%mm2 \n\t" // 4L1 - 4L2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1063 "psllw $2, %%mm3 \n\t" // 4H1 - 4H2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1064 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1065 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1066 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1067 "movq (%%eax, %1, 2), %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1068 "movq %%mm2, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1069 "punpcklbw %%mm7, %%mm2 \n\t" // L3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1070 "punpckhbw %%mm7, %%mm3 \n\t" // H3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1071 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1072 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1073 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1074 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1075 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1076 "movq %%mm0, temp0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1077 "movq %%mm1, temp1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1078 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1079 "movq (%0, %1, 4), %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1080 "movq %%mm0, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1081 "punpcklbw %%mm7, %%mm0 \n\t" // L4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1082 "punpckhbw %%mm7, %%mm1 \n\t" // H4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1083 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1084 "psubw %%mm0, %%mm2 \n\t" // L3 - L4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1085 "psubw %%mm1, %%mm3 \n\t" // H3 - H4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1086 "movq %%mm2, temp2 \n\t" // L3 - L4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1087 "movq %%mm3, temp3 \n\t" // H3 - H4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1088 "paddw %%mm4, %%mm4 \n\t" // 2L2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1089 "paddw %%mm5, %%mm5 \n\t" // 2H2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1090 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1091 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1092 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1093 "psllw $2, %%mm2 \n\t" // 4L3 - 4L4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1094 "psllw $2, %%mm3 \n\t" // 4H3 - 4H4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1095 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1096 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1097 //50 opcodes so far |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1098 "movq (%%ebx), %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1099 "movq %%mm2, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1100 "punpcklbw %%mm7, %%mm2 \n\t" // L5 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1101 "punpckhbw %%mm7, %%mm3 \n\t" // H5 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1102 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1103 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1104 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1105 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1106 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1107 "movq (%%ebx, %1), %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1108 "punpcklbw %%mm7, %%mm6 \n\t" // L6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1109 "psubw %%mm6, %%mm2 \n\t" // L5 - L6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1110 "movq (%%ebx, %1), %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1111 "punpckhbw %%mm7, %%mm6 \n\t" // H6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1112 "psubw %%mm6, %%mm3 \n\t" // H5 - H6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1113 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1114 "paddw %%mm0, %%mm0 \n\t" // 2L4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1115 "paddw %%mm1, %%mm1 \n\t" // 2H4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1116 "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1117 "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1118 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1119 "psllw $2, %%mm2 \n\t" // 4L5 - 4L6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1120 "psllw $2, %%mm3 \n\t" // 4H5 - 4H6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1121 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1122 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1123 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1124 "movq (%%ebx, %1, 2), %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1125 "movq %%mm2, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1126 "punpcklbw %%mm7, %%mm2 \n\t" // L7 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1127 "punpckhbw %%mm7, %%mm3 \n\t" // H7 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1128 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1129 "paddw %%mm2, %%mm2 \n\t" // 2L7 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1130 "paddw %%mm3, %%mm3 \n\t" // 2H7 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1131 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1132 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1133 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1134 "movq temp0, %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1135 "movq temp1, %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1136 //FIXME pxor, psubw, pmax for abs |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1137 "movq %%mm7, %%mm6 \n\t" // 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1138 "pcmpgtw %%mm0, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1139 "pxor %%mm6, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1140 "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1141 "movq %%mm7, %%mm6 \n\t" // 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1142 "pcmpgtw %%mm1, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1143 "pxor %%mm6, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1144 "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7| |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1145 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1146 "movq %%mm7, %%mm6 \n\t" // 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1147 "pcmpgtw %%mm2, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1148 "pxor %%mm6, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1149 "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3| |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1150 "movq %%mm7, %%mm6 \n\t" // 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1151 "pcmpgtw %%mm3, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1152 "pxor %%mm6, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1153 "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3| |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1154 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1155 #ifdef HAVE_MMX2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1156 "pminsw %%mm2, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1157 "pminsw %%mm3, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1158 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1159 "movq %%mm0, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1160 "psubusw %%mm2, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1161 "psubw %%mm6, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1162 "movq %%mm1, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1163 "psubusw %%mm3, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1164 "psubw %%mm6, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1165 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1166 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1167 "movq %%mm7, %%mm6 \n\t" // 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1168 "pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1169 "pxor %%mm6, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1170 "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5| |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1171 "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1172 "pxor %%mm7, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1173 "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5| |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1174 // 100 opcodes |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1175 "movd %2, %%mm2 \n\t" // QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1176 "punpcklwd %%mm2, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1177 "punpcklwd %%mm2, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1178 "psllw $3, %%mm2 \n\t" // 8QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1179 "movq %%mm2, %%mm3 \n\t" // 8QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1180 "pcmpgtw %%mm4, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1181 "pcmpgtw %%mm5, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1182 "pand %%mm2, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1183 "pand %%mm3, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1184 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1185 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1186 "psubusw %%mm0, %%mm4 \n\t" // hd |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1187 "psubusw %%mm1, %%mm5 \n\t" // ld |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1188 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1189 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1190 "movq w05, %%mm2 \n\t" // 5 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1191 "pmullw %%mm2, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1192 "pmullw %%mm2, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1193 "movq w20, %%mm2 \n\t" // 32 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1194 "paddw %%mm2, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1195 "paddw %%mm2, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1196 "psrlw $6, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1197 "psrlw $6, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1198 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1199 /* |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1200 "movq w06, %%mm2 \n\t" // 6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1201 "paddw %%mm2, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1202 "paddw %%mm2, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1203 "movq w1400, %%mm2 \n\t" // 1400h = 5120 = 5/64*2^16 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1204 //FIXME if *5/64 is supposed to be /13 then we should use 5041 instead of 5120 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1205 "pmulhw %%mm2, %%mm4 \n\t" // hd/13 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1206 "pmulhw %%mm2, %%mm5 \n\t" // ld/13 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1207 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1208 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1209 "movq temp2, %%mm0 \n\t" // L3 - L4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1210 "movq temp3, %%mm1 \n\t" // H3 - H4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1211 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1212 "pxor %%mm2, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1213 "pxor %%mm3, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1214 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1215 // FIXME rounding error |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1216 "psraw $1, %%mm0 \n\t" // (L3 - L4)/2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1217 "psraw $1, %%mm1 \n\t" // (H3 - H4)/2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1218 "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1219 "pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1220 "pxor %%mm2, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1221 "pxor %%mm3, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1222 "psubw %%mm2, %%mm0 \n\t" // |L3-L4| |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1223 "psubw %%mm3, %%mm1 \n\t" // |H3-H4| |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1224 // "psrlw $1, %%mm0 \n\t" // |L3 - L4|/2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1225 // "psrlw $1, %%mm1 \n\t" // |H3 - H4|/2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1226 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1227 "pxor %%mm6, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1228 "pxor %%mm7, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1229 "pand %%mm2, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1230 "pand %%mm3, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1231 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1232 #ifdef HAVE_MMX2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1233 "pminsw %%mm0, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1234 "pminsw %%mm1, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1235 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1236 "movq %%mm4, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1237 "psubusw %%mm0, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1238 "psubw %%mm2, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1239 "movq %%mm5, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1240 "psubusw %%mm1, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1241 "psubw %%mm2, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1242 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1243 "pxor %%mm6, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1244 "pxor %%mm7, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1245 "psubw %%mm6, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1246 "psubw %%mm7, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1247 "packsswb %%mm5, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1248 "movq (%%eax, %1, 2), %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1249 "paddb %%mm4, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1250 "movq %%mm0, (%%eax, %1, 2) \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1251 "movq (%0, %1, 4), %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1252 "psubb %%mm4, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1253 "movq %%mm0, (%0, %1, 4) \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1254 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1255 : |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1256 : "r" (src), "r" (stride), "r" (QP) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1257 : "%eax", "%ebx" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1258 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1259 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1260 const int l1= stride; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1261 const int l2= stride + l1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1262 const int l3= stride + l2; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1263 const int l4= stride + l3; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1264 const int l5= stride + l4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1265 const int l6= stride + l5; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1266 const int l7= stride + l6; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1267 const int l8= stride + l7; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1268 // const int l9= stride + l8; |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
1269 int x; |
111 | 1270 src+= stride*3; |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
1271 for(x=0; x<BLOCK_SIZE; x++) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1272 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1273 const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1274 if(ABS(middleEnergy) < 8*QP) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1275 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1276 const int q=(src[l4] - src[l5])/2; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1277 const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1278 const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1279 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1280 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1281 d= MAX(d, 0); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1282 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1283 d= (5*d + 32) >> 6; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1284 d*= SIGN(-middleEnergy); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1285 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1286 if(q>0) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1287 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1288 d= d<0 ? 0 : d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1289 d= d>q ? q : d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1290 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1291 else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1292 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1293 d= d>0 ? 0 : d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1294 d= d<q ? q : d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1295 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1296 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1297 src[l4]-= d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1298 src[l5]+= d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1299 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1300 src++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1301 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1302 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1303 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1304 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1305 //FIXME? |255-0| = 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1306 /** |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1307 * Check if the given 8x8 Block is mostly "flat" and copy the unaliged data into tempBlock. |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1308 */ |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
1309 static inline int isHorizDCAndCopy2Temp(uint8_t src[], int stride) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1310 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1311 // src++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1312 int numEq= 0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1313 #ifdef HAVE_MMX |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1314 asm volatile ( |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1315 // "int $3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1316 "pushl %1\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1317 "movq b7E, %%mm7 \n\t" // mm7 = 0x7F |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1318 "movq b7C, %%mm6 \n\t" // mm6 = 0x7D |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1319 "leal tempBlock, %%eax \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1320 "pxor %%mm0, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1321 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1322 #define HDC_CHECK_AND_CPY(i) \ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1323 "movq -4(%1), %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1324 "psrlq $32, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1325 "punpckldq 4(%1), %%mm2 \n\t" /* (%1) */\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1326 "movq %%mm2, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1327 "psrlq $8, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1328 "psubb %%mm1, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1329 "paddb %%mm7, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1330 "pcmpgtb %%mm6, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1331 "paddb %%mm2, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1332 "movq %%mm1," #i "(%%eax) \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1333 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1334 HDC_CHECK_AND_CPY(0) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1335 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1336 HDC_CHECK_AND_CPY(8) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1337 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1338 HDC_CHECK_AND_CPY(16) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1339 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1340 HDC_CHECK_AND_CPY(24) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1341 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1342 HDC_CHECK_AND_CPY(32) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1343 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1344 HDC_CHECK_AND_CPY(40) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1345 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1346 HDC_CHECK_AND_CPY(48) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1347 "addl %2, %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1348 HDC_CHECK_AND_CPY(56) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1349 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1350 "psllq $8, %%mm0 \n\t" // remove dummy value |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1351 "movq %%mm0, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1352 "psrlw $8, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1353 "paddb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1354 "movq %%mm0, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1355 "psrlq $16, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1356 "paddb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1357 "movq %%mm0, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1358 "psrlq $32, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1359 "paddb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1360 "popl %1\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1361 "movd %%mm0, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1362 : "=r" (numEq) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1363 : "r" (src), "r" (stride) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1364 : "%eax" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1365 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1366 // printf("%d\n", numEq); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1367 numEq= (256 - (numEq & 0xFF)) &0xFF; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1368 #else |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
1369 int y; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
1370 for(y=0; y<BLOCK_SIZE; y++) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1371 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1372 if(((src[0] - src[1] + 1) & 0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1373 if(((src[1] - src[2] + 1) & 0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1374 if(((src[2] - src[3] + 1) & 0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1375 if(((src[3] - src[4] + 1) & 0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1376 if(((src[4] - src[5] + 1) & 0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1377 if(((src[5] - src[6] + 1) & 0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1378 if(((src[6] - src[7] + 1) & 0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1379 tempBlock[0 + y*TEMP_STRIDE] = src[0]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1380 tempBlock[1 + y*TEMP_STRIDE] = src[1]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1381 tempBlock[2 + y*TEMP_STRIDE] = src[2]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1382 tempBlock[3 + y*TEMP_STRIDE] = src[3]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1383 tempBlock[4 + y*TEMP_STRIDE] = src[4]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1384 tempBlock[5 + y*TEMP_STRIDE] = src[5]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1385 tempBlock[6 + y*TEMP_STRIDE] = src[6]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1386 tempBlock[7 + y*TEMP_STRIDE] = src[7]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1387 src+= stride; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1388 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1389 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1390 /* if(abs(numEq - asmEq) > 0) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1391 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1392 // printf("\nasm:%d c:%d\n", asmEq, numEq); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1393 for(int y=0; y<8; y++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1394 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1395 for(int x=0; x<8; x++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1396 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1397 printf("%d ", src[x + y*stride]); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1398 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1399 printf("\n"); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1400 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1401 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1402 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1403 // printf("%d\n", numEq); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1404 return numEq > hFlatnessThreshold; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1405 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1406 |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
1407 static inline int isHorizMinMaxOk(uint8_t src[], int stride, int QP) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1408 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1409 #ifdef MMX_FIXME |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1410 FIXME |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1411 int isOk; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1412 asm volatile( |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1413 // "int $3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1414 "movq (%1, %2), %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1415 "movq (%1, %2, 8), %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1416 "movq %%mm0, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1417 "psubusb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1418 "psubusb %%mm2, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1419 "por %%mm1, %%mm0 \n\t" // ABS Diff |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1420 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1421 "movq pQPb, %%mm7 \n\t" // QP,..., QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1422 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1423 "psubusb %%mm7, %%mm0 \n\t" // Diff <= 2QP -> 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1424 "pcmpeqd b00, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1425 "psrlq $16, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1426 "pcmpeqd bFF, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1427 // "movd %%mm0, (%1, %2, 4)\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1428 "movd %%mm0, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1429 : "=r" (isOk) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1430 : "r" (src), "r" (stride) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1431 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1432 return isOk; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1433 #else |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
1434 if(abs(src[0] - src[7]) > 2*QP) return 0; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1435 |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
1436 return 1; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1437 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1438 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1439 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1440 static inline void doHorizDefFilterAndCopyBack(uint8_t dst[], int stride, int QP) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1441 { |
96 | 1442 #ifdef HAVE_MMX |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1443 asm volatile( |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1444 "pushl %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1445 "pxor %%mm7, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1446 "movq bm00001000, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1447 "movd %2, %%mm5 \n\t" // QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1448 "movq %%mm5, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1449 "paddusb %%mm5, %%mm5 \n\t" // 2QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1450 "paddusb %%mm5, %%mm4 \n\t" // 3QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1451 "psllq $24, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1452 "pxor %%mm5, %%mm5 \n\t" // 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1453 "psubb %%mm4, %%mm5 \n\t" // -QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1454 "leal tempBlock, %%eax \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1455 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1456 //FIXME? "unroll by 2" and mix |
96 | 1457 #ifdef HAVE_MMX2 |
1458 #define HDF(i) \ | |
1459 "movq " #i "(%%eax), %%mm0 \n\t"\ | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1460 "movq %%mm0, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1461 "movq %%mm0, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1462 "psrlq $8, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1463 "psubusb %%mm1, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1464 "psubusb %%mm0, %%mm1 \n\t"\ |
96 | 1465 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ |
1466 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
1467 "pshufw $0x00, %%mm1, %%mm3 \n\t" /* p´5 = |p1 - p2| */\ | |
1468 "pminub %%mm1, %%mm3 \n\t" /* p´5 = min(|p2-p1|, |p6-p5|)*/\ | |
1469 "psrlq $16, %%mm3 \n\t" /* p´3 = min(|p2-p1|, |p6-p5|)*/\ | |
1470 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-min(|p1-p2|,|p5-p6|) */\ | |
1471 "paddb %%mm5, %%mm1 \n\t"\ | |
1472 "psubusb %%mm5, %%mm1 \n\t"\ | |
1473 "psrlw $2, %%mm1 \n\t"\ | |
1474 "pxor %%mm2, %%mm1 \n\t"\ | |
1475 "psubb %%mm2, %%mm1 \n\t"\ | |
1476 "pand %%mm6, %%mm1 \n\t"\ | |
1477 "psubb %%mm1, %%mm0 \n\t"\ | |
1478 "psllq $8, %%mm1 \n\t"\ | |
1479 "paddb %%mm1, %%mm0 \n\t"\ | |
1480 "movd %%mm0, (%0) \n\t"\ | |
1481 "psrlq $32, %%mm0 \n\t"\ | |
1482 "movd %%mm0, 4(%0) \n\t" | |
1483 #else | |
1484 #define HDF(i)\ | |
1485 "movq " #i "(%%eax), %%mm0 \n\t"\ | |
1486 "movq %%mm0, %%mm1 \n\t"\ | |
1487 "movq %%mm0, %%mm2 \n\t"\ | |
1488 "psrlq $8, %%mm1 \n\t"\ | |
1489 "psubusb %%mm1, %%mm2 \n\t"\ | |
1490 "psubusb %%mm0, %%mm1 \n\t"\ | |
1491 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ | |
1492 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
1493 "movq %%mm1, %%mm3 \n\t"\ | |
1494 "psllq $32, %%mm3 \n\t"\ | |
1495 "movq %%mm3, %%mm4 \n\t"\ | |
1496 "psubusb %%mm1, %%mm4 \n\t"\ | |
1497 "psubb %%mm4, %%mm3 \n\t"\ | |
1498 "psrlq $16, %%mm3 \n\t" /* p´3 = min(|p2-p1|, |p6-p5|)*/\ | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1499 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-min(|p1-p2|,|p5,ü6|) */\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1500 "paddb %%mm5, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1501 "psubusb %%mm5, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1502 "psrlw $2, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1503 "pxor %%mm2, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1504 "psubb %%mm2, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1505 "pand %%mm6, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1506 "psubb %%mm1, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1507 "psllq $8, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1508 "paddb %%mm1, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1509 "movd %%mm0, (%0) \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1510 "psrlq $32, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1511 "movd %%mm0, 4(%0) \n\t" |
96 | 1512 #endif |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1513 HDF(0) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1514 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1515 HDF(8) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1516 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1517 HDF(16) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1518 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1519 HDF(24) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1520 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1521 HDF(32) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1522 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1523 HDF(40) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1524 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1525 HDF(48) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1526 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1527 HDF(56) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1528 "popl %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1529 : |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1530 : "r" (dst), "r" (stride), "r" (QP) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1531 : "%eax" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1532 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1533 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1534 uint8_t *src= tempBlock; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1535 |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
1536 int y; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
1537 for(y=0; y<BLOCK_SIZE; y++) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1538 { |
105 | 1539 const int middleEnergy= 5*(src[4] - src[5]) + 2*(src[2] - src[5]); |
1540 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1541 dst[0] = src[0]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1542 dst[1] = src[1]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1543 dst[2] = src[2]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1544 dst[3] = src[3]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1545 dst[4] = src[4]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1546 dst[5] = src[5]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1547 dst[6] = src[6]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1548 dst[7] = src[7]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1549 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1550 if(ABS(middleEnergy) < 8*QP) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1551 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1552 const int q=(src[3] - src[4])/2; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1553 const int leftEnergy= 5*(src[2] - src[1]) + 2*(src[0] - src[3]); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1554 const int rightEnergy= 5*(src[6] - src[5]) + 2*(src[4] - src[7]); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1555 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1556 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1557 d= MAX(d, 0); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1558 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1559 d= (5*d + 32) >> 6; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1560 d*= SIGN(-middleEnergy); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1561 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1562 if(q>0) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1563 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1564 d= d<0 ? 0 : d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1565 d= d>q ? q : d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1566 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1567 else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1568 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1569 d= d>0 ? 0 : d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1570 d= d<q ? q : d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1571 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1572 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1573 dst[3]-= d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1574 dst[4]+= d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1575 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1576 dst+= stride; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1577 src+= TEMP_STRIDE; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1578 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1579 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1580 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1581 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1582 /** |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1583 * Do a horizontal low pass filter on the 10x8 block (dst points to middle 8x8 Block) |
107 | 1584 * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 (C version) |
1585 * using the 7-Tap Filter (2,2,2,4,2,2,2)/16 (MMX2/3DNOW version) | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1586 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1587 static inline void doHorizLowPassAndCopyBack(uint8_t dst[], int stride, int QP) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1588 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1589 //return; |
96 | 1590 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1591 asm volatile( //"movv %0 %1 %2\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1592 "pushl %0\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1593 "pxor %%mm7, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1594 "leal tempBlock, %%eax \n\t" |
96 | 1595 /* |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1596 #define HLP1 "movq (%0), %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1597 "movq %%mm0, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1598 "psllq $8, %%mm0 \n\t"\ |
96 | 1599 PAVGB(%%mm1, %%mm0)\ |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1600 "psrlw $8, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1601 "pxor %%mm1, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1602 "packuswb %%mm1, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1603 "movq %%mm0, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1604 "movq %%mm0, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1605 "psllq $32, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1606 "paddb %%mm0, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1607 "psllq $16, %%mm2 \n\t"\ |
96 | 1608 PAVGB(%%mm2, %%mm0)\ |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1609 "movq %%mm0, %%mm3 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1610 "pand bm11001100, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1611 "paddusb %%mm0, %%mm3 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1612 "psrlq $8, %%mm3 \n\t"\ |
96 | 1613 PAVGB(%%mm1, %%mm4)\ |
1614 PAVGB(%%mm3, %%mm2)\ | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1615 "psrlq $16, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1616 "punpcklbw %%mm2, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1617 "movq %%mm2, (%0) \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1618 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1619 #define HLP2 "movq (%0), %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1620 "movq %%mm0, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1621 "psllq $8, %%mm0 \n\t"\ |
96 | 1622 PAVGB(%%mm1, %%mm0)\ |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1623 "psrlw $8, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1624 "pxor %%mm1, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1625 "packuswb %%mm1, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1626 "movq %%mm0, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1627 "psllq $32, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1628 "psllq $16, %%mm2 \n\t"\ |
96 | 1629 PAVGB(%%mm2, %%mm0)\ |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1630 "movq %%mm0, %%mm3 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1631 "pand bm11001100, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1632 "paddusb %%mm0, %%mm3 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1633 "psrlq $8, %%mm3 \n\t"\ |
96 | 1634 PAVGB(%%mm3, %%mm2)\ |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1635 "psrlq $16, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1636 "punpcklbw %%mm2, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1637 "movq %%mm2, (%0) \n\t"\ |
96 | 1638 */ |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1639 // approximately a 7-Tap Filter with Vector (1,2,3,4,3,2,1)/16 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1640 /* |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1641 Implemented Exact 7-Tap |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1642 9421 A321 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1643 36421 64321 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1644 334321 = |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1645 1234321 = |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1646 1234321 = |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1647 123433 = |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1648 12463 12346 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1649 1249 123A |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1650 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1651 */ |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1652 |
96 | 1653 #ifdef HAVE_MMX2 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1654 #define HLP3(i) "movq " #i "(%%eax), %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1655 "movq %%mm0, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1656 "movq %%mm0, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1657 "movq %%mm0, %%mm3 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1658 "movq %%mm0, %%mm4 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1659 "psllq $8, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1660 "psrlq $8, %%mm2 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1661 "pand bm00000001, %%mm3 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1662 "pand bm10000000, %%mm4 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1663 "por %%mm3, %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1664 "por %%mm4, %%mm2 \n\t"\ |
96 | 1665 PAVGB(%%mm2, %%mm1)\ |
1666 PAVGB(%%mm1, %%mm0)\ | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1667 \ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1668 "pshufw $0xF9, %%mm0, %%mm3 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1669 "pshufw $0x90, %%mm0, %%mm4 \n\t"\ |
96 | 1670 PAVGB(%%mm3, %%mm4)\ |
1671 PAVGB(%%mm4, %%mm0)\ | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1672 "movd %%mm0, (%0) \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1673 "psrlq $32, %%mm0 \n\t"\ |
96 | 1674 "movd %%mm0, 4(%0) \n\t" |
1675 #else | |
1676 #define HLP3(i) "movq " #i "(%%eax), %%mm0 \n\t"\ | |
1677 "movq %%mm0, %%mm1 \n\t"\ | |
1678 "movq %%mm0, %%mm2 \n\t"\ | |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1679 "movd -4(%0), %%mm3 \n\t" /*0001000*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1680 "movd 8(%0), %%mm4 \n\t" /*0001000*/\ |
96 | 1681 "psllq $8, %%mm1 \n\t"\ |
1682 "psrlq $8, %%mm2 \n\t"\ | |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1683 "psrlq $24, %%mm3 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1684 "psllq $56, %%mm4 \n\t"\ |
96 | 1685 "por %%mm3, %%mm1 \n\t"\ |
1686 "por %%mm4, %%mm2 \n\t"\ | |
1687 PAVGB(%%mm2, %%mm1)\ | |
1688 PAVGB(%%mm1, %%mm0)\ | |
1689 \ | |
1690 "movq %%mm0, %%mm3 \n\t"\ | |
1691 "movq %%mm0, %%mm4 \n\t"\ | |
1692 "movq %%mm0, %%mm5 \n\t"\ | |
1693 "psrlq $16, %%mm3 \n\t"\ | |
1694 "psllq $16, %%mm4 \n\t"\ | |
1695 "pand bm11000000, %%mm5 \n\t"\ | |
1696 "por %%mm5, %%mm3 \n\t"\ | |
1697 "movq %%mm0, %%mm5 \n\t"\ | |
1698 "pand bm00000011, %%mm5 \n\t"\ | |
1699 "por %%mm5, %%mm4 \n\t"\ | |
1700 PAVGB(%%mm3, %%mm4)\ | |
1701 PAVGB(%%mm4, %%mm0)\ | |
1702 "movd %%mm0, (%0) \n\t"\ | |
1703 "psrlq $32, %%mm0 \n\t"\ | |
1704 "movd %%mm0, 4(%0) \n\t" | |
1705 #endif | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1706 |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1707 /* uses the 7-Tap Filter: 1112111 */ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1708 #define NEW_HLP(i)\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1709 "movq " #i "(%%eax), %%mm0 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1710 "movq %%mm0, %%mm1 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1711 "movq %%mm0, %%mm2 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1712 "movd -4(%0), %%mm3 \n\t" /*0001000*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1713 "movd 8(%0), %%mm4 \n\t" /*0001000*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1714 "psllq $8, %%mm1 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1715 "psrlq $8, %%mm2 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1716 "psrlq $24, %%mm3 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1717 "psllq $56, %%mm4 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1718 "por %%mm3, %%mm1 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1719 "por %%mm4, %%mm2 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1720 "movq %%mm1, %%mm5 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1721 PAVGB(%%mm2, %%mm1)\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1722 PAVGB(%%mm1, %%mm0)\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1723 "psllq $8, %%mm5 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1724 "psrlq $8, %%mm2 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1725 "por %%mm3, %%mm5 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1726 "por %%mm4, %%mm2 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1727 "movq %%mm5, %%mm1 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1728 PAVGB(%%mm2, %%mm5)\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1729 "psllq $8, %%mm1 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1730 "psrlq $8, %%mm2 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1731 "por %%mm3, %%mm1 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1732 "por %%mm4, %%mm2 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1733 PAVGB(%%mm2, %%mm1)\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1734 PAVGB(%%mm1, %%mm5)\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1735 PAVGB(%%mm5, %%mm0)\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1736 "movd %%mm0, (%0) \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1737 "psrlq $32, %%mm0 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1738 "movd %%mm0, 4(%0) \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1739 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1740 /* uses the 9-Tap Filter: 112242211 */ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1741 #define NEW_HLP2(i)\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1742 "movq " #i "(%%eax), %%mm0 \n\t" /*0001000*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1743 "movq %%mm0, %%mm1 \n\t" /*0001000*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1744 "movq %%mm0, %%mm2 \n\t" /*0001000*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1745 "movd -4(%0), %%mm3 \n\t" /*0001000*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1746 "movd 8(%0), %%mm4 \n\t" /*0001000*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1747 "psllq $8, %%mm1 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1748 "psrlq $8, %%mm2 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1749 "psrlq $24, %%mm3 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1750 "psllq $56, %%mm4 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1751 "por %%mm3, %%mm1 \n\t" /*0010000*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1752 "por %%mm4, %%mm2 \n\t" /*0000100*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1753 "movq %%mm1, %%mm5 \n\t" /*0010000*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1754 PAVGB(%%mm2, %%mm1) /*0010100*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1755 PAVGB(%%mm1, %%mm0) /*0012100*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1756 "psllq $8, %%mm5 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1757 "psrlq $8, %%mm2 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1758 "por %%mm3, %%mm5 \n\t" /*0100000*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1759 "por %%mm4, %%mm2 \n\t" /*0000010*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1760 "movq %%mm5, %%mm1 \n\t" /*0100000*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1761 PAVGB(%%mm2, %%mm5) /*0100010*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1762 "psllq $8, %%mm1 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1763 "psrlq $8, %%mm2 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1764 "por %%mm3, %%mm1 \n\t" /*1000000*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1765 "por %%mm4, %%mm2 \n\t" /*0000001*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1766 "movq %%mm1, %%mm6 \n\t" /*1000000*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1767 PAVGB(%%mm2, %%mm1) /*1000001*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1768 "psllq $8, %%mm6 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1769 "psrlq $8, %%mm2 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1770 "por %%mm3, %%mm6 \n\t"/*100000000*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1771 "por %%mm4, %%mm2 \n\t"/*000000001*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1772 PAVGB(%%mm2, %%mm6) /*100000001*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1773 PAVGB(%%mm6, %%mm1) /*110000011*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1774 PAVGB(%%mm1, %%mm5) /*112000211*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1775 PAVGB(%%mm5, %%mm0) /*112242211*/\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1776 "movd %%mm0, (%0) \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1777 "psrlq $32, %%mm0 \n\t"\ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1778 "movd %%mm0, 4(%0) \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1779 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1780 #define HLP(i) NEW_HLP(i) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1781 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1782 HLP(0) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1783 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1784 HLP(8) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1785 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1786 HLP(16) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1787 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1788 HLP(24) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1789 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1790 HLP(32) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1791 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1792 HLP(40) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1793 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1794 HLP(48) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1795 "addl %1, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1796 HLP(56) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1797 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1798 "popl %0\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1799 : |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1800 : "r" (dst), "r" (stride) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1801 : "%eax", "%ebx" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1802 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1803 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1804 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1805 uint8_t *temp= tempBlock; |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
1806 int y; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
1807 for(y=0; y<BLOCK_SIZE; y++) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1808 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1809 const int first= ABS(dst[-1] - dst[0]) < QP ? dst[-1] : dst[0]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1810 const int last= ABS(dst[8] - dst[7]) < QP ? dst[8] : dst[7]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1811 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1812 int sums[9]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1813 sums[0] = first + temp[0]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1814 sums[1] = temp[0] + temp[1]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1815 sums[2] = temp[1] + temp[2]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1816 sums[3] = temp[2] + temp[3]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1817 sums[4] = temp[3] + temp[4]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1818 sums[5] = temp[4] + temp[5]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1819 sums[6] = temp[5] + temp[6]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1820 sums[7] = temp[6] + temp[7]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1821 sums[8] = temp[7] + last; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1822 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1823 dst[0]= ((sums[0]<<2) + ((first + sums[2])<<1) + sums[4] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1824 dst[1]= ((dst[1]<<2) + (first + sums[0] + sums[3]<<1) + sums[5] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1825 dst[2]= ((dst[2]<<2) + (first + sums[1] + sums[4]<<1) + sums[6] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1826 dst[3]= ((dst[3]<<2) + (sums[2] + sums[5]<<1) + sums[0] + sums[7] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1827 dst[4]= ((dst[4]<<2) + (sums[3] + sums[6]<<1) + sums[1] + sums[8] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1828 dst[5]= ((dst[5]<<2) + (last + sums[7] + sums[4]<<1) + sums[2] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1829 dst[6]= ((last + dst[6]<<2) + (dst[7] + sums[5]<<1) + sums[3] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1830 dst[7]= ((sums[8]<<2) + (last + sums[6]<<1) + sums[4] + 8)>>4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1831 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1832 dst+= stride; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1833 temp+= TEMP_STRIDE; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1834 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1835 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1836 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1837 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1838 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1839 static inline void dering(uint8_t src[], int stride, int QP) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1840 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1841 //FIXME |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1842 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1843 #ifdef HAVE_MMX2X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1844 asm volatile( |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1845 "leal (%0, %1), %%eax \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1846 "leal (%%eax, %1, 4), %%ebx \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1847 // 0 1 2 3 4 5 6 7 8 9 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1848 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1849 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1850 "pcmpeq %%mm6, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1851 "pxor %%mm7, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1852 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1853 #define FIND_MIN_MAX(addr)\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1854 "movq (" #addr "), %%mm0, \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1855 "pminub %%mm0, %%mm6 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1856 "pmaxub %%mm0, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1857 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1858 FIND_MIN_MAX(%0) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1859 FIND_MIN_MAX(%%eax) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1860 FIND_MIN_MAX(%%eax, %1) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1861 FIND_MIN_MAX(%%eax, %1, 2) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1862 FIND_MIN_MAX(%0, %1, 4) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1863 FIND_MIN_MAX(%%ebx) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1864 FIND_MIN_MAX(%%ebx, %1) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1865 FIND_MIN_MAX(%%ebx, %1, 2) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1866 FIND_MIN_MAX(%0, %1, 8) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1867 FIND_MIN_MAX(%%ebx, %1, 2) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1868 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1869 "movq %%mm6, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1870 "psrlq $32, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1871 "pminub %%mm4, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1872 "movq %%mm6, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1873 "psrlq $16, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1874 "pminub %%mm4, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1875 "movq %%mm6, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1876 "psrlq $8, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1877 "pminub %%mm4, %%mm6 \n\t" // min of pixels |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1878 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1879 "movq %%mm7, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1880 "psrlq $32, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1881 "pmaxub %%mm4, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1882 "movq %%mm7, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1883 "psrlq $16, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1884 "pmaxub %%mm4, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1885 "movq %%mm7, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1886 "psrlq $8, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1887 "pmaxub %%mm4, %%mm7 \n\t" // max of pixels |
96 | 1888 PAVGB(%%mm6, %%mm7) // (max + min)/2 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1889 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1890 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1891 : : "r" (src), "r" (stride), "r" (QP) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1892 : "%eax", "%ebx" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1893 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1894 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1895 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1896 //FIXME |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1897 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1898 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1899 |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1900 /** |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1901 * Deinterlaces the given block |
111 | 1902 * will be called for every 8x8 block, and can read & write into an 8x16 block |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1903 */ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1904 static inline void deInterlaceInterpolateLinear(uint8_t src[], int stride) |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1905 { |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1906 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1907 asm volatile( |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1908 "leal (%0, %1), %%eax \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1909 "leal (%%eax, %1, 4), %%ebx \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1910 // 0 1 2 3 4 5 6 7 8 9 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1911 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1912 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1913 "movq (%0), %%mm0 \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1914 "movq (%%eax, %1), %%mm1 \n\t" |
111 | 1915 PAVGB(%%mm1, %%mm0) |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1916 "movq %%mm0, (%%eax) \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1917 "movq (%0, %1, 4), %%mm0 \n\t" |
111 | 1918 PAVGB(%%mm0, %%mm1) |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1919 "movq %%mm1, (%%eax, %1, 2) \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1920 "movq (%%ebx, %1), %%mm1 \n\t" |
111 | 1921 PAVGB(%%mm1, %%mm0) |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1922 "movq %%mm0, (%%ebx) \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1923 "movq (%0, %1, 8), %%mm0 \n\t" |
111 | 1924 PAVGB(%%mm0, %%mm1) |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1925 "movq %%mm1, (%%ebx, %1, 2) \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1926 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1927 : : "r" (src), "r" (stride) |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1928 : "%eax", "%ebx" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1929 ); |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1930 #else |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1931 int x; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1932 for(x=0; x<8; x++) |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1933 { |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1934 src[stride] = (src[0] + src[stride*2])>>1; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1935 src[stride*3] = (src[stride*2] + src[stride*4])>>1; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1936 src[stride*5] = (src[stride*4] + src[stride*6])>>1; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1937 src[stride*7] = (src[stride*6] + src[stride*8])>>1; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1938 src++; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1939 } |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1940 #endif |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1941 } |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1942 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1943 /** |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1944 * Deinterlaces the given block |
111 | 1945 * will be called for every 8x8 block, and can read & write into an 8x16 block |
1946 * no cliping in C version | |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1947 */ |
111 | 1948 static inline void deInterlaceInterpolateCubic(uint8_t src[], int stride) |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1949 { |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1950 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1951 asm volatile( |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1952 "leal (%0, %1), %%eax \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1953 "leal (%%eax, %1, 4), %%ebx \n\t" |
111 | 1954 "leal (%%ebx, %1, 4), %%ecx \n\t" |
1955 "addl %1, %%ecx \n\t" | |
1956 "pxor %%mm7, %%mm7 \n\t" | |
1957 // 0 1 2 3 4 5 6 7 8 9 10 | |
1958 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 ecx | |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1959 |
111 | 1960 #define DEINT_CUBIC(a,b,c,d,e)\ |
1961 "movq " #a ", %%mm0 \n\t"\ | |
1962 "movq " #b ", %%mm1 \n\t"\ | |
1963 "movq " #d ", %%mm2 \n\t"\ | |
1964 "movq " #e ", %%mm3 \n\t"\ | |
1965 PAVGB(%%mm2, %%mm1) /* (b+d) /2 */\ | |
1966 PAVGB(%%mm3, %%mm0) /* a(a+e) /2 */\ | |
1967 "movq %%mm0, %%mm2 \n\t"\ | |
1968 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
1969 "punpckhbw %%mm7, %%mm2 \n\t"\ | |
1970 "movq %%mm1, %%mm3 \n\t"\ | |
1971 "punpcklbw %%mm7, %%mm1 \n\t"\ | |
1972 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
1973 "psubw %%mm1, %%mm0 \n\t" /* L(a+e - (b+d))/2 */\ | |
1974 "psubw %%mm3, %%mm2 \n\t" /* H(a+e - (b+d))/2 */\ | |
1975 "psraw $3, %%mm0 \n\t" /* L(a+e - (b+d))/16 */\ | |
1976 "psraw $3, %%mm2 \n\t" /* H(a+e - (b+d))/16 */\ | |
1977 "psubw %%mm0, %%mm1 \n\t" /* L(9b + 9d - a - e)/16 */\ | |
1978 "psubw %%mm2, %%mm3 \n\t" /* H(9b + 9d - a - e)/16 */\ | |
1979 "packuswb %%mm3, %%mm1 \n\t"\ | |
1980 "movq %%mm1, " #c " \n\t" | |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1981 |
111 | 1982 DEINT_CUBIC((%0), (%%eax, %1), (%%eax, %1, 2), (%0, %1, 4), (%%ebx, %1)) |
1983 DEINT_CUBIC((%%eax, %1), (%0, %1, 4), (%%ebx), (%%ebx, %1), (%0, %1, 8)) | |
1984 DEINT_CUBIC((%0, %1, 4), (%%ebx, %1), (%%ebx, %1, 2), (%0, %1, 8), (%%ecx)) | |
1985 DEINT_CUBIC((%%ebx, %1), (%0, %1, 8), (%%ebx, %1, 4), (%%ecx), (%%ecx, %1, 2)) | |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1986 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1987 : : "r" (src), "r" (stride) |
111 | 1988 : "%eax", "%ebx", "ecx" |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1989 ); |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1990 #else |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1991 int x; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1992 for(x=0; x<8; x++) |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1993 { |
111 | 1994 src[stride*3] = (-src[0] + 9*src[stride*2] + 9*src[stride*4] - src[stride*6])>>4; |
1995 src[stride*5] = (-src[stride*2] + 9*src[stride*4] + 9*src[stride*6] - src[stride*8])>>4; | |
1996 src[stride*7] = (-src[stride*4] + 9*src[stride*6] + 9*src[stride*8] - src[stride*10])>>4; | |
1997 src[stride*9] = (-src[stride*6] + 9*src[stride*8] + 9*src[stride*10] - src[stride*12])>>4; | |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1998 src++; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1999 } |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2000 #endif |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2001 } |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2002 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2003 /** |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2004 * Deinterlaces the given block |
111 | 2005 * will be called for every 8x8 block, and can read & write into an 8x16 block |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2006 * will shift the image up by 1 line (FIXME if this is a problem) |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2007 */ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2008 static inline void deInterlaceBlendLinear(uint8_t src[], int stride) |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2009 { |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2010 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2011 asm volatile( |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2012 "leal (%0, %1), %%eax \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2013 "leal (%%eax, %1, 4), %%ebx \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2014 // 0 1 2 3 4 5 6 7 8 9 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2015 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2016 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2017 "movq (%0), %%mm0 \n\t" // L0 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2018 "movq (%%eax, %1), %%mm1 \n\t" // L2 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2019 PAVGB(%%mm1, %%mm0) // L0+L2 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2020 "movq (%%eax), %%mm2 \n\t" // L1 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2021 PAVGB(%%mm2, %%mm0) |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2022 "movq %%mm0, (%0) \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2023 "movq (%%eax, %1, 2), %%mm0 \n\t" // L3 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2024 PAVGB(%%mm0, %%mm2) // L1+L3 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2025 PAVGB(%%mm1, %%mm2) // 2L2 + L1 + L3 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2026 "movq %%mm2, (%%eax) \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2027 "movq (%0, %1, 4), %%mm2 \n\t" // L4 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2028 PAVGB(%%mm2, %%mm1) // L2+L4 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2029 PAVGB(%%mm0, %%mm1) // 2L3 + L2 + L4 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2030 "movq %%mm1, (%%eax, %1) \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2031 "movq (%%ebx), %%mm1 \n\t" // L5 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2032 PAVGB(%%mm1, %%mm0) // L3+L5 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2033 PAVGB(%%mm2, %%mm0) // 2L4 + L3 + L5 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2034 "movq %%mm0, (%%eax, %1, 2) \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2035 "movq (%%ebx, %1), %%mm0 \n\t" // L6 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2036 PAVGB(%%mm0, %%mm2) // L4+L6 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2037 PAVGB(%%mm1, %%mm2) // 2L5 + L4 + L6 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2038 "movq %%mm2, (%0, %1, 4) \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2039 "movq (%%ebx, %1, 2), %%mm2 \n\t" // L7 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2040 PAVGB(%%mm2, %%mm1) // L5+L7 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2041 PAVGB(%%mm0, %%mm1) // 2L6 + L5 + L7 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2042 "movq %%mm1, (%%ebx) \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2043 "movq (%0, %1, 8), %%mm1 \n\t" // L8 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2044 PAVGB(%%mm1, %%mm0) // L6+L8 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2045 PAVGB(%%mm2, %%mm0) // 2L7 + L6 + L8 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2046 "movq %%mm0, (%%ebx, %1) \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2047 "movq (%%ebx, %1, 4), %%mm0 \n\t" // L9 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2048 PAVGB(%%mm0, %%mm2) // L7+L9 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2049 PAVGB(%%mm1, %%mm2) // 2L8 + L7 + L9 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2050 "movq %%mm2, (%%ebx, %1, 2) \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2051 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2052 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2053 : : "r" (src), "r" (stride) |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2054 : "%eax", "%ebx" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2055 ); |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2056 #else |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2057 int x; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2058 for(x=0; x<8; x++) |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2059 { |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2060 src[0 ] = (src[0 ] + 2*src[stride ] + src[stride*2])>>2; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2061 src[stride ] = (src[stride ] + 2*src[stride*2] + src[stride*3])>>2; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2062 src[stride*2] = (src[stride*2] + 2*src[stride*3] + src[stride*4])>>2; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2063 src[stride*3] = (src[stride*3] + 2*src[stride*4] + src[stride*5])>>2; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2064 src[stride*4] = (src[stride*4] + 2*src[stride*5] + src[stride*6])>>2; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2065 src[stride*5] = (src[stride*5] + 2*src[stride*6] + src[stride*7])>>2; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2066 src[stride*6] = (src[stride*6] + 2*src[stride*7] + src[stride*8])>>2; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2067 src[stride*7] = (src[stride*7] + 2*src[stride*8] + src[stride*9])>>2; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2068 src++; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2069 } |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2070 #endif |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2071 } |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2072 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2073 /** |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2074 * Deinterlaces the given block |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2075 * will be called for every 8x8 block, except the last row, and can read & write into an 8x16 block |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2076 */ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2077 static inline void deInterlaceMedian(uint8_t src[], int stride) |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2078 { |
107 | 2079 #ifdef HAVE_MMX |
2080 #ifdef HAVE_MMX2 | |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2081 asm volatile( |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2082 "leal (%0, %1), %%eax \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2083 "leal (%%eax, %1, 4), %%ebx \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2084 // 0 1 2 3 4 5 6 7 8 9 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2085 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2086 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2087 "movq (%0), %%mm0 \n\t" // |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2088 "movq (%%eax, %1), %%mm2 \n\t" // |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2089 "movq (%%eax), %%mm1 \n\t" // |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2090 "movq %%mm0, %%mm3 \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2091 "pmaxub %%mm1, %%mm0 \n\t" // |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2092 "pminub %%mm3, %%mm1 \n\t" // |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2093 "pmaxub %%mm2, %%mm1 \n\t" // |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2094 "pminub %%mm1, %%mm0 \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2095 "movq %%mm0, (%%eax) \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2096 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2097 "movq (%0, %1, 4), %%mm0 \n\t" // |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2098 "movq (%%eax, %1, 2), %%mm1 \n\t" // |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2099 "movq %%mm2, %%mm3 \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2100 "pmaxub %%mm1, %%mm2 \n\t" // |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2101 "pminub %%mm3, %%mm1 \n\t" // |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2102 "pmaxub %%mm0, %%mm1 \n\t" // |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2103 "pminub %%mm1, %%mm2 \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2104 "movq %%mm2, (%%eax, %1, 2) \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2105 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2106 "movq (%%ebx), %%mm2 \n\t" // |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2107 "movq (%%ebx, %1), %%mm1 \n\t" // |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2108 "movq %%mm2, %%mm3 \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2109 "pmaxub %%mm0, %%mm2 \n\t" // |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2110 "pminub %%mm3, %%mm0 \n\t" // |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2111 "pmaxub %%mm1, %%mm0 \n\t" // |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2112 "pminub %%mm0, %%mm2 \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2113 "movq %%mm2, (%%ebx) \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2114 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2115 "movq (%%ebx, %1, 2), %%mm2 \n\t" // |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2116 "movq (%0, %1, 8), %%mm0 \n\t" // |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2117 "movq %%mm2, %%mm3 \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2118 "pmaxub %%mm0, %%mm2 \n\t" // |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2119 "pminub %%mm3, %%mm0 \n\t" // |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2120 "pmaxub %%mm1, %%mm0 \n\t" // |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2121 "pminub %%mm0, %%mm2 \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2122 "movq %%mm2, (%%ebx, %1, 2) \n\t" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2123 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2124 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2125 : : "r" (src), "r" (stride) |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2126 : "%eax", "%ebx" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2127 ); |
107 | 2128 |
2129 #else // MMX without MMX2 | |
2130 asm volatile( | |
2131 "leal (%0, %1), %%eax \n\t" | |
2132 "leal (%%eax, %1, 4), %%ebx \n\t" | |
2133 // 0 1 2 3 4 5 6 7 8 9 | |
2134 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 | |
2135 "pxor %%mm7, %%mm7 \n\t" | |
2136 | |
2137 #define MEDIAN(a,b,c)\ | |
2138 "movq " #a ", %%mm0 \n\t"\ | |
2139 "movq " #b ", %%mm2 \n\t"\ | |
2140 "movq " #c ", %%mm1 \n\t"\ | |
2141 "movq %%mm0, %%mm3 \n\t"\ | |
2142 "movq %%mm1, %%mm4 \n\t"\ | |
2143 "movq %%mm2, %%mm5 \n\t"\ | |
2144 "psubusb %%mm1, %%mm3 \n\t"\ | |
2145 "psubusb %%mm2, %%mm4 \n\t"\ | |
2146 "psubusb %%mm0, %%mm5 \n\t"\ | |
2147 "pcmpeqb %%mm7, %%mm3 \n\t"\ | |
2148 "pcmpeqb %%mm7, %%mm4 \n\t"\ | |
2149 "pcmpeqb %%mm7, %%mm5 \n\t"\ | |
2150 "movq %%mm3, %%mm6 \n\t"\ | |
2151 "pxor %%mm4, %%mm3 \n\t"\ | |
2152 "pxor %%mm5, %%mm4 \n\t"\ | |
2153 "pxor %%mm6, %%mm5 \n\t"\ | |
2154 "por %%mm3, %%mm1 \n\t"\ | |
2155 "por %%mm4, %%mm2 \n\t"\ | |
2156 "por %%mm5, %%mm0 \n\t"\ | |
2157 "pand %%mm2, %%mm0 \n\t"\ | |
2158 "pand %%mm1, %%mm0 \n\t"\ | |
2159 "movq %%mm0, " #b " \n\t" | |
2160 | |
2161 MEDIAN((%0), (%%eax), (%%eax, %1)) | |
2162 MEDIAN((%%eax, %1), (%%eax, %1, 2), (%0, %1, 4)) | |
2163 MEDIAN((%0, %1, 4), (%%ebx), (%%ebx, %1)) | |
2164 MEDIAN((%%ebx, %1), (%%ebx, %1, 2), (%0, %1, 8)) | |
2165 | |
2166 : : "r" (src), "r" (stride) | |
2167 : "%eax", "%ebx" | |
2168 ); | |
2169 #endif // MMX | |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2170 #else |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2171 //FIXME |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2172 int x; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2173 for(x=0; x<8; x++) |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2174 { |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2175 src[0 ] = (src[0 ] + 2*src[stride ] + src[stride*2])>>2; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2176 src[stride ] = (src[stride ] + 2*src[stride*2] + src[stride*3])>>2; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2177 src[stride*2] = (src[stride*2] + 2*src[stride*3] + src[stride*4])>>2; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2178 src[stride*3] = (src[stride*3] + 2*src[stride*4] + src[stride*5])>>2; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2179 src[stride*4] = (src[stride*4] + 2*src[stride*5] + src[stride*6])>>2; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2180 src[stride*5] = (src[stride*5] + 2*src[stride*6] + src[stride*7])>>2; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2181 src[stride*6] = (src[stride*6] + 2*src[stride*7] + src[stride*8])>>2; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2182 src[stride*7] = (src[stride*7] + 2*src[stride*8] + src[stride*9])>>2; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2183 src++; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2184 } |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2185 #endif |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2186 } |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2187 |
102 | 2188 #ifdef HAVE_ODIVX_POSTPROCESS |
2189 #include "../opendivx/postprocess.h" | |
2190 int use_old_pp=0; | |
2191 #endif | |
96 | 2192 |
102 | 2193 static void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height, |
2194 QP_STORE_T QPs[], int QPStride, int isColor, int mode); | |
96 | 2195 |
116 | 2196 /* -pp Command line Help |
2197 NOTE/FIXME: put this at an appropriate place (--help, html docs, man mplayer)? | |
2198 | |
2199 -pp <filterName>[:<option>[:<option>...]][,[-]<filterName>[:<option>...]]... | |
2200 | |
2201 long form example: | |
2202 -pp vdeblock:autoq,hdeblock:autoq,linblenddeint -pp default,-vdeblock | |
2203 short form example: | |
2204 -pp vb:a,hb:a,lb -pp de,-vb | |
2205 | |
2206 Filters Options | |
2207 short long name short long option Description | |
2208 * * a autoq cpu power dependant enabler | |
2209 c chrom chrominance filtring enabled | |
2210 y nochrom chrominance filtring disabled | |
2211 hb hdeblock horizontal deblocking filter | |
2212 vb vdeblock vertical deblocking filter | |
2213 vr rkvdeblock | |
2214 h1 x1hdeblock Experimental horizontal deblock filter 1 | |
2215 v1 x1vdeblock Experimental vertical deblock filter 1 | |
2216 dr dering not implemented yet | |
2217 al autolevels automatic brightness / contrast fixer | |
2218 f fullyrange stretch luminance range to (0..255) | |
2219 lb linblenddeint linear blend deinterlacer | |
2220 li linipoldeint linear interpolating deinterlacer | |
2221 ci cubicipoldeint cubic interpolating deinterlacer | |
2222 md mediandeint median deinterlacer | |
2223 de default hdeblock:a,vdeblock:a,dering:a,autolevels | |
2224 fa fast x1hdeblock:a,x1vdeblock:a,dering:a,autolevels | |
2225 */ | |
2226 | |
2227 /** | |
2228 * returns a PPMode struct which will have a non 0 error variable if an error occured | |
2229 * name is the string after "-pp" on the command line | |
2230 * quality is a number from 0 to GET_PP_QUALITY_MAX | |
2231 */ | |
2232 struct PPMode getPPModeByNameAndQuality(char *name, int quality) | |
2233 { | |
2234 char temp[GET_MODE_BUFFER_SIZE]; | |
2235 char *p= temp; | |
2236 char *filterDelimiters= ","; | |
2237 char *optionDelimiters= ":"; | |
2238 struct PPMode ppMode= {0,0,0,0,0,0}; | |
2239 char *filterToken; | |
2240 | |
2241 strncpy(temp, name, GET_MODE_BUFFER_SIZE); | |
2242 | |
2243 for(;;){ | |
2244 char *p2; | |
2245 char *filterName; | |
2246 int q= GET_PP_QUALITY_MAX; | |
2247 int chrom=-1; | |
2248 char *option; | |
2249 char *options[OPTIONS_ARRAY_SIZE]; | |
2250 int i; | |
2251 int filterNameOk=0; | |
2252 int numOfUnknownOptions=0; | |
2253 int enable=1; //does the user want us to enabled or disabled the filter | |
2254 | |
2255 filterToken= strtok(p, filterDelimiters); | |
2256 if(filterToken == NULL) break; | |
2257 p+= strlen(filterToken) + 1; | |
2258 filterName= strtok(filterToken, optionDelimiters); | |
2259 printf("%s::%s\n", filterToken, filterName); | |
2260 | |
2261 if(*filterName == '-') | |
2262 { | |
2263 enable=0; | |
2264 filterName++; | |
2265 } | |
2266 for(;;){ //for all options | |
2267 option= strtok(NULL, optionDelimiters); | |
2268 if(option == NULL) break; | |
2269 | |
2270 printf("%s\n", option); | |
2271 if(!strcmp("autoq", option) || !strcmp("a", option)) q= quality; | |
2272 else if(!strcmp("nochrom", option) || !strcmp("y", option)) chrom=0; | |
2273 else if(!strcmp("chrom", option) || !strcmp("c", option)) chrom=1; | |
2274 else | |
2275 { | |
2276 options[numOfUnknownOptions] = option; | |
2277 numOfUnknownOptions++; | |
2278 options[numOfUnknownOptions] = NULL; | |
2279 } | |
2280 if(numOfUnknownOptions >= OPTIONS_ARRAY_SIZE-1) break; | |
2281 } | |
2282 | |
2283 /* replace stuff from the replace Table */ | |
2284 for(i=0; replaceTable[2*i]!=NULL; i++) | |
2285 { | |
2286 if(!strcmp(replaceTable[2*i], filterName)) | |
2287 { | |
2288 int newlen= strlen(replaceTable[2*i + 1]); | |
2289 int plen; | |
2290 int spaceLeft; | |
2291 | |
2292 if(p==NULL) p= temp, *p=0; //last filter | |
2293 else p--, *p=','; //not last filter | |
2294 | |
2295 plen= strlen(p); | |
2296 spaceLeft= (int)p - (int)temp + plen; | |
2297 if(spaceLeft + newlen >= GET_MODE_BUFFER_SIZE) | |
2298 { | |
2299 ppMode.error++; | |
2300 break; | |
2301 } | |
2302 memmove(p + newlen, p, plen+1); | |
2303 memcpy(p, replaceTable[2*i + 1], newlen); | |
2304 filterNameOk=1; | |
2305 } | |
2306 } | |
2307 | |
2308 for(i=0; filters[i].shortName!=NULL; i++) | |
2309 { | |
2310 if( !strcmp(filters[i].longName, filterName) | |
2311 || !strcmp(filters[i].shortName, filterName)) | |
2312 { | |
2313 ppMode.lumMode &= ~filters[i].mask; | |
2314 ppMode.chromMode &= ~filters[i].mask; | |
2315 | |
2316 filterNameOk=1; | |
2317 if(!enable) break; // user wants to disable it | |
2318 | |
2319 if(q >= filters[i].minLumQuality) | |
2320 ppMode.lumMode|= filters[i].mask; | |
2321 if(chrom==1 || (chrom==-1 && filters[i].chromDefault)) | |
2322 if(q >= filters[i].minChromQuality) | |
2323 ppMode.chromMode|= filters[i].mask; | |
2324 | |
2325 if(filters[i].mask == LEVEL_FIX) | |
2326 { | |
2327 int o; | |
2328 ppMode.minAllowedY= 16; | |
2329 ppMode.maxAllowedY= 234; | |
2330 for(o=0; options[o]!=NULL; o++) | |
2331 if( !strcmp(options[o],"fullyrange") | |
2332 ||!strcmp(options[o],"f")) | |
2333 { | |
2334 ppMode.minAllowedY= 0; | |
2335 ppMode.maxAllowedY= 255; | |
2336 numOfUnknownOptions--; | |
2337 } | |
2338 } | |
2339 } | |
2340 } | |
2341 if(!filterNameOk) ppMode.error++; | |
2342 ppMode.error += numOfUnknownOptions; | |
2343 } | |
2344 | |
2345 if(ppMode.lumMode & H_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_Y_H; | |
2346 if(ppMode.lumMode & V_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_Y_V; | |
2347 if(ppMode.chromMode & H_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_C_H; | |
2348 if(ppMode.chromMode & V_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_C_V; | |
2349 if(ppMode.lumMode & DERING) ppMode.oldMode |= PP_DERING_Y; | |
2350 if(ppMode.chromMode & DERING) ppMode.oldMode |= PP_DERING_C; | |
2351 | |
2352 return ppMode; | |
2353 } | |
2354 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2355 /** |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2356 * ... |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2357 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2358 void postprocess(unsigned char * src[], int src_stride, |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2359 unsigned char * dst[], int dst_stride, |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2360 int horizontal_size, int vertical_size, |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2361 QP_STORE_T *QP_store, int QP_stride, |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2362 int mode) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2363 { |
116 | 2364 /* |
2365 static int qual=0; | |
2366 | |
2367 struct PPMode ppMode= getPPModeByNameAndQuality("fast,default,-hdeblock,-vdeblock", qual); | |
2368 qual++; | |
2369 qual%=7; | |
2370 printf("\n%d %d %d %d\n", ppMode.lumMode, ppMode.chromMode, ppMode.oldMode, ppMode.error); | |
2371 postprocess2(src, src_stride, dst, dst_stride, | |
2372 horizontal_size, vertical_size, QP_store, QP_stride, &ppMode); | |
2373 | |
2374 return; | |
2375 */ | |
96 | 2376 |
102 | 2377 #ifdef HAVE_ODIVX_POSTPROCESS |
2378 // Note: I could make this shit outside of this file, but it would mean one | |
2379 // more function call... | |
2380 if(use_old_pp){ | |
2381 odivx_postprocess(src,src_stride,dst,dst_stride,horizontal_size,vertical_size,QP_store,QP_stride,mode); | |
2382 return; | |
2383 } | |
2384 #endif | |
2385 | |
96 | 2386 postProcess(src[0], src_stride, dst[0], dst_stride, |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2387 horizontal_size, vertical_size, QP_store, QP_stride, 0, mode); |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2388 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2389 horizontal_size >>= 1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2390 vertical_size >>= 1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2391 src_stride >>= 1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2392 dst_stride >>= 1; |
99 | 2393 mode= ((mode&0xFF)>>4) | (mode&0xFFFFFF00); |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2394 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2395 if(1) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2396 { |
96 | 2397 postProcess(src[1], src_stride, dst[1], dst_stride, |
99 | 2398 horizontal_size, vertical_size, QP_store, QP_stride, 1, mode); |
96 | 2399 postProcess(src[2], src_stride, dst[2], dst_stride, |
116 | 2400 horizontal_size, vertical_size, QP_store, QP_stride, 2, mode); |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2401 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2402 else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2403 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2404 memcpy(dst[1], src[1], src_stride*horizontal_size); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2405 memcpy(dst[2], src[2], src_stride*horizontal_size); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2406 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2407 } |
102 | 2408 |
116 | 2409 void postprocess2(unsigned char * src[], int src_stride, |
2410 unsigned char * dst[], int dst_stride, | |
2411 int horizontal_size, int vertical_size, | |
2412 QP_STORE_T *QP_store, int QP_stride, | |
2413 struct PPMode *mode) | |
2414 { | |
2415 | |
2416 #ifdef HAVE_ODIVX_POSTPROCESS | |
2417 // Note: I could make this shit outside of this file, but it would mean one | |
2418 // more function call... | |
2419 if(use_old_pp){ | |
2420 odivx_postprocess(src,src_stride,dst,dst_stride,horizontal_size,vertical_size,QP_store,QP_stride, | |
2421 mode->oldMode); | |
2422 return; | |
2423 } | |
2424 #endif | |
2425 | |
2426 postProcess(src[0], src_stride, dst[0], dst_stride, | |
2427 horizontal_size, vertical_size, QP_store, QP_stride, 0, mode->lumMode); | |
2428 | |
2429 horizontal_size >>= 1; | |
2430 vertical_size >>= 1; | |
2431 src_stride >>= 1; | |
2432 dst_stride >>= 1; | |
2433 | |
2434 postProcess(src[1], src_stride, dst[1], dst_stride, | |
2435 horizontal_size, vertical_size, QP_store, QP_stride, 1, mode->chromMode); | |
2436 postProcess(src[2], src_stride, dst[2], dst_stride, | |
2437 horizontal_size, vertical_size, QP_store, QP_stride, 2, mode->chromMode); | |
2438 } | |
2439 | |
2440 | |
96 | 2441 /** |
2442 * gets the mode flags for a given quality (larger values mean slower but better postprocessing) | |
102 | 2443 * 0 <= quality <= 6 |
96 | 2444 */ |
102 | 2445 int getPpModeForQuality(int quality){ |
2446 int modes[1+GET_PP_QUALITY_MAX]= { | |
2447 0, | |
2448 #if 1 | |
2449 // horizontal filters first | |
2450 LUM_H_DEBLOCK, | |
2451 LUM_H_DEBLOCK | LUM_V_DEBLOCK, | |
2452 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK, | |
2453 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK, | |
2454 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK | LUM_DERING, | |
2455 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK | LUM_DERING | CHROM_DERING | |
2456 #else | |
2457 // vertical filters first | |
96 | 2458 LUM_V_DEBLOCK, |
2459 LUM_V_DEBLOCK | LUM_H_DEBLOCK, | |
2460 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK, | |
2461 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK, | |
2462 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING, | |
2463 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING | CHROM_DERING | |
102 | 2464 #endif |
2465 }; | |
96 | 2466 |
102 | 2467 #ifdef HAVE_ODIVX_POSTPROCESS |
2468 int odivx_modes[1+GET_PP_QUALITY_MAX]= { | |
2469 0, | |
2470 PP_DEBLOCK_Y_H, | |
2471 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V, | |
2472 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H, | |
2473 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V, | |
2474 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V|PP_DERING_Y, | |
2475 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V|PP_DERING_Y|PP_DERING_C | |
2476 }; | |
2477 if(use_old_pp) return odivx_modes[quality]; | |
2478 #endif | |
2479 return modes[quality]; | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2480 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2481 |
98 | 2482 //} // extern "C" |
96 | 2483 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2484 /** |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2485 * Copies a block from src to dst and fixes the blacklevel |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2486 * numLines must be a multiple of 4 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2487 * levelFix == 0 -> dont touch the brighness & contrast |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2488 */ |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2489 static inline void blockCopy(uint8_t dst[], int dstStride, uint8_t src[], int srcStride, |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2490 int numLines, int levelFix) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2491 { |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2492 int i; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2493 if(levelFix) |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2494 { |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2495 #ifdef HAVE_MMX |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2496 asm volatile( |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2497 "movl %4, %%eax \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2498 "movl %%eax, temp0\n\t" |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2499 "pushl %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2500 "pushl %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2501 "leal (%2,%2), %%eax \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2502 "leal (%3,%3), %%ebx \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2503 "movq packedYOffset, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2504 "movq packedYScale, %%mm3 \n\t" |
101
fcf4e8fcb34b
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
100
diff
changeset
|
2505 "pxor %%mm4, %%mm4 \n\t" |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2506 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2507 #define SCALED_CPY \ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2508 "movq (%0), %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2509 "movq (%0,%2), %%mm1 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2510 "psubusb %%mm2, %%mm0 \n\t"\ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2511 "psubusb %%mm2, %%mm1 \n\t"\ |
101
fcf4e8fcb34b
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
100
diff
changeset
|
2512 "movq %%mm0, %%mm5 \n\t"\ |
fcf4e8fcb34b
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
100
diff
changeset
|
2513 "punpcklbw %%mm4, %%mm0 \n\t"\ |
fcf4e8fcb34b
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
100
diff
changeset
|
2514 "punpckhbw %%mm4, %%mm5 \n\t"\ |
fcf4e8fcb34b
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
100
diff
changeset
|
2515 "psllw $7, %%mm0 \n\t"\ |
fcf4e8fcb34b
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
100
diff
changeset
|
2516 "psllw $7, %%mm5 \n\t"\ |
fcf4e8fcb34b
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
100
diff
changeset
|
2517 "pmulhw %%mm3, %%mm0 \n\t"\ |
fcf4e8fcb34b
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
100
diff
changeset
|
2518 "pmulhw %%mm3, %%mm5 \n\t"\ |
fcf4e8fcb34b
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
100
diff
changeset
|
2519 "packuswb %%mm5, %%mm0 \n\t"\ |
fcf4e8fcb34b
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
100
diff
changeset
|
2520 "movq %%mm0, (%1) \n\t"\ |
fcf4e8fcb34b
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
100
diff
changeset
|
2521 "movq %%mm1, %%mm5 \n\t"\ |
fcf4e8fcb34b
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
100
diff
changeset
|
2522 "punpcklbw %%mm4, %%mm1 \n\t"\ |
fcf4e8fcb34b
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
100
diff
changeset
|
2523 "punpckhbw %%mm4, %%mm5 \n\t"\ |
fcf4e8fcb34b
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
100
diff
changeset
|
2524 "psllw $7, %%mm1 \n\t"\ |
fcf4e8fcb34b
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
100
diff
changeset
|
2525 "psllw $7, %%mm5 \n\t"\ |
fcf4e8fcb34b
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
100
diff
changeset
|
2526 "pmulhw %%mm3, %%mm1 \n\t"\ |
fcf4e8fcb34b
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
100
diff
changeset
|
2527 "pmulhw %%mm3, %%mm5 \n\t"\ |
fcf4e8fcb34b
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
100
diff
changeset
|
2528 "packuswb %%mm5, %%mm1 \n\t"\ |
fcf4e8fcb34b
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
100
diff
changeset
|
2529 "movq %%mm1, (%1, %3) \n\t"\ |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2530 |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2531 "1: \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2532 SCALED_CPY |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2533 "addl %%eax, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2534 "addl %%ebx, %1 \n\t" |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2535 SCALED_CPY |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2536 "addl %%eax, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2537 "addl %%ebx, %1 \n\t" |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2538 "decl temp0 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2539 "jnz 1b \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2540 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2541 "popl %1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2542 "popl %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2543 : : "r" (src), |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2544 "r" (dst), |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2545 "r" (srcStride), |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2546 "r" (dstStride), |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2547 "m" (numLines>>2) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2548 : "%eax", "%ebx" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2549 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2550 #else |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2551 for(i=0; i<numLines; i++) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2552 memcpy( &(dst[dstStride*i]), |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2553 &(src[srcStride*i]), BLOCK_SIZE); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2554 #endif |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2555 } |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2556 else |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2557 { |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2558 #ifdef HAVE_MMX |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2559 asm volatile( |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2560 "movl %4, %%eax \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2561 "movl %%eax, temp0\n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2562 "pushl %0 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2563 "pushl %1 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2564 "leal (%2,%2), %%eax \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2565 "leal (%3,%3), %%ebx \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2566 "movq packedYOffset, %%mm2 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2567 "movq packedYScale, %%mm3 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2568 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2569 #define SIMPLE_CPY \ |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2570 "movq (%0), %%mm0 \n\t"\ |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2571 "movq (%0,%2), %%mm1 \n\t"\ |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2572 "movq %%mm0, (%1) \n\t"\ |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2573 "movq %%mm1, (%1, %3) \n\t"\ |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2574 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2575 "1: \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2576 SIMPLE_CPY |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2577 "addl %%eax, %0 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2578 "addl %%ebx, %1 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2579 SIMPLE_CPY |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2580 "addl %%eax, %0 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2581 "addl %%ebx, %1 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2582 "decl temp0 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2583 "jnz 1b \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2584 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2585 "popl %1 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2586 "popl %0 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2587 : : "r" (src), |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2588 "r" (dst), |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2589 "r" (srcStride), |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2590 "r" (dstStride), |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2591 "m" (numLines>>2) |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2592 : "%eax", "%ebx" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2593 ); |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2594 #else |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2595 for(i=0; i<numLines; i++) |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2596 memcpy( &(dst[dstStride*i]), |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2597 &(src[srcStride*i]), BLOCK_SIZE); |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2598 #endif |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2599 } |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2600 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2601 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2602 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2603 /** |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2604 * Filters array of bytes (Y or U or V values) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2605 */ |
102 | 2606 static void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height, |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2607 QP_STORE_T QPs[], int QPStride, int isColor, int mode) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2608 { |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2609 int x,y; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2610 /* we need 64bit here otherwise we´ll going to have a problem |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2611 after watching a black picture for 5 hours*/ |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2612 static uint64_t *yHistogram= NULL; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2613 int black=0, white=255; // blackest black and whitest white in the picture |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2614 |
111 | 2615 /* Temporary buffers for handling the last row(s) */ |
2616 static uint8_t *tempDst= NULL; | |
2617 static uint8_t *tempSrc= NULL; | |
2618 | |
112 | 2619 /* Temporary buffers for handling the last block */ |
2620 static uint8_t *tempDstBlock= NULL; | |
2621 static uint8_t *tempSrcBlock= NULL; | |
2622 | |
2623 uint8_t *dstBlockPtrBackup; | |
2624 uint8_t *srcBlockPtrBackup; | |
2625 | |
107 | 2626 #ifdef TIMING |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2627 long long T0, T1, memcpyTime=0, vertTime=0, horizTime=0, sumTime, diffTime=0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2628 sumTime= rdtsc(); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2629 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2630 |
111 | 2631 if(tempDst==NULL) |
2632 { | |
2633 tempDst= (uint8_t*)memalign(8, 1024*24); | |
2634 tempSrc= (uint8_t*)memalign(8, 1024*24); | |
112 | 2635 tempDstBlock= (uint8_t*)memalign(8, 1024*24); |
2636 tempSrcBlock= (uint8_t*)memalign(8, 1024*24); | |
111 | 2637 } |
2638 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2639 if(!yHistogram) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2640 { |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2641 int i; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2642 yHistogram= (uint64_t*)malloc(8*256); |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2643 for(i=0; i<256; i++) yHistogram[i]= width*height/64*15/256; |
112 | 2644 |
2645 if(mode & FULL_Y_RANGE) | |
2646 { | |
2647 maxAllowedY=255; | |
2648 minAllowedY=0; | |
2649 } | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2650 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2651 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2652 if(!isColor) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2653 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2654 uint64_t sum= 0; |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2655 int i; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2656 static int framenum= -1; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2657 uint64_t maxClipped; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2658 uint64_t clipped; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2659 double scale; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2660 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2661 framenum++; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2662 if(framenum == 1) yHistogram[0]= width*height/64*15/256; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2663 |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2664 for(i=0; i<256; i++) |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2665 { |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2666 sum+= yHistogram[i]; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2667 // printf("%d ", yHistogram[i]); |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2668 } |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2669 // printf("\n\n"); |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2670 |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2671 /* we allways get a completly black picture first */ |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2672 maxClipped= (uint64_t)(sum * maxClippedThreshold); |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2673 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2674 clipped= sum; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2675 for(black=255; black>0; black--) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2676 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2677 if(clipped < maxClipped) break; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2678 clipped-= yHistogram[black]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2679 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2680 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2681 clipped= sum; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2682 for(white=0; white<256; white++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2683 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2684 if(clipped < maxClipped) break; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2685 clipped-= yHistogram[white]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2686 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2687 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2688 // we cant handle negative correctures |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2689 packedYOffset= MAX(black - minAllowedY, 0); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2690 packedYOffset|= packedYOffset<<32; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2691 packedYOffset|= packedYOffset<<16; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2692 packedYOffset|= packedYOffset<<8; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2693 |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2694 scale= (double)(maxAllowedY - minAllowedY) / (double)(white-black); |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2695 |
101
fcf4e8fcb34b
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
100
diff
changeset
|
2696 packedYScale= (uint16_t)(scale*512.0 + 0.5); |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2697 packedYScale|= packedYScale<<32; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2698 packedYScale|= packedYScale<<16; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2699 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2700 else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2701 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2702 packedYScale= 0x0100010001000100LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2703 packedYOffset= 0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2704 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2705 |
111 | 2706 /* copy first row of 8x8 blocks */ |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2707 for(x=0; x<width; x+=BLOCK_SIZE) |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2708 blockCopy(dst + x, dstStride, src + x, srcStride, 8, mode & LEVEL_FIX); |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2709 |
111 | 2710 for(y=0; y<height; y+=BLOCK_SIZE) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2711 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2712 //1% speedup if these are here instead of the inner loop |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2713 uint8_t *srcBlock= &(src[y*srcStride]); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2714 uint8_t *dstBlock= &(dst[y*dstStride]); |
111 | 2715 |
2716 /* can we mess with a 8x16 block from srcBlock/dstBlock downwards, if not | |
2717 than use a temporary buffer */ | |
2718 if(y+15 >= height) | |
2719 { | |
2720 /* copy from line 5 to 12 of src, these will e copied with | |
2721 blockcopy to dst later */ | |
2722 memcpy(tempSrc + srcStride*5, srcBlock + srcStride*5, | |
2723 srcStride*MAX(height-y-5, 0) ); | |
2724 | |
2725 /* duplicate last line to fill the void upto line 12 */ | |
2726 if(y+12 >= height) | |
2727 { | |
2728 int i; | |
2729 for(i=height-y; i<=12; i++) | |
2730 memcpy(tempSrc + srcStride*i, | |
2731 src + srcStride*(height-1), srcStride); | |
2732 } | |
2733 | |
2734 | |
2735 /* copy up to 5 lines of dst */ | |
2736 memcpy(tempDst, dstBlock, dstStride*MIN(height-y, 5) ); | |
2737 dstBlock= tempDst; | |
2738 srcBlock= tempSrc; | |
2739 } | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2740 |
112 | 2741 // From this point on it is guranteed that we can read and write 16 lines downward |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2742 // finish 1 block before the next otherwise we´ll might have a problem |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2743 // with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2744 for(x=0; x<width; x+=BLOCK_SIZE) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2745 { |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2746 const int stride= dstStride; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2747 int QP= isColor ? |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2748 QPs[(y>>3)*QPStride + (x>>3)]: |
100 | 2749 QPs[(y>>4)*QPStride + (x>>4)]; |
2750 if(!isColor && (mode & LEVEL_FIX)) QP= (QP* (packedYScale &0xFFFF))>>8; | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2751 #ifdef HAVE_MMX |
111 | 2752 asm volatile( |
2753 "movd %0, %%mm7 \n\t" | |
2754 "packuswb %%mm7, %%mm7 \n\t" // 0, 0, 0, QP, 0, 0, 0, QP | |
2755 "packuswb %%mm7, %%mm7 \n\t" // 0,QP, 0, QP, 0,QP, 0, QP | |
2756 "packuswb %%mm7, %%mm7 \n\t" // QP,..., QP | |
2757 "movq %%mm7, pQPb \n\t" | |
2758 : : "r" (QP) | |
2759 ); | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2760 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2761 |
107 | 2762 #ifdef MORE_TIMING |
111 | 2763 T0= rdtsc(); |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2764 #endif |
96 | 2765 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2766 #ifdef HAVE_MMX2 |
111 | 2767 prefetchnta(srcBlock + (((x>>3)&3) + 5)*srcStride + 32); |
2768 prefetchnta(srcBlock + (((x>>3)&3) + 9)*srcStride + 32); | |
2769 prefetcht0(dstBlock + (((x>>3)&3) + 5)*dstStride + 32); | |
2770 prefetcht0(dstBlock + (((x>>3)&3) + 9)*dstStride + 32); | |
96 | 2771 #elif defined(HAVE_3DNOW) |
2772 //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ... | |
111 | 2773 /* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32); |
2774 prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32); | |
2775 prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32); | |
2776 prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32); | |
96 | 2777 */ |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2778 #endif |
111 | 2779 |
2780 if(!isColor) yHistogram[ srcBlock[srcStride*5] ]++; | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2781 |
113 | 2782 #ifdef PP_FUNNY_STRIDE |
112 | 2783 //can we mess with a 8x16 block, if not use a temp buffer, yes again |
2784 if(x+7 >= width) | |
2785 { | |
2786 int i; | |
2787 dstBlockPtrBackup= dstBlock; | |
2788 srcBlockPtrBackup= srcBlock; | |
2789 | |
2790 for(i=0;i<BLOCK_SIZE*2; i++) | |
2791 { | |
2792 memcpy(tempSrcBlock+i*srcStride, srcBlock+i*srcStride, width-x); | |
2793 memcpy(tempDstBlock+i*dstStride, dstBlock+i*dstStride, width-x); | |
2794 } | |
2795 | |
2796 dstBlock= tempDstBlock; | |
2797 srcBlock= tempSrcBlock; | |
2798 } | |
113 | 2799 #endif |
112 | 2800 |
111 | 2801 blockCopy(dstBlock + dstStride*5, dstStride, |
2802 srcBlock + srcStride*5, srcStride, 8, mode & LEVEL_FIX); | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2803 |
111 | 2804 if(mode & LINEAR_IPOL_DEINT_FILTER) |
2805 deInterlaceInterpolateLinear(dstBlock, dstStride); | |
2806 else if(mode & LINEAR_BLEND_DEINT_FILTER) | |
2807 deInterlaceBlendLinear(dstBlock, dstStride); | |
2808 else if(mode & MEDIAN_DEINT_FILTER) | |
2809 deInterlaceMedian(dstBlock, dstStride); | |
2810 else if(mode & CUBIC_IPOL_DEINT_FILTER) | |
2811 deInterlaceInterpolateCubic(dstBlock, dstStride); | |
2812 /* else if(mode & CUBIC_BLEND_DEINT_FILTER) | |
2813 deInterlaceBlendCubic(dstBlock, dstStride); | |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2814 */ |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2815 |
111 | 2816 /* only deblock if we have 2 blocks */ |
2817 if(y + 8 < height) | |
2818 { | |
107 | 2819 #ifdef MORE_TIMING |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2820 T1= rdtsc(); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2821 memcpyTime+= T1-T0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2822 T0=T1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2823 #endif |
115
4514b8e7f0f1
more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents:
113
diff
changeset
|
2824 if(mode & V_RK1_FILTER) |
4514b8e7f0f1
more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents:
113
diff
changeset
|
2825 vertRK1Filter(dstBlock, stride, QP); |
4514b8e7f0f1
more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents:
113
diff
changeset
|
2826 else if(mode & V_X1_FILTER) |
4514b8e7f0f1
more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents:
113
diff
changeset
|
2827 vertX1Filter(dstBlock, stride, QP); |
4514b8e7f0f1
more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents:
113
diff
changeset
|
2828 else if(mode & V_DEBLOCK) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2829 { |
115
4514b8e7f0f1
more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents:
113
diff
changeset
|
2830 if( isVertDC(dstBlock, stride)) |
96 | 2831 { |
115
4514b8e7f0f1
more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents:
113
diff
changeset
|
2832 if(isVertMinMaxOk(dstBlock, stride, QP)) |
4514b8e7f0f1
more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents:
113
diff
changeset
|
2833 doVertLowPass(dstBlock, stride, QP); |
96 | 2834 } |
115
4514b8e7f0f1
more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents:
113
diff
changeset
|
2835 else |
4514b8e7f0f1
more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents:
113
diff
changeset
|
2836 doVertDefFilter(dstBlock, stride, QP); |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2837 } |
107 | 2838 #ifdef MORE_TIMING |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2839 T1= rdtsc(); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2840 vertTime+= T1-T0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2841 T0=T1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2842 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2843 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2844 |
111 | 2845 /* check if we have a previous block to deblock it with dstBlock */ |
112 | 2846 if(x - 8 >= 0) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2847 { |
107 | 2848 #ifdef MORE_TIMING |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2849 T0= rdtsc(); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2850 #endif |
115
4514b8e7f0f1
more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents:
113
diff
changeset
|
2851 if(mode & H_X1_FILTER) |
4514b8e7f0f1
more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents:
113
diff
changeset
|
2852 horizX1Filter(dstBlock-4, stride, QP); |
4514b8e7f0f1
more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents:
113
diff
changeset
|
2853 else if(mode & H_DEBLOCK) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2854 { |
115
4514b8e7f0f1
more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents:
113
diff
changeset
|
2855 if( isHorizDCAndCopy2Temp(dstBlock-4, stride)) |
96 | 2856 { |
115
4514b8e7f0f1
more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents:
113
diff
changeset
|
2857 if(isHorizMinMaxOk(tempBlock, TEMP_STRIDE, QP)) |
4514b8e7f0f1
more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents:
113
diff
changeset
|
2858 doHorizLowPassAndCopyBack(dstBlock-4, stride, QP); |
96 | 2859 } |
115
4514b8e7f0f1
more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents:
113
diff
changeset
|
2860 else |
4514b8e7f0f1
more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents:
113
diff
changeset
|
2861 doHorizDefFilterAndCopyBack(dstBlock-4, stride, QP); |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2862 } |
107 | 2863 #ifdef MORE_TIMING |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2864 T1= rdtsc(); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2865 horizTime+= T1-T0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2866 T0=T1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2867 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2868 dering(dstBlock - 9 - stride, stride, QP); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2869 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2870 else if(y!=0) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2871 dering(dstBlock - stride*9 + width-9, stride, QP); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2872 //FIXME dering filter will not be applied to last block (bottom right) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2873 |
113 | 2874 #ifdef PP_FUNNY_STRIDE |
112 | 2875 /* did we use a tmp-block buffer */ |
2876 if(x+7 >= width) | |
2877 { | |
2878 int i; | |
2879 dstBlock= dstBlockPtrBackup; | |
2880 srcBlock= srcBlockPtrBackup; | |
2881 | |
2882 for(i=0;i<BLOCK_SIZE*2; i++) | |
2883 { | |
2884 memcpy(dstBlock+i*dstStride, tempDstBlock+i*dstStride, width-x); | |
2885 } | |
2886 } | |
113 | 2887 #endif |
112 | 2888 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2889 dstBlock+=8; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2890 srcBlock+=8; |
111 | 2891 } |
2892 | |
2893 /* did we use a tmp buffer */ | |
112 | 2894 if(y+15 >= height) |
111 | 2895 { |
2896 uint8_t *dstBlock= &(dst[y*dstStride]); | |
2897 memcpy(dstBlock, tempDst, dstStride*(height-y) ); | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2898 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2899 } |
96 | 2900 #ifdef HAVE_3DNOW |
2901 asm volatile("femms"); | |
2902 #elif defined (HAVE_MMX) | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2903 asm volatile("emms"); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2904 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2905 |
107 | 2906 #ifdef TIMING |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2907 // FIXME diff is mostly the time spent for rdtsc (should subtract that but ...) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2908 sumTime= rdtsc() - sumTime; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2909 if(!isColor) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2910 printf("cpy:%4dk, vert:%4dk, horiz:%4dk, sum:%4dk, diff:%4dk, color: %d/%d \r", |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2911 (int)(memcpyTime/1000), (int)(vertTime/1000), (int)(horizTime/1000), |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
2912 (int)(sumTime/1000), (int)((sumTime-memcpyTime-vertTime-horizTime)/1000) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2913 , black, white); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2914 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2915 } |