Mercurial > mplayer.hg
annotate postproc/postprocess_template.c @ 3013:71384f064a3e
faster mmx2 / 3dnow deblocking filter
brightness_debug (draws luminance histogram & autodetected white/black level)
author | michael |
---|---|
date | Mon, 19 Nov 2001 22:20:30 +0000 |
parents | 6885bf566271 |
children | 86e1a0f4f0bc |
rev | line source |
---|---|
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2 Copyright (C) 2001 Michael Niedermayer (michaelni@gmx.at) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4 This program is free software; you can redistribute it and/or modify |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
5 it under the terms of the GNU General Public License as published by |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
6 the Free Software Foundation; either version 2 of the License, or |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
7 (at your option) any later version. |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
8 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
9 This program is distributed in the hope that it will be useful, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
10 but WITHOUT ANY WARRANTY; without even the implied warranty of |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
12 GNU General Public License for more details. |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
13 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
14 You should have received a copy of the GNU General Public License |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
15 along with this program; if not, write to the Free Software |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
16 Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
17 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
18 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
19 /* |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
20 C MMX MMX2 3DNow |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
21 isVertDC Ec Ec |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
22 isVertMinMaxOk Ec Ec |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
23 doVertLowPass E e e |
3013 | 24 doVertDefFilter Ec Ec e e |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
25 isHorizDC Ec Ec |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
26 isHorizMinMaxOk a E |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
27 doHorizLowPass E e e |
3013 | 28 doHorizDefFilter Ec Ec e e |
2477 | 29 deRing E e e* |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
30 Vertical RKAlgo1 E a a |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
31 Horizontal RKAlgo1 a a |
2860 | 32 Vertical X1# a E E |
33 Horizontal X1# a E E | |
2246 | 34 LinIpolDeinterlace e E E* |
35 CubicIpolDeinterlace a e e* | |
36 LinBlendDeinterlace e E E* | |
2860 | 37 MedianDeinterlace# Ec Ec |
2895
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
38 TempDeNoiser# E e e |
2860 | 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 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
42 E = Exact implementation |
2246 | 43 e = allmost exact implementation (slightly different rounding,...) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
44 a = alternative / approximate impl |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
45 c = checked against the other implementations (-vo md5) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
46 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
47 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
48 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
49 TODO: |
2179 | 50 verify that everything workes as it should (how?) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
51 reduce the time wasted on the mem transfer |
2159 | 52 implement everything in C at least (done at the moment but ...) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
53 unroll stuff if instructions depend too much on the prior one |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
54 we use 8x8 blocks for the horizontal filters, opendivx seems to use 8x4? |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
55 move YScale thing to the end instead of fixing QP |
2159 | 56 write a faster and higher quality deblocking filter :) |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
57 make the mainloop more flexible (variable number of blocks at once |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
58 (the if/else stuff per block is slowing things down) |
2179 | 59 compare the quality & speed of all filters |
60 split this huge file | |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
61 border remover |
2570 | 62 optimize c versions |
2860 | 63 try to unroll inner for(x=0 ... loop to avoid these damn if(x ... checks |
2895
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
64 smart blur |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
65 ... |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
66 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
67 |
2221 | 68 //Changelog: use the CVS log |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
69 |
2775 | 70 #include "../config.h" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
71 #include <inttypes.h> |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
72 #include <stdio.h> |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
73 #include <stdlib.h> |
2389 | 74 #include <string.h> |
2476 | 75 #ifdef HAVE_MALLOC_H |
76 #include <malloc.h> | |
77 #endif | |
2159 | 78 //#undef HAVE_MMX2 |
79 //#define HAVE_3DNOW | |
80 //#undef HAVE_MMX | |
3013 | 81 //#define DEBUG_BRIGHTNESS |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
82 #include "postprocess.h" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
83 |
2189 | 84 #define MIN(a,b) ((a) > (b) ? (b) : (a)) |
85 #define MAX(a,b) ((a) < (b) ? (b) : (a)) | |
86 #define ABS(a) ((a) > 0 ? (a) : (-(a))) | |
87 #define SIGN(a) ((a) > 0 ? 1 : -1) | |
88 | |
89 #ifdef HAVE_MMX2 | |
90 #define PAVGB(a,b) "pavgb " #a ", " #b " \n\t" | |
91 #elif defined (HAVE_3DNOW) | |
92 #define PAVGB(a,b) "pavgusb " #a ", " #b " \n\t" | |
93 #endif | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
94 |
2477 | 95 #ifdef HAVE_MMX2 |
96 #define PMINUB(a,b,t) "pminub " #a ", " #b " \n\t" | |
97 #elif defined (HAVE_MMX) | |
98 #define PMINUB(b,a,t) \ | |
99 "movq " #a ", " #t " \n\t"\ | |
100 "psubusb " #b ", " #t " \n\t"\ | |
101 "psubb " #t ", " #a " \n\t" | |
102 #endif | |
103 | |
104 #ifdef HAVE_MMX2 | |
105 #define PMAXUB(a,b) "pmaxub " #a ", " #b " \n\t" | |
106 #elif defined (HAVE_MMX) | |
107 #define PMAXUB(a,b) \ | |
108 "psubusb " #a ", " #b " \n\t"\ | |
109 "paddb " #a ", " #b " \n\t" | |
110 #endif | |
111 | |
112 | |
2389 | 113 #define GET_MODE_BUFFER_SIZE 500 |
114 #define OPTIONS_ARRAY_SIZE 10 | |
115 | |
2742
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
116 #ifdef HAVE_MMX |
2750
9ef09e232505
gcc does optimize writes to non volatile variables away if it didnt know that they were read in between
michael
parents:
2742
diff
changeset
|
117 static volatile uint64_t __attribute__((aligned(8))) packedYOffset= 0x0000000000000000LL; |
9ef09e232505
gcc does optimize writes to non volatile variables away if it didnt know that they were read in between
michael
parents:
2742
diff
changeset
|
118 static volatile uint64_t __attribute__((aligned(8))) packedYScale= 0x0100010001000100LL; |
2742
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
119 static uint64_t __attribute__((aligned(8))) w05= 0x0005000500050005LL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
120 static uint64_t __attribute__((aligned(8))) w20= 0x0020002000200020LL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
121 static uint64_t __attribute__((aligned(8))) w1400= 0x1400140014001400LL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
122 static uint64_t __attribute__((aligned(8))) bm00000001= 0x00000000000000FFLL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
123 static uint64_t __attribute__((aligned(8))) bm00010000= 0x000000FF00000000LL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
124 static uint64_t __attribute__((aligned(8))) bm00001000= 0x00000000FF000000LL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
125 static uint64_t __attribute__((aligned(8))) bm10000000= 0xFF00000000000000LL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
126 static uint64_t __attribute__((aligned(8))) bm10000001= 0xFF000000000000FFLL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
127 static uint64_t __attribute__((aligned(8))) bm11000011= 0xFFFF00000000FFFFLL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
128 static uint64_t __attribute__((aligned(8))) bm00000011= 0x000000000000FFFFLL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
129 static uint64_t __attribute__((aligned(8))) bm11111110= 0xFFFFFFFFFFFFFF00LL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
130 static uint64_t __attribute__((aligned(8))) bm11000000= 0xFFFF000000000000LL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
131 static uint64_t __attribute__((aligned(8))) bm00011000= 0x000000FFFF000000LL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
132 static uint64_t __attribute__((aligned(8))) bm00110011= 0x0000FFFF0000FFFFLL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
133 static uint64_t __attribute__((aligned(8))) bm11001100= 0xFFFF0000FFFF0000LL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
134 static uint64_t __attribute__((aligned(8))) b00= 0x0000000000000000LL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
135 static uint64_t __attribute__((aligned(8))) b01= 0x0101010101010101LL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
136 static uint64_t __attribute__((aligned(8))) b02= 0x0202020202020202LL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
137 static uint64_t __attribute__((aligned(8))) b0F= 0x0F0F0F0F0F0F0F0FLL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
138 static uint64_t __attribute__((aligned(8))) b04= 0x0404040404040404LL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
139 static uint64_t __attribute__((aligned(8))) b08= 0x0808080808080808LL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
140 static uint64_t __attribute__((aligned(8))) bFF= 0xFFFFFFFFFFFFFFFFLL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
141 static uint64_t __attribute__((aligned(8))) b20= 0x2020202020202020LL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
142 static uint64_t __attribute__((aligned(8))) b80= 0x8080808080808080LL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
143 static uint64_t __attribute__((aligned(8))) b7E= 0x7E7E7E7E7E7E7E7ELL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
144 static uint64_t __attribute__((aligned(8))) b7C= 0x7C7C7C7C7C7C7C7CLL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
145 static uint64_t __attribute__((aligned(8))) b3F= 0x3F3F3F3F3F3F3F3FLL; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
146 static uint64_t __attribute__((aligned(8))) temp0=0; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
147 static uint64_t __attribute__((aligned(8))) temp1=0; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
148 static uint64_t __attribute__((aligned(8))) temp2=0; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
149 static uint64_t __attribute__((aligned(8))) temp3=0; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
150 static uint64_t __attribute__((aligned(8))) temp4=0; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
151 static uint64_t __attribute__((aligned(8))) temp5=0; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
152 static uint64_t __attribute__((aligned(8))) pQPb=0; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
153 static uint64_t __attribute__((aligned(8))) pQPb2=0; |
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
154 static uint8_t __attribute__((aligned(8))) tempBlocks[8*16*2]; //used for the horizontal code |
2899 | 155 static uint32_t __attribute__((aligned(4))) maxTmpNoise[4]; |
2742
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
156 #else |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
157 static uint64_t packedYOffset= 0x0000000000000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
158 static uint64_t packedYScale= 0x0100010001000100LL; |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
159 static uint8_t tempBlocks[8*16*2]; //used for the horizontal code |
2742
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
160 #endif |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
161 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
162 int hFlatnessThreshold= 56 - 16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
163 int vFlatnessThreshold= 56 - 16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
164 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
165 //amount of "black" u r willing to loose to get a brightness corrected picture |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
166 double maxClippedThreshold= 0.01; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
167 |
2389 | 168 int maxAllowedY=234; |
2285 | 169 int minAllowedY=16; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
170 |
2389 | 171 static struct PPFilter filters[]= |
172 { | |
173 {"hb", "hdeblock", 1, 1, 3, H_DEBLOCK}, | |
174 {"vb", "vdeblock", 1, 2, 4, V_DEBLOCK}, | |
175 {"vr", "rkvdeblock", 1, 2, 4, H_RK1_FILTER}, | |
176 {"h1", "x1hdeblock", 1, 1, 3, H_X1_FILTER}, | |
177 {"v1", "x1vdeblock", 1, 2, 4, V_X1_FILTER}, | |
178 {"dr", "dering", 1, 5, 6, DERING}, | |
179 {"al", "autolevels", 0, 1, 2, LEVEL_FIX}, | |
180 {"lb", "linblenddeint", 0, 1, 6, LINEAR_BLEND_DEINT_FILTER}, | |
181 {"li", "linipoldeint", 0, 1, 6, LINEAR_IPOL_DEINT_FILTER}, | |
182 {"ci", "cubicipoldeint", 0, 1, 6, CUBIC_IPOL_DEINT_FILTER}, | |
183 {"md", "mediandeint", 0, 1, 6, MEDIAN_DEINT_FILTER}, | |
2860 | 184 {"tn", "tmpnoise", 1, 7, 8, TEMP_NOISE_FILTER}, |
2389 | 185 {NULL, NULL,0,0,0,0} //End Marker |
186 }; | |
187 | |
188 static char *replaceTable[]= | |
189 { | |
2860 | 190 "default", "hdeblock:a,vdeblock:a,dering:a,autolevels,tmpnoise:a:150:200:400", |
191 "de", "hdeblock:a,vdeblock:a,dering:a,autolevels,tmpnoise:a:150:200:400", | |
192 "fast", "x1hdeblock:a,x1vdeblock:a,dering:a,autolevels,tmpnoise:a:150:200:400", | |
193 "fa", "x1hdeblock:a,x1vdeblock:a,dering:a,autolevels,tmpnoise:a:150:200:400", | |
2389 | 194 NULL //End Marker |
195 }; | |
196 | |
2742
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
197 #ifdef HAVE_MMX |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
198 static inline void unusedVariableWarningFixer() |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
199 { |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
200 if( |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
201 packedYOffset + packedYScale + w05 + w20 + w1400 + bm00000001 + bm00010000 |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
202 + bm00001000 + bm10000000 + bm10000001 + bm11000011 + bm00000011 + bm11111110 |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
203 + bm11000000 + bm00011000 + bm00110011 + bm11001100 + b00 + b01 + b02 + b0F |
2742
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
204 + bFF + b20 + b04+ b08 + pQPb2 + b80 + b7E + b7C + b3F + temp0 + temp1 + temp2 + temp3 + temp4 |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
205 + temp5 + pQPb== 0) b00=0; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
206 } |
2742
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
207 #endif |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
208 |
2221 | 209 #ifdef TIMING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
210 static inline long long rdtsc() |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
211 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
212 long long l; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
213 asm volatile( "rdtsc\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
214 : "=A" (l) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
215 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
216 // printf("%d\n", int(l/1000)); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
217 return l; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
218 } |
2185 | 219 #endif |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
220 |
2185 | 221 #ifdef HAVE_MMX2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
222 static inline void prefetchnta(void *p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
223 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
224 asm volatile( "prefetchnta (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
225 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
226 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
227 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
228 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
229 static inline void prefetcht0(void *p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
230 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
231 asm volatile( "prefetcht0 (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
232 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
233 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
234 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
235 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
236 static inline void prefetcht1(void *p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
237 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
238 asm volatile( "prefetcht1 (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
239 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
240 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
241 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
242 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
243 static inline void prefetcht2(void *p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
244 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
245 asm volatile( "prefetcht2 (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
246 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
247 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
248 } |
2185 | 249 #endif |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
250 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
251 //FIXME? |255-0| = 1 (shouldnt be a problem ...) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
252 /** |
2246 | 253 * Check if the middle 8x8 Block in the given 8x16 block is flat |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
254 */ |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
255 static inline int isVertDC(uint8_t src[], int stride){ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
256 int numEq= 0; |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
257 #ifndef HAVE_MMX |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
258 int y; |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
259 #endif |
2246 | 260 src+= stride*4; // src points to begin of the 8x8 Block |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
261 #ifdef HAVE_MMX |
2413 | 262 asm volatile( |
263 "leal (%1, %2), %%eax \n\t" | |
264 "leal (%%eax, %2, 4), %%ebx \n\t" | |
265 // 0 1 2 3 4 5 6 7 8 9 | |
266 // %1 eax eax+%2 eax+2%2 %1+4%2 ebx ebx+%2 ebx+2%2 %1+8%2 ebx+4%2 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
267 "movq b7E, %%mm7 \n\t" // mm7 = 0x7F |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
268 "movq b7C, %%mm6 \n\t" // mm6 = 0x7D |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
269 "movq (%1), %%mm0 \n\t" |
2413 | 270 "movq (%%eax), %%mm1 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
271 "psubb %%mm1, %%mm0 \n\t" // mm0 = differnece |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
272 "paddb %%mm7, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
273 "pcmpgtb %%mm6, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
274 |
2413 | 275 "movq (%%eax,%2), %%mm2 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
276 "psubb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
277 "paddb %%mm7, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
278 "pcmpgtb %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
279 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
280 |
2413 | 281 "movq (%%eax, %2, 2), %%mm1 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
282 "psubb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
283 "paddb %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
284 "pcmpgtb %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
285 "paddb %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
286 |
2413 | 287 "movq (%1, %2, 4), %%mm2 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
288 "psubb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
289 "paddb %%mm7, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
290 "pcmpgtb %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
291 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
292 |
2413 | 293 "movq (%%ebx), %%mm1 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
294 "psubb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
295 "paddb %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
296 "pcmpgtb %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
297 "paddb %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
298 |
2413 | 299 "movq (%%ebx, %2), %%mm2 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
300 "psubb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
301 "paddb %%mm7, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
302 "pcmpgtb %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
303 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
304 |
2413 | 305 "movq (%%ebx, %2, 2), %%mm1 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
306 "psubb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
307 "paddb %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
308 "pcmpgtb %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
309 "paddb %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
310 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
311 " \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
312 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
313 "psrlw $8, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
314 "paddb %%mm1, %%mm0 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
315 #ifdef HAVE_MMX2 |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
316 "pshufw $0xF9, %%mm0, %%mm1 \n\t" |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
317 "paddb %%mm1, %%mm0 \n\t" |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
318 "pshufw $0xFE, %%mm0, %%mm1 \n\t" |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
319 #else |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
320 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
321 "psrlq $16, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
322 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
323 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
324 "psrlq $32, %%mm0 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
325 #endif |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
326 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
327 "movd %%mm0, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
328 : "=r" (numEq) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
329 : "r" (src), "r" (stride) |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
330 : "%eax", "%ebx" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
331 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
332 |
2413 | 333 numEq= (256 - numEq) &0xFF; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
334 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
335 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
336 for(y=0; y<BLOCK_SIZE-1; y++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
337 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
338 if(((src[0] - src[0+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
339 if(((src[1] - src[1+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
340 if(((src[2] - src[2+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
341 if(((src[3] - src[3+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
342 if(((src[4] - src[4+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
343 if(((src[5] - src[5+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
344 if(((src[6] - src[6+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
345 if(((src[7] - src[7+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
346 src+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
347 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
348 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
349 /* if(abs(numEq - asmEq) > 0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
350 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
351 printf("\nasm:%d c:%d\n", asmEq, numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
352 for(int y=0; y<8; y++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
353 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
354 for(int x=0; x<8; x++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
355 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
356 printf("%d ", temp[x + y*stride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
357 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
358 printf("\n"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
359 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
360 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
361 */ |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
362 // for(int i=0; i<numEq/8; i++) src[i]=255; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
363 return (numEq > vFlatnessThreshold) ? 1 : 0; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
364 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
365 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
366 static inline int isVertMinMaxOk(uint8_t src[], int stride, int QP) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
367 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
368 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
369 int isOk; |
2246 | 370 src+= stride*3; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
371 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
372 // "int $3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
373 "movq (%1, %2), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
374 "movq (%1, %2, 8), %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
375 "movq %%mm0, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
376 "psubusb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
377 "psubusb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
378 "por %%mm1, %%mm0 \n\t" // ABS Diff |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
379 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
380 "movq pQPb, %%mm7 \n\t" // QP,..., QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
381 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
382 "psubusb %%mm7, %%mm0 \n\t" // Diff <= 2QP -> 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
383 "pcmpeqd b00, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
384 "psrlq $16, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
385 "pcmpeqd bFF, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
386 // "movd %%mm0, (%1, %2, 4)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
387 "movd %%mm0, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
388 : "=r" (isOk) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
389 : "r" (src), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
390 ); |
2416 | 391 return isOk; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
392 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
393 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
394 int isOk2= 1; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
395 int x; |
2246 | 396 src+= stride*3; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
397 for(x=0; x<BLOCK_SIZE; x++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
398 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
399 if(abs((int)src[x + stride] - (int)src[x + (stride<<3)]) > 2*QP) isOk2=0; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
400 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
401 /* if(isOk && !isOk2 || !isOk && isOk2) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
402 { |
508468a75be0
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); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
404 for(int y=0; y<9; y++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
405 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
406 for(int x=0; x<8; x++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
407 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
408 printf("%d ", src[x + y*stride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
409 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
410 printf("\n"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
411 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
412 } */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
413 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
414 return isOk2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
415 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
416 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
417 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
418 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
419 /** |
2246 | 420 * Do a vertical low pass filter on the 8x16 block (only write to the 8x8 block in the middle) |
2221 | 421 * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
422 */ |
508468a75be0
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) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
424 { |
2159 | 425 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2246 | 426 src+= stride*3; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
427 asm volatile( //"movv %0 %1 %2\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
428 "movq pQPb, %%mm0 \n\t" // QP,..., QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
429 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
430 "movq (%0), %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
431 "movq (%0, %1), %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
432 "movq %%mm5, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
433 "movq %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
434 "psubusb %%mm6, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
435 "psubusb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
436 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
437 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
438 "pcmpeqb b00, %%mm2 \n\t" // diff <= QP -> FF |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
439 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
440 "pand %%mm2, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
441 "pandn %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
442 "por %%mm2, %%mm6 \n\t"// First Line to Filter |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
443 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
444 "movq (%0, %1, 8), %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
445 "leal (%0, %1, 4), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
446 "leal (%0, %1, 8), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
447 "subl %1, %%ebx \n\t" |
508468a75be0
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 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
449 "movq (%0, %1, 8), %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
450 "movq %%mm5, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
451 "movq %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
452 "psubusb %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
453 "psubusb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
454 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
455 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
456 "pcmpeqb b00, %%mm2 \n\t" // diff <= QP -> FF |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
457 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
458 "pand %%mm2, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
459 "pandn %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
460 "por %%mm2, %%mm7 \n\t" // First Line to Filter |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
461 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
462 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
463 // 1 2 3 4 5 6 7 8 |
508468a75be0
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 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
465 // 6 4 2 2 1 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
466 // 6 4 4 2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
467 // 6 8 2 |
2246 | 468 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
469 "movq (%0, %1), %%mm0 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
470 "movq %%mm0, %%mm1 \n\t" // 1 |
2159 | 471 PAVGB(%%mm6, %%mm0) //1 1 /2 |
472 PAVGB(%%mm6, %%mm0) //3 1 /4 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
473 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
474 "movq (%0, %1, 4), %%mm2 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
475 "movq %%mm2, %%mm5 \n\t" // 1 |
2159 | 476 PAVGB((%%eax), %%mm2) // 11 /2 |
477 PAVGB((%0, %1, 2), %%mm2) // 211 /4 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
478 "movq %%mm2, %%mm3 \n\t" // 211 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
479 "movq (%0), %%mm4 \n\t" // 1 |
2159 | 480 PAVGB(%%mm4, %%mm3) // 4 211 /8 |
481 PAVGB(%%mm0, %%mm3) //642211 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
482 "movq %%mm3, (%0) \n\t" // X |
508468a75be0
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 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
484 "movq %%mm1, %%mm0 \n\t" // 1 |
2159 | 485 PAVGB(%%mm6, %%mm0) //1 1 /2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
486 "movq %%mm4, %%mm3 \n\t" // 1 |
2159 | 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 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
492 "movq %%mm3, (%0,%1) \n\t" // X |
508468a75be0
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 |
2159 | 494 PAVGB(%%mm4, %%mm6) //11 /2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
495 "movq (%%ebx), %%mm0 \n\t" // 1 |
2159 | 496 PAVGB((%%eax, %1, 2), %%mm0) // 11/2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
497 "movq %%mm0, %%mm3 \n\t" // 11/2 |
2159 | 498 PAVGB(%%mm1, %%mm0) // 2 11/4 |
499 PAVGB(%%mm6, %%mm0) //222 11/8 | |
500 PAVGB(%%mm2, %%mm0) //22242211/16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
501 "movq (%0, %1, 2), %%mm2 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
502 "movq %%mm0, (%0, %1, 2) \n\t" // X |
508468a75be0
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 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
504 "movq (%%eax, %1, 4), %%mm0 \n\t" // 1 |
2159 | 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 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
511 "movq (%%eax), %%mm5 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
512 "movq %%mm6, (%%eax) \n\t" // X |
508468a75be0
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 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
514 "movq (%%eax, %1, 4), %%mm6 \n\t" // 1 |
2159 | 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 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
519 "movq (%0, %1, 4), %%mm4 \n\t" // 1 |
2159 | 520 PAVGB(%%mm4, %%mm2) // 112 /4 |
521 PAVGB(%%mm2, %%mm6) // 112242211 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
522 "movq %%mm6, (%0, %1, 4) \n\t" // X |
508468a75be0
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 |
2159 | 524 PAVGB(%%mm7, %%mm1) // 11 2 /4 |
525 PAVGB(%%mm4, %%mm5) // 11 /2 | |
526 PAVGB(%%mm5, %%mm0) // 11 11 /4 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
527 "movq (%%eax, %1, 2), %%mm6 \n\t" // 1 |
2159 | 528 PAVGB(%%mm6, %%mm1) // 11 4 2 /8 |
529 PAVGB(%%mm0, %%mm1) // 11224222 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
530 "movq %%mm1, (%%eax, %1, 2) \n\t" // X |
508468a75be0
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 |
2159 | 532 PAVGB((%%ebx), %%mm2) // 112 4 /8 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
533 "movq (%%eax, %1, 4), %%mm0 \n\t" // 1 |
2159 | 534 PAVGB(%%mm0, %%mm6) // 1 1 /2 |
535 PAVGB(%%mm7, %%mm6) // 1 12 /4 | |
536 PAVGB(%%mm2, %%mm6) // 1122424 /4 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
537 "movq %%mm6, (%%ebx) \n\t" // X |
508468a75be0
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 |
2159 | 539 PAVGB(%%mm7, %%mm5) // 11 2 /4 |
540 PAVGB(%%mm7, %%mm5) // 11 6 /8 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
541 |
2159 | 542 PAVGB(%%mm3, %%mm0) // 112 /4 |
543 PAVGB(%%mm0, %%mm5) // 112246 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
544 "movq %%mm5, (%%eax, %1, 4) \n\t" // X |
2570 | 545 "subl %1, %0 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
546 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
547 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
548 : "r" (src), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
549 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
550 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
551 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
552 const int l1= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
553 const int l2= stride + l1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
554 const int l3= stride + l2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
555 const int l4= stride + l3; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
556 const int l5= stride + l4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
557 const int l6= stride + l5; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
558 const int l7= stride + l6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
559 const int l8= stride + l7; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
560 const int l9= stride + l8; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
561 int x; |
2246 | 562 src+= stride*3; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
563 for(x=0; x<BLOCK_SIZE; x++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
564 { |
508468a75be0
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]; |
508468a75be0
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]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
567 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
568 int sums[9]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
569 sums[0] = first + src[l1]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
570 sums[1] = src[l1] + src[l2]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
571 sums[2] = src[l2] + src[l3]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
572 sums[3] = src[l3] + src[l4]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
573 sums[4] = src[l4] + src[l5]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
574 sums[5] = src[l5] + src[l6]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
575 sums[6] = src[l6] + src[l7]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
576 sums[7] = src[l7] + src[l8]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
577 sums[8] = src[l8] + last; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
578 |
508468a75be0
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; |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
580 src[l2]= ((src[l2]<<2) + ((first + sums[0] + sums[3])<<1) + sums[5] + 8)>>4; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
581 src[l3]= ((src[l3]<<2) + ((first + sums[1] + sums[4])<<1) + sums[6] + 8)>>4; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
582 src[l4]= ((src[l4]<<2) + ((sums[2] + sums[5])<<1) + sums[0] + sums[7] + 8)>>4; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
583 src[l5]= ((src[l5]<<2) + ((sums[3] + sums[6])<<1) + sums[1] + sums[8] + 8)>>4; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
584 src[l6]= ((src[l6]<<2) + ((last + sums[7] + sums[4])<<1) + sums[2] + 8)>>4; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
585 src[l7]= (((last + src[l7])<<2) + ((src[l8] + sums[5])<<1) + sums[3] + 8)>>4; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
586 src[l8]= ((sums[8]<<2) + ((last + sums[6])<<1) + sums[4] + 8)>>4; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
587 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
588 src++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
589 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
590 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
591 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
592 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
593 |
2159 | 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 */ | |
2179 | 605 static inline void vertRK1Filter(uint8_t *src, int stride, int QP) |
2159 | 606 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
607 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2246 | 608 src+= stride*3; |
2159 | 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 | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
628 PAVGB(%%mm3, %%mm5) |
2159 | 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; | |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
676 // const int l7= stride + l6; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
677 // const int l8= stride + l7; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
678 // const int l9= stride + l8; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
679 int x; |
2586 | 680 const int QP15= QP + (QP>>2); |
2246 | 681 src+= stride*3; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
682 for(x=0; x<BLOCK_SIZE; x++) |
2159 | 683 { |
2586 | 684 const int v = (src[x+l5] - src[x+l4]); |
685 if(ABS(v) < QP15) | |
2159 | 686 { |
2586 | 687 src[x+l3] +=v>>3; |
688 src[x+l4] +=v>>1; | |
689 src[x+l5] -=v>>1; | |
690 src[x+l6] -=v>>3; | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
691 |
2159 | 692 } |
693 } | |
694 | |
695 #endif | |
696 } | |
697 | |
698 /** | |
699 * Experimental Filter 1 | |
2179 | 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 | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
702 * can only smooth blocks at the expected locations (it cant smooth them if they did move) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
703 * MMX2 version does correct clipping C version doesnt |
2159 | 704 */ |
705 static inline void vertX1Filter(uint8_t *src, int stride, int QP) | |
706 { | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
707 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2246 | 708 src+= stride*3; |
709 | |
2159 | 710 asm volatile( |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
711 "pxor %%mm7, %%mm7 \n\t" // 0 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
712 // "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
713 "leal (%0, %1), %%eax \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
714 "leal (%%eax, %1, 4), %%ebx \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
715 // 0 1 2 3 4 5 6 7 8 9 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
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 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
717 "movq (%%eax, %1, 2), %%mm0 \n\t" // line 3 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
718 "movq (%0, %1, 4), %%mm1 \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
719 "movq %%mm1, %%mm2 \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
720 "psubusb %%mm0, %%mm1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
721 "psubusb %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
722 "por %%mm1, %%mm0 \n\t" // |l2 - l3| |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
723 "movq (%%ebx), %%mm3 \n\t" // line 5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
724 "movq (%%ebx, %1), %%mm4 \n\t" // line 6 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
725 "movq %%mm3, %%mm5 \n\t" // line 5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
726 "psubusb %%mm4, %%mm3 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
727 "psubusb %%mm5, %%mm4 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
728 "por %%mm4, %%mm3 \n\t" // |l5 - l6| |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
729 PAVGB(%%mm3, %%mm0) // (|l2 - l3| + |l5 - l6|)/2 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
730 "movq %%mm2, %%mm1 \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
731 "psubusb %%mm5, %%mm2 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
732 "movq %%mm2, %%mm4 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
733 "pcmpeqb %%mm7, %%mm2 \n\t" // (l4 - l5) <= 0 ? -1 : 0 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
734 "psubusb %%mm1, %%mm5 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
735 "por %%mm5, %%mm4 \n\t" // |l4 - l5| |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
736 "psubusb %%mm0, %%mm4 \n\t" //d = MAX(0, |l4-l5| - (|l2-l3| + |l5-l6|)/2) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
737 "movq %%mm4, %%mm3 \n\t" // d |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
738 "psubusb pQPb, %%mm4 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
739 "pcmpeqb %%mm7, %%mm4 \n\t" // d <= QP ? -1 : 0 |
2179 | 740 "psubusb b01, %%mm3 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
741 "pand %%mm4, %%mm3 \n\t" // d <= QP ? d : 0 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
742 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
743 PAVGB(%%mm7, %%mm3) // d/2 |
2179 | 744 "movq %%mm3, %%mm1 \n\t" // d/2 |
745 PAVGB(%%mm7, %%mm3) // d/4 | |
746 PAVGB(%%mm1, %%mm3) // 3*d/8 | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
747 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
748 "movq (%0, %1, 4), %%mm0 \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
749 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
750 "psubusb %%mm3, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
751 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
752 "movq %%mm0, (%0, %1, 4) \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
753 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
754 "movq (%%ebx), %%mm0 \n\t" // line 5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
755 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
756 "paddusb %%mm3, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
757 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
758 "movq %%mm0, (%%ebx) \n\t" // line 5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
759 |
2179 | 760 PAVGB(%%mm7, %%mm1) // d/4 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
761 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
762 "movq (%%eax, %1, 2), %%mm0 \n\t" // line 3 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
763 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4 |
2179 | 764 "psubusb %%mm1, %%mm0 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
765 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
766 "movq %%mm0, (%%eax, %1, 2) \n\t" // line 3 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
767 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
768 "movq (%%ebx, %1), %%mm0 \n\t" // line 6 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
769 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5 |
2179 | 770 "paddusb %%mm1, %%mm0 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
771 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
772 "movq %%mm0, (%%ebx, %1) \n\t" // line 6 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
773 |
2179 | 774 PAVGB(%%mm7, %%mm1) // d/8 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
775 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
776 "movq (%%eax, %1), %%mm0 \n\t" // line 2 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
777 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l2-1 : l2 |
2179 | 778 "psubusb %%mm1, %%mm0 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
779 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
780 "movq %%mm0, (%%eax, %1) \n\t" // line 2 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
781 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
782 "movq (%%ebx, %1, 2), %%mm0 \n\t" // line 7 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
783 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l7-1 : l7 |
2179 | 784 "paddusb %%mm1, %%mm0 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
785 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
786 "movq %%mm0, (%%ebx, %1, 2) \n\t" // line 7 |
2159 | 787 |
788 : | |
789 : "r" (src), "r" (stride) | |
790 : "%eax", "%ebx" | |
791 ); | |
792 #else | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
793 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
794 const int l1= stride; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
795 const int l2= stride + l1; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
796 const int l3= stride + l2; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
797 const int l4= stride + l3; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
798 const int l5= stride + l4; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
799 const int l6= stride + l5; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
800 const int l7= stride + l6; |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
801 // const int l8= stride + l7; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
802 // const int l9= stride + l8; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
803 int x; |
2246 | 804 |
805 src+= stride*3; | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
806 for(x=0; x<BLOCK_SIZE; x++) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
807 { |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
808 int a= src[l3] - src[l4]; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
809 int b= src[l4] - src[l5]; |
2179 | 810 int c= src[l5] - src[l6]; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
811 |
2586 | 812 int d= ABS(b) - ((ABS(a) + ABS(c))>>1); |
813 d= MAX(d, 0); | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
814 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
815 if(d < QP) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
816 { |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
817 int v = d * SIGN(-b); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
818 |
2586 | 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; | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
825 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
826 } |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
827 src++; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
828 } |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
829 /* |
2159 | 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 } | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
857 */ |
2159 | 858 #endif |
859 } | |
860 | |
2179 | 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 static uint64_t *lut= NULL; | |
873 if(lut==NULL) | |
874 { | |
875 int i; | |
876 lut= (uint64_t*)memalign(8, 256*8); | |
877 for(i=0; i<256; i++) | |
878 { | |
879 int v= i < 128 ? 2*i : 2*(i-256); | |
880 /* | |
881 //Simulate 112242211 9-Tap filter | |
882 uint64_t a= (v/16) & 0xFF; | |
883 uint64_t b= (v/8) & 0xFF; | |
884 uint64_t c= (v/4) & 0xFF; | |
885 uint64_t d= (3*v/8) & 0xFF; | |
886 */ | |
887 //Simulate piecewise linear interpolation | |
888 uint64_t a= (v/16) & 0xFF; | |
889 uint64_t b= (v*3/16) & 0xFF; | |
890 uint64_t c= (v*5/16) & 0xFF; | |
891 uint64_t d= (7*v/16) & 0xFF; | |
892 uint64_t A= (0x100 - a)&0xFF; | |
893 uint64_t B= (0x100 - b)&0xFF; | |
894 uint64_t C= (0x100 - c)&0xFF; | |
895 uint64_t D= (0x100 - c)&0xFF; | |
896 | |
897 lut[i] = (a<<56) | (b<<48) | (c<<40) | (d<<32) | | |
898 (D<<24) | (C<<16) | (B<<8) | (A); | |
899 //lut[i] = (v<<32) | (v<<24); | |
900 } | |
901 } | |
902 | |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
903 #if 0 |
2179 | 904 asm volatile( |
905 "pxor %%mm7, %%mm7 \n\t" // 0 | |
906 // "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE | |
907 "leal (%0, %1), %%eax \n\t" | |
908 "leal (%%eax, %1, 4), %%ebx \n\t" | |
909 | |
910 "movq b80, %%mm6 \n\t" | |
2180 | 911 "movd pQPb, %%mm5 \n\t" // QP |
2179 | 912 "movq %%mm5, %%mm4 \n\t" |
913 "paddusb %%mm5, %%mm5 \n\t" // 2QP | |
914 "paddusb %%mm5, %%mm4 \n\t" // 3QP | |
915 "pxor %%mm5, %%mm5 \n\t" // 0 | |
916 "psubb %%mm4, %%mm5 \n\t" // -3QP | |
917 "por bm11111110, %%mm5 \n\t" // ...,FF,FF,-3QP | |
918 "psllq $24, %%mm5 \n\t" | |
919 | |
920 // 0 1 2 3 4 5 6 7 8 9 | |
921 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 | |
922 | |
923 #define HX1old(a) \ | |
924 "movd " #a ", %%mm0 \n\t"\ | |
925 "movd 4" #a ", %%mm1 \n\t"\ | |
926 "punpckldq %%mm1, %%mm0 \n\t"\ | |
927 "movq %%mm0, %%mm1 \n\t"\ | |
928 "movq %%mm0, %%mm2 \n\t"\ | |
929 "psrlq $8, %%mm1 \n\t"\ | |
930 "psubusb %%mm1, %%mm2 \n\t"\ | |
931 "psubusb %%mm0, %%mm1 \n\t"\ | |
932 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ | |
933 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
934 "pshufw $0x00, %%mm1, %%mm3 \n\t" /* p´5 = |p1 - p2| */\ | |
935 PAVGB(%%mm1, %%mm3) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\ | |
936 "psrlq $16, %%mm3 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\ | |
937 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\ | |
938 "paddb %%mm5, %%mm1 \n\t"\ | |
939 "psubusb %%mm5, %%mm1 \n\t"\ | |
940 PAVGB(%%mm7, %%mm1)\ | |
941 "pxor %%mm2, %%mm1 \n\t"\ | |
942 "psubb %%mm2, %%mm1 \n\t"\ | |
943 "psrlq $24, %%mm1 \n\t"\ | |
944 "movd %%mm1, %%ecx \n\t"\ | |
945 "paddb %%mm6, %%mm0 \n\t"\ | |
946 "paddsb (%3, %%ecx, 8), %%mm0 \n\t"\ | |
947 "paddb %%mm6, %%mm0 \n\t"\ | |
948 "movq %%mm0, " #a " \n\t"\ | |
949 | |
950 /* | |
951 HX1old((%0)) | |
952 HX1old((%%eax)) | |
953 HX1old((%%eax, %1)) | |
954 HX1old((%%eax, %1, 2)) | |
955 HX1old((%0, %1, 4)) | |
956 HX1old((%%ebx)) | |
957 HX1old((%%ebx, %1)) | |
958 HX1old((%%ebx, %1, 2)) | |
959 */ | |
960 | |
961 //FIXME add some comments, its unreadable ... | |
962 #define HX1b(a, c, b, d) \ | |
963 "movd " #a ", %%mm0 \n\t"\ | |
964 "movd 4" #a ", %%mm1 \n\t"\ | |
965 "punpckldq %%mm1, %%mm0 \n\t"\ | |
966 "movd " #b ", %%mm4 \n\t"\ | |
967 "movq %%mm0, %%mm1 \n\t"\ | |
968 "movq %%mm0, %%mm2 \n\t"\ | |
969 "psrlq $8, %%mm1 \n\t"\ | |
970 "movd 4" #b ", %%mm3 \n\t"\ | |
971 "psubusb %%mm1, %%mm2 \n\t"\ | |
972 "psubusb %%mm0, %%mm1 \n\t"\ | |
973 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ | |
974 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
975 "punpckldq %%mm3, %%mm4 \n\t"\ | |
976 "movq %%mm1, %%mm3 \n\t"\ | |
977 "psllq $32, %%mm3 \n\t" /* p´5 = |p1 - p2| */\ | |
978 PAVGB(%%mm1, %%mm3) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\ | |
979 "paddb %%mm6, %%mm0 \n\t"\ | |
980 "psrlq $16, %%mm3 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\ | |
981 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\ | |
982 "movq %%mm4, %%mm3 \n\t"\ | |
983 "paddb %%mm5, %%mm1 \n\t"\ | |
984 "psubusb %%mm5, %%mm1 \n\t"\ | |
985 "psrlq $8, %%mm3 \n\t"\ | |
986 PAVGB(%%mm7, %%mm1)\ | |
987 "pxor %%mm2, %%mm1 \n\t"\ | |
988 "psubb %%mm2, %%mm1 \n\t"\ | |
989 "movq %%mm4, %%mm2 \n\t"\ | |
990 "psrlq $24, %%mm1 \n\t"\ | |
991 "psubusb %%mm3, %%mm2 \n\t"\ | |
992 "movd %%mm1, %%ecx \n\t"\ | |
993 "psubusb %%mm4, %%mm3 \n\t"\ | |
2180 | 994 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\ |
2179 | 995 "por %%mm2, %%mm3 \n\t" /* p´x = |px - p(x+1)| */\ |
996 "paddb %%mm6, %%mm0 \n\t"\ | |
997 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
998 "movq %%mm3, %%mm1 \n\t"\ | |
999 "psllq $32, %%mm1 \n\t" /* p´5 = |p1 - p2| */\ | |
1000 "movq %%mm0, " #a " \n\t"\ | |
1001 PAVGB(%%mm3, %%mm1) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\ | |
1002 "paddb %%mm6, %%mm4 \n\t"\ | |
1003 "psrlq $16, %%mm1 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\ | |
1004 "psubusb %%mm1, %%mm3 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\ | |
1005 "paddb %%mm5, %%mm3 \n\t"\ | |
1006 "psubusb %%mm5, %%mm3 \n\t"\ | |
1007 PAVGB(%%mm7, %%mm3)\ | |
1008 "pxor %%mm2, %%mm3 \n\t"\ | |
1009 "psubb %%mm2, %%mm3 \n\t"\ | |
1010 "psrlq $24, %%mm3 \n\t"\ | |
1011 "movd " #c ", %%mm0 \n\t"\ | |
1012 "movd 4" #c ", %%mm1 \n\t"\ | |
1013 "punpckldq %%mm1, %%mm0 \n\t"\ | |
1014 "paddb %%mm6, %%mm0 \n\t"\ | |
2180 | 1015 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\ |
2179 | 1016 "paddb %%mm6, %%mm0 \n\t"\ |
1017 "movq %%mm0, " #c " \n\t"\ | |
1018 "movd %%mm3, %%ecx \n\t"\ | |
1019 "movd " #d ", %%mm0 \n\t"\ | |
2180 | 1020 "paddsb (%2, %%ecx, 8), %%mm4 \n\t"\ |
2179 | 1021 "movd 4" #d ", %%mm1 \n\t"\ |
1022 "paddb %%mm6, %%mm4 \n\t"\ | |
1023 "punpckldq %%mm1, %%mm0 \n\t"\ | |
1024 "movq %%mm4, " #b " \n\t"\ | |
1025 "paddb %%mm6, %%mm0 \n\t"\ | |
2180 | 1026 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\ |
2179 | 1027 "paddb %%mm6, %%mm0 \n\t"\ |
1028 "movq %%mm0, " #d " \n\t"\ | |
1029 | |
1030 HX1b((%0),(%%eax),(%%eax, %1),(%%eax, %1, 2)) | |
1031 HX1b((%0, %1, 4),(%%ebx),(%%ebx, %1),(%%ebx, %1, 2)) | |
1032 | |
1033 | |
1034 : | |
2180 | 1035 : "r" (src), "r" (stride), "r" (lut) |
2179 | 1036 : "%eax", "%ebx", "%ecx" |
1037 ); | |
1038 #else | |
1039 | |
1040 //FIXME (has little in common with the mmx2 version) | |
1041 for(y=0; y<BLOCK_SIZE; y++) | |
1042 { | |
1043 int a= src[1] - src[2]; | |
1044 int b= src[3] - src[4]; | |
1045 int c= src[5] - src[6]; | |
1046 | |
1047 int d= MAX(ABS(b) - (ABS(a) + ABS(c))/2, 0); | |
1048 | |
1049 if(d < QP) | |
1050 { | |
1051 int v = d * SIGN(-b); | |
1052 | |
1053 src[1] +=v/8; | |
1054 src[2] +=v/4; | |
1055 src[3] +=3*v/8; | |
1056 src[4] -=3*v/8; | |
1057 src[5] -=v/4; | |
1058 src[6] -=v/8; | |
1059 | |
1060 } | |
1061 src+=stride; | |
1062 } | |
1063 #endif | |
1064 } | |
1065 | |
2159 | 1066 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1067 static inline void doVertDefFilter(uint8_t src[], int stride, int QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1068 { |
3013 | 1069 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
1070 /* | |
1071 uint8_t tmp[16]; | |
1072 const int l1= stride; | |
1073 const int l2= stride + l1; | |
1074 const int l3= stride + l2; | |
1075 const int l4= (int)tmp - (int)src - stride*3; | |
1076 const int l5= (int)tmp - (int)src - stride*3 + 8; | |
1077 const int l6= stride*3 + l3; | |
1078 const int l7= stride + l6; | |
1079 const int l8= stride + l7; | |
1080 | |
1081 memcpy(tmp, src+stride*7, 8); | |
1082 memcpy(tmp+8, src+stride*8, 8); | |
1083 */ | |
2246 | 1084 src+= stride*4; |
3013 | 1085 asm volatile( |
1086 | |
1087 #if 0 //sligtly more accurate and slightly slower | |
1088 "pxor %%mm7, %%mm7 \n\t" // 0 | |
1089 "leal (%0, %1), %%eax \n\t" | |
1090 "leal (%%eax, %1, 4), %%ebx \n\t" | |
1091 // 0 1 2 3 4 5 6 7 | |
1092 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ebx+%1 ebx+2%1 | |
1093 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 | |
1094 | |
1095 | |
1096 "movq (%0, %1, 2), %%mm0 \n\t" // l2 | |
1097 "movq (%0), %%mm1 \n\t" // l0 | |
1098 "movq %%mm0, %%mm2 \n\t" // l2 | |
1099 PAVGB(%%mm7, %%mm0) // ~l2/2 | |
1100 PAVGB(%%mm1, %%mm0) // ~(l2 + 2l0)/4 | |
1101 PAVGB(%%mm2, %%mm0) // ~(5l2 + 2l0)/8 | |
1102 | |
1103 "movq (%%eax), %%mm1 \n\t" // l1 | |
1104 "movq (%%eax, %1, 2), %%mm3 \n\t" // l3 | |
1105 "movq %%mm1, %%mm4 \n\t" // l1 | |
1106 PAVGB(%%mm7, %%mm1) // ~l1/2 | |
1107 PAVGB(%%mm3, %%mm1) // ~(l1 + 2l3)/4 | |
1108 PAVGB(%%mm4, %%mm1) // ~(5l1 + 2l3)/8 | |
1109 | |
1110 "movq %%mm0, %%mm4 \n\t" // ~(5l2 + 2l0)/8 | |
1111 "psubusb %%mm1, %%mm0 \n\t" | |
1112 "psubusb %%mm4, %%mm1 \n\t" | |
1113 "por %%mm0, %%mm1 \n\t" // ~|2l0 - 5l1 + 5l2 - 2l3|/8 | |
1114 // mm1= |lenergy|, mm2= l2, mm3= l3, mm7=0 | |
1115 | |
1116 "movq (%0, %1, 4), %%mm0 \n\t" // l4 | |
1117 "movq %%mm0, %%mm4 \n\t" // l4 | |
1118 PAVGB(%%mm7, %%mm0) // ~l4/2 | |
1119 PAVGB(%%mm2, %%mm0) // ~(l4 + 2l2)/4 | |
1120 PAVGB(%%mm4, %%mm0) // ~(5l4 + 2l2)/8 | |
1121 | |
1122 "movq (%%ebx), %%mm2 \n\t" // l5 | |
1123 "movq %%mm3, %%mm5 \n\t" // l3 | |
1124 PAVGB(%%mm7, %%mm3) // ~l3/2 | |
1125 PAVGB(%%mm2, %%mm3) // ~(l3 + 2l5)/4 | |
1126 PAVGB(%%mm5, %%mm3) // ~(5l3 + 2l5)/8 | |
1127 | |
1128 "movq %%mm0, %%mm6 \n\t" // ~(5l4 + 2l2)/8 | |
1129 "psubusb %%mm3, %%mm0 \n\t" | |
1130 "psubusb %%mm6, %%mm3 \n\t" | |
1131 "por %%mm0, %%mm3 \n\t" // ~|2l2 - 5l3 + 5l4 - 2l5|/8 | |
1132 "pcmpeqb %%mm7, %%mm0 \n\t" // SIGN(2l2 - 5l3 + 5l4 - 2l5) | |
1133 // mm0= SIGN(menergy), mm1= |lenergy|, mm2= l5, mm3= |menergy|, mm4=l4, mm5= l3, mm7=0 | |
1134 | |
1135 "movq (%%ebx, %1), %%mm6 \n\t" // l6 | |
1136 "movq %%mm6, %%mm5 \n\t" // l6 | |
1137 PAVGB(%%mm7, %%mm6) // ~l6/2 | |
1138 PAVGB(%%mm4, %%mm6) // ~(l6 + 2l4)/4 | |
1139 PAVGB(%%mm5, %%mm6) // ~(5l6 + 2l4)/8 | |
1140 | |
1141 "movq (%%ebx, %1, 2), %%mm5 \n\t" // l7 | |
1142 "movq %%mm2, %%mm4 \n\t" // l5 | |
1143 PAVGB(%%mm7, %%mm2) // ~l5/2 | |
1144 PAVGB(%%mm5, %%mm2) // ~(l5 + 2l7)/4 | |
1145 PAVGB(%%mm4, %%mm2) // ~(5l5 + 2l7)/8 | |
1146 | |
1147 "movq %%mm6, %%mm4 \n\t" // ~(5l6 + 2l4)/8 | |
1148 "psubusb %%mm2, %%mm6 \n\t" | |
1149 "psubusb %%mm4, %%mm2 \n\t" | |
1150 "por %%mm6, %%mm2 \n\t" // ~|2l4 - 5l5 + 5l6 - 2l7|/8 | |
1151 // mm0= SIGN(menergy), mm1= |lenergy|/8, mm2= |renergy|/8, mm3= |menergy|/8, mm7=0 | |
1152 | |
1153 | |
1154 PMINUB(%%mm2, %%mm1, %%mm4) // MIN(|lenergy|,|renergy|)/8 | |
1155 "movq pQPb, %%mm4 \n\t" // QP //FIXME QP+1 ? | |
1156 "paddusb b01, %%mm4 \n\t" | |
1157 "pcmpgtb %%mm3, %%mm4 \n\t" // |menergy|/8 < QP | |
1158 "psubusb %%mm1, %%mm3 \n\t" // d=|menergy|/8-MIN(|lenergy|,|renergy|)/8 | |
1159 "pand %%mm4, %%mm3 \n\t" | |
1160 | |
1161 "movq %%mm3, %%mm1 \n\t" | |
1162 // "psubusb b01, %%mm3 \n\t" | |
1163 PAVGB(%%mm7, %%mm3) | |
1164 PAVGB(%%mm7, %%mm3) | |
1165 "paddusb %%mm1, %%mm3 \n\t" | |
1166 // "paddusb b01, %%mm3 \n\t" | |
1167 | |
1168 "movq (%%eax, %1, 2), %%mm6 \n\t" //l3 | |
1169 "movq (%0, %1, 4), %%mm5 \n\t" //l4 | |
1170 "movq (%0, %1, 4), %%mm4 \n\t" //l4 | |
1171 "psubusb %%mm6, %%mm5 \n\t" | |
1172 "psubusb %%mm4, %%mm6 \n\t" | |
1173 "por %%mm6, %%mm5 \n\t" // |l3-l4| | |
1174 "pcmpeqb %%mm7, %%mm6 \n\t" // SIGN(l3-l4) | |
1175 "pxor %%mm6, %%mm0 \n\t" | |
1176 "pand %%mm0, %%mm3 \n\t" | |
1177 PMINUB(%%mm5, %%mm3, %%mm0) | |
1178 | |
1179 "psubusb b01, %%mm3 \n\t" | |
1180 PAVGB(%%mm7, %%mm3) | |
1181 | |
1182 "movq (%%eax, %1, 2), %%mm0 \n\t" | |
1183 "movq (%0, %1, 4), %%mm2 \n\t" | |
1184 "pxor %%mm6, %%mm0 \n\t" | |
1185 "pxor %%mm6, %%mm2 \n\t" | |
1186 "psubb %%mm3, %%mm0 \n\t" | |
1187 "paddb %%mm3, %%mm2 \n\t" | |
1188 "pxor %%mm6, %%mm0 \n\t" | |
1189 "pxor %%mm6, %%mm2 \n\t" | |
1190 "movq %%mm0, (%%eax, %1, 2) \n\t" | |
1191 "movq %%mm2, (%0, %1, 4) \n\t" | |
1192 #endif | |
1193 | |
1194 "leal (%0, %1), %%eax \n\t" | |
1195 "pcmpeqb %%mm6, %%mm6 \n\t" // -1 | |
1196 // 0 1 2 3 4 5 6 7 | |
1197 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ebx+%1 ebx+2%1 | |
1198 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 | |
1199 | |
1200 | |
1201 "movq (%%eax, %1, 2), %%mm1 \n\t" // l3 | |
1202 "movq (%0, %1, 4), %%mm0 \n\t" // l4 | |
1203 "pxor %%mm6, %%mm1 \n\t" // -l3-1 | |
1204 PAVGB(%%mm1, %%mm0) // -q+128 = (l4-l3+256)/2 | |
1205 // mm1=-l3-1, mm0=128-q | |
1206 | |
1207 "movq (%%eax, %1, 4), %%mm2 \n\t" // l5 | |
1208 "movq (%%eax, %1), %%mm3 \n\t" // l2 | |
1209 "pxor %%mm6, %%mm2 \n\t" // -l5-1 | |
1210 "movq %%mm2, %%mm5 \n\t" // -l5-1 | |
1211 "movq b80, %%mm4 \n\t" // 128 | |
1212 "leal (%%eax, %1, 4), %%ebx \n\t" | |
1213 PAVGB(%%mm3, %%mm2) // (l2-l5+256)/2 | |
1214 PAVGB(%%mm0, %%mm4) // ~(l4-l3)/4 + 128 | |
1215 PAVGB(%%mm2, %%mm4) // ~(l2-l5)/4 +(l4-l3)/8 + 128 | |
1216 PAVGB(%%mm0, %%mm4) // ~(l2-l5)/8 +5(l4-l3)/16 + 128 | |
1217 // mm1=-l3-1, mm0=128-q, mm3=l2, mm4=menergy/16 + 128, mm5= -l5-1 | |
1218 | |
1219 "movq (%%eax), %%mm2 \n\t" // l1 | |
1220 "pxor %%mm6, %%mm2 \n\t" // -l1-1 | |
1221 PAVGB(%%mm3, %%mm2) // (l2-l1+256)/2 | |
1222 PAVGB((%0), %%mm1) // (l0-l3+256)/2 | |
1223 "movq b80, %%mm3 \n\t" // 128 | |
1224 PAVGB(%%mm2, %%mm3) // ~(l2-l1)/4 + 128 | |
1225 PAVGB(%%mm1, %%mm3) // ~(l0-l3)/4 +(l2-l1)/8 + 128 | |
1226 PAVGB(%%mm2, %%mm3) // ~(l0-l3)/8 +5(l2-l1)/16 + 128 | |
1227 // mm0=128-q, mm3=lenergy/16 + 128, mm4= menergy/16 + 128, mm5= -l5-1 | |
1228 | |
1229 PAVGB((%%ebx, %1), %%mm5) // (l6-l5+256)/2 | |
1230 "movq (%%ebx, %1, 2), %%mm1 \n\t" // l7 | |
1231 "pxor %%mm6, %%mm1 \n\t" // -l7-1 | |
1232 PAVGB((%0, %1, 4), %%mm1) // (l4-l7+256)/2 | |
1233 "movq b80, %%mm2 \n\t" // 128 | |
1234 PAVGB(%%mm5, %%mm2) // ~(l6-l5)/4 + 128 | |
1235 PAVGB(%%mm1, %%mm2) // ~(l4-l7)/4 +(l6-l5)/8 + 128 | |
1236 PAVGB(%%mm5, %%mm2) // ~(l4-l7)/8 +5(l6-l5)/16 + 128 | |
1237 // mm0=128-q, mm2=renergy/16 + 128, mm3=lenergy/16 + 128, mm4= menergy/16 + 128 | |
1238 | |
1239 "movq b00, %%mm1 \n\t" // 0 | |
1240 "movq b00, %%mm5 \n\t" // 0 | |
1241 "psubb %%mm2, %%mm1 \n\t" // 128 - renergy/16 | |
1242 "psubb %%mm3, %%mm5 \n\t" // 128 - lenergy/16 | |
1243 PMAXUB(%%mm1, %%mm2) // 128 + |renergy/16| | |
1244 PMAXUB(%%mm5, %%mm3) // 128 + |lenergy/16| | |
1245 PMINUB(%%mm2, %%mm3, %%mm1) // 128 + MIN(|lenergy|,|renergy|)/16 | |
1246 | |
1247 // mm0=128-q, mm3=128 + MIN(|lenergy|,|renergy|)/16, mm4= menergy/16 + 128 | |
1248 | |
1249 "movq b00, %%mm7 \n\t" // 0 | |
1250 "movq pQPb, %%mm2 \n\t" // QP | |
1251 PAVGB(%%mm6, %%mm2) // 128 + QP/2 | |
1252 "psubb %%mm6, %%mm2 \n\t" | |
1253 | |
1254 "movq %%mm4, %%mm1 \n\t" | |
1255 "pcmpgtb %%mm7, %%mm1 \n\t" // SIGN(menergy) | |
1256 "pxor %%mm1, %%mm4 \n\t" | |
1257 "psubb %%mm1, %%mm4 \n\t" // 128 + |menergy|/16 | |
1258 "pcmpgtb %%mm4, %%mm2 \n\t" // |menergy|/16 < QP/2 | |
1259 "psubusb %%mm3, %%mm4 \n\t" //d=|menergy|/16 - MIN(|lenergy|,|renergy|)/16 | |
1260 // mm0=128-q, mm1= SIGN(menergy), mm2= |menergy|/16 < QP/2, mm4= d/16 | |
1261 | |
1262 "movq %%mm4, %%mm3 \n\t" // d | |
1263 "psubusb b01, %%mm4 \n\t" | |
1264 PAVGB(%%mm7, %%mm4) // d/32 | |
1265 PAVGB(%%mm7, %%mm4) // (d + 32)/64 | |
1266 "paddb %%mm3, %%mm4 \n\t" // 5d/64 | |
1267 "pand %%mm2, %%mm4 \n\t" | |
1268 | |
1269 "movq b80, %%mm5 \n\t" // 128 | |
1270 "psubb %%mm0, %%mm5 \n\t" // q | |
1271 "paddsb %%mm6, %%mm5 \n\t" // fix bad rounding | |
1272 "pcmpgtb %%mm5, %%mm7 \n\t" // SIGN(q) | |
1273 "pxor %%mm7, %%mm5 \n\t" | |
1274 | |
1275 PMINUB(%%mm5, %%mm4, %%mm3) // MIN(|q|, 5d/64) | |
1276 "pxor %%mm1, %%mm7 \n\t" // SIGN(d*q) | |
1277 | |
1278 "pand %%mm7, %%mm4 \n\t" | |
1279 "movq (%%eax, %1, 2), %%mm0 \n\t" | |
1280 "movq (%0, %1, 4), %%mm2 \n\t" | |
1281 "pxor %%mm1, %%mm0 \n\t" | |
1282 "pxor %%mm1, %%mm2 \n\t" | |
1283 "paddb %%mm4, %%mm0 \n\t" | |
1284 "psubb %%mm4, %%mm2 \n\t" | |
1285 "pxor %%mm1, %%mm0 \n\t" | |
1286 "pxor %%mm1, %%mm2 \n\t" | |
1287 "movq %%mm0, (%%eax, %1, 2) \n\t" | |
1288 "movq %%mm2, (%0, %1, 4) \n\t" | |
1289 | |
1290 : | |
1291 : "r" (src), "r" (stride) | |
1292 : "%eax", "%ebx" | |
1293 ); | |
1294 | |
1295 /* | |
1296 { | |
1297 int x; | |
1298 src-= stride; | |
1299 for(x=0; x<BLOCK_SIZE; x++) | |
1300 { | |
1301 const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]); | |
1302 if(ABS(middleEnergy)< 8*QP) | |
1303 { | |
1304 const int q=(src[l4] - src[l5])/2; | |
1305 const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]); | |
1306 const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]); | |
1307 | |
1308 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) ); | |
1309 d= MAX(d, 0); | |
1310 | |
1311 d= (5*d + 32) >> 6; | |
1312 d*= SIGN(-middleEnergy); | |
1313 | |
1314 if(q>0) | |
1315 { | |
1316 d= d<0 ? 0 : d; | |
1317 d= d>q ? q : d; | |
1318 } | |
1319 else | |
1320 { | |
1321 d= d>0 ? 0 : d; | |
1322 d= d<q ? q : d; | |
1323 } | |
1324 | |
1325 src[l4]-= d; | |
1326 src[l5]+= d; | |
1327 } | |
1328 src++; | |
1329 } | |
1330 src-=8; | |
1331 for(x=0; x<8; x++) | |
1332 { | |
1333 int y; | |
1334 for(y=4; y<6; y++) | |
1335 { | |
1336 int d= src[x+y*stride] - tmp[x+(y-4)*8]; | |
1337 int ad= ABS(d); | |
1338 static int max=0; | |
1339 static int sum=0; | |
1340 static int num=0; | |
1341 static int bias=0; | |
1342 | |
1343 if(max<ad) max=ad; | |
1344 sum+= ad>3 ? 1 : 0; | |
1345 if(ad>3) | |
1346 { | |
1347 src[0] = src[7] = src[stride*7] = src[(stride+1)*7]=255; | |
1348 } | |
1349 if(y==4) bias+=d; | |
1350 num++; | |
1351 if(num%1000000 == 0) | |
1352 { | |
1353 printf(" %d %d %d %d\n", num, sum, max, bias); | |
1354 } | |
1355 } | |
1356 } | |
1357 } | |
1358 */ | |
1359 #elif defined (HAVE_MMX) | |
1360 src+= stride*4; | |
1361 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1362 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1363 "pxor %%mm7, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1364 "leal (%0, %1), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1365 "leal (%%eax, %1, 4), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1366 // 0 1 2 3 4 5 6 7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1367 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ebx+%1 ebx+2%1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1368 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1369 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1370 "movq (%0), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1371 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1372 "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1373 "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1374 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1375 "movq (%%eax), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1376 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1377 "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1378 "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1379 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1380 "movq (%%eax, %1), %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1381 "movq %%mm4, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1382 "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1383 "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1384 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1385 "paddw %%mm0, %%mm0 \n\t" // 2L0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1386 "paddw %%mm1, %%mm1 \n\t" // 2H0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1387 "psubw %%mm4, %%mm2 \n\t" // L1 - L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1388 "psubw %%mm5, %%mm3 \n\t" // H1 - H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1389 "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1390 "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1391 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1392 "psllw $2, %%mm2 \n\t" // 4L1 - 4L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1393 "psllw $2, %%mm3 \n\t" // 4H1 - 4H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1394 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1395 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1396 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1397 "movq (%%eax, %1, 2), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1398 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1399 "punpcklbw %%mm7, %%mm2 \n\t" // L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1400 "punpckhbw %%mm7, %%mm3 \n\t" // H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1401 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1402 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1403 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1404 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1405 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1406 "movq %%mm0, temp0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1407 "movq %%mm1, temp1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1408 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1409 "movq (%0, %1, 4), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1410 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1411 "punpcklbw %%mm7, %%mm0 \n\t" // L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1412 "punpckhbw %%mm7, %%mm1 \n\t" // H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1413 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1414 "psubw %%mm0, %%mm2 \n\t" // L3 - L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1415 "psubw %%mm1, %%mm3 \n\t" // H3 - H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1416 "movq %%mm2, temp2 \n\t" // L3 - L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1417 "movq %%mm3, temp3 \n\t" // H3 - H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1418 "paddw %%mm4, %%mm4 \n\t" // 2L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1419 "paddw %%mm5, %%mm5 \n\t" // 2H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1420 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1421 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1422 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1423 "psllw $2, %%mm2 \n\t" // 4L3 - 4L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1424 "psllw $2, %%mm3 \n\t" // 4H3 - 4H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1425 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1426 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1427 //50 opcodes so far |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1428 "movq (%%ebx), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1429 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1430 "punpcklbw %%mm7, %%mm2 \n\t" // L5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1431 "punpckhbw %%mm7, %%mm3 \n\t" // H5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1432 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1433 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1434 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1435 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1436 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1437 "movq (%%ebx, %1), %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1438 "punpcklbw %%mm7, %%mm6 \n\t" // L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1439 "psubw %%mm6, %%mm2 \n\t" // L5 - L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1440 "movq (%%ebx, %1), %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1441 "punpckhbw %%mm7, %%mm6 \n\t" // H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1442 "psubw %%mm6, %%mm3 \n\t" // H5 - H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1443 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1444 "paddw %%mm0, %%mm0 \n\t" // 2L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1445 "paddw %%mm1, %%mm1 \n\t" // 2H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1446 "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1447 "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1448 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1449 "psllw $2, %%mm2 \n\t" // 4L5 - 4L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1450 "psllw $2, %%mm3 \n\t" // 4H5 - 4H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1451 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1452 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1453 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1454 "movq (%%ebx, %1, 2), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1455 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1456 "punpcklbw %%mm7, %%mm2 \n\t" // L7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1457 "punpckhbw %%mm7, %%mm3 \n\t" // H7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1458 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1459 "paddw %%mm2, %%mm2 \n\t" // 2L7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1460 "paddw %%mm3, %%mm3 \n\t" // 2H7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1461 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1462 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1463 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1464 "movq temp0, %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1465 "movq temp1, %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
2570 | 1466 |
1467 #ifdef HAVE_MMX2 | |
1468 "movq %%mm7, %%mm6 \n\t" // 0 | |
1469 "psubw %%mm0, %%mm6 \n\t" | |
1470 "pmaxsw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| | |
1471 "movq %%mm7, %%mm6 \n\t" // 0 | |
1472 "psubw %%mm1, %%mm6 \n\t" | |
1473 "pmaxsw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7| | |
1474 "movq %%mm7, %%mm6 \n\t" // 0 | |
1475 "psubw %%mm2, %%mm6 \n\t" | |
1476 "pmaxsw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3| | |
1477 "movq %%mm7, %%mm6 \n\t" // 0 | |
1478 "psubw %%mm3, %%mm6 \n\t" | |
1479 "pmaxsw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3| | |
1480 #else | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1481 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1482 "pcmpgtw %%mm0, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1483 "pxor %%mm6, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1484 "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1485 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1486 "pcmpgtw %%mm1, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1487 "pxor %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1488 "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1489 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1490 "pcmpgtw %%mm2, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1491 "pxor %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1492 "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1493 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1494 "pcmpgtw %%mm3, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1495 "pxor %%mm6, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1496 "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3| |
2570 | 1497 #endif |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1498 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1499 #ifdef HAVE_MMX2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1500 "pminsw %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1501 "pminsw %%mm3, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1502 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1503 "movq %%mm0, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1504 "psubusw %%mm2, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1505 "psubw %%mm6, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1506 "movq %%mm1, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1507 "psubusw %%mm3, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1508 "psubw %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1509 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1510 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1511 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1512 "pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1513 "pxor %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1514 "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1515 "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1516 "pxor %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1517 "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1518 // 100 opcodes |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1519 "movd %2, %%mm2 \n\t" // QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1520 "punpcklwd %%mm2, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1521 "punpcklwd %%mm2, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1522 "psllw $3, %%mm2 \n\t" // 8QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1523 "movq %%mm2, %%mm3 \n\t" // 8QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1524 "pcmpgtw %%mm4, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1525 "pcmpgtw %%mm5, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1526 "pand %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1527 "pand %%mm3, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1528 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1529 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1530 "psubusw %%mm0, %%mm4 \n\t" // hd |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1531 "psubusw %%mm1, %%mm5 \n\t" // ld |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1532 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1533 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1534 "movq w05, %%mm2 \n\t" // 5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1535 "pmullw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1536 "pmullw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1537 "movq w20, %%mm2 \n\t" // 32 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1538 "paddw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1539 "paddw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1540 "psrlw $6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1541 "psrlw $6, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1542 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1543 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1544 "movq w06, %%mm2 \n\t" // 6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1545 "paddw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1546 "paddw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1547 "movq w1400, %%mm2 \n\t" // 1400h = 5120 = 5/64*2^16 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1548 //FIXME if *5/64 is supposed to be /13 then we should use 5041 instead of 5120 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1549 "pmulhw %%mm2, %%mm4 \n\t" // hd/13 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1550 "pmulhw %%mm2, %%mm5 \n\t" // ld/13 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1551 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1552 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1553 "movq temp2, %%mm0 \n\t" // L3 - L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1554 "movq temp3, %%mm1 \n\t" // H3 - H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1555 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1556 "pxor %%mm2, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1557 "pxor %%mm3, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1558 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1559 "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1560 "pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1561 "pxor %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1562 "pxor %%mm3, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1563 "psubw %%mm2, %%mm0 \n\t" // |L3-L4| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1564 "psubw %%mm3, %%mm1 \n\t" // |H3-H4| |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1565 "psrlw $1, %%mm0 \n\t" // |L3 - L4|/2 |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1566 "psrlw $1, %%mm1 \n\t" // |H3 - H4|/2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1567 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1568 "pxor %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1569 "pxor %%mm7, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1570 "pand %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1571 "pand %%mm3, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1572 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1573 #ifdef HAVE_MMX2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1574 "pminsw %%mm0, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1575 "pminsw %%mm1, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1576 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1577 "movq %%mm4, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1578 "psubusw %%mm0, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1579 "psubw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1580 "movq %%mm5, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1581 "psubusw %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1582 "psubw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1583 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1584 "pxor %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1585 "pxor %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1586 "psubw %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1587 "psubw %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1588 "packsswb %%mm5, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1589 "movq (%%eax, %1, 2), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1590 "paddb %%mm4, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1591 "movq %%mm0, (%%eax, %1, 2) \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1592 "movq (%0, %1, 4), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1593 "psubb %%mm4, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1594 "movq %%mm0, (%0, %1, 4) \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1595 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1596 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1597 : "r" (src), "r" (stride), "r" (QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1598 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1599 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1600 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1601 const int l1= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1602 const int l2= stride + l1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1603 const int l3= stride + l2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1604 const int l4= stride + l3; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1605 const int l5= stride + l4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1606 const int l6= stride + l5; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1607 const int l7= stride + l6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1608 const int l8= stride + l7; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1609 // const int l9= stride + l8; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1610 int x; |
2246 | 1611 src+= stride*3; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1612 for(x=0; x<BLOCK_SIZE; x++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1613 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1614 const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1615 if(ABS(middleEnergy) < 8*QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1616 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1617 const int q=(src[l4] - src[l5])/2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1618 const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1619 const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1620 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1621 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1622 d= MAX(d, 0); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1623 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1624 d= (5*d + 32) >> 6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1625 d*= SIGN(-middleEnergy); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1626 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1627 if(q>0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1628 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1629 d= d<0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1630 d= d>q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1631 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1632 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1633 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1634 d= d>0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1635 d= d<q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1636 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1637 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1638 src[l4]-= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1639 src[l5]+= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1640 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1641 src++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1642 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1643 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1644 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1645 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1646 //FIXME? |255-0| = 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1647 /** |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1648 * Check if the given 8x8 Block is mostly "flat" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1649 */ |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1650 static inline int isHorizDC(uint8_t src[], int stride) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1651 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1652 // src++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1653 int numEq= 0; |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1654 #if 0 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1655 asm volatile ( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1656 // "int $3 \n\t" |
2416 | 1657 "leal (%1, %2), %%ecx \n\t" |
1658 "leal (%%ecx, %2, 4), %%ebx \n\t" | |
1659 // 0 1 2 3 4 5 6 7 8 9 | |
1660 // %1 ecx ecx+%2 ecx+2%2 %1+4%2 ebx ebx+%2 ebx+2%2 %1+8%2 ebx+4%2 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1661 "movq b7E, %%mm7 \n\t" // mm7 = 0x7F |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1662 "movq b7C, %%mm6 \n\t" // mm6 = 0x7D |
2416 | 1663 "pxor %%mm0, %%mm0 \n\t" |
1664 "movl %1, %%eax \n\t" | |
1665 "andl $0x1F, %%eax \n\t" | |
1666 "cmpl $24, %%eax \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1667 "leal tempBlock, %%eax \n\t" |
2416 | 1668 "jb 1f \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1669 |
2416 | 1670 #define HDC_CHECK_AND_CPY(src, dst) \ |
1671 "movd " #src ", %%mm2 \n\t"\ | |
1672 "punpckldq 4" #src ", %%mm2 \n\t" /* (%1) */\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1673 "movq %%mm2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1674 "psrlq $8, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1675 "psubb %%mm1, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1676 "paddb %%mm7, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1677 "pcmpgtb %%mm6, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1678 "paddb %%mm2, %%mm0 \n\t"\ |
2416 | 1679 "movq %%mm1," #dst "(%%eax) \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1680 |
2416 | 1681 HDC_CHECK_AND_CPY((%1),0) |
1682 HDC_CHECK_AND_CPY((%%ecx),8) | |
1683 HDC_CHECK_AND_CPY((%%ecx, %2),16) | |
1684 HDC_CHECK_AND_CPY((%%ecx, %2, 2),24) | |
1685 HDC_CHECK_AND_CPY((%1, %2, 4),32) | |
1686 HDC_CHECK_AND_CPY((%%ebx),40) | |
1687 HDC_CHECK_AND_CPY((%%ebx, %2),48) | |
1688 HDC_CHECK_AND_CPY((%%ebx, %2, 2),56) | |
1689 "jmp 2f \n\t" | |
1690 "1: \n\t" | |
1691 // src does not cross a 32 byte cache line so dont waste time with alignment | |
1692 #define HDC_CHECK_AND_CPY2(src, dst) \ | |
1693 "movq " #src ", %%mm2 \n\t"\ | |
1694 "movq " #src ", %%mm1 \n\t"\ | |
1695 "psrlq $8, %%mm2 \n\t"\ | |
1696 "psubb %%mm1, %%mm2 \n\t"\ | |
1697 "paddb %%mm7, %%mm2 \n\t"\ | |
1698 "pcmpgtb %%mm6, %%mm2 \n\t"\ | |
1699 "paddb %%mm2, %%mm0 \n\t"\ | |
1700 "movq %%mm1," #dst "(%%eax) \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1701 |
2416 | 1702 HDC_CHECK_AND_CPY2((%1),0) |
1703 HDC_CHECK_AND_CPY2((%%ecx),8) | |
1704 HDC_CHECK_AND_CPY2((%%ecx, %2),16) | |
1705 HDC_CHECK_AND_CPY2((%%ecx, %2, 2),24) | |
1706 HDC_CHECK_AND_CPY2((%1, %2, 4),32) | |
1707 HDC_CHECK_AND_CPY2((%%ebx),40) | |
1708 HDC_CHECK_AND_CPY2((%%ebx, %2),48) | |
1709 HDC_CHECK_AND_CPY2((%%ebx, %2, 2),56) | |
1710 "2: \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1711 "psllq $8, %%mm0 \n\t" // remove dummy value |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1712 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1713 "psrlw $8, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1714 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1715 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1716 "psrlq $16, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1717 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1718 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1719 "psrlq $32, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1720 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1721 "movd %%mm0, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1722 : "=r" (numEq) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1723 : "r" (src), "r" (stride) |
2416 | 1724 : "%eax", "%ebx", "%ecx" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1725 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1726 // printf("%d\n", numEq); |
2416 | 1727 numEq= (256 - numEq) &0xFF; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1728 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1729 int y; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1730 for(y=0; y<BLOCK_SIZE; y++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1731 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1732 if(((src[0] - src[1] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1733 if(((src[1] - src[2] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1734 if(((src[2] - src[3] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1735 if(((src[3] - src[4] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1736 if(((src[4] - src[5] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1737 if(((src[5] - src[6] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1738 if(((src[6] - src[7] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1739 src+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1740 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1741 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1742 /* if(abs(numEq - asmEq) > 0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1743 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1744 // printf("\nasm:%d c:%d\n", asmEq, numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1745 for(int y=0; y<8; y++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1746 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1747 for(int x=0; x<8; x++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1748 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1749 printf("%d ", src[x + y*stride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1750 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1751 printf("\n"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1752 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1753 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1754 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1755 // printf("%d\n", numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1756 return numEq > hFlatnessThreshold; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1757 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1758 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1759 static inline int isHorizMinMaxOk(uint8_t src[], int stride, int QP) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1760 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1761 if(abs(src[0] - src[7]) > 2*QP) return 0; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1762 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1763 return 1; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1764 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1765 |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1766 static inline void doHorizDefFilter(uint8_t dst[], int stride, int QP) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1767 { |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1768 #if 0 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1769 asm volatile( |
2428 | 1770 "leal (%0, %1), %%ecx \n\t" |
1771 "leal (%%ecx, %1, 4), %%ebx \n\t" | |
1772 // 0 1 2 3 4 5 6 7 8 9 | |
1773 // %0 ecx ecx+%1 ecx+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1774 "pxor %%mm7, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1775 "movq bm00001000, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1776 "movd %2, %%mm5 \n\t" // QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1777 "movq %%mm5, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1778 "paddusb %%mm5, %%mm5 \n\t" // 2QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1779 "paddusb %%mm5, %%mm4 \n\t" // 3QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1780 "psllq $24, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1781 "pxor %%mm5, %%mm5 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1782 "psubb %%mm4, %%mm5 \n\t" // -QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1783 "leal tempBlock, %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1784 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1785 //FIXME? "unroll by 2" and mix |
2159 | 1786 #ifdef HAVE_MMX2 |
2428 | 1787 #define HDF(src, dst) \ |
1788 "movq " #src "(%%eax), %%mm0 \n\t"\ | |
1789 "movq " #src "(%%eax), %%mm1 \n\t"\ | |
1790 "movq " #src "(%%eax), %%mm2 \n\t"\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1791 "psrlq $8, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1792 "psubusb %%mm1, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1793 "psubusb %%mm0, %%mm1 \n\t"\ |
2159 | 1794 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ |
1795 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
1796 "pshufw $0x00, %%mm1, %%mm3 \n\t" /* p´5 = |p1 - p2| */\ | |
1797 "pminub %%mm1, %%mm3 \n\t" /* p´5 = min(|p2-p1|, |p6-p5|)*/\ | |
1798 "psrlq $16, %%mm3 \n\t" /* p´3 = min(|p2-p1|, |p6-p5|)*/\ | |
1799 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-min(|p1-p2|,|p5-p6|) */\ | |
1800 "paddb %%mm5, %%mm1 \n\t"\ | |
1801 "psubusb %%mm5, %%mm1 \n\t"\ | |
1802 "psrlw $2, %%mm1 \n\t"\ | |
1803 "pxor %%mm2, %%mm1 \n\t"\ | |
1804 "psubb %%mm2, %%mm1 \n\t"\ | |
1805 "pand %%mm6, %%mm1 \n\t"\ | |
1806 "psubb %%mm1, %%mm0 \n\t"\ | |
1807 "psllq $8, %%mm1 \n\t"\ | |
1808 "paddb %%mm1, %%mm0 \n\t"\ | |
2428 | 1809 "movd %%mm0, " #dst" \n\t"\ |
2159 | 1810 "psrlq $32, %%mm0 \n\t"\ |
2428 | 1811 "movd %%mm0, 4" #dst" \n\t" |
2159 | 1812 #else |
2428 | 1813 #define HDF(src, dst)\ |
1814 "movq " #src "(%%eax), %%mm0 \n\t"\ | |
2159 | 1815 "movq %%mm0, %%mm1 \n\t"\ |
1816 "movq %%mm0, %%mm2 \n\t"\ | |
1817 "psrlq $8, %%mm1 \n\t"\ | |
1818 "psubusb %%mm1, %%mm2 \n\t"\ | |
1819 "psubusb %%mm0, %%mm1 \n\t"\ | |
1820 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ | |
1821 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
1822 "movq %%mm1, %%mm3 \n\t"\ | |
1823 "psllq $32, %%mm3 \n\t"\ | |
1824 "movq %%mm3, %%mm4 \n\t"\ | |
1825 "psubusb %%mm1, %%mm4 \n\t"\ | |
1826 "psubb %%mm4, %%mm3 \n\t"\ | |
1827 "psrlq $16, %%mm3 \n\t" /* p´3 = min(|p2-p1|, |p6-p5|)*/\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1828 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-min(|p1-p2|,|p5,ü6|) */\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1829 "paddb %%mm5, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1830 "psubusb %%mm5, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1831 "psrlw $2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1832 "pxor %%mm2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1833 "psubb %%mm2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1834 "pand %%mm6, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1835 "psubb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1836 "psllq $8, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1837 "paddb %%mm1, %%mm0 \n\t"\ |
2428 | 1838 "movd %%mm0, " #dst " \n\t"\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1839 "psrlq $32, %%mm0 \n\t"\ |
2428 | 1840 "movd %%mm0, 4" #dst " \n\t" |
2159 | 1841 #endif |
2428 | 1842 HDF(0,(%0)) |
1843 HDF(8,(%%ecx)) | |
1844 HDF(16,(%%ecx, %1)) | |
1845 HDF(24,(%%ecx, %1, 2)) | |
1846 HDF(32,(%0, %1, 4)) | |
1847 HDF(40,(%%ebx)) | |
1848 HDF(48,(%%ebx, %1)) | |
1849 HDF(56,(%%ebx, %1, 2)) | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1850 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1851 : "r" (dst), "r" (stride), "r" (QP) |
2428 | 1852 : "%eax", "%ebx", "%ecx" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1853 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1854 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1855 int y; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1856 for(y=0; y<BLOCK_SIZE; y++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1857 { |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1858 const int middleEnergy= 5*(dst[4] - dst[5]) + 2*(dst[2] - dst[5]); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1859 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1860 if(ABS(middleEnergy) < 8*QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1861 { |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1862 const int q=(dst[3] - dst[4])/2; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1863 const int leftEnergy= 5*(dst[2] - dst[1]) + 2*(dst[0] - dst[3]); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1864 const int rightEnergy= 5*(dst[6] - dst[5]) + 2*(dst[4] - dst[7]); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1865 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1866 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1867 d= MAX(d, 0); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1868 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1869 d= (5*d + 32) >> 6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1870 d*= SIGN(-middleEnergy); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1871 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1872 if(q>0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1873 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1874 d= d<0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1875 d= d>q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1876 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1877 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1878 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1879 d= d>0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1880 d= d<q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1881 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1882 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1883 dst[3]-= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1884 dst[4]+= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1885 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1886 dst+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1887 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1888 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1889 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1890 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1891 /** |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1892 * Do a horizontal low pass filter on the 10x8 block (dst points to middle 8x8 Block) |
2221 | 1893 * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 (C version) |
1894 * using the 7-Tap Filter (2,2,2,4,2,2,2)/16 (MMX2/3DNOW version) | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1895 */ |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1896 static inline void doHorizLowPass(uint8_t dst[], int stride, int QP) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1897 { |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1898 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1899 #if 0 |
2428 | 1900 asm volatile( |
1901 "leal (%0, %1), %%ecx \n\t" | |
1902 "leal (%%ecx, %1, 4), %%ebx \n\t" | |
1903 // 0 1 2 3 4 5 6 7 8 9 | |
1904 // %0 ecx ecx+%1 ecx+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1905 "pxor %%mm7, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1906 "leal tempBlock, %%eax \n\t" |
2159 | 1907 /* |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1908 #define HLP1 "movq (%0), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1909 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1910 "psllq $8, %%mm0 \n\t"\ |
2159 | 1911 PAVGB(%%mm1, %%mm0)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1912 "psrlw $8, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1913 "pxor %%mm1, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1914 "packuswb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1915 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1916 "movq %%mm0, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1917 "psllq $32, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1918 "paddb %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1919 "psllq $16, %%mm2 \n\t"\ |
2159 | 1920 PAVGB(%%mm2, %%mm0)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1921 "movq %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1922 "pand bm11001100, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1923 "paddusb %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1924 "psrlq $8, %%mm3 \n\t"\ |
2159 | 1925 PAVGB(%%mm1, %%mm4)\ |
1926 PAVGB(%%mm3, %%mm2)\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1927 "psrlq $16, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1928 "punpcklbw %%mm2, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1929 "movq %%mm2, (%0) \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1930 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1931 #define HLP2 "movq (%0), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1932 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1933 "psllq $8, %%mm0 \n\t"\ |
2159 | 1934 PAVGB(%%mm1, %%mm0)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1935 "psrlw $8, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1936 "pxor %%mm1, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1937 "packuswb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1938 "movq %%mm0, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1939 "psllq $32, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1940 "psllq $16, %%mm2 \n\t"\ |
2159 | 1941 PAVGB(%%mm2, %%mm0)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1942 "movq %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1943 "pand bm11001100, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1944 "paddusb %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1945 "psrlq $8, %%mm3 \n\t"\ |
2159 | 1946 PAVGB(%%mm3, %%mm2)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1947 "psrlq $16, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1948 "punpcklbw %%mm2, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1949 "movq %%mm2, (%0) \n\t"\ |
2159 | 1950 */ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1951 // approximately a 7-Tap Filter with Vector (1,2,3,4,3,2,1)/16 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1952 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1953 Implemented Exact 7-Tap |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1954 9421 A321 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1955 36421 64321 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1956 334321 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1957 1234321 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1958 1234321 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1959 123433 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1960 12463 12346 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1961 1249 123A |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1962 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1963 */ |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1964 |
2159 | 1965 #ifdef HAVE_MMX2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1966 #define HLP3(i) "movq " #i "(%%eax), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1967 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1968 "movq %%mm0, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1969 "movq %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1970 "movq %%mm0, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1971 "psllq $8, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1972 "psrlq $8, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1973 "pand bm00000001, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1974 "pand bm10000000, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1975 "por %%mm3, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1976 "por %%mm4, %%mm2 \n\t"\ |
2159 | 1977 PAVGB(%%mm2, %%mm1)\ |
1978 PAVGB(%%mm1, %%mm0)\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1979 \ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1980 "pshufw $0xF9, %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1981 "pshufw $0x90, %%mm0, %%mm4 \n\t"\ |
2159 | 1982 PAVGB(%%mm3, %%mm4)\ |
1983 PAVGB(%%mm4, %%mm0)\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1984 "movd %%mm0, (%0) \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1985 "psrlq $32, %%mm0 \n\t"\ |
2159 | 1986 "movd %%mm0, 4(%0) \n\t" |
1987 #else | |
1988 #define HLP3(i) "movq " #i "(%%eax), %%mm0 \n\t"\ | |
1989 "movq %%mm0, %%mm1 \n\t"\ | |
1990 "movq %%mm0, %%mm2 \n\t"\ | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1991 "movd -4(%0), %%mm3 \n\t" /*0001000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1992 "movd 8(%0), %%mm4 \n\t" /*0001000*/\ |
2159 | 1993 "psllq $8, %%mm1 \n\t"\ |
1994 "psrlq $8, %%mm2 \n\t"\ | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1995 "psrlq $24, %%mm3 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1996 "psllq $56, %%mm4 \n\t"\ |
2159 | 1997 "por %%mm3, %%mm1 \n\t"\ |
1998 "por %%mm4, %%mm2 \n\t"\ | |
1999 PAVGB(%%mm2, %%mm1)\ | |
2000 PAVGB(%%mm1, %%mm0)\ | |
2001 \ | |
2002 "movq %%mm0, %%mm3 \n\t"\ | |
2003 "movq %%mm0, %%mm4 \n\t"\ | |
2004 "movq %%mm0, %%mm5 \n\t"\ | |
2005 "psrlq $16, %%mm3 \n\t"\ | |
2006 "psllq $16, %%mm4 \n\t"\ | |
2007 "pand bm11000000, %%mm5 \n\t"\ | |
2008 "por %%mm5, %%mm3 \n\t"\ | |
2009 "movq %%mm0, %%mm5 \n\t"\ | |
2010 "pand bm00000011, %%mm5 \n\t"\ | |
2011 "por %%mm5, %%mm4 \n\t"\ | |
2012 PAVGB(%%mm3, %%mm4)\ | |
2013 PAVGB(%%mm4, %%mm0)\ | |
2014 "movd %%mm0, (%0) \n\t"\ | |
2015 "psrlq $32, %%mm0 \n\t"\ | |
2016 "movd %%mm0, 4(%0) \n\t" | |
2017 #endif | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2018 |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2019 /* uses the 7-Tap Filter: 1112111 */ |
2428 | 2020 #define NEW_HLP(src, dst)\ |
2021 "movq " #src "(%%eax), %%mm1 \n\t"\ | |
2022 "movq " #src "(%%eax), %%mm2 \n\t"\ | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2023 "psllq $8, %%mm1 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2024 "psrlq $8, %%mm2 \n\t"\ |
2428 | 2025 "movd -4" #dst ", %%mm3 \n\t" /*0001000*/\ |
2026 "movd 8" #dst ", %%mm4 \n\t" /*0001000*/\ | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2027 "psrlq $24, %%mm3 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2028 "psllq $56, %%mm4 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2029 "por %%mm3, %%mm1 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2030 "por %%mm4, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2031 "movq %%mm1, %%mm5 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2032 PAVGB(%%mm2, %%mm1)\ |
2428 | 2033 "movq " #src "(%%eax), %%mm0 \n\t"\ |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2034 PAVGB(%%mm1, %%mm0)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2035 "psllq $8, %%mm5 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2036 "psrlq $8, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2037 "por %%mm3, %%mm5 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2038 "por %%mm4, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2039 "movq %%mm5, %%mm1 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2040 PAVGB(%%mm2, %%mm5)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2041 "psllq $8, %%mm1 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2042 "psrlq $8, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2043 "por %%mm3, %%mm1 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2044 "por %%mm4, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2045 PAVGB(%%mm2, %%mm1)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2046 PAVGB(%%mm1, %%mm5)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2047 PAVGB(%%mm5, %%mm0)\ |
2428 | 2048 "movd %%mm0, " #dst " \n\t"\ |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2049 "psrlq $32, %%mm0 \n\t"\ |
2428 | 2050 "movd %%mm0, 4" #dst " \n\t" |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2051 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2052 /* uses the 9-Tap Filter: 112242211 */ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2053 #define NEW_HLP2(i)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2054 "movq " #i "(%%eax), %%mm0 \n\t" /*0001000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2055 "movq %%mm0, %%mm1 \n\t" /*0001000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2056 "movq %%mm0, %%mm2 \n\t" /*0001000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2057 "movd -4(%0), %%mm3 \n\t" /*0001000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2058 "movd 8(%0), %%mm4 \n\t" /*0001000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2059 "psllq $8, %%mm1 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2060 "psrlq $8, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2061 "psrlq $24, %%mm3 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2062 "psllq $56, %%mm4 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2063 "por %%mm3, %%mm1 \n\t" /*0010000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2064 "por %%mm4, %%mm2 \n\t" /*0000100*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2065 "movq %%mm1, %%mm5 \n\t" /*0010000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2066 PAVGB(%%mm2, %%mm1) /*0010100*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2067 PAVGB(%%mm1, %%mm0) /*0012100*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2068 "psllq $8, %%mm5 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2069 "psrlq $8, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2070 "por %%mm3, %%mm5 \n\t" /*0100000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2071 "por %%mm4, %%mm2 \n\t" /*0000010*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2072 "movq %%mm5, %%mm1 \n\t" /*0100000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2073 PAVGB(%%mm2, %%mm5) /*0100010*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2074 "psllq $8, %%mm1 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2075 "psrlq $8, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2076 "por %%mm3, %%mm1 \n\t" /*1000000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2077 "por %%mm4, %%mm2 \n\t" /*0000001*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2078 "movq %%mm1, %%mm6 \n\t" /*1000000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2079 PAVGB(%%mm2, %%mm1) /*1000001*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2080 "psllq $8, %%mm6 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2081 "psrlq $8, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2082 "por %%mm3, %%mm6 \n\t"/*100000000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2083 "por %%mm4, %%mm2 \n\t"/*000000001*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2084 PAVGB(%%mm2, %%mm6) /*100000001*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2085 PAVGB(%%mm6, %%mm1) /*110000011*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2086 PAVGB(%%mm1, %%mm5) /*112000211*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2087 PAVGB(%%mm5, %%mm0) /*112242211*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2088 "movd %%mm0, (%0) \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2089 "psrlq $32, %%mm0 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2090 "movd %%mm0, 4(%0) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2091 |
2428 | 2092 #define HLP(src, dst) NEW_HLP(src, dst) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2093 |
2428 | 2094 HLP(0, (%0)) |
2095 HLP(8, (%%ecx)) | |
2096 HLP(16, (%%ecx, %1)) | |
2097 HLP(24, (%%ecx, %1, 2)) | |
2098 HLP(32, (%0, %1, 4)) | |
2099 HLP(40, (%%ebx)) | |
2100 HLP(48, (%%ebx, %1)) | |
2101 HLP(56, (%%ebx, %1, 2)) | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2102 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2103 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2104 : "r" (dst), "r" (stride) |
2428 | 2105 : "%eax", "%ebx", "%ecx" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2106 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2107 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2108 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2109 int y; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2110 for(y=0; y<BLOCK_SIZE; y++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2111 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2112 const int first= ABS(dst[-1] - dst[0]) < QP ? dst[-1] : dst[0]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2113 const int last= ABS(dst[8] - dst[7]) < QP ? dst[8] : dst[7]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2114 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2115 int sums[9]; |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2116 sums[0] = first + dst[0]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2117 sums[1] = dst[0] + dst[1]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2118 sums[2] = dst[1] + dst[2]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2119 sums[3] = dst[2] + dst[3]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2120 sums[4] = dst[3] + dst[4]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2121 sums[5] = dst[4] + dst[5]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2122 sums[6] = dst[5] + dst[6]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2123 sums[7] = dst[6] + dst[7]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2124 sums[8] = dst[7] + last; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2125 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2126 dst[0]= ((sums[0]<<2) + ((first + sums[2])<<1) + sums[4] + 8)>>4; |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2127 dst[1]= ((dst[1]<<2) + ((first + sums[0] + sums[3])<<1) + sums[5] + 8)>>4; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2128 dst[2]= ((dst[2]<<2) + ((first + sums[1] + sums[4])<<1) + sums[6] + 8)>>4; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2129 dst[3]= ((dst[3]<<2) + ((sums[2] + sums[5])<<1) + sums[0] + sums[7] + 8)>>4; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2130 dst[4]= ((dst[4]<<2) + ((sums[3] + sums[6])<<1) + sums[1] + sums[8] + 8)>>4; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2131 dst[5]= ((dst[5]<<2) + ((last + sums[7] + sums[4])<<1) + sums[2] + 8)>>4; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2132 dst[6]= (((last + dst[6])<<2) + ((dst[7] + sums[5])<<1) + sums[3] + 8)>>4; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2133 dst[7]= ((sums[8]<<2) + ((last + sums[6])<<1) + sums[4] + 8)>>4; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2134 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2135 dst+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2136 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2137 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2138 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2139 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2140 static inline void dering(uint8_t src[], int stride, int QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2141 { |
2475 | 2142 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2143 asm volatile( |
2473 | 2144 "movq pQPb, %%mm0 \n\t" |
2145 "paddusb %%mm0, %%mm0 \n\t" | |
2146 "movq %%mm0, pQPb2 \n\t" | |
2147 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2148 "leal (%0, %1), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2149 "leal (%%eax, %1, 4), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2150 // 0 1 2 3 4 5 6 7 8 9 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2151 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2152 |
2473 | 2153 "pcmpeqb %%mm6, %%mm6 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2154 "pxor %%mm7, %%mm7 \n\t" |
2475 | 2155 #ifdef HAVE_MMX2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2156 #define FIND_MIN_MAX(addr)\ |
2473 | 2157 "movq " #addr ", %%mm0 \n\t"\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2158 "pminub %%mm0, %%mm6 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2159 "pmaxub %%mm0, %%mm7 \n\t" |
2475 | 2160 #else |
2161 #define FIND_MIN_MAX(addr)\ | |
2162 "movq " #addr ", %%mm0 \n\t"\ | |
2163 "movq %%mm6, %%mm1 \n\t"\ | |
2164 "psubusb %%mm0, %%mm7 \n\t"\ | |
2165 "paddb %%mm0, %%mm7 \n\t"\ | |
2166 "psubusb %%mm0, %%mm1 \n\t"\ | |
2167 "psubb %%mm1, %%mm6 \n\t" | |
2168 #endif | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2169 |
2473 | 2170 FIND_MIN_MAX((%%eax)) |
2171 FIND_MIN_MAX((%%eax, %1)) | |
2172 FIND_MIN_MAX((%%eax, %1, 2)) | |
2173 FIND_MIN_MAX((%0, %1, 4)) | |
2174 FIND_MIN_MAX((%%ebx)) | |
2175 FIND_MIN_MAX((%%ebx, %1)) | |
2176 FIND_MIN_MAX((%%ebx, %1, 2)) | |
2177 FIND_MIN_MAX((%0, %1, 8)) | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2178 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2179 "movq %%mm6, %%mm4 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2180 "psrlq $8, %%mm6 \n\t" |
2475 | 2181 #ifdef HAVE_MMX2 |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2182 "pminub %%mm4, %%mm6 \n\t" // min of pixels |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2183 "pshufw $0xF9, %%mm6, %%mm4 \n\t" |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2184 "pminub %%mm4, %%mm6 \n\t" // min of pixels |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2185 "pshufw $0xFE, %%mm6, %%mm4 \n\t" |
2475 | 2186 "pminub %%mm4, %%mm6 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2187 #else |
2475 | 2188 "movq %%mm6, %%mm1 \n\t" |
2189 "psubusb %%mm4, %%mm1 \n\t" | |
2190 "psubb %%mm1, %%mm6 \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2191 "movq %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2192 "psrlq $16, %%mm6 \n\t" |
2475 | 2193 "movq %%mm6, %%mm1 \n\t" |
2194 "psubusb %%mm4, %%mm1 \n\t" | |
2195 "psubb %%mm1, %%mm6 \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2196 "movq %%mm6, %%mm4 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2197 "psrlq $32, %%mm6 \n\t" |
2475 | 2198 "movq %%mm6, %%mm1 \n\t" |
2199 "psubusb %%mm4, %%mm1 \n\t" | |
2200 "psubb %%mm1, %%mm6 \n\t" | |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2201 #endif |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2202 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2203 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2204 "movq %%mm7, %%mm4 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2205 "psrlq $8, %%mm7 \n\t" |
2475 | 2206 #ifdef HAVE_MMX2 |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2207 "pmaxub %%mm4, %%mm7 \n\t" // max of pixels |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2208 "pshufw $0xF9, %%mm7, %%mm4 \n\t" |
2475 | 2209 "pmaxub %%mm4, %%mm7 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2210 "pshufw $0xFE, %%mm7, %%mm4 \n\t" |
2475 | 2211 "pmaxub %%mm4, %%mm7 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2212 #else |
2475 | 2213 "psubusb %%mm4, %%mm7 \n\t" |
2214 "paddb %%mm4, %%mm7 \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2215 "movq %%mm7, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2216 "psrlq $16, %%mm7 \n\t" |
2475 | 2217 "psubusb %%mm4, %%mm7 \n\t" |
2218 "paddb %%mm4, %%mm7 \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2219 "movq %%mm7, %%mm4 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2220 "psrlq $32, %%mm7 \n\t" |
2475 | 2221 "psubusb %%mm4, %%mm7 \n\t" |
2222 "paddb %%mm4, %%mm7 \n\t" | |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2223 #endif |
2473 | 2224 PAVGB(%%mm6, %%mm7) // a=(max + min)/2 |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2225 "punpcklbw %%mm7, %%mm7 \n\t" |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2226 "punpcklbw %%mm7, %%mm7 \n\t" |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2227 "punpcklbw %%mm7, %%mm7 \n\t" |
2473 | 2228 "movq %%mm7, temp0 \n\t" |
2229 | |
2230 "movq (%0), %%mm0 \n\t" // L10 | |
2231 "movq %%mm0, %%mm1 \n\t" // L10 | |
2232 "movq %%mm0, %%mm2 \n\t" // L10 | |
2233 "psllq $8, %%mm1 \n\t" | |
2234 "psrlq $8, %%mm2 \n\t" | |
2235 "movd -4(%0), %%mm3 \n\t" | |
2236 "movd 8(%0), %%mm4 \n\t" | |
2237 "psrlq $24, %%mm3 \n\t" | |
2238 "psllq $56, %%mm4 \n\t" | |
2239 "por %%mm3, %%mm1 \n\t" // L00 | |
2240 "por %%mm4, %%mm2 \n\t" // L20 | |
2241 "movq %%mm1, %%mm3 \n\t" // L00 | |
2242 PAVGB(%%mm2, %%mm1) // (L20 + L00)/2 | |
2243 PAVGB(%%mm0, %%mm1) // (L20 + L00 + 2L10)/4 | |
2244 "psubusb %%mm7, %%mm0 \n\t" | |
2245 "psubusb %%mm7, %%mm2 \n\t" | |
2246 "psubusb %%mm7, %%mm3 \n\t" | |
2247 "pcmpeqb b00, %%mm0 \n\t" // L10 > a ? 0 : -1 | |
2248 "pcmpeqb b00, %%mm2 \n\t" // L20 > a ? 0 : -1 | |
2249 "pcmpeqb b00, %%mm3 \n\t" // L00 > a ? 0 : -1 | |
2250 "paddb %%mm2, %%mm0 \n\t" | |
2251 "paddb %%mm3, %%mm0 \n\t" | |
2252 | |
2253 "movq (%%eax), %%mm2 \n\t" // L11 | |
2254 "movq %%mm2, %%mm3 \n\t" // L11 | |
2255 "movq %%mm2, %%mm4 \n\t" // L11 | |
2256 "psllq $8, %%mm3 \n\t" | |
2257 "psrlq $8, %%mm4 \n\t" | |
2258 "movd -4(%%eax), %%mm5 \n\t" | |
2259 "movd 8(%%eax), %%mm6 \n\t" | |
2260 "psrlq $24, %%mm5 \n\t" | |
2261 "psllq $56, %%mm6 \n\t" | |
2262 "por %%mm5, %%mm3 \n\t" // L01 | |
2263 "por %%mm6, %%mm4 \n\t" // L21 | |
2264 "movq %%mm3, %%mm5 \n\t" // L01 | |
2265 PAVGB(%%mm4, %%mm3) // (L21 + L01)/2 | |
2266 PAVGB(%%mm2, %%mm3) // (L21 + L01 + 2L11)/4 | |
2267 "psubusb %%mm7, %%mm2 \n\t" | |
2268 "psubusb %%mm7, %%mm4 \n\t" | |
2269 "psubusb %%mm7, %%mm5 \n\t" | |
2270 "pcmpeqb b00, %%mm2 \n\t" // L11 > a ? 0 : -1 | |
2271 "pcmpeqb b00, %%mm4 \n\t" // L21 > a ? 0 : -1 | |
2272 "pcmpeqb b00, %%mm5 \n\t" // L01 > a ? 0 : -1 | |
2273 "paddb %%mm4, %%mm2 \n\t" | |
2274 "paddb %%mm5, %%mm2 \n\t" | |
2275 // 0, 2, 3, 1 | |
2276 #define DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \ | |
2277 "movq " #src ", " #sx " \n\t" /* src[0] */\ | |
2278 "movq " #sx ", " #lx " \n\t" /* src[0] */\ | |
2279 "movq " #sx ", " #t0 " \n\t" /* src[0] */\ | |
2280 "psllq $8, " #lx " \n\t"\ | |
2281 "psrlq $8, " #t0 " \n\t"\ | |
2282 "movd -4" #src ", " #t1 " \n\t"\ | |
2283 "psrlq $24, " #t1 " \n\t"\ | |
2284 "por " #t1 ", " #lx " \n\t" /* src[-1] */\ | |
2285 "movd 8" #src ", " #t1 " \n\t"\ | |
2286 "psllq $56, " #t1 " \n\t"\ | |
2287 "por " #t1 ", " #t0 " \n\t" /* src[+1] */\ | |
2288 "movq " #lx ", " #t1 " \n\t" /* src[-1] */\ | |
2289 PAVGB(t0, lx) /* (src[-1] + src[+1])/2 */\ | |
2290 PAVGB(sx, lx) /* (src[-1] + 2src[0] + src[+1])/4 */\ | |
2478 | 2291 PAVGB(lx, pplx) \ |
2292 "movq " #lx ", temp1 \n\t"\ | |
2293 "movq temp0, " #lx " \n\t"\ | |
2570 | 2294 "psubusb " #lx ", " #t1 " \n\t"\ |
2295 "psubusb " #lx ", " #t0 " \n\t"\ | |
2296 "psubusb " #lx ", " #sx " \n\t"\ | |
2478 | 2297 "movq b00, " #lx " \n\t"\ |
2570 | 2298 "pcmpeqb " #lx ", " #t1 " \n\t" /* src[-1] > a ? 0 : -1*/\ |
2299 "pcmpeqb " #lx ", " #t0 " \n\t" /* src[+1] > a ? 0 : -1*/\ | |
2300 "pcmpeqb " #lx ", " #sx " \n\t" /* src[0] > a ? 0 : -1*/\ | |
2473 | 2301 "paddb " #t1 ", " #t0 " \n\t"\ |
2302 "paddb " #t0 ", " #sx " \n\t"\ | |
2303 \ | |
2304 PAVGB(plx, pplx) /* filtered */\ | |
2305 "movq " #dst ", " #t0 " \n\t" /* dst */\ | |
2477 | 2306 "movq " #t0 ", " #t1 " \n\t" /* dst */\ |
2307 "psubusb pQPb2, " #t0 " \n\t"\ | |
2308 "paddusb pQPb2, " #t1 " \n\t"\ | |
2309 PMAXUB(t0, pplx)\ | |
2310 PMINUB(t1, pplx, t0)\ | |
2473 | 2311 "paddb " #sx ", " #ppsx " \n\t"\ |
2312 "paddb " #psx ", " #ppsx " \n\t"\ | |
2313 "#paddb b02, " #ppsx " \n\t"\ | |
2314 "pand b08, " #ppsx " \n\t"\ | |
2570 | 2315 "pcmpeqb " #lx ", " #ppsx " \n\t"\ |
2477 | 2316 "pand " #ppsx ", " #pplx " \n\t"\ |
2473 | 2317 "pandn " #dst ", " #ppsx " \n\t"\ |
2570 | 2318 "por " #pplx ", " #ppsx " \n\t"\ |
2478 | 2319 "movq " #ppsx ", " #dst " \n\t"\ |
2320 "movq temp1, " #lx " \n\t" | |
2477 | 2321 |
2473 | 2322 /* |
2323 0000000 | |
2324 1111111 | |
2325 | |
2326 1111110 | |
2327 1111101 | |
2328 1111100 | |
2329 1111011 | |
2330 1111010 | |
2331 1111001 | |
2332 | |
2333 1111000 | |
2334 1110111 | |
2335 | |
2336 */ | |
2337 //DERING_CORE(dst,src ,ppsx ,psx ,sx ,pplx ,plx ,lx ,t0 ,t1) | |
2338 DERING_CORE((%%eax),(%%eax, %1) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7) | |
2339 DERING_CORE((%%eax, %1),(%%eax, %1, 2) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7) | |
2340 DERING_CORE((%%eax, %1, 2),(%0, %1, 4) ,%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7) | |
2341 DERING_CORE((%0, %1, 4),(%%ebx) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7) | |
2342 DERING_CORE((%%ebx),(%%ebx, %1) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7) | |
2343 DERING_CORE((%%ebx, %1), (%%ebx, %1, 2),%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7) | |
2344 DERING_CORE((%%ebx, %1, 2),(%0, %1, 8) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7) | |
2345 DERING_CORE((%0, %1, 8),(%%ebx, %1, 4) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7) | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2346 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2347 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2348 : : "r" (src), "r" (stride), "r" (QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2349 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2350 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2351 #else |
2477 | 2352 int y; |
2353 int min=255; | |
2354 int max=0; | |
2355 int avg; | |
2356 uint8_t *p; | |
2357 int s[10]; | |
2358 | |
2359 for(y=1; y<9; y++) | |
2360 { | |
2361 int x; | |
2362 p= src + stride*y; | |
2363 for(x=1; x<9; x++) | |
2364 { | |
2365 p++; | |
2366 if(*p > max) max= *p; | |
2367 if(*p < min) min= *p; | |
2368 } | |
2369 } | |
2370 avg= (min + max + 1)/2; | |
2371 | |
2372 for(y=0; y<10; y++) | |
2373 { | |
2374 int x; | |
2375 int t = 0; | |
2376 p= src + stride*y; | |
2377 for(x=0; x<10; x++) | |
2378 { | |
2379 if(*p > avg) t |= (1<<x); | |
2380 p++; | |
2381 } | |
2382 t |= (~t)<<16; | |
2383 t &= (t<<1) & (t>>1); | |
2384 s[y] = t; | |
2385 } | |
2386 | |
2387 for(y=1; y<9; y++) | |
2388 { | |
2389 int x; | |
2390 int t = s[y-1] & s[y] & s[y+1]; | |
2391 t|= t>>16; | |
2392 | |
2393 p= src + stride*y; | |
2394 for(x=1; x<9; x++) | |
2395 { | |
2396 p++; | |
2397 if(t & (1<<x)) | |
2398 { | |
2399 int f= (*(p-stride-1)) + 2*(*(p-stride)) + (*(p-stride+1)) | |
2400 +2*(*(p -1)) + 4*(*p ) + 2*(*(p +1)) | |
2401 +(*(p+stride-1)) + 2*(*(p+stride)) + (*(p+stride+1)); | |
2402 f= (f + 8)>>4; | |
2403 | |
2404 if (*p + 2*QP < f) *p= *p + 2*QP; | |
2405 else if(*p - 2*QP > f) *p= *p - 2*QP; | |
2406 else *p=f; | |
2407 } | |
2408 } | |
2409 } | |
2410 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2411 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2412 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2413 |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2414 /** |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2415 * Deinterlaces the given block |
2595 | 2416 * will be called for every 8x8 block and can read & write from line 4-15 |
2417 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too | |
2418 * lines 4-12 will be read into the deblocking filter and should be deinterlaced | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2419 */ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2420 static inline void deInterlaceInterpolateLinear(uint8_t src[], int stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2421 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2422 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2595 | 2423 src+= 4*stride; |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2424 asm volatile( |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2425 "leal (%0, %1), %%eax \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2426 "leal (%%eax, %1, 4), %%ebx \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2427 // 0 1 2 3 4 5 6 7 8 9 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2428 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2429 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2430 "movq (%0), %%mm0 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2431 "movq (%%eax, %1), %%mm1 \n\t" |
2246 | 2432 PAVGB(%%mm1, %%mm0) |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2433 "movq %%mm0, (%%eax) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2434 "movq (%0, %1, 4), %%mm0 \n\t" |
2246 | 2435 PAVGB(%%mm0, %%mm1) |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2436 "movq %%mm1, (%%eax, %1, 2) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2437 "movq (%%ebx, %1), %%mm1 \n\t" |
2246 | 2438 PAVGB(%%mm1, %%mm0) |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2439 "movq %%mm0, (%%ebx) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2440 "movq (%0, %1, 8), %%mm0 \n\t" |
2246 | 2441 PAVGB(%%mm0, %%mm1) |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2442 "movq %%mm1, (%%ebx, %1, 2) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2443 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2444 : : "r" (src), "r" (stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2445 : "%eax", "%ebx" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2446 ); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2447 #else |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2448 int x; |
2595 | 2449 src+= 4*stride; |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2450 for(x=0; x<8; x++) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2451 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2452 src[stride] = (src[0] + src[stride*2])>>1; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2453 src[stride*3] = (src[stride*2] + src[stride*4])>>1; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2454 src[stride*5] = (src[stride*4] + src[stride*6])>>1; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2455 src[stride*7] = (src[stride*6] + src[stride*8])>>1; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2456 src++; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2457 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2458 #endif |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2459 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2460 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2461 /** |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2462 * Deinterlaces the given block |
2595 | 2463 * will be called for every 8x8 block and can read & write from line 4-15 |
2464 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too | |
2465 * lines 4-12 will be read into the deblocking filter and should be deinterlaced | |
2466 * this filter will read lines 3-15 and write 7-13 | |
2246 | 2467 * no cliping in C version |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2468 */ |
2246 | 2469 static inline void deInterlaceInterpolateCubic(uint8_t src[], int stride) |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2470 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2471 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2595 | 2472 src+= stride*3; |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2473 asm volatile( |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2474 "leal (%0, %1), %%eax \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2475 "leal (%%eax, %1, 4), %%ebx \n\t" |
2246 | 2476 "leal (%%ebx, %1, 4), %%ecx \n\t" |
2477 "addl %1, %%ecx \n\t" | |
2478 "pxor %%mm7, %%mm7 \n\t" | |
2479 // 0 1 2 3 4 5 6 7 8 9 10 | |
2480 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 ecx | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2481 |
2246 | 2482 #define DEINT_CUBIC(a,b,c,d,e)\ |
2483 "movq " #a ", %%mm0 \n\t"\ | |
2484 "movq " #b ", %%mm1 \n\t"\ | |
2485 "movq " #d ", %%mm2 \n\t"\ | |
2486 "movq " #e ", %%mm3 \n\t"\ | |
2487 PAVGB(%%mm2, %%mm1) /* (b+d) /2 */\ | |
2488 PAVGB(%%mm3, %%mm0) /* a(a+e) /2 */\ | |
2489 "movq %%mm0, %%mm2 \n\t"\ | |
2490 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
2491 "punpckhbw %%mm7, %%mm2 \n\t"\ | |
2492 "movq %%mm1, %%mm3 \n\t"\ | |
2493 "punpcklbw %%mm7, %%mm1 \n\t"\ | |
2494 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
2495 "psubw %%mm1, %%mm0 \n\t" /* L(a+e - (b+d))/2 */\ | |
2496 "psubw %%mm3, %%mm2 \n\t" /* H(a+e - (b+d))/2 */\ | |
2497 "psraw $3, %%mm0 \n\t" /* L(a+e - (b+d))/16 */\ | |
2498 "psraw $3, %%mm2 \n\t" /* H(a+e - (b+d))/16 */\ | |
2499 "psubw %%mm0, %%mm1 \n\t" /* L(9b + 9d - a - e)/16 */\ | |
2500 "psubw %%mm2, %%mm3 \n\t" /* H(9b + 9d - a - e)/16 */\ | |
2501 "packuswb %%mm3, %%mm1 \n\t"\ | |
2502 "movq %%mm1, " #c " \n\t" | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2503 |
2246 | 2504 DEINT_CUBIC((%0), (%%eax, %1), (%%eax, %1, 2), (%0, %1, 4), (%%ebx, %1)) |
2505 DEINT_CUBIC((%%eax, %1), (%0, %1, 4), (%%ebx), (%%ebx, %1), (%0, %1, 8)) | |
2506 DEINT_CUBIC((%0, %1, 4), (%%ebx, %1), (%%ebx, %1, 2), (%0, %1, 8), (%%ecx)) | |
2507 DEINT_CUBIC((%%ebx, %1), (%0, %1, 8), (%%ebx, %1, 4), (%%ecx), (%%ecx, %1, 2)) | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2508 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2509 : : "r" (src), "r" (stride) |
2246 | 2510 : "%eax", "%ebx", "ecx" |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2511 ); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2512 #else |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2513 int x; |
2595 | 2514 src+= stride*3; |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2515 for(x=0; x<8; x++) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2516 { |
2246 | 2517 src[stride*3] = (-src[0] + 9*src[stride*2] + 9*src[stride*4] - src[stride*6])>>4; |
2518 src[stride*5] = (-src[stride*2] + 9*src[stride*4] + 9*src[stride*6] - src[stride*8])>>4; | |
2519 src[stride*7] = (-src[stride*4] + 9*src[stride*6] + 9*src[stride*8] - src[stride*10])>>4; | |
2520 src[stride*9] = (-src[stride*6] + 9*src[stride*8] + 9*src[stride*10] - src[stride*12])>>4; | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2521 src++; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2522 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2523 #endif |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2524 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2525 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2526 /** |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2527 * Deinterlaces the given block |
2595 | 2528 * will be called for every 8x8 block and can read & write from line 4-15 |
2529 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too | |
2530 * lines 4-12 will be read into the deblocking filter and should be deinterlaced | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2531 * will shift the image up by 1 line (FIXME if this is a problem) |
2595 | 2532 * this filter will read lines 4-13 and write 4-11 |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2533 */ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2534 static inline void deInterlaceBlendLinear(uint8_t src[], int stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2535 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2536 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2595 | 2537 src+= 4*stride; |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2538 asm volatile( |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2539 "leal (%0, %1), %%eax \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2540 "leal (%%eax, %1, 4), %%ebx \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2541 // 0 1 2 3 4 5 6 7 8 9 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2542 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2543 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2544 "movq (%0), %%mm0 \n\t" // L0 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2545 "movq (%%eax, %1), %%mm1 \n\t" // L2 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2546 PAVGB(%%mm1, %%mm0) // L0+L2 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2547 "movq (%%eax), %%mm2 \n\t" // L1 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2548 PAVGB(%%mm2, %%mm0) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2549 "movq %%mm0, (%0) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2550 "movq (%%eax, %1, 2), %%mm0 \n\t" // L3 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2551 PAVGB(%%mm0, %%mm2) // L1+L3 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2552 PAVGB(%%mm1, %%mm2) // 2L2 + L1 + L3 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2553 "movq %%mm2, (%%eax) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2554 "movq (%0, %1, 4), %%mm2 \n\t" // L4 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2555 PAVGB(%%mm2, %%mm1) // L2+L4 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2556 PAVGB(%%mm0, %%mm1) // 2L3 + L2 + L4 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2557 "movq %%mm1, (%%eax, %1) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2558 "movq (%%ebx), %%mm1 \n\t" // L5 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2559 PAVGB(%%mm1, %%mm0) // L3+L5 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2560 PAVGB(%%mm2, %%mm0) // 2L4 + L3 + L5 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2561 "movq %%mm0, (%%eax, %1, 2) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2562 "movq (%%ebx, %1), %%mm0 \n\t" // L6 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2563 PAVGB(%%mm0, %%mm2) // L4+L6 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2564 PAVGB(%%mm1, %%mm2) // 2L5 + L4 + L6 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2565 "movq %%mm2, (%0, %1, 4) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2566 "movq (%%ebx, %1, 2), %%mm2 \n\t" // L7 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2567 PAVGB(%%mm2, %%mm1) // L5+L7 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2568 PAVGB(%%mm0, %%mm1) // 2L6 + L5 + L7 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2569 "movq %%mm1, (%%ebx) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2570 "movq (%0, %1, 8), %%mm1 \n\t" // L8 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2571 PAVGB(%%mm1, %%mm0) // L6+L8 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2572 PAVGB(%%mm2, %%mm0) // 2L7 + L6 + L8 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2573 "movq %%mm0, (%%ebx, %1) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2574 "movq (%%ebx, %1, 4), %%mm0 \n\t" // L9 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2575 PAVGB(%%mm0, %%mm2) // L7+L9 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2576 PAVGB(%%mm1, %%mm2) // 2L8 + L7 + L9 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2577 "movq %%mm2, (%%ebx, %1, 2) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2578 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2579 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2580 : : "r" (src), "r" (stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2581 : "%eax", "%ebx" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2582 ); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2583 #else |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2584 int x; |
2595 | 2585 src+= 4*stride; |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2586 for(x=0; x<8; x++) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2587 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2588 src[0 ] = (src[0 ] + 2*src[stride ] + src[stride*2])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2589 src[stride ] = (src[stride ] + 2*src[stride*2] + src[stride*3])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2590 src[stride*2] = (src[stride*2] + 2*src[stride*3] + src[stride*4])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2591 src[stride*3] = (src[stride*3] + 2*src[stride*4] + src[stride*5])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2592 src[stride*4] = (src[stride*4] + 2*src[stride*5] + src[stride*6])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2593 src[stride*5] = (src[stride*5] + 2*src[stride*6] + src[stride*7])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2594 src[stride*6] = (src[stride*6] + 2*src[stride*7] + src[stride*8])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2595 src[stride*7] = (src[stride*7] + 2*src[stride*8] + src[stride*9])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2596 src++; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2597 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2598 #endif |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2599 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2600 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2601 /** |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2602 * Deinterlaces the given block |
2595 | 2603 * will be called for every 8x8 block and can read & write from line 4-15, |
2604 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too | |
2605 * lines 4-12 will be read into the deblocking filter and should be deinterlaced | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2606 */ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2607 static inline void deInterlaceMedian(uint8_t src[], int stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2608 { |
2221 | 2609 #ifdef HAVE_MMX |
2595 | 2610 src+= 4*stride; |
2221 | 2611 #ifdef HAVE_MMX2 |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2612 asm volatile( |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2613 "leal (%0, %1), %%eax \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2614 "leal (%%eax, %1, 4), %%ebx \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2615 // 0 1 2 3 4 5 6 7 8 9 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2616 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2617 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2618 "movq (%0), %%mm0 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2619 "movq (%%eax, %1), %%mm2 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2620 "movq (%%eax), %%mm1 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2621 "movq %%mm0, %%mm3 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2622 "pmaxub %%mm1, %%mm0 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2623 "pminub %%mm3, %%mm1 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2624 "pmaxub %%mm2, %%mm1 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2625 "pminub %%mm1, %%mm0 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2626 "movq %%mm0, (%%eax) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2627 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2628 "movq (%0, %1, 4), %%mm0 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2629 "movq (%%eax, %1, 2), %%mm1 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2630 "movq %%mm2, %%mm3 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2631 "pmaxub %%mm1, %%mm2 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2632 "pminub %%mm3, %%mm1 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2633 "pmaxub %%mm0, %%mm1 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2634 "pminub %%mm1, %%mm2 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2635 "movq %%mm2, (%%eax, %1, 2) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2636 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2637 "movq (%%ebx), %%mm2 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2638 "movq (%%ebx, %1), %%mm1 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2639 "movq %%mm2, %%mm3 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2640 "pmaxub %%mm0, %%mm2 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2641 "pminub %%mm3, %%mm0 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2642 "pmaxub %%mm1, %%mm0 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2643 "pminub %%mm0, %%mm2 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2644 "movq %%mm2, (%%ebx) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2645 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2646 "movq (%%ebx, %1, 2), %%mm2 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2647 "movq (%0, %1, 8), %%mm0 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2648 "movq %%mm2, %%mm3 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2649 "pmaxub %%mm0, %%mm2 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2650 "pminub %%mm3, %%mm0 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2651 "pmaxub %%mm1, %%mm0 \n\t" // |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2652 "pminub %%mm0, %%mm2 \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2653 "movq %%mm2, (%%ebx, %1, 2) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2654 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2655 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2656 : : "r" (src), "r" (stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2657 : "%eax", "%ebx" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2658 ); |
2221 | 2659 |
2660 #else // MMX without MMX2 | |
2661 asm volatile( | |
2662 "leal (%0, %1), %%eax \n\t" | |
2663 "leal (%%eax, %1, 4), %%ebx \n\t" | |
2664 // 0 1 2 3 4 5 6 7 8 9 | |
2665 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 | |
2666 "pxor %%mm7, %%mm7 \n\t" | |
2667 | |
2668 #define MEDIAN(a,b,c)\ | |
2669 "movq " #a ", %%mm0 \n\t"\ | |
2670 "movq " #b ", %%mm2 \n\t"\ | |
2671 "movq " #c ", %%mm1 \n\t"\ | |
2672 "movq %%mm0, %%mm3 \n\t"\ | |
2673 "movq %%mm1, %%mm4 \n\t"\ | |
2674 "movq %%mm2, %%mm5 \n\t"\ | |
2675 "psubusb %%mm1, %%mm3 \n\t"\ | |
2676 "psubusb %%mm2, %%mm4 \n\t"\ | |
2677 "psubusb %%mm0, %%mm5 \n\t"\ | |
2678 "pcmpeqb %%mm7, %%mm3 \n\t"\ | |
2679 "pcmpeqb %%mm7, %%mm4 \n\t"\ | |
2680 "pcmpeqb %%mm7, %%mm5 \n\t"\ | |
2681 "movq %%mm3, %%mm6 \n\t"\ | |
2682 "pxor %%mm4, %%mm3 \n\t"\ | |
2683 "pxor %%mm5, %%mm4 \n\t"\ | |
2684 "pxor %%mm6, %%mm5 \n\t"\ | |
2685 "por %%mm3, %%mm1 \n\t"\ | |
2686 "por %%mm4, %%mm2 \n\t"\ | |
2687 "por %%mm5, %%mm0 \n\t"\ | |
2688 "pand %%mm2, %%mm0 \n\t"\ | |
2689 "pand %%mm1, %%mm0 \n\t"\ | |
2690 "movq %%mm0, " #b " \n\t" | |
2691 | |
2692 MEDIAN((%0), (%%eax), (%%eax, %1)) | |
2693 MEDIAN((%%eax, %1), (%%eax, %1, 2), (%0, %1, 4)) | |
2694 MEDIAN((%0, %1, 4), (%%ebx), (%%ebx, %1)) | |
2695 MEDIAN((%%ebx, %1), (%%ebx, %1, 2), (%0, %1, 8)) | |
2696 | |
2697 : : "r" (src), "r" (stride) | |
2698 : "%eax", "%ebx" | |
2699 ); | |
2700 #endif // MMX | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2701 #else |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2702 //FIXME |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2703 int x; |
2595 | 2704 src+= 4*stride; |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2705 for(x=0; x<8; x++) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2706 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2707 src[0 ] = (src[0 ] + 2*src[stride ] + src[stride*2])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2708 src[stride ] = (src[stride ] + 2*src[stride*2] + src[stride*3])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2709 src[stride*2] = (src[stride*2] + 2*src[stride*3] + src[stride*4])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2710 src[stride*3] = (src[stride*3] + 2*src[stride*4] + src[stride*5])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2711 src[stride*4] = (src[stride*4] + 2*src[stride*5] + src[stride*6])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2712 src[stride*5] = (src[stride*5] + 2*src[stride*6] + src[stride*7])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2713 src[stride*6] = (src[stride*6] + 2*src[stride*7] + src[stride*8])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2714 src[stride*7] = (src[stride*7] + 2*src[stride*8] + src[stride*9])>>2; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2715 src++; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2716 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2717 #endif |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2718 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2719 |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2720 #ifdef HAVE_MMX |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2721 /** |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2722 * transposes and shift the given 8x8 Block into dst1 and dst2 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2723 */ |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2724 static inline void transpose1(uint8_t *dst1, uint8_t *dst2, uint8_t *src, int srcStride) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2725 { |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2726 asm( |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2727 "leal (%0, %1), %%eax \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2728 "leal (%%eax, %1, 4), %%ebx \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2729 // 0 1 2 3 4 5 6 7 8 9 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2730 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2731 "movq (%0), %%mm0 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2732 "movq (%%eax), %%mm1 \n\t" // abcdefgh |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2733 "movq %%mm0, %%mm2 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2734 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2735 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2736 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2737 "movq (%%eax, %1), %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2738 "movq (%%eax, %1, 2), %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2739 "movq %%mm1, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2740 "punpcklbw %%mm3, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2741 "punpckhbw %%mm3, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2742 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2743 "movq %%mm0, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2744 "punpcklwd %%mm1, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2745 "punpckhwd %%mm1, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2746 "movq %%mm2, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2747 "punpcklwd %%mm4, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2748 "punpckhwd %%mm4, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2749 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2750 "movd %%mm0, 128(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2751 "psrlq $32, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2752 "movd %%mm0, 144(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2753 "movd %%mm3, 160(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2754 "psrlq $32, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2755 "movd %%mm3, 176(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2756 "movd %%mm3, 48(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2757 "movd %%mm2, 192(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2758 "movd %%mm2, 64(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2759 "psrlq $32, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2760 "movd %%mm2, 80(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2761 "movd %%mm1, 96(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2762 "psrlq $32, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2763 "movd %%mm1, 112(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2764 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2765 "movq (%0, %1, 4), %%mm0 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2766 "movq (%%ebx), %%mm1 \n\t" // abcdefgh |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2767 "movq %%mm0, %%mm2 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2768 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2769 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2770 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2771 "movq (%%ebx, %1), %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2772 "movq (%%ebx, %1, 2), %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2773 "movq %%mm1, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2774 "punpcklbw %%mm3, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2775 "punpckhbw %%mm3, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2776 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2777 "movq %%mm0, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2778 "punpcklwd %%mm1, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2779 "punpckhwd %%mm1, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2780 "movq %%mm2, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2781 "punpcklwd %%mm4, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2782 "punpckhwd %%mm4, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2783 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2784 "movd %%mm0, 132(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2785 "psrlq $32, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2786 "movd %%mm0, 148(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2787 "movd %%mm3, 164(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2788 "psrlq $32, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2789 "movd %%mm3, 180(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2790 "movd %%mm3, 52(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2791 "movd %%mm2, 196(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2792 "movd %%mm2, 68(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2793 "psrlq $32, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2794 "movd %%mm2, 84(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2795 "movd %%mm1, 100(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2796 "psrlq $32, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2797 "movd %%mm1, 116(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2798 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2799 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2800 :: "r" (src), "r" (srcStride), "r" (dst1), "r" (dst2) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2801 : "%eax", "%ebx" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2802 ); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2803 } |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2804 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2805 /** |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2806 * transposes the given 8x8 block |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2807 */ |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2808 static inline void transpose2(uint8_t *dst, int dstStride, uint8_t *src) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2809 { |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2810 asm( |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2811 "leal (%0, %1), %%eax \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2812 "leal (%%eax, %1, 4), %%ebx \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2813 // 0 1 2 3 4 5 6 7 8 9 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2814 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2815 "movq (%2), %%mm0 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2816 "movq 16(%2), %%mm1 \n\t" // abcdefgh |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2817 "movq %%mm0, %%mm2 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2818 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2819 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2820 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2821 "movq 32(%2), %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2822 "movq 48(%2), %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2823 "movq %%mm1, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2824 "punpcklbw %%mm3, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2825 "punpckhbw %%mm3, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2826 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2827 "movq %%mm0, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2828 "punpcklwd %%mm1, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2829 "punpckhwd %%mm1, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2830 "movq %%mm2, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2831 "punpcklwd %%mm4, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2832 "punpckhwd %%mm4, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2833 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2834 "movd %%mm0, (%0) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2835 "psrlq $32, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2836 "movd %%mm0, (%%eax) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2837 "movd %%mm3, (%%eax, %1) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2838 "psrlq $32, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2839 "movd %%mm3, (%%eax, %1, 2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2840 "movd %%mm2, (%0, %1, 4) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2841 "psrlq $32, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2842 "movd %%mm2, (%%ebx) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2843 "movd %%mm1, (%%ebx, %1) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2844 "psrlq $32, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2845 "movd %%mm1, (%%ebx, %1, 2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2846 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2847 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2848 "movq 64(%2), %%mm0 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2849 "movq 80(%2), %%mm1 \n\t" // abcdefgh |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2850 "movq %%mm0, %%mm2 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2851 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2852 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2853 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2854 "movq 96(%2), %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2855 "movq 112(%2), %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2856 "movq %%mm1, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2857 "punpcklbw %%mm3, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2858 "punpckhbw %%mm3, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2859 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2860 "movq %%mm0, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2861 "punpcklwd %%mm1, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2862 "punpckhwd %%mm1, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2863 "movq %%mm2, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2864 "punpcklwd %%mm4, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2865 "punpckhwd %%mm4, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2866 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2867 "movd %%mm0, 4(%0) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2868 "psrlq $32, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2869 "movd %%mm0, 4(%%eax) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2870 "movd %%mm3, 4(%%eax, %1) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2871 "psrlq $32, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2872 "movd %%mm3, 4(%%eax, %1, 2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2873 "movd %%mm2, 4(%0, %1, 4) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2874 "psrlq $32, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2875 "movd %%mm2, 4(%%ebx) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2876 "movd %%mm1, 4(%%ebx, %1) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2877 "psrlq $32, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2878 "movd %%mm1, 4(%%ebx, %1, 2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2879 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2880 :: "r" (dst), "r" (dstStride), "r" (src) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2881 : "%eax", "%ebx" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2882 ); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2883 } |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2884 #endif |
2895
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2885 //static int test=0; |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2886 |
2860 | 2887 static void inline tempNoiseReducer(uint8_t *src, int stride, |
2899 | 2888 uint8_t *tempBlured, uint32_t *tempBluredPast, int *maxNoise) |
2860 | 2889 { |
2895
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2890 #define FAST_L2_DIFF |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2891 //#define L1_DIFF //u should change the thresholds too if u try that one |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2892 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2893 asm volatile( |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2894 "leal (%2, %2, 2), %%eax \n\t" // 3*stride |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2895 "leal (%2, %2, 4), %%ebx \n\t" // 5*stride |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2896 "leal (%%ebx, %2, 2), %%ecx \n\t" // 7*stride |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2897 // 0 1 2 3 4 5 6 7 8 9 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2898 // %x %x+%2 %x+2%2 %x+eax %x+4%2 %x+ebx %x+2eax %x+ecx %x+8%2 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2899 //FIXME reorder? |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2900 #ifdef L1_DIFF //needs mmx2 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2901 "movq (%0), %%mm0 \n\t" // L0 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2902 "psadbw (%1), %%mm0 \n\t" // |L0-R0| |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2903 "movq (%0, %2), %%mm1 \n\t" // L1 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2904 "psadbw (%1, %2), %%mm1 \n\t" // |L1-R1| |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2905 "movq (%0, %2, 2), %%mm2 \n\t" // L2 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2906 "psadbw (%1, %2, 2), %%mm2 \n\t" // |L2-R2| |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2907 "movq (%0, %%eax), %%mm3 \n\t" // L3 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2908 "psadbw (%1, %%eax), %%mm3 \n\t" // |L3-R3| |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2909 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2910 "movq (%0, %2, 4), %%mm4 \n\t" // L4 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2911 "paddw %%mm1, %%mm0 \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2912 "psadbw (%1, %2, 4), %%mm4 \n\t" // |L4-R4| |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2913 "movq (%0, %%ebx), %%mm5 \n\t" // L5 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2914 "paddw %%mm2, %%mm0 \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2915 "psadbw (%1, %%ebx), %%mm5 \n\t" // |L5-R5| |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2916 "movq (%0, %%eax, 2), %%mm6 \n\t" // L6 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2917 "paddw %%mm3, %%mm0 \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2918 "psadbw (%1, %%eax, 2), %%mm6 \n\t" // |L6-R6| |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2919 "movq (%0, %%ecx), %%mm7 \n\t" // L7 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2920 "paddw %%mm4, %%mm0 \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2921 "psadbw (%1, %%ecx), %%mm7 \n\t" // |L7-R7| |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2922 "paddw %%mm5, %%mm6 \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2923 "paddw %%mm7, %%mm6 \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2924 "paddw %%mm6, %%mm0 \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2925 #elif defined (FAST_L2_DIFF) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2926 "pcmpeqb %%mm7, %%mm7 \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2927 "movq b80, %%mm6 \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2928 "pxor %%mm0, %%mm0 \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2929 #define L2_DIFF_CORE(a, b)\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2930 "movq " #a ", %%mm5 \n\t"\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2931 "movq " #b ", %%mm2 \n\t"\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2932 "pxor %%mm7, %%mm2 \n\t"\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2933 PAVGB(%%mm2, %%mm5)\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2934 "paddb %%mm6, %%mm5 \n\t"\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2935 "movq %%mm5, %%mm2 \n\t"\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2936 "psllw $8, %%mm5 \n\t"\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2937 "pmaddwd %%mm5, %%mm5 \n\t"\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2938 "pmaddwd %%mm2, %%mm2 \n\t"\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2939 "paddd %%mm2, %%mm5 \n\t"\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2940 "psrld $14, %%mm5 \n\t"\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2941 "paddd %%mm5, %%mm0 \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2942 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2943 L2_DIFF_CORE((%0), (%1)) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2944 L2_DIFF_CORE((%0, %2), (%1, %2)) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2945 L2_DIFF_CORE((%0, %2, 2), (%1, %2, 2)) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2946 L2_DIFF_CORE((%0, %%eax), (%1, %%eax)) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2947 L2_DIFF_CORE((%0, %2, 4), (%1, %2, 4)) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2948 L2_DIFF_CORE((%0, %%ebx), (%1, %%ebx)) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2949 L2_DIFF_CORE((%0, %%eax,2), (%1, %%eax,2)) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2950 L2_DIFF_CORE((%0, %%ecx), (%1, %%ecx)) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2951 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2952 #else |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2953 "pxor %%mm7, %%mm7 \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2954 "pxor %%mm0, %%mm0 \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2955 #define L2_DIFF_CORE(a, b)\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2956 "movq " #a ", %%mm5 \n\t"\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2957 "movq " #b ", %%mm2 \n\t"\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2958 "movq %%mm5, %%mm1 \n\t"\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2959 "movq %%mm2, %%mm3 \n\t"\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2960 "punpcklbw %%mm7, %%mm5 \n\t"\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2961 "punpckhbw %%mm7, %%mm1 \n\t"\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2962 "punpcklbw %%mm7, %%mm2 \n\t"\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2963 "punpckhbw %%mm7, %%mm3 \n\t"\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2964 "psubw %%mm2, %%mm5 \n\t"\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2965 "psubw %%mm3, %%mm1 \n\t"\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2966 "pmaddwd %%mm5, %%mm5 \n\t"\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2967 "pmaddwd %%mm1, %%mm1 \n\t"\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2968 "paddd %%mm1, %%mm5 \n\t"\ |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2969 "paddd %%mm5, %%mm0 \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2970 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2971 L2_DIFF_CORE((%0), (%1)) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2972 L2_DIFF_CORE((%0, %2), (%1, %2)) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2973 L2_DIFF_CORE((%0, %2, 2), (%1, %2, 2)) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2974 L2_DIFF_CORE((%0, %%eax), (%1, %%eax)) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2975 L2_DIFF_CORE((%0, %2, 4), (%1, %2, 4)) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2976 L2_DIFF_CORE((%0, %%ebx), (%1, %%ebx)) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2977 L2_DIFF_CORE((%0, %%eax,2), (%1, %%eax,2)) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2978 L2_DIFF_CORE((%0, %%ecx), (%1, %%ecx)) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2979 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2980 #endif |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2981 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2982 "movq %%mm0, %%mm4 \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2983 "psrlq $32, %%mm0 \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2984 "paddd %%mm0, %%mm4 \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2985 "movd %%mm4, %%ecx \n\t" |
2899 | 2986 "shll $2, %%ecx \n\t" |
2987 "movl %3, %%ebx \n\t" | |
2988 "addl -4(%%ebx), %%ecx \n\t" | |
2989 "addl 4(%%ebx), %%ecx \n\t" | |
2990 "addl -1024(%%ebx), %%ecx \n\t" | |
2991 "addl $4, %%ecx \n\t" | |
2992 "addl 1024(%%ebx), %%ecx \n\t" | |
2993 "shrl $3, %%ecx \n\t" | |
2994 "movl %%ecx, (%%ebx) \n\t" | |
2995 "leal (%%eax, %2, 2), %%ebx \n\t" // 5*stride | |
2996 | |
2895
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2997 // "movl %3, %%ecx \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2998 // "movl %%ecx, test \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
2999 // "jmp 4f \n\t" |
2899 | 3000 "cmpl 4+maxTmpNoise, %%ecx \n\t" |
2895
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3001 " jb 2f \n\t" |
2899 | 3002 "cmpl 8+maxTmpNoise, %%ecx \n\t" |
2895
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3003 " jb 1f \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3004 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3005 "leal (%%ebx, %2, 2), %%ecx \n\t" // 7*stride |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3006 "movq (%0), %%mm0 \n\t" // L0 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3007 "movq (%0, %2), %%mm1 \n\t" // L1 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3008 "movq (%0, %2, 2), %%mm2 \n\t" // L2 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3009 "movq (%0, %%eax), %%mm3 \n\t" // L3 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3010 "movq (%0, %2, 4), %%mm4 \n\t" // L4 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3011 "movq (%0, %%ebx), %%mm5 \n\t" // L5 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3012 "movq (%0, %%eax, 2), %%mm6 \n\t" // L6 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3013 "movq (%0, %%ecx), %%mm7 \n\t" // L7 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3014 "movq %%mm0, (%1) \n\t" // L0 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3015 "movq %%mm1, (%1, %2) \n\t" // L1 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3016 "movq %%mm2, (%1, %2, 2) \n\t" // L2 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3017 "movq %%mm3, (%1, %%eax) \n\t" // L3 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3018 "movq %%mm4, (%1, %2, 4) \n\t" // L4 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3019 "movq %%mm5, (%1, %%ebx) \n\t" // L5 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3020 "movq %%mm6, (%1, %%eax, 2) \n\t" // L6 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3021 "movq %%mm7, (%1, %%ecx) \n\t" // L7 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3022 "jmp 4f \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3023 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3024 "1: \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3025 "leal (%%ebx, %2, 2), %%ecx \n\t" // 7*stride |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3026 "movq (%0), %%mm0 \n\t" // L0 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3027 "pavgb (%1), %%mm0 \n\t" // L0 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3028 "movq (%0, %2), %%mm1 \n\t" // L1 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3029 "pavgb (%1, %2), %%mm1 \n\t" // L1 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3030 "movq (%0, %2, 2), %%mm2 \n\t" // L2 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3031 "pavgb (%1, %2, 2), %%mm2 \n\t" // L2 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3032 "movq (%0, %%eax), %%mm3 \n\t" // L3 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3033 "pavgb (%1, %%eax), %%mm3 \n\t" // L3 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3034 "movq (%0, %2, 4), %%mm4 \n\t" // L4 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3035 "pavgb (%1, %2, 4), %%mm4 \n\t" // L4 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3036 "movq (%0, %%ebx), %%mm5 \n\t" // L5 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3037 "pavgb (%1, %%ebx), %%mm5 \n\t" // L5 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3038 "movq (%0, %%eax, 2), %%mm6 \n\t" // L6 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3039 "pavgb (%1, %%eax, 2), %%mm6 \n\t" // L6 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3040 "movq (%0, %%ecx), %%mm7 \n\t" // L7 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3041 "pavgb (%1, %%ecx), %%mm7 \n\t" // L7 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3042 "movq %%mm0, (%1) \n\t" // R0 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3043 "movq %%mm1, (%1, %2) \n\t" // R1 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3044 "movq %%mm2, (%1, %2, 2) \n\t" // R2 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3045 "movq %%mm3, (%1, %%eax) \n\t" // R3 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3046 "movq %%mm4, (%1, %2, 4) \n\t" // R4 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3047 "movq %%mm5, (%1, %%ebx) \n\t" // R5 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3048 "movq %%mm6, (%1, %%eax, 2) \n\t" // R6 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3049 "movq %%mm7, (%1, %%ecx) \n\t" // R7 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3050 "movq %%mm0, (%0) \n\t" // L0 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3051 "movq %%mm1, (%0, %2) \n\t" // L1 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3052 "movq %%mm2, (%0, %2, 2) \n\t" // L2 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3053 "movq %%mm3, (%0, %%eax) \n\t" // L3 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3054 "movq %%mm4, (%0, %2, 4) \n\t" // L4 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3055 "movq %%mm5, (%0, %%ebx) \n\t" // L5 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3056 "movq %%mm6, (%0, %%eax, 2) \n\t" // L6 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3057 "movq %%mm7, (%0, %%ecx) \n\t" // L7 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3058 "jmp 4f \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3059 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3060 "2: \n\t" |
2899 | 3061 "cmpl maxTmpNoise, %%ecx \n\t" |
2895
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3062 " jb 3f \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3063 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3064 "leal (%%ebx, %2, 2), %%ecx \n\t" // 7*stride |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3065 "movq (%0), %%mm0 \n\t" // L0 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3066 "movq (%0, %2), %%mm1 \n\t" // L1 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3067 "movq (%0, %2, 2), %%mm2 \n\t" // L2 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3068 "movq (%0, %%eax), %%mm3 \n\t" // L3 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3069 "movq (%1), %%mm4 \n\t" // R0 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3070 "movq (%1, %2), %%mm5 \n\t" // R1 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3071 "movq (%1, %2, 2), %%mm6 \n\t" // R2 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3072 "movq (%1, %%eax), %%mm7 \n\t" // R3 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3073 PAVGB(%%mm4, %%mm0) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3074 PAVGB(%%mm5, %%mm1) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3075 PAVGB(%%mm6, %%mm2) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3076 PAVGB(%%mm7, %%mm3) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3077 PAVGB(%%mm4, %%mm0) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3078 PAVGB(%%mm5, %%mm1) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3079 PAVGB(%%mm6, %%mm2) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3080 PAVGB(%%mm7, %%mm3) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3081 "movq %%mm0, (%1) \n\t" // R0 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3082 "movq %%mm1, (%1, %2) \n\t" // R1 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3083 "movq %%mm2, (%1, %2, 2) \n\t" // R2 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3084 "movq %%mm3, (%1, %%eax) \n\t" // R3 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3085 "movq %%mm0, (%0) \n\t" // L0 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3086 "movq %%mm1, (%0, %2) \n\t" // L1 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3087 "movq %%mm2, (%0, %2, 2) \n\t" // L2 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3088 "movq %%mm3, (%0, %%eax) \n\t" // L3 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3089 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3090 "movq (%0, %2, 4), %%mm0 \n\t" // L4 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3091 "movq (%0, %%ebx), %%mm1 \n\t" // L5 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3092 "movq (%0, %%eax, 2), %%mm2 \n\t" // L6 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3093 "movq (%0, %%ecx), %%mm3 \n\t" // L7 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3094 "movq (%1, %2, 4), %%mm4 \n\t" // R4 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3095 "movq (%1, %%ebx), %%mm5 \n\t" // R5 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3096 "movq (%1, %%eax, 2), %%mm6 \n\t" // R6 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3097 "movq (%1, %%ecx), %%mm7 \n\t" // R7 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3098 PAVGB(%%mm4, %%mm0) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3099 PAVGB(%%mm5, %%mm1) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3100 PAVGB(%%mm6, %%mm2) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3101 PAVGB(%%mm7, %%mm3) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3102 PAVGB(%%mm4, %%mm0) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3103 PAVGB(%%mm5, %%mm1) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3104 PAVGB(%%mm6, %%mm2) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3105 PAVGB(%%mm7, %%mm3) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3106 "movq %%mm0, (%1, %2, 4) \n\t" // R4 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3107 "movq %%mm1, (%1, %%ebx) \n\t" // R5 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3108 "movq %%mm2, (%1, %%eax, 2) \n\t" // R6 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3109 "movq %%mm3, (%1, %%ecx) \n\t" // R7 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3110 "movq %%mm0, (%0, %2, 4) \n\t" // L4 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3111 "movq %%mm1, (%0, %%ebx) \n\t" // L5 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3112 "movq %%mm2, (%0, %%eax, 2) \n\t" // L6 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3113 "movq %%mm3, (%0, %%ecx) \n\t" // L7 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3114 "jmp 4f \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3115 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3116 "3: \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3117 "leal (%%ebx, %2, 2), %%ecx \n\t" // 7*stride |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3118 "movq (%0), %%mm0 \n\t" // L0 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3119 "movq (%0, %2), %%mm1 \n\t" // L1 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3120 "movq (%0, %2, 2), %%mm2 \n\t" // L2 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3121 "movq (%0, %%eax), %%mm3 \n\t" // L3 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3122 "movq (%1), %%mm4 \n\t" // R0 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3123 "movq (%1, %2), %%mm5 \n\t" // R1 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3124 "movq (%1, %2, 2), %%mm6 \n\t" // R2 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3125 "movq (%1, %%eax), %%mm7 \n\t" // R3 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3126 PAVGB(%%mm4, %%mm0) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3127 PAVGB(%%mm5, %%mm1) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3128 PAVGB(%%mm6, %%mm2) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3129 PAVGB(%%mm7, %%mm3) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3130 PAVGB(%%mm4, %%mm0) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3131 PAVGB(%%mm5, %%mm1) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3132 PAVGB(%%mm6, %%mm2) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3133 PAVGB(%%mm7, %%mm3) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3134 PAVGB(%%mm4, %%mm0) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3135 PAVGB(%%mm5, %%mm1) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3136 PAVGB(%%mm6, %%mm2) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3137 PAVGB(%%mm7, %%mm3) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3138 "movq %%mm0, (%1) \n\t" // R0 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3139 "movq %%mm1, (%1, %2) \n\t" // R1 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3140 "movq %%mm2, (%1, %2, 2) \n\t" // R2 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3141 "movq %%mm3, (%1, %%eax) \n\t" // R3 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3142 "movq %%mm0, (%0) \n\t" // L0 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3143 "movq %%mm1, (%0, %2) \n\t" // L1 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3144 "movq %%mm2, (%0, %2, 2) \n\t" // L2 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3145 "movq %%mm3, (%0, %%eax) \n\t" // L3 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3146 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3147 "movq (%0, %2, 4), %%mm0 \n\t" // L4 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3148 "movq (%0, %%ebx), %%mm1 \n\t" // L5 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3149 "movq (%0, %%eax, 2), %%mm2 \n\t" // L6 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3150 "movq (%0, %%ecx), %%mm3 \n\t" // L7 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3151 "movq (%1, %2, 4), %%mm4 \n\t" // R4 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3152 "movq (%1, %%ebx), %%mm5 \n\t" // R5 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3153 "movq (%1, %%eax, 2), %%mm6 \n\t" // R6 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3154 "movq (%1, %%ecx), %%mm7 \n\t" // R7 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3155 PAVGB(%%mm4, %%mm0) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3156 PAVGB(%%mm5, %%mm1) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3157 PAVGB(%%mm6, %%mm2) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3158 PAVGB(%%mm7, %%mm3) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3159 PAVGB(%%mm4, %%mm0) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3160 PAVGB(%%mm5, %%mm1) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3161 PAVGB(%%mm6, %%mm2) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3162 PAVGB(%%mm7, %%mm3) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3163 PAVGB(%%mm4, %%mm0) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3164 PAVGB(%%mm5, %%mm1) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3165 PAVGB(%%mm6, %%mm2) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3166 PAVGB(%%mm7, %%mm3) |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3167 "movq %%mm0, (%1, %2, 4) \n\t" // R4 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3168 "movq %%mm1, (%1, %%ebx) \n\t" // R5 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3169 "movq %%mm2, (%1, %%eax, 2) \n\t" // R6 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3170 "movq %%mm3, (%1, %%ecx) \n\t" // R7 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3171 "movq %%mm0, (%0, %2, 4) \n\t" // L4 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3172 "movq %%mm1, (%0, %%ebx) \n\t" // L5 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3173 "movq %%mm2, (%0, %%eax, 2) \n\t" // L6 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3174 "movq %%mm3, (%0, %%ecx) \n\t" // L7 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3175 |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3176 "4: \n\t" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3177 |
2899 | 3178 :: "r" (src), "r" (tempBlured), "r"(stride), "m" (tempBluredPast) |
2895
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3179 : "%eax", "%ebx", "%ecx", "memory" |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3180 ); |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3181 //printf("%d\n", test); |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3182 #else |
2860 | 3183 int y; |
3184 int d=0; | |
3185 int sysd=0; | |
2899 | 3186 int i; |
2860 | 3187 |
3188 for(y=0; y<8; y++) | |
3189 { | |
3190 int x; | |
3191 for(x=0; x<8; x++) | |
3192 { | |
3193 int ref= tempBlured[ x + y*stride ]; | |
3194 int cur= src[ x + y*stride ]; | |
3195 int d1=ref - cur; | |
2895
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3196 // if(x==0 || x==7) d1+= d1>>1; |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3197 // if(y==0 || y==7) d1+= d1>>1; |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3198 // d+= ABS(d1); |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3199 d+= d1*d1; |
2860 | 3200 sysd+= d1; |
3201 } | |
3202 } | |
2899 | 3203 i=d; |
3204 d= ( | |
3205 4*d | |
3206 +(*(tempBluredPast-256)) | |
3207 +(*(tempBluredPast-1))+ (*(tempBluredPast+1)) | |
3208 +(*(tempBluredPast+256)) | |
3209 +4)>>3; | |
3210 *tempBluredPast=i; | |
3211 // ((*tempBluredPast)*3 + d + 2)>>2; | |
3212 | |
2860 | 3213 //printf("%d %d %d\n", maxNoise[0], maxNoise[1], maxNoise[2]); |
3214 /* | |
3215 Switch between | |
3216 1 0 0 0 0 0 0 (0) | |
3217 64 32 16 8 4 2 1 (1) | |
3218 64 48 36 27 20 15 11 (33) (approx) | |
3219 64 56 49 43 37 33 29 (200) (approx) | |
3220 */ | |
3221 if(d > maxNoise[1]) | |
3222 { | |
3223 if(d < maxNoise[2]) | |
3224 { | |
3225 for(y=0; y<8; y++) | |
3226 { | |
3227 int x; | |
3228 for(x=0; x<8; x++) | |
3229 { | |
3230 int ref= tempBlured[ x + y*stride ]; | |
3231 int cur= src[ x + y*stride ]; | |
3232 tempBlured[ x + y*stride ]= | |
3233 src[ x + y*stride ]= | |
3234 (ref + cur + 1)>>1; | |
3235 } | |
3236 } | |
3237 } | |
3238 else | |
3239 { | |
3240 for(y=0; y<8; y++) | |
3241 { | |
3242 int x; | |
3243 for(x=0; x<8; x++) | |
3244 { | |
3245 tempBlured[ x + y*stride ]= src[ x + y*stride ]; | |
3246 } | |
3247 } | |
3248 } | |
3249 } | |
3250 else | |
3251 { | |
3252 if(d < maxNoise[0]) | |
3253 { | |
3254 for(y=0; y<8; y++) | |
3255 { | |
3256 int x; | |
3257 for(x=0; x<8; x++) | |
3258 { | |
3259 int ref= tempBlured[ x + y*stride ]; | |
3260 int cur= src[ x + y*stride ]; | |
3261 tempBlured[ x + y*stride ]= | |
3262 src[ x + y*stride ]= | |
3263 (ref*7 + cur + 4)>>3; | |
3264 } | |
3265 } | |
3266 } | |
3267 else | |
3268 { | |
3269 for(y=0; y<8; y++) | |
3270 { | |
3271 int x; | |
3272 for(x=0; x<8; x++) | |
3273 { | |
3274 int ref= tempBlured[ x + y*stride ]; | |
3275 int cur= src[ x + y*stride ]; | |
3276 tempBlured[ x + y*stride ]= | |
3277 src[ x + y*stride ]= | |
3278 (ref*3 + cur + 2)>>2; | |
3279 } | |
3280 } | |
3281 } | |
3282 } | |
2895
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3283 #endif |
2860 | 3284 } |
3285 | |
2185 | 3286 #ifdef HAVE_ODIVX_POSTPROCESS |
3287 #include "../opendivx/postprocess.h" | |
3288 int use_old_pp=0; | |
3289 #endif | |
2159 | 3290 |
2185 | 3291 static void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height, |
2860 | 3292 QP_STORE_T QPs[], int QPStride, int isColor, struct PPMode *ppMode); |
2159 | 3293 |
2389 | 3294 /* -pp Command line Help |
3295 NOTE/FIXME: put this at an appropriate place (--help, html docs, man mplayer)? | |
3296 | |
3297 -pp <filterName>[:<option>[:<option>...]][,[-]<filterName>[:<option>...]]... | |
3298 | |
3299 long form example: | |
3300 -pp vdeblock:autoq,hdeblock:autoq,linblenddeint -pp default,-vdeblock | |
3301 short form example: | |
3302 -pp vb:a,hb:a,lb -pp de,-vb | |
2860 | 3303 more examples: |
3304 -pp tn:64:128:256 | |
2389 | 3305 |
3306 Filters Options | |
3307 short long name short long option Description | |
3308 * * a autoq cpu power dependant enabler | |
3309 c chrom chrominance filtring enabled | |
3310 y nochrom chrominance filtring disabled | |
3311 hb hdeblock horizontal deblocking filter | |
3312 vb vdeblock vertical deblocking filter | |
3313 vr rkvdeblock | |
3314 h1 x1hdeblock Experimental horizontal deblock filter 1 | |
3315 v1 x1vdeblock Experimental vertical deblock filter 1 | |
3316 dr dering not implemented yet | |
3317 al autolevels automatic brightness / contrast fixer | |
3318 f fullyrange stretch luminance range to (0..255) | |
3319 lb linblenddeint linear blend deinterlacer | |
3320 li linipoldeint linear interpolating deinterlacer | |
3321 ci cubicipoldeint cubic interpolating deinterlacer | |
3322 md mediandeint median deinterlacer | |
3323 de default hdeblock:a,vdeblock:a,dering:a,autolevels | |
3324 fa fast x1hdeblock:a,x1vdeblock:a,dering:a,autolevels | |
2860 | 3325 tn tmpnoise (3 Thresholds) Temporal Noise Reducer |
2389 | 3326 */ |
3327 | |
3328 /** | |
3329 * returns a PPMode struct which will have a non 0 error variable if an error occured | |
3330 * name is the string after "-pp" on the command line | |
3331 * quality is a number from 0 to GET_PP_QUALITY_MAX | |
3332 */ | |
3333 struct PPMode getPPModeByNameAndQuality(char *name, int quality) | |
3334 { | |
3335 char temp[GET_MODE_BUFFER_SIZE]; | |
3336 char *p= temp; | |
3337 char *filterDelimiters= ","; | |
3338 char *optionDelimiters= ":"; | |
2860 | 3339 struct PPMode ppMode= {0,0,0,0,0,0,{150,200,400}}; |
2389 | 3340 char *filterToken; |
3341 | |
3342 strncpy(temp, name, GET_MODE_BUFFER_SIZE); | |
3343 | |
2860 | 3344 printf("%s\n", name); |
3345 | |
2389 | 3346 for(;;){ |
3347 char *filterName; | |
2860 | 3348 int q= 1000000; //GET_PP_QUALITY_MAX; |
2389 | 3349 int chrom=-1; |
3350 char *option; | |
3351 char *options[OPTIONS_ARRAY_SIZE]; | |
3352 int i; | |
3353 int filterNameOk=0; | |
3354 int numOfUnknownOptions=0; | |
3355 int enable=1; //does the user want us to enabled or disabled the filter | |
3356 | |
3357 filterToken= strtok(p, filterDelimiters); | |
3358 if(filterToken == NULL) break; | |
2860 | 3359 p+= strlen(filterToken) + 1; // p points to next filterToken |
2389 | 3360 filterName= strtok(filterToken, optionDelimiters); |
3361 printf("%s::%s\n", filterToken, filterName); | |
3362 | |
3363 if(*filterName == '-') | |
3364 { | |
3365 enable=0; | |
3366 filterName++; | |
3367 } | |
2860 | 3368 |
2389 | 3369 for(;;){ //for all options |
3370 option= strtok(NULL, optionDelimiters); | |
3371 if(option == NULL) break; | |
3372 | |
3373 printf("%s\n", option); | |
3374 if(!strcmp("autoq", option) || !strcmp("a", option)) q= quality; | |
3375 else if(!strcmp("nochrom", option) || !strcmp("y", option)) chrom=0; | |
3376 else if(!strcmp("chrom", option) || !strcmp("c", option)) chrom=1; | |
3377 else | |
3378 { | |
3379 options[numOfUnknownOptions] = option; | |
3380 numOfUnknownOptions++; | |
3381 } | |
3382 if(numOfUnknownOptions >= OPTIONS_ARRAY_SIZE-1) break; | |
3383 } | |
2860 | 3384 options[numOfUnknownOptions] = NULL; |
2389 | 3385 |
3386 /* replace stuff from the replace Table */ | |
3387 for(i=0; replaceTable[2*i]!=NULL; i++) | |
3388 { | |
3389 if(!strcmp(replaceTable[2*i], filterName)) | |
3390 { | |
3391 int newlen= strlen(replaceTable[2*i + 1]); | |
3392 int plen; | |
3393 int spaceLeft; | |
3394 | |
3395 if(p==NULL) p= temp, *p=0; //last filter | |
3396 else p--, *p=','; //not last filter | |
3397 | |
3398 plen= strlen(p); | |
3399 spaceLeft= (int)p - (int)temp + plen; | |
3400 if(spaceLeft + newlen >= GET_MODE_BUFFER_SIZE) | |
3401 { | |
3402 ppMode.error++; | |
3403 break; | |
3404 } | |
3405 memmove(p + newlen, p, plen+1); | |
3406 memcpy(p, replaceTable[2*i + 1], newlen); | |
3407 filterNameOk=1; | |
3408 } | |
3409 } | |
3410 | |
3411 for(i=0; filters[i].shortName!=NULL; i++) | |
3412 { | |
2860 | 3413 // printf("Compareing %s, %s, %s\n", filters[i].shortName,filters[i].longName, filterName); |
2389 | 3414 if( !strcmp(filters[i].longName, filterName) |
3415 || !strcmp(filters[i].shortName, filterName)) | |
3416 { | |
3417 ppMode.lumMode &= ~filters[i].mask; | |
3418 ppMode.chromMode &= ~filters[i].mask; | |
3419 | |
3420 filterNameOk=1; | |
3421 if(!enable) break; // user wants to disable it | |
3422 | |
3423 if(q >= filters[i].minLumQuality) | |
3424 ppMode.lumMode|= filters[i].mask; | |
3425 if(chrom==1 || (chrom==-1 && filters[i].chromDefault)) | |
3426 if(q >= filters[i].minChromQuality) | |
3427 ppMode.chromMode|= filters[i].mask; | |
3428 | |
3429 if(filters[i].mask == LEVEL_FIX) | |
3430 { | |
3431 int o; | |
3432 ppMode.minAllowedY= 16; | |
3433 ppMode.maxAllowedY= 234; | |
3434 for(o=0; options[o]!=NULL; o++) | |
3435 if( !strcmp(options[o],"fullyrange") | |
3436 ||!strcmp(options[o],"f")) | |
3437 { | |
3438 ppMode.minAllowedY= 0; | |
3439 ppMode.maxAllowedY= 255; | |
3440 numOfUnknownOptions--; | |
3441 } | |
3442 } | |
2860 | 3443 else if(filters[i].mask == TEMP_NOISE_FILTER) |
3444 { | |
3445 int o; | |
3446 int numOfNoises=0; | |
3447 ppMode.maxTmpNoise[0]= 150; | |
3448 ppMode.maxTmpNoise[1]= 200; | |
3449 ppMode.maxTmpNoise[2]= 400; | |
3450 | |
3451 for(o=0; options[o]!=NULL; o++) | |
3452 { | |
3453 char *tail; | |
3454 ppMode.maxTmpNoise[numOfNoises]= | |
3455 strtol(options[o], &tail, 0); | |
3456 if(tail!=options[o]) | |
3457 { | |
3458 numOfNoises++; | |
3459 numOfUnknownOptions--; | |
3460 if(numOfNoises >= 3) break; | |
3461 } | |
3462 } | |
3463 } | |
2389 | 3464 } |
3465 } | |
3466 if(!filterNameOk) ppMode.error++; | |
3467 ppMode.error += numOfUnknownOptions; | |
3468 } | |
3469 | |
2808 | 3470 #ifdef HAVE_ODIVX_POSTPROCESS |
2389 | 3471 if(ppMode.lumMode & H_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_Y_H; |
3472 if(ppMode.lumMode & V_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_Y_V; | |
3473 if(ppMode.chromMode & H_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_C_H; | |
3474 if(ppMode.chromMode & V_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_C_V; | |
3475 if(ppMode.lumMode & DERING) ppMode.oldMode |= PP_DERING_Y; | |
3476 if(ppMode.chromMode & DERING) ppMode.oldMode |= PP_DERING_C; | |
2808 | 3477 #endif |
2389 | 3478 |
3479 return ppMode; | |
3480 } | |
3481 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3482 /** |
2860 | 3483 * Obsolete, dont use it, use postprocess2() instead |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3484 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3485 void postprocess(unsigned char * src[], int src_stride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3486 unsigned char * dst[], int dst_stride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3487 int horizontal_size, int vertical_size, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3488 QP_STORE_T *QP_store, int QP_stride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3489 int mode) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3490 { |
2860 | 3491 struct PPMode ppMode; |
3492 static QP_STORE_T zeroArray[2048/8]; | |
2389 | 3493 /* |
3494 static int qual=0; | |
3495 | |
2860 | 3496 ppMode= getPPModeByNameAndQuality("fast,default,-hdeblock,-vdeblock,tmpnoise:150:200:300", qual); |
3497 printf("OK\n"); | |
2389 | 3498 qual++; |
3499 qual%=7; | |
2860 | 3500 printf("\n%X %X %X %X :%d: %d %d %d\n", ppMode.lumMode, ppMode.chromMode, ppMode.oldMode, ppMode.error, |
3501 qual, ppMode.maxTmpNoise[0], ppMode.maxTmpNoise[1], ppMode.maxTmpNoise[2]); | |
2389 | 3502 postprocess2(src, src_stride, dst, dst_stride, |
3503 horizontal_size, vertical_size, QP_store, QP_stride, &ppMode); | |
3504 | |
3505 return; | |
3506 */ | |
2808 | 3507 if(QP_store==NULL) |
3508 { | |
3509 QP_store= zeroArray; | |
3510 QP_stride= 0; | |
3511 } | |
2159 | 3512 |
2860 | 3513 ppMode.lumMode= mode; |
3514 mode= ((mode&0xFF)>>4) | (mode&0xFFFFFF00); | |
3515 ppMode.chromMode= mode; | |
2895
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3516 ppMode.maxTmpNoise[0]= 700; |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3517 ppMode.maxTmpNoise[1]= 1500; |
dd3fabd01df0
temp denoiser changes: (a-b)^2 instead of |a-b| and MMX2/3DNOW version
michael
parents:
2860
diff
changeset
|
3518 ppMode.maxTmpNoise[2]= 3000; |
2860 | 3519 |
2185 | 3520 #ifdef HAVE_ODIVX_POSTPROCESS |
3521 // Note: I could make this shit outside of this file, but it would mean one | |
3522 // more function call... | |
3523 if(use_old_pp){ | |
3524 odivx_postprocess(src,src_stride,dst,dst_stride,horizontal_size,vertical_size,QP_store,QP_stride,mode); | |
3525 return; | |
3526 } | |
3527 #endif | |
3528 | |
2159 | 3529 postProcess(src[0], src_stride, dst[0], dst_stride, |
2860 | 3530 horizontal_size, vertical_size, QP_store, QP_stride, 0, &ppMode); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3531 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3532 horizontal_size >>= 1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3533 vertical_size >>= 1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3534 src_stride >>= 1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3535 dst_stride >>= 1; |
2595 | 3536 // mode&= ~(LINEAR_IPOL_DEINT_FILTER | LINEAR_BLEND_DEINT_FILTER | |
3537 // MEDIAN_DEINT_FILTER | CUBIC_IPOL_DEINT_FILTER); | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3538 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3539 if(1) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3540 { |
2159 | 3541 postProcess(src[1], src_stride, dst[1], dst_stride, |
2860 | 3542 horizontal_size, vertical_size, QP_store, QP_stride, 1, &ppMode); |
2159 | 3543 postProcess(src[2], src_stride, dst[2], dst_stride, |
2860 | 3544 horizontal_size, vertical_size, QP_store, QP_stride, 2, &ppMode); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3545 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3546 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3547 { |
2860 | 3548 memset(dst[1], 128, dst_stride*vertical_size); |
3549 memset(dst[2], 128, dst_stride*vertical_size); | |
3550 // memcpy(dst[1], src[1], src_stride*horizontal_size); | |
3551 // memcpy(dst[2], src[2], src_stride*horizontal_size); | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3552 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3553 } |
2185 | 3554 |
2389 | 3555 void postprocess2(unsigned char * src[], int src_stride, |
3556 unsigned char * dst[], int dst_stride, | |
3557 int horizontal_size, int vertical_size, | |
3558 QP_STORE_T *QP_store, int QP_stride, | |
3559 struct PPMode *mode) | |
3560 { | |
3561 | |
2808 | 3562 static QP_STORE_T zeroArray[2048/8]; |
3563 if(QP_store==NULL) | |
3564 { | |
3565 QP_store= zeroArray; | |
3566 QP_stride= 0; | |
3567 } | |
3568 | |
2389 | 3569 #ifdef HAVE_ODIVX_POSTPROCESS |
3570 // Note: I could make this shit outside of this file, but it would mean one | |
3571 // more function call... | |
3572 if(use_old_pp){ | |
3573 odivx_postprocess(src,src_stride,dst,dst_stride,horizontal_size,vertical_size,QP_store,QP_stride, | |
3574 mode->oldMode); | |
3575 return; | |
3576 } | |
3577 #endif | |
3578 | |
3579 postProcess(src[0], src_stride, dst[0], dst_stride, | |
2860 | 3580 horizontal_size, vertical_size, QP_store, QP_stride, 0, mode); |
2389 | 3581 |
3582 horizontal_size >>= 1; | |
3583 vertical_size >>= 1; | |
3584 src_stride >>= 1; | |
3585 dst_stride >>= 1; | |
3586 | |
3587 postProcess(src[1], src_stride, dst[1], dst_stride, | |
2860 | 3588 horizontal_size, vertical_size, QP_store, QP_stride, 1, mode); |
2389 | 3589 postProcess(src[2], src_stride, dst[2], dst_stride, |
2860 | 3590 horizontal_size, vertical_size, QP_store, QP_stride, 2, mode); |
2389 | 3591 } |
3592 | |
3593 | |
2159 | 3594 /** |
3595 * gets the mode flags for a given quality (larger values mean slower but better postprocessing) | |
2185 | 3596 * 0 <= quality <= 6 |
2159 | 3597 */ |
2185 | 3598 int getPpModeForQuality(int quality){ |
3599 int modes[1+GET_PP_QUALITY_MAX]= { | |
3600 0, | |
3601 #if 1 | |
3602 // horizontal filters first | |
3603 LUM_H_DEBLOCK, | |
3604 LUM_H_DEBLOCK | LUM_V_DEBLOCK, | |
3605 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK, | |
3606 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK, | |
3607 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK | LUM_DERING, | |
3608 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK | LUM_DERING | CHROM_DERING | |
3609 #else | |
3610 // vertical filters first | |
2159 | 3611 LUM_V_DEBLOCK, |
3612 LUM_V_DEBLOCK | LUM_H_DEBLOCK, | |
3613 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK, | |
3614 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK, | |
3615 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING, | |
3616 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING | CHROM_DERING | |
2185 | 3617 #endif |
3618 }; | |
2159 | 3619 |
2185 | 3620 #ifdef HAVE_ODIVX_POSTPROCESS |
3621 int odivx_modes[1+GET_PP_QUALITY_MAX]= { | |
3622 0, | |
3623 PP_DEBLOCK_Y_H, | |
3624 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V, | |
3625 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H, | |
3626 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V, | |
3627 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V|PP_DERING_Y, | |
3628 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V|PP_DERING_Y|PP_DERING_C | |
3629 }; | |
3630 if(use_old_pp) return odivx_modes[quality]; | |
3631 #endif | |
3632 return modes[quality]; | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3633 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3634 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3635 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3636 * Copies a block from src to dst and fixes the blacklevel |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3637 * numLines must be a multiple of 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3638 * levelFix == 0 -> dont touch the brighness & contrast |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3639 */ |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3640 static inline void blockCopy(uint8_t dst[], int dstStride, uint8_t src[], int srcStride, |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3641 int numLines, int levelFix) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3642 { |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
3643 #ifndef HAVE_MMX |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3644 int i; |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
3645 #endif |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3646 if(levelFix) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3647 { |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3648 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3649 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3650 "leal (%2,%2), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3651 "leal (%3,%3), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3652 "movq packedYOffset, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3653 "movq packedYScale, %%mm3 \n\t" |
2181
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
3654 "pxor %%mm4, %%mm4 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3655 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3656 #define SCALED_CPY \ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3657 "movq (%0), %%mm0 \n\t"\ |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
3658 "movq (%0), %%mm5 \n\t"\ |
2181
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
3659 "punpcklbw %%mm4, %%mm0 \n\t"\ |
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
3660 "punpckhbw %%mm4, %%mm5 \n\t"\ |
2394 | 3661 "psubw %%mm2, %%mm0 \n\t"\ |
3662 "psubw %%mm2, %%mm5 \n\t"\ | |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
3663 "movq (%0,%2), %%mm1 \n\t"\ |
2394 | 3664 "psllw $6, %%mm0 \n\t"\ |
3665 "psllw $6, %%mm5 \n\t"\ | |
2181
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
3666 "pmulhw %%mm3, %%mm0 \n\t"\ |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
3667 "movq (%0,%2), %%mm6 \n\t"\ |
2181
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
3668 "pmulhw %%mm3, %%mm5 \n\t"\ |
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
3669 "punpcklbw %%mm4, %%mm1 \n\t"\ |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
3670 "punpckhbw %%mm4, %%mm6 \n\t"\ |
2394 | 3671 "psubw %%mm2, %%mm1 \n\t"\ |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
3672 "psubw %%mm2, %%mm6 \n\t"\ |
2394 | 3673 "psllw $6, %%mm1 \n\t"\ |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
3674 "psllw $6, %%mm6 \n\t"\ |
2181
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
3675 "pmulhw %%mm3, %%mm1 \n\t"\ |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
3676 "pmulhw %%mm3, %%mm6 \n\t"\ |
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
3677 "addl %%eax, %0 \n\t"\ |
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
3678 "packuswb %%mm5, %%mm0 \n\t"\ |
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
3679 "packuswb %%mm6, %%mm1 \n\t"\ |
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
3680 "movq %%mm0, (%1) \n\t"\ |
2181
d90f8fc7ead6
fixed a sig4 bug an non mmx2 cpus (in case of more sig4 errors please send me a "disassemble $eip-16 $eip+16" from gdb)
michael
parents:
2180
diff
changeset
|
3681 "movq %%mm1, (%1, %3) \n\t"\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3682 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3683 SCALED_CPY |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3684 "addl %%ebx, %1 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3685 SCALED_CPY |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3686 "addl %%ebx, %1 \n\t" |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
3687 SCALED_CPY |
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
3688 "addl %%ebx, %1 \n\t" |
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
3689 SCALED_CPY |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3690 |
2413 | 3691 : "+r"(src), |
3692 "+r"(dst) | |
3693 :"r" (srcStride), | |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
3694 "r" (dstStride) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3695 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3696 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3697 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3698 for(i=0; i<numLines; i++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3699 memcpy( &(dst[dstStride*i]), |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3700 &(src[srcStride*i]), BLOCK_SIZE); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3701 #endif |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3702 } |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3703 else |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3704 { |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3705 #ifdef HAVE_MMX |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3706 asm volatile( |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3707 "movl %4, %%eax \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3708 "movl %%eax, temp0\n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3709 "pushl %0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3710 "pushl %1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3711 "leal (%2,%2), %%eax \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3712 "leal (%3,%3), %%ebx \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3713 "movq packedYOffset, %%mm2 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3714 "movq packedYScale, %%mm3 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3715 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3716 #define SIMPLE_CPY \ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3717 "movq (%0), %%mm0 \n\t"\ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3718 "movq (%0,%2), %%mm1 \n\t"\ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3719 "movq %%mm0, (%1) \n\t"\ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3720 "movq %%mm1, (%1, %3) \n\t"\ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3721 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3722 "1: \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3723 SIMPLE_CPY |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3724 "addl %%eax, %0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3725 "addl %%ebx, %1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3726 SIMPLE_CPY |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3727 "addl %%eax, %0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3728 "addl %%ebx, %1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3729 "decl temp0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3730 "jnz 1b \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3731 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3732 "popl %1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3733 "popl %0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3734 : : "r" (src), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3735 "r" (dst), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3736 "r" (srcStride), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3737 "r" (dstStride), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3738 "m" (numLines>>2) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3739 : "%eax", "%ebx" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3740 ); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3741 #else |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3742 for(i=0; i<numLines; i++) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3743 memcpy( &(dst[dstStride*i]), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3744 &(src[srcStride*i]), BLOCK_SIZE); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3745 #endif |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3746 } |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3747 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3748 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3749 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3750 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3751 * Filters array of bytes (Y or U or V values) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3752 */ |
2185 | 3753 static void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height, |
2860 | 3754 QP_STORE_T QPs[], int QPStride, int isColor, struct PPMode *ppMode) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3755 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3756 int x,y; |
2860 | 3757 const int mode= isColor ? ppMode->chromMode : ppMode->lumMode; |
3758 | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3759 /* we need 64bit here otherwise we´ll going to have a problem |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3760 after watching a black picture for 5 hours*/ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3761 static uint64_t *yHistogram= NULL; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3762 int black=0, white=255; // blackest black and whitest white in the picture |
2437 | 3763 int QPCorrecture= 256; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3764 |
2246 | 3765 /* Temporary buffers for handling the last row(s) */ |
3766 static uint8_t *tempDst= NULL; | |
3767 static uint8_t *tempSrc= NULL; | |
3768 | |
2285 | 3769 /* Temporary buffers for handling the last block */ |
3770 static uint8_t *tempDstBlock= NULL; | |
3771 static uint8_t *tempSrcBlock= NULL; | |
3772 | |
2860 | 3773 /* Temporal noise reducing buffers */ |
3774 static uint8_t *tempBlured[3]= {NULL,NULL,NULL}; | |
2899 | 3775 static uint32_t *tempBluredPast[3]= {NULL,NULL,NULL}; |
2860 | 3776 |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
3777 #ifdef PP_FUNNY_STRIDE |
2285 | 3778 uint8_t *dstBlockPtrBackup; |
3779 uint8_t *srcBlockPtrBackup; | |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
3780 #endif |
2285 | 3781 |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
3782 #ifdef MORE_TIMING |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
3783 long long T0, T1, diffTime=0; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
3784 #endif |
2221 | 3785 #ifdef TIMING |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
3786 long long memcpyTime=0, vertTime=0, horizTime=0, sumTime; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3787 sumTime= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3788 #endif |
2570 | 3789 //mode= 0x7F; |
2899 | 3790 #ifdef HAVE_MMX |
3791 maxTmpNoise[0]= ppMode->maxTmpNoise[0]; | |
3792 maxTmpNoise[1]= ppMode->maxTmpNoise[1]; | |
3793 maxTmpNoise[2]= ppMode->maxTmpNoise[2]; | |
3794 #endif | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3795 |
2246 | 3796 if(tempDst==NULL) |
3797 { | |
3798 tempDst= (uint8_t*)memalign(8, 1024*24); | |
3799 tempSrc= (uint8_t*)memalign(8, 1024*24); | |
2285 | 3800 tempDstBlock= (uint8_t*)memalign(8, 1024*24); |
3801 tempSrcBlock= (uint8_t*)memalign(8, 1024*24); | |
2246 | 3802 } |
3803 | |
2860 | 3804 if(tempBlured[isColor]==NULL && (mode & TEMP_NOISE_FILTER)) |
3805 { | |
3806 // printf("%d %d %d\n", isColor, dstStride, height); | |
3807 //FIXME works only as long as the size doesnt increase | |
3808 //Note:the +17*1024 is just there so i dont have to worry about r/w over te end | |
3809 tempBlured[isColor]= (uint8_t*)memalign(8, dstStride*((height+7)&(~7)) + 17*1024); | |
2899 | 3810 tempBluredPast[isColor]= (uint32_t*)memalign(8, 256*((height+7)&(~7))/2 + 17*1024); |
2860 | 3811 |
3812 memset(tempBlured[isColor], 0, dstStride*((height+7)&(~7)) + 17*1024); | |
2899 | 3813 memset(tempBluredPast[isColor], 0, 256*((height+7)&(~7))/2 + 17*1024); |
2860 | 3814 } |
3815 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3816 if(!yHistogram) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3817 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3818 int i; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3819 yHistogram= (uint64_t*)malloc(8*256); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3820 for(i=0; i<256; i++) yHistogram[i]= width*height/64*15/256; |
2285 | 3821 |
3822 if(mode & FULL_Y_RANGE) | |
3823 { | |
3824 maxAllowedY=255; | |
3825 minAllowedY=0; | |
3826 } | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3827 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3828 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3829 if(!isColor) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3830 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3831 uint64_t sum= 0; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3832 int i; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3833 static int framenum= -1; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3834 uint64_t maxClipped; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3835 uint64_t clipped; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3836 double scale; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3837 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3838 framenum++; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3839 if(framenum == 1) yHistogram[0]= width*height/64*15/256; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3840 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3841 for(i=0; i<256; i++) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3842 { |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3843 sum+= yHistogram[i]; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3844 // printf("%d ", yHistogram[i]); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3845 } |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3846 // printf("\n\n"); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3847 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3848 /* we allways get a completly black picture first */ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3849 maxClipped= (uint64_t)(sum * maxClippedThreshold); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3850 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3851 clipped= sum; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3852 for(black=255; black>0; black--) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3853 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3854 if(clipped < maxClipped) break; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3855 clipped-= yHistogram[black]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3856 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3857 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3858 clipped= sum; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3859 for(white=0; white<256; white++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3860 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3861 if(clipped < maxClipped) break; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3862 clipped-= yHistogram[white]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3863 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3864 |
2394 | 3865 packedYOffset= (black - minAllowedY) & 0xFFFF; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3866 packedYOffset|= packedYOffset<<32; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3867 packedYOffset|= packedYOffset<<16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3868 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3869 scale= (double)(maxAllowedY - minAllowedY) / (double)(white-black); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3870 |
2394 | 3871 packedYScale= (uint16_t)(scale*1024.0 + 0.5); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3872 packedYScale|= packedYScale<<32; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3873 packedYScale|= packedYScale<<16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3874 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3875 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3876 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3877 packedYScale= 0x0100010001000100LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3878 packedYOffset= 0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3879 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3880 |
2437 | 3881 if(mode & LEVEL_FIX) QPCorrecture= packedYScale &0xFFFF; |
3882 else QPCorrecture= 256; | |
3883 | |
2742
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
3884 /* copy & deinterlace first row of blocks */ |
2595 | 3885 y=-BLOCK_SIZE; |
3886 { | |
3887 //1% speedup if these are here instead of the inner loop | |
3888 uint8_t *srcBlock= &(src[y*srcStride]); | |
3889 uint8_t *dstBlock= &(dst[y*dstStride]); | |
3890 | |
3891 dstBlock= tempDst + dstStride; | |
3892 | |
3893 // From this point on it is guranteed that we can read and write 16 lines downward | |
3894 // finish 1 block before the next otherwise we´ll might have a problem | |
3895 // with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing | |
3896 for(x=0; x<width; x+=BLOCK_SIZE) | |
3897 { | |
3898 | |
3899 #ifdef HAVE_MMX2 | |
3900 /* | |
3901 prefetchnta(srcBlock + (((x>>3)&3) + 5)*srcStride + 32); | |
3902 prefetchnta(srcBlock + (((x>>3)&3) + 9)*srcStride + 32); | |
3903 prefetcht0(dstBlock + (((x>>3)&3) + 5)*dstStride + 32); | |
3904 prefetcht0(dstBlock + (((x>>3)&3) + 9)*dstStride + 32); | |
3905 */ | |
3906 /* | |
3907 prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32); | |
3908 prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32); | |
3909 prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32); | |
3910 prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32); | |
3911 */ | |
3912 | |
3913 asm( | |
3914 "movl %4, %%eax \n\t" | |
3915 "shrl $2, %%eax \n\t" | |
3916 "andl $6, %%eax \n\t" | |
3917 "addl $8, %%eax \n\t" | |
3918 "movl %%eax, %%ebx \n\t" | |
3919 "imul %1, %%eax \n\t" | |
3920 "imul %3, %%ebx \n\t" | |
3921 "prefetchnta 32(%%eax, %0) \n\t" | |
3922 "prefetcht0 32(%%ebx, %2) \n\t" | |
3923 "addl %1, %%eax \n\t" | |
3924 "addl %3, %%ebx \n\t" | |
3925 "prefetchnta 32(%%eax, %0) \n\t" | |
3926 "prefetcht0 32(%%ebx, %2) \n\t" | |
3927 :: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride), | |
3928 "m" (x) | |
3929 : "%eax", "%ebx" | |
3930 ); | |
3931 | |
3932 #elif defined(HAVE_3DNOW) | |
3933 //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ... | |
3934 /* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32); | |
3935 prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32); | |
3936 prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32); | |
3937 prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32); | |
3938 */ | |
3939 #endif | |
3940 | |
3941 blockCopy(dstBlock + dstStride*8, dstStride, | |
3942 srcBlock + srcStride*8, srcStride, 8, mode & LEVEL_FIX); | |
3943 | |
3944 if(mode & LINEAR_IPOL_DEINT_FILTER) | |
3945 deInterlaceInterpolateLinear(dstBlock, dstStride); | |
3946 else if(mode & LINEAR_BLEND_DEINT_FILTER) | |
3947 deInterlaceBlendLinear(dstBlock, dstStride); | |
3948 else if(mode & MEDIAN_DEINT_FILTER) | |
3949 deInterlaceMedian(dstBlock, dstStride); | |
3950 else if(mode & CUBIC_IPOL_DEINT_FILTER) | |
3951 deInterlaceInterpolateCubic(dstBlock, dstStride); | |
3952 /* else if(mode & CUBIC_BLEND_DEINT_FILTER) | |
3953 deInterlaceBlendCubic(dstBlock, dstStride); | |
3954 */ | |
3955 dstBlock+=8; | |
3956 srcBlock+=8; | |
3957 } | |
3958 memcpy(&(dst[y*dstStride]) + 8*dstStride, tempDst + 9*dstStride, 8*dstStride ); | |
3959 } | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3960 |
2246 | 3961 for(y=0; y<height; y+=BLOCK_SIZE) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3962 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3963 //1% speedup if these are here instead of the inner loop |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3964 uint8_t *srcBlock= &(src[y*srcStride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3965 uint8_t *dstBlock= &(dst[y*dstStride]); |
2437 | 3966 #ifdef ARCH_X86 |
3967 int *QPptr= isColor ? &QPs[(y>>3)*QPStride] :&QPs[(y>>4)*QPStride]; | |
3968 int QPDelta= isColor ? 1<<(32-3) : 1<<(32-4); | |
3969 int QPFrac= QPDelta; | |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3970 uint8_t *tempBlock1= tempBlocks; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3971 uint8_t *tempBlock2= tempBlocks + 8; |
2437 | 3972 #endif |
2860 | 3973 int QP=0; |
2473 | 3974 /* can we mess with a 8x16 block from srcBlock/dstBlock downwards and 1 line upwards |
3975 if not than use a temporary buffer */ | |
2246 | 3976 if(y+15 >= height) |
3977 { | |
2860 | 3978 int i; |
2595 | 3979 /* copy from line 8 to 15 of src, these will be copied with |
2246 | 3980 blockcopy to dst later */ |
2595 | 3981 memcpy(tempSrc + srcStride*8, srcBlock + srcStride*8, |
3982 srcStride*MAX(height-y-8, 0) ); | |
3983 | |
2860 | 3984 /* duplicate last line of src to fill the void upto line 15 */ |
3985 for(i=MAX(height-y, 8); i<=15; i++) | |
3986 memcpy(tempSrc + srcStride*i, src + srcStride*(height-1), srcStride); | |
3987 | |
3988 /* copy up to 9 lines of dst (line -1 to 7)*/ | |
2595 | 3989 memcpy(tempDst, dstBlock - dstStride, dstStride*MIN(height-y+1, 9) ); |
2860 | 3990 |
3991 /* duplicate last line of dst to fill the void upto line 8 */ | |
3992 for(i=height-y+1; i<=8; i++) | |
3993 memcpy(tempDst + dstStride*i, dst + dstStride*(height-1), dstStride); | |
3994 | |
2473 | 3995 dstBlock= tempDst + dstStride; |
2246 | 3996 srcBlock= tempSrc; |
3997 } | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3998 |
2285 | 3999 // From this point on it is guranteed that we can read and write 16 lines downward |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4000 // finish 1 block before the next otherwise we´ll might have a problem |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4001 // with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
4002 for(x=0; x<width; x+=BLOCK_SIZE) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4003 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
4004 const int stride= dstStride; |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4005 uint8_t *tmpXchg; |
2437 | 4006 #ifdef ARCH_X86 |
2860 | 4007 QP= *QPptr; |
2437 | 4008 asm volatile( |
4009 "addl %2, %1 \n\t" | |
4010 "sbbl %%eax, %%eax \n\t" | |
4011 "shll $2, %%eax \n\t" | |
4012 "subl %%eax, %0 \n\t" | |
4013 : "+r" (QPptr), "+m" (QPFrac) | |
4014 : "r" (QPDelta) | |
4015 : "%eax" | |
4016 ); | |
4017 #else | |
2860 | 4018 QP= isColor ? |
2437 | 4019 QPs[(y>>3)*QPStride + (x>>3)]: |
4020 QPs[(y>>4)*QPStride + (x>>4)]; | |
4021 #endif | |
4022 if(!isColor) | |
2428 | 4023 { |
2437 | 4024 QP= (QP* QPCorrecture)>>8; |
2742
d5636499cafd
minor changes (fixed some warnings, added attribute aligned(8) stuff)
michael
parents:
2595
diff
changeset
|
4025 yHistogram[ srcBlock[srcStride*12 + 4] ]++; |
2428 | 4026 } |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4027 #ifdef HAVE_MMX |
2246 | 4028 asm volatile( |
4029 "movd %0, %%mm7 \n\t" | |
4030 "packuswb %%mm7, %%mm7 \n\t" // 0, 0, 0, QP, 0, 0, 0, QP | |
4031 "packuswb %%mm7, %%mm7 \n\t" // 0,QP, 0, QP, 0,QP, 0, QP | |
4032 "packuswb %%mm7, %%mm7 \n\t" // QP,..., QP | |
4033 "movq %%mm7, pQPb \n\t" | |
4034 : : "r" (QP) | |
4035 ); | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4036 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4037 |
2221 | 4038 #ifdef MORE_TIMING |
2246 | 4039 T0= rdtsc(); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4040 #endif |
2159 | 4041 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4042 #ifdef HAVE_MMX2 |
2437 | 4043 /* |
2246 | 4044 prefetchnta(srcBlock + (((x>>3)&3) + 5)*srcStride + 32); |
4045 prefetchnta(srcBlock + (((x>>3)&3) + 9)*srcStride + 32); | |
4046 prefetcht0(dstBlock + (((x>>3)&3) + 5)*dstStride + 32); | |
4047 prefetcht0(dstBlock + (((x>>3)&3) + 9)*dstStride + 32); | |
2437 | 4048 */ |
4049 /* | |
4050 prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32); | |
4051 prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32); | |
4052 prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32); | |
4053 prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32); | |
4054 */ | |
4055 | |
4056 asm( | |
4057 "movl %4, %%eax \n\t" | |
4058 "shrl $2, %%eax \n\t" | |
4059 "andl $6, %%eax \n\t" | |
2595 | 4060 "addl $8, %%eax \n\t" |
2437 | 4061 "movl %%eax, %%ebx \n\t" |
4062 "imul %1, %%eax \n\t" | |
4063 "imul %3, %%ebx \n\t" | |
4064 "prefetchnta 32(%%eax, %0) \n\t" | |
4065 "prefetcht0 32(%%ebx, %2) \n\t" | |
4066 "addl %1, %%eax \n\t" | |
4067 "addl %3, %%ebx \n\t" | |
4068 "prefetchnta 32(%%eax, %0) \n\t" | |
4069 "prefetcht0 32(%%ebx, %2) \n\t" | |
4070 :: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride), | |
4071 "m" (x) | |
4072 : "%eax", "%ebx" | |
4073 ); | |
4074 | |
2159 | 4075 #elif defined(HAVE_3DNOW) |
4076 //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ... | |
2246 | 4077 /* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32); |
4078 prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32); | |
4079 prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32); | |
4080 prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32); | |
2159 | 4081 */ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4082 #endif |
2246 | 4083 |
2286 | 4084 #ifdef PP_FUNNY_STRIDE |
2285 | 4085 //can we mess with a 8x16 block, if not use a temp buffer, yes again |
4086 if(x+7 >= width) | |
4087 { | |
4088 int i; | |
4089 dstBlockPtrBackup= dstBlock; | |
4090 srcBlockPtrBackup= srcBlock; | |
4091 | |
4092 for(i=0;i<BLOCK_SIZE*2; i++) | |
4093 { | |
4094 memcpy(tempSrcBlock+i*srcStride, srcBlock+i*srcStride, width-x); | |
4095 memcpy(tempDstBlock+i*dstStride, dstBlock+i*dstStride, width-x); | |
4096 } | |
4097 | |
4098 dstBlock= tempDstBlock; | |
4099 srcBlock= tempSrcBlock; | |
4100 } | |
2286 | 4101 #endif |
2285 | 4102 |
2595 | 4103 blockCopy(dstBlock + dstStride*8, dstStride, |
4104 srcBlock + srcStride*8, srcStride, 8, mode & LEVEL_FIX); | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4105 |
2246 | 4106 if(mode & LINEAR_IPOL_DEINT_FILTER) |
4107 deInterlaceInterpolateLinear(dstBlock, dstStride); | |
4108 else if(mode & LINEAR_BLEND_DEINT_FILTER) | |
4109 deInterlaceBlendLinear(dstBlock, dstStride); | |
4110 else if(mode & MEDIAN_DEINT_FILTER) | |
4111 deInterlaceMedian(dstBlock, dstStride); | |
4112 else if(mode & CUBIC_IPOL_DEINT_FILTER) | |
4113 deInterlaceInterpolateCubic(dstBlock, dstStride); | |
4114 /* else if(mode & CUBIC_BLEND_DEINT_FILTER) | |
4115 deInterlaceBlendCubic(dstBlock, dstStride); | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
4116 */ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4117 |
2246 | 4118 /* only deblock if we have 2 blocks */ |
4119 if(y + 8 < height) | |
4120 { | |
2221 | 4121 #ifdef MORE_TIMING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4122 T1= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4123 memcpyTime+= T1-T0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4124 T0=T1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4125 #endif |
2300
e10f7dc4938f
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:
2286
diff
changeset
|
4126 if(mode & V_RK1_FILTER) |
e10f7dc4938f
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:
2286
diff
changeset
|
4127 vertRK1Filter(dstBlock, stride, QP); |
e10f7dc4938f
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:
2286
diff
changeset
|
4128 else if(mode & V_X1_FILTER) |
e10f7dc4938f
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:
2286
diff
changeset
|
4129 vertX1Filter(dstBlock, stride, QP); |
e10f7dc4938f
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:
2286
diff
changeset
|
4130 else if(mode & V_DEBLOCK) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4131 { |
2300
e10f7dc4938f
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:
2286
diff
changeset
|
4132 if( isVertDC(dstBlock, stride)) |
2159 | 4133 { |
2300
e10f7dc4938f
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:
2286
diff
changeset
|
4134 if(isVertMinMaxOk(dstBlock, stride, QP)) |
e10f7dc4938f
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:
2286
diff
changeset
|
4135 doVertLowPass(dstBlock, stride, QP); |
2159 | 4136 } |
2300
e10f7dc4938f
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:
2286
diff
changeset
|
4137 else |
e10f7dc4938f
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:
2286
diff
changeset
|
4138 doVertDefFilter(dstBlock, stride, QP); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4139 } |
2221 | 4140 #ifdef MORE_TIMING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4141 T1= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4142 vertTime+= T1-T0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4143 T0=T1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4144 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4145 } |
2473 | 4146 |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4147 #ifdef HAVE_MMX |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4148 transpose1(tempBlock1, tempBlock2, dstBlock, dstStride); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4149 #endif |
2246 | 4150 /* check if we have a previous block to deblock it with dstBlock */ |
2285 | 4151 if(x - 8 >= 0) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4152 { |
2221 | 4153 #ifdef MORE_TIMING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4154 T0= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4155 #endif |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4156 #ifdef HAVE_MMX |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4157 if(mode & H_RK1_FILTER) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4158 vertRK1Filter(tempBlock1, 16, QP); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4159 else if(mode & H_X1_FILTER) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4160 vertX1Filter(tempBlock1, 16, QP); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4161 else if(mode & H_DEBLOCK) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4162 { |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4163 if( isVertDC(tempBlock1, 16)) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4164 { |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4165 if(isVertMinMaxOk(tempBlock1, 16, QP)) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4166 doVertLowPass(tempBlock1, 16, QP); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4167 } |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4168 else |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4169 doVertDefFilter(tempBlock1, 16, QP); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4170 } |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4171 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4172 transpose2(dstBlock-4, dstStride, tempBlock1 + 4*16); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4173 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4174 #else |
2300
e10f7dc4938f
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:
2286
diff
changeset
|
4175 if(mode & H_X1_FILTER) |
e10f7dc4938f
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:
2286
diff
changeset
|
4176 horizX1Filter(dstBlock-4, stride, QP); |
e10f7dc4938f
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:
2286
diff
changeset
|
4177 else if(mode & H_DEBLOCK) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4178 { |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4179 if( isHorizDC(dstBlock-4, stride)) |
2159 | 4180 { |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4181 if(isHorizMinMaxOk(dstBlock-4, stride, QP)) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4182 doHorizLowPass(dstBlock-4, stride, QP); |
2159 | 4183 } |
2300
e10f7dc4938f
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:
2286
diff
changeset
|
4184 else |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4185 doHorizDefFilter(dstBlock-4, stride, QP); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4186 } |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4187 #endif |
2221 | 4188 #ifdef MORE_TIMING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4189 T1= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4190 horizTime+= T1-T0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4191 T0=T1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4192 #endif |
2473 | 4193 if(mode & DERING) |
4194 { | |
4195 //FIXME filter first line | |
4196 if(y>0) dering(dstBlock - stride - 8, stride, QP); | |
4197 } | |
2860 | 4198 |
4199 if(mode & TEMP_NOISE_FILTER) | |
4200 { | |
4201 tempNoiseReducer(dstBlock-8, stride, | |
4202 tempBlured[isColor] + y*dstStride + x, | |
2899 | 4203 tempBluredPast[isColor] + (y>>3)*256 + (x>>3), |
2860 | 4204 ppMode->maxTmpNoise); |
4205 } | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4206 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4207 |
2286 | 4208 #ifdef PP_FUNNY_STRIDE |
2285 | 4209 /* did we use a tmp-block buffer */ |
4210 if(x+7 >= width) | |
4211 { | |
4212 int i; | |
4213 dstBlock= dstBlockPtrBackup; | |
4214 srcBlock= srcBlockPtrBackup; | |
4215 | |
4216 for(i=0;i<BLOCK_SIZE*2; i++) | |
4217 { | |
4218 memcpy(dstBlock+i*dstStride, tempDstBlock+i*dstStride, width-x); | |
4219 } | |
4220 } | |
2286 | 4221 #endif |
2285 | 4222 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4223 dstBlock+=8; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4224 srcBlock+=8; |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4225 |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
4226 #ifdef HAVE_MMX |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4227 tmpXchg= tempBlock1; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4228 tempBlock1= tempBlock2; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
4229 tempBlock2 = tmpXchg; |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
4230 #endif |
2246 | 4231 } |
4232 | |
2860 | 4233 if(mode & DERING) |
4234 { | |
4235 if(y > 0) dering(dstBlock - dstStride - 8, dstStride, QP); | |
4236 } | |
4237 | |
4238 if((mode & TEMP_NOISE_FILTER)) | |
4239 { | |
4240 tempNoiseReducer(dstBlock-8, dstStride, | |
4241 tempBlured[isColor] + y*dstStride + x, | |
2899 | 4242 tempBluredPast[isColor] + (y>>3)*256 + (x>>3), |
2860 | 4243 ppMode->maxTmpNoise); |
4244 } | |
4245 | |
2595 | 4246 /* did we use a tmp buffer for the last lines*/ |
2285 | 4247 if(y+15 >= height) |
2246 | 4248 { |
4249 uint8_t *dstBlock= &(dst[y*dstStride]); | |
2473 | 4250 memcpy(dstBlock, tempDst + dstStride, dstStride*(height-y) ); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4251 } |
3013 | 4252 /* |
4253 for(x=0; x<width; x+=32) | |
4254 { | |
4255 int i; | |
4256 i+= + dstBlock[x + 7*dstStride] + dstBlock[x + 8*dstStride] | |
4257 + dstBlock[x + 9*dstStride] + dstBlock[x +10*dstStride] | |
4258 + dstBlock[x +11*dstStride] + dstBlock[x +12*dstStride] | |
4259 + dstBlock[x +13*dstStride] + dstBlock[x +14*dstStride] | |
4260 + dstBlock[x +15*dstStride]; | |
4261 } | |
4262 */ } | |
2159 | 4263 #ifdef HAVE_3DNOW |
4264 asm volatile("femms"); | |
4265 #elif defined (HAVE_MMX) | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4266 asm volatile("emms"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4267 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4268 |
2221 | 4269 #ifdef TIMING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4270 // FIXME diff is mostly the time spent for rdtsc (should subtract that but ...) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4271 sumTime= rdtsc() - sumTime; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4272 if(!isColor) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4273 printf("cpy:%4dk, vert:%4dk, horiz:%4dk, sum:%4dk, diff:%4dk, color: %d/%d \r", |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
4274 (int)(memcpyTime/1000), (int)(vertTime/1000), (int)(horizTime/1000), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
4275 (int)(sumTime/1000), (int)((sumTime-memcpyTime-vertTime-horizTime)/1000) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4276 , black, white); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4277 #endif |
3013 | 4278 #ifdef DEBUG_BRIGHTNESS |
4279 if(!isColor) | |
4280 { | |
4281 int max=1; | |
4282 int i; | |
4283 for(i=0; i<256; i++) | |
4284 if(yHistogram[i] > max) max=yHistogram[i]; | |
4285 | |
4286 for(i=1; i<256; i++) | |
4287 { | |
4288 int x; | |
4289 int start=yHistogram[i-1]/(max/256+1); | |
4290 int end=yHistogram[i]/(max/256+1); | |
4291 int inc= end > start ? 1 : -1; | |
4292 for(x=start; x!=end+inc; x+=inc) | |
4293 dst[ i*dstStride + x]+=128; | |
4294 } | |
4295 | |
4296 for(i=0; i<100; i+=2) | |
4297 { | |
4298 dst[ (white)*dstStride + i]+=128; | |
4299 dst[ (black)*dstStride + i]+=128; | |
4300 } | |
4301 | |
4302 } | |
4303 #endif | |
4304 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4305 } |