Mercurial > libavcodec.hg
annotate libpostproc/postprocess_template.c @ 167:2d97f0157a79 libavcodec
faster dering
author | michael |
---|---|
date | Sat, 24 Nov 2001 01:38:30 +0000 |
parents | ec349ac7869b |
children | 712c7a115164 |
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 |
163 | 24 doVertDefFilter Ec Ec e e |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
25 isHorizDC Ec Ec |
128
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
26 isHorizMinMaxOk a E |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
27 doHorizLowPass E e e |
163 | 28 doHorizDefFilter Ec Ec e e |
134 | 29 deRing E e e* |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
30 Vertical RKAlgo1 E a a |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
31 Horizontal RKAlgo1 a a |
156 | 32 Vertical X1# a E E |
33 Horizontal X1# a E E | |
111 | 34 LinIpolDeinterlace e E E* |
35 CubicIpolDeinterlace a e e* | |
36 LinBlendDeinterlace e E E* | |
156 | 37 MedianDeinterlace# Ec Ec |
157
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
38 TempDeNoiser# E e e |
156 | 39 |
40 * i dont have a 3dnow CPU -> its untested, but noone said it doesnt work so it seems to work | |
41 # more or less selfinvented filters so the exactness isnt too meaningfull | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
42 E = Exact implementation |
111 | 43 e = allmost exact implementation (slightly different rounding,...) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
44 a = alternative / approximate impl |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
45 c = checked against the other implementations (-vo md5) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
46 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
47 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
48 /* |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
49 TODO: |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
50 reduce the time wasted on the mem transfer |
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 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
|
57 (the if/else stuff per block is slowing things down) |
99 | 58 compare the quality & speed of all filters |
59 split this huge file | |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
60 border remover |
140 | 61 optimize c versions |
156 | 62 try to unroll inner for(x=0 ... loop to avoid these damn if(x ... checks |
157
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
63 smart blur |
167 | 64 commandline option for the deblock / dering thresholds |
65 memcpy chrominance if no chroma filtering is done | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
66 ... |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
67 */ |
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 |
150 | 71 #include "../config.h" |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
72 #include <inttypes.h> |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
73 #include <stdio.h> |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
74 #include <stdlib.h> |
116 | 75 #include <string.h> |
133 | 76 #ifdef HAVE_MALLOC_H |
77 #include <malloc.h> | |
78 #endif | |
96 | 79 //#undef HAVE_MMX2 |
80 //#define HAVE_3DNOW | |
81 //#undef HAVE_MMX | |
163 | 82 //#define DEBUG_BRIGHTNESS |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
83 #include "postprocess.h" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
84 |
104 | 85 #define MIN(a,b) ((a) > (b) ? (b) : (a)) |
86 #define MAX(a,b) ((a) < (b) ? (b) : (a)) | |
87 #define ABS(a) ((a) > 0 ? (a) : (-(a))) | |
88 #define SIGN(a) ((a) > 0 ? 1 : -1) | |
89 | |
90 #ifdef HAVE_MMX2 | |
91 #define PAVGB(a,b) "pavgb " #a ", " #b " \n\t" | |
92 #elif defined (HAVE_3DNOW) | |
93 #define PAVGB(a,b) "pavgusb " #a ", " #b " \n\t" | |
94 #endif | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
95 |
134 | 96 #ifdef HAVE_MMX2 |
97 #define PMINUB(a,b,t) "pminub " #a ", " #b " \n\t" | |
98 #elif defined (HAVE_MMX) | |
99 #define PMINUB(b,a,t) \ | |
100 "movq " #a ", " #t " \n\t"\ | |
101 "psubusb " #b ", " #t " \n\t"\ | |
102 "psubb " #t ", " #a " \n\t" | |
103 #endif | |
104 | |
105 #ifdef HAVE_MMX2 | |
106 #define PMAXUB(a,b) "pmaxub " #a ", " #b " \n\t" | |
107 #elif defined (HAVE_MMX) | |
108 #define PMAXUB(a,b) \ | |
109 "psubusb " #a ", " #b " \n\t"\ | |
110 "paddb " #a ", " #b " \n\t" | |
111 #endif | |
112 | |
113 | |
116 | 114 #define GET_MODE_BUFFER_SIZE 500 |
115 #define OPTIONS_ARRAY_SIZE 10 | |
116 | |
148
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
117 #ifdef HAVE_MMX |
149
3fc494e92f0d
gcc does optimize writes to non volatile variables away if it didnt know that they were read in between
michael
parents:
148
diff
changeset
|
118 static volatile uint64_t __attribute__((aligned(8))) packedYOffset= 0x0000000000000000LL; |
3fc494e92f0d
gcc does optimize writes to non volatile variables away if it didnt know that they were read in between
michael
parents:
148
diff
changeset
|
119 static volatile uint64_t __attribute__((aligned(8))) packedYScale= 0x0100010001000100LL; |
148
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
120 static uint64_t __attribute__((aligned(8))) w05= 0x0005000500050005LL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
121 static uint64_t __attribute__((aligned(8))) w20= 0x0020002000200020LL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
122 static uint64_t __attribute__((aligned(8))) w1400= 0x1400140014001400LL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
123 static uint64_t __attribute__((aligned(8))) bm00000001= 0x00000000000000FFLL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
124 static uint64_t __attribute__((aligned(8))) bm00010000= 0x000000FF00000000LL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
125 static uint64_t __attribute__((aligned(8))) bm00001000= 0x00000000FF000000LL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
126 static uint64_t __attribute__((aligned(8))) bm10000000= 0xFF00000000000000LL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
127 static uint64_t __attribute__((aligned(8))) bm10000001= 0xFF000000000000FFLL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
128 static uint64_t __attribute__((aligned(8))) bm11000011= 0xFFFF00000000FFFFLL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
129 static uint64_t __attribute__((aligned(8))) bm00000011= 0x000000000000FFFFLL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
130 static uint64_t __attribute__((aligned(8))) bm11111110= 0xFFFFFFFFFFFFFF00LL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
131 static uint64_t __attribute__((aligned(8))) bm11000000= 0xFFFF000000000000LL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
132 static uint64_t __attribute__((aligned(8))) bm00011000= 0x000000FFFF000000LL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
133 static uint64_t __attribute__((aligned(8))) bm00110011= 0x0000FFFF0000FFFFLL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
134 static uint64_t __attribute__((aligned(8))) bm11001100= 0xFFFF0000FFFF0000LL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
135 static uint64_t __attribute__((aligned(8))) b00= 0x0000000000000000LL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
136 static uint64_t __attribute__((aligned(8))) b01= 0x0101010101010101LL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
137 static uint64_t __attribute__((aligned(8))) b02= 0x0202020202020202LL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
138 static uint64_t __attribute__((aligned(8))) b0F= 0x0F0F0F0F0F0F0F0FLL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
139 static uint64_t __attribute__((aligned(8))) b04= 0x0404040404040404LL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
140 static uint64_t __attribute__((aligned(8))) b08= 0x0808080808080808LL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
141 static uint64_t __attribute__((aligned(8))) bFF= 0xFFFFFFFFFFFFFFFFLL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
142 static uint64_t __attribute__((aligned(8))) b20= 0x2020202020202020LL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
143 static uint64_t __attribute__((aligned(8))) b80= 0x8080808080808080LL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
144 static uint64_t __attribute__((aligned(8))) b7E= 0x7E7E7E7E7E7E7E7ELL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
145 static uint64_t __attribute__((aligned(8))) b7C= 0x7C7C7C7C7C7C7C7CLL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
146 static uint64_t __attribute__((aligned(8))) b3F= 0x3F3F3F3F3F3F3F3FLL; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
147 static uint64_t __attribute__((aligned(8))) temp0=0; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
148 static uint64_t __attribute__((aligned(8))) temp1=0; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
149 static uint64_t __attribute__((aligned(8))) temp2=0; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
150 static uint64_t __attribute__((aligned(8))) temp3=0; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
151 static uint64_t __attribute__((aligned(8))) temp4=0; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
152 static uint64_t __attribute__((aligned(8))) temp5=0; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
153 static uint64_t __attribute__((aligned(8))) pQPb=0; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
154 static uint64_t __attribute__((aligned(8))) pQPb2=0; |
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
155 static uint8_t __attribute__((aligned(8))) tempBlocks[8*16*2]; //used for the horizontal code |
158 | 156 static uint32_t __attribute__((aligned(4))) maxTmpNoise[4]; |
148
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
157 #else |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
158 static uint64_t packedYOffset= 0x0000000000000000LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
159 static uint64_t packedYScale= 0x0100010001000100LL; |
128
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
160 static uint8_t tempBlocks[8*16*2]; //used for the horizontal code |
148
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
161 #endif |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
162 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
163 int hFlatnessThreshold= 56 - 16; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
164 int vFlatnessThreshold= 56 - 16; |
167 | 165 int deringThreshold= 20; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
166 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
167 //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
|
168 double maxClippedThreshold= 0.01; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
169 |
116 | 170 int maxAllowedY=234; |
112 | 171 int minAllowedY=16; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
172 |
116 | 173 static struct PPFilter filters[]= |
174 { | |
175 {"hb", "hdeblock", 1, 1, 3, H_DEBLOCK}, | |
176 {"vb", "vdeblock", 1, 2, 4, V_DEBLOCK}, | |
177 {"vr", "rkvdeblock", 1, 2, 4, H_RK1_FILTER}, | |
178 {"h1", "x1hdeblock", 1, 1, 3, H_X1_FILTER}, | |
179 {"v1", "x1vdeblock", 1, 2, 4, V_X1_FILTER}, | |
180 {"dr", "dering", 1, 5, 6, DERING}, | |
181 {"al", "autolevels", 0, 1, 2, LEVEL_FIX}, | |
182 {"lb", "linblenddeint", 0, 1, 6, LINEAR_BLEND_DEINT_FILTER}, | |
183 {"li", "linipoldeint", 0, 1, 6, LINEAR_IPOL_DEINT_FILTER}, | |
184 {"ci", "cubicipoldeint", 0, 1, 6, CUBIC_IPOL_DEINT_FILTER}, | |
185 {"md", "mediandeint", 0, 1, 6, MEDIAN_DEINT_FILTER}, | |
156 | 186 {"tn", "tmpnoise", 1, 7, 8, TEMP_NOISE_FILTER}, |
116 | 187 {NULL, NULL,0,0,0,0} //End Marker |
188 }; | |
189 | |
190 static char *replaceTable[]= | |
191 { | |
156 | 192 "default", "hdeblock:a,vdeblock:a,dering:a,autolevels,tmpnoise:a:150:200:400", |
193 "de", "hdeblock:a,vdeblock:a,dering:a,autolevels,tmpnoise:a:150:200:400", | |
194 "fast", "x1hdeblock:a,x1vdeblock:a,dering:a,autolevels,tmpnoise:a:150:200:400", | |
195 "fa", "x1hdeblock:a,x1vdeblock:a,dering:a,autolevels,tmpnoise:a:150:200:400", | |
116 | 196 NULL //End Marker |
197 }; | |
198 | |
148
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
199 #ifdef HAVE_MMX |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
200 static inline void unusedVariableWarningFixer() |
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
201 { |
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
202 if( |
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
203 packedYOffset + packedYScale + w05 + w20 + w1400 + bm00000001 + bm00010000 |
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
204 + bm00001000 + bm10000000 + bm10000001 + bm11000011 + bm00000011 + bm11111110 |
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
205 + bm11000000 + bm00011000 + bm00110011 + bm11001100 + b00 + b01 + b02 + b0F |
148
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
206 + bFF + b20 + b04+ b08 + pQPb2 + b80 + b7E + b7C + b3F + temp0 + temp1 + temp2 + temp3 + temp4 |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
207 + temp5 + pQPb== 0) b00=0; |
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
208 } |
148
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
209 #endif |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
210 |
107 | 211 #ifdef TIMING |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
212 static inline long long rdtsc() |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
213 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
214 long long l; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
215 asm volatile( "rdtsc\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
216 : "=A" (l) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
217 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
218 // printf("%d\n", int(l/1000)); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
219 return l; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
220 } |
102 | 221 #endif |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
222 |
102 | 223 #ifdef HAVE_MMX2 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
224 static inline void prefetchnta(void *p) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
225 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
226 asm volatile( "prefetchnta (%0)\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
227 : : "r" (p) |
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 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
230 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
231 static inline void prefetcht0(void *p) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
232 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
233 asm volatile( "prefetcht0 (%0)\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
234 : : "r" (p) |
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 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
237 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
238 static inline void prefetcht1(void *p) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
239 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
240 asm volatile( "prefetcht1 (%0)\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
241 : : "r" (p) |
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 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
244 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
245 static inline void prefetcht2(void *p) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
246 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
247 asm volatile( "prefetcht2 (%0)\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
248 : : "r" (p) |
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 } |
102 | 251 #endif |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
252 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
253 //FIXME? |255-0| = 1 (shouldnt be a problem ...) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
254 /** |
111 | 255 * 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
|
256 */ |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
257 static inline int isVertDC(uint8_t src[], int stride){ |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
258 int numEq= 0; |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
259 #ifndef HAVE_MMX |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
260 int y; |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
261 #endif |
111 | 262 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
|
263 #ifdef HAVE_MMX |
119 | 264 asm volatile( |
265 "leal (%1, %2), %%eax \n\t" | |
266 "leal (%%eax, %2, 4), %%ebx \n\t" | |
267 // 0 1 2 3 4 5 6 7 8 9 | |
268 // %1 eax eax+%2 eax+2%2 %1+4%2 ebx ebx+%2 ebx+2%2 %1+8%2 ebx+4%2 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
269 "movq b7E, %%mm7 \n\t" // mm7 = 0x7F |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
270 "movq b7C, %%mm6 \n\t" // mm6 = 0x7D |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
271 "movq (%1), %%mm0 \n\t" |
119 | 272 "movq (%%eax), %%mm1 \n\t" |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
273 "psubb %%mm1, %%mm0 \n\t" // mm0 = differnece |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
274 "paddb %%mm7, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
275 "pcmpgtb %%mm6, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
276 |
119 | 277 "movq (%%eax,%2), %%mm2 \n\t" |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
278 "psubb %%mm2, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
279 "paddb %%mm7, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
280 "pcmpgtb %%mm6, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
281 "paddb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
282 |
119 | 283 "movq (%%eax, %2, 2), %%mm1 \n\t" |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
284 "psubb %%mm1, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
285 "paddb %%mm7, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
286 "pcmpgtb %%mm6, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
287 "paddb %%mm2, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
288 |
119 | 289 "movq (%1, %2, 4), %%mm2 \n\t" |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
290 "psubb %%mm2, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
291 "paddb %%mm7, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
292 "pcmpgtb %%mm6, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
293 "paddb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
294 |
119 | 295 "movq (%%ebx), %%mm1 \n\t" |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
296 "psubb %%mm1, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
297 "paddb %%mm7, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
298 "pcmpgtb %%mm6, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
299 "paddb %%mm2, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
300 |
119 | 301 "movq (%%ebx, %2), %%mm2 \n\t" |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
302 "psubb %%mm2, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
303 "paddb %%mm7, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
304 "pcmpgtb %%mm6, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
305 "paddb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
306 |
119 | 307 "movq (%%ebx, %2, 2), %%mm1 \n\t" |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
308 "psubb %%mm1, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
309 "paddb %%mm7, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
310 "pcmpgtb %%mm6, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
311 "paddb %%mm2, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
312 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
313 " \n\t" |
167 | 314 #ifdef HAVE_MMX2 |
315 "pxor %%mm7, %%mm7 \n\t" | |
316 "psadbw %%mm7, %%mm0 \n\t" | |
317 #else | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
318 "movq %%mm0, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
319 "psrlw $8, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
320 "paddb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
321 "movq %%mm0, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
322 "psrlq $16, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
323 "paddb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
324 "movq %%mm0, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
325 "psrlq $32, %%mm0 \n\t" |
167 | 326 "paddb %%mm1, %%mm0 \n\t" |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
327 #endif |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
328 "movd %%mm0, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
329 : "=r" (numEq) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
330 : "r" (src), "r" (stride) |
167 | 331 : "%ebx" |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
332 ); |
167 | 333 numEq= (-numEq) &0xFF; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
334 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
335 #else |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
336 for(y=0; y<BLOCK_SIZE-1; y++) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
337 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
338 if(((src[0] - src[0+stride] + 1)&0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
339 if(((src[1] - src[1+stride] + 1)&0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
340 if(((src[2] - src[2+stride] + 1)&0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
341 if(((src[3] - src[3+stride] + 1)&0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
342 if(((src[4] - src[4+stride] + 1)&0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
343 if(((src[5] - src[5+stride] + 1)&0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
344 if(((src[6] - src[6+stride] + 1)&0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
345 if(((src[7] - src[7+stride] + 1)&0xFFFF) < 3) numEq++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
346 src+= stride; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
347 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
348 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
349 /* if(abs(numEq - asmEq) > 0) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
350 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
351 printf("\nasm:%d c:%d\n", asmEq, numEq); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
352 for(int y=0; y<8; y++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
353 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
354 for(int x=0; x<8; x++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
355 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
356 printf("%d ", temp[x + y*stride]); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
357 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
358 printf("\n"); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
359 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
360 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
361 */ |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
362 // for(int i=0; i<numEq/8; i++) src[i]=255; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
363 return (numEq > vFlatnessThreshold) ? 1 : 0; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
364 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
365 |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
366 static inline int isVertMinMaxOk(uint8_t src[], int stride, int QP) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
367 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
368 #ifdef HAVE_MMX |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
369 int isOk; |
111 | 370 src+= stride*3; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
371 asm volatile( |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
372 // "int $3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
373 "movq (%1, %2), %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
374 "movq (%1, %2, 8), %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
375 "movq %%mm0, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
376 "psubusb %%mm1, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
377 "psubusb %%mm2, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
378 "por %%mm1, %%mm0 \n\t" // ABS Diff |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
379 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
380 "movq pQPb, %%mm7 \n\t" // QP,..., QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
381 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
382 "psubusb %%mm7, %%mm0 \n\t" // Diff <= 2QP -> 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
383 "pcmpeqd b00, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
384 "psrlq $16, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
385 "pcmpeqd bFF, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
386 // "movd %%mm0, (%1, %2, 4)\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
387 "movd %%mm0, %0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
388 : "=r" (isOk) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
389 : "r" (src), "r" (stride) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
390 ); |
120 | 391 return isOk; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
392 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
393 |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
394 int isOk2= 1; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
395 int x; |
111 | 396 src+= stride*3; |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
397 for(x=0; x<BLOCK_SIZE; x++) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
398 { |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
399 if(abs((int)src[x + stride] - (int)src[x + (stride<<3)]) > 2*QP) isOk2=0; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
400 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
401 /* if(isOk && !isOk2 || !isOk && isOk2) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
402 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
403 printf("\nasm:%d c:%d QP:%d\n", isOk, isOk2, QP); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
404 for(int y=0; y<9; y++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
405 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
406 for(int x=0; x<8; x++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
407 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
408 printf("%d ", src[x + y*stride]); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
409 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
410 printf("\n"); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
411 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
412 } */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
413 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
414 return isOk2; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
415 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
416 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
417 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
418 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
419 /** |
111 | 420 * Do a vertical low pass filter on the 8x16 block (only write to the 8x8 block in the middle) |
107 | 421 * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
422 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
423 static inline void doVertLowPass(uint8_t *src, int stride, int QP) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
424 { |
96 | 425 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
111 | 426 src+= stride*3; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
427 asm volatile( //"movv %0 %1 %2\n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
428 "movq pQPb, %%mm0 \n\t" // QP,..., QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
429 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
430 "movq (%0), %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
431 "movq (%0, %1), %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
432 "movq %%mm5, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
433 "movq %%mm6, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
434 "psubusb %%mm6, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
435 "psubusb %%mm1, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
436 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
437 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
438 "pcmpeqb b00, %%mm2 \n\t" // diff <= QP -> FF |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
439 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
440 "pand %%mm2, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
441 "pandn %%mm1, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
442 "por %%mm2, %%mm6 \n\t"// First Line to Filter |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
443 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
444 "movq (%0, %1, 8), %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
445 "leal (%0, %1, 4), %%eax \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
446 "leal (%0, %1, 8), %%ebx \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
447 "subl %1, %%ebx \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
448 "addl %1, %0 \n\t" // %0 points to line 1 not 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
449 "movq (%0, %1, 8), %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
450 "movq %%mm5, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
451 "movq %%mm7, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
452 "psubusb %%mm7, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
453 "psubusb %%mm1, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
454 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
455 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
456 "pcmpeqb b00, %%mm2 \n\t" // diff <= QP -> FF |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
457 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
458 "pand %%mm2, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
459 "pandn %%mm1, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
460 "por %%mm2, %%mm7 \n\t" // First Line to Filter |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
461 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
462 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
463 // 1 2 3 4 5 6 7 8 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
464 // %0 %0+%1 %0+2%1 eax %0+4%1 eax+2%1 ebx eax+4%1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
465 // 6 4 2 2 1 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
466 // 6 4 4 2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
467 // 6 8 2 |
111 | 468 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
469 "movq (%0, %1), %%mm0 \n\t" // 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
470 "movq %%mm0, %%mm1 \n\t" // 1 |
96 | 471 PAVGB(%%mm6, %%mm0) //1 1 /2 |
472 PAVGB(%%mm6, %%mm0) //3 1 /4 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
473 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
474 "movq (%0, %1, 4), %%mm2 \n\t" // 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
475 "movq %%mm2, %%mm5 \n\t" // 1 |
96 | 476 PAVGB((%%eax), %%mm2) // 11 /2 |
477 PAVGB((%0, %1, 2), %%mm2) // 211 /4 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
478 "movq %%mm2, %%mm3 \n\t" // 211 /4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
479 "movq (%0), %%mm4 \n\t" // 1 |
96 | 480 PAVGB(%%mm4, %%mm3) // 4 211 /8 |
481 PAVGB(%%mm0, %%mm3) //642211 /16 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
482 "movq %%mm3, (%0) \n\t" // X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
483 // mm1=2 mm2=3(211) mm4=1 mm5=5 mm6=0 mm7=9 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
484 "movq %%mm1, %%mm0 \n\t" // 1 |
96 | 485 PAVGB(%%mm6, %%mm0) //1 1 /2 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
486 "movq %%mm4, %%mm3 \n\t" // 1 |
96 | 487 PAVGB((%0,%1,2), %%mm3) // 1 1 /2 |
488 PAVGB((%%eax,%1,2), %%mm5) // 11 /2 | |
489 PAVGB((%%eax), %%mm5) // 211 /4 | |
490 PAVGB(%%mm5, %%mm3) // 2 2211 /8 | |
491 PAVGB(%%mm0, %%mm3) //4242211 /16 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
492 "movq %%mm3, (%0,%1) \n\t" // X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
493 // mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9 |
96 | 494 PAVGB(%%mm4, %%mm6) //11 /2 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
495 "movq (%%ebx), %%mm0 \n\t" // 1 |
96 | 496 PAVGB((%%eax, %1, 2), %%mm0) // 11/2 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
497 "movq %%mm0, %%mm3 \n\t" // 11/2 |
96 | 498 PAVGB(%%mm1, %%mm0) // 2 11/4 |
499 PAVGB(%%mm6, %%mm0) //222 11/8 | |
500 PAVGB(%%mm2, %%mm0) //22242211/16 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
501 "movq (%0, %1, 2), %%mm2 \n\t" // 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
502 "movq %%mm0, (%0, %1, 2) \n\t" // X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
503 // mm1=2 mm2=3 mm3=6(11) mm4=1 mm5=4(211) mm6=0(11) mm7=9 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
504 "movq (%%eax, %1, 4), %%mm0 \n\t" // 1 |
96 | 505 PAVGB((%%ebx), %%mm0) // 11 /2 |
506 PAVGB(%%mm0, %%mm6) //11 11 /4 | |
507 PAVGB(%%mm1, %%mm4) // 11 /2 | |
508 PAVGB(%%mm2, %%mm1) // 11 /2 | |
509 PAVGB(%%mm1, %%mm6) //1122 11 /8 | |
510 PAVGB(%%mm5, %%mm6) //112242211 /16 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
511 "movq (%%eax), %%mm5 \n\t" // 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
512 "movq %%mm6, (%%eax) \n\t" // X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
513 // mm0=7(11) mm1=2(11) mm2=3 mm3=6(11) mm4=1(11) mm5=4 mm7=9 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
514 "movq (%%eax, %1, 4), %%mm6 \n\t" // 1 |
96 | 515 PAVGB(%%mm7, %%mm6) // 11 /2 |
516 PAVGB(%%mm4, %%mm6) // 11 11 /4 | |
517 PAVGB(%%mm3, %%mm6) // 11 2211 /8 | |
518 PAVGB(%%mm5, %%mm2) // 11 /2 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
519 "movq (%0, %1, 4), %%mm4 \n\t" // 1 |
96 | 520 PAVGB(%%mm4, %%mm2) // 112 /4 |
521 PAVGB(%%mm2, %%mm6) // 112242211 /16 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
522 "movq %%mm6, (%0, %1, 4) \n\t" // X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
523 // mm0=7(11) mm1=2(11) mm2=3(112) mm3=6(11) mm4=5 mm5=4 mm7=9 |
96 | 524 PAVGB(%%mm7, %%mm1) // 11 2 /4 |
525 PAVGB(%%mm4, %%mm5) // 11 /2 | |
526 PAVGB(%%mm5, %%mm0) // 11 11 /4 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
527 "movq (%%eax, %1, 2), %%mm6 \n\t" // 1 |
96 | 528 PAVGB(%%mm6, %%mm1) // 11 4 2 /8 |
529 PAVGB(%%mm0, %%mm1) // 11224222 /16 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
530 "movq %%mm1, (%%eax, %1, 2) \n\t" // X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
531 // mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9 |
96 | 532 PAVGB((%%ebx), %%mm2) // 112 4 /8 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
533 "movq (%%eax, %1, 4), %%mm0 \n\t" // 1 |
96 | 534 PAVGB(%%mm0, %%mm6) // 1 1 /2 |
535 PAVGB(%%mm7, %%mm6) // 1 12 /4 | |
536 PAVGB(%%mm2, %%mm6) // 1122424 /4 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
537 "movq %%mm6, (%%ebx) \n\t" // X |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
538 // mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9 |
96 | 539 PAVGB(%%mm7, %%mm5) // 11 2 /4 |
540 PAVGB(%%mm7, %%mm5) // 11 6 /8 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
541 |
96 | 542 PAVGB(%%mm3, %%mm0) // 112 /4 |
543 PAVGB(%%mm0, %%mm5) // 112246 /16 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
544 "movq %%mm5, (%%eax, %1, 4) \n\t" // X |
140 | 545 "subl %1, %0 \n\t" |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
546 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
547 : |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
548 : "r" (src), "r" (stride) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
549 : "%eax", "%ebx" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
550 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
551 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
552 const int l1= stride; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
553 const int l2= stride + l1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
554 const int l3= stride + l2; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
555 const int l4= stride + l3; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
556 const int l5= stride + l4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
557 const int l6= stride + l5; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
558 const int l7= stride + l6; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
559 const int l8= stride + l7; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
560 const int l9= stride + l8; |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
561 int x; |
111 | 562 src+= stride*3; |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
563 for(x=0; x<BLOCK_SIZE; x++) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
564 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
565 const int first= ABS(src[0] - src[l1]) < QP ? src[0] : src[l1]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
566 const int last= ABS(src[l8] - src[l9]) < QP ? src[l9] : src[l8]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
567 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
568 int sums[9]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
569 sums[0] = first + src[l1]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
570 sums[1] = src[l1] + src[l2]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
571 sums[2] = src[l2] + src[l3]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
572 sums[3] = src[l3] + src[l4]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
573 sums[4] = src[l4] + src[l5]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
574 sums[5] = src[l5] + src[l6]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
575 sums[6] = src[l6] + src[l7]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
576 sums[7] = src[l7] + src[l8]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
577 sums[8] = src[l8] + last; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
578 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
579 src[l1]= ((sums[0]<<2) + ((first + sums[2])<<1) + sums[4] + 8)>>4; |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
580 src[l2]= ((src[l2]<<2) + ((first + sums[0] + sums[3])<<1) + sums[5] + 8)>>4; |
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
581 src[l3]= ((src[l3]<<2) + ((first + sums[1] + sums[4])<<1) + sums[6] + 8)>>4; |
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
582 src[l4]= ((src[l4]<<2) + ((sums[2] + sums[5])<<1) + sums[0] + sums[7] + 8)>>4; |
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
583 src[l5]= ((src[l5]<<2) + ((sums[3] + sums[6])<<1) + sums[1] + sums[8] + 8)>>4; |
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
584 src[l6]= ((src[l6]<<2) + ((last + sums[7] + sums[4])<<1) + sums[2] + 8)>>4; |
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
585 src[l7]= (((last + src[l7])<<2) + ((src[l8] + sums[5])<<1) + sums[3] + 8)>>4; |
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
586 src[l8]= ((sums[8]<<2) + ((last + sums[6])<<1) + sums[4] + 8)>>4; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
587 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
588 src++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
589 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
590 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
591 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
592 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
593 |
96 | 594 /** |
595 * Experimental implementation of the filter (Algorithm 1) described in a paper from Ramkishor & Karandikar | |
596 * values are correctly clipped (MMX2) | |
597 * values are wraparound (C) | |
598 * conclusion: its fast, but introduces ugly horizontal patterns if there is a continious gradient | |
599 0 8 16 24 | |
600 x = 8 | |
601 x/2 = 4 | |
602 x/8 = 1 | |
603 1 12 12 23 | |
604 */ | |
99 | 605 static inline void vertRK1Filter(uint8_t *src, int stride, int QP) |
96 | 606 { |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
607 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
111 | 608 src+= stride*3; |
96 | 609 // FIXME rounding |
610 asm volatile( | |
611 "pxor %%mm7, %%mm7 \n\t" // 0 | |
612 "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE | |
613 "leal (%0, %1), %%eax \n\t" | |
614 "leal (%%eax, %1, 4), %%ebx \n\t" | |
615 // 0 1 2 3 4 5 6 7 8 9 | |
616 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 | |
617 "movq pQPb, %%mm0 \n\t" // QP,..., QP | |
618 "movq %%mm0, %%mm1 \n\t" // QP,..., QP | |
619 "paddusb b02, %%mm0 \n\t" | |
620 "psrlw $2, %%mm0 \n\t" | |
621 "pand b3F, %%mm0 \n\t" // QP/4,..., QP/4 | |
622 "paddusb %%mm1, %%mm0 \n\t" // QP*1.25 ... | |
623 "movq (%0, %1, 4), %%mm2 \n\t" // line 4 | |
624 "movq (%%ebx), %%mm3 \n\t" // line 5 | |
625 "movq %%mm2, %%mm4 \n\t" // line 4 | |
626 "pcmpeqb %%mm5, %%mm5 \n\t" // -1 | |
627 "pxor %%mm2, %%mm5 \n\t" // -line 4 - 1 | |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
628 PAVGB(%%mm3, %%mm5) |
96 | 629 "paddb %%mm6, %%mm5 \n\t" // (l5-l4)/2 |
630 "psubusb %%mm3, %%mm4 \n\t" | |
631 "psubusb %%mm2, %%mm3 \n\t" | |
632 "por %%mm3, %%mm4 \n\t" // |l4 - l5| | |
633 "psubusb %%mm0, %%mm4 \n\t" | |
634 "pcmpeqb %%mm7, %%mm4 \n\t" | |
635 "pand %%mm4, %%mm5 \n\t" // d/2 | |
636 | |
637 // "paddb %%mm6, %%mm2 \n\t" // line 4 + 0x80 | |
638 "paddb %%mm5, %%mm2 \n\t" | |
639 // "psubb %%mm6, %%mm2 \n\t" | |
640 "movq %%mm2, (%0,%1, 4) \n\t" | |
641 | |
642 "movq (%%ebx), %%mm2 \n\t" | |
643 // "paddb %%mm6, %%mm2 \n\t" // line 5 + 0x80 | |
644 "psubb %%mm5, %%mm2 \n\t" | |
645 // "psubb %%mm6, %%mm2 \n\t" | |
646 "movq %%mm2, (%%ebx) \n\t" | |
647 | |
648 "paddb %%mm6, %%mm5 \n\t" | |
649 "psrlw $2, %%mm5 \n\t" | |
650 "pand b3F, %%mm5 \n\t" | |
651 "psubb b20, %%mm5 \n\t" // (l5-l4)/8 | |
652 | |
653 "movq (%%eax, %1, 2), %%mm2 \n\t" | |
654 "paddb %%mm6, %%mm2 \n\t" // line 3 + 0x80 | |
655 "paddsb %%mm5, %%mm2 \n\t" | |
656 "psubb %%mm6, %%mm2 \n\t" | |
657 "movq %%mm2, (%%eax, %1, 2) \n\t" | |
658 | |
659 "movq (%%ebx, %1), %%mm2 \n\t" | |
660 "paddb %%mm6, %%mm2 \n\t" // line 6 + 0x80 | |
661 "psubsb %%mm5, %%mm2 \n\t" | |
662 "psubb %%mm6, %%mm2 \n\t" | |
663 "movq %%mm2, (%%ebx, %1) \n\t" | |
664 | |
665 : | |
666 : "r" (src), "r" (stride) | |
667 : "%eax", "%ebx" | |
668 ); | |
669 #else | |
670 const int l1= stride; | |
671 const int l2= stride + l1; | |
672 const int l3= stride + l2; | |
673 const int l4= stride + l3; | |
674 const int l5= stride + l4; | |
675 const int l6= stride + l5; | |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
676 // const int l7= stride + l6; |
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
677 // const int l8= stride + l7; |
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
678 // const int l9= stride + l8; |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
679 int x; |
141 | 680 const int QP15= QP + (QP>>2); |
111 | 681 src+= stride*3; |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
682 for(x=0; x<BLOCK_SIZE; x++) |
96 | 683 { |
141 | 684 const int v = (src[x+l5] - src[x+l4]); |
685 if(ABS(v) < QP15) | |
96 | 686 { |
141 | 687 src[x+l3] +=v>>3; |
688 src[x+l4] +=v>>1; | |
689 src[x+l5] -=v>>1; | |
690 src[x+l6] -=v>>3; | |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
691 |
96 | 692 } |
693 } | |
694 | |
695 #endif | |
696 } | |
697 | |
698 /** | |
699 * Experimental Filter 1 | |
99 | 700 * will not damage linear gradients |
701 * Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter | |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
702 * can only smooth blocks at the expected locations (it cant smooth them if they did move) |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
703 * MMX2 version does correct clipping C version doesnt |
96 | 704 */ |
705 static inline void vertX1Filter(uint8_t *src, int stride, int QP) | |
706 { | |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
707 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
111 | 708 src+= stride*3; |
709 | |
96 | 710 asm volatile( |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
711 "pxor %%mm7, %%mm7 \n\t" // 0 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
712 // "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
713 "leal (%0, %1), %%eax \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
714 "leal (%%eax, %1, 4), %%ebx \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
715 // 0 1 2 3 4 5 6 7 8 9 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
716 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
717 "movq (%%eax, %1, 2), %%mm0 \n\t" // line 3 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
718 "movq (%0, %1, 4), %%mm1 \n\t" // line 4 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
719 "movq %%mm1, %%mm2 \n\t" // line 4 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
720 "psubusb %%mm0, %%mm1 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
721 "psubusb %%mm2, %%mm0 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
722 "por %%mm1, %%mm0 \n\t" // |l2 - l3| |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
723 "movq (%%ebx), %%mm3 \n\t" // line 5 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
724 "movq (%%ebx, %1), %%mm4 \n\t" // line 6 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
725 "movq %%mm3, %%mm5 \n\t" // line 5 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
726 "psubusb %%mm4, %%mm3 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
727 "psubusb %%mm5, %%mm4 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
728 "por %%mm4, %%mm3 \n\t" // |l5 - l6| |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
729 PAVGB(%%mm3, %%mm0) // (|l2 - l3| + |l5 - l6|)/2 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
730 "movq %%mm2, %%mm1 \n\t" // line 4 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
731 "psubusb %%mm5, %%mm2 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
732 "movq %%mm2, %%mm4 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
733 "pcmpeqb %%mm7, %%mm2 \n\t" // (l4 - l5) <= 0 ? -1 : 0 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
734 "psubusb %%mm1, %%mm5 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
735 "por %%mm5, %%mm4 \n\t" // |l4 - l5| |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
736 "psubusb %%mm0, %%mm4 \n\t" //d = MAX(0, |l4-l5| - (|l2-l3| + |l5-l6|)/2) |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
737 "movq %%mm4, %%mm3 \n\t" // d |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
738 "psubusb pQPb, %%mm4 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
739 "pcmpeqb %%mm7, %%mm4 \n\t" // d <= QP ? -1 : 0 |
99 | 740 "psubusb b01, %%mm3 \n\t" |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
741 "pand %%mm4, %%mm3 \n\t" // d <= QP ? d : 0 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
742 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
743 PAVGB(%%mm7, %%mm3) // d/2 |
99 | 744 "movq %%mm3, %%mm1 \n\t" // d/2 |
745 PAVGB(%%mm7, %%mm3) // d/4 | |
746 PAVGB(%%mm1, %%mm3) // 3*d/8 | |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
747 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
748 "movq (%0, %1, 4), %%mm0 \n\t" // line 4 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
749 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
750 "psubusb %%mm3, %%mm0 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
751 "pxor %%mm2, %%mm0 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
752 "movq %%mm0, (%0, %1, 4) \n\t" // line 4 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
753 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
754 "movq (%%ebx), %%mm0 \n\t" // line 5 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
755 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
756 "paddusb %%mm3, %%mm0 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
757 "pxor %%mm2, %%mm0 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
758 "movq %%mm0, (%%ebx) \n\t" // line 5 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
759 |
99 | 760 PAVGB(%%mm7, %%mm1) // d/4 |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
761 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
762 "movq (%%eax, %1, 2), %%mm0 \n\t" // line 3 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
763 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4 |
99 | 764 "psubusb %%mm1, %%mm0 \n\t" |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
765 "pxor %%mm2, %%mm0 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
766 "movq %%mm0, (%%eax, %1, 2) \n\t" // line 3 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
767 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
768 "movq (%%ebx, %1), %%mm0 \n\t" // line 6 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
769 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5 |
99 | 770 "paddusb %%mm1, %%mm0 \n\t" |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
771 "pxor %%mm2, %%mm0 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
772 "movq %%mm0, (%%ebx, %1) \n\t" // line 6 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
773 |
99 | 774 PAVGB(%%mm7, %%mm1) // d/8 |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
775 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
776 "movq (%%eax, %1), %%mm0 \n\t" // line 2 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
777 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l2-1 : l2 |
99 | 778 "psubusb %%mm1, %%mm0 \n\t" |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
779 "pxor %%mm2, %%mm0 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
780 "movq %%mm0, (%%eax, %1) \n\t" // line 2 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
781 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
782 "movq (%%ebx, %1, 2), %%mm0 \n\t" // line 7 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
783 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l7-1 : l7 |
99 | 784 "paddusb %%mm1, %%mm0 \n\t" |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
785 "pxor %%mm2, %%mm0 \n\t" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
786 "movq %%mm0, (%%ebx, %1, 2) \n\t" // line 7 |
96 | 787 |
788 : | |
789 : "r" (src), "r" (stride) | |
790 : "%eax", "%ebx" | |
791 ); | |
792 #else | |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
793 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
794 const int l1= stride; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
795 const int l2= stride + l1; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
796 const int l3= stride + l2; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
797 const int l4= stride + l3; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
798 const int l5= stride + l4; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
799 const int l6= stride + l5; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
800 const int l7= stride + l6; |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
801 // const int l8= stride + l7; |
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
802 // const int l9= stride + l8; |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
803 int x; |
111 | 804 |
805 src+= stride*3; | |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
806 for(x=0; x<BLOCK_SIZE; x++) |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
807 { |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
808 int a= src[l3] - src[l4]; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
809 int b= src[l4] - src[l5]; |
99 | 810 int c= src[l5] - src[l6]; |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
811 |
141 | 812 int d= ABS(b) - ((ABS(a) + ABS(c))>>1); |
813 d= MAX(d, 0); | |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
814 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
815 if(d < QP) |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
816 { |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
817 int v = d * SIGN(-b); |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
818 |
141 | 819 src[l2] +=v>>3; |
820 src[l3] +=v>>2; | |
821 src[l4] +=(3*v)>>3; | |
822 src[l5] -=(3*v)>>3; | |
823 src[l6] -=v>>2; | |
824 src[l7] -=v>>3; | |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
825 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
826 } |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
827 src++; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
828 } |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
829 /* |
96 | 830 const int l1= stride; |
831 const int l2= stride + l1; | |
832 const int l3= stride + l2; | |
833 const int l4= stride + l3; | |
834 const int l5= stride + l4; | |
835 const int l6= stride + l5; | |
836 const int l7= stride + l6; | |
837 const int l8= stride + l7; | |
838 const int l9= stride + l8; | |
839 for(int x=0; x<BLOCK_SIZE; x++) | |
840 { | |
841 int v2= src[l2]; | |
842 int v3= src[l3]; | |
843 int v4= src[l4]; | |
844 int v5= src[l5]; | |
845 int v6= src[l6]; | |
846 int v7= src[l7]; | |
847 | |
848 if(ABS(v4-v5)<QP && ABS(v4-v5) - (ABS(v3-v4) + ABS(v5-v6))>0 ) | |
849 { | |
850 src[l3] = (6*v2 + 4*v3 + 3*v4 + 2*v5 + v6 )/16; | |
851 src[l4] = (3*v2 + 3*v3 + 4*v4 + 3*v5 + 2*v6 + v7 )/16; | |
852 src[l5] = (1*v2 + 2*v3 + 3*v4 + 4*v5 + 3*v6 + 3*v7)/16; | |
853 src[l6] = ( 1*v3 + 2*v4 + 3*v5 + 4*v6 + 6*v7)/16; | |
854 } | |
855 src++; | |
856 } | |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
857 */ |
96 | 858 #endif |
859 } | |
860 | |
165 | 861 /** |
862 * Experimental Filter 1 (Horizontal) | |
863 * will not damage linear gradients | |
864 * Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter | |
865 * can only smooth blocks at the expected locations (it cant smooth them if they did move) | |
866 * MMX2 version does correct clipping C version doesnt | |
867 * not identical with the vertical one | |
868 */ | |
869 static inline void horizX1Filter(uint8_t *src, int stride, int QP) | |
870 { | |
871 int y; | |
872 //FIXME (has little in common with the mmx2 version) | |
873 for(y=0; y<BLOCK_SIZE; y++) | |
874 { | |
875 int a= src[1] - src[2]; | |
876 int b= src[3] - src[4]; | |
877 int c= src[5] - src[6]; | |
878 | |
879 int d= MAX(ABS(b) - (ABS(a) + ABS(c))/2, 0); | |
880 | |
881 if(d < QP) | |
882 { | |
883 int v = d * SIGN(-b); | |
884 | |
885 src[1] +=v/8; | |
886 src[2] +=v/4; | |
887 src[3] +=3*v/8; | |
888 src[4] -=3*v/8; | |
889 src[5] -=v/4; | |
890 src[6] -=v/8; | |
891 | |
892 } | |
893 src+=stride; | |
894 } | |
895 } | |
896 | |
897 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
898 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
|
899 { |
163 | 900 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
901 /* | |
902 uint8_t tmp[16]; | |
903 const int l1= stride; | |
904 const int l2= stride + l1; | |
905 const int l3= stride + l2; | |
906 const int l4= (int)tmp - (int)src - stride*3; | |
907 const int l5= (int)tmp - (int)src - stride*3 + 8; | |
908 const int l6= stride*3 + l3; | |
909 const int l7= stride + l6; | |
910 const int l8= stride + l7; | |
911 | |
912 memcpy(tmp, src+stride*7, 8); | |
913 memcpy(tmp+8, src+stride*8, 8); | |
914 */ | |
111 | 915 src+= stride*4; |
163 | 916 asm volatile( |
917 | |
918 #if 0 //sligtly more accurate and slightly slower | |
919 "pxor %%mm7, %%mm7 \n\t" // 0 | |
920 "leal (%0, %1), %%eax \n\t" | |
921 "leal (%%eax, %1, 4), %%ebx \n\t" | |
922 // 0 1 2 3 4 5 6 7 | |
923 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ebx+%1 ebx+2%1 | |
924 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 | |
925 | |
926 | |
927 "movq (%0, %1, 2), %%mm0 \n\t" // l2 | |
928 "movq (%0), %%mm1 \n\t" // l0 | |
929 "movq %%mm0, %%mm2 \n\t" // l2 | |
930 PAVGB(%%mm7, %%mm0) // ~l2/2 | |
931 PAVGB(%%mm1, %%mm0) // ~(l2 + 2l0)/4 | |
932 PAVGB(%%mm2, %%mm0) // ~(5l2 + 2l0)/8 | |
933 | |
934 "movq (%%eax), %%mm1 \n\t" // l1 | |
935 "movq (%%eax, %1, 2), %%mm3 \n\t" // l3 | |
936 "movq %%mm1, %%mm4 \n\t" // l1 | |
937 PAVGB(%%mm7, %%mm1) // ~l1/2 | |
938 PAVGB(%%mm3, %%mm1) // ~(l1 + 2l3)/4 | |
939 PAVGB(%%mm4, %%mm1) // ~(5l1 + 2l3)/8 | |
940 | |
941 "movq %%mm0, %%mm4 \n\t" // ~(5l2 + 2l0)/8 | |
942 "psubusb %%mm1, %%mm0 \n\t" | |
943 "psubusb %%mm4, %%mm1 \n\t" | |
944 "por %%mm0, %%mm1 \n\t" // ~|2l0 - 5l1 + 5l2 - 2l3|/8 | |
945 // mm1= |lenergy|, mm2= l2, mm3= l3, mm7=0 | |
946 | |
947 "movq (%0, %1, 4), %%mm0 \n\t" // l4 | |
948 "movq %%mm0, %%mm4 \n\t" // l4 | |
949 PAVGB(%%mm7, %%mm0) // ~l4/2 | |
950 PAVGB(%%mm2, %%mm0) // ~(l4 + 2l2)/4 | |
951 PAVGB(%%mm4, %%mm0) // ~(5l4 + 2l2)/8 | |
952 | |
953 "movq (%%ebx), %%mm2 \n\t" // l5 | |
954 "movq %%mm3, %%mm5 \n\t" // l3 | |
955 PAVGB(%%mm7, %%mm3) // ~l3/2 | |
956 PAVGB(%%mm2, %%mm3) // ~(l3 + 2l5)/4 | |
957 PAVGB(%%mm5, %%mm3) // ~(5l3 + 2l5)/8 | |
958 | |
959 "movq %%mm0, %%mm6 \n\t" // ~(5l4 + 2l2)/8 | |
960 "psubusb %%mm3, %%mm0 \n\t" | |
961 "psubusb %%mm6, %%mm3 \n\t" | |
962 "por %%mm0, %%mm3 \n\t" // ~|2l2 - 5l3 + 5l4 - 2l5|/8 | |
963 "pcmpeqb %%mm7, %%mm0 \n\t" // SIGN(2l2 - 5l3 + 5l4 - 2l5) | |
964 // mm0= SIGN(menergy), mm1= |lenergy|, mm2= l5, mm3= |menergy|, mm4=l4, mm5= l3, mm7=0 | |
965 | |
966 "movq (%%ebx, %1), %%mm6 \n\t" // l6 | |
967 "movq %%mm6, %%mm5 \n\t" // l6 | |
968 PAVGB(%%mm7, %%mm6) // ~l6/2 | |
969 PAVGB(%%mm4, %%mm6) // ~(l6 + 2l4)/4 | |
970 PAVGB(%%mm5, %%mm6) // ~(5l6 + 2l4)/8 | |
971 | |
972 "movq (%%ebx, %1, 2), %%mm5 \n\t" // l7 | |
973 "movq %%mm2, %%mm4 \n\t" // l5 | |
974 PAVGB(%%mm7, %%mm2) // ~l5/2 | |
975 PAVGB(%%mm5, %%mm2) // ~(l5 + 2l7)/4 | |
976 PAVGB(%%mm4, %%mm2) // ~(5l5 + 2l7)/8 | |
977 | |
978 "movq %%mm6, %%mm4 \n\t" // ~(5l6 + 2l4)/8 | |
979 "psubusb %%mm2, %%mm6 \n\t" | |
980 "psubusb %%mm4, %%mm2 \n\t" | |
981 "por %%mm6, %%mm2 \n\t" // ~|2l4 - 5l5 + 5l6 - 2l7|/8 | |
982 // mm0= SIGN(menergy), mm1= |lenergy|/8, mm2= |renergy|/8, mm3= |menergy|/8, mm7=0 | |
983 | |
984 | |
985 PMINUB(%%mm2, %%mm1, %%mm4) // MIN(|lenergy|,|renergy|)/8 | |
986 "movq pQPb, %%mm4 \n\t" // QP //FIXME QP+1 ? | |
987 "paddusb b01, %%mm4 \n\t" | |
988 "pcmpgtb %%mm3, %%mm4 \n\t" // |menergy|/8 < QP | |
989 "psubusb %%mm1, %%mm3 \n\t" // d=|menergy|/8-MIN(|lenergy|,|renergy|)/8 | |
990 "pand %%mm4, %%mm3 \n\t" | |
991 | |
992 "movq %%mm3, %%mm1 \n\t" | |
993 // "psubusb b01, %%mm3 \n\t" | |
994 PAVGB(%%mm7, %%mm3) | |
995 PAVGB(%%mm7, %%mm3) | |
996 "paddusb %%mm1, %%mm3 \n\t" | |
997 // "paddusb b01, %%mm3 \n\t" | |
998 | |
999 "movq (%%eax, %1, 2), %%mm6 \n\t" //l3 | |
1000 "movq (%0, %1, 4), %%mm5 \n\t" //l4 | |
1001 "movq (%0, %1, 4), %%mm4 \n\t" //l4 | |
1002 "psubusb %%mm6, %%mm5 \n\t" | |
1003 "psubusb %%mm4, %%mm6 \n\t" | |
1004 "por %%mm6, %%mm5 \n\t" // |l3-l4| | |
1005 "pcmpeqb %%mm7, %%mm6 \n\t" // SIGN(l3-l4) | |
1006 "pxor %%mm6, %%mm0 \n\t" | |
1007 "pand %%mm0, %%mm3 \n\t" | |
1008 PMINUB(%%mm5, %%mm3, %%mm0) | |
1009 | |
1010 "psubusb b01, %%mm3 \n\t" | |
1011 PAVGB(%%mm7, %%mm3) | |
1012 | |
1013 "movq (%%eax, %1, 2), %%mm0 \n\t" | |
1014 "movq (%0, %1, 4), %%mm2 \n\t" | |
1015 "pxor %%mm6, %%mm0 \n\t" | |
1016 "pxor %%mm6, %%mm2 \n\t" | |
1017 "psubb %%mm3, %%mm0 \n\t" | |
1018 "paddb %%mm3, %%mm2 \n\t" | |
1019 "pxor %%mm6, %%mm0 \n\t" | |
1020 "pxor %%mm6, %%mm2 \n\t" | |
1021 "movq %%mm0, (%%eax, %1, 2) \n\t" | |
1022 "movq %%mm2, (%0, %1, 4) \n\t" | |
1023 #endif | |
1024 | |
1025 "leal (%0, %1), %%eax \n\t" | |
1026 "pcmpeqb %%mm6, %%mm6 \n\t" // -1 | |
1027 // 0 1 2 3 4 5 6 7 | |
1028 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ebx+%1 ebx+2%1 | |
1029 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 | |
1030 | |
1031 | |
1032 "movq (%%eax, %1, 2), %%mm1 \n\t" // l3 | |
1033 "movq (%0, %1, 4), %%mm0 \n\t" // l4 | |
1034 "pxor %%mm6, %%mm1 \n\t" // -l3-1 | |
1035 PAVGB(%%mm1, %%mm0) // -q+128 = (l4-l3+256)/2 | |
1036 // mm1=-l3-1, mm0=128-q | |
1037 | |
1038 "movq (%%eax, %1, 4), %%mm2 \n\t" // l5 | |
1039 "movq (%%eax, %1), %%mm3 \n\t" // l2 | |
1040 "pxor %%mm6, %%mm2 \n\t" // -l5-1 | |
1041 "movq %%mm2, %%mm5 \n\t" // -l5-1 | |
1042 "movq b80, %%mm4 \n\t" // 128 | |
1043 "leal (%%eax, %1, 4), %%ebx \n\t" | |
1044 PAVGB(%%mm3, %%mm2) // (l2-l5+256)/2 | |
1045 PAVGB(%%mm0, %%mm4) // ~(l4-l3)/4 + 128 | |
1046 PAVGB(%%mm2, %%mm4) // ~(l2-l5)/4 +(l4-l3)/8 + 128 | |
1047 PAVGB(%%mm0, %%mm4) // ~(l2-l5)/8 +5(l4-l3)/16 + 128 | |
1048 // mm1=-l3-1, mm0=128-q, mm3=l2, mm4=menergy/16 + 128, mm5= -l5-1 | |
1049 | |
1050 "movq (%%eax), %%mm2 \n\t" // l1 | |
1051 "pxor %%mm6, %%mm2 \n\t" // -l1-1 | |
1052 PAVGB(%%mm3, %%mm2) // (l2-l1+256)/2 | |
1053 PAVGB((%0), %%mm1) // (l0-l3+256)/2 | |
1054 "movq b80, %%mm3 \n\t" // 128 | |
1055 PAVGB(%%mm2, %%mm3) // ~(l2-l1)/4 + 128 | |
1056 PAVGB(%%mm1, %%mm3) // ~(l0-l3)/4 +(l2-l1)/8 + 128 | |
1057 PAVGB(%%mm2, %%mm3) // ~(l0-l3)/8 +5(l2-l1)/16 + 128 | |
1058 // mm0=128-q, mm3=lenergy/16 + 128, mm4= menergy/16 + 128, mm5= -l5-1 | |
1059 | |
1060 PAVGB((%%ebx, %1), %%mm5) // (l6-l5+256)/2 | |
1061 "movq (%%ebx, %1, 2), %%mm1 \n\t" // l7 | |
1062 "pxor %%mm6, %%mm1 \n\t" // -l7-1 | |
1063 PAVGB((%0, %1, 4), %%mm1) // (l4-l7+256)/2 | |
1064 "movq b80, %%mm2 \n\t" // 128 | |
1065 PAVGB(%%mm5, %%mm2) // ~(l6-l5)/4 + 128 | |
1066 PAVGB(%%mm1, %%mm2) // ~(l4-l7)/4 +(l6-l5)/8 + 128 | |
1067 PAVGB(%%mm5, %%mm2) // ~(l4-l7)/8 +5(l6-l5)/16 + 128 | |
1068 // mm0=128-q, mm2=renergy/16 + 128, mm3=lenergy/16 + 128, mm4= menergy/16 + 128 | |
1069 | |
1070 "movq b00, %%mm1 \n\t" // 0 | |
1071 "movq b00, %%mm5 \n\t" // 0 | |
1072 "psubb %%mm2, %%mm1 \n\t" // 128 - renergy/16 | |
1073 "psubb %%mm3, %%mm5 \n\t" // 128 - lenergy/16 | |
1074 PMAXUB(%%mm1, %%mm2) // 128 + |renergy/16| | |
1075 PMAXUB(%%mm5, %%mm3) // 128 + |lenergy/16| | |
1076 PMINUB(%%mm2, %%mm3, %%mm1) // 128 + MIN(|lenergy|,|renergy|)/16 | |
1077 | |
1078 // mm0=128-q, mm3=128 + MIN(|lenergy|,|renergy|)/16, mm4= menergy/16 + 128 | |
1079 | |
1080 "movq b00, %%mm7 \n\t" // 0 | |
1081 "movq pQPb, %%mm2 \n\t" // QP | |
1082 PAVGB(%%mm6, %%mm2) // 128 + QP/2 | |
1083 "psubb %%mm6, %%mm2 \n\t" | |
1084 | |
1085 "movq %%mm4, %%mm1 \n\t" | |
1086 "pcmpgtb %%mm7, %%mm1 \n\t" // SIGN(menergy) | |
1087 "pxor %%mm1, %%mm4 \n\t" | |
1088 "psubb %%mm1, %%mm4 \n\t" // 128 + |menergy|/16 | |
1089 "pcmpgtb %%mm4, %%mm2 \n\t" // |menergy|/16 < QP/2 | |
1090 "psubusb %%mm3, %%mm4 \n\t" //d=|menergy|/16 - MIN(|lenergy|,|renergy|)/16 | |
1091 // mm0=128-q, mm1= SIGN(menergy), mm2= |menergy|/16 < QP/2, mm4= d/16 | |
1092 | |
1093 "movq %%mm4, %%mm3 \n\t" // d | |
1094 "psubusb b01, %%mm4 \n\t" | |
1095 PAVGB(%%mm7, %%mm4) // d/32 | |
1096 PAVGB(%%mm7, %%mm4) // (d + 32)/64 | |
1097 "paddb %%mm3, %%mm4 \n\t" // 5d/64 | |
1098 "pand %%mm2, %%mm4 \n\t" | |
1099 | |
1100 "movq b80, %%mm5 \n\t" // 128 | |
1101 "psubb %%mm0, %%mm5 \n\t" // q | |
1102 "paddsb %%mm6, %%mm5 \n\t" // fix bad rounding | |
1103 "pcmpgtb %%mm5, %%mm7 \n\t" // SIGN(q) | |
1104 "pxor %%mm7, %%mm5 \n\t" | |
1105 | |
1106 PMINUB(%%mm5, %%mm4, %%mm3) // MIN(|q|, 5d/64) | |
1107 "pxor %%mm1, %%mm7 \n\t" // SIGN(d*q) | |
1108 | |
1109 "pand %%mm7, %%mm4 \n\t" | |
1110 "movq (%%eax, %1, 2), %%mm0 \n\t" | |
1111 "movq (%0, %1, 4), %%mm2 \n\t" | |
1112 "pxor %%mm1, %%mm0 \n\t" | |
1113 "pxor %%mm1, %%mm2 \n\t" | |
1114 "paddb %%mm4, %%mm0 \n\t" | |
1115 "psubb %%mm4, %%mm2 \n\t" | |
1116 "pxor %%mm1, %%mm0 \n\t" | |
1117 "pxor %%mm1, %%mm2 \n\t" | |
1118 "movq %%mm0, (%%eax, %1, 2) \n\t" | |
1119 "movq %%mm2, (%0, %1, 4) \n\t" | |
1120 | |
1121 : | |
1122 : "r" (src), "r" (stride) | |
1123 : "%eax", "%ebx" | |
1124 ); | |
1125 | |
1126 /* | |
1127 { | |
1128 int x; | |
1129 src-= stride; | |
1130 for(x=0; x<BLOCK_SIZE; x++) | |
1131 { | |
1132 const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]); | |
1133 if(ABS(middleEnergy)< 8*QP) | |
1134 { | |
1135 const int q=(src[l4] - src[l5])/2; | |
1136 const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]); | |
1137 const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]); | |
1138 | |
1139 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) ); | |
1140 d= MAX(d, 0); | |
1141 | |
1142 d= (5*d + 32) >> 6; | |
1143 d*= SIGN(-middleEnergy); | |
1144 | |
1145 if(q>0) | |
1146 { | |
1147 d= d<0 ? 0 : d; | |
1148 d= d>q ? q : d; | |
1149 } | |
1150 else | |
1151 { | |
1152 d= d>0 ? 0 : d; | |
1153 d= d<q ? q : d; | |
1154 } | |
1155 | |
1156 src[l4]-= d; | |
1157 src[l5]+= d; | |
1158 } | |
1159 src++; | |
1160 } | |
1161 src-=8; | |
1162 for(x=0; x<8; x++) | |
1163 { | |
1164 int y; | |
1165 for(y=4; y<6; y++) | |
1166 { | |
1167 int d= src[x+y*stride] - tmp[x+(y-4)*8]; | |
1168 int ad= ABS(d); | |
1169 static int max=0; | |
1170 static int sum=0; | |
1171 static int num=0; | |
1172 static int bias=0; | |
1173 | |
1174 if(max<ad) max=ad; | |
1175 sum+= ad>3 ? 1 : 0; | |
1176 if(ad>3) | |
1177 { | |
1178 src[0] = src[7] = src[stride*7] = src[(stride+1)*7]=255; | |
1179 } | |
1180 if(y==4) bias+=d; | |
1181 num++; | |
1182 if(num%1000000 == 0) | |
1183 { | |
1184 printf(" %d %d %d %d\n", num, sum, max, bias); | |
1185 } | |
1186 } | |
1187 } | |
1188 } | |
1189 */ | |
1190 #elif defined (HAVE_MMX) | |
1191 src+= stride*4; | |
1192 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1193 asm volatile( |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1194 "pxor %%mm7, %%mm7 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1195 "leal (%0, %1), %%eax \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1196 "leal (%%eax, %1, 4), %%ebx \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1197 // 0 1 2 3 4 5 6 7 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1198 // %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
|
1199 // %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
|
1200 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1201 "movq (%0), %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1202 "movq %%mm0, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1203 "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1204 "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1205 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1206 "movq (%%eax), %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1207 "movq %%mm2, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1208 "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1209 "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1210 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1211 "movq (%%eax, %1), %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1212 "movq %%mm4, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1213 "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1214 "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1215 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1216 "paddw %%mm0, %%mm0 \n\t" // 2L0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1217 "paddw %%mm1, %%mm1 \n\t" // 2H0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1218 "psubw %%mm4, %%mm2 \n\t" // L1 - L2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1219 "psubw %%mm5, %%mm3 \n\t" // H1 - H2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1220 "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1221 "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1222 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1223 "psllw $2, %%mm2 \n\t" // 4L1 - 4L2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1224 "psllw $2, %%mm3 \n\t" // 4H1 - 4H2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1225 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1226 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1227 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1228 "movq (%%eax, %1, 2), %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1229 "movq %%mm2, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1230 "punpcklbw %%mm7, %%mm2 \n\t" // L3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1231 "punpckhbw %%mm7, %%mm3 \n\t" // H3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1232 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1233 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1234 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1235 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1236 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1237 "movq %%mm0, temp0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1238 "movq %%mm1, temp1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1239 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1240 "movq (%0, %1, 4), %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1241 "movq %%mm0, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1242 "punpcklbw %%mm7, %%mm0 \n\t" // L4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1243 "punpckhbw %%mm7, %%mm1 \n\t" // H4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1244 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1245 "psubw %%mm0, %%mm2 \n\t" // L3 - L4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1246 "psubw %%mm1, %%mm3 \n\t" // H3 - H4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1247 "movq %%mm2, temp2 \n\t" // L3 - L4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1248 "movq %%mm3, temp3 \n\t" // H3 - H4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1249 "paddw %%mm4, %%mm4 \n\t" // 2L2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1250 "paddw %%mm5, %%mm5 \n\t" // 2H2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1251 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1252 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1253 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1254 "psllw $2, %%mm2 \n\t" // 4L3 - 4L4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1255 "psllw $2, %%mm3 \n\t" // 4H3 - 4H4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1256 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1257 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1258 //50 opcodes so far |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1259 "movq (%%ebx), %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1260 "movq %%mm2, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1261 "punpcklbw %%mm7, %%mm2 \n\t" // L5 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1262 "punpckhbw %%mm7, %%mm3 \n\t" // H5 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1263 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1264 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1265 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1266 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1267 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1268 "movq (%%ebx, %1), %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1269 "punpcklbw %%mm7, %%mm6 \n\t" // L6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1270 "psubw %%mm6, %%mm2 \n\t" // L5 - L6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1271 "movq (%%ebx, %1), %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1272 "punpckhbw %%mm7, %%mm6 \n\t" // H6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1273 "psubw %%mm6, %%mm3 \n\t" // H5 - H6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1274 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1275 "paddw %%mm0, %%mm0 \n\t" // 2L4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1276 "paddw %%mm1, %%mm1 \n\t" // 2H4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1277 "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1278 "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6 |
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 "psllw $2, %%mm2 \n\t" // 4L5 - 4L6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1281 "psllw $2, %%mm3 \n\t" // 4H5 - 4H6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1282 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1283 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1284 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1285 "movq (%%ebx, %1, 2), %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1286 "movq %%mm2, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1287 "punpcklbw %%mm7, %%mm2 \n\t" // L7 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1288 "punpckhbw %%mm7, %%mm3 \n\t" // H7 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1289 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1290 "paddw %%mm2, %%mm2 \n\t" // 2L7 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1291 "paddw %%mm3, %%mm3 \n\t" // 2H7 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1292 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1293 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1294 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1295 "movq temp0, %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1296 "movq temp1, %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
140 | 1297 |
1298 #ifdef HAVE_MMX2 | |
1299 "movq %%mm7, %%mm6 \n\t" // 0 | |
1300 "psubw %%mm0, %%mm6 \n\t" | |
1301 "pmaxsw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| | |
1302 "movq %%mm7, %%mm6 \n\t" // 0 | |
1303 "psubw %%mm1, %%mm6 \n\t" | |
1304 "pmaxsw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7| | |
1305 "movq %%mm7, %%mm6 \n\t" // 0 | |
1306 "psubw %%mm2, %%mm6 \n\t" | |
1307 "pmaxsw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3| | |
1308 "movq %%mm7, %%mm6 \n\t" // 0 | |
1309 "psubw %%mm3, %%mm6 \n\t" | |
1310 "pmaxsw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3| | |
1311 #else | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1312 "movq %%mm7, %%mm6 \n\t" // 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1313 "pcmpgtw %%mm0, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1314 "pxor %%mm6, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1315 "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1316 "movq %%mm7, %%mm6 \n\t" // 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1317 "pcmpgtw %%mm1, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1318 "pxor %%mm6, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1319 "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7| |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1320 "movq %%mm7, %%mm6 \n\t" // 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1321 "pcmpgtw %%mm2, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1322 "pxor %%mm6, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1323 "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3| |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1324 "movq %%mm7, %%mm6 \n\t" // 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1325 "pcmpgtw %%mm3, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1326 "pxor %%mm6, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1327 "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3| |
140 | 1328 #endif |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1329 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1330 #ifdef HAVE_MMX2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1331 "pminsw %%mm2, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1332 "pminsw %%mm3, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1333 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1334 "movq %%mm0, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1335 "psubusw %%mm2, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1336 "psubw %%mm6, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1337 "movq %%mm1, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1338 "psubusw %%mm3, %%mm6 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1339 "psubw %%mm6, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1340 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1341 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1342 "movq %%mm7, %%mm6 \n\t" // 0 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1343 "pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1344 "pxor %%mm6, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1345 "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5| |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1346 "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1347 "pxor %%mm7, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1348 "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5| |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1349 // 100 opcodes |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1350 "movd %2, %%mm2 \n\t" // QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1351 "punpcklwd %%mm2, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1352 "punpcklwd %%mm2, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1353 "psllw $3, %%mm2 \n\t" // 8QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1354 "movq %%mm2, %%mm3 \n\t" // 8QP |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1355 "pcmpgtw %%mm4, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1356 "pcmpgtw %%mm5, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1357 "pand %%mm2, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1358 "pand %%mm3, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1359 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1360 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1361 "psubusw %%mm0, %%mm4 \n\t" // hd |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1362 "psubusw %%mm1, %%mm5 \n\t" // ld |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1363 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1364 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1365 "movq w05, %%mm2 \n\t" // 5 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1366 "pmullw %%mm2, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1367 "pmullw %%mm2, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1368 "movq w20, %%mm2 \n\t" // 32 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1369 "paddw %%mm2, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1370 "paddw %%mm2, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1371 "psrlw $6, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1372 "psrlw $6, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1373 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1374 /* |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1375 "movq w06, %%mm2 \n\t" // 6 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1376 "paddw %%mm2, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1377 "paddw %%mm2, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1378 "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
|
1379 //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
|
1380 "pmulhw %%mm2, %%mm4 \n\t" // hd/13 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1381 "pmulhw %%mm2, %%mm5 \n\t" // ld/13 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1382 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1383 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1384 "movq temp2, %%mm0 \n\t" // L3 - L4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1385 "movq temp3, %%mm1 \n\t" // H3 - H4 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1386 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1387 "pxor %%mm2, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1388 "pxor %%mm3, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1389 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1390 "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1391 "pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1392 "pxor %%mm2, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1393 "pxor %%mm3, %%mm1 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1394 "psubw %%mm2, %%mm0 \n\t" // |L3-L4| |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1395 "psubw %%mm3, %%mm1 \n\t" // |H3-H4| |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
1396 "psrlw $1, %%mm0 \n\t" // |L3 - L4|/2 |
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
1397 "psrlw $1, %%mm1 \n\t" // |H3 - H4|/2 |
95
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 "pxor %%mm6, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1400 "pxor %%mm7, %%mm3 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1401 "pand %%mm2, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1402 "pand %%mm3, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1403 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1404 #ifdef HAVE_MMX2 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1405 "pminsw %%mm0, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1406 "pminsw %%mm1, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1407 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1408 "movq %%mm4, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1409 "psubusw %%mm0, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1410 "psubw %%mm2, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1411 "movq %%mm5, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1412 "psubusw %%mm1, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1413 "psubw %%mm2, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1414 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1415 "pxor %%mm6, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1416 "pxor %%mm7, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1417 "psubw %%mm6, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1418 "psubw %%mm7, %%mm5 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1419 "packsswb %%mm5, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1420 "movq (%%eax, %1, 2), %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1421 "paddb %%mm4, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1422 "movq %%mm0, (%%eax, %1, 2) \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1423 "movq (%0, %1, 4), %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1424 "psubb %%mm4, %%mm0 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1425 "movq %%mm0, (%0, %1, 4) \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1426 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1427 : |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1428 : "r" (src), "r" (stride), "r" (QP) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1429 : "%eax", "%ebx" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1430 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1431 #else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1432 const int l1= stride; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1433 const int l2= stride + l1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1434 const int l3= stride + l2; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1435 const int l4= stride + l3; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1436 const int l5= stride + l4; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1437 const int l6= stride + l5; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1438 const int l7= stride + l6; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1439 const int l8= stride + l7; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1440 // const int l9= stride + l8; |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
1441 int x; |
111 | 1442 src+= stride*3; |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
1443 for(x=0; x<BLOCK_SIZE; x++) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1444 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1445 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
|
1446 if(ABS(middleEnergy) < 8*QP) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1447 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1448 const int q=(src[l4] - src[l5])/2; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1449 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
|
1450 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
|
1451 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1452 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1453 d= MAX(d, 0); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1454 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1455 d= (5*d + 32) >> 6; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1456 d*= SIGN(-middleEnergy); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1457 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1458 if(q>0) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1459 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1460 d= d<0 ? 0 : d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1461 d= d>q ? q : d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1462 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1463 else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1464 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1465 d= d>0 ? 0 : d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1466 d= d<q ? q : d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1467 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1468 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1469 src[l4]-= d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1470 src[l5]+= d; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1471 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1472 src++; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1473 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1474 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1475 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1476 |
165 | 1477 /** |
1478 * Check if the given 8x8 Block is mostly "flat" | |
1479 */ | |
1480 static inline int isHorizDC(uint8_t src[], int stride) | |
1481 { | |
1482 int numEq= 0; | |
1483 int y; | |
1484 for(y=0; y<BLOCK_SIZE; y++) | |
1485 { | |
1486 if(((src[0] - src[1] + 1) & 0xFFFF) < 3) numEq++; | |
1487 if(((src[1] - src[2] + 1) & 0xFFFF) < 3) numEq++; | |
1488 if(((src[2] - src[3] + 1) & 0xFFFF) < 3) numEq++; | |
1489 if(((src[3] - src[4] + 1) & 0xFFFF) < 3) numEq++; | |
1490 if(((src[4] - src[5] + 1) & 0xFFFF) < 3) numEq++; | |
1491 if(((src[5] - src[6] + 1) & 0xFFFF) < 3) numEq++; | |
1492 if(((src[6] - src[7] + 1) & 0xFFFF) < 3) numEq++; | |
1493 src+= stride; | |
1494 } | |
1495 return numEq > hFlatnessThreshold; | |
1496 } | |
1497 | |
1498 static inline int isHorizMinMaxOk(uint8_t src[], int stride, int QP) | |
1499 { | |
1500 if(abs(src[0] - src[7]) > 2*QP) return 0; | |
1501 | |
1502 return 1; | |
1503 } | |
1504 | |
1505 static inline void doHorizDefFilter(uint8_t dst[], int stride, int QP) | |
1506 { | |
1507 int y; | |
1508 for(y=0; y<BLOCK_SIZE; y++) | |
1509 { | |
1510 const int middleEnergy= 5*(dst[4] - dst[5]) + 2*(dst[2] - dst[5]); | |
1511 | |
1512 if(ABS(middleEnergy) < 8*QP) | |
1513 { | |
1514 const int q=(dst[3] - dst[4])/2; | |
1515 const int leftEnergy= 5*(dst[2] - dst[1]) + 2*(dst[0] - dst[3]); | |
1516 const int rightEnergy= 5*(dst[6] - dst[5]) + 2*(dst[4] - dst[7]); | |
1517 | |
1518 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) ); | |
1519 d= MAX(d, 0); | |
1520 | |
1521 d= (5*d + 32) >> 6; | |
1522 d*= SIGN(-middleEnergy); | |
1523 | |
1524 if(q>0) | |
1525 { | |
1526 d= d<0 ? 0 : d; | |
1527 d= d>q ? q : d; | |
1528 } | |
1529 else | |
1530 { | |
1531 d= d>0 ? 0 : d; | |
1532 d= d<q ? q : d; | |
1533 } | |
1534 | |
1535 dst[3]-= d; | |
1536 dst[4]+= d; | |
1537 } | |
1538 dst+= stride; | |
1539 } | |
1540 } | |
1541 | |
1542 /** | |
1543 * Do a horizontal low pass filter on the 10x8 block (dst points to middle 8x8 Block) | |
1544 * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 (C version) | |
1545 */ | |
1546 static inline void doHorizLowPass(uint8_t dst[], int stride, int QP) | |
1547 { | |
1548 | |
1549 int y; | |
1550 for(y=0; y<BLOCK_SIZE; y++) | |
1551 { | |
1552 const int first= ABS(dst[-1] - dst[0]) < QP ? dst[-1] : dst[0]; | |
1553 const int last= ABS(dst[8] - dst[7]) < QP ? dst[8] : dst[7]; | |
1554 | |
1555 int sums[9]; | |
1556 sums[0] = first + dst[0]; | |
1557 sums[1] = dst[0] + dst[1]; | |
1558 sums[2] = dst[1] + dst[2]; | |
1559 sums[3] = dst[2] + dst[3]; | |
1560 sums[4] = dst[3] + dst[4]; | |
1561 sums[5] = dst[4] + dst[5]; | |
1562 sums[6] = dst[5] + dst[6]; | |
1563 sums[7] = dst[6] + dst[7]; | |
1564 sums[8] = dst[7] + last; | |
1565 | |
1566 dst[0]= ((sums[0]<<2) + ((first + sums[2])<<1) + sums[4] + 8)>>4; | |
1567 dst[1]= ((dst[1]<<2) + ((first + sums[0] + sums[3])<<1) + sums[5] + 8)>>4; | |
1568 dst[2]= ((dst[2]<<2) + ((first + sums[1] + sums[4])<<1) + sums[6] + 8)>>4; | |
1569 dst[3]= ((dst[3]<<2) + ((sums[2] + sums[5])<<1) + sums[0] + sums[7] + 8)>>4; | |
1570 dst[4]= ((dst[4]<<2) + ((sums[3] + sums[6])<<1) + sums[1] + sums[8] + 8)>>4; | |
1571 dst[5]= ((dst[5]<<2) + ((last + sums[7] + sums[4])<<1) + sums[2] + 8)>>4; | |
1572 dst[6]= (((last + dst[6])<<2) + ((dst[7] + sums[5])<<1) + sums[3] + 8)>>4; | |
1573 dst[7]= ((sums[8]<<2) + ((last + sums[6])<<1) + sums[4] + 8)>>4; | |
1574 | |
1575 dst+= stride; | |
1576 } | |
1577 } | |
1578 | |
1579 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1580 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
|
1581 { |
132 | 1582 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1583 asm volatile( |
130 | 1584 "movq pQPb, %%mm0 \n\t" |
1585 "paddusb %%mm0, %%mm0 \n\t" | |
1586 "movq %%mm0, pQPb2 \n\t" | |
1587 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1588 "leal (%0, %1), %%eax \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1589 "leal (%%eax, %1, 4), %%ebx \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1590 // 0 1 2 3 4 5 6 7 8 9 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1591 // %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
|
1592 |
167 | 1593 "pcmpeqb %%mm7, %%mm7 \n\t" |
1594 "pxor %%mm6, %%mm6 \n\t" | |
132 | 1595 #ifdef HAVE_MMX2 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1596 #define FIND_MIN_MAX(addr)\ |
130 | 1597 "movq " #addr ", %%mm0 \n\t"\ |
167 | 1598 "pminub %%mm0, %%mm7 \n\t"\ |
1599 "pmaxub %%mm0, %%mm6 \n\t" | |
132 | 1600 #else |
1601 #define FIND_MIN_MAX(addr)\ | |
1602 "movq " #addr ", %%mm0 \n\t"\ | |
167 | 1603 "movq %%mm7, %%mm1 \n\t"\ |
1604 "psubusb %%mm0, %%mm6 \n\t"\ | |
1605 "paddb %%mm0, %%mm6 \n\t"\ | |
132 | 1606 "psubusb %%mm0, %%mm1 \n\t"\ |
167 | 1607 "psubb %%mm1, %%mm7 \n\t" |
132 | 1608 #endif |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1609 |
130 | 1610 FIND_MIN_MAX((%%eax)) |
1611 FIND_MIN_MAX((%%eax, %1)) | |
1612 FIND_MIN_MAX((%%eax, %1, 2)) | |
1613 FIND_MIN_MAX((%0, %1, 4)) | |
1614 FIND_MIN_MAX((%%ebx)) | |
1615 FIND_MIN_MAX((%%ebx, %1)) | |
1616 FIND_MIN_MAX((%%ebx, %1, 2)) | |
1617 FIND_MIN_MAX((%0, %1, 8)) | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1618 |
167 | 1619 "movq %%mm7, %%mm4 \n\t" |
1620 "psrlq $8, %%mm7 \n\t" | |
1621 #ifdef HAVE_MMX2 | |
1622 "pminub %%mm4, %%mm7 \n\t" // min of pixels | |
1623 "pshufw $0xF9, %%mm7, %%mm4 \n\t" | |
1624 "pminub %%mm4, %%mm7 \n\t" // min of pixels | |
1625 "pshufw $0xFE, %%mm7, %%mm4 \n\t" | |
1626 "pminub %%mm4, %%mm7 \n\t" | |
1627 #else | |
1628 "movq %%mm7, %%mm1 \n\t" | |
1629 "psubusb %%mm4, %%mm1 \n\t" | |
1630 "psubb %%mm1, %%mm7 \n\t" | |
1631 "movq %%mm7, %%mm4 \n\t" | |
1632 "psrlq $16, %%mm7 \n\t" | |
1633 "movq %%mm7, %%mm1 \n\t" | |
1634 "psubusb %%mm4, %%mm1 \n\t" | |
1635 "psubb %%mm1, %%mm7 \n\t" | |
1636 "movq %%mm7, %%mm4 \n\t" | |
1637 "psrlq $32, %%mm7 \n\t" | |
1638 "movq %%mm7, %%mm1 \n\t" | |
1639 "psubusb %%mm4, %%mm1 \n\t" | |
1640 "psubb %%mm1, %%mm7 \n\t" | |
1641 #endif | |
1642 | |
1643 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1644 "movq %%mm6, %%mm4 \n\t" |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
1645 "psrlq $8, %%mm6 \n\t" |
132 | 1646 #ifdef HAVE_MMX2 |
167 | 1647 "pmaxub %%mm4, %%mm6 \n\t" // max of pixels |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
1648 "pshufw $0xF9, %%mm6, %%mm4 \n\t" |
167 | 1649 "pmaxub %%mm4, %%mm6 \n\t" |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
1650 "pshufw $0xFE, %%mm6, %%mm4 \n\t" |
167 | 1651 "pmaxub %%mm4, %%mm6 \n\t" |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
1652 #else |
167 | 1653 "psubusb %%mm4, %%mm6 \n\t" |
1654 "paddb %%mm4, %%mm6 \n\t" | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1655 "movq %%mm6, %%mm4 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1656 "psrlq $16, %%mm6 \n\t" |
167 | 1657 "psubusb %%mm4, %%mm6 \n\t" |
1658 "paddb %%mm4, %%mm6 \n\t" | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1659 "movq %%mm6, %%mm4 \n\t" |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
1660 "psrlq $32, %%mm6 \n\t" |
167 | 1661 "psubusb %%mm4, %%mm6 \n\t" |
1662 "paddb %%mm4, %%mm6 \n\t" | |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
1663 #endif |
167 | 1664 "movq %%mm6, %%mm0 \n\t" // max |
1665 "psubb %%mm7, %%mm6 \n\t" // max - min | |
1666 "movd %%mm6, %%ecx \n\t" | |
1667 "cmpb deringThreshold, %%cl \n\t" | |
1668 " jb 1f \n\t" | |
1669 PAVGB(%%mm0, %%mm7) // a=(max + min)/2 | |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
1670 "punpcklbw %%mm7, %%mm7 \n\t" |
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
1671 "punpcklbw %%mm7, %%mm7 \n\t" |
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
1672 "punpcklbw %%mm7, %%mm7 \n\t" |
130 | 1673 "movq %%mm7, temp0 \n\t" |
1674 | |
1675 "movq (%0), %%mm0 \n\t" // L10 | |
1676 "movq %%mm0, %%mm1 \n\t" // L10 | |
1677 "movq %%mm0, %%mm2 \n\t" // L10 | |
1678 "psllq $8, %%mm1 \n\t" | |
1679 "psrlq $8, %%mm2 \n\t" | |
1680 "movd -4(%0), %%mm3 \n\t" | |
1681 "movd 8(%0), %%mm4 \n\t" | |
1682 "psrlq $24, %%mm3 \n\t" | |
1683 "psllq $56, %%mm4 \n\t" | |
1684 "por %%mm3, %%mm1 \n\t" // L00 | |
1685 "por %%mm4, %%mm2 \n\t" // L20 | |
1686 "movq %%mm1, %%mm3 \n\t" // L00 | |
1687 PAVGB(%%mm2, %%mm1) // (L20 + L00)/2 | |
1688 PAVGB(%%mm0, %%mm1) // (L20 + L00 + 2L10)/4 | |
1689 "psubusb %%mm7, %%mm0 \n\t" | |
1690 "psubusb %%mm7, %%mm2 \n\t" | |
1691 "psubusb %%mm7, %%mm3 \n\t" | |
1692 "pcmpeqb b00, %%mm0 \n\t" // L10 > a ? 0 : -1 | |
1693 "pcmpeqb b00, %%mm2 \n\t" // L20 > a ? 0 : -1 | |
1694 "pcmpeqb b00, %%mm3 \n\t" // L00 > a ? 0 : -1 | |
1695 "paddb %%mm2, %%mm0 \n\t" | |
1696 "paddb %%mm3, %%mm0 \n\t" | |
1697 | |
1698 "movq (%%eax), %%mm2 \n\t" // L11 | |
1699 "movq %%mm2, %%mm3 \n\t" // L11 | |
1700 "movq %%mm2, %%mm4 \n\t" // L11 | |
1701 "psllq $8, %%mm3 \n\t" | |
1702 "psrlq $8, %%mm4 \n\t" | |
1703 "movd -4(%%eax), %%mm5 \n\t" | |
1704 "movd 8(%%eax), %%mm6 \n\t" | |
1705 "psrlq $24, %%mm5 \n\t" | |
1706 "psllq $56, %%mm6 \n\t" | |
1707 "por %%mm5, %%mm3 \n\t" // L01 | |
1708 "por %%mm6, %%mm4 \n\t" // L21 | |
1709 "movq %%mm3, %%mm5 \n\t" // L01 | |
1710 PAVGB(%%mm4, %%mm3) // (L21 + L01)/2 | |
1711 PAVGB(%%mm2, %%mm3) // (L21 + L01 + 2L11)/4 | |
1712 "psubusb %%mm7, %%mm2 \n\t" | |
1713 "psubusb %%mm7, %%mm4 \n\t" | |
1714 "psubusb %%mm7, %%mm5 \n\t" | |
1715 "pcmpeqb b00, %%mm2 \n\t" // L11 > a ? 0 : -1 | |
1716 "pcmpeqb b00, %%mm4 \n\t" // L21 > a ? 0 : -1 | |
1717 "pcmpeqb b00, %%mm5 \n\t" // L01 > a ? 0 : -1 | |
1718 "paddb %%mm4, %%mm2 \n\t" | |
1719 "paddb %%mm5, %%mm2 \n\t" | |
1720 // 0, 2, 3, 1 | |
1721 #define DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \ | |
1722 "movq " #src ", " #sx " \n\t" /* src[0] */\ | |
1723 "movq " #sx ", " #lx " \n\t" /* src[0] */\ | |
1724 "movq " #sx ", " #t0 " \n\t" /* src[0] */\ | |
1725 "psllq $8, " #lx " \n\t"\ | |
1726 "psrlq $8, " #t0 " \n\t"\ | |
1727 "movd -4" #src ", " #t1 " \n\t"\ | |
1728 "psrlq $24, " #t1 " \n\t"\ | |
1729 "por " #t1 ", " #lx " \n\t" /* src[-1] */\ | |
1730 "movd 8" #src ", " #t1 " \n\t"\ | |
1731 "psllq $56, " #t1 " \n\t"\ | |
1732 "por " #t1 ", " #t0 " \n\t" /* src[+1] */\ | |
1733 "movq " #lx ", " #t1 " \n\t" /* src[-1] */\ | |
1734 PAVGB(t0, lx) /* (src[-1] + src[+1])/2 */\ | |
1735 PAVGB(sx, lx) /* (src[-1] + 2src[0] + src[+1])/4 */\ | |
135 | 1736 PAVGB(lx, pplx) \ |
1737 "movq " #lx ", temp1 \n\t"\ | |
1738 "movq temp0, " #lx " \n\t"\ | |
140 | 1739 "psubusb " #lx ", " #t1 " \n\t"\ |
1740 "psubusb " #lx ", " #t0 " \n\t"\ | |
1741 "psubusb " #lx ", " #sx " \n\t"\ | |
135 | 1742 "movq b00, " #lx " \n\t"\ |
140 | 1743 "pcmpeqb " #lx ", " #t1 " \n\t" /* src[-1] > a ? 0 : -1*/\ |
1744 "pcmpeqb " #lx ", " #t0 " \n\t" /* src[+1] > a ? 0 : -1*/\ | |
1745 "pcmpeqb " #lx ", " #sx " \n\t" /* src[0] > a ? 0 : -1*/\ | |
130 | 1746 "paddb " #t1 ", " #t0 " \n\t"\ |
1747 "paddb " #t0 ", " #sx " \n\t"\ | |
1748 \ | |
1749 PAVGB(plx, pplx) /* filtered */\ | |
1750 "movq " #dst ", " #t0 " \n\t" /* dst */\ | |
134 | 1751 "movq " #t0 ", " #t1 " \n\t" /* dst */\ |
1752 "psubusb pQPb2, " #t0 " \n\t"\ | |
1753 "paddusb pQPb2, " #t1 " \n\t"\ | |
1754 PMAXUB(t0, pplx)\ | |
1755 PMINUB(t1, pplx, t0)\ | |
130 | 1756 "paddb " #sx ", " #ppsx " \n\t"\ |
1757 "paddb " #psx ", " #ppsx " \n\t"\ | |
1758 "#paddb b02, " #ppsx " \n\t"\ | |
1759 "pand b08, " #ppsx " \n\t"\ | |
140 | 1760 "pcmpeqb " #lx ", " #ppsx " \n\t"\ |
134 | 1761 "pand " #ppsx ", " #pplx " \n\t"\ |
130 | 1762 "pandn " #dst ", " #ppsx " \n\t"\ |
140 | 1763 "por " #pplx ", " #ppsx " \n\t"\ |
135 | 1764 "movq " #ppsx ", " #dst " \n\t"\ |
1765 "movq temp1, " #lx " \n\t" | |
134 | 1766 |
130 | 1767 /* |
1768 0000000 | |
1769 1111111 | |
1770 | |
1771 1111110 | |
1772 1111101 | |
1773 1111100 | |
1774 1111011 | |
1775 1111010 | |
1776 1111001 | |
1777 | |
1778 1111000 | |
1779 1110111 | |
1780 | |
1781 */ | |
1782 //DERING_CORE(dst,src ,ppsx ,psx ,sx ,pplx ,plx ,lx ,t0 ,t1) | |
1783 DERING_CORE((%%eax),(%%eax, %1) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7) | |
1784 DERING_CORE((%%eax, %1),(%%eax, %1, 2) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7) | |
1785 DERING_CORE((%%eax, %1, 2),(%0, %1, 4) ,%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7) | |
1786 DERING_CORE((%0, %1, 4),(%%ebx) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7) | |
1787 DERING_CORE((%%ebx),(%%ebx, %1) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7) | |
1788 DERING_CORE((%%ebx, %1), (%%ebx, %1, 2),%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7) | |
1789 DERING_CORE((%%ebx, %1, 2),(%0, %1, 8) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7) | |
1790 DERING_CORE((%0, %1, 8),(%%ebx, %1, 4) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7) | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1791 |
167 | 1792 "1: \n\t" |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1793 : : "r" (src), "r" (stride), "r" (QP) |
167 | 1794 : "%eax", "%ebx", "%ecx" |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1795 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1796 #else |
134 | 1797 int y; |
1798 int min=255; | |
1799 int max=0; | |
1800 int avg; | |
1801 uint8_t *p; | |
1802 int s[10]; | |
1803 | |
1804 for(y=1; y<9; y++) | |
1805 { | |
1806 int x; | |
1807 p= src + stride*y; | |
1808 for(x=1; x<9; x++) | |
1809 { | |
1810 p++; | |
1811 if(*p > max) max= *p; | |
1812 if(*p < min) min= *p; | |
1813 } | |
1814 } | |
1815 avg= (min + max + 1)/2; | |
1816 | |
167 | 1817 if(max - min <deringThreshold) return; |
1818 | |
134 | 1819 for(y=0; y<10; y++) |
1820 { | |
1821 int x; | |
1822 int t = 0; | |
1823 p= src + stride*y; | |
1824 for(x=0; x<10; x++) | |
1825 { | |
1826 if(*p > avg) t |= (1<<x); | |
1827 p++; | |
1828 } | |
1829 t |= (~t)<<16; | |
1830 t &= (t<<1) & (t>>1); | |
1831 s[y] = t; | |
1832 } | |
1833 | |
1834 for(y=1; y<9; y++) | |
1835 { | |
1836 int x; | |
1837 int t = s[y-1] & s[y] & s[y+1]; | |
1838 t|= t>>16; | |
1839 | |
1840 p= src + stride*y; | |
1841 for(x=1; x<9; x++) | |
1842 { | |
1843 p++; | |
1844 if(t & (1<<x)) | |
1845 { | |
1846 int f= (*(p-stride-1)) + 2*(*(p-stride)) + (*(p-stride+1)) | |
1847 +2*(*(p -1)) + 4*(*p ) + 2*(*(p +1)) | |
1848 +(*(p+stride-1)) + 2*(*(p+stride)) + (*(p+stride+1)); | |
1849 f= (f + 8)>>4; | |
1850 | |
167 | 1851 #ifdef DEBUG_DERING_THRESHOLD |
1852 asm volatile("emms\n\t":); | |
1853 { | |
1854 static long long numPixels=0; | |
1855 if(x!=1 && x!=8 && y!=1 && y!=8) numPixels++; | |
1856 // if((max-min)<20 || (max-min)*QP<200) | |
1857 // if((max-min)*QP < 500) | |
1858 // if(max-min<QP/2) | |
1859 if(max-min < 20) | |
1860 { | |
1861 static int numSkiped=0; | |
1862 static int errorSum=0; | |
1863 static int worstQP=0; | |
1864 static int worstRange=0; | |
1865 static int worstDiff=0; | |
1866 int diff= (f - *p); | |
1867 int absDiff= ABS(diff); | |
1868 int error= diff*diff; | |
1869 | |
1870 if(x==1 || x==8 || y==1 || y==8) continue; | |
1871 | |
1872 numSkiped++; | |
1873 if(absDiff > worstDiff) | |
1874 { | |
1875 worstDiff= absDiff; | |
1876 worstQP= QP; | |
1877 worstRange= max-min; | |
1878 } | |
1879 errorSum+= error; | |
1880 | |
1881 if(1024LL*1024LL*1024LL % numSkiped == 0) | |
1882 { | |
1883 printf( "sum:%1.3f, skip:%d, wQP:%d, " | |
1884 "wRange:%d, wDiff:%d, relSkip:%1.3f\n", | |
1885 (float)errorSum/numSkiped, numSkiped, worstQP, worstRange, | |
1886 worstDiff, (float)numSkiped/numPixels); | |
1887 } | |
1888 } | |
1889 } | |
1890 #endif | |
134 | 1891 if (*p + 2*QP < f) *p= *p + 2*QP; |
1892 else if(*p - 2*QP > f) *p= *p - 2*QP; | |
1893 else *p=f; | |
1894 } | |
1895 } | |
1896 } | |
167 | 1897 #ifdef DEBUG_DERING_THRESHOLD |
1898 if(max-min < 20) | |
1899 { | |
1900 for(y=1; y<9; y++) | |
1901 { | |
1902 int x; | |
1903 int t = 0; | |
1904 p= src + stride*y; | |
1905 for(x=1; x<9; x++) | |
1906 { | |
1907 p++; | |
1908 *p = MIN(*p + 20, 255); | |
1909 } | |
1910 } | |
1911 // src[0] = src[7]=src[stride*7]=src[stride*7 + 7]=255; | |
1912 } | |
1913 #endif | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1914 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1915 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1916 |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1917 /** |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1918 * Deinterlaces the given block |
142 | 1919 * will be called for every 8x8 block and can read & write from line 4-15 |
1920 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too | |
1921 * lines 4-12 will be read into the deblocking filter and should be deinterlaced | |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1922 */ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1923 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
|
1924 { |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1925 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
142 | 1926 src+= 4*stride; |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1927 asm volatile( |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1928 "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
|
1929 "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
|
1930 // 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
|
1931 // %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
|
1932 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1933 "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
|
1934 "movq (%%eax, %1), %%mm1 \n\t" |
111 | 1935 PAVGB(%%mm1, %%mm0) |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1936 "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
|
1937 "movq (%0, %1, 4), %%mm0 \n\t" |
111 | 1938 PAVGB(%%mm0, %%mm1) |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1939 "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
|
1940 "movq (%%ebx, %1), %%mm1 \n\t" |
111 | 1941 PAVGB(%%mm1, %%mm0) |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1942 "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
|
1943 "movq (%0, %1, 8), %%mm0 \n\t" |
111 | 1944 PAVGB(%%mm0, %%mm1) |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1945 "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
|
1946 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1947 : : "r" (src), "r" (stride) |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1948 : "%eax", "%ebx" |
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 #else |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1951 int x; |
142 | 1952 src+= 4*stride; |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1953 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
|
1954 { |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1955 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
|
1956 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
|
1957 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
|
1958 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
|
1959 src++; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1960 } |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1961 #endif |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1962 } |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1963 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1964 /** |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1965 * Deinterlaces the given block |
142 | 1966 * will be called for every 8x8 block and can read & write from line 4-15 |
1967 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too | |
1968 * lines 4-12 will be read into the deblocking filter and should be deinterlaced | |
1969 * this filter will read lines 3-15 and write 7-13 | |
111 | 1970 * 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
|
1971 */ |
111 | 1972 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
|
1973 { |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1974 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
142 | 1975 src+= stride*3; |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1976 asm volatile( |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
1977 "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
|
1978 "leal (%%eax, %1, 4), %%ebx \n\t" |
111 | 1979 "leal (%%ebx, %1, 4), %%ecx \n\t" |
1980 "addl %1, %%ecx \n\t" | |
1981 "pxor %%mm7, %%mm7 \n\t" | |
1982 // 0 1 2 3 4 5 6 7 8 9 10 | |
1983 // %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
|
1984 |
111 | 1985 #define DEINT_CUBIC(a,b,c,d,e)\ |
1986 "movq " #a ", %%mm0 \n\t"\ | |
1987 "movq " #b ", %%mm1 \n\t"\ | |
1988 "movq " #d ", %%mm2 \n\t"\ | |
1989 "movq " #e ", %%mm3 \n\t"\ | |
1990 PAVGB(%%mm2, %%mm1) /* (b+d) /2 */\ | |
1991 PAVGB(%%mm3, %%mm0) /* a(a+e) /2 */\ | |
1992 "movq %%mm0, %%mm2 \n\t"\ | |
1993 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
1994 "punpckhbw %%mm7, %%mm2 \n\t"\ | |
1995 "movq %%mm1, %%mm3 \n\t"\ | |
1996 "punpcklbw %%mm7, %%mm1 \n\t"\ | |
1997 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
1998 "psubw %%mm1, %%mm0 \n\t" /* L(a+e - (b+d))/2 */\ | |
1999 "psubw %%mm3, %%mm2 \n\t" /* H(a+e - (b+d))/2 */\ | |
2000 "psraw $3, %%mm0 \n\t" /* L(a+e - (b+d))/16 */\ | |
2001 "psraw $3, %%mm2 \n\t" /* H(a+e - (b+d))/16 */\ | |
2002 "psubw %%mm0, %%mm1 \n\t" /* L(9b + 9d - a - e)/16 */\ | |
2003 "psubw %%mm2, %%mm3 \n\t" /* H(9b + 9d - a - e)/16 */\ | |
2004 "packuswb %%mm3, %%mm1 \n\t"\ | |
2005 "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
|
2006 |
111 | 2007 DEINT_CUBIC((%0), (%%eax, %1), (%%eax, %1, 2), (%0, %1, 4), (%%ebx, %1)) |
2008 DEINT_CUBIC((%%eax, %1), (%0, %1, 4), (%%ebx), (%%ebx, %1), (%0, %1, 8)) | |
2009 DEINT_CUBIC((%0, %1, 4), (%%ebx, %1), (%%ebx, %1, 2), (%0, %1, 8), (%%ecx)) | |
2010 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
|
2011 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2012 : : "r" (src), "r" (stride) |
111 | 2013 : "%eax", "%ebx", "ecx" |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2014 ); |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2015 #else |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2016 int x; |
142 | 2017 src+= stride*3; |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2018 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
|
2019 { |
111 | 2020 src[stride*3] = (-src[0] + 9*src[stride*2] + 9*src[stride*4] - src[stride*6])>>4; |
2021 src[stride*5] = (-src[stride*2] + 9*src[stride*4] + 9*src[stride*6] - src[stride*8])>>4; | |
2022 src[stride*7] = (-src[stride*4] + 9*src[stride*6] + 9*src[stride*8] - src[stride*10])>>4; | |
2023 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
|
2024 src++; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2025 } |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2026 #endif |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2027 } |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2028 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2029 /** |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2030 * Deinterlaces the given block |
142 | 2031 * will be called for every 8x8 block and can read & write from line 4-15 |
2032 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too | |
2033 * lines 4-12 will be read into the deblocking filter and should be deinterlaced | |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2034 * will shift the image up by 1 line (FIXME if this is a problem) |
142 | 2035 * this filter will read lines 4-13 and write 4-11 |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2036 */ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2037 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
|
2038 { |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2039 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
142 | 2040 src+= 4*stride; |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2041 asm volatile( |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2042 "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
|
2043 "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
|
2044 // 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
|
2045 // %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
|
2046 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2047 "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
|
2048 "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
|
2049 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
|
2050 "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
|
2051 PAVGB(%%mm2, %%mm0) |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2052 "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
|
2053 "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
|
2054 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
|
2055 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
|
2056 "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
|
2057 "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
|
2058 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
|
2059 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
|
2060 "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
|
2061 "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
|
2062 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
|
2063 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
|
2064 "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
|
2065 "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
|
2066 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
|
2067 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
|
2068 "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
|
2069 "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
|
2070 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
|
2071 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
|
2072 "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
|
2073 "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
|
2074 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
|
2075 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
|
2076 "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
|
2077 "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
|
2078 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
|
2079 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
|
2080 "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
|
2081 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2082 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2083 : : "r" (src), "r" (stride) |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2084 : "%eax", "%ebx" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2085 ); |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2086 #else |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2087 int x; |
142 | 2088 src+= 4*stride; |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2089 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
|
2090 { |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2091 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
|
2092 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
|
2093 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
|
2094 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
|
2095 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
|
2096 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
|
2097 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
|
2098 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
|
2099 src++; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2100 } |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2101 #endif |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2102 } |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2103 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2104 /** |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2105 * Deinterlaces the given block |
142 | 2106 * will be called for every 8x8 block and can read & write from line 4-15, |
2107 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too | |
2108 * lines 4-12 will be read into the deblocking filter and should be deinterlaced | |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2109 */ |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2110 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
|
2111 { |
107 | 2112 #ifdef HAVE_MMX |
142 | 2113 src+= 4*stride; |
107 | 2114 #ifdef HAVE_MMX2 |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2115 asm volatile( |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2116 "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
|
2117 "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
|
2118 // 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
|
2119 // %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
|
2120 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2121 "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
|
2122 "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
|
2123 "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
|
2124 "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
|
2125 "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
|
2126 "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
|
2127 "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
|
2128 "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
|
2129 "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
|
2130 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2131 "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
|
2132 "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
|
2133 "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
|
2134 "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
|
2135 "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
|
2136 "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
|
2137 "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
|
2138 "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
|
2139 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2140 "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
|
2141 "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
|
2142 "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
|
2143 "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
|
2144 "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
|
2145 "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
|
2146 "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
|
2147 "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
|
2148 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2149 "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
|
2150 "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
|
2151 "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
|
2152 "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
|
2153 "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
|
2154 "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
|
2155 "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
|
2156 "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
|
2157 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2158 |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2159 : : "r" (src), "r" (stride) |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2160 : "%eax", "%ebx" |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2161 ); |
107 | 2162 |
2163 #else // MMX without MMX2 | |
2164 asm volatile( | |
2165 "leal (%0, %1), %%eax \n\t" | |
2166 "leal (%%eax, %1, 4), %%ebx \n\t" | |
2167 // 0 1 2 3 4 5 6 7 8 9 | |
2168 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 | |
2169 "pxor %%mm7, %%mm7 \n\t" | |
2170 | |
2171 #define MEDIAN(a,b,c)\ | |
2172 "movq " #a ", %%mm0 \n\t"\ | |
2173 "movq " #b ", %%mm2 \n\t"\ | |
2174 "movq " #c ", %%mm1 \n\t"\ | |
2175 "movq %%mm0, %%mm3 \n\t"\ | |
2176 "movq %%mm1, %%mm4 \n\t"\ | |
2177 "movq %%mm2, %%mm5 \n\t"\ | |
2178 "psubusb %%mm1, %%mm3 \n\t"\ | |
2179 "psubusb %%mm2, %%mm4 \n\t"\ | |
2180 "psubusb %%mm0, %%mm5 \n\t"\ | |
2181 "pcmpeqb %%mm7, %%mm3 \n\t"\ | |
2182 "pcmpeqb %%mm7, %%mm4 \n\t"\ | |
2183 "pcmpeqb %%mm7, %%mm5 \n\t"\ | |
2184 "movq %%mm3, %%mm6 \n\t"\ | |
2185 "pxor %%mm4, %%mm3 \n\t"\ | |
2186 "pxor %%mm5, %%mm4 \n\t"\ | |
2187 "pxor %%mm6, %%mm5 \n\t"\ | |
2188 "por %%mm3, %%mm1 \n\t"\ | |
2189 "por %%mm4, %%mm2 \n\t"\ | |
2190 "por %%mm5, %%mm0 \n\t"\ | |
2191 "pand %%mm2, %%mm0 \n\t"\ | |
2192 "pand %%mm1, %%mm0 \n\t"\ | |
2193 "movq %%mm0, " #b " \n\t" | |
2194 | |
2195 MEDIAN((%0), (%%eax), (%%eax, %1)) | |
2196 MEDIAN((%%eax, %1), (%%eax, %1, 2), (%0, %1, 4)) | |
2197 MEDIAN((%0, %1, 4), (%%ebx), (%%ebx, %1)) | |
2198 MEDIAN((%%ebx, %1), (%%ebx, %1, 2), (%0, %1, 8)) | |
2199 | |
2200 : : "r" (src), "r" (stride) | |
2201 : "%eax", "%ebx" | |
2202 ); | |
2203 #endif // MMX | |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2204 #else |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2205 //FIXME |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2206 int x; |
142 | 2207 src+= 4*stride; |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2208 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
|
2209 { |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2210 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
|
2211 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
|
2212 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
|
2213 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
|
2214 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
|
2215 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
|
2216 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
|
2217 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
|
2218 src++; |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2219 } |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2220 #endif |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2221 } |
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
2222 |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
2223 #ifdef HAVE_MMX |
128
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2224 /** |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2225 * transposes and shift the given 8x8 Block into dst1 and dst2 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2226 */ |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2227 static inline void transpose1(uint8_t *dst1, uint8_t *dst2, uint8_t *src, int srcStride) |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2228 { |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2229 asm( |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2230 "leal (%0, %1), %%eax \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2231 "leal (%%eax, %1, 4), %%ebx \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2232 // 0 1 2 3 4 5 6 7 8 9 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2233 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2234 "movq (%0), %%mm0 \n\t" // 12345678 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2235 "movq (%%eax), %%mm1 \n\t" // abcdefgh |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2236 "movq %%mm0, %%mm2 \n\t" // 12345678 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2237 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2238 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2239 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2240 "movq (%%eax, %1), %%mm1 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2241 "movq (%%eax, %1, 2), %%mm3 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2242 "movq %%mm1, %%mm4 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2243 "punpcklbw %%mm3, %%mm1 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2244 "punpckhbw %%mm3, %%mm4 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2245 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2246 "movq %%mm0, %%mm3 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2247 "punpcklwd %%mm1, %%mm0 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2248 "punpckhwd %%mm1, %%mm3 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2249 "movq %%mm2, %%mm1 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2250 "punpcklwd %%mm4, %%mm2 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2251 "punpckhwd %%mm4, %%mm1 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2252 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2253 "movd %%mm0, 128(%2) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2254 "psrlq $32, %%mm0 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2255 "movd %%mm0, 144(%2) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2256 "movd %%mm3, 160(%2) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2257 "psrlq $32, %%mm3 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2258 "movd %%mm3, 176(%2) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2259 "movd %%mm3, 48(%3) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2260 "movd %%mm2, 192(%2) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2261 "movd %%mm2, 64(%3) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2262 "psrlq $32, %%mm2 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2263 "movd %%mm2, 80(%3) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2264 "movd %%mm1, 96(%3) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2265 "psrlq $32, %%mm1 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2266 "movd %%mm1, 112(%3) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2267 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2268 "movq (%0, %1, 4), %%mm0 \n\t" // 12345678 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2269 "movq (%%ebx), %%mm1 \n\t" // abcdefgh |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2270 "movq %%mm0, %%mm2 \n\t" // 12345678 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2271 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2272 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2273 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2274 "movq (%%ebx, %1), %%mm1 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2275 "movq (%%ebx, %1, 2), %%mm3 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2276 "movq %%mm1, %%mm4 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2277 "punpcklbw %%mm3, %%mm1 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2278 "punpckhbw %%mm3, %%mm4 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2279 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2280 "movq %%mm0, %%mm3 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2281 "punpcklwd %%mm1, %%mm0 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2282 "punpckhwd %%mm1, %%mm3 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2283 "movq %%mm2, %%mm1 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2284 "punpcklwd %%mm4, %%mm2 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2285 "punpckhwd %%mm4, %%mm1 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2286 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2287 "movd %%mm0, 132(%2) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2288 "psrlq $32, %%mm0 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2289 "movd %%mm0, 148(%2) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2290 "movd %%mm3, 164(%2) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2291 "psrlq $32, %%mm3 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2292 "movd %%mm3, 180(%2) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2293 "movd %%mm3, 52(%3) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2294 "movd %%mm2, 196(%2) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2295 "movd %%mm2, 68(%3) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2296 "psrlq $32, %%mm2 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2297 "movd %%mm2, 84(%3) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2298 "movd %%mm1, 100(%3) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2299 "psrlq $32, %%mm1 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2300 "movd %%mm1, 116(%3) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2301 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2302 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2303 :: "r" (src), "r" (srcStride), "r" (dst1), "r" (dst2) |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2304 : "%eax", "%ebx" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2305 ); |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2306 } |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2307 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2308 /** |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2309 * transposes the given 8x8 block |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2310 */ |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2311 static inline void transpose2(uint8_t *dst, int dstStride, uint8_t *src) |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2312 { |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2313 asm( |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2314 "leal (%0, %1), %%eax \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2315 "leal (%%eax, %1, 4), %%ebx \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2316 // 0 1 2 3 4 5 6 7 8 9 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2317 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2318 "movq (%2), %%mm0 \n\t" // 12345678 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2319 "movq 16(%2), %%mm1 \n\t" // abcdefgh |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2320 "movq %%mm0, %%mm2 \n\t" // 12345678 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2321 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2322 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2323 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2324 "movq 32(%2), %%mm1 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2325 "movq 48(%2), %%mm3 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2326 "movq %%mm1, %%mm4 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2327 "punpcklbw %%mm3, %%mm1 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2328 "punpckhbw %%mm3, %%mm4 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2329 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2330 "movq %%mm0, %%mm3 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2331 "punpcklwd %%mm1, %%mm0 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2332 "punpckhwd %%mm1, %%mm3 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2333 "movq %%mm2, %%mm1 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2334 "punpcklwd %%mm4, %%mm2 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2335 "punpckhwd %%mm4, %%mm1 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2336 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2337 "movd %%mm0, (%0) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2338 "psrlq $32, %%mm0 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2339 "movd %%mm0, (%%eax) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2340 "movd %%mm3, (%%eax, %1) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2341 "psrlq $32, %%mm3 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2342 "movd %%mm3, (%%eax, %1, 2) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2343 "movd %%mm2, (%0, %1, 4) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2344 "psrlq $32, %%mm2 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2345 "movd %%mm2, (%%ebx) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2346 "movd %%mm1, (%%ebx, %1) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2347 "psrlq $32, %%mm1 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2348 "movd %%mm1, (%%ebx, %1, 2) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2349 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2350 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2351 "movq 64(%2), %%mm0 \n\t" // 12345678 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2352 "movq 80(%2), %%mm1 \n\t" // abcdefgh |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2353 "movq %%mm0, %%mm2 \n\t" // 12345678 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2354 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2355 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2356 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2357 "movq 96(%2), %%mm1 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2358 "movq 112(%2), %%mm3 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2359 "movq %%mm1, %%mm4 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2360 "punpcklbw %%mm3, %%mm1 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2361 "punpckhbw %%mm3, %%mm4 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2362 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2363 "movq %%mm0, %%mm3 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2364 "punpcklwd %%mm1, %%mm0 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2365 "punpckhwd %%mm1, %%mm3 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2366 "movq %%mm2, %%mm1 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2367 "punpcklwd %%mm4, %%mm2 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2368 "punpckhwd %%mm4, %%mm1 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2369 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2370 "movd %%mm0, 4(%0) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2371 "psrlq $32, %%mm0 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2372 "movd %%mm0, 4(%%eax) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2373 "movd %%mm3, 4(%%eax, %1) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2374 "psrlq $32, %%mm3 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2375 "movd %%mm3, 4(%%eax, %1, 2) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2376 "movd %%mm2, 4(%0, %1, 4) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2377 "psrlq $32, %%mm2 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2378 "movd %%mm2, 4(%%ebx) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2379 "movd %%mm1, 4(%%ebx, %1) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2380 "psrlq $32, %%mm1 \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2381 "movd %%mm1, 4(%%ebx, %1, 2) \n\t" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2382 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2383 :: "r" (dst), "r" (dstStride), "r" (src) |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2384 : "%eax", "%ebx" |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2385 ); |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2386 } |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
2387 #endif |
157
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2388 //static int test=0; |
128
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
2389 |
156 | 2390 static void inline tempNoiseReducer(uint8_t *src, int stride, |
158 | 2391 uint8_t *tempBlured, uint32_t *tempBluredPast, int *maxNoise) |
156 | 2392 { |
157
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2393 #define FAST_L2_DIFF |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2394 //#define L1_DIFF //u should change the thresholds too if u try that one |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2395 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2396 asm volatile( |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2397 "leal (%2, %2, 2), %%eax \n\t" // 3*stride |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2398 "leal (%2, %2, 4), %%ebx \n\t" // 5*stride |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2399 "leal (%%ebx, %2, 2), %%ecx \n\t" // 7*stride |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2400 // 0 1 2 3 4 5 6 7 8 9 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2401 // %x %x+%2 %x+2%2 %x+eax %x+4%2 %x+ebx %x+2eax %x+ecx %x+8%2 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2402 //FIXME reorder? |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2403 #ifdef L1_DIFF //needs mmx2 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2404 "movq (%0), %%mm0 \n\t" // L0 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2405 "psadbw (%1), %%mm0 \n\t" // |L0-R0| |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2406 "movq (%0, %2), %%mm1 \n\t" // L1 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2407 "psadbw (%1, %2), %%mm1 \n\t" // |L1-R1| |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2408 "movq (%0, %2, 2), %%mm2 \n\t" // L2 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2409 "psadbw (%1, %2, 2), %%mm2 \n\t" // |L2-R2| |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2410 "movq (%0, %%eax), %%mm3 \n\t" // L3 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2411 "psadbw (%1, %%eax), %%mm3 \n\t" // |L3-R3| |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2412 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2413 "movq (%0, %2, 4), %%mm4 \n\t" // L4 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2414 "paddw %%mm1, %%mm0 \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2415 "psadbw (%1, %2, 4), %%mm4 \n\t" // |L4-R4| |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2416 "movq (%0, %%ebx), %%mm5 \n\t" // L5 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2417 "paddw %%mm2, %%mm0 \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2418 "psadbw (%1, %%ebx), %%mm5 \n\t" // |L5-R5| |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2419 "movq (%0, %%eax, 2), %%mm6 \n\t" // L6 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2420 "paddw %%mm3, %%mm0 \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2421 "psadbw (%1, %%eax, 2), %%mm6 \n\t" // |L6-R6| |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2422 "movq (%0, %%ecx), %%mm7 \n\t" // L7 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2423 "paddw %%mm4, %%mm0 \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2424 "psadbw (%1, %%ecx), %%mm7 \n\t" // |L7-R7| |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2425 "paddw %%mm5, %%mm6 \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2426 "paddw %%mm7, %%mm6 \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2427 "paddw %%mm6, %%mm0 \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2428 #elif defined (FAST_L2_DIFF) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2429 "pcmpeqb %%mm7, %%mm7 \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2430 "movq b80, %%mm6 \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2431 "pxor %%mm0, %%mm0 \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2432 #define L2_DIFF_CORE(a, b)\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2433 "movq " #a ", %%mm5 \n\t"\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2434 "movq " #b ", %%mm2 \n\t"\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2435 "pxor %%mm7, %%mm2 \n\t"\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2436 PAVGB(%%mm2, %%mm5)\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2437 "paddb %%mm6, %%mm5 \n\t"\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2438 "movq %%mm5, %%mm2 \n\t"\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2439 "psllw $8, %%mm5 \n\t"\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2440 "pmaddwd %%mm5, %%mm5 \n\t"\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2441 "pmaddwd %%mm2, %%mm2 \n\t"\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2442 "paddd %%mm2, %%mm5 \n\t"\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2443 "psrld $14, %%mm5 \n\t"\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2444 "paddd %%mm5, %%mm0 \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2445 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2446 L2_DIFF_CORE((%0), (%1)) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2447 L2_DIFF_CORE((%0, %2), (%1, %2)) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2448 L2_DIFF_CORE((%0, %2, 2), (%1, %2, 2)) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2449 L2_DIFF_CORE((%0, %%eax), (%1, %%eax)) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2450 L2_DIFF_CORE((%0, %2, 4), (%1, %2, 4)) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2451 L2_DIFF_CORE((%0, %%ebx), (%1, %%ebx)) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2452 L2_DIFF_CORE((%0, %%eax,2), (%1, %%eax,2)) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2453 L2_DIFF_CORE((%0, %%ecx), (%1, %%ecx)) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2454 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2455 #else |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2456 "pxor %%mm7, %%mm7 \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2457 "pxor %%mm0, %%mm0 \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2458 #define L2_DIFF_CORE(a, b)\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2459 "movq " #a ", %%mm5 \n\t"\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2460 "movq " #b ", %%mm2 \n\t"\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2461 "movq %%mm5, %%mm1 \n\t"\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2462 "movq %%mm2, %%mm3 \n\t"\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2463 "punpcklbw %%mm7, %%mm5 \n\t"\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2464 "punpckhbw %%mm7, %%mm1 \n\t"\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2465 "punpcklbw %%mm7, %%mm2 \n\t"\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2466 "punpckhbw %%mm7, %%mm3 \n\t"\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2467 "psubw %%mm2, %%mm5 \n\t"\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2468 "psubw %%mm3, %%mm1 \n\t"\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2469 "pmaddwd %%mm5, %%mm5 \n\t"\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2470 "pmaddwd %%mm1, %%mm1 \n\t"\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2471 "paddd %%mm1, %%mm5 \n\t"\ |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2472 "paddd %%mm5, %%mm0 \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2473 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2474 L2_DIFF_CORE((%0), (%1)) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2475 L2_DIFF_CORE((%0, %2), (%1, %2)) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2476 L2_DIFF_CORE((%0, %2, 2), (%1, %2, 2)) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2477 L2_DIFF_CORE((%0, %%eax), (%1, %%eax)) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2478 L2_DIFF_CORE((%0, %2, 4), (%1, %2, 4)) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2479 L2_DIFF_CORE((%0, %%ebx), (%1, %%ebx)) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2480 L2_DIFF_CORE((%0, %%eax,2), (%1, %%eax,2)) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2481 L2_DIFF_CORE((%0, %%ecx), (%1, %%ecx)) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2482 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2483 #endif |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2484 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2485 "movq %%mm0, %%mm4 \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2486 "psrlq $32, %%mm0 \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2487 "paddd %%mm0, %%mm4 \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2488 "movd %%mm4, %%ecx \n\t" |
158 | 2489 "shll $2, %%ecx \n\t" |
2490 "movl %3, %%ebx \n\t" | |
2491 "addl -4(%%ebx), %%ecx \n\t" | |
2492 "addl 4(%%ebx), %%ecx \n\t" | |
2493 "addl -1024(%%ebx), %%ecx \n\t" | |
2494 "addl $4, %%ecx \n\t" | |
2495 "addl 1024(%%ebx), %%ecx \n\t" | |
2496 "shrl $3, %%ecx \n\t" | |
2497 "movl %%ecx, (%%ebx) \n\t" | |
2498 "leal (%%eax, %2, 2), %%ebx \n\t" // 5*stride | |
2499 | |
157
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2500 // "movl %3, %%ecx \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2501 // "movl %%ecx, test \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2502 // "jmp 4f \n\t" |
158 | 2503 "cmpl 4+maxTmpNoise, %%ecx \n\t" |
157
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2504 " jb 2f \n\t" |
158 | 2505 "cmpl 8+maxTmpNoise, %%ecx \n\t" |
157
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2506 " jb 1f \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2507 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2508 "leal (%%ebx, %2, 2), %%ecx \n\t" // 7*stride |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2509 "movq (%0), %%mm0 \n\t" // L0 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2510 "movq (%0, %2), %%mm1 \n\t" // L1 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2511 "movq (%0, %2, 2), %%mm2 \n\t" // L2 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2512 "movq (%0, %%eax), %%mm3 \n\t" // L3 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2513 "movq (%0, %2, 4), %%mm4 \n\t" // L4 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2514 "movq (%0, %%ebx), %%mm5 \n\t" // L5 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2515 "movq (%0, %%eax, 2), %%mm6 \n\t" // L6 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2516 "movq (%0, %%ecx), %%mm7 \n\t" // L7 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2517 "movq %%mm0, (%1) \n\t" // L0 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2518 "movq %%mm1, (%1, %2) \n\t" // L1 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2519 "movq %%mm2, (%1, %2, 2) \n\t" // L2 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2520 "movq %%mm3, (%1, %%eax) \n\t" // L3 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2521 "movq %%mm4, (%1, %2, 4) \n\t" // L4 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2522 "movq %%mm5, (%1, %%ebx) \n\t" // L5 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2523 "movq %%mm6, (%1, %%eax, 2) \n\t" // L6 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2524 "movq %%mm7, (%1, %%ecx) \n\t" // L7 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2525 "jmp 4f \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2526 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2527 "1: \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2528 "leal (%%ebx, %2, 2), %%ecx \n\t" // 7*stride |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2529 "movq (%0), %%mm0 \n\t" // L0 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2530 "pavgb (%1), %%mm0 \n\t" // L0 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2531 "movq (%0, %2), %%mm1 \n\t" // L1 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2532 "pavgb (%1, %2), %%mm1 \n\t" // L1 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2533 "movq (%0, %2, 2), %%mm2 \n\t" // L2 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2534 "pavgb (%1, %2, 2), %%mm2 \n\t" // L2 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2535 "movq (%0, %%eax), %%mm3 \n\t" // L3 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2536 "pavgb (%1, %%eax), %%mm3 \n\t" // L3 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2537 "movq (%0, %2, 4), %%mm4 \n\t" // L4 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2538 "pavgb (%1, %2, 4), %%mm4 \n\t" // L4 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2539 "movq (%0, %%ebx), %%mm5 \n\t" // L5 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2540 "pavgb (%1, %%ebx), %%mm5 \n\t" // L5 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2541 "movq (%0, %%eax, 2), %%mm6 \n\t" // L6 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2542 "pavgb (%1, %%eax, 2), %%mm6 \n\t" // L6 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2543 "movq (%0, %%ecx), %%mm7 \n\t" // L7 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2544 "pavgb (%1, %%ecx), %%mm7 \n\t" // L7 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2545 "movq %%mm0, (%1) \n\t" // R0 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2546 "movq %%mm1, (%1, %2) \n\t" // R1 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2547 "movq %%mm2, (%1, %2, 2) \n\t" // R2 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2548 "movq %%mm3, (%1, %%eax) \n\t" // R3 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2549 "movq %%mm4, (%1, %2, 4) \n\t" // R4 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2550 "movq %%mm5, (%1, %%ebx) \n\t" // R5 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2551 "movq %%mm6, (%1, %%eax, 2) \n\t" // R6 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2552 "movq %%mm7, (%1, %%ecx) \n\t" // R7 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2553 "movq %%mm0, (%0) \n\t" // L0 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2554 "movq %%mm1, (%0, %2) \n\t" // L1 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2555 "movq %%mm2, (%0, %2, 2) \n\t" // L2 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2556 "movq %%mm3, (%0, %%eax) \n\t" // L3 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2557 "movq %%mm4, (%0, %2, 4) \n\t" // L4 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2558 "movq %%mm5, (%0, %%ebx) \n\t" // L5 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2559 "movq %%mm6, (%0, %%eax, 2) \n\t" // L6 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2560 "movq %%mm7, (%0, %%ecx) \n\t" // L7 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2561 "jmp 4f \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2562 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2563 "2: \n\t" |
158 | 2564 "cmpl maxTmpNoise, %%ecx \n\t" |
157
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2565 " jb 3f \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2566 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2567 "leal (%%ebx, %2, 2), %%ecx \n\t" // 7*stride |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2568 "movq (%0), %%mm0 \n\t" // L0 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2569 "movq (%0, %2), %%mm1 \n\t" // L1 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2570 "movq (%0, %2, 2), %%mm2 \n\t" // L2 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2571 "movq (%0, %%eax), %%mm3 \n\t" // L3 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2572 "movq (%1), %%mm4 \n\t" // R0 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2573 "movq (%1, %2), %%mm5 \n\t" // R1 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2574 "movq (%1, %2, 2), %%mm6 \n\t" // R2 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2575 "movq (%1, %%eax), %%mm7 \n\t" // R3 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2576 PAVGB(%%mm4, %%mm0) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2577 PAVGB(%%mm5, %%mm1) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2578 PAVGB(%%mm6, %%mm2) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2579 PAVGB(%%mm7, %%mm3) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2580 PAVGB(%%mm4, %%mm0) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2581 PAVGB(%%mm5, %%mm1) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2582 PAVGB(%%mm6, %%mm2) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2583 PAVGB(%%mm7, %%mm3) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2584 "movq %%mm0, (%1) \n\t" // R0 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2585 "movq %%mm1, (%1, %2) \n\t" // R1 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2586 "movq %%mm2, (%1, %2, 2) \n\t" // R2 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2587 "movq %%mm3, (%1, %%eax) \n\t" // R3 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2588 "movq %%mm0, (%0) \n\t" // L0 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2589 "movq %%mm1, (%0, %2) \n\t" // L1 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2590 "movq %%mm2, (%0, %2, 2) \n\t" // L2 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2591 "movq %%mm3, (%0, %%eax) \n\t" // L3 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2592 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2593 "movq (%0, %2, 4), %%mm0 \n\t" // L4 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2594 "movq (%0, %%ebx), %%mm1 \n\t" // L5 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2595 "movq (%0, %%eax, 2), %%mm2 \n\t" // L6 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2596 "movq (%0, %%ecx), %%mm3 \n\t" // L7 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2597 "movq (%1, %2, 4), %%mm4 \n\t" // R4 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2598 "movq (%1, %%ebx), %%mm5 \n\t" // R5 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2599 "movq (%1, %%eax, 2), %%mm6 \n\t" // R6 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2600 "movq (%1, %%ecx), %%mm7 \n\t" // R7 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2601 PAVGB(%%mm4, %%mm0) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2602 PAVGB(%%mm5, %%mm1) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2603 PAVGB(%%mm6, %%mm2) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2604 PAVGB(%%mm7, %%mm3) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2605 PAVGB(%%mm4, %%mm0) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2606 PAVGB(%%mm5, %%mm1) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2607 PAVGB(%%mm6, %%mm2) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2608 PAVGB(%%mm7, %%mm3) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2609 "movq %%mm0, (%1, %2, 4) \n\t" // R4 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2610 "movq %%mm1, (%1, %%ebx) \n\t" // R5 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2611 "movq %%mm2, (%1, %%eax, 2) \n\t" // R6 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2612 "movq %%mm3, (%1, %%ecx) \n\t" // R7 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2613 "movq %%mm0, (%0, %2, 4) \n\t" // L4 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2614 "movq %%mm1, (%0, %%ebx) \n\t" // L5 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2615 "movq %%mm2, (%0, %%eax, 2) \n\t" // L6 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2616 "movq %%mm3, (%0, %%ecx) \n\t" // L7 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2617 "jmp 4f \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2618 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2619 "3: \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2620 "leal (%%ebx, %2, 2), %%ecx \n\t" // 7*stride |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2621 "movq (%0), %%mm0 \n\t" // L0 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2622 "movq (%0, %2), %%mm1 \n\t" // L1 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2623 "movq (%0, %2, 2), %%mm2 \n\t" // L2 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2624 "movq (%0, %%eax), %%mm3 \n\t" // L3 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2625 "movq (%1), %%mm4 \n\t" // R0 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2626 "movq (%1, %2), %%mm5 \n\t" // R1 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2627 "movq (%1, %2, 2), %%mm6 \n\t" // R2 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2628 "movq (%1, %%eax), %%mm7 \n\t" // R3 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2629 PAVGB(%%mm4, %%mm0) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2630 PAVGB(%%mm5, %%mm1) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2631 PAVGB(%%mm6, %%mm2) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2632 PAVGB(%%mm7, %%mm3) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2633 PAVGB(%%mm4, %%mm0) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2634 PAVGB(%%mm5, %%mm1) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2635 PAVGB(%%mm6, %%mm2) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2636 PAVGB(%%mm7, %%mm3) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2637 PAVGB(%%mm4, %%mm0) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2638 PAVGB(%%mm5, %%mm1) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2639 PAVGB(%%mm6, %%mm2) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2640 PAVGB(%%mm7, %%mm3) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2641 "movq %%mm0, (%1) \n\t" // R0 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2642 "movq %%mm1, (%1, %2) \n\t" // R1 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2643 "movq %%mm2, (%1, %2, 2) \n\t" // R2 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2644 "movq %%mm3, (%1, %%eax) \n\t" // R3 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2645 "movq %%mm0, (%0) \n\t" // L0 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2646 "movq %%mm1, (%0, %2) \n\t" // L1 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2647 "movq %%mm2, (%0, %2, 2) \n\t" // L2 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2648 "movq %%mm3, (%0, %%eax) \n\t" // L3 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2649 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2650 "movq (%0, %2, 4), %%mm0 \n\t" // L4 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2651 "movq (%0, %%ebx), %%mm1 \n\t" // L5 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2652 "movq (%0, %%eax, 2), %%mm2 \n\t" // L6 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2653 "movq (%0, %%ecx), %%mm3 \n\t" // L7 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2654 "movq (%1, %2, 4), %%mm4 \n\t" // R4 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2655 "movq (%1, %%ebx), %%mm5 \n\t" // R5 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2656 "movq (%1, %%eax, 2), %%mm6 \n\t" // R6 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2657 "movq (%1, %%ecx), %%mm7 \n\t" // R7 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2658 PAVGB(%%mm4, %%mm0) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2659 PAVGB(%%mm5, %%mm1) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2660 PAVGB(%%mm6, %%mm2) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2661 PAVGB(%%mm7, %%mm3) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2662 PAVGB(%%mm4, %%mm0) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2663 PAVGB(%%mm5, %%mm1) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2664 PAVGB(%%mm6, %%mm2) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2665 PAVGB(%%mm7, %%mm3) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2666 PAVGB(%%mm4, %%mm0) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2667 PAVGB(%%mm5, %%mm1) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2668 PAVGB(%%mm6, %%mm2) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2669 PAVGB(%%mm7, %%mm3) |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2670 "movq %%mm0, (%1, %2, 4) \n\t" // R4 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2671 "movq %%mm1, (%1, %%ebx) \n\t" // R5 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2672 "movq %%mm2, (%1, %%eax, 2) \n\t" // R6 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2673 "movq %%mm3, (%1, %%ecx) \n\t" // R7 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2674 "movq %%mm0, (%0, %2, 4) \n\t" // L4 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2675 "movq %%mm1, (%0, %%ebx) \n\t" // L5 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2676 "movq %%mm2, (%0, %%eax, 2) \n\t" // L6 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2677 "movq %%mm3, (%0, %%ecx) \n\t" // L7 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2678 |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2679 "4: \n\t" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2680 |
158 | 2681 :: "r" (src), "r" (tempBlured), "r"(stride), "m" (tempBluredPast) |
157
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2682 : "%eax", "%ebx", "%ecx", "memory" |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2683 ); |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2684 //printf("%d\n", test); |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2685 #else |
156 | 2686 int y; |
2687 int d=0; | |
2688 int sysd=0; | |
158 | 2689 int i; |
156 | 2690 |
2691 for(y=0; y<8; y++) | |
2692 { | |
2693 int x; | |
2694 for(x=0; x<8; x++) | |
2695 { | |
2696 int ref= tempBlured[ x + y*stride ]; | |
2697 int cur= src[ x + y*stride ]; | |
2698 int d1=ref - cur; | |
157
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2699 // if(x==0 || x==7) d1+= d1>>1; |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2700 // if(y==0 || y==7) d1+= d1>>1; |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2701 // d+= ABS(d1); |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2702 d+= d1*d1; |
156 | 2703 sysd+= d1; |
2704 } | |
2705 } | |
158 | 2706 i=d; |
2707 d= ( | |
2708 4*d | |
2709 +(*(tempBluredPast-256)) | |
2710 +(*(tempBluredPast-1))+ (*(tempBluredPast+1)) | |
2711 +(*(tempBluredPast+256)) | |
2712 +4)>>3; | |
2713 *tempBluredPast=i; | |
2714 // ((*tempBluredPast)*3 + d + 2)>>2; | |
2715 | |
156 | 2716 //printf("%d %d %d\n", maxNoise[0], maxNoise[1], maxNoise[2]); |
2717 /* | |
2718 Switch between | |
2719 1 0 0 0 0 0 0 (0) | |
2720 64 32 16 8 4 2 1 (1) | |
2721 64 48 36 27 20 15 11 (33) (approx) | |
2722 64 56 49 43 37 33 29 (200) (approx) | |
2723 */ | |
2724 if(d > maxNoise[1]) | |
2725 { | |
2726 if(d < maxNoise[2]) | |
2727 { | |
2728 for(y=0; y<8; y++) | |
2729 { | |
2730 int x; | |
2731 for(x=0; x<8; x++) | |
2732 { | |
2733 int ref= tempBlured[ x + y*stride ]; | |
2734 int cur= src[ x + y*stride ]; | |
2735 tempBlured[ x + y*stride ]= | |
2736 src[ x + y*stride ]= | |
2737 (ref + cur + 1)>>1; | |
2738 } | |
2739 } | |
2740 } | |
2741 else | |
2742 { | |
2743 for(y=0; y<8; y++) | |
2744 { | |
2745 int x; | |
2746 for(x=0; x<8; x++) | |
2747 { | |
2748 tempBlured[ x + y*stride ]= src[ x + y*stride ]; | |
2749 } | |
2750 } | |
2751 } | |
2752 } | |
2753 else | |
2754 { | |
2755 if(d < maxNoise[0]) | |
2756 { | |
2757 for(y=0; y<8; y++) | |
2758 { | |
2759 int x; | |
2760 for(x=0; x<8; x++) | |
2761 { | |
2762 int ref= tempBlured[ x + y*stride ]; | |
2763 int cur= src[ x + y*stride ]; | |
2764 tempBlured[ x + y*stride ]= | |
2765 src[ x + y*stride ]= | |
2766 (ref*7 + cur + 4)>>3; | |
2767 } | |
2768 } | |
2769 } | |
2770 else | |
2771 { | |
2772 for(y=0; y<8; y++) | |
2773 { | |
2774 int x; | |
2775 for(x=0; x<8; x++) | |
2776 { | |
2777 int ref= tempBlured[ x + y*stride ]; | |
2778 int cur= src[ x + y*stride ]; | |
2779 tempBlured[ x + y*stride ]= | |
2780 src[ x + y*stride ]= | |
2781 (ref*3 + cur + 2)>>2; | |
2782 } | |
2783 } | |
2784 } | |
2785 } | |
157
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
2786 #endif |
156 | 2787 } |
2788 | |
102 | 2789 #ifdef HAVE_ODIVX_POSTPROCESS |
2790 #include "../opendivx/postprocess.h" | |
2791 int use_old_pp=0; | |
2792 #endif | |
96 | 2793 |
102 | 2794 static void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height, |
156 | 2795 QP_STORE_T QPs[], int QPStride, int isColor, struct PPMode *ppMode); |
96 | 2796 |
116 | 2797 /* -pp Command line Help |
2798 NOTE/FIXME: put this at an appropriate place (--help, html docs, man mplayer)? | |
2799 | |
2800 -pp <filterName>[:<option>[:<option>...]][,[-]<filterName>[:<option>...]]... | |
2801 | |
2802 long form example: | |
2803 -pp vdeblock:autoq,hdeblock:autoq,linblenddeint -pp default,-vdeblock | |
2804 short form example: | |
2805 -pp vb:a,hb:a,lb -pp de,-vb | |
156 | 2806 more examples: |
2807 -pp tn:64:128:256 | |
116 | 2808 |
2809 Filters Options | |
2810 short long name short long option Description | |
2811 * * a autoq cpu power dependant enabler | |
2812 c chrom chrominance filtring enabled | |
2813 y nochrom chrominance filtring disabled | |
2814 hb hdeblock horizontal deblocking filter | |
2815 vb vdeblock vertical deblocking filter | |
2816 vr rkvdeblock | |
2817 h1 x1hdeblock Experimental horizontal deblock filter 1 | |
2818 v1 x1vdeblock Experimental vertical deblock filter 1 | |
2819 dr dering not implemented yet | |
2820 al autolevels automatic brightness / contrast fixer | |
2821 f fullyrange stretch luminance range to (0..255) | |
2822 lb linblenddeint linear blend deinterlacer | |
2823 li linipoldeint linear interpolating deinterlacer | |
2824 ci cubicipoldeint cubic interpolating deinterlacer | |
2825 md mediandeint median deinterlacer | |
2826 de default hdeblock:a,vdeblock:a,dering:a,autolevels | |
2827 fa fast x1hdeblock:a,x1vdeblock:a,dering:a,autolevels | |
156 | 2828 tn tmpnoise (3 Thresholds) Temporal Noise Reducer |
116 | 2829 */ |
2830 | |
2831 /** | |
2832 * returns a PPMode struct which will have a non 0 error variable if an error occured | |
2833 * name is the string after "-pp" on the command line | |
2834 * quality is a number from 0 to GET_PP_QUALITY_MAX | |
2835 */ | |
2836 struct PPMode getPPModeByNameAndQuality(char *name, int quality) | |
2837 { | |
2838 char temp[GET_MODE_BUFFER_SIZE]; | |
2839 char *p= temp; | |
2840 char *filterDelimiters= ","; | |
2841 char *optionDelimiters= ":"; | |
156 | 2842 struct PPMode ppMode= {0,0,0,0,0,0,{150,200,400}}; |
116 | 2843 char *filterToken; |
2844 | |
2845 strncpy(temp, name, GET_MODE_BUFFER_SIZE); | |
2846 | |
156 | 2847 printf("%s\n", name); |
2848 | |
116 | 2849 for(;;){ |
2850 char *filterName; | |
156 | 2851 int q= 1000000; //GET_PP_QUALITY_MAX; |
116 | 2852 int chrom=-1; |
2853 char *option; | |
2854 char *options[OPTIONS_ARRAY_SIZE]; | |
2855 int i; | |
2856 int filterNameOk=0; | |
2857 int numOfUnknownOptions=0; | |
2858 int enable=1; //does the user want us to enabled or disabled the filter | |
2859 | |
2860 filterToken= strtok(p, filterDelimiters); | |
2861 if(filterToken == NULL) break; | |
156 | 2862 p+= strlen(filterToken) + 1; // p points to next filterToken |
116 | 2863 filterName= strtok(filterToken, optionDelimiters); |
2864 printf("%s::%s\n", filterToken, filterName); | |
2865 | |
2866 if(*filterName == '-') | |
2867 { | |
2868 enable=0; | |
2869 filterName++; | |
2870 } | |
156 | 2871 |
116 | 2872 for(;;){ //for all options |
2873 option= strtok(NULL, optionDelimiters); | |
2874 if(option == NULL) break; | |
2875 | |
2876 printf("%s\n", option); | |
2877 if(!strcmp("autoq", option) || !strcmp("a", option)) q= quality; | |
2878 else if(!strcmp("nochrom", option) || !strcmp("y", option)) chrom=0; | |
2879 else if(!strcmp("chrom", option) || !strcmp("c", option)) chrom=1; | |
2880 else | |
2881 { | |
2882 options[numOfUnknownOptions] = option; | |
2883 numOfUnknownOptions++; | |
2884 } | |
2885 if(numOfUnknownOptions >= OPTIONS_ARRAY_SIZE-1) break; | |
2886 } | |
156 | 2887 options[numOfUnknownOptions] = NULL; |
116 | 2888 |
2889 /* replace stuff from the replace Table */ | |
2890 for(i=0; replaceTable[2*i]!=NULL; i++) | |
2891 { | |
2892 if(!strcmp(replaceTable[2*i], filterName)) | |
2893 { | |
2894 int newlen= strlen(replaceTable[2*i + 1]); | |
2895 int plen; | |
2896 int spaceLeft; | |
2897 | |
2898 if(p==NULL) p= temp, *p=0; //last filter | |
2899 else p--, *p=','; //not last filter | |
2900 | |
2901 plen= strlen(p); | |
2902 spaceLeft= (int)p - (int)temp + plen; | |
2903 if(spaceLeft + newlen >= GET_MODE_BUFFER_SIZE) | |
2904 { | |
2905 ppMode.error++; | |
2906 break; | |
2907 } | |
2908 memmove(p + newlen, p, plen+1); | |
2909 memcpy(p, replaceTable[2*i + 1], newlen); | |
2910 filterNameOk=1; | |
2911 } | |
2912 } | |
2913 | |
2914 for(i=0; filters[i].shortName!=NULL; i++) | |
2915 { | |
156 | 2916 // printf("Compareing %s, %s, %s\n", filters[i].shortName,filters[i].longName, filterName); |
116 | 2917 if( !strcmp(filters[i].longName, filterName) |
2918 || !strcmp(filters[i].shortName, filterName)) | |
2919 { | |
2920 ppMode.lumMode &= ~filters[i].mask; | |
2921 ppMode.chromMode &= ~filters[i].mask; | |
2922 | |
2923 filterNameOk=1; | |
2924 if(!enable) break; // user wants to disable it | |
2925 | |
2926 if(q >= filters[i].minLumQuality) | |
2927 ppMode.lumMode|= filters[i].mask; | |
2928 if(chrom==1 || (chrom==-1 && filters[i].chromDefault)) | |
2929 if(q >= filters[i].minChromQuality) | |
2930 ppMode.chromMode|= filters[i].mask; | |
2931 | |
2932 if(filters[i].mask == LEVEL_FIX) | |
2933 { | |
2934 int o; | |
2935 ppMode.minAllowedY= 16; | |
2936 ppMode.maxAllowedY= 234; | |
2937 for(o=0; options[o]!=NULL; o++) | |
2938 if( !strcmp(options[o],"fullyrange") | |
2939 ||!strcmp(options[o],"f")) | |
2940 { | |
2941 ppMode.minAllowedY= 0; | |
2942 ppMode.maxAllowedY= 255; | |
2943 numOfUnknownOptions--; | |
2944 } | |
2945 } | |
156 | 2946 else if(filters[i].mask == TEMP_NOISE_FILTER) |
2947 { | |
2948 int o; | |
2949 int numOfNoises=0; | |
2950 ppMode.maxTmpNoise[0]= 150; | |
2951 ppMode.maxTmpNoise[1]= 200; | |
2952 ppMode.maxTmpNoise[2]= 400; | |
2953 | |
2954 for(o=0; options[o]!=NULL; o++) | |
2955 { | |
2956 char *tail; | |
2957 ppMode.maxTmpNoise[numOfNoises]= | |
2958 strtol(options[o], &tail, 0); | |
2959 if(tail!=options[o]) | |
2960 { | |
2961 numOfNoises++; | |
2962 numOfUnknownOptions--; | |
2963 if(numOfNoises >= 3) break; | |
2964 } | |
2965 } | |
2966 } | |
116 | 2967 } |
2968 } | |
2969 if(!filterNameOk) ppMode.error++; | |
2970 ppMode.error += numOfUnknownOptions; | |
2971 } | |
2972 | |
152 | 2973 #ifdef HAVE_ODIVX_POSTPROCESS |
116 | 2974 if(ppMode.lumMode & H_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_Y_H; |
2975 if(ppMode.lumMode & V_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_Y_V; | |
2976 if(ppMode.chromMode & H_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_C_H; | |
2977 if(ppMode.chromMode & V_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_C_V; | |
2978 if(ppMode.lumMode & DERING) ppMode.oldMode |= PP_DERING_Y; | |
2979 if(ppMode.chromMode & DERING) ppMode.oldMode |= PP_DERING_C; | |
152 | 2980 #endif |
116 | 2981 |
2982 return ppMode; | |
2983 } | |
2984 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2985 /** |
156 | 2986 * Obsolete, dont use it, use postprocess2() instead |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2987 */ |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2988 void postprocess(unsigned char * src[], int src_stride, |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2989 unsigned char * dst[], int dst_stride, |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2990 int horizontal_size, int vertical_size, |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2991 QP_STORE_T *QP_store, int QP_stride, |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2992 int mode) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2993 { |
156 | 2994 struct PPMode ppMode; |
2995 static QP_STORE_T zeroArray[2048/8]; | |
116 | 2996 /* |
2997 static int qual=0; | |
2998 | |
156 | 2999 ppMode= getPPModeByNameAndQuality("fast,default,-hdeblock,-vdeblock,tmpnoise:150:200:300", qual); |
3000 printf("OK\n"); | |
116 | 3001 qual++; |
3002 qual%=7; | |
156 | 3003 printf("\n%X %X %X %X :%d: %d %d %d\n", ppMode.lumMode, ppMode.chromMode, ppMode.oldMode, ppMode.error, |
3004 qual, ppMode.maxTmpNoise[0], ppMode.maxTmpNoise[1], ppMode.maxTmpNoise[2]); | |
116 | 3005 postprocess2(src, src_stride, dst, dst_stride, |
3006 horizontal_size, vertical_size, QP_store, QP_stride, &ppMode); | |
3007 | |
3008 return; | |
3009 */ | |
152 | 3010 if(QP_store==NULL) |
3011 { | |
3012 QP_store= zeroArray; | |
3013 QP_stride= 0; | |
3014 } | |
96 | 3015 |
156 | 3016 ppMode.lumMode= mode; |
3017 mode= ((mode&0xFF)>>4) | (mode&0xFFFFFF00); | |
3018 ppMode.chromMode= mode; | |
157
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
3019 ppMode.maxTmpNoise[0]= 700; |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
3020 ppMode.maxTmpNoise[1]= 1500; |
bc12fd7e6153
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
156
diff
changeset
|
3021 ppMode.maxTmpNoise[2]= 3000; |
156 | 3022 |
102 | 3023 #ifdef HAVE_ODIVX_POSTPROCESS |
3024 // Note: I could make this shit outside of this file, but it would mean one | |
3025 // more function call... | |
3026 if(use_old_pp){ | |
3027 odivx_postprocess(src,src_stride,dst,dst_stride,horizontal_size,vertical_size,QP_store,QP_stride,mode); | |
3028 return; | |
3029 } | |
3030 #endif | |
3031 | |
96 | 3032 postProcess(src[0], src_stride, dst[0], dst_stride, |
156 | 3033 horizontal_size, vertical_size, QP_store, QP_stride, 0, &ppMode); |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3034 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3035 horizontal_size >>= 1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3036 vertical_size >>= 1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3037 src_stride >>= 1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3038 dst_stride >>= 1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3039 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3040 if(1) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3041 { |
96 | 3042 postProcess(src[1], src_stride, dst[1], dst_stride, |
156 | 3043 horizontal_size, vertical_size, QP_store, QP_stride, 1, &ppMode); |
96 | 3044 postProcess(src[2], src_stride, dst[2], dst_stride, |
156 | 3045 horizontal_size, vertical_size, QP_store, QP_stride, 2, &ppMode); |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3046 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3047 else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3048 { |
156 | 3049 memset(dst[1], 128, dst_stride*vertical_size); |
3050 memset(dst[2], 128, dst_stride*vertical_size); | |
3051 // memcpy(dst[1], src[1], src_stride*horizontal_size); | |
3052 // memcpy(dst[2], src[2], src_stride*horizontal_size); | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3053 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3054 } |
102 | 3055 |
116 | 3056 void postprocess2(unsigned char * src[], int src_stride, |
3057 unsigned char * dst[], int dst_stride, | |
3058 int horizontal_size, int vertical_size, | |
3059 QP_STORE_T *QP_store, int QP_stride, | |
3060 struct PPMode *mode) | |
3061 { | |
3062 | |
152 | 3063 static QP_STORE_T zeroArray[2048/8]; |
3064 if(QP_store==NULL) | |
3065 { | |
3066 QP_store= zeroArray; | |
3067 QP_stride= 0; | |
3068 } | |
3069 | |
116 | 3070 #ifdef HAVE_ODIVX_POSTPROCESS |
3071 // Note: I could make this shit outside of this file, but it would mean one | |
3072 // more function call... | |
3073 if(use_old_pp){ | |
3074 odivx_postprocess(src,src_stride,dst,dst_stride,horizontal_size,vertical_size,QP_store,QP_stride, | |
3075 mode->oldMode); | |
3076 return; | |
3077 } | |
3078 #endif | |
3079 | |
3080 postProcess(src[0], src_stride, dst[0], dst_stride, | |
156 | 3081 horizontal_size, vertical_size, QP_store, QP_stride, 0, mode); |
116 | 3082 |
3083 horizontal_size >>= 1; | |
3084 vertical_size >>= 1; | |
3085 src_stride >>= 1; | |
3086 dst_stride >>= 1; | |
3087 | |
3088 postProcess(src[1], src_stride, dst[1], dst_stride, | |
156 | 3089 horizontal_size, vertical_size, QP_store, QP_stride, 1, mode); |
116 | 3090 postProcess(src[2], src_stride, dst[2], dst_stride, |
156 | 3091 horizontal_size, vertical_size, QP_store, QP_stride, 2, mode); |
116 | 3092 } |
3093 | |
3094 | |
96 | 3095 /** |
3096 * gets the mode flags for a given quality (larger values mean slower but better postprocessing) | |
102 | 3097 * 0 <= quality <= 6 |
96 | 3098 */ |
102 | 3099 int getPpModeForQuality(int quality){ |
3100 int modes[1+GET_PP_QUALITY_MAX]= { | |
3101 0, | |
3102 #if 1 | |
3103 // horizontal filters first | |
3104 LUM_H_DEBLOCK, | |
3105 LUM_H_DEBLOCK | LUM_V_DEBLOCK, | |
3106 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK, | |
3107 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK, | |
3108 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK | LUM_DERING, | |
3109 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK | LUM_DERING | CHROM_DERING | |
3110 #else | |
3111 // vertical filters first | |
96 | 3112 LUM_V_DEBLOCK, |
3113 LUM_V_DEBLOCK | LUM_H_DEBLOCK, | |
3114 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK, | |
3115 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK, | |
3116 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING, | |
3117 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING | CHROM_DERING | |
102 | 3118 #endif |
3119 }; | |
96 | 3120 |
102 | 3121 #ifdef HAVE_ODIVX_POSTPROCESS |
3122 int odivx_modes[1+GET_PP_QUALITY_MAX]= { | |
3123 0, | |
3124 PP_DEBLOCK_Y_H, | |
3125 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V, | |
3126 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H, | |
3127 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V, | |
3128 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V|PP_DERING_Y, | |
3129 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V|PP_DERING_Y|PP_DERING_C | |
3130 }; | |
3131 if(use_old_pp) return odivx_modes[quality]; | |
3132 #endif | |
3133 return modes[quality]; | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3134 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3135 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3136 /** |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3137 * 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
|
3138 * numLines must be a multiple of 4 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3139 * levelFix == 0 -> dont touch the brighness & contrast |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3140 */ |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3141 static inline void blockCopy(uint8_t dst[], int dstStride, uint8_t src[], int srcStride, |
164 | 3142 int levelFix) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3143 { |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
3144 #ifndef HAVE_MMX |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3145 int i; |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
3146 #endif |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3147 if(levelFix) |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3148 { |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3149 #ifdef HAVE_MMX |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3150 asm volatile( |
166 | 3151 "leal (%0,%2), %%eax \n\t" |
3152 "leal (%1,%3), %%ebx \n\t" | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3153 "movq packedYOffset, %%mm2 \n\t" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3154 "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
|
3155 "pxor %%mm4, %%mm4 \n\t" |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3156 |
166 | 3157 #define SCALED_CPY(src1, src2, dst1, dst2) \ |
3158 "movq " #src1 ", %%mm0 \n\t"\ | |
3159 "movq " #src1 ", %%mm5 \n\t"\ | |
101
fcf4e8fcb34b
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
100
diff
changeset
|
3160 "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
|
3161 "punpckhbw %%mm4, %%mm5 \n\t"\ |
117 | 3162 "psubw %%mm2, %%mm0 \n\t"\ |
3163 "psubw %%mm2, %%mm5 \n\t"\ | |
166 | 3164 "movq " #src2 ", %%mm1 \n\t"\ |
117 | 3165 "psllw $6, %%mm0 \n\t"\ |
3166 "psllw $6, %%mm5 \n\t"\ | |
101
fcf4e8fcb34b
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
100
diff
changeset
|
3167 "pmulhw %%mm3, %%mm0 \n\t"\ |
166 | 3168 "movq " #src2 ", %%mm6 \n\t"\ |
101
fcf4e8fcb34b
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
100
diff
changeset
|
3169 "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
|
3170 "punpcklbw %%mm4, %%mm1 \n\t"\ |
118
3dd1950ac98d
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
117
diff
changeset
|
3171 "punpckhbw %%mm4, %%mm6 \n\t"\ |
117 | 3172 "psubw %%mm2, %%mm1 \n\t"\ |
118
3dd1950ac98d
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
117
diff
changeset
|
3173 "psubw %%mm2, %%mm6 \n\t"\ |
117 | 3174 "psllw $6, %%mm1 \n\t"\ |
118
3dd1950ac98d
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
117
diff
changeset
|
3175 "psllw $6, %%mm6 \n\t"\ |
101
fcf4e8fcb34b
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
100
diff
changeset
|
3176 "pmulhw %%mm3, %%mm1 \n\t"\ |
118
3dd1950ac98d
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
117
diff
changeset
|
3177 "pmulhw %%mm3, %%mm6 \n\t"\ |
3dd1950ac98d
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
117
diff
changeset
|
3178 "packuswb %%mm5, %%mm0 \n\t"\ |
3dd1950ac98d
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
117
diff
changeset
|
3179 "packuswb %%mm6, %%mm1 \n\t"\ |
166 | 3180 "movq %%mm0, " #dst1 " \n\t"\ |
3181 "movq %%mm1, " #dst2 " \n\t"\ | |
3182 | |
3183 SCALED_CPY((%0) , (%0, %2) , (%1) , (%1, %3)) | |
3184 SCALED_CPY((%0, %2, 2), (%%eax, %2, 2), (%1, %3, 2), (%%ebx, %3, 2)) | |
3185 SCALED_CPY((%0, %2, 4), (%%eax, %2, 4), (%1, %3, 4), (%%ebx, %3, 4)) | |
3186 "leal (%%eax,%2,4), %%eax \n\t" | |
3187 "leal (%%ebx,%3,4), %%ebx \n\t" | |
3188 SCALED_CPY((%%eax, %2), (%%eax, %2, 2), (%%ebx, %3), (%%ebx, %3, 2)) | |
3189 | |
3190 | |
3191 : : "r"(src), | |
3192 "r"(dst), | |
3193 "r" (srcStride), | |
118
3dd1950ac98d
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
117
diff
changeset
|
3194 "r" (dstStride) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3195 : "%eax", "%ebx" |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3196 ); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3197 #else |
164 | 3198 for(i=0; i<8; i++) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3199 memcpy( &(dst[dstStride*i]), |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3200 &(src[srcStride*i]), BLOCK_SIZE); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3201 #endif |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3202 } |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3203 else |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3204 { |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3205 #ifdef HAVE_MMX |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3206 asm volatile( |
166 | 3207 "leal (%0,%2), %%eax \n\t" |
3208 "leal (%1,%3), %%ebx \n\t" | |
3209 | |
3210 #define SIMPLE_CPY(src1, src2, dst1, dst2) \ | |
3211 "movq " #src1 ", %%mm0 \n\t"\ | |
3212 "movq " #src2 ", %%mm1 \n\t"\ | |
3213 "movq %%mm0, " #dst1 " \n\t"\ | |
3214 "movq %%mm1, " #dst2 " \n\t"\ | |
3215 | |
3216 SIMPLE_CPY((%0) , (%0, %2) , (%1) , (%1, %3)) | |
3217 SIMPLE_CPY((%0, %2, 2), (%%eax, %2, 2), (%1, %3, 2), (%%ebx, %3, 2)) | |
3218 SIMPLE_CPY((%0, %2, 4), (%%eax, %2, 4), (%1, %3, 4), (%%ebx, %3, 4)) | |
3219 "leal (%%eax,%2,4), %%eax \n\t" | |
3220 "leal (%%ebx,%3,4), %%ebx \n\t" | |
3221 SIMPLE_CPY((%%eax, %2), (%%eax, %2, 2), (%%ebx, %3), (%%ebx, %3, 2)) | |
3222 | |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3223 : : "r" (src), |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3224 "r" (dst), |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3225 "r" (srcStride), |
164 | 3226 "r" (dstStride) |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3227 : "%eax", "%ebx" |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3228 ); |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3229 #else |
164 | 3230 for(i=0; i<8; i++) |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3231 memcpy( &(dst[dstStride*i]), |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3232 &(src[srcStride*i]), BLOCK_SIZE); |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3233 #endif |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3234 } |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3235 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3236 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3237 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3238 /** |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3239 * Filters array of bytes (Y or U or V values) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3240 */ |
102 | 3241 static void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height, |
156 | 3242 QP_STORE_T QPs[], int QPStride, int isColor, struct PPMode *ppMode) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3243 { |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3244 int x,y; |
156 | 3245 const int mode= isColor ? ppMode->chromMode : ppMode->lumMode; |
3246 | |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3247 /* 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
|
3248 after watching a black picture for 5 hours*/ |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3249 static uint64_t *yHistogram= NULL; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3250 int black=0, white=255; // blackest black and whitest white in the picture |
126 | 3251 int QPCorrecture= 256; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3252 |
111 | 3253 /* Temporary buffers for handling the last row(s) */ |
3254 static uint8_t *tempDst= NULL; | |
3255 static uint8_t *tempSrc= NULL; | |
3256 | |
112 | 3257 /* Temporary buffers for handling the last block */ |
3258 static uint8_t *tempDstBlock= NULL; | |
3259 static uint8_t *tempSrcBlock= NULL; | |
3260 | |
156 | 3261 /* Temporal noise reducing buffers */ |
3262 static uint8_t *tempBlured[3]= {NULL,NULL,NULL}; | |
158 | 3263 static uint32_t *tempBluredPast[3]= {NULL,NULL,NULL}; |
156 | 3264 |
164 | 3265 int copyAhead; |
3266 | |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
3267 #ifdef PP_FUNNY_STRIDE |
112 | 3268 uint8_t *dstBlockPtrBackup; |
3269 uint8_t *srcBlockPtrBackup; | |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
3270 #endif |
112 | 3271 |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
3272 #ifdef MORE_TIMING |
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
3273 long long T0, T1, diffTime=0; |
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
3274 #endif |
107 | 3275 #ifdef TIMING |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
3276 long long memcpyTime=0, vertTime=0, horizTime=0, sumTime; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3277 sumTime= rdtsc(); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3278 #endif |
140 | 3279 //mode= 0x7F; |
158 | 3280 #ifdef HAVE_MMX |
3281 maxTmpNoise[0]= ppMode->maxTmpNoise[0]; | |
3282 maxTmpNoise[1]= ppMode->maxTmpNoise[1]; | |
3283 maxTmpNoise[2]= ppMode->maxTmpNoise[2]; | |
3284 #endif | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3285 |
164 | 3286 if(mode & CUBIC_IPOL_DEINT_FILTER) copyAhead=16; |
3287 else if(mode & LINEAR_BLEND_DEINT_FILTER) copyAhead=14; | |
3288 else if( (mode & V_DEBLOCK) | |
3289 || (mode & LINEAR_IPOL_DEINT_FILTER) | |
3290 || (mode & MEDIAN_DEINT_FILTER)) copyAhead=13; | |
3291 else if(mode & V_X1_FILTER) copyAhead=11; | |
3292 else if(mode & V_RK1_FILTER) copyAhead=10; | |
3293 else if(mode & DERING) copyAhead=9; | |
3294 else copyAhead=8; | |
3295 | |
3296 copyAhead-= 8; | |
3297 | |
111 | 3298 if(tempDst==NULL) |
3299 { | |
3300 tempDst= (uint8_t*)memalign(8, 1024*24); | |
3301 tempSrc= (uint8_t*)memalign(8, 1024*24); | |
112 | 3302 tempDstBlock= (uint8_t*)memalign(8, 1024*24); |
3303 tempSrcBlock= (uint8_t*)memalign(8, 1024*24); | |
111 | 3304 } |
3305 | |
156 | 3306 if(tempBlured[isColor]==NULL && (mode & TEMP_NOISE_FILTER)) |
3307 { | |
3308 // printf("%d %d %d\n", isColor, dstStride, height); | |
3309 //FIXME works only as long as the size doesnt increase | |
3310 //Note:the +17*1024 is just there so i dont have to worry about r/w over te end | |
3311 tempBlured[isColor]= (uint8_t*)memalign(8, dstStride*((height+7)&(~7)) + 17*1024); | |
158 | 3312 tempBluredPast[isColor]= (uint32_t*)memalign(8, 256*((height+7)&(~7))/2 + 17*1024); |
156 | 3313 |
3314 memset(tempBlured[isColor], 0, dstStride*((height+7)&(~7)) + 17*1024); | |
158 | 3315 memset(tempBluredPast[isColor], 0, 256*((height+7)&(~7))/2 + 17*1024); |
156 | 3316 } |
3317 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3318 if(!yHistogram) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3319 { |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3320 int i; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3321 yHistogram= (uint64_t*)malloc(8*256); |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3322 for(i=0; i<256; i++) yHistogram[i]= width*height/64*15/256; |
112 | 3323 |
3324 if(mode & FULL_Y_RANGE) | |
3325 { | |
3326 maxAllowedY=255; | |
3327 minAllowedY=0; | |
3328 } | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3329 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3330 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3331 if(!isColor) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3332 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3333 uint64_t sum= 0; |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3334 int i; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3335 static int framenum= -1; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3336 uint64_t maxClipped; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3337 uint64_t clipped; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3338 double scale; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3339 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3340 framenum++; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3341 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
|
3342 |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3343 for(i=0; i<256; i++) |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3344 { |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3345 sum+= yHistogram[i]; |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3346 // printf("%d ", yHistogram[i]); |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3347 } |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3348 // printf("\n\n"); |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3349 |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3350 /* we allways get a completly black picture first */ |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3351 maxClipped= (uint64_t)(sum * maxClippedThreshold); |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3352 |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3353 clipped= sum; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3354 for(black=255; black>0; black--) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3355 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3356 if(clipped < maxClipped) break; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3357 clipped-= yHistogram[black]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3358 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3359 |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3360 clipped= sum; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3361 for(white=0; white<256; white++) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3362 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3363 if(clipped < maxClipped) break; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3364 clipped-= yHistogram[white]; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3365 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3366 |
117 | 3367 packedYOffset= (black - minAllowedY) & 0xFFFF; |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3368 packedYOffset|= packedYOffset<<32; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3369 packedYOffset|= packedYOffset<<16; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3370 |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3371 scale= (double)(maxAllowedY - minAllowedY) / (double)(white-black); |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3372 |
117 | 3373 packedYScale= (uint16_t)(scale*1024.0 + 0.5); |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3374 packedYScale|= packedYScale<<32; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3375 packedYScale|= packedYScale<<16; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3376 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3377 else |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3378 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3379 packedYScale= 0x0100010001000100LL; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3380 packedYOffset= 0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3381 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3382 |
126 | 3383 if(mode & LEVEL_FIX) QPCorrecture= packedYScale &0xFFFF; |
3384 else QPCorrecture= 256; | |
3385 | |
148
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
3386 /* copy & deinterlace first row of blocks */ |
142 | 3387 y=-BLOCK_SIZE; |
3388 { | |
3389 //1% speedup if these are here instead of the inner loop | |
3390 uint8_t *srcBlock= &(src[y*srcStride]); | |
3391 uint8_t *dstBlock= &(dst[y*dstStride]); | |
3392 | |
3393 dstBlock= tempDst + dstStride; | |
3394 | |
3395 // From this point on it is guranteed that we can read and write 16 lines downward | |
3396 // finish 1 block before the next otherwise we´ll might have a problem | |
3397 // with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing | |
3398 for(x=0; x<width; x+=BLOCK_SIZE) | |
3399 { | |
3400 | |
3401 #ifdef HAVE_MMX2 | |
3402 /* | |
3403 prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32); | |
3404 prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32); | |
3405 prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32); | |
3406 prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32); | |
3407 */ | |
3408 | |
3409 asm( | |
3410 "movl %4, %%eax \n\t" | |
3411 "shrl $2, %%eax \n\t" | |
3412 "andl $6, %%eax \n\t" | |
164 | 3413 "addl %5, %%eax \n\t" |
142 | 3414 "movl %%eax, %%ebx \n\t" |
3415 "imul %1, %%eax \n\t" | |
3416 "imul %3, %%ebx \n\t" | |
3417 "prefetchnta 32(%%eax, %0) \n\t" | |
3418 "prefetcht0 32(%%ebx, %2) \n\t" | |
3419 "addl %1, %%eax \n\t" | |
3420 "addl %3, %%ebx \n\t" | |
3421 "prefetchnta 32(%%eax, %0) \n\t" | |
3422 "prefetcht0 32(%%ebx, %2) \n\t" | |
3423 :: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride), | |
164 | 3424 "m" (x), "m" (copyAhead) |
142 | 3425 : "%eax", "%ebx" |
3426 ); | |
3427 | |
3428 #elif defined(HAVE_3DNOW) | |
3429 //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ... | |
3430 /* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32); | |
3431 prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32); | |
3432 prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32); | |
3433 prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32); | |
3434 */ | |
3435 #endif | |
3436 | |
164 | 3437 blockCopy(dstBlock + dstStride*copyAhead, dstStride, |
3438 srcBlock + srcStride*copyAhead, srcStride, mode & LEVEL_FIX); | |
142 | 3439 |
3440 if(mode & LINEAR_IPOL_DEINT_FILTER) | |
3441 deInterlaceInterpolateLinear(dstBlock, dstStride); | |
3442 else if(mode & LINEAR_BLEND_DEINT_FILTER) | |
3443 deInterlaceBlendLinear(dstBlock, dstStride); | |
3444 else if(mode & MEDIAN_DEINT_FILTER) | |
3445 deInterlaceMedian(dstBlock, dstStride); | |
3446 else if(mode & CUBIC_IPOL_DEINT_FILTER) | |
3447 deInterlaceInterpolateCubic(dstBlock, dstStride); | |
3448 /* else if(mode & CUBIC_BLEND_DEINT_FILTER) | |
3449 deInterlaceBlendCubic(dstBlock, dstStride); | |
3450 */ | |
3451 dstBlock+=8; | |
3452 srcBlock+=8; | |
3453 } | |
164 | 3454 memcpy(&(dst[y*dstStride]) + 8*dstStride, tempDst + 9*dstStride, copyAhead*dstStride ); |
142 | 3455 } |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3456 |
111 | 3457 for(y=0; y<height; y+=BLOCK_SIZE) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3458 { |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3459 //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
|
3460 uint8_t *srcBlock= &(src[y*srcStride]); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3461 uint8_t *dstBlock= &(dst[y*dstStride]); |
126 | 3462 #ifdef ARCH_X86 |
3463 int *QPptr= isColor ? &QPs[(y>>3)*QPStride] :&QPs[(y>>4)*QPStride]; | |
3464 int QPDelta= isColor ? 1<<(32-3) : 1<<(32-4); | |
3465 int QPFrac= QPDelta; | |
128
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3466 uint8_t *tempBlock1= tempBlocks; |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3467 uint8_t *tempBlock2= tempBlocks + 8; |
126 | 3468 #endif |
156 | 3469 int QP=0; |
130 | 3470 /* can we mess with a 8x16 block from srcBlock/dstBlock downwards and 1 line upwards |
3471 if not than use a temporary buffer */ | |
111 | 3472 if(y+15 >= height) |
3473 { | |
156 | 3474 int i; |
164 | 3475 /* copy from line (copyAhead) to (copyAhead+7) of src, these will be copied with |
111 | 3476 blockcopy to dst later */ |
164 | 3477 memcpy(tempSrc + srcStride*copyAhead, srcBlock + srcStride*copyAhead, |
3478 srcStride*MAX(height-y-copyAhead, 0) ); | |
3479 | |
3480 /* duplicate last line of src to fill the void upto line (copyAhead+7) */ | |
3481 for(i=MAX(height-y, 8); i<copyAhead+8; i++) | |
156 | 3482 memcpy(tempSrc + srcStride*i, src + srcStride*(height-1), srcStride); |
3483 | |
164 | 3484 /* copy up to (copyAhead+1) lines of dst (line -1 to (copyAhead-1))*/ |
3485 memcpy(tempDst, dstBlock - dstStride, dstStride*MIN(height-y+1, copyAhead+1) ); | |
3486 | |
3487 /* duplicate last line of dst to fill the void upto line (copyAhead) */ | |
3488 for(i=height-y+1; i<=copyAhead; i++) | |
156 | 3489 memcpy(tempDst + dstStride*i, dst + dstStride*(height-1), dstStride); |
3490 | |
130 | 3491 dstBlock= tempDst + dstStride; |
111 | 3492 srcBlock= tempSrc; |
3493 } | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3494 |
112 | 3495 // 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
|
3496 // 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
|
3497 // 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
|
3498 for(x=0; x<width; x+=BLOCK_SIZE) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3499 { |
97
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3500 const int stride= dstStride; |
128
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3501 uint8_t *tmpXchg; |
126 | 3502 #ifdef ARCH_X86 |
156 | 3503 QP= *QPptr; |
126 | 3504 asm volatile( |
3505 "addl %2, %1 \n\t" | |
3506 "sbbl %%eax, %%eax \n\t" | |
3507 "shll $2, %%eax \n\t" | |
3508 "subl %%eax, %0 \n\t" | |
3509 : "+r" (QPptr), "+m" (QPFrac) | |
3510 : "r" (QPDelta) | |
3511 : "%eax" | |
3512 ); | |
3513 #else | |
156 | 3514 QP= isColor ? |
126 | 3515 QPs[(y>>3)*QPStride + (x>>3)]: |
3516 QPs[(y>>4)*QPStride + (x>>4)]; | |
3517 #endif | |
3518 if(!isColor) | |
121 | 3519 { |
126 | 3520 QP= (QP* QPCorrecture)>>8; |
148
1cfc4d567c0a
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
142
diff
changeset
|
3521 yHistogram[ srcBlock[srcStride*12 + 4] ]++; |
121 | 3522 } |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3523 #ifdef HAVE_MMX |
111 | 3524 asm volatile( |
3525 "movd %0, %%mm7 \n\t" | |
3526 "packuswb %%mm7, %%mm7 \n\t" // 0, 0, 0, QP, 0, 0, 0, QP | |
3527 "packuswb %%mm7, %%mm7 \n\t" // 0,QP, 0, QP, 0,QP, 0, QP | |
3528 "packuswb %%mm7, %%mm7 \n\t" // QP,..., QP | |
3529 "movq %%mm7, pQPb \n\t" | |
3530 : : "r" (QP) | |
3531 ); | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3532 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3533 |
107 | 3534 #ifdef MORE_TIMING |
111 | 3535 T0= rdtsc(); |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3536 #endif |
96 | 3537 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3538 #ifdef HAVE_MMX2 |
126 | 3539 /* |
3540 prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32); | |
3541 prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32); | |
3542 prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32); | |
3543 prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32); | |
3544 */ | |
3545 | |
3546 asm( | |
3547 "movl %4, %%eax \n\t" | |
3548 "shrl $2, %%eax \n\t" | |
3549 "andl $6, %%eax \n\t" | |
164 | 3550 "addl %5, %%eax \n\t" |
126 | 3551 "movl %%eax, %%ebx \n\t" |
3552 "imul %1, %%eax \n\t" | |
3553 "imul %3, %%ebx \n\t" | |
3554 "prefetchnta 32(%%eax, %0) \n\t" | |
3555 "prefetcht0 32(%%ebx, %2) \n\t" | |
3556 "addl %1, %%eax \n\t" | |
3557 "addl %3, %%ebx \n\t" | |
3558 "prefetchnta 32(%%eax, %0) \n\t" | |
3559 "prefetcht0 32(%%ebx, %2) \n\t" | |
3560 :: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride), | |
164 | 3561 "m" (x), "m" (copyAhead) |
126 | 3562 : "%eax", "%ebx" |
3563 ); | |
3564 | |
96 | 3565 #elif defined(HAVE_3DNOW) |
3566 //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ... | |
111 | 3567 /* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32); |
3568 prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32); | |
3569 prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32); | |
3570 prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32); | |
96 | 3571 */ |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3572 #endif |
111 | 3573 |
113 | 3574 #ifdef PP_FUNNY_STRIDE |
112 | 3575 //can we mess with a 8x16 block, if not use a temp buffer, yes again |
3576 if(x+7 >= width) | |
3577 { | |
3578 int i; | |
3579 dstBlockPtrBackup= dstBlock; | |
3580 srcBlockPtrBackup= srcBlock; | |
3581 | |
3582 for(i=0;i<BLOCK_SIZE*2; i++) | |
3583 { | |
3584 memcpy(tempSrcBlock+i*srcStride, srcBlock+i*srcStride, width-x); | |
3585 memcpy(tempDstBlock+i*dstStride, dstBlock+i*dstStride, width-x); | |
3586 } | |
3587 | |
3588 dstBlock= tempDstBlock; | |
3589 srcBlock= tempSrcBlock; | |
3590 } | |
113 | 3591 #endif |
112 | 3592 |
164 | 3593 blockCopy(dstBlock + dstStride*copyAhead, dstStride, |
3594 srcBlock + srcStride*copyAhead, srcStride, mode & LEVEL_FIX); | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3595 |
111 | 3596 if(mode & LINEAR_IPOL_DEINT_FILTER) |
3597 deInterlaceInterpolateLinear(dstBlock, dstStride); | |
3598 else if(mode & LINEAR_BLEND_DEINT_FILTER) | |
3599 deInterlaceBlendLinear(dstBlock, dstStride); | |
3600 else if(mode & MEDIAN_DEINT_FILTER) | |
3601 deInterlaceMedian(dstBlock, dstStride); | |
3602 else if(mode & CUBIC_IPOL_DEINT_FILTER) | |
3603 deInterlaceInterpolateCubic(dstBlock, dstStride); | |
3604 /* else if(mode & CUBIC_BLEND_DEINT_FILTER) | |
3605 deInterlaceBlendCubic(dstBlock, dstStride); | |
106
389391a6d0bf
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
105
diff
changeset
|
3606 */ |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3607 |
111 | 3608 /* only deblock if we have 2 blocks */ |
3609 if(y + 8 < height) | |
3610 { | |
107 | 3611 #ifdef MORE_TIMING |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3612 T1= rdtsc(); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3613 memcpyTime+= T1-T0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3614 T0=T1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3615 #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
|
3616 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
|
3617 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
|
3618 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
|
3619 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
|
3620 else if(mode & V_DEBLOCK) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3621 { |
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
|
3622 if( isVertDC(dstBlock, stride)) |
96 | 3623 { |
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
|
3624 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
|
3625 doVertLowPass(dstBlock, stride, QP); |
96 | 3626 } |
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
|
3627 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
|
3628 doVertDefFilter(dstBlock, stride, QP); |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3629 } |
107 | 3630 #ifdef MORE_TIMING |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3631 T1= rdtsc(); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3632 vertTime+= T1-T0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3633 T0=T1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3634 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3635 } |
130 | 3636 |
128
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3637 #ifdef HAVE_MMX |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3638 transpose1(tempBlock1, tempBlock2, dstBlock, dstStride); |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3639 #endif |
111 | 3640 /* check if we have a previous block to deblock it with dstBlock */ |
112 | 3641 if(x - 8 >= 0) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3642 { |
107 | 3643 #ifdef MORE_TIMING |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3644 T0= rdtsc(); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3645 #endif |
128
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3646 #ifdef HAVE_MMX |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3647 if(mode & H_RK1_FILTER) |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3648 vertRK1Filter(tempBlock1, 16, QP); |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3649 else if(mode & H_X1_FILTER) |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3650 vertX1Filter(tempBlock1, 16, QP); |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3651 else if(mode & H_DEBLOCK) |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3652 { |
164 | 3653 if( isVertDC(tempBlock1, 16) ) |
128
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3654 { |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3655 if(isVertMinMaxOk(tempBlock1, 16, QP)) |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3656 doVertLowPass(tempBlock1, 16, QP); |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3657 } |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3658 else |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3659 doVertDefFilter(tempBlock1, 16, QP); |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3660 } |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3661 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3662 transpose2(dstBlock-4, dstStride, tempBlock1 + 4*16); |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3663 |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3664 #else |
115
4514b8e7f0f1
more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents:
113
diff
changeset
|
3665 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
|
3666 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
|
3667 else if(mode & H_DEBLOCK) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3668 { |
128
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3669 if( isHorizDC(dstBlock-4, stride)) |
96 | 3670 { |
128
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3671 if(isHorizMinMaxOk(dstBlock-4, stride, QP)) |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3672 doHorizLowPass(dstBlock-4, stride, QP); |
96 | 3673 } |
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
|
3674 else |
128
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3675 doHorizDefFilter(dstBlock-4, stride, QP); |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3676 } |
128
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3677 #endif |
107 | 3678 #ifdef MORE_TIMING |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3679 T1= rdtsc(); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3680 horizTime+= T1-T0; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3681 T0=T1; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3682 #endif |
130 | 3683 if(mode & DERING) |
3684 { | |
3685 //FIXME filter first line | |
3686 if(y>0) dering(dstBlock - stride - 8, stride, QP); | |
3687 } | |
156 | 3688 |
3689 if(mode & TEMP_NOISE_FILTER) | |
3690 { | |
3691 tempNoiseReducer(dstBlock-8, stride, | |
3692 tempBlured[isColor] + y*dstStride + x, | |
158 | 3693 tempBluredPast[isColor] + (y>>3)*256 + (x>>3), |
156 | 3694 ppMode->maxTmpNoise); |
3695 } | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3696 } |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3697 |
113 | 3698 #ifdef PP_FUNNY_STRIDE |
112 | 3699 /* did we use a tmp-block buffer */ |
3700 if(x+7 >= width) | |
3701 { | |
3702 int i; | |
3703 dstBlock= dstBlockPtrBackup; | |
3704 srcBlock= srcBlockPtrBackup; | |
3705 | |
3706 for(i=0;i<BLOCK_SIZE*2; i++) | |
3707 { | |
3708 memcpy(dstBlock+i*dstStride, tempDstBlock+i*dstStride, width-x); | |
3709 } | |
3710 } | |
113 | 3711 #endif |
112 | 3712 |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3713 dstBlock+=8; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3714 srcBlock+=8; |
128
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3715 |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
3716 #ifdef HAVE_MMX |
128
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3717 tmpXchg= tempBlock1; |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3718 tempBlock1= tempBlock2; |
e5266b8e79be
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
126
diff
changeset
|
3719 tempBlock2 = tmpXchg; |
129
be35346e27c1
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
128
diff
changeset
|
3720 #endif |
111 | 3721 } |
3722 | |
156 | 3723 if(mode & DERING) |
3724 { | |
3725 if(y > 0) dering(dstBlock - dstStride - 8, dstStride, QP); | |
3726 } | |
3727 | |
3728 if((mode & TEMP_NOISE_FILTER)) | |
3729 { | |
3730 tempNoiseReducer(dstBlock-8, dstStride, | |
3731 tempBlured[isColor] + y*dstStride + x, | |
158 | 3732 tempBluredPast[isColor] + (y>>3)*256 + (x>>3), |
156 | 3733 ppMode->maxTmpNoise); |
3734 } | |
3735 | |
142 | 3736 /* did we use a tmp buffer for the last lines*/ |
112 | 3737 if(y+15 >= height) |
111 | 3738 { |
3739 uint8_t *dstBlock= &(dst[y*dstStride]); | |
130 | 3740 memcpy(dstBlock, tempDst + dstStride, dstStride*(height-y) ); |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3741 } |
163 | 3742 /* |
3743 for(x=0; x<width; x+=32) | |
3744 { | |
164 | 3745 volatile int i; |
163 | 3746 i+= + dstBlock[x + 7*dstStride] + dstBlock[x + 8*dstStride] |
3747 + dstBlock[x + 9*dstStride] + dstBlock[x +10*dstStride] | |
164 | 3748 + dstBlock[x +11*dstStride] + dstBlock[x +12*dstStride]; |
3749 // + dstBlock[x +13*dstStride] | |
3750 // + dstBlock[x +14*dstStride] + dstBlock[x +15*dstStride]; | |
3751 }*/ | |
3752 } | |
96 | 3753 #ifdef HAVE_3DNOW |
3754 asm volatile("femms"); | |
3755 #elif defined (HAVE_MMX) | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3756 asm volatile("emms"); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3757 #endif |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3758 |
107 | 3759 #ifdef TIMING |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3760 // 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
|
3761 sumTime= rdtsc() - sumTime; |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3762 if(!isColor) |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3763 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
|
3764 (int)(memcpyTime/1000), (int)(vertTime/1000), (int)(horizTime/1000), |
e57b1d38d71f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
96
diff
changeset
|
3765 (int)(sumTime/1000), (int)((sumTime-memcpyTime-vertTime-horizTime)/1000) |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3766 , black, white); |
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3767 #endif |
163 | 3768 #ifdef DEBUG_BRIGHTNESS |
3769 if(!isColor) | |
3770 { | |
3771 int max=1; | |
3772 int i; | |
3773 for(i=0; i<256; i++) | |
3774 if(yHistogram[i] > max) max=yHistogram[i]; | |
3775 | |
3776 for(i=1; i<256; i++) | |
3777 { | |
3778 int x; | |
3779 int start=yHistogram[i-1]/(max/256+1); | |
3780 int end=yHistogram[i]/(max/256+1); | |
3781 int inc= end > start ? 1 : -1; | |
3782 for(x=start; x!=end+inc; x+=inc) | |
3783 dst[ i*dstStride + x]+=128; | |
3784 } | |
3785 | |
3786 for(i=0; i<100; i+=2) | |
3787 { | |
3788 dst[ (white)*dstStride + i]+=128; | |
3789 dst[ (black)*dstStride + i]+=128; | |
3790 } | |
3791 | |
3792 } | |
3793 #endif | |
3794 | |
95
8bce253b537c
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3795 } |