Mercurial > mplayer.hg
annotate postproc/postprocess_template.c @ 2476:a6c5a537f30a
a few warning fixes (missing #include's)
author | pl |
---|---|
date | Thu, 25 Oct 2001 23:34:14 +0000 |
parents | 3369845d92f4 |
children | 830c0e171d7d |
rev | line source |
---|---|
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2 Copyright (C) 2001 Michael Niedermayer (michaelni@gmx.at) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4 This program is free software; you can redistribute it and/or modify |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
5 it under the terms of the GNU General Public License as published by |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
6 the Free Software Foundation; either version 2 of the License, or |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
7 (at your option) any later version. |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
8 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
9 This program is distributed in the hope that it will be useful, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
10 but WITHOUT ANY WARRANTY; without even the implied warranty of |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
12 GNU General Public License for more details. |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
13 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
14 You should have received a copy of the GNU General Public License |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
15 along with this program; if not, write to the Free Software |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
16 Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
17 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
18 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
19 /* |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
20 C MMX MMX2 3DNow |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
21 isVertDC Ec Ec |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
22 isVertMinMaxOk Ec Ec |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
23 doVertLowPass E e e |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
24 doVertDefFilter Ec Ec Ec |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
25 isHorizDC Ec Ec |
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 |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
28 doHorizDefFilter Ec Ec Ec |
2475 | 29 deRing 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 |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
32 Vertical X1 a E E |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
33 Horizontal X1 a E E |
2246 | 34 LinIpolDeinterlace e E E* |
35 CubicIpolDeinterlace a e e* | |
36 LinBlendDeinterlace e E E* | |
2221 | 37 MedianDeinterlace Ec Ec |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
38 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
39 |
2159 | 40 * i dont have a 3dnow CPU -> its untested |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
41 E = Exact implementation |
2246 | 42 e = allmost exact implementation (slightly different rounding,...) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
43 a = alternative / approximate impl |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
44 c = checked against the other implementations (-vo md5) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
45 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
46 |
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 TODO: |
2179 | 49 verify that everything workes as it should (how?) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
50 reduce the time wasted on the mem transfer |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
51 implement dering |
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 do something about the speed of the horizontal filters |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
58 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
|
59 (the if/else stuff per block is slowing things down) |
2179 | 60 compare the quality & speed of all filters |
61 split this huge file | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
62 fix warnings (unused vars, ...) |
2221 | 63 noise reduction filters |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
64 border remover |
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 Notes: |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
68 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
69 |
2221 | 70 //Changelog: use the CVS log |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
71 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
72 #include <inttypes.h> |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
73 #include <stdio.h> |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
74 #include <stdlib.h> |
2389 | 75 #include <string.h> |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
76 #include "../config.h" |
2476 | 77 #ifdef HAVE_MALLOC_H |
78 #include <malloc.h> | |
79 #endif | |
2159 | 80 //#undef HAVE_MMX2 |
81 //#define HAVE_3DNOW | |
82 //#undef HAVE_MMX | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
83 #include "postprocess.h" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
84 |
2189 | 85 #define MIN(a,b) ((a) > (b) ? (b) : (a)) |
86 #define MAX(a,b) ((a) < (b) ? (b) : (a)) | |
87 #define ABS(a) ((a) > 0 ? (a) : (-(a))) | |
88 #define SIGN(a) ((a) > 0 ? 1 : -1) | |
89 | |
90 #ifdef HAVE_MMX2 | |
91 #define PAVGB(a,b) "pavgb " #a ", " #b " \n\t" | |
92 #elif defined (HAVE_3DNOW) | |
93 #define PAVGB(a,b) "pavgusb " #a ", " #b " \n\t" | |
94 #endif | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
95 |
2389 | 96 #define GET_MODE_BUFFER_SIZE 500 |
97 #define OPTIONS_ARRAY_SIZE 10 | |
98 | |
99 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
100 static uint64_t packedYOffset= 0x0000000000000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
101 static uint64_t packedYScale= 0x0100010001000100LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
102 static uint64_t w05= 0x0005000500050005LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
103 static uint64_t w20= 0x0020002000200020LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
104 static uint64_t w1400= 0x1400140014001400LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
105 static uint64_t bm00000001= 0x00000000000000FFLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
106 static uint64_t bm00010000= 0x000000FF00000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
107 static uint64_t bm00001000= 0x00000000FF000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
108 static uint64_t bm10000000= 0xFF00000000000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
109 static uint64_t bm10000001= 0xFF000000000000FFLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
110 static uint64_t bm11000011= 0xFFFF00000000FFFFLL; |
2159 | 111 static uint64_t bm00000011= 0x000000000000FFFFLL; |
2179 | 112 static uint64_t bm11111110= 0xFFFFFFFFFFFFFF00LL; |
2159 | 113 static uint64_t bm11000000= 0xFFFF000000000000LL; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
114 static uint64_t bm00011000= 0x000000FFFF000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
115 static uint64_t bm00110011= 0x0000FFFF0000FFFFLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
116 static uint64_t bm11001100= 0xFFFF0000FFFF0000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
117 static uint64_t b00= 0x0000000000000000LL; |
2179 | 118 static uint64_t b01= 0x0101010101010101LL; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
119 static uint64_t b02= 0x0202020202020202LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
120 static uint64_t b0F= 0x0F0F0F0F0F0F0F0FLL; |
2473 | 121 static uint64_t b04= 0x0404040404040404LL; |
122 static uint64_t b08= 0x0808080808080808LL; | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
123 static uint64_t bFF= 0xFFFFFFFFFFFFFFFFLL; |
2159 | 124 static uint64_t b20= 0x2020202020202020LL; |
125 static uint64_t b80= 0x8080808080808080LL; | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
126 static uint64_t b7E= 0x7E7E7E7E7E7E7E7ELL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
127 static uint64_t b7C= 0x7C7C7C7C7C7C7C7CLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
128 static uint64_t b3F= 0x3F3F3F3F3F3F3F3FLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
129 static uint64_t temp0=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
130 static uint64_t temp1=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
131 static uint64_t temp2=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
132 static uint64_t temp3=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
133 static uint64_t temp4=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
134 static uint64_t temp5=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
135 static uint64_t pQPb=0; |
2473 | 136 static uint64_t pQPb2=0; |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
137 static uint8_t tempBlocks[8*16*2]; //used for the horizontal code |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
138 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
139 int hFlatnessThreshold= 56 - 16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
140 int vFlatnessThreshold= 56 - 16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
141 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
142 //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
|
143 double maxClippedThreshold= 0.01; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
144 |
2389 | 145 int maxAllowedY=234; |
2285 | 146 int minAllowedY=16; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
147 |
2389 | 148 static struct PPFilter filters[]= |
149 { | |
150 {"hb", "hdeblock", 1, 1, 3, H_DEBLOCK}, | |
151 {"vb", "vdeblock", 1, 2, 4, V_DEBLOCK}, | |
152 {"vr", "rkvdeblock", 1, 2, 4, H_RK1_FILTER}, | |
153 {"h1", "x1hdeblock", 1, 1, 3, H_X1_FILTER}, | |
154 {"v1", "x1vdeblock", 1, 2, 4, V_X1_FILTER}, | |
155 {"dr", "dering", 1, 5, 6, DERING}, | |
156 {"al", "autolevels", 0, 1, 2, LEVEL_FIX}, | |
157 {"lb", "linblenddeint", 0, 1, 6, LINEAR_BLEND_DEINT_FILTER}, | |
158 {"li", "linipoldeint", 0, 1, 6, LINEAR_IPOL_DEINT_FILTER}, | |
159 {"ci", "cubicipoldeint", 0, 1, 6, CUBIC_IPOL_DEINT_FILTER}, | |
160 {"md", "mediandeint", 0, 1, 6, MEDIAN_DEINT_FILTER}, | |
161 {NULL, NULL,0,0,0,0} //End Marker | |
162 }; | |
163 | |
164 static char *replaceTable[]= | |
165 { | |
166 "default", "hdeblock:a,vdeblock:a,dering:a,autolevels", | |
167 "de", "hdeblock:a,vdeblock:a,dering:a,autolevels", | |
168 "fast", "x1hdeblock:a,x1vdeblock:a,dering:a,autolevels", | |
169 "fa", "x1hdeblock:a,x1vdeblock:a,dering:a,autolevels", | |
170 NULL //End Marker | |
171 }; | |
172 | |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
173 static inline void unusedVariableWarningFixer() |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
174 { |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
175 if( |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
176 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
|
177 + bm00001000 + bm10000000 + bm10000001 + bm11000011 + bm00000011 + bm11111110 |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
178 + bm11000000 + bm00011000 + bm00110011 + bm11001100 + b00 + b01 + b02 + b0F |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
179 + bFF + b20 + b80 + b7E + b7C + b3F + temp0 + temp1 + temp2 + temp3 + temp4 |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
180 + temp5 + pQPb== 0) b00=0; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
181 } |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
182 |
2221 | 183 #ifdef TIMING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
184 static inline long long rdtsc() |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
185 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
186 long long l; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
187 asm volatile( "rdtsc\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
188 : "=A" (l) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
189 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
190 // printf("%d\n", int(l/1000)); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
191 return l; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
192 } |
2185 | 193 #endif |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
194 |
2185 | 195 #ifdef HAVE_MMX2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
196 static inline void prefetchnta(void *p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
197 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
198 asm volatile( "prefetchnta (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
199 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
200 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
201 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
202 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
203 static inline void prefetcht0(void *p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
204 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
205 asm volatile( "prefetcht0 (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
206 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
207 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
208 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
209 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
210 static inline void prefetcht1(void *p) |
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 asm volatile( "prefetcht1 (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
213 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
214 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
215 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
216 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
217 static inline void prefetcht2(void *p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
218 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
219 asm volatile( "prefetcht2 (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
220 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
221 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
222 } |
2185 | 223 #endif |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
224 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
225 //FIXME? |255-0| = 1 (shouldnt be a problem ...) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
226 /** |
2246 | 227 * 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
|
228 */ |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
229 static inline int isVertDC(uint8_t src[], int stride){ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
230 int numEq= 0; |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
231 #ifndef HAVE_MMX |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
232 int y; |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
233 #endif |
2246 | 234 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
|
235 #ifdef HAVE_MMX |
2413 | 236 asm volatile( |
237 "leal (%1, %2), %%eax \n\t" | |
238 "leal (%%eax, %2, 4), %%ebx \n\t" | |
239 // 0 1 2 3 4 5 6 7 8 9 | |
240 // %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
|
241 "movq b7E, %%mm7 \n\t" // mm7 = 0x7F |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
242 "movq b7C, %%mm6 \n\t" // mm6 = 0x7D |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
243 "movq (%1), %%mm0 \n\t" |
2413 | 244 "movq (%%eax), %%mm1 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
245 "psubb %%mm1, %%mm0 \n\t" // mm0 = differnece |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
246 "paddb %%mm7, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
247 "pcmpgtb %%mm6, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
248 |
2413 | 249 "movq (%%eax,%2), %%mm2 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
250 "psubb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
251 "paddb %%mm7, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
252 "pcmpgtb %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
253 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
254 |
2413 | 255 "movq (%%eax, %2, 2), %%mm1 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
256 "psubb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
257 "paddb %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
258 "pcmpgtb %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
259 "paddb %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
260 |
2413 | 261 "movq (%1, %2, 4), %%mm2 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
262 "psubb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
263 "paddb %%mm7, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
264 "pcmpgtb %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
265 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
266 |
2413 | 267 "movq (%%ebx), %%mm1 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
268 "psubb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
269 "paddb %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
270 "pcmpgtb %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
271 "paddb %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
272 |
2413 | 273 "movq (%%ebx, %2), %%mm2 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
274 "psubb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
275 "paddb %%mm7, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
276 "pcmpgtb %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
277 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
278 |
2413 | 279 "movq (%%ebx, %2, 2), %%mm1 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
280 "psubb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
281 "paddb %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
282 "pcmpgtb %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
283 "paddb %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
284 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
285 " \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
286 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
287 "psrlw $8, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
288 "paddb %%mm1, %%mm0 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
289 #ifdef HAVE_MMX2 |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
290 "pshufw $0xF9, %%mm0, %%mm1 \n\t" |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
291 "paddb %%mm1, %%mm0 \n\t" |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
292 "pshufw $0xFE, %%mm0, %%mm1 \n\t" |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
293 #else |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
294 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
295 "psrlq $16, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
296 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
297 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
298 "psrlq $32, %%mm0 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
299 #endif |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
300 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
301 "movd %%mm0, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
302 : "=r" (numEq) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
303 : "r" (src), "r" (stride) |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
304 : "%eax", "%ebx" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
305 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
306 |
2413 | 307 numEq= (256 - numEq) &0xFF; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
308 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
309 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
310 for(y=0; y<BLOCK_SIZE-1; y++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
311 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
312 if(((src[0] - src[0+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
313 if(((src[1] - src[1+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
314 if(((src[2] - src[2+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
315 if(((src[3] - src[3+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
316 if(((src[4] - src[4+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
317 if(((src[5] - src[5+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
318 if(((src[6] - src[6+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
319 if(((src[7] - src[7+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
320 src+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
321 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
322 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
323 /* if(abs(numEq - asmEq) > 0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
324 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
325 printf("\nasm:%d c:%d\n", asmEq, numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
326 for(int y=0; y<8; y++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
327 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
328 for(int x=0; x<8; x++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
329 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
330 printf("%d ", temp[x + y*stride]); |
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 printf("\n"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
333 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
334 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
335 */ |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
336 // 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
|
337 return (numEq > vFlatnessThreshold) ? 1 : 0; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
338 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
339 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
340 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
|
341 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
342 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
343 int isOk; |
2246 | 344 src+= stride*3; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
345 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
346 // "int $3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
347 "movq (%1, %2), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
348 "movq (%1, %2, 8), %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
349 "movq %%mm0, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
350 "psubusb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
351 "psubusb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
352 "por %%mm1, %%mm0 \n\t" // ABS Diff |
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 "movq pQPb, %%mm7 \n\t" // QP,..., QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
355 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
356 "psubusb %%mm7, %%mm0 \n\t" // Diff <= 2QP -> 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
357 "pcmpeqd b00, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
358 "psrlq $16, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
359 "pcmpeqd bFF, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
360 // "movd %%mm0, (%1, %2, 4)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
361 "movd %%mm0, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
362 : "=r" (isOk) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
363 : "r" (src), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
364 ); |
2416 | 365 return isOk; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
366 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
367 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
368 int isOk2= 1; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
369 int x; |
2246 | 370 src+= stride*3; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
371 for(x=0; x<BLOCK_SIZE; x++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
372 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
373 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
|
374 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
375 /* if(isOk && !isOk2 || !isOk && isOk2) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
376 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
377 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
|
378 for(int y=0; y<9; y++) |
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 for(int x=0; x<8; x++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
381 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
382 printf("%d ", src[x + y*stride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
383 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
384 printf("\n"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
385 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
386 } */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
387 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
388 return isOk2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
389 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
390 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
391 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
392 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
393 /** |
2246 | 394 * Do a vertical low pass filter on the 8x16 block (only write to the 8x8 block in the middle) |
2221 | 395 * 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
|
396 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
397 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
|
398 { |
2159 | 399 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2246 | 400 src+= stride*3; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
401 asm volatile( //"movv %0 %1 %2\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
402 "pushl %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
403 "movq pQPb, %%mm0 \n\t" // QP,..., QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
404 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
405 "movq (%0), %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
406 "movq (%0, %1), %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
407 "movq %%mm5, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
408 "movq %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
409 "psubusb %%mm6, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
410 "psubusb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
411 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
412 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
413 "pcmpeqb b00, %%mm2 \n\t" // diff <= QP -> FF |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
414 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
415 "pand %%mm2, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
416 "pandn %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
417 "por %%mm2, %%mm6 \n\t"// First Line to Filter |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
418 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
419 "movq (%0, %1, 8), %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
420 "leal (%0, %1, 4), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
421 "leal (%0, %1, 8), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
422 "subl %1, %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
423 "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
|
424 "movq (%0, %1, 8), %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
425 "movq %%mm5, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
426 "movq %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
427 "psubusb %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
428 "psubusb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
429 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
430 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
431 "pcmpeqb b00, %%mm2 \n\t" // diff <= QP -> FF |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
432 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
433 "pand %%mm2, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
434 "pandn %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
435 "por %%mm2, %%mm7 \n\t" // First Line to Filter |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
436 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
437 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
438 // 1 2 3 4 5 6 7 8 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
439 // %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
|
440 // 6 4 2 2 1 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
441 // 6 4 4 2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
442 // 6 8 2 |
2246 | 443 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
444 "movq (%0, %1), %%mm0 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
445 "movq %%mm0, %%mm1 \n\t" // 1 |
2159 | 446 PAVGB(%%mm6, %%mm0) //1 1 /2 |
447 PAVGB(%%mm6, %%mm0) //3 1 /4 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
448 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
449 "movq (%0, %1, 4), %%mm2 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
450 "movq %%mm2, %%mm5 \n\t" // 1 |
2159 | 451 PAVGB((%%eax), %%mm2) // 11 /2 |
452 PAVGB((%0, %1, 2), %%mm2) // 211 /4 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
453 "movq %%mm2, %%mm3 \n\t" // 211 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
454 "movq (%0), %%mm4 \n\t" // 1 |
2159 | 455 PAVGB(%%mm4, %%mm3) // 4 211 /8 |
456 PAVGB(%%mm0, %%mm3) //642211 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
457 "movq %%mm3, (%0) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
458 // 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
|
459 "movq %%mm1, %%mm0 \n\t" // 1 |
2159 | 460 PAVGB(%%mm6, %%mm0) //1 1 /2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
461 "movq %%mm4, %%mm3 \n\t" // 1 |
2159 | 462 PAVGB((%0,%1,2), %%mm3) // 1 1 /2 |
463 PAVGB((%%eax,%1,2), %%mm5) // 11 /2 | |
464 PAVGB((%%eax), %%mm5) // 211 /4 | |
465 PAVGB(%%mm5, %%mm3) // 2 2211 /8 | |
466 PAVGB(%%mm0, %%mm3) //4242211 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
467 "movq %%mm3, (%0,%1) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
468 // mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9 |
2159 | 469 PAVGB(%%mm4, %%mm6) //11 /2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
470 "movq (%%ebx), %%mm0 \n\t" // 1 |
2159 | 471 PAVGB((%%eax, %1, 2), %%mm0) // 11/2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
472 "movq %%mm0, %%mm3 \n\t" // 11/2 |
2159 | 473 PAVGB(%%mm1, %%mm0) // 2 11/4 |
474 PAVGB(%%mm6, %%mm0) //222 11/8 | |
475 PAVGB(%%mm2, %%mm0) //22242211/16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
476 "movq (%0, %1, 2), %%mm2 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
477 "movq %%mm0, (%0, %1, 2) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
478 // 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
|
479 "movq (%%eax, %1, 4), %%mm0 \n\t" // 1 |
2159 | 480 PAVGB((%%ebx), %%mm0) // 11 /2 |
481 PAVGB(%%mm0, %%mm6) //11 11 /4 | |
482 PAVGB(%%mm1, %%mm4) // 11 /2 | |
483 PAVGB(%%mm2, %%mm1) // 11 /2 | |
484 PAVGB(%%mm1, %%mm6) //1122 11 /8 | |
485 PAVGB(%%mm5, %%mm6) //112242211 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
486 "movq (%%eax), %%mm5 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
487 "movq %%mm6, (%%eax) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
488 // 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
|
489 "movq (%%eax, %1, 4), %%mm6 \n\t" // 1 |
2159 | 490 PAVGB(%%mm7, %%mm6) // 11 /2 |
491 PAVGB(%%mm4, %%mm6) // 11 11 /4 | |
492 PAVGB(%%mm3, %%mm6) // 11 2211 /8 | |
493 PAVGB(%%mm5, %%mm2) // 11 /2 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
494 "movq (%0, %1, 4), %%mm4 \n\t" // 1 |
2159 | 495 PAVGB(%%mm4, %%mm2) // 112 /4 |
496 PAVGB(%%mm2, %%mm6) // 112242211 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
497 "movq %%mm6, (%0, %1, 4) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
498 // mm0=7(11) mm1=2(11) mm2=3(112) mm3=6(11) mm4=5 mm5=4 mm7=9 |
2159 | 499 PAVGB(%%mm7, %%mm1) // 11 2 /4 |
500 PAVGB(%%mm4, %%mm5) // 11 /2 | |
501 PAVGB(%%mm5, %%mm0) // 11 11 /4 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
502 "movq (%%eax, %1, 2), %%mm6 \n\t" // 1 |
2159 | 503 PAVGB(%%mm6, %%mm1) // 11 4 2 /8 |
504 PAVGB(%%mm0, %%mm1) // 11224222 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
505 "movq %%mm1, (%%eax, %1, 2) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
506 // mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9 |
2159 | 507 PAVGB((%%ebx), %%mm2) // 112 4 /8 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
508 "movq (%%eax, %1, 4), %%mm0 \n\t" // 1 |
2159 | 509 PAVGB(%%mm0, %%mm6) // 1 1 /2 |
510 PAVGB(%%mm7, %%mm6) // 1 12 /4 | |
511 PAVGB(%%mm2, %%mm6) // 1122424 /4 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
512 "movq %%mm6, (%%ebx) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
513 // mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9 |
2159 | 514 PAVGB(%%mm7, %%mm5) // 11 2 /4 |
515 PAVGB(%%mm7, %%mm5) // 11 6 /8 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
516 |
2159 | 517 PAVGB(%%mm3, %%mm0) // 112 /4 |
518 PAVGB(%%mm0, %%mm5) // 112246 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
519 "movq %%mm5, (%%eax, %1, 4) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
520 "popl %0\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
521 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
522 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
523 : "r" (src), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
524 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
525 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
526 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
527 const int l1= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
528 const int l2= stride + l1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
529 const int l3= stride + l2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
530 const int l4= stride + l3; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
531 const int l5= stride + l4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
532 const int l6= stride + l5; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
533 const int l7= stride + l6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
534 const int l8= stride + l7; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
535 const int l9= stride + l8; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
536 int x; |
2246 | 537 src+= stride*3; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
538 for(x=0; x<BLOCK_SIZE; x++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
539 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
540 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
|
541 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
|
542 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
543 int sums[9]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
544 sums[0] = first + src[l1]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
545 sums[1] = src[l1] + src[l2]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
546 sums[2] = src[l2] + src[l3]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
547 sums[3] = src[l3] + src[l4]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
548 sums[4] = src[l4] + src[l5]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
549 sums[5] = src[l5] + src[l6]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
550 sums[6] = src[l6] + src[l7]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
551 sums[7] = src[l7] + src[l8]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
552 sums[8] = src[l8] + last; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
553 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
554 src[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
|
555 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
|
556 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
|
557 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
|
558 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
|
559 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
|
560 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
|
561 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
|
562 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
563 src++; |
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 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
566 #endif |
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 |
2159 | 569 /** |
570 * Experimental implementation of the filter (Algorithm 1) described in a paper from Ramkishor & Karandikar | |
571 * values are correctly clipped (MMX2) | |
572 * values are wraparound (C) | |
573 * conclusion: its fast, but introduces ugly horizontal patterns if there is a continious gradient | |
574 0 8 16 24 | |
575 x = 8 | |
576 x/2 = 4 | |
577 x/8 = 1 | |
578 1 12 12 23 | |
579 */ | |
2179 | 580 static inline void vertRK1Filter(uint8_t *src, int stride, int QP) |
2159 | 581 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
582 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2246 | 583 src+= stride*3; |
2159 | 584 // FIXME rounding |
585 asm volatile( | |
586 "pxor %%mm7, %%mm7 \n\t" // 0 | |
587 "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE | |
588 "leal (%0, %1), %%eax \n\t" | |
589 "leal (%%eax, %1, 4), %%ebx \n\t" | |
590 // 0 1 2 3 4 5 6 7 8 9 | |
591 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 | |
592 "movq pQPb, %%mm0 \n\t" // QP,..., QP | |
593 "movq %%mm0, %%mm1 \n\t" // QP,..., QP | |
594 "paddusb b02, %%mm0 \n\t" | |
595 "psrlw $2, %%mm0 \n\t" | |
596 "pand b3F, %%mm0 \n\t" // QP/4,..., QP/4 | |
597 "paddusb %%mm1, %%mm0 \n\t" // QP*1.25 ... | |
598 "movq (%0, %1, 4), %%mm2 \n\t" // line 4 | |
599 "movq (%%ebx), %%mm3 \n\t" // line 5 | |
600 "movq %%mm2, %%mm4 \n\t" // line 4 | |
601 "pcmpeqb %%mm5, %%mm5 \n\t" // -1 | |
602 "pxor %%mm2, %%mm5 \n\t" // -line 4 - 1 | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
603 PAVGB(%%mm3, %%mm5) |
2159 | 604 "paddb %%mm6, %%mm5 \n\t" // (l5-l4)/2 |
605 "psubusb %%mm3, %%mm4 \n\t" | |
606 "psubusb %%mm2, %%mm3 \n\t" | |
607 "por %%mm3, %%mm4 \n\t" // |l4 - l5| | |
608 "psubusb %%mm0, %%mm4 \n\t" | |
609 "pcmpeqb %%mm7, %%mm4 \n\t" | |
610 "pand %%mm4, %%mm5 \n\t" // d/2 | |
611 | |
612 // "paddb %%mm6, %%mm2 \n\t" // line 4 + 0x80 | |
613 "paddb %%mm5, %%mm2 \n\t" | |
614 // "psubb %%mm6, %%mm2 \n\t" | |
615 "movq %%mm2, (%0,%1, 4) \n\t" | |
616 | |
617 "movq (%%ebx), %%mm2 \n\t" | |
618 // "paddb %%mm6, %%mm2 \n\t" // line 5 + 0x80 | |
619 "psubb %%mm5, %%mm2 \n\t" | |
620 // "psubb %%mm6, %%mm2 \n\t" | |
621 "movq %%mm2, (%%ebx) \n\t" | |
622 | |
623 "paddb %%mm6, %%mm5 \n\t" | |
624 "psrlw $2, %%mm5 \n\t" | |
625 "pand b3F, %%mm5 \n\t" | |
626 "psubb b20, %%mm5 \n\t" // (l5-l4)/8 | |
627 | |
628 "movq (%%eax, %1, 2), %%mm2 \n\t" | |
629 "paddb %%mm6, %%mm2 \n\t" // line 3 + 0x80 | |
630 "paddsb %%mm5, %%mm2 \n\t" | |
631 "psubb %%mm6, %%mm2 \n\t" | |
632 "movq %%mm2, (%%eax, %1, 2) \n\t" | |
633 | |
634 "movq (%%ebx, %1), %%mm2 \n\t" | |
635 "paddb %%mm6, %%mm2 \n\t" // line 6 + 0x80 | |
636 "psubsb %%mm5, %%mm2 \n\t" | |
637 "psubb %%mm6, %%mm2 \n\t" | |
638 "movq %%mm2, (%%ebx, %1) \n\t" | |
639 | |
640 : | |
641 : "r" (src), "r" (stride) | |
642 : "%eax", "%ebx" | |
643 ); | |
644 #else | |
645 const int l1= stride; | |
646 const int l2= stride + l1; | |
647 const int l3= stride + l2; | |
648 const int l4= stride + l3; | |
649 const int l5= stride + l4; | |
650 const int l6= stride + l5; | |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
651 // const int l7= stride + l6; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
652 // const int l8= stride + l7; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
653 // const int l9= stride + l8; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
654 int x; |
2246 | 655 src+= stride*3; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
656 for(x=0; x<BLOCK_SIZE; x++) |
2159 | 657 { |
658 if(ABS(src[l4]-src[l5]) < QP + QP/4) | |
659 { | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
660 int v = (src[l5] - src[l4]); |
2159 | 661 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
662 src[l3] +=v/8; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
663 src[l4] +=v/2; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
664 src[l5] -=v/2; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
665 src[l6] -=v/8; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
666 |
2159 | 667 } |
668 src++; | |
669 } | |
670 | |
671 #endif | |
672 } | |
673 | |
674 /** | |
675 * Experimental Filter 1 | |
2179 | 676 * will not damage linear gradients |
677 * 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
|
678 * 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
|
679 * MMX2 version does correct clipping C version doesnt |
2159 | 680 */ |
681 static inline void vertX1Filter(uint8_t *src, int stride, int QP) | |
682 { | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
683 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2246 | 684 src+= stride*3; |
685 | |
2159 | 686 asm volatile( |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
687 "pxor %%mm7, %%mm7 \n\t" // 0 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
688 // "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
689 "leal (%0, %1), %%eax \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
690 "leal (%%eax, %1, 4), %%ebx \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
691 // 0 1 2 3 4 5 6 7 8 9 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
692 // %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
|
693 "movq (%%eax, %1, 2), %%mm0 \n\t" // line 3 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
694 "movq (%0, %1, 4), %%mm1 \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
695 "movq %%mm1, %%mm2 \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
696 "psubusb %%mm0, %%mm1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
697 "psubusb %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
698 "por %%mm1, %%mm0 \n\t" // |l2 - l3| |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
699 "movq (%%ebx), %%mm3 \n\t" // line 5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
700 "movq (%%ebx, %1), %%mm4 \n\t" // line 6 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
701 "movq %%mm3, %%mm5 \n\t" // line 5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
702 "psubusb %%mm4, %%mm3 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
703 "psubusb %%mm5, %%mm4 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
704 "por %%mm4, %%mm3 \n\t" // |l5 - l6| |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
705 PAVGB(%%mm3, %%mm0) // (|l2 - l3| + |l5 - l6|)/2 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
706 "movq %%mm2, %%mm1 \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
707 "psubusb %%mm5, %%mm2 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
708 "movq %%mm2, %%mm4 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
709 "pcmpeqb %%mm7, %%mm2 \n\t" // (l4 - l5) <= 0 ? -1 : 0 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
710 "psubusb %%mm1, %%mm5 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
711 "por %%mm5, %%mm4 \n\t" // |l4 - l5| |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
712 "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
|
713 "movq %%mm4, %%mm3 \n\t" // d |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
714 "psubusb pQPb, %%mm4 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
715 "pcmpeqb %%mm7, %%mm4 \n\t" // d <= QP ? -1 : 0 |
2179 | 716 "psubusb b01, %%mm3 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
717 "pand %%mm4, %%mm3 \n\t" // d <= QP ? d : 0 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
718 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
719 PAVGB(%%mm7, %%mm3) // d/2 |
2179 | 720 "movq %%mm3, %%mm1 \n\t" // d/2 |
721 PAVGB(%%mm7, %%mm3) // d/4 | |
722 PAVGB(%%mm1, %%mm3) // 3*d/8 | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
723 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
724 "movq (%0, %1, 4), %%mm0 \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
725 "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
|
726 "psubusb %%mm3, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
727 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
728 "movq %%mm0, (%0, %1, 4) \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
729 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
730 "movq (%%ebx), %%mm0 \n\t" // line 5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
731 "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
|
732 "paddusb %%mm3, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
733 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
734 "movq %%mm0, (%%ebx) \n\t" // line 5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
735 |
2179 | 736 PAVGB(%%mm7, %%mm1) // d/4 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
737 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
738 "movq (%%eax, %1, 2), %%mm0 \n\t" // line 3 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
739 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4 |
2179 | 740 "psubusb %%mm1, %%mm0 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
741 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
742 "movq %%mm0, (%%eax, %1, 2) \n\t" // line 3 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
743 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
744 "movq (%%ebx, %1), %%mm0 \n\t" // line 6 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
745 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5 |
2179 | 746 "paddusb %%mm1, %%mm0 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
747 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
748 "movq %%mm0, (%%ebx, %1) \n\t" // line 6 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
749 |
2179 | 750 PAVGB(%%mm7, %%mm1) // d/8 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
751 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
752 "movq (%%eax, %1), %%mm0 \n\t" // line 2 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
753 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l2-1 : l2 |
2179 | 754 "psubusb %%mm1, %%mm0 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
755 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
756 "movq %%mm0, (%%eax, %1) \n\t" // line 2 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
757 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
758 "movq (%%ebx, %1, 2), %%mm0 \n\t" // line 7 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
759 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l7-1 : l7 |
2179 | 760 "paddusb %%mm1, %%mm0 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
761 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
762 "movq %%mm0, (%%ebx, %1, 2) \n\t" // line 7 |
2159 | 763 |
764 : | |
765 : "r" (src), "r" (stride) | |
766 : "%eax", "%ebx" | |
767 ); | |
768 #else | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
769 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
770 const int l1= stride; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
771 const int l2= stride + l1; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
772 const int l3= stride + l2; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
773 const int l4= stride + l3; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
774 const int l5= stride + l4; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
775 const int l6= stride + l5; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
776 const int l7= stride + l6; |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
777 // const int l8= stride + l7; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
778 // const int l9= stride + l8; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
779 int x; |
2246 | 780 |
781 src+= stride*3; | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
782 for(x=0; x<BLOCK_SIZE; x++) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
783 { |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
784 int a= src[l3] - src[l4]; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
785 int b= src[l4] - src[l5]; |
2179 | 786 int c= src[l5] - src[l6]; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
787 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
788 int d= MAX(ABS(b) - (ABS(a) + ABS(c))/2, 0); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
789 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
790 if(d < QP) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
791 { |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
792 int v = d * SIGN(-b); |
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 src[l2] +=v/8; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
795 src[l3] +=v/4; |
2179 | 796 src[l4] +=3*v/8; |
797 src[l5] -=3*v/8; | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
798 src[l6] -=v/4; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
799 src[l7] -=v/8; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
800 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
801 } |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
802 src++; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
803 } |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
804 /* |
2159 | 805 const int l1= stride; |
806 const int l2= stride + l1; | |
807 const int l3= stride + l2; | |
808 const int l4= stride + l3; | |
809 const int l5= stride + l4; | |
810 const int l6= stride + l5; | |
811 const int l7= stride + l6; | |
812 const int l8= stride + l7; | |
813 const int l9= stride + l8; | |
814 for(int x=0; x<BLOCK_SIZE; x++) | |
815 { | |
816 int v2= src[l2]; | |
817 int v3= src[l3]; | |
818 int v4= src[l4]; | |
819 int v5= src[l5]; | |
820 int v6= src[l6]; | |
821 int v7= src[l7]; | |
822 | |
823 if(ABS(v4-v5)<QP && ABS(v4-v5) - (ABS(v3-v4) + ABS(v5-v6))>0 ) | |
824 { | |
825 src[l3] = (6*v2 + 4*v3 + 3*v4 + 2*v5 + v6 )/16; | |
826 src[l4] = (3*v2 + 3*v3 + 4*v4 + 3*v5 + 2*v6 + v7 )/16; | |
827 src[l5] = (1*v2 + 2*v3 + 3*v4 + 4*v5 + 3*v6 + 3*v7)/16; | |
828 src[l6] = ( 1*v3 + 2*v4 + 3*v5 + 4*v6 + 6*v7)/16; | |
829 } | |
830 src++; | |
831 } | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
832 */ |
2159 | 833 #endif |
834 } | |
835 | |
2179 | 836 /** |
837 * Experimental Filter 1 (Horizontal) | |
838 * will not damage linear gradients | |
839 * Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter | |
840 * can only smooth blocks at the expected locations (it cant smooth them if they did move) | |
841 * MMX2 version does correct clipping C version doesnt | |
842 * not identical with the vertical one | |
843 */ | |
844 static inline void horizX1Filter(uint8_t *src, int stride, int QP) | |
845 { | |
846 int y; | |
847 static uint64_t *lut= NULL; | |
848 if(lut==NULL) | |
849 { | |
850 int i; | |
851 lut= (uint64_t*)memalign(8, 256*8); | |
852 for(i=0; i<256; i++) | |
853 { | |
854 int v= i < 128 ? 2*i : 2*(i-256); | |
855 /* | |
856 //Simulate 112242211 9-Tap filter | |
857 uint64_t a= (v/16) & 0xFF; | |
858 uint64_t b= (v/8) & 0xFF; | |
859 uint64_t c= (v/4) & 0xFF; | |
860 uint64_t d= (3*v/8) & 0xFF; | |
861 */ | |
862 //Simulate piecewise linear interpolation | |
863 uint64_t a= (v/16) & 0xFF; | |
864 uint64_t b= (v*3/16) & 0xFF; | |
865 uint64_t c= (v*5/16) & 0xFF; | |
866 uint64_t d= (7*v/16) & 0xFF; | |
867 uint64_t A= (0x100 - a)&0xFF; | |
868 uint64_t B= (0x100 - b)&0xFF; | |
869 uint64_t C= (0x100 - c)&0xFF; | |
870 uint64_t D= (0x100 - c)&0xFF; | |
871 | |
872 lut[i] = (a<<56) | (b<<48) | (c<<40) | (d<<32) | | |
873 (D<<24) | (C<<16) | (B<<8) | (A); | |
874 //lut[i] = (v<<32) | (v<<24); | |
875 } | |
876 } | |
877 | |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
878 #if 0 |
2179 | 879 asm volatile( |
880 "pxor %%mm7, %%mm7 \n\t" // 0 | |
881 // "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE | |
882 "leal (%0, %1), %%eax \n\t" | |
883 "leal (%%eax, %1, 4), %%ebx \n\t" | |
884 | |
885 "movq b80, %%mm6 \n\t" | |
2180 | 886 "movd pQPb, %%mm5 \n\t" // QP |
2179 | 887 "movq %%mm5, %%mm4 \n\t" |
888 "paddusb %%mm5, %%mm5 \n\t" // 2QP | |
889 "paddusb %%mm5, %%mm4 \n\t" // 3QP | |
890 "pxor %%mm5, %%mm5 \n\t" // 0 | |
891 "psubb %%mm4, %%mm5 \n\t" // -3QP | |
892 "por bm11111110, %%mm5 \n\t" // ...,FF,FF,-3QP | |
893 "psllq $24, %%mm5 \n\t" | |
894 | |
895 // 0 1 2 3 4 5 6 7 8 9 | |
896 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 | |
897 | |
898 #define HX1old(a) \ | |
899 "movd " #a ", %%mm0 \n\t"\ | |
900 "movd 4" #a ", %%mm1 \n\t"\ | |
901 "punpckldq %%mm1, %%mm0 \n\t"\ | |
902 "movq %%mm0, %%mm1 \n\t"\ | |
903 "movq %%mm0, %%mm2 \n\t"\ | |
904 "psrlq $8, %%mm1 \n\t"\ | |
905 "psubusb %%mm1, %%mm2 \n\t"\ | |
906 "psubusb %%mm0, %%mm1 \n\t"\ | |
907 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ | |
908 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
909 "pshufw $0x00, %%mm1, %%mm3 \n\t" /* p´5 = |p1 - p2| */\ | |
910 PAVGB(%%mm1, %%mm3) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\ | |
911 "psrlq $16, %%mm3 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\ | |
912 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\ | |
913 "paddb %%mm5, %%mm1 \n\t"\ | |
914 "psubusb %%mm5, %%mm1 \n\t"\ | |
915 PAVGB(%%mm7, %%mm1)\ | |
916 "pxor %%mm2, %%mm1 \n\t"\ | |
917 "psubb %%mm2, %%mm1 \n\t"\ | |
918 "psrlq $24, %%mm1 \n\t"\ | |
919 "movd %%mm1, %%ecx \n\t"\ | |
920 "paddb %%mm6, %%mm0 \n\t"\ | |
921 "paddsb (%3, %%ecx, 8), %%mm0 \n\t"\ | |
922 "paddb %%mm6, %%mm0 \n\t"\ | |
923 "movq %%mm0, " #a " \n\t"\ | |
924 | |
925 /* | |
926 HX1old((%0)) | |
927 HX1old((%%eax)) | |
928 HX1old((%%eax, %1)) | |
929 HX1old((%%eax, %1, 2)) | |
930 HX1old((%0, %1, 4)) | |
931 HX1old((%%ebx)) | |
932 HX1old((%%ebx, %1)) | |
933 HX1old((%%ebx, %1, 2)) | |
934 */ | |
935 | |
936 //FIXME add some comments, its unreadable ... | |
937 #define HX1b(a, c, b, d) \ | |
938 "movd " #a ", %%mm0 \n\t"\ | |
939 "movd 4" #a ", %%mm1 \n\t"\ | |
940 "punpckldq %%mm1, %%mm0 \n\t"\ | |
941 "movd " #b ", %%mm4 \n\t"\ | |
942 "movq %%mm0, %%mm1 \n\t"\ | |
943 "movq %%mm0, %%mm2 \n\t"\ | |
944 "psrlq $8, %%mm1 \n\t"\ | |
945 "movd 4" #b ", %%mm3 \n\t"\ | |
946 "psubusb %%mm1, %%mm2 \n\t"\ | |
947 "psubusb %%mm0, %%mm1 \n\t"\ | |
948 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ | |
949 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
950 "punpckldq %%mm3, %%mm4 \n\t"\ | |
951 "movq %%mm1, %%mm3 \n\t"\ | |
952 "psllq $32, %%mm3 \n\t" /* p´5 = |p1 - p2| */\ | |
953 PAVGB(%%mm1, %%mm3) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\ | |
954 "paddb %%mm6, %%mm0 \n\t"\ | |
955 "psrlq $16, %%mm3 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\ | |
956 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\ | |
957 "movq %%mm4, %%mm3 \n\t"\ | |
958 "paddb %%mm5, %%mm1 \n\t"\ | |
959 "psubusb %%mm5, %%mm1 \n\t"\ | |
960 "psrlq $8, %%mm3 \n\t"\ | |
961 PAVGB(%%mm7, %%mm1)\ | |
962 "pxor %%mm2, %%mm1 \n\t"\ | |
963 "psubb %%mm2, %%mm1 \n\t"\ | |
964 "movq %%mm4, %%mm2 \n\t"\ | |
965 "psrlq $24, %%mm1 \n\t"\ | |
966 "psubusb %%mm3, %%mm2 \n\t"\ | |
967 "movd %%mm1, %%ecx \n\t"\ | |
968 "psubusb %%mm4, %%mm3 \n\t"\ | |
2180 | 969 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\ |
2179 | 970 "por %%mm2, %%mm3 \n\t" /* p´x = |px - p(x+1)| */\ |
971 "paddb %%mm6, %%mm0 \n\t"\ | |
972 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
973 "movq %%mm3, %%mm1 \n\t"\ | |
974 "psllq $32, %%mm1 \n\t" /* p´5 = |p1 - p2| */\ | |
975 "movq %%mm0, " #a " \n\t"\ | |
976 PAVGB(%%mm3, %%mm1) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\ | |
977 "paddb %%mm6, %%mm4 \n\t"\ | |
978 "psrlq $16, %%mm1 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\ | |
979 "psubusb %%mm1, %%mm3 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\ | |
980 "paddb %%mm5, %%mm3 \n\t"\ | |
981 "psubusb %%mm5, %%mm3 \n\t"\ | |
982 PAVGB(%%mm7, %%mm3)\ | |
983 "pxor %%mm2, %%mm3 \n\t"\ | |
984 "psubb %%mm2, %%mm3 \n\t"\ | |
985 "psrlq $24, %%mm3 \n\t"\ | |
986 "movd " #c ", %%mm0 \n\t"\ | |
987 "movd 4" #c ", %%mm1 \n\t"\ | |
988 "punpckldq %%mm1, %%mm0 \n\t"\ | |
989 "paddb %%mm6, %%mm0 \n\t"\ | |
2180 | 990 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\ |
2179 | 991 "paddb %%mm6, %%mm0 \n\t"\ |
992 "movq %%mm0, " #c " \n\t"\ | |
993 "movd %%mm3, %%ecx \n\t"\ | |
994 "movd " #d ", %%mm0 \n\t"\ | |
2180 | 995 "paddsb (%2, %%ecx, 8), %%mm4 \n\t"\ |
2179 | 996 "movd 4" #d ", %%mm1 \n\t"\ |
997 "paddb %%mm6, %%mm4 \n\t"\ | |
998 "punpckldq %%mm1, %%mm0 \n\t"\ | |
999 "movq %%mm4, " #b " \n\t"\ | |
1000 "paddb %%mm6, %%mm0 \n\t"\ | |
2180 | 1001 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\ |
2179 | 1002 "paddb %%mm6, %%mm0 \n\t"\ |
1003 "movq %%mm0, " #d " \n\t"\ | |
1004 | |
1005 HX1b((%0),(%%eax),(%%eax, %1),(%%eax, %1, 2)) | |
1006 HX1b((%0, %1, 4),(%%ebx),(%%ebx, %1),(%%ebx, %1, 2)) | |
1007 | |
1008 | |
1009 : | |
2180 | 1010 : "r" (src), "r" (stride), "r" (lut) |
2179 | 1011 : "%eax", "%ebx", "%ecx" |
1012 ); | |
1013 #else | |
1014 | |
1015 //FIXME (has little in common with the mmx2 version) | |
1016 for(y=0; y<BLOCK_SIZE; y++) | |
1017 { | |
1018 int a= src[1] - src[2]; | |
1019 int b= src[3] - src[4]; | |
1020 int c= src[5] - src[6]; | |
1021 | |
1022 int d= MAX(ABS(b) - (ABS(a) + ABS(c))/2, 0); | |
1023 | |
1024 if(d < QP) | |
1025 { | |
1026 int v = d * SIGN(-b); | |
1027 | |
1028 src[1] +=v/8; | |
1029 src[2] +=v/4; | |
1030 src[3] +=3*v/8; | |
1031 src[4] -=3*v/8; | |
1032 src[5] -=v/4; | |
1033 src[6] -=v/8; | |
1034 | |
1035 } | |
1036 src+=stride; | |
1037 } | |
1038 #endif | |
1039 } | |
1040 | |
2159 | 1041 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1042 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
|
1043 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1044 #ifdef HAVE_MMX |
2246 | 1045 src+= stride*4; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1046 //FIXME try pmul for *5 stuff |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1047 // src[0]=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1048 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1049 "pxor %%mm7, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1050 "leal (%0, %1), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1051 "leal (%%eax, %1, 4), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1052 // 0 1 2 3 4 5 6 7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1053 // %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
|
1054 // %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
|
1055 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1056 "movq (%0), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1057 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1058 "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1059 "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1060 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1061 "movq (%%eax), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1062 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1063 "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1064 "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1065 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1066 "movq (%%eax, %1), %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1067 "movq %%mm4, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1068 "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1069 "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1070 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1071 "paddw %%mm0, %%mm0 \n\t" // 2L0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1072 "paddw %%mm1, %%mm1 \n\t" // 2H0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1073 "psubw %%mm4, %%mm2 \n\t" // L1 - L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1074 "psubw %%mm5, %%mm3 \n\t" // H1 - H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1075 "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1076 "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1077 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1078 "psllw $2, %%mm2 \n\t" // 4L1 - 4L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1079 "psllw $2, %%mm3 \n\t" // 4H1 - 4H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1080 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1081 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1082 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1083 "movq (%%eax, %1, 2), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1084 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1085 "punpcklbw %%mm7, %%mm2 \n\t" // L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1086 "punpckhbw %%mm7, %%mm3 \n\t" // H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1087 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1088 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1089 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1090 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1091 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1092 "movq %%mm0, temp0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1093 "movq %%mm1, temp1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1094 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1095 "movq (%0, %1, 4), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1096 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1097 "punpcklbw %%mm7, %%mm0 \n\t" // L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1098 "punpckhbw %%mm7, %%mm1 \n\t" // H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1099 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1100 "psubw %%mm0, %%mm2 \n\t" // L3 - L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1101 "psubw %%mm1, %%mm3 \n\t" // H3 - H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1102 "movq %%mm2, temp2 \n\t" // L3 - L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1103 "movq %%mm3, temp3 \n\t" // H3 - H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1104 "paddw %%mm4, %%mm4 \n\t" // 2L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1105 "paddw %%mm5, %%mm5 \n\t" // 2H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1106 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1107 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1108 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1109 "psllw $2, %%mm2 \n\t" // 4L3 - 4L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1110 "psllw $2, %%mm3 \n\t" // 4H3 - 4H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1111 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1112 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1113 //50 opcodes so far |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1114 "movq (%%ebx), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1115 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1116 "punpcklbw %%mm7, %%mm2 \n\t" // L5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1117 "punpckhbw %%mm7, %%mm3 \n\t" // H5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1118 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1119 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1120 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1121 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1122 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1123 "movq (%%ebx, %1), %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1124 "punpcklbw %%mm7, %%mm6 \n\t" // L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1125 "psubw %%mm6, %%mm2 \n\t" // L5 - L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1126 "movq (%%ebx, %1), %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1127 "punpckhbw %%mm7, %%mm6 \n\t" // H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1128 "psubw %%mm6, %%mm3 \n\t" // H5 - H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1129 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1130 "paddw %%mm0, %%mm0 \n\t" // 2L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1131 "paddw %%mm1, %%mm1 \n\t" // 2H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1132 "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1133 "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1134 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1135 "psllw $2, %%mm2 \n\t" // 4L5 - 4L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1136 "psllw $2, %%mm3 \n\t" // 4H5 - 4H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1137 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1138 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1139 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1140 "movq (%%ebx, %1, 2), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1141 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1142 "punpcklbw %%mm7, %%mm2 \n\t" // L7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1143 "punpckhbw %%mm7, %%mm3 \n\t" // H7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1144 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1145 "paddw %%mm2, %%mm2 \n\t" // 2L7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1146 "paddw %%mm3, %%mm3 \n\t" // 2H7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1147 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1148 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1149 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1150 "movq temp0, %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1151 "movq temp1, %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1152 //FIXME pxor, psubw, pmax for abs |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1153 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1154 "pcmpgtw %%mm0, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1155 "pxor %%mm6, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1156 "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1157 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1158 "pcmpgtw %%mm1, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1159 "pxor %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1160 "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1161 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1162 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1163 "pcmpgtw %%mm2, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1164 "pxor %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1165 "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1166 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1167 "pcmpgtw %%mm3, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1168 "pxor %%mm6, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1169 "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1170 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1171 #ifdef HAVE_MMX2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1172 "pminsw %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1173 "pminsw %%mm3, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1174 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1175 "movq %%mm0, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1176 "psubusw %%mm2, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1177 "psubw %%mm6, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1178 "movq %%mm1, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1179 "psubusw %%mm3, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1180 "psubw %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1181 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1182 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1183 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1184 "pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1185 "pxor %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1186 "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1187 "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1188 "pxor %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1189 "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1190 // 100 opcodes |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1191 "movd %2, %%mm2 \n\t" // QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1192 "punpcklwd %%mm2, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1193 "punpcklwd %%mm2, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1194 "psllw $3, %%mm2 \n\t" // 8QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1195 "movq %%mm2, %%mm3 \n\t" // 8QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1196 "pcmpgtw %%mm4, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1197 "pcmpgtw %%mm5, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1198 "pand %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1199 "pand %%mm3, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1200 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1201 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1202 "psubusw %%mm0, %%mm4 \n\t" // hd |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1203 "psubusw %%mm1, %%mm5 \n\t" // ld |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1204 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1205 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1206 "movq w05, %%mm2 \n\t" // 5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1207 "pmullw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1208 "pmullw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1209 "movq w20, %%mm2 \n\t" // 32 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1210 "paddw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1211 "paddw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1212 "psrlw $6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1213 "psrlw $6, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1214 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1215 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1216 "movq w06, %%mm2 \n\t" // 6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1217 "paddw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1218 "paddw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1219 "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
|
1220 //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
|
1221 "pmulhw %%mm2, %%mm4 \n\t" // hd/13 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1222 "pmulhw %%mm2, %%mm5 \n\t" // ld/13 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1223 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1224 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1225 "movq temp2, %%mm0 \n\t" // L3 - L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1226 "movq temp3, %%mm1 \n\t" // H3 - H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1227 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1228 "pxor %%mm2, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1229 "pxor %%mm3, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1230 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1231 "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1232 "pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1233 "pxor %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1234 "pxor %%mm3, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1235 "psubw %%mm2, %%mm0 \n\t" // |L3-L4| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1236 "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
|
1237 "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
|
1238 "psrlw $1, %%mm1 \n\t" // |H3 - H4|/2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1239 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1240 "pxor %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1241 "pxor %%mm7, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1242 "pand %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1243 "pand %%mm3, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1244 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1245 #ifdef HAVE_MMX2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1246 "pminsw %%mm0, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1247 "pminsw %%mm1, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1248 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1249 "movq %%mm4, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1250 "psubusw %%mm0, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1251 "psubw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1252 "movq %%mm5, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1253 "psubusw %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1254 "psubw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1255 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1256 "pxor %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1257 "pxor %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1258 "psubw %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1259 "psubw %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1260 "packsswb %%mm5, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1261 "movq (%%eax, %1, 2), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1262 "paddb %%mm4, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1263 "movq %%mm0, (%%eax, %1, 2) \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1264 "movq (%0, %1, 4), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1265 "psubb %%mm4, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1266 "movq %%mm0, (%0, %1, 4) \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1267 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1268 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1269 : "r" (src), "r" (stride), "r" (QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1270 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1271 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1272 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1273 const int l1= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1274 const int l2= stride + l1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1275 const int l3= stride + l2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1276 const int l4= stride + l3; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1277 const int l5= stride + l4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1278 const int l6= stride + l5; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1279 const int l7= stride + l6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1280 const int l8= stride + l7; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1281 // const int l9= stride + l8; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1282 int x; |
2246 | 1283 src+= stride*3; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1284 for(x=0; x<BLOCK_SIZE; x++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1285 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1286 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
|
1287 if(ABS(middleEnergy) < 8*QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1288 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1289 const int q=(src[l4] - src[l5])/2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1290 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
|
1291 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
|
1292 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1293 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1294 d= MAX(d, 0); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1295 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1296 d= (5*d + 32) >> 6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1297 d*= SIGN(-middleEnergy); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1298 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1299 if(q>0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1300 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1301 d= d<0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1302 d= d>q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1303 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1304 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1305 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1306 d= d>0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1307 d= d<q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1308 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1309 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1310 src[l4]-= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1311 src[l5]+= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1312 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1313 src++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1314 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1315 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1316 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1317 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1318 //FIXME? |255-0| = 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1319 /** |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1320 * Check if the given 8x8 Block is mostly "flat" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1321 */ |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1322 static inline int isHorizDC(uint8_t src[], int stride) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1323 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1324 // src++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1325 int numEq= 0; |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1326 #if 0 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1327 asm volatile ( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1328 // "int $3 \n\t" |
2416 | 1329 "leal (%1, %2), %%ecx \n\t" |
1330 "leal (%%ecx, %2, 4), %%ebx \n\t" | |
1331 // 0 1 2 3 4 5 6 7 8 9 | |
1332 // %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
|
1333 "movq b7E, %%mm7 \n\t" // mm7 = 0x7F |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1334 "movq b7C, %%mm6 \n\t" // mm6 = 0x7D |
2416 | 1335 "pxor %%mm0, %%mm0 \n\t" |
1336 "movl %1, %%eax \n\t" | |
1337 "andl $0x1F, %%eax \n\t" | |
1338 "cmpl $24, %%eax \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1339 "leal tempBlock, %%eax \n\t" |
2416 | 1340 "jb 1f \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1341 |
2416 | 1342 #define HDC_CHECK_AND_CPY(src, dst) \ |
1343 "movd " #src ", %%mm2 \n\t"\ | |
1344 "punpckldq 4" #src ", %%mm2 \n\t" /* (%1) */\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1345 "movq %%mm2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1346 "psrlq $8, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1347 "psubb %%mm1, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1348 "paddb %%mm7, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1349 "pcmpgtb %%mm6, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1350 "paddb %%mm2, %%mm0 \n\t"\ |
2416 | 1351 "movq %%mm1," #dst "(%%eax) \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1352 |
2416 | 1353 HDC_CHECK_AND_CPY((%1),0) |
1354 HDC_CHECK_AND_CPY((%%ecx),8) | |
1355 HDC_CHECK_AND_CPY((%%ecx, %2),16) | |
1356 HDC_CHECK_AND_CPY((%%ecx, %2, 2),24) | |
1357 HDC_CHECK_AND_CPY((%1, %2, 4),32) | |
1358 HDC_CHECK_AND_CPY((%%ebx),40) | |
1359 HDC_CHECK_AND_CPY((%%ebx, %2),48) | |
1360 HDC_CHECK_AND_CPY((%%ebx, %2, 2),56) | |
1361 "jmp 2f \n\t" | |
1362 "1: \n\t" | |
1363 // src does not cross a 32 byte cache line so dont waste time with alignment | |
1364 #define HDC_CHECK_AND_CPY2(src, dst) \ | |
1365 "movq " #src ", %%mm2 \n\t"\ | |
1366 "movq " #src ", %%mm1 \n\t"\ | |
1367 "psrlq $8, %%mm2 \n\t"\ | |
1368 "psubb %%mm1, %%mm2 \n\t"\ | |
1369 "paddb %%mm7, %%mm2 \n\t"\ | |
1370 "pcmpgtb %%mm6, %%mm2 \n\t"\ | |
1371 "paddb %%mm2, %%mm0 \n\t"\ | |
1372 "movq %%mm1," #dst "(%%eax) \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1373 |
2416 | 1374 HDC_CHECK_AND_CPY2((%1),0) |
1375 HDC_CHECK_AND_CPY2((%%ecx),8) | |
1376 HDC_CHECK_AND_CPY2((%%ecx, %2),16) | |
1377 HDC_CHECK_AND_CPY2((%%ecx, %2, 2),24) | |
1378 HDC_CHECK_AND_CPY2((%1, %2, 4),32) | |
1379 HDC_CHECK_AND_CPY2((%%ebx),40) | |
1380 HDC_CHECK_AND_CPY2((%%ebx, %2),48) | |
1381 HDC_CHECK_AND_CPY2((%%ebx, %2, 2),56) | |
1382 "2: \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1383 "psllq $8, %%mm0 \n\t" // remove dummy value |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1384 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1385 "psrlw $8, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1386 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1387 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1388 "psrlq $16, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1389 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1390 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1391 "psrlq $32, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1392 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1393 "movd %%mm0, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1394 : "=r" (numEq) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1395 : "r" (src), "r" (stride) |
2416 | 1396 : "%eax", "%ebx", "%ecx" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1397 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1398 // printf("%d\n", numEq); |
2416 | 1399 numEq= (256 - numEq) &0xFF; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1400 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1401 int y; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1402 for(y=0; y<BLOCK_SIZE; y++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1403 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1404 if(((src[0] - src[1] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1405 if(((src[1] - src[2] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1406 if(((src[2] - src[3] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1407 if(((src[3] - src[4] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1408 if(((src[4] - src[5] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1409 if(((src[5] - src[6] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1410 if(((src[6] - src[7] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1411 src+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1412 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1413 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1414 /* if(abs(numEq - asmEq) > 0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1415 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1416 // printf("\nasm:%d c:%d\n", asmEq, numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1417 for(int y=0; y<8; y++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1418 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1419 for(int x=0; x<8; x++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1420 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1421 printf("%d ", src[x + y*stride]); |
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 printf("\n"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1424 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1425 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1426 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1427 // printf("%d\n", numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1428 return numEq > hFlatnessThreshold; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1429 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1430 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1431 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
|
1432 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1433 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
|
1434 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1435 return 1; |
2158
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 |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1438 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
|
1439 { |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1440 #if 0 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1441 asm volatile( |
2428 | 1442 "leal (%0, %1), %%ecx \n\t" |
1443 "leal (%%ecx, %1, 4), %%ebx \n\t" | |
1444 // 0 1 2 3 4 5 6 7 8 9 | |
1445 // %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
|
1446 "pxor %%mm7, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1447 "movq bm00001000, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1448 "movd %2, %%mm5 \n\t" // QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1449 "movq %%mm5, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1450 "paddusb %%mm5, %%mm5 \n\t" // 2QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1451 "paddusb %%mm5, %%mm4 \n\t" // 3QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1452 "psllq $24, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1453 "pxor %%mm5, %%mm5 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1454 "psubb %%mm4, %%mm5 \n\t" // -QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1455 "leal tempBlock, %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1456 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1457 //FIXME? "unroll by 2" and mix |
2159 | 1458 #ifdef HAVE_MMX2 |
2428 | 1459 #define HDF(src, dst) \ |
1460 "movq " #src "(%%eax), %%mm0 \n\t"\ | |
1461 "movq " #src "(%%eax), %%mm1 \n\t"\ | |
1462 "movq " #src "(%%eax), %%mm2 \n\t"\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1463 "psrlq $8, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1464 "psubusb %%mm1, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1465 "psubusb %%mm0, %%mm1 \n\t"\ |
2159 | 1466 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ |
1467 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
1468 "pshufw $0x00, %%mm1, %%mm3 \n\t" /* p´5 = |p1 - p2| */\ | |
1469 "pminub %%mm1, %%mm3 \n\t" /* p´5 = min(|p2-p1|, |p6-p5|)*/\ | |
1470 "psrlq $16, %%mm3 \n\t" /* p´3 = min(|p2-p1|, |p6-p5|)*/\ | |
1471 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-min(|p1-p2|,|p5-p6|) */\ | |
1472 "paddb %%mm5, %%mm1 \n\t"\ | |
1473 "psubusb %%mm5, %%mm1 \n\t"\ | |
1474 "psrlw $2, %%mm1 \n\t"\ | |
1475 "pxor %%mm2, %%mm1 \n\t"\ | |
1476 "psubb %%mm2, %%mm1 \n\t"\ | |
1477 "pand %%mm6, %%mm1 \n\t"\ | |
1478 "psubb %%mm1, %%mm0 \n\t"\ | |
1479 "psllq $8, %%mm1 \n\t"\ | |
1480 "paddb %%mm1, %%mm0 \n\t"\ | |
2428 | 1481 "movd %%mm0, " #dst" \n\t"\ |
2159 | 1482 "psrlq $32, %%mm0 \n\t"\ |
2428 | 1483 "movd %%mm0, 4" #dst" \n\t" |
2159 | 1484 #else |
2428 | 1485 #define HDF(src, dst)\ |
1486 "movq " #src "(%%eax), %%mm0 \n\t"\ | |
2159 | 1487 "movq %%mm0, %%mm1 \n\t"\ |
1488 "movq %%mm0, %%mm2 \n\t"\ | |
1489 "psrlq $8, %%mm1 \n\t"\ | |
1490 "psubusb %%mm1, %%mm2 \n\t"\ | |
1491 "psubusb %%mm0, %%mm1 \n\t"\ | |
1492 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ | |
1493 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
1494 "movq %%mm1, %%mm3 \n\t"\ | |
1495 "psllq $32, %%mm3 \n\t"\ | |
1496 "movq %%mm3, %%mm4 \n\t"\ | |
1497 "psubusb %%mm1, %%mm4 \n\t"\ | |
1498 "psubb %%mm4, %%mm3 \n\t"\ | |
1499 "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
|
1500 "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
|
1501 "paddb %%mm5, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1502 "psubusb %%mm5, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1503 "psrlw $2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1504 "pxor %%mm2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1505 "psubb %%mm2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1506 "pand %%mm6, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1507 "psubb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1508 "psllq $8, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1509 "paddb %%mm1, %%mm0 \n\t"\ |
2428 | 1510 "movd %%mm0, " #dst " \n\t"\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1511 "psrlq $32, %%mm0 \n\t"\ |
2428 | 1512 "movd %%mm0, 4" #dst " \n\t" |
2159 | 1513 #endif |
2428 | 1514 HDF(0,(%0)) |
1515 HDF(8,(%%ecx)) | |
1516 HDF(16,(%%ecx, %1)) | |
1517 HDF(24,(%%ecx, %1, 2)) | |
1518 HDF(32,(%0, %1, 4)) | |
1519 HDF(40,(%%ebx)) | |
1520 HDF(48,(%%ebx, %1)) | |
1521 HDF(56,(%%ebx, %1, 2)) | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1522 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1523 : "r" (dst), "r" (stride), "r" (QP) |
2428 | 1524 : "%eax", "%ebx", "%ecx" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1525 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1526 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1527 int y; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1528 for(y=0; y<BLOCK_SIZE; y++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1529 { |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1530 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
|
1531 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1532 if(ABS(middleEnergy) < 8*QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1533 { |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1534 const int q=(dst[3] - dst[4])/2; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1535 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
|
1536 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
|
1537 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1538 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1539 d= MAX(d, 0); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1540 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1541 d= (5*d + 32) >> 6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1542 d*= SIGN(-middleEnergy); |
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 if(q>0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1545 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1546 d= d<0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1547 d= d>q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1548 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1549 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1550 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1551 d= d>0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1552 d= d<q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1553 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1554 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1555 dst[3]-= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1556 dst[4]+= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1557 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1558 dst+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1559 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1560 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1561 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1562 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1563 /** |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1564 * Do a horizontal low pass filter on the 10x8 block (dst points to middle 8x8 Block) |
2221 | 1565 * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 (C version) |
1566 * 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
|
1567 */ |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1568 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
|
1569 { |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1570 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1571 #if 0 |
2428 | 1572 asm volatile( |
1573 "leal (%0, %1), %%ecx \n\t" | |
1574 "leal (%%ecx, %1, 4), %%ebx \n\t" | |
1575 // 0 1 2 3 4 5 6 7 8 9 | |
1576 // %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
|
1577 "pxor %%mm7, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1578 "leal tempBlock, %%eax \n\t" |
2159 | 1579 /* |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1580 #define HLP1 "movq (%0), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1581 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1582 "psllq $8, %%mm0 \n\t"\ |
2159 | 1583 PAVGB(%%mm1, %%mm0)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1584 "psrlw $8, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1585 "pxor %%mm1, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1586 "packuswb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1587 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1588 "movq %%mm0, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1589 "psllq $32, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1590 "paddb %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1591 "psllq $16, %%mm2 \n\t"\ |
2159 | 1592 PAVGB(%%mm2, %%mm0)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1593 "movq %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1594 "pand bm11001100, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1595 "paddusb %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1596 "psrlq $8, %%mm3 \n\t"\ |
2159 | 1597 PAVGB(%%mm1, %%mm4)\ |
1598 PAVGB(%%mm3, %%mm2)\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1599 "psrlq $16, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1600 "punpcklbw %%mm2, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1601 "movq %%mm2, (%0) \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1602 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1603 #define HLP2 "movq (%0), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1604 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1605 "psllq $8, %%mm0 \n\t"\ |
2159 | 1606 PAVGB(%%mm1, %%mm0)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1607 "psrlw $8, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1608 "pxor %%mm1, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1609 "packuswb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1610 "movq %%mm0, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1611 "psllq $32, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1612 "psllq $16, %%mm2 \n\t"\ |
2159 | 1613 PAVGB(%%mm2, %%mm0)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1614 "movq %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1615 "pand bm11001100, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1616 "paddusb %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1617 "psrlq $8, %%mm3 \n\t"\ |
2159 | 1618 PAVGB(%%mm3, %%mm2)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1619 "psrlq $16, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1620 "punpcklbw %%mm2, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1621 "movq %%mm2, (%0) \n\t"\ |
2159 | 1622 */ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1623 // 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
|
1624 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1625 Implemented Exact 7-Tap |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1626 9421 A321 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1627 36421 64321 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1628 334321 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1629 1234321 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1630 1234321 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1631 123433 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1632 12463 12346 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1633 1249 123A |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1634 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1635 */ |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1636 |
2159 | 1637 #ifdef HAVE_MMX2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1638 #define HLP3(i) "movq " #i "(%%eax), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1639 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1640 "movq %%mm0, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1641 "movq %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1642 "movq %%mm0, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1643 "psllq $8, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1644 "psrlq $8, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1645 "pand bm00000001, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1646 "pand bm10000000, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1647 "por %%mm3, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1648 "por %%mm4, %%mm2 \n\t"\ |
2159 | 1649 PAVGB(%%mm2, %%mm1)\ |
1650 PAVGB(%%mm1, %%mm0)\ | |
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 "pshufw $0xF9, %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1653 "pshufw $0x90, %%mm0, %%mm4 \n\t"\ |
2159 | 1654 PAVGB(%%mm3, %%mm4)\ |
1655 PAVGB(%%mm4, %%mm0)\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1656 "movd %%mm0, (%0) \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1657 "psrlq $32, %%mm0 \n\t"\ |
2159 | 1658 "movd %%mm0, 4(%0) \n\t" |
1659 #else | |
1660 #define HLP3(i) "movq " #i "(%%eax), %%mm0 \n\t"\ | |
1661 "movq %%mm0, %%mm1 \n\t"\ | |
1662 "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
|
1663 "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
|
1664 "movd 8(%0), %%mm4 \n\t" /*0001000*/\ |
2159 | 1665 "psllq $8, %%mm1 \n\t"\ |
1666 "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
|
1667 "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
|
1668 "psllq $56, %%mm4 \n\t"\ |
2159 | 1669 "por %%mm3, %%mm1 \n\t"\ |
1670 "por %%mm4, %%mm2 \n\t"\ | |
1671 PAVGB(%%mm2, %%mm1)\ | |
1672 PAVGB(%%mm1, %%mm0)\ | |
1673 \ | |
1674 "movq %%mm0, %%mm3 \n\t"\ | |
1675 "movq %%mm0, %%mm4 \n\t"\ | |
1676 "movq %%mm0, %%mm5 \n\t"\ | |
1677 "psrlq $16, %%mm3 \n\t"\ | |
1678 "psllq $16, %%mm4 \n\t"\ | |
1679 "pand bm11000000, %%mm5 \n\t"\ | |
1680 "por %%mm5, %%mm3 \n\t"\ | |
1681 "movq %%mm0, %%mm5 \n\t"\ | |
1682 "pand bm00000011, %%mm5 \n\t"\ | |
1683 "por %%mm5, %%mm4 \n\t"\ | |
1684 PAVGB(%%mm3, %%mm4)\ | |
1685 PAVGB(%%mm4, %%mm0)\ | |
1686 "movd %%mm0, (%0) \n\t"\ | |
1687 "psrlq $32, %%mm0 \n\t"\ | |
1688 "movd %%mm0, 4(%0) \n\t" | |
1689 #endif | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1690 |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1691 /* uses the 7-Tap Filter: 1112111 */ |
2428 | 1692 #define NEW_HLP(src, dst)\ |
1693 "movq " #src "(%%eax), %%mm1 \n\t"\ | |
1694 "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
|
1695 "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
|
1696 "psrlq $8, %%mm2 \n\t"\ |
2428 | 1697 "movd -4" #dst ", %%mm3 \n\t" /*0001000*/\ |
1698 "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
|
1699 "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
|
1700 "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
|
1701 "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
|
1702 "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
|
1703 "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
|
1704 PAVGB(%%mm2, %%mm1)\ |
2428 | 1705 "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
|
1706 PAVGB(%%mm1, %%mm0)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1707 "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
|
1708 "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
|
1709 "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
|
1710 "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
|
1711 "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
|
1712 PAVGB(%%mm2, %%mm5)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1713 "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
|
1714 "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
|
1715 "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
|
1716 "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
|
1717 PAVGB(%%mm2, %%mm1)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1718 PAVGB(%%mm1, %%mm5)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1719 PAVGB(%%mm5, %%mm0)\ |
2428 | 1720 "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
|
1721 "psrlq $32, %%mm0 \n\t"\ |
2428 | 1722 "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
|
1723 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1724 /* 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
|
1725 #define NEW_HLP2(i)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1726 "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
|
1727 "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
|
1728 "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
|
1729 "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
|
1730 "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
|
1731 "psllq $8, %%mm1 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1732 "psrlq $8, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1733 "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
|
1734 "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
|
1735 "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
|
1736 "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
|
1737 "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
|
1738 PAVGB(%%mm2, %%mm1) /*0010100*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1739 PAVGB(%%mm1, %%mm0) /*0012100*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1740 "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
|
1741 "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
|
1742 "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
|
1743 "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
|
1744 "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
|
1745 PAVGB(%%mm2, %%mm5) /*0100010*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1746 "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
|
1747 "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
|
1748 "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
|
1749 "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
|
1750 "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
|
1751 PAVGB(%%mm2, %%mm1) /*1000001*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1752 "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
|
1753 "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
|
1754 "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
|
1755 "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
|
1756 PAVGB(%%mm2, %%mm6) /*100000001*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1757 PAVGB(%%mm6, %%mm1) /*110000011*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1758 PAVGB(%%mm1, %%mm5) /*112000211*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1759 PAVGB(%%mm5, %%mm0) /*112242211*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1760 "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
|
1761 "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
|
1762 "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
|
1763 |
2428 | 1764 #define HLP(src, dst) NEW_HLP(src, dst) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1765 |
2428 | 1766 HLP(0, (%0)) |
1767 HLP(8, (%%ecx)) | |
1768 HLP(16, (%%ecx, %1)) | |
1769 HLP(24, (%%ecx, %1, 2)) | |
1770 HLP(32, (%0, %1, 4)) | |
1771 HLP(40, (%%ebx)) | |
1772 HLP(48, (%%ebx, %1)) | |
1773 HLP(56, (%%ebx, %1, 2)) | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1774 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1775 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1776 : "r" (dst), "r" (stride) |
2428 | 1777 : "%eax", "%ebx", "%ecx" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1778 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1779 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1780 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1781 int y; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1782 for(y=0; y<BLOCK_SIZE; y++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1783 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1784 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
|
1785 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
|
1786 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1787 int sums[9]; |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1788 sums[0] = first + dst[0]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1789 sums[1] = dst[0] + dst[1]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1790 sums[2] = dst[1] + dst[2]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1791 sums[3] = dst[2] + dst[3]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1792 sums[4] = dst[3] + dst[4]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1793 sums[5] = dst[4] + dst[5]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1794 sums[6] = dst[5] + dst[6]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1795 sums[7] = dst[6] + dst[7]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1796 sums[8] = dst[7] + last; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1797 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1798 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
|
1799 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
|
1800 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
|
1801 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
|
1802 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
|
1803 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
|
1804 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
|
1805 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
|
1806 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1807 dst+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1808 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1809 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1810 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1811 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1812 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
|
1813 { |
2475 | 1814 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1815 asm volatile( |
2473 | 1816 "movq pQPb, %%mm0 \n\t" |
1817 "paddusb %%mm0, %%mm0 \n\t" | |
1818 "movq %%mm0, pQPb2 \n\t" | |
1819 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1820 "leal (%0, %1), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1821 "leal (%%eax, %1, 4), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1822 // 0 1 2 3 4 5 6 7 8 9 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1823 // %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
|
1824 |
2473 | 1825 "pcmpeqb %%mm6, %%mm6 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1826 "pxor %%mm7, %%mm7 \n\t" |
2475 | 1827 #ifdef HAVE_MMX2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1828 #define FIND_MIN_MAX(addr)\ |
2473 | 1829 "movq " #addr ", %%mm0 \n\t"\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1830 "pminub %%mm0, %%mm6 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1831 "pmaxub %%mm0, %%mm7 \n\t" |
2475 | 1832 #else |
1833 #define FIND_MIN_MAX(addr)\ | |
1834 "movq " #addr ", %%mm0 \n\t"\ | |
1835 "movq %%mm6, %%mm1 \n\t"\ | |
1836 "psubusb %%mm0, %%mm7 \n\t"\ | |
1837 "paddb %%mm0, %%mm7 \n\t"\ | |
1838 "psubusb %%mm0, %%mm1 \n\t"\ | |
1839 "psubb %%mm1, %%mm6 \n\t" | |
1840 #endif | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1841 |
2473 | 1842 FIND_MIN_MAX((%%eax)) |
1843 FIND_MIN_MAX((%%eax, %1)) | |
1844 FIND_MIN_MAX((%%eax, %1, 2)) | |
1845 FIND_MIN_MAX((%0, %1, 4)) | |
1846 FIND_MIN_MAX((%%ebx)) | |
1847 FIND_MIN_MAX((%%ebx, %1)) | |
1848 FIND_MIN_MAX((%%ebx, %1, 2)) | |
1849 FIND_MIN_MAX((%0, %1, 8)) | |
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 "movq %%mm6, %%mm4 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1852 "psrlq $8, %%mm6 \n\t" |
2475 | 1853 #ifdef HAVE_MMX2 |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1854 "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
|
1855 "pshufw $0xF9, %%mm6, %%mm4 \n\t" |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1856 "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
|
1857 "pshufw $0xFE, %%mm6, %%mm4 \n\t" |
2475 | 1858 "pminub %%mm4, %%mm6 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1859 #else |
2475 | 1860 "movq %%mm6, %%mm1 \n\t" |
1861 "psubusb %%mm4, %%mm1 \n\t" | |
1862 "psubb %%mm1, %%mm6 \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1863 "movq %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1864 "psrlq $16, %%mm6 \n\t" |
2475 | 1865 "movq %%mm6, %%mm1 \n\t" |
1866 "psubusb %%mm4, %%mm1 \n\t" | |
1867 "psubb %%mm1, %%mm6 \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1868 "movq %%mm6, %%mm4 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1869 "psrlq $32, %%mm6 \n\t" |
2475 | 1870 "movq %%mm6, %%mm1 \n\t" |
1871 "psubusb %%mm4, %%mm1 \n\t" | |
1872 "psubb %%mm1, %%mm6 \n\t" | |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1873 #endif |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1874 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1875 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1876 "movq %%mm7, %%mm4 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1877 "psrlq $8, %%mm7 \n\t" |
2475 | 1878 #ifdef HAVE_MMX2 |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1879 "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
|
1880 "pshufw $0xF9, %%mm7, %%mm4 \n\t" |
2475 | 1881 "pmaxub %%mm4, %%mm7 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1882 "pshufw $0xFE, %%mm7, %%mm4 \n\t" |
2475 | 1883 "pmaxub %%mm4, %%mm7 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1884 #else |
2475 | 1885 "psubusb %%mm4, %%mm7 \n\t" |
1886 "paddb %%mm4, %%mm7 \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1887 "movq %%mm7, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1888 "psrlq $16, %%mm7 \n\t" |
2475 | 1889 "psubusb %%mm4, %%mm7 \n\t" |
1890 "paddb %%mm4, %%mm7 \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1891 "movq %%mm7, %%mm4 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1892 "psrlq $32, %%mm7 \n\t" |
2475 | 1893 "psubusb %%mm4, %%mm7 \n\t" |
1894 "paddb %%mm4, %%mm7 \n\t" | |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1895 #endif |
2473 | 1896 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
|
1897 "punpcklbw %%mm7, %%mm7 \n\t" |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1898 "punpcklbw %%mm7, %%mm7 \n\t" |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1899 "punpcklbw %%mm7, %%mm7 \n\t" |
2473 | 1900 "movq %%mm7, temp0 \n\t" |
1901 | |
1902 "movq (%0), %%mm0 \n\t" // L10 | |
1903 "movq %%mm0, %%mm1 \n\t" // L10 | |
1904 "movq %%mm0, %%mm2 \n\t" // L10 | |
1905 "psllq $8, %%mm1 \n\t" | |
1906 "psrlq $8, %%mm2 \n\t" | |
1907 "movd -4(%0), %%mm3 \n\t" | |
1908 "movd 8(%0), %%mm4 \n\t" | |
1909 "psrlq $24, %%mm3 \n\t" | |
1910 "psllq $56, %%mm4 \n\t" | |
1911 "por %%mm3, %%mm1 \n\t" // L00 | |
1912 "por %%mm4, %%mm2 \n\t" // L20 | |
1913 "movq %%mm1, %%mm3 \n\t" // L00 | |
1914 PAVGB(%%mm2, %%mm1) // (L20 + L00)/2 | |
1915 PAVGB(%%mm0, %%mm1) // (L20 + L00 + 2L10)/4 | |
1916 "psubusb %%mm7, %%mm0 \n\t" | |
1917 "psubusb %%mm7, %%mm2 \n\t" | |
1918 "psubusb %%mm7, %%mm3 \n\t" | |
1919 "pcmpeqb b00, %%mm0 \n\t" // L10 > a ? 0 : -1 | |
1920 "pcmpeqb b00, %%mm2 \n\t" // L20 > a ? 0 : -1 | |
1921 "pcmpeqb b00, %%mm3 \n\t" // L00 > a ? 0 : -1 | |
1922 "paddb %%mm2, %%mm0 \n\t" | |
1923 "paddb %%mm3, %%mm0 \n\t" | |
1924 | |
1925 "movq (%%eax), %%mm2 \n\t" // L11 | |
1926 "movq %%mm2, %%mm3 \n\t" // L11 | |
1927 "movq %%mm2, %%mm4 \n\t" // L11 | |
1928 "psllq $8, %%mm3 \n\t" | |
1929 "psrlq $8, %%mm4 \n\t" | |
1930 "movd -4(%%eax), %%mm5 \n\t" | |
1931 "movd 8(%%eax), %%mm6 \n\t" | |
1932 "psrlq $24, %%mm5 \n\t" | |
1933 "psllq $56, %%mm6 \n\t" | |
1934 "por %%mm5, %%mm3 \n\t" // L01 | |
1935 "por %%mm6, %%mm4 \n\t" // L21 | |
1936 "movq %%mm3, %%mm5 \n\t" // L01 | |
1937 PAVGB(%%mm4, %%mm3) // (L21 + L01)/2 | |
1938 PAVGB(%%mm2, %%mm3) // (L21 + L01 + 2L11)/4 | |
1939 "psubusb %%mm7, %%mm2 \n\t" | |
1940 "psubusb %%mm7, %%mm4 \n\t" | |
1941 "psubusb %%mm7, %%mm5 \n\t" | |
1942 "pcmpeqb b00, %%mm2 \n\t" // L11 > a ? 0 : -1 | |
1943 "pcmpeqb b00, %%mm4 \n\t" // L21 > a ? 0 : -1 | |
1944 "pcmpeqb b00, %%mm5 \n\t" // L01 > a ? 0 : -1 | |
1945 "paddb %%mm4, %%mm2 \n\t" | |
1946 "paddb %%mm5, %%mm2 \n\t" | |
1947 // 0, 2, 3, 1 | |
1948 #define DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \ | |
1949 "movq " #src ", " #sx " \n\t" /* src[0] */\ | |
1950 "movq " #sx ", " #lx " \n\t" /* src[0] */\ | |
1951 "movq " #sx ", " #t0 " \n\t" /* src[0] */\ | |
1952 "psllq $8, " #lx " \n\t"\ | |
1953 "psrlq $8, " #t0 " \n\t"\ | |
1954 "movd -4" #src ", " #t1 " \n\t"\ | |
1955 "psrlq $24, " #t1 " \n\t"\ | |
1956 "por " #t1 ", " #lx " \n\t" /* src[-1] */\ | |
1957 "movd 8" #src ", " #t1 " \n\t"\ | |
1958 "psllq $56, " #t1 " \n\t"\ | |
1959 "por " #t1 ", " #t0 " \n\t" /* src[+1] */\ | |
1960 "movq " #lx ", " #t1 " \n\t" /* src[-1] */\ | |
1961 PAVGB(t0, lx) /* (src[-1] + src[+1])/2 */\ | |
1962 PAVGB(sx, lx) /* (src[-1] + 2src[0] + src[+1])/4 */\ | |
1963 "psubusb temp0, " #t1 " \n\t"\ | |
1964 "psubusb temp0, " #t0 " \n\t"\ | |
1965 "psubusb temp0, " #sx " \n\t"\ | |
1966 "pcmpeqb b00, " #t1 " \n\t" /* src[-1] > a ? 0 : -1*/\ | |
1967 "pcmpeqb b00, " #t0 " \n\t" /* src[+1] > a ? 0 : -1*/\ | |
1968 "pcmpeqb b00, " #sx " \n\t" /* src[0] > a ? 0 : -1*/\ | |
1969 "paddb " #t1 ", " #t0 " \n\t"\ | |
1970 "paddb " #t0 ", " #sx " \n\t"\ | |
1971 \ | |
1972 PAVGB(lx, pplx) \ | |
1973 PAVGB(plx, pplx) /* filtered */\ | |
1974 "movq " #dst ", " #t0 " \n\t" /* dst */\ | |
1975 "movq " #pplx ", " #t1 " \n\t"\ | |
1976 "psubusb " #t0 ", " #pplx " \n\t"\ | |
1977 "psubusb " #t1 ", " #t0 " \n\t"\ | |
1978 "por " #t0 ", " #pplx " \n\t" /* |filtered - dst| */\ | |
1979 "psubusb pQPb2, " #pplx " \n\t"\ | |
1980 "pcmpeqb b00, " #pplx " \n\t"\ | |
1981 "paddb " #sx ", " #ppsx " \n\t"\ | |
1982 "paddb " #psx ", " #ppsx " \n\t"\ | |
1983 "#paddb b02, " #ppsx " \n\t"\ | |
1984 "pand b08, " #ppsx " \n\t"\ | |
1985 "pcmpeqb b00, " #ppsx " \n\t"\ | |
1986 "pand " #pplx ", " #ppsx " \n\t"\ | |
1987 "pand " #ppsx ", " #t1 " \n\t"\ | |
1988 "pandn " #dst ", " #ppsx " \n\t"\ | |
1989 "por " #t1 ", " #ppsx " \n\t"\ | |
1990 "movq " #ppsx ", " #dst " \n\t" | |
1991 /* | |
1992 0000000 | |
1993 1111111 | |
1994 | |
1995 1111110 | |
1996 1111101 | |
1997 1111100 | |
1998 1111011 | |
1999 1111010 | |
2000 1111001 | |
2001 | |
2002 1111000 | |
2003 1110111 | |
2004 | |
2005 */ | |
2006 //DERING_CORE(dst,src ,ppsx ,psx ,sx ,pplx ,plx ,lx ,t0 ,t1) | |
2007 DERING_CORE((%%eax),(%%eax, %1) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7) | |
2008 DERING_CORE((%%eax, %1),(%%eax, %1, 2) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7) | |
2009 DERING_CORE((%%eax, %1, 2),(%0, %1, 4) ,%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7) | |
2010 DERING_CORE((%0, %1, 4),(%%ebx) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7) | |
2011 DERING_CORE((%%ebx),(%%ebx, %1) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7) | |
2012 DERING_CORE((%%ebx, %1), (%%ebx, %1, 2),%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7) | |
2013 DERING_CORE((%%ebx, %1, 2),(%0, %1, 8) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7) | |
2014 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
|
2015 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2016 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2017 : : "r" (src), "r" (stride), "r" (QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2018 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2019 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2020 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2021 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2022 //FIXME |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2023 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2024 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2025 |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2026 /** |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2027 * Deinterlaces the given block |
2246 | 2028 * will be called for every 8x8 block, and can read & write into an 8x16 block |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2029 */ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2030 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
|
2031 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2032 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2033 asm volatile( |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2034 "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
|
2035 "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
|
2036 // 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
|
2037 // %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
|
2038 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2039 "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
|
2040 "movq (%%eax, %1), %%mm1 \n\t" |
2246 | 2041 PAVGB(%%mm1, %%mm0) |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2042 "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
|
2043 "movq (%0, %1, 4), %%mm0 \n\t" |
2246 | 2044 PAVGB(%%mm0, %%mm1) |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2045 "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
|
2046 "movq (%%ebx, %1), %%mm1 \n\t" |
2246 | 2047 PAVGB(%%mm1, %%mm0) |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2048 "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
|
2049 "movq (%0, %1, 8), %%mm0 \n\t" |
2246 | 2050 PAVGB(%%mm0, %%mm1) |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2051 "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
|
2052 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2053 : : "r" (src), "r" (stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2054 : "%eax", "%ebx" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2055 ); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2056 #else |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2057 int x; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2058 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
|
2059 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2060 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
|
2061 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
|
2062 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
|
2063 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
|
2064 src++; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2065 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2066 #endif |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2067 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2068 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2069 /** |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2070 * Deinterlaces the given block |
2246 | 2071 * will be called for every 8x8 block, and can read & write into an 8x16 block |
2072 * 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
|
2073 */ |
2246 | 2074 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
|
2075 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2076 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2077 asm volatile( |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2078 "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
|
2079 "leal (%%eax, %1, 4), %%ebx \n\t" |
2246 | 2080 "leal (%%ebx, %1, 4), %%ecx \n\t" |
2081 "addl %1, %%ecx \n\t" | |
2082 "pxor %%mm7, %%mm7 \n\t" | |
2083 // 0 1 2 3 4 5 6 7 8 9 10 | |
2084 // %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
|
2085 |
2246 | 2086 #define DEINT_CUBIC(a,b,c,d,e)\ |
2087 "movq " #a ", %%mm0 \n\t"\ | |
2088 "movq " #b ", %%mm1 \n\t"\ | |
2089 "movq " #d ", %%mm2 \n\t"\ | |
2090 "movq " #e ", %%mm3 \n\t"\ | |
2091 PAVGB(%%mm2, %%mm1) /* (b+d) /2 */\ | |
2092 PAVGB(%%mm3, %%mm0) /* a(a+e) /2 */\ | |
2093 "movq %%mm0, %%mm2 \n\t"\ | |
2094 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
2095 "punpckhbw %%mm7, %%mm2 \n\t"\ | |
2096 "movq %%mm1, %%mm3 \n\t"\ | |
2097 "punpcklbw %%mm7, %%mm1 \n\t"\ | |
2098 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
2099 "psubw %%mm1, %%mm0 \n\t" /* L(a+e - (b+d))/2 */\ | |
2100 "psubw %%mm3, %%mm2 \n\t" /* H(a+e - (b+d))/2 */\ | |
2101 "psraw $3, %%mm0 \n\t" /* L(a+e - (b+d))/16 */\ | |
2102 "psraw $3, %%mm2 \n\t" /* H(a+e - (b+d))/16 */\ | |
2103 "psubw %%mm0, %%mm1 \n\t" /* L(9b + 9d - a - e)/16 */\ | |
2104 "psubw %%mm2, %%mm3 \n\t" /* H(9b + 9d - a - e)/16 */\ | |
2105 "packuswb %%mm3, %%mm1 \n\t"\ | |
2106 "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
|
2107 |
2246 | 2108 DEINT_CUBIC((%0), (%%eax, %1), (%%eax, %1, 2), (%0, %1, 4), (%%ebx, %1)) |
2109 DEINT_CUBIC((%%eax, %1), (%0, %1, 4), (%%ebx), (%%ebx, %1), (%0, %1, 8)) | |
2110 DEINT_CUBIC((%0, %1, 4), (%%ebx, %1), (%%ebx, %1, 2), (%0, %1, 8), (%%ecx)) | |
2111 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
|
2112 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2113 : : "r" (src), "r" (stride) |
2246 | 2114 : "%eax", "%ebx", "ecx" |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2115 ); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2116 #else |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2117 int x; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2118 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
|
2119 { |
2246 | 2120 src[stride*3] = (-src[0] + 9*src[stride*2] + 9*src[stride*4] - src[stride*6])>>4; |
2121 src[stride*5] = (-src[stride*2] + 9*src[stride*4] + 9*src[stride*6] - src[stride*8])>>4; | |
2122 src[stride*7] = (-src[stride*4] + 9*src[stride*6] + 9*src[stride*8] - src[stride*10])>>4; | |
2123 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
|
2124 src++; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2125 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2126 #endif |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2127 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2128 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2129 /** |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2130 * Deinterlaces the given block |
2246 | 2131 * will be called for every 8x8 block, and can read & write into an 8x16 block |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2132 * will shift the image up by 1 line (FIXME if this is a problem) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2133 */ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2134 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
|
2135 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2136 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2137 asm volatile( |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2138 "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
|
2139 "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
|
2140 // 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
|
2141 // %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
|
2142 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2143 "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
|
2144 "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
|
2145 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
|
2146 "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
|
2147 PAVGB(%%mm2, %%mm0) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2148 "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
|
2149 "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
|
2150 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
|
2151 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
|
2152 "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
|
2153 "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
|
2154 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
|
2155 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
|
2156 "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
|
2157 "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
|
2158 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
|
2159 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
|
2160 "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
|
2161 "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
|
2162 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
|
2163 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
|
2164 "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
|
2165 "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
|
2166 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
|
2167 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
|
2168 "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
|
2169 "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
|
2170 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
|
2171 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
|
2172 "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
|
2173 "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
|
2174 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
|
2175 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
|
2176 "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
|
2177 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2178 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2179 : : "r" (src), "r" (stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2180 : "%eax", "%ebx" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2181 ); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2182 #else |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2183 int x; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2184 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
|
2185 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2186 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
|
2187 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
|
2188 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
|
2189 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
|
2190 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
|
2191 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
|
2192 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
|
2193 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
|
2194 src++; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2195 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2196 #endif |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2197 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2198 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2199 /** |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2200 * Deinterlaces the given block |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2201 * will be called for every 8x8 block, except the last row, and can read & write into an 8x16 block |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2202 */ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2203 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
|
2204 { |
2221 | 2205 #ifdef HAVE_MMX |
2206 #ifdef HAVE_MMX2 | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2207 asm volatile( |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2208 "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
|
2209 "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
|
2210 // 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
|
2211 // %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
|
2212 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2213 "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
|
2214 "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
|
2215 "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
|
2216 "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
|
2217 "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
|
2218 "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
|
2219 "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
|
2220 "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
|
2221 "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
|
2222 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2223 "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
|
2224 "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
|
2225 "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
|
2226 "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
|
2227 "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
|
2228 "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
|
2229 "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
|
2230 "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
|
2231 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2232 "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
|
2233 "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
|
2234 "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
|
2235 "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
|
2236 "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
|
2237 "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
|
2238 "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
|
2239 "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
|
2240 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2241 "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
|
2242 "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
|
2243 "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
|
2244 "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
|
2245 "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
|
2246 "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
|
2247 "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
|
2248 "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
|
2249 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2250 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2251 : : "r" (src), "r" (stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2252 : "%eax", "%ebx" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2253 ); |
2221 | 2254 |
2255 #else // MMX without MMX2 | |
2256 asm volatile( | |
2257 "leal (%0, %1), %%eax \n\t" | |
2258 "leal (%%eax, %1, 4), %%ebx \n\t" | |
2259 // 0 1 2 3 4 5 6 7 8 9 | |
2260 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 | |
2261 "pxor %%mm7, %%mm7 \n\t" | |
2262 | |
2263 #define MEDIAN(a,b,c)\ | |
2264 "movq " #a ", %%mm0 \n\t"\ | |
2265 "movq " #b ", %%mm2 \n\t"\ | |
2266 "movq " #c ", %%mm1 \n\t"\ | |
2267 "movq %%mm0, %%mm3 \n\t"\ | |
2268 "movq %%mm1, %%mm4 \n\t"\ | |
2269 "movq %%mm2, %%mm5 \n\t"\ | |
2270 "psubusb %%mm1, %%mm3 \n\t"\ | |
2271 "psubusb %%mm2, %%mm4 \n\t"\ | |
2272 "psubusb %%mm0, %%mm5 \n\t"\ | |
2273 "pcmpeqb %%mm7, %%mm3 \n\t"\ | |
2274 "pcmpeqb %%mm7, %%mm4 \n\t"\ | |
2275 "pcmpeqb %%mm7, %%mm5 \n\t"\ | |
2276 "movq %%mm3, %%mm6 \n\t"\ | |
2277 "pxor %%mm4, %%mm3 \n\t"\ | |
2278 "pxor %%mm5, %%mm4 \n\t"\ | |
2279 "pxor %%mm6, %%mm5 \n\t"\ | |
2280 "por %%mm3, %%mm1 \n\t"\ | |
2281 "por %%mm4, %%mm2 \n\t"\ | |
2282 "por %%mm5, %%mm0 \n\t"\ | |
2283 "pand %%mm2, %%mm0 \n\t"\ | |
2284 "pand %%mm1, %%mm0 \n\t"\ | |
2285 "movq %%mm0, " #b " \n\t" | |
2286 | |
2287 MEDIAN((%0), (%%eax), (%%eax, %1)) | |
2288 MEDIAN((%%eax, %1), (%%eax, %1, 2), (%0, %1, 4)) | |
2289 MEDIAN((%0, %1, 4), (%%ebx), (%%ebx, %1)) | |
2290 MEDIAN((%%ebx, %1), (%%ebx, %1, 2), (%0, %1, 8)) | |
2291 | |
2292 : : "r" (src), "r" (stride) | |
2293 : "%eax", "%ebx" | |
2294 ); | |
2295 #endif // MMX | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2296 #else |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2297 //FIXME |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2298 int x; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2299 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
|
2300 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2301 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
|
2302 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
|
2303 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
|
2304 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
|
2305 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
|
2306 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
|
2307 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
|
2308 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
|
2309 src++; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2310 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2311 #endif |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2312 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2313 |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2314 #ifdef HAVE_MMX |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2315 /** |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2316 * 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
|
2317 */ |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2318 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
|
2319 { |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2320 asm( |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2321 "leal (%0, %1), %%eax \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2322 "leal (%%eax, %1, 4), %%ebx \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2323 // 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
|
2324 // %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
|
2325 "movq (%0), %%mm0 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2326 "movq (%%eax), %%mm1 \n\t" // abcdefgh |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2327 "movq %%mm0, %%mm2 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2328 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2329 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2330 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2331 "movq (%%eax, %1), %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2332 "movq (%%eax, %1, 2), %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2333 "movq %%mm1, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2334 "punpcklbw %%mm3, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2335 "punpckhbw %%mm3, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2336 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2337 "movq %%mm0, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2338 "punpcklwd %%mm1, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2339 "punpckhwd %%mm1, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2340 "movq %%mm2, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2341 "punpcklwd %%mm4, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2342 "punpckhwd %%mm4, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2343 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2344 "movd %%mm0, 128(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2345 "psrlq $32, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2346 "movd %%mm0, 144(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2347 "movd %%mm3, 160(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2348 "psrlq $32, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2349 "movd %%mm3, 176(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2350 "movd %%mm3, 48(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2351 "movd %%mm2, 192(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2352 "movd %%mm2, 64(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2353 "psrlq $32, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2354 "movd %%mm2, 80(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2355 "movd %%mm1, 96(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2356 "psrlq $32, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2357 "movd %%mm1, 112(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2358 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2359 "movq (%0, %1, 4), %%mm0 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2360 "movq (%%ebx), %%mm1 \n\t" // abcdefgh |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2361 "movq %%mm0, %%mm2 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2362 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2363 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2364 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2365 "movq (%%ebx, %1), %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2366 "movq (%%ebx, %1, 2), %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2367 "movq %%mm1, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2368 "punpcklbw %%mm3, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2369 "punpckhbw %%mm3, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2370 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2371 "movq %%mm0, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2372 "punpcklwd %%mm1, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2373 "punpckhwd %%mm1, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2374 "movq %%mm2, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2375 "punpcklwd %%mm4, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2376 "punpckhwd %%mm4, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2377 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2378 "movd %%mm0, 132(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2379 "psrlq $32, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2380 "movd %%mm0, 148(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2381 "movd %%mm3, 164(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2382 "psrlq $32, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2383 "movd %%mm3, 180(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2384 "movd %%mm3, 52(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2385 "movd %%mm2, 196(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2386 "movd %%mm2, 68(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2387 "psrlq $32, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2388 "movd %%mm2, 84(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2389 "movd %%mm1, 100(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2390 "psrlq $32, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2391 "movd %%mm1, 116(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2392 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2393 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2394 :: "r" (src), "r" (srcStride), "r" (dst1), "r" (dst2) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2395 : "%eax", "%ebx" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2396 ); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2397 } |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2398 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2399 /** |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2400 * transposes the given 8x8 block |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2401 */ |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2402 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
|
2403 { |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2404 asm( |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2405 "leal (%0, %1), %%eax \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2406 "leal (%%eax, %1, 4), %%ebx \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2407 // 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
|
2408 // %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
|
2409 "movq (%2), %%mm0 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2410 "movq 16(%2), %%mm1 \n\t" // abcdefgh |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2411 "movq %%mm0, %%mm2 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2412 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2413 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2414 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2415 "movq 32(%2), %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2416 "movq 48(%2), %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2417 "movq %%mm1, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2418 "punpcklbw %%mm3, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2419 "punpckhbw %%mm3, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2420 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2421 "movq %%mm0, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2422 "punpcklwd %%mm1, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2423 "punpckhwd %%mm1, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2424 "movq %%mm2, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2425 "punpcklwd %%mm4, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2426 "punpckhwd %%mm4, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2427 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2428 "movd %%mm0, (%0) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2429 "psrlq $32, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2430 "movd %%mm0, (%%eax) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2431 "movd %%mm3, (%%eax, %1) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2432 "psrlq $32, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2433 "movd %%mm3, (%%eax, %1, 2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2434 "movd %%mm2, (%0, %1, 4) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2435 "psrlq $32, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2436 "movd %%mm2, (%%ebx) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2437 "movd %%mm1, (%%ebx, %1) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2438 "psrlq $32, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2439 "movd %%mm1, (%%ebx, %1, 2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2440 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2441 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2442 "movq 64(%2), %%mm0 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2443 "movq 80(%2), %%mm1 \n\t" // abcdefgh |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2444 "movq %%mm0, %%mm2 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2445 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2446 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2447 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2448 "movq 96(%2), %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2449 "movq 112(%2), %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2450 "movq %%mm1, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2451 "punpcklbw %%mm3, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2452 "punpckhbw %%mm3, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2453 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2454 "movq %%mm0, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2455 "punpcklwd %%mm1, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2456 "punpckhwd %%mm1, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2457 "movq %%mm2, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2458 "punpcklwd %%mm4, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2459 "punpckhwd %%mm4, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2460 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2461 "movd %%mm0, 4(%0) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2462 "psrlq $32, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2463 "movd %%mm0, 4(%%eax) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2464 "movd %%mm3, 4(%%eax, %1) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2465 "psrlq $32, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2466 "movd %%mm3, 4(%%eax, %1, 2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2467 "movd %%mm2, 4(%0, %1, 4) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2468 "psrlq $32, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2469 "movd %%mm2, 4(%%ebx) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2470 "movd %%mm1, 4(%%ebx, %1) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2471 "psrlq $32, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2472 "movd %%mm1, 4(%%ebx, %1, 2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2473 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2474 :: "r" (dst), "r" (dstStride), "r" (src) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2475 : "%eax", "%ebx" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2476 ); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2477 } |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2478 #endif |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2479 |
2185 | 2480 #ifdef HAVE_ODIVX_POSTPROCESS |
2481 #include "../opendivx/postprocess.h" | |
2482 int use_old_pp=0; | |
2483 #endif | |
2159 | 2484 |
2185 | 2485 static void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height, |
2486 QP_STORE_T QPs[], int QPStride, int isColor, int mode); | |
2159 | 2487 |
2389 | 2488 /* -pp Command line Help |
2489 NOTE/FIXME: put this at an appropriate place (--help, html docs, man mplayer)? | |
2490 | |
2491 -pp <filterName>[:<option>[:<option>...]][,[-]<filterName>[:<option>...]]... | |
2492 | |
2493 long form example: | |
2494 -pp vdeblock:autoq,hdeblock:autoq,linblenddeint -pp default,-vdeblock | |
2495 short form example: | |
2496 -pp vb:a,hb:a,lb -pp de,-vb | |
2497 | |
2498 Filters Options | |
2499 short long name short long option Description | |
2500 * * a autoq cpu power dependant enabler | |
2501 c chrom chrominance filtring enabled | |
2502 y nochrom chrominance filtring disabled | |
2503 hb hdeblock horizontal deblocking filter | |
2504 vb vdeblock vertical deblocking filter | |
2505 vr rkvdeblock | |
2506 h1 x1hdeblock Experimental horizontal deblock filter 1 | |
2507 v1 x1vdeblock Experimental vertical deblock filter 1 | |
2508 dr dering not implemented yet | |
2509 al autolevels automatic brightness / contrast fixer | |
2510 f fullyrange stretch luminance range to (0..255) | |
2511 lb linblenddeint linear blend deinterlacer | |
2512 li linipoldeint linear interpolating deinterlacer | |
2513 ci cubicipoldeint cubic interpolating deinterlacer | |
2514 md mediandeint median deinterlacer | |
2515 de default hdeblock:a,vdeblock:a,dering:a,autolevels | |
2516 fa fast x1hdeblock:a,x1vdeblock:a,dering:a,autolevels | |
2517 */ | |
2518 | |
2519 /** | |
2520 * returns a PPMode struct which will have a non 0 error variable if an error occured | |
2521 * name is the string after "-pp" on the command line | |
2522 * quality is a number from 0 to GET_PP_QUALITY_MAX | |
2523 */ | |
2524 struct PPMode getPPModeByNameAndQuality(char *name, int quality) | |
2525 { | |
2526 char temp[GET_MODE_BUFFER_SIZE]; | |
2527 char *p= temp; | |
2528 char *filterDelimiters= ","; | |
2529 char *optionDelimiters= ":"; | |
2530 struct PPMode ppMode= {0,0,0,0,0,0}; | |
2531 char *filterToken; | |
2532 | |
2533 strncpy(temp, name, GET_MODE_BUFFER_SIZE); | |
2534 | |
2535 for(;;){ | |
2536 char *filterName; | |
2537 int q= GET_PP_QUALITY_MAX; | |
2538 int chrom=-1; | |
2539 char *option; | |
2540 char *options[OPTIONS_ARRAY_SIZE]; | |
2541 int i; | |
2542 int filterNameOk=0; | |
2543 int numOfUnknownOptions=0; | |
2544 int enable=1; //does the user want us to enabled or disabled the filter | |
2545 | |
2546 filterToken= strtok(p, filterDelimiters); | |
2547 if(filterToken == NULL) break; | |
2548 p+= strlen(filterToken) + 1; | |
2549 filterName= strtok(filterToken, optionDelimiters); | |
2550 printf("%s::%s\n", filterToken, filterName); | |
2551 | |
2552 if(*filterName == '-') | |
2553 { | |
2554 enable=0; | |
2555 filterName++; | |
2556 } | |
2557 for(;;){ //for all options | |
2558 option= strtok(NULL, optionDelimiters); | |
2559 if(option == NULL) break; | |
2560 | |
2561 printf("%s\n", option); | |
2562 if(!strcmp("autoq", option) || !strcmp("a", option)) q= quality; | |
2563 else if(!strcmp("nochrom", option) || !strcmp("y", option)) chrom=0; | |
2564 else if(!strcmp("chrom", option) || !strcmp("c", option)) chrom=1; | |
2565 else | |
2566 { | |
2567 options[numOfUnknownOptions] = option; | |
2568 numOfUnknownOptions++; | |
2569 options[numOfUnknownOptions] = NULL; | |
2570 } | |
2571 if(numOfUnknownOptions >= OPTIONS_ARRAY_SIZE-1) break; | |
2572 } | |
2573 | |
2574 /* replace stuff from the replace Table */ | |
2575 for(i=0; replaceTable[2*i]!=NULL; i++) | |
2576 { | |
2577 if(!strcmp(replaceTable[2*i], filterName)) | |
2578 { | |
2579 int newlen= strlen(replaceTable[2*i + 1]); | |
2580 int plen; | |
2581 int spaceLeft; | |
2582 | |
2583 if(p==NULL) p= temp, *p=0; //last filter | |
2584 else p--, *p=','; //not last filter | |
2585 | |
2586 plen= strlen(p); | |
2587 spaceLeft= (int)p - (int)temp + plen; | |
2588 if(spaceLeft + newlen >= GET_MODE_BUFFER_SIZE) | |
2589 { | |
2590 ppMode.error++; | |
2591 break; | |
2592 } | |
2593 memmove(p + newlen, p, plen+1); | |
2594 memcpy(p, replaceTable[2*i + 1], newlen); | |
2595 filterNameOk=1; | |
2596 } | |
2597 } | |
2598 | |
2599 for(i=0; filters[i].shortName!=NULL; i++) | |
2600 { | |
2601 if( !strcmp(filters[i].longName, filterName) | |
2602 || !strcmp(filters[i].shortName, filterName)) | |
2603 { | |
2604 ppMode.lumMode &= ~filters[i].mask; | |
2605 ppMode.chromMode &= ~filters[i].mask; | |
2606 | |
2607 filterNameOk=1; | |
2608 if(!enable) break; // user wants to disable it | |
2609 | |
2610 if(q >= filters[i].minLumQuality) | |
2611 ppMode.lumMode|= filters[i].mask; | |
2612 if(chrom==1 || (chrom==-1 && filters[i].chromDefault)) | |
2613 if(q >= filters[i].minChromQuality) | |
2614 ppMode.chromMode|= filters[i].mask; | |
2615 | |
2616 if(filters[i].mask == LEVEL_FIX) | |
2617 { | |
2618 int o; | |
2619 ppMode.minAllowedY= 16; | |
2620 ppMode.maxAllowedY= 234; | |
2621 for(o=0; options[o]!=NULL; o++) | |
2622 if( !strcmp(options[o],"fullyrange") | |
2623 ||!strcmp(options[o],"f")) | |
2624 { | |
2625 ppMode.minAllowedY= 0; | |
2626 ppMode.maxAllowedY= 255; | |
2627 numOfUnknownOptions--; | |
2628 } | |
2629 } | |
2630 } | |
2631 } | |
2632 if(!filterNameOk) ppMode.error++; | |
2633 ppMode.error += numOfUnknownOptions; | |
2634 } | |
2635 | |
2636 if(ppMode.lumMode & H_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_Y_H; | |
2637 if(ppMode.lumMode & V_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_Y_V; | |
2638 if(ppMode.chromMode & H_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_C_H; | |
2639 if(ppMode.chromMode & V_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_C_V; | |
2640 if(ppMode.lumMode & DERING) ppMode.oldMode |= PP_DERING_Y; | |
2641 if(ppMode.chromMode & DERING) ppMode.oldMode |= PP_DERING_C; | |
2642 | |
2643 return ppMode; | |
2644 } | |
2645 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2646 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2647 * ... |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2648 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2649 void postprocess(unsigned char * src[], int src_stride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2650 unsigned char * dst[], int dst_stride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2651 int horizontal_size, int vertical_size, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2652 QP_STORE_T *QP_store, int QP_stride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2653 int mode) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2654 { |
2389 | 2655 /* |
2656 static int qual=0; | |
2657 | |
2658 struct PPMode ppMode= getPPModeByNameAndQuality("fast,default,-hdeblock,-vdeblock", qual); | |
2659 qual++; | |
2660 qual%=7; | |
2661 printf("\n%d %d %d %d\n", ppMode.lumMode, ppMode.chromMode, ppMode.oldMode, ppMode.error); | |
2662 postprocess2(src, src_stride, dst, dst_stride, | |
2663 horizontal_size, vertical_size, QP_store, QP_stride, &ppMode); | |
2664 | |
2665 return; | |
2666 */ | |
2159 | 2667 |
2185 | 2668 #ifdef HAVE_ODIVX_POSTPROCESS |
2669 // Note: I could make this shit outside of this file, but it would mean one | |
2670 // more function call... | |
2671 if(use_old_pp){ | |
2672 odivx_postprocess(src,src_stride,dst,dst_stride,horizontal_size,vertical_size,QP_store,QP_stride,mode); | |
2673 return; | |
2674 } | |
2675 #endif | |
2676 | |
2159 | 2677 postProcess(src[0], src_stride, dst[0], dst_stride, |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2678 horizontal_size, vertical_size, QP_store, QP_stride, 0, mode); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2679 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2680 horizontal_size >>= 1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2681 vertical_size >>= 1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2682 src_stride >>= 1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2683 dst_stride >>= 1; |
2179 | 2684 mode= ((mode&0xFF)>>4) | (mode&0xFFFFFF00); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2685 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2686 if(1) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2687 { |
2159 | 2688 postProcess(src[1], src_stride, dst[1], dst_stride, |
2179 | 2689 horizontal_size, vertical_size, QP_store, QP_stride, 1, mode); |
2159 | 2690 postProcess(src[2], src_stride, dst[2], dst_stride, |
2389 | 2691 horizontal_size, vertical_size, QP_store, QP_stride, 2, mode); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2692 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2693 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2694 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2695 memcpy(dst[1], src[1], src_stride*horizontal_size); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2696 memcpy(dst[2], src[2], src_stride*horizontal_size); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2697 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2698 } |
2185 | 2699 |
2389 | 2700 void postprocess2(unsigned char * src[], int src_stride, |
2701 unsigned char * dst[], int dst_stride, | |
2702 int horizontal_size, int vertical_size, | |
2703 QP_STORE_T *QP_store, int QP_stride, | |
2704 struct PPMode *mode) | |
2705 { | |
2706 | |
2707 #ifdef HAVE_ODIVX_POSTPROCESS | |
2708 // Note: I could make this shit outside of this file, but it would mean one | |
2709 // more function call... | |
2710 if(use_old_pp){ | |
2711 odivx_postprocess(src,src_stride,dst,dst_stride,horizontal_size,vertical_size,QP_store,QP_stride, | |
2712 mode->oldMode); | |
2713 return; | |
2714 } | |
2715 #endif | |
2716 | |
2717 postProcess(src[0], src_stride, dst[0], dst_stride, | |
2718 horizontal_size, vertical_size, QP_store, QP_stride, 0, mode->lumMode); | |
2719 | |
2720 horizontal_size >>= 1; | |
2721 vertical_size >>= 1; | |
2722 src_stride >>= 1; | |
2723 dst_stride >>= 1; | |
2724 | |
2725 postProcess(src[1], src_stride, dst[1], dst_stride, | |
2726 horizontal_size, vertical_size, QP_store, QP_stride, 1, mode->chromMode); | |
2727 postProcess(src[2], src_stride, dst[2], dst_stride, | |
2728 horizontal_size, vertical_size, QP_store, QP_stride, 2, mode->chromMode); | |
2729 } | |
2730 | |
2731 | |
2159 | 2732 /** |
2733 * gets the mode flags for a given quality (larger values mean slower but better postprocessing) | |
2185 | 2734 * 0 <= quality <= 6 |
2159 | 2735 */ |
2185 | 2736 int getPpModeForQuality(int quality){ |
2737 int modes[1+GET_PP_QUALITY_MAX]= { | |
2738 0, | |
2739 #if 1 | |
2740 // horizontal filters first | |
2741 LUM_H_DEBLOCK, | |
2742 LUM_H_DEBLOCK | LUM_V_DEBLOCK, | |
2743 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK, | |
2744 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK, | |
2745 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK | LUM_DERING, | |
2746 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK | LUM_DERING | CHROM_DERING | |
2747 #else | |
2748 // vertical filters first | |
2159 | 2749 LUM_V_DEBLOCK, |
2750 LUM_V_DEBLOCK | LUM_H_DEBLOCK, | |
2751 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK, | |
2752 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK, | |
2753 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING, | |
2754 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING | CHROM_DERING | |
2185 | 2755 #endif |
2756 }; | |
2159 | 2757 |
2185 | 2758 #ifdef HAVE_ODIVX_POSTPROCESS |
2759 int odivx_modes[1+GET_PP_QUALITY_MAX]= { | |
2760 0, | |
2761 PP_DEBLOCK_Y_H, | |
2762 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V, | |
2763 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H, | |
2764 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V, | |
2765 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V|PP_DERING_Y, | |
2766 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V|PP_DERING_Y|PP_DERING_C | |
2767 }; | |
2768 if(use_old_pp) return odivx_modes[quality]; | |
2769 #endif | |
2770 return modes[quality]; | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2771 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2772 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2773 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2774 * 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
|
2775 * numLines must be a multiple of 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2776 * levelFix == 0 -> dont touch the brighness & contrast |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2777 */ |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2778 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
|
2779 int numLines, int levelFix) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2780 { |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2781 #ifndef HAVE_MMX |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2782 int i; |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2783 #endif |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2784 if(levelFix) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2785 { |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2786 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2787 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2788 "leal (%2,%2), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2789 "leal (%3,%3), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2790 "movq packedYOffset, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2791 "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
|
2792 "pxor %%mm4, %%mm4 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2793 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2794 #define SCALED_CPY \ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2795 "movq (%0), %%mm0 \n\t"\ |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2796 "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
|
2797 "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
|
2798 "punpckhbw %%mm4, %%mm5 \n\t"\ |
2394 | 2799 "psubw %%mm2, %%mm0 \n\t"\ |
2800 "psubw %%mm2, %%mm5 \n\t"\ | |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2801 "movq (%0,%2), %%mm1 \n\t"\ |
2394 | 2802 "psllw $6, %%mm0 \n\t"\ |
2803 "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
|
2804 "pmulhw %%mm3, %%mm0 \n\t"\ |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2805 "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
|
2806 "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
|
2807 "punpcklbw %%mm4, %%mm1 \n\t"\ |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2808 "punpckhbw %%mm4, %%mm6 \n\t"\ |
2394 | 2809 "psubw %%mm2, %%mm1 \n\t"\ |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2810 "psubw %%mm2, %%mm6 \n\t"\ |
2394 | 2811 "psllw $6, %%mm1 \n\t"\ |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2812 "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
|
2813 "pmulhw %%mm3, %%mm1 \n\t"\ |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2814 "pmulhw %%mm3, %%mm6 \n\t"\ |
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2815 "addl %%eax, %0 \n\t"\ |
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2816 "packuswb %%mm5, %%mm0 \n\t"\ |
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2817 "packuswb %%mm6, %%mm1 \n\t"\ |
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2818 "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
|
2819 "movq %%mm1, (%1, %3) \n\t"\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2820 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2821 SCALED_CPY |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2822 "addl %%ebx, %1 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2823 SCALED_CPY |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2824 "addl %%ebx, %1 \n\t" |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2825 SCALED_CPY |
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2826 "addl %%ebx, %1 \n\t" |
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2827 SCALED_CPY |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2828 |
2413 | 2829 : "+r"(src), |
2830 "+r"(dst) | |
2831 :"r" (srcStride), | |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2832 "r" (dstStride) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2833 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2834 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2835 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2836 for(i=0; i<numLines; i++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2837 memcpy( &(dst[dstStride*i]), |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2838 &(src[srcStride*i]), BLOCK_SIZE); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2839 #endif |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2840 } |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2841 else |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2842 { |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2843 #ifdef HAVE_MMX |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2844 asm volatile( |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2845 "movl %4, %%eax \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2846 "movl %%eax, temp0\n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2847 "pushl %0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2848 "pushl %1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2849 "leal (%2,%2), %%eax \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2850 "leal (%3,%3), %%ebx \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2851 "movq packedYOffset, %%mm2 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2852 "movq packedYScale, %%mm3 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2853 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2854 #define SIMPLE_CPY \ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2855 "movq (%0), %%mm0 \n\t"\ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2856 "movq (%0,%2), %%mm1 \n\t"\ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2857 "movq %%mm0, (%1) \n\t"\ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2858 "movq %%mm1, (%1, %3) \n\t"\ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2859 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2860 "1: \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2861 SIMPLE_CPY |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2862 "addl %%eax, %0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2863 "addl %%ebx, %1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2864 SIMPLE_CPY |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2865 "addl %%eax, %0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2866 "addl %%ebx, %1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2867 "decl temp0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2868 "jnz 1b \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2869 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2870 "popl %1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2871 "popl %0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2872 : : "r" (src), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2873 "r" (dst), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2874 "r" (srcStride), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2875 "r" (dstStride), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2876 "m" (numLines>>2) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2877 : "%eax", "%ebx" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2878 ); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2879 #else |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2880 for(i=0; i<numLines; i++) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2881 memcpy( &(dst[dstStride*i]), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2882 &(src[srcStride*i]), BLOCK_SIZE); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2883 #endif |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2884 } |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2885 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2886 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2887 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2888 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2889 * Filters array of bytes (Y or U or V values) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2890 */ |
2185 | 2891 static void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height, |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2892 QP_STORE_T QPs[], int QPStride, int isColor, int mode) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2893 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2894 int x,y; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2895 /* 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
|
2896 after watching a black picture for 5 hours*/ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2897 static uint64_t *yHistogram= NULL; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2898 int black=0, white=255; // blackest black and whitest white in the picture |
2437 | 2899 int QPCorrecture= 256; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2900 |
2246 | 2901 /* Temporary buffers for handling the last row(s) */ |
2902 static uint8_t *tempDst= NULL; | |
2903 static uint8_t *tempSrc= NULL; | |
2904 | |
2285 | 2905 /* Temporary buffers for handling the last block */ |
2906 static uint8_t *tempDstBlock= NULL; | |
2907 static uint8_t *tempSrcBlock= NULL; | |
2908 | |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2909 #ifdef PP_FUNNY_STRIDE |
2285 | 2910 uint8_t *dstBlockPtrBackup; |
2911 uint8_t *srcBlockPtrBackup; | |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2912 #endif |
2285 | 2913 |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2914 #ifdef MORE_TIMING |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2915 long long T0, T1, diffTime=0; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2916 #endif |
2221 | 2917 #ifdef TIMING |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2918 long long memcpyTime=0, vertTime=0, horizTime=0, sumTime; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2919 sumTime= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2920 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2921 |
2246 | 2922 if(tempDst==NULL) |
2923 { | |
2924 tempDst= (uint8_t*)memalign(8, 1024*24); | |
2925 tempSrc= (uint8_t*)memalign(8, 1024*24); | |
2285 | 2926 tempDstBlock= (uint8_t*)memalign(8, 1024*24); |
2927 tempSrcBlock= (uint8_t*)memalign(8, 1024*24); | |
2246 | 2928 } |
2929 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2930 if(!yHistogram) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2931 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2932 int i; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2933 yHistogram= (uint64_t*)malloc(8*256); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2934 for(i=0; i<256; i++) yHistogram[i]= width*height/64*15/256; |
2285 | 2935 |
2936 if(mode & FULL_Y_RANGE) | |
2937 { | |
2938 maxAllowedY=255; | |
2939 minAllowedY=0; | |
2940 } | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2941 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2942 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2943 if(!isColor) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2944 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2945 uint64_t sum= 0; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2946 int i; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2947 static int framenum= -1; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2948 uint64_t maxClipped; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2949 uint64_t clipped; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2950 double scale; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2951 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2952 framenum++; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2953 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
|
2954 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2955 for(i=0; i<256; i++) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2956 { |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2957 sum+= yHistogram[i]; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2958 // printf("%d ", yHistogram[i]); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2959 } |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2960 // printf("\n\n"); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2961 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2962 /* we allways get a completly black picture first */ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2963 maxClipped= (uint64_t)(sum * maxClippedThreshold); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2964 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2965 clipped= sum; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2966 for(black=255; black>0; black--) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2967 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2968 if(clipped < maxClipped) break; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2969 clipped-= yHistogram[black]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2970 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2971 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2972 clipped= sum; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2973 for(white=0; white<256; white++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2974 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2975 if(clipped < maxClipped) break; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2976 clipped-= yHistogram[white]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2977 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2978 |
2394 | 2979 packedYOffset= (black - minAllowedY) & 0xFFFF; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2980 packedYOffset|= packedYOffset<<32; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2981 packedYOffset|= packedYOffset<<16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2982 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2983 scale= (double)(maxAllowedY - minAllowedY) / (double)(white-black); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2984 |
2394 | 2985 packedYScale= (uint16_t)(scale*1024.0 + 0.5); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2986 packedYScale|= packedYScale<<32; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2987 packedYScale|= packedYScale<<16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2988 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2989 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2990 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2991 packedYScale= 0x0100010001000100LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2992 packedYOffset= 0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2993 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2994 |
2437 | 2995 if(mode & LEVEL_FIX) QPCorrecture= packedYScale &0xFFFF; |
2996 else QPCorrecture= 256; | |
2997 | |
2246 | 2998 /* copy first row of 8x8 blocks */ |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2999 for(x=0; x<width; x+=BLOCK_SIZE) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3000 blockCopy(dst + x, dstStride, src + x, srcStride, 8, mode & LEVEL_FIX); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3001 |
2246 | 3002 for(y=0; y<height; y+=BLOCK_SIZE) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3003 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3004 //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
|
3005 uint8_t *srcBlock= &(src[y*srcStride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3006 uint8_t *dstBlock= &(dst[y*dstStride]); |
2437 | 3007 #ifdef ARCH_X86 |
3008 int *QPptr= isColor ? &QPs[(y>>3)*QPStride] :&QPs[(y>>4)*QPStride]; | |
3009 int QPDelta= isColor ? 1<<(32-3) : 1<<(32-4); | |
3010 int QPFrac= QPDelta; | |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3011 uint8_t *tempBlock1= tempBlocks; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3012 uint8_t *tempBlock2= tempBlocks + 8; |
2437 | 3013 #endif |
2473 | 3014 /* can we mess with a 8x16 block from srcBlock/dstBlock downwards and 1 line upwards |
3015 if not than use a temporary buffer */ | |
2246 | 3016 if(y+15 >= height) |
3017 { | |
2473 | 3018 /* copy from line 5 to 12 of src, these will be copied with |
2246 | 3019 blockcopy to dst later */ |
3020 memcpy(tempSrc + srcStride*5, srcBlock + srcStride*5, | |
3021 srcStride*MAX(height-y-5, 0) ); | |
3022 | |
3023 /* duplicate last line to fill the void upto line 12 */ | |
3024 if(y+12 >= height) | |
3025 { | |
3026 int i; | |
3027 for(i=height-y; i<=12; i++) | |
3028 memcpy(tempSrc + srcStride*i, | |
3029 src + srcStride*(height-1), srcStride); | |
3030 } | |
3031 | |
3032 | |
2473 | 3033 /* copy up to 6 lines of dst */ |
3034 memcpy(tempDst, dstBlock - dstStride, dstStride*MIN(height-y+1, 6) ); | |
3035 dstBlock= tempDst + dstStride; | |
2246 | 3036 srcBlock= tempSrc; |
3037 } | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3038 |
2285 | 3039 // 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
|
3040 // 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
|
3041 // 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
|
3042 for(x=0; x<width; x+=BLOCK_SIZE) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3043 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3044 const int stride= dstStride; |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3045 uint8_t *tmpXchg; |
2437 | 3046 #ifdef ARCH_X86 |
3047 int QP= *QPptr; | |
3048 asm volatile( | |
3049 "addl %2, %1 \n\t" | |
3050 "sbbl %%eax, %%eax \n\t" | |
3051 "shll $2, %%eax \n\t" | |
3052 "subl %%eax, %0 \n\t" | |
3053 : "+r" (QPptr), "+m" (QPFrac) | |
3054 : "r" (QPDelta) | |
3055 : "%eax" | |
3056 ); | |
3057 #else | |
3058 int QP= isColor ? | |
3059 QPs[(y>>3)*QPStride + (x>>3)]: | |
3060 QPs[(y>>4)*QPStride + (x>>4)]; | |
3061 #endif | |
3062 if(!isColor) | |
2428 | 3063 { |
2437 | 3064 QP= (QP* QPCorrecture)>>8; |
3065 yHistogram[ srcBlock[srcStride*4 + 4] ]++; | |
2428 | 3066 } |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3067 #ifdef HAVE_MMX |
2246 | 3068 asm volatile( |
3069 "movd %0, %%mm7 \n\t" | |
3070 "packuswb %%mm7, %%mm7 \n\t" // 0, 0, 0, QP, 0, 0, 0, QP | |
3071 "packuswb %%mm7, %%mm7 \n\t" // 0,QP, 0, QP, 0,QP, 0, QP | |
3072 "packuswb %%mm7, %%mm7 \n\t" // QP,..., QP | |
3073 "movq %%mm7, pQPb \n\t" | |
3074 : : "r" (QP) | |
3075 ); | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3076 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3077 |
2221 | 3078 #ifdef MORE_TIMING |
2246 | 3079 T0= rdtsc(); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3080 #endif |
2159 | 3081 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3082 #ifdef HAVE_MMX2 |
2437 | 3083 /* |
2246 | 3084 prefetchnta(srcBlock + (((x>>3)&3) + 5)*srcStride + 32); |
3085 prefetchnta(srcBlock + (((x>>3)&3) + 9)*srcStride + 32); | |
3086 prefetcht0(dstBlock + (((x>>3)&3) + 5)*dstStride + 32); | |
3087 prefetcht0(dstBlock + (((x>>3)&3) + 9)*dstStride + 32); | |
2437 | 3088 */ |
3089 /* | |
3090 prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32); | |
3091 prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32); | |
3092 prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32); | |
3093 prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32); | |
3094 */ | |
3095 | |
3096 asm( | |
3097 "movl %4, %%eax \n\t" | |
3098 "shrl $2, %%eax \n\t" | |
3099 "andl $6, %%eax \n\t" | |
3100 "addl $5, %%eax \n\t" | |
3101 "movl %%eax, %%ebx \n\t" | |
3102 "imul %1, %%eax \n\t" | |
3103 "imul %3, %%ebx \n\t" | |
3104 "prefetchnta 32(%%eax, %0) \n\t" | |
3105 "prefetcht0 32(%%ebx, %2) \n\t" | |
3106 "addl %1, %%eax \n\t" | |
3107 "addl %3, %%ebx \n\t" | |
3108 "prefetchnta 32(%%eax, %0) \n\t" | |
3109 "prefetcht0 32(%%ebx, %2) \n\t" | |
3110 :: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride), | |
3111 "m" (x) | |
3112 : "%eax", "%ebx" | |
3113 ); | |
3114 | |
2159 | 3115 #elif defined(HAVE_3DNOW) |
3116 //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ... | |
2246 | 3117 /* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32); |
3118 prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32); | |
3119 prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32); | |
3120 prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32); | |
2159 | 3121 */ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3122 #endif |
2246 | 3123 |
2286 | 3124 #ifdef PP_FUNNY_STRIDE |
2285 | 3125 //can we mess with a 8x16 block, if not use a temp buffer, yes again |
3126 if(x+7 >= width) | |
3127 { | |
3128 int i; | |
3129 dstBlockPtrBackup= dstBlock; | |
3130 srcBlockPtrBackup= srcBlock; | |
3131 | |
3132 for(i=0;i<BLOCK_SIZE*2; i++) | |
3133 { | |
3134 memcpy(tempSrcBlock+i*srcStride, srcBlock+i*srcStride, width-x); | |
3135 memcpy(tempDstBlock+i*dstStride, dstBlock+i*dstStride, width-x); | |
3136 } | |
3137 | |
3138 dstBlock= tempDstBlock; | |
3139 srcBlock= tempSrcBlock; | |
3140 } | |
2286 | 3141 #endif |
2285 | 3142 |
2246 | 3143 blockCopy(dstBlock + dstStride*5, dstStride, |
3144 srcBlock + srcStride*5, srcStride, 8, mode & LEVEL_FIX); | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3145 |
2246 | 3146 if(mode & LINEAR_IPOL_DEINT_FILTER) |
3147 deInterlaceInterpolateLinear(dstBlock, dstStride); | |
3148 else if(mode & LINEAR_BLEND_DEINT_FILTER) | |
3149 deInterlaceBlendLinear(dstBlock, dstStride); | |
3150 else if(mode & MEDIAN_DEINT_FILTER) | |
3151 deInterlaceMedian(dstBlock, dstStride); | |
3152 else if(mode & CUBIC_IPOL_DEINT_FILTER) | |
3153 deInterlaceInterpolateCubic(dstBlock, dstStride); | |
3154 /* else if(mode & CUBIC_BLEND_DEINT_FILTER) | |
3155 deInterlaceBlendCubic(dstBlock, dstStride); | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
3156 */ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3157 |
2246 | 3158 /* only deblock if we have 2 blocks */ |
3159 if(y + 8 < height) | |
3160 { | |
2221 | 3161 #ifdef MORE_TIMING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3162 T1= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3163 memcpyTime+= T1-T0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3164 T0=T1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3165 #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
|
3166 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
|
3167 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
|
3168 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
|
3169 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
|
3170 else if(mode & V_DEBLOCK) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3171 { |
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
|
3172 if( isVertDC(dstBlock, stride)) |
2159 | 3173 { |
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
|
3174 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
|
3175 doVertLowPass(dstBlock, stride, QP); |
2159 | 3176 } |
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
|
3177 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
|
3178 doVertDefFilter(dstBlock, stride, QP); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3179 } |
2221 | 3180 #ifdef MORE_TIMING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3181 T1= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3182 vertTime+= T1-T0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3183 T0=T1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3184 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3185 } |
2473 | 3186 |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3187 #ifdef HAVE_MMX |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3188 transpose1(tempBlock1, tempBlock2, dstBlock, dstStride); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3189 #endif |
2246 | 3190 /* check if we have a previous block to deblock it with dstBlock */ |
2285 | 3191 if(x - 8 >= 0) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3192 { |
2221 | 3193 #ifdef MORE_TIMING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3194 T0= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3195 #endif |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3196 #ifdef HAVE_MMX |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3197 if(mode & H_RK1_FILTER) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3198 vertRK1Filter(tempBlock1, 16, QP); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3199 else if(mode & H_X1_FILTER) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3200 vertX1Filter(tempBlock1, 16, QP); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3201 else if(mode & H_DEBLOCK) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3202 { |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3203 if( isVertDC(tempBlock1, 16)) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3204 { |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3205 if(isVertMinMaxOk(tempBlock1, 16, QP)) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3206 doVertLowPass(tempBlock1, 16, QP); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3207 } |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3208 else |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3209 doVertDefFilter(tempBlock1, 16, QP); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3210 } |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3211 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3212 transpose2(dstBlock-4, dstStride, tempBlock1 + 4*16); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3213 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3214 #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
|
3215 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
|
3216 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
|
3217 else if(mode & H_DEBLOCK) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3218 { |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3219 if( isHorizDC(dstBlock-4, stride)) |
2159 | 3220 { |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3221 if(isHorizMinMaxOk(dstBlock-4, stride, QP)) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3222 doHorizLowPass(dstBlock-4, stride, QP); |
2159 | 3223 } |
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
|
3224 else |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3225 doHorizDefFilter(dstBlock-4, stride, QP); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3226 } |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3227 #endif |
2221 | 3228 #ifdef MORE_TIMING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3229 T1= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3230 horizTime+= T1-T0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3231 T0=T1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3232 #endif |
2473 | 3233 if(mode & DERING) |
3234 { | |
3235 //FIXME filter first line | |
3236 if(y>0) dering(dstBlock - stride - 8, stride, QP); | |
3237 } | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3238 } |
2473 | 3239 else if(mode & DERING) |
3240 { | |
3241 //FIXME y+15 is required cuz of the tempBuffer thing -> bottom right block isnt filtered | |
3242 if(y > 8 && y+15 < height) dering(dstBlock - stride*9 + width - 8, stride, QP); | |
3243 } | |
3244 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3245 |
2286 | 3246 #ifdef PP_FUNNY_STRIDE |
2285 | 3247 /* did we use a tmp-block buffer */ |
3248 if(x+7 >= width) | |
3249 { | |
3250 int i; | |
3251 dstBlock= dstBlockPtrBackup; | |
3252 srcBlock= srcBlockPtrBackup; | |
3253 | |
3254 for(i=0;i<BLOCK_SIZE*2; i++) | |
3255 { | |
3256 memcpy(dstBlock+i*dstStride, tempDstBlock+i*dstStride, width-x); | |
3257 } | |
3258 } | |
2286 | 3259 #endif |
2285 | 3260 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3261 dstBlock+=8; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3262 srcBlock+=8; |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3263 |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
3264 #ifdef HAVE_MMX |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3265 tmpXchg= tempBlock1; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3266 tempBlock1= tempBlock2; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3267 tempBlock2 = tmpXchg; |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
3268 #endif |
2246 | 3269 } |
3270 | |
3271 /* did we use a tmp buffer */ | |
2285 | 3272 if(y+15 >= height) |
2246 | 3273 { |
3274 uint8_t *dstBlock= &(dst[y*dstStride]); | |
2473 | 3275 memcpy(dstBlock, tempDst + dstStride, dstStride*(height-y) ); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3276 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3277 } |
2159 | 3278 #ifdef HAVE_3DNOW |
3279 asm volatile("femms"); | |
3280 #elif defined (HAVE_MMX) | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3281 asm volatile("emms"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3282 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3283 |
2221 | 3284 #ifdef TIMING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3285 // 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
|
3286 sumTime= rdtsc() - sumTime; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3287 if(!isColor) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3288 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
|
3289 (int)(memcpyTime/1000), (int)(vertTime/1000), (int)(horizTime/1000), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3290 (int)(sumTime/1000), (int)((sumTime-memcpyTime-vertTime-horizTime)/1000) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3291 , black, white); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3292 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3293 } |