Mercurial > mplayer.hg
annotate postproc/postprocess.c @ 2566:a350d8bed636
bugfixes
small speedup for the non x86 stuff
author | michael |
---|---|
date | Tue, 30 Oct 2001 20:30:24 +0000 |
parents | 42d5846eeb51 |
children | af43a83122fc |
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 |
2477 | 29 deRing E e e* |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
30 Vertical RKAlgo1 E a a |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
31 Horizontal RKAlgo1 a a |
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 |
2477 | 96 #ifdef HAVE_MMX2 |
97 #define PMINUB(a,b,t) "pminub " #a ", " #b " \n\t" | |
98 #elif defined (HAVE_MMX) | |
99 #define PMINUB(b,a,t) \ | |
100 "movq " #a ", " #t " \n\t"\ | |
101 "psubusb " #b ", " #t " \n\t"\ | |
102 "psubb " #t ", " #a " \n\t" | |
103 #endif | |
104 | |
105 #ifdef HAVE_MMX2 | |
106 #define PMAXUB(a,b) "pmaxub " #a ", " #b " \n\t" | |
107 #elif defined (HAVE_MMX) | |
108 #define PMAXUB(a,b) \ | |
109 "psubusb " #a ", " #b " \n\t"\ | |
110 "paddb " #a ", " #b " \n\t" | |
111 #endif | |
112 | |
113 | |
2389 | 114 #define GET_MODE_BUFFER_SIZE 500 |
115 #define OPTIONS_ARRAY_SIZE 10 | |
116 | |
117 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
118 static uint64_t packedYOffset= 0x0000000000000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
119 static uint64_t packedYScale= 0x0100010001000100LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
120 static uint64_t w05= 0x0005000500050005LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
121 static uint64_t w20= 0x0020002000200020LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
122 static uint64_t w1400= 0x1400140014001400LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
123 static uint64_t bm00000001= 0x00000000000000FFLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
124 static uint64_t bm00010000= 0x000000FF00000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
125 static uint64_t bm00001000= 0x00000000FF000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
126 static uint64_t bm10000000= 0xFF00000000000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
127 static uint64_t bm10000001= 0xFF000000000000FFLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
128 static uint64_t bm11000011= 0xFFFF00000000FFFFLL; |
2159 | 129 static uint64_t bm00000011= 0x000000000000FFFFLL; |
2179 | 130 static uint64_t bm11111110= 0xFFFFFFFFFFFFFF00LL; |
2159 | 131 static uint64_t bm11000000= 0xFFFF000000000000LL; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
132 static uint64_t bm00011000= 0x000000FFFF000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
133 static uint64_t bm00110011= 0x0000FFFF0000FFFFLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
134 static uint64_t bm11001100= 0xFFFF0000FFFF0000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
135 static uint64_t b00= 0x0000000000000000LL; |
2179 | 136 static uint64_t b01= 0x0101010101010101LL; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
137 static uint64_t b02= 0x0202020202020202LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
138 static uint64_t b0F= 0x0F0F0F0F0F0F0F0FLL; |
2473 | 139 static uint64_t b04= 0x0404040404040404LL; |
140 static uint64_t b08= 0x0808080808080808LL; | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
141 static uint64_t bFF= 0xFFFFFFFFFFFFFFFFLL; |
2159 | 142 static uint64_t b20= 0x2020202020202020LL; |
143 static uint64_t b80= 0x8080808080808080LL; | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
144 static uint64_t b7E= 0x7E7E7E7E7E7E7E7ELL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
145 static uint64_t b7C= 0x7C7C7C7C7C7C7C7CLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
146 static uint64_t b3F= 0x3F3F3F3F3F3F3F3FLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
147 static uint64_t temp0=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
148 static uint64_t temp1=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
149 static uint64_t temp2=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
150 static uint64_t temp3=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
151 static uint64_t temp4=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
152 static uint64_t temp5=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
153 static uint64_t pQPb=0; |
2473 | 154 static uint64_t pQPb2=0; |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
155 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
|
156 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
157 int hFlatnessThreshold= 56 - 16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
158 int vFlatnessThreshold= 56 - 16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
159 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
160 //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
|
161 double maxClippedThreshold= 0.01; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
162 |
2389 | 163 int maxAllowedY=234; |
2285 | 164 int minAllowedY=16; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
165 |
2389 | 166 static struct PPFilter filters[]= |
167 { | |
168 {"hb", "hdeblock", 1, 1, 3, H_DEBLOCK}, | |
169 {"vb", "vdeblock", 1, 2, 4, V_DEBLOCK}, | |
170 {"vr", "rkvdeblock", 1, 2, 4, H_RK1_FILTER}, | |
171 {"h1", "x1hdeblock", 1, 1, 3, H_X1_FILTER}, | |
172 {"v1", "x1vdeblock", 1, 2, 4, V_X1_FILTER}, | |
173 {"dr", "dering", 1, 5, 6, DERING}, | |
174 {"al", "autolevels", 0, 1, 2, LEVEL_FIX}, | |
175 {"lb", "linblenddeint", 0, 1, 6, LINEAR_BLEND_DEINT_FILTER}, | |
176 {"li", "linipoldeint", 0, 1, 6, LINEAR_IPOL_DEINT_FILTER}, | |
177 {"ci", "cubicipoldeint", 0, 1, 6, CUBIC_IPOL_DEINT_FILTER}, | |
178 {"md", "mediandeint", 0, 1, 6, MEDIAN_DEINT_FILTER}, | |
179 {NULL, NULL,0,0,0,0} //End Marker | |
180 }; | |
181 | |
182 static char *replaceTable[]= | |
183 { | |
184 "default", "hdeblock:a,vdeblock:a,dering:a,autolevels", | |
185 "de", "hdeblock:a,vdeblock:a,dering:a,autolevels", | |
186 "fast", "x1hdeblock:a,x1vdeblock:a,dering:a,autolevels", | |
187 "fa", "x1hdeblock:a,x1vdeblock:a,dering:a,autolevels", | |
188 NULL //End Marker | |
189 }; | |
190 | |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
191 static inline void unusedVariableWarningFixer() |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
192 { |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
193 if( |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
194 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
|
195 + bm00001000 + bm10000000 + bm10000001 + bm11000011 + bm00000011 + bm11111110 |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
196 + 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
|
197 + 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
|
198 + temp5 + pQPb== 0) b00=0; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
199 } |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
200 |
2221 | 201 #ifdef TIMING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
202 static inline long long rdtsc() |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
203 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
204 long long l; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
205 asm volatile( "rdtsc\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
206 : "=A" (l) |
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 // printf("%d\n", int(l/1000)); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
209 return l; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
210 } |
2185 | 211 #endif |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
212 |
2185 | 213 #ifdef HAVE_MMX2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
214 static inline void prefetchnta(void *p) |
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 asm volatile( "prefetchnta (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
217 : : "r" (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 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
220 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
221 static inline void prefetcht0(void *p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
222 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
223 asm volatile( "prefetcht0 (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
224 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
225 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
226 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
227 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
228 static inline void prefetcht1(void *p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
229 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
230 asm volatile( "prefetcht1 (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
231 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
232 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
233 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
234 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
235 static inline void prefetcht2(void *p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
236 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
237 asm volatile( "prefetcht2 (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
238 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
239 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
240 } |
2185 | 241 #endif |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
242 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
243 //FIXME? |255-0| = 1 (shouldnt be a problem ...) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
244 /** |
2246 | 245 * 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
|
246 */ |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
247 static inline int isVertDC(uint8_t src[], int stride){ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
248 int numEq= 0; |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
249 #ifndef HAVE_MMX |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
250 int y; |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
251 #endif |
2246 | 252 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
|
253 #ifdef HAVE_MMX |
2413 | 254 asm volatile( |
255 "leal (%1, %2), %%eax \n\t" | |
256 "leal (%%eax, %2, 4), %%ebx \n\t" | |
257 // 0 1 2 3 4 5 6 7 8 9 | |
258 // %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
|
259 "movq b7E, %%mm7 \n\t" // mm7 = 0x7F |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
260 "movq b7C, %%mm6 \n\t" // mm6 = 0x7D |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
261 "movq (%1), %%mm0 \n\t" |
2413 | 262 "movq (%%eax), %%mm1 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
263 "psubb %%mm1, %%mm0 \n\t" // mm0 = differnece |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
264 "paddb %%mm7, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
265 "pcmpgtb %%mm6, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
266 |
2413 | 267 "movq (%%eax,%2), %%mm2 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
268 "psubb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
269 "paddb %%mm7, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
270 "pcmpgtb %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
271 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
272 |
2413 | 273 "movq (%%eax, %2, 2), %%mm1 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
274 "psubb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
275 "paddb %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
276 "pcmpgtb %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
277 "paddb %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
278 |
2413 | 279 "movq (%1, %2, 4), %%mm2 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
280 "psubb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
281 "paddb %%mm7, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
282 "pcmpgtb %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
283 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
284 |
2413 | 285 "movq (%%ebx), %%mm1 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
286 "psubb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
287 "paddb %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
288 "pcmpgtb %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
289 "paddb %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
290 |
2413 | 291 "movq (%%ebx, %2), %%mm2 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
292 "psubb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
293 "paddb %%mm7, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
294 "pcmpgtb %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
295 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
296 |
2413 | 297 "movq (%%ebx, %2, 2), %%mm1 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
298 "psubb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
299 "paddb %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
300 "pcmpgtb %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
301 "paddb %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
302 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
303 " \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
304 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
305 "psrlw $8, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
306 "paddb %%mm1, %%mm0 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
307 #ifdef HAVE_MMX2 |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
308 "pshufw $0xF9, %%mm0, %%mm1 \n\t" |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
309 "paddb %%mm1, %%mm0 \n\t" |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
310 "pshufw $0xFE, %%mm0, %%mm1 \n\t" |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
311 #else |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
312 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
313 "psrlq $16, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
314 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
315 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
316 "psrlq $32, %%mm0 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
317 #endif |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
318 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
319 "movd %%mm0, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
320 : "=r" (numEq) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
321 : "r" (src), "r" (stride) |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
322 : "%eax", "%ebx" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
323 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
324 |
2413 | 325 numEq= (256 - numEq) &0xFF; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
326 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
327 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
328 for(y=0; y<BLOCK_SIZE-1; y++) |
2158
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 if(((src[0] - src[0+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
331 if(((src[1] - src[1+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
332 if(((src[2] - src[2+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
333 if(((src[3] - src[3+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
334 if(((src[4] - src[4+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
335 if(((src[5] - src[5+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
336 if(((src[6] - src[6+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
337 if(((src[7] - src[7+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
338 src+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
339 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
340 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
341 /* if(abs(numEq - asmEq) > 0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
342 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
343 printf("\nasm:%d c:%d\n", asmEq, numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
344 for(int y=0; y<8; y++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
345 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
346 for(int x=0; x<8; x++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
347 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
348 printf("%d ", temp[x + y*stride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
349 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
350 printf("\n"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
351 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
352 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
353 */ |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
354 // 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
|
355 return (numEq > vFlatnessThreshold) ? 1 : 0; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
356 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
357 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
358 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
|
359 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
360 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
361 int isOk; |
2246 | 362 src+= stride*3; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
363 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
364 // "int $3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
365 "movq (%1, %2), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
366 "movq (%1, %2, 8), %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
367 "movq %%mm0, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
368 "psubusb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
369 "psubusb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
370 "por %%mm1, %%mm0 \n\t" // ABS Diff |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
371 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
372 "movq pQPb, %%mm7 \n\t" // QP,..., QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
373 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
374 "psubusb %%mm7, %%mm0 \n\t" // Diff <= 2QP -> 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
375 "pcmpeqd b00, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
376 "psrlq $16, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
377 "pcmpeqd bFF, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
378 // "movd %%mm0, (%1, %2, 4)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
379 "movd %%mm0, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
380 : "=r" (isOk) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
381 : "r" (src), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
382 ); |
2416 | 383 return isOk; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
384 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
385 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
386 int isOk2= 1; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
387 int x; |
2246 | 388 src+= stride*3; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
389 for(x=0; x<BLOCK_SIZE; x++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
390 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
391 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
|
392 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
393 /* if(isOk && !isOk2 || !isOk && isOk2) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
394 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
395 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
|
396 for(int y=0; y<9; y++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
397 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
398 for(int x=0; x<8; x++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
399 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
400 printf("%d ", src[x + y*stride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
401 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
402 printf("\n"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
403 } |
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 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
406 return isOk2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
407 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
408 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
409 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
410 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
411 /** |
2246 | 412 * Do a vertical low pass filter on the 8x16 block (only write to the 8x8 block in the middle) |
2221 | 413 * 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
|
414 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
415 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
|
416 { |
2159 | 417 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2246 | 418 src+= stride*3; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
419 asm volatile( //"movv %0 %1 %2\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
420 "pushl %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
421 "movq pQPb, %%mm0 \n\t" // QP,..., QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
422 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
423 "movq (%0), %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
424 "movq (%0, %1), %%mm5 \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 %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
427 "psubusb %%mm6, %%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, %%mm6 \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, %%mm6 \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 "movq (%0, %1, 8), %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
438 "leal (%0, %1, 4), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
439 "leal (%0, %1, 8), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
440 "subl %1, %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
441 "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
|
442 "movq (%0, %1, 8), %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
443 "movq %%mm5, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
444 "movq %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
445 "psubusb %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
446 "psubusb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
447 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
448 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
449 "pcmpeqb b00, %%mm2 \n\t" // diff <= QP -> FF |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
450 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
451 "pand %%mm2, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
452 "pandn %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
453 "por %%mm2, %%mm7 \n\t" // First Line to Filter |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
454 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
455 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
456 // 1 2 3 4 5 6 7 8 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
457 // %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
|
458 // 6 4 2 2 1 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
459 // 6 4 4 2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
460 // 6 8 2 |
2246 | 461 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
462 "movq (%0, %1), %%mm0 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
463 "movq %%mm0, %%mm1 \n\t" // 1 |
2159 | 464 PAVGB(%%mm6, %%mm0) //1 1 /2 |
465 PAVGB(%%mm6, %%mm0) //3 1 /4 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
466 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
467 "movq (%0, %1, 4), %%mm2 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
468 "movq %%mm2, %%mm5 \n\t" // 1 |
2159 | 469 PAVGB((%%eax), %%mm2) // 11 /2 |
470 PAVGB((%0, %1, 2), %%mm2) // 211 /4 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
471 "movq %%mm2, %%mm3 \n\t" // 211 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
472 "movq (%0), %%mm4 \n\t" // 1 |
2159 | 473 PAVGB(%%mm4, %%mm3) // 4 211 /8 |
474 PAVGB(%%mm0, %%mm3) //642211 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
475 "movq %%mm3, (%0) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
476 // 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
|
477 "movq %%mm1, %%mm0 \n\t" // 1 |
2159 | 478 PAVGB(%%mm6, %%mm0) //1 1 /2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
479 "movq %%mm4, %%mm3 \n\t" // 1 |
2159 | 480 PAVGB((%0,%1,2), %%mm3) // 1 1 /2 |
481 PAVGB((%%eax,%1,2), %%mm5) // 11 /2 | |
482 PAVGB((%%eax), %%mm5) // 211 /4 | |
483 PAVGB(%%mm5, %%mm3) // 2 2211 /8 | |
484 PAVGB(%%mm0, %%mm3) //4242211 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
485 "movq %%mm3, (%0,%1) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
486 // mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9 |
2159 | 487 PAVGB(%%mm4, %%mm6) //11 /2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
488 "movq (%%ebx), %%mm0 \n\t" // 1 |
2159 | 489 PAVGB((%%eax, %1, 2), %%mm0) // 11/2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
490 "movq %%mm0, %%mm3 \n\t" // 11/2 |
2159 | 491 PAVGB(%%mm1, %%mm0) // 2 11/4 |
492 PAVGB(%%mm6, %%mm0) //222 11/8 | |
493 PAVGB(%%mm2, %%mm0) //22242211/16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
494 "movq (%0, %1, 2), %%mm2 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
495 "movq %%mm0, (%0, %1, 2) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
496 // 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
|
497 "movq (%%eax, %1, 4), %%mm0 \n\t" // 1 |
2159 | 498 PAVGB((%%ebx), %%mm0) // 11 /2 |
499 PAVGB(%%mm0, %%mm6) //11 11 /4 | |
500 PAVGB(%%mm1, %%mm4) // 11 /2 | |
501 PAVGB(%%mm2, %%mm1) // 11 /2 | |
502 PAVGB(%%mm1, %%mm6) //1122 11 /8 | |
503 PAVGB(%%mm5, %%mm6) //112242211 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
504 "movq (%%eax), %%mm5 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
505 "movq %%mm6, (%%eax) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
506 // 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
|
507 "movq (%%eax, %1, 4), %%mm6 \n\t" // 1 |
2159 | 508 PAVGB(%%mm7, %%mm6) // 11 /2 |
509 PAVGB(%%mm4, %%mm6) // 11 11 /4 | |
510 PAVGB(%%mm3, %%mm6) // 11 2211 /8 | |
511 PAVGB(%%mm5, %%mm2) // 11 /2 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
512 "movq (%0, %1, 4), %%mm4 \n\t" // 1 |
2159 | 513 PAVGB(%%mm4, %%mm2) // 112 /4 |
514 PAVGB(%%mm2, %%mm6) // 112242211 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
515 "movq %%mm6, (%0, %1, 4) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
516 // mm0=7(11) mm1=2(11) mm2=3(112) mm3=6(11) mm4=5 mm5=4 mm7=9 |
2159 | 517 PAVGB(%%mm7, %%mm1) // 11 2 /4 |
518 PAVGB(%%mm4, %%mm5) // 11 /2 | |
519 PAVGB(%%mm5, %%mm0) // 11 11 /4 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
520 "movq (%%eax, %1, 2), %%mm6 \n\t" // 1 |
2159 | 521 PAVGB(%%mm6, %%mm1) // 11 4 2 /8 |
522 PAVGB(%%mm0, %%mm1) // 11224222 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
523 "movq %%mm1, (%%eax, %1, 2) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
524 // mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9 |
2159 | 525 PAVGB((%%ebx), %%mm2) // 112 4 /8 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
526 "movq (%%eax, %1, 4), %%mm0 \n\t" // 1 |
2159 | 527 PAVGB(%%mm0, %%mm6) // 1 1 /2 |
528 PAVGB(%%mm7, %%mm6) // 1 12 /4 | |
529 PAVGB(%%mm2, %%mm6) // 1122424 /4 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
530 "movq %%mm6, (%%ebx) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
531 // mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9 |
2159 | 532 PAVGB(%%mm7, %%mm5) // 11 2 /4 |
533 PAVGB(%%mm7, %%mm5) // 11 6 /8 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
534 |
2159 | 535 PAVGB(%%mm3, %%mm0) // 112 /4 |
536 PAVGB(%%mm0, %%mm5) // 112246 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
537 "movq %%mm5, (%%eax, %1, 4) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
538 "popl %0\n\t" |
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 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
541 : "r" (src), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
542 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
543 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
544 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
545 const int l1= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
546 const int l2= stride + l1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
547 const int l3= stride + l2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
548 const int l4= stride + l3; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
549 const int l5= stride + l4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
550 const int l6= stride + l5; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
551 const int l7= stride + l6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
552 const int l8= stride + l7; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
553 const int l9= stride + l8; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
554 int x; |
2246 | 555 src+= stride*3; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
556 for(x=0; x<BLOCK_SIZE; x++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
557 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
558 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
|
559 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
|
560 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
561 int sums[9]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
562 sums[0] = first + src[l1]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
563 sums[1] = src[l1] + src[l2]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
564 sums[2] = src[l2] + src[l3]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
565 sums[3] = src[l3] + src[l4]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
566 sums[4] = src[l4] + src[l5]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
567 sums[5] = src[l5] + src[l6]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
568 sums[6] = src[l6] + src[l7]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
569 sums[7] = src[l7] + src[l8]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
570 sums[8] = src[l8] + last; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
571 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
572 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
|
573 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
|
574 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
|
575 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
|
576 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
|
577 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
|
578 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
|
579 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
|
580 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
581 src++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
582 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
583 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
584 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
585 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
586 |
2159 | 587 /** |
588 * Experimental implementation of the filter (Algorithm 1) described in a paper from Ramkishor & Karandikar | |
589 * values are correctly clipped (MMX2) | |
590 * values are wraparound (C) | |
591 * conclusion: its fast, but introduces ugly horizontal patterns if there is a continious gradient | |
592 0 8 16 24 | |
593 x = 8 | |
594 x/2 = 4 | |
595 x/8 = 1 | |
596 1 12 12 23 | |
597 */ | |
2179 | 598 static inline void vertRK1Filter(uint8_t *src, int stride, int QP) |
2159 | 599 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
600 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2246 | 601 src+= stride*3; |
2159 | 602 // FIXME rounding |
603 asm volatile( | |
604 "pxor %%mm7, %%mm7 \n\t" // 0 | |
605 "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE | |
606 "leal (%0, %1), %%eax \n\t" | |
607 "leal (%%eax, %1, 4), %%ebx \n\t" | |
608 // 0 1 2 3 4 5 6 7 8 9 | |
609 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 | |
610 "movq pQPb, %%mm0 \n\t" // QP,..., QP | |
611 "movq %%mm0, %%mm1 \n\t" // QP,..., QP | |
612 "paddusb b02, %%mm0 \n\t" | |
613 "psrlw $2, %%mm0 \n\t" | |
614 "pand b3F, %%mm0 \n\t" // QP/4,..., QP/4 | |
615 "paddusb %%mm1, %%mm0 \n\t" // QP*1.25 ... | |
616 "movq (%0, %1, 4), %%mm2 \n\t" // line 4 | |
617 "movq (%%ebx), %%mm3 \n\t" // line 5 | |
618 "movq %%mm2, %%mm4 \n\t" // line 4 | |
619 "pcmpeqb %%mm5, %%mm5 \n\t" // -1 | |
620 "pxor %%mm2, %%mm5 \n\t" // -line 4 - 1 | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
621 PAVGB(%%mm3, %%mm5) |
2159 | 622 "paddb %%mm6, %%mm5 \n\t" // (l5-l4)/2 |
623 "psubusb %%mm3, %%mm4 \n\t" | |
624 "psubusb %%mm2, %%mm3 \n\t" | |
625 "por %%mm3, %%mm4 \n\t" // |l4 - l5| | |
626 "psubusb %%mm0, %%mm4 \n\t" | |
627 "pcmpeqb %%mm7, %%mm4 \n\t" | |
628 "pand %%mm4, %%mm5 \n\t" // d/2 | |
629 | |
630 // "paddb %%mm6, %%mm2 \n\t" // line 4 + 0x80 | |
631 "paddb %%mm5, %%mm2 \n\t" | |
632 // "psubb %%mm6, %%mm2 \n\t" | |
633 "movq %%mm2, (%0,%1, 4) \n\t" | |
634 | |
635 "movq (%%ebx), %%mm2 \n\t" | |
636 // "paddb %%mm6, %%mm2 \n\t" // line 5 + 0x80 | |
637 "psubb %%mm5, %%mm2 \n\t" | |
638 // "psubb %%mm6, %%mm2 \n\t" | |
639 "movq %%mm2, (%%ebx) \n\t" | |
640 | |
641 "paddb %%mm6, %%mm5 \n\t" | |
642 "psrlw $2, %%mm5 \n\t" | |
643 "pand b3F, %%mm5 \n\t" | |
644 "psubb b20, %%mm5 \n\t" // (l5-l4)/8 | |
645 | |
646 "movq (%%eax, %1, 2), %%mm2 \n\t" | |
647 "paddb %%mm6, %%mm2 \n\t" // line 3 + 0x80 | |
648 "paddsb %%mm5, %%mm2 \n\t" | |
649 "psubb %%mm6, %%mm2 \n\t" | |
650 "movq %%mm2, (%%eax, %1, 2) \n\t" | |
651 | |
652 "movq (%%ebx, %1), %%mm2 \n\t" | |
653 "paddb %%mm6, %%mm2 \n\t" // line 6 + 0x80 | |
654 "psubsb %%mm5, %%mm2 \n\t" | |
655 "psubb %%mm6, %%mm2 \n\t" | |
656 "movq %%mm2, (%%ebx, %1) \n\t" | |
657 | |
658 : | |
659 : "r" (src), "r" (stride) | |
660 : "%eax", "%ebx" | |
661 ); | |
662 #else | |
663 const int l1= stride; | |
664 const int l2= stride + l1; | |
665 const int l3= stride + l2; | |
666 const int l4= stride + l3; | |
667 const int l5= stride + l4; | |
668 const int l6= stride + l5; | |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
669 // const int l7= stride + l6; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
670 // const int l8= stride + l7; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
671 // const int l9= stride + l8; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
672 int x; |
2246 | 673 src+= stride*3; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
674 for(x=0; x<BLOCK_SIZE; x++) |
2159 | 675 { |
676 if(ABS(src[l4]-src[l5]) < QP + QP/4) | |
677 { | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
678 int v = (src[l5] - src[l4]); |
2159 | 679 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
680 src[l3] +=v/8; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
681 src[l4] +=v/2; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
682 src[l5] -=v/2; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
683 src[l6] -=v/8; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
684 |
2159 | 685 } |
686 src++; | |
687 } | |
688 | |
689 #endif | |
690 } | |
691 | |
692 /** | |
693 * Experimental Filter 1 | |
2179 | 694 * will not damage linear gradients |
695 * 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
|
696 * 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
|
697 * MMX2 version does correct clipping C version doesnt |
2159 | 698 */ |
699 static inline void vertX1Filter(uint8_t *src, int stride, int QP) | |
700 { | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
701 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2246 | 702 src+= stride*3; |
703 | |
2159 | 704 asm volatile( |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
705 "pxor %%mm7, %%mm7 \n\t" // 0 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
706 // "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
707 "leal (%0, %1), %%eax \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
708 "leal (%%eax, %1, 4), %%ebx \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
709 // 0 1 2 3 4 5 6 7 8 9 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
710 // %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
|
711 "movq (%%eax, %1, 2), %%mm0 \n\t" // line 3 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
712 "movq (%0, %1, 4), %%mm1 \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
713 "movq %%mm1, %%mm2 \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
714 "psubusb %%mm0, %%mm1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
715 "psubusb %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
716 "por %%mm1, %%mm0 \n\t" // |l2 - l3| |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
717 "movq (%%ebx), %%mm3 \n\t" // line 5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
718 "movq (%%ebx, %1), %%mm4 \n\t" // line 6 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
719 "movq %%mm3, %%mm5 \n\t" // line 5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
720 "psubusb %%mm4, %%mm3 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
721 "psubusb %%mm5, %%mm4 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
722 "por %%mm4, %%mm3 \n\t" // |l5 - l6| |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
723 PAVGB(%%mm3, %%mm0) // (|l2 - l3| + |l5 - l6|)/2 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
724 "movq %%mm2, %%mm1 \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
725 "psubusb %%mm5, %%mm2 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
726 "movq %%mm2, %%mm4 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
727 "pcmpeqb %%mm7, %%mm2 \n\t" // (l4 - l5) <= 0 ? -1 : 0 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
728 "psubusb %%mm1, %%mm5 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
729 "por %%mm5, %%mm4 \n\t" // |l4 - l5| |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
730 "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
|
731 "movq %%mm4, %%mm3 \n\t" // d |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
732 "psubusb pQPb, %%mm4 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
733 "pcmpeqb %%mm7, %%mm4 \n\t" // d <= QP ? -1 : 0 |
2179 | 734 "psubusb b01, %%mm3 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
735 "pand %%mm4, %%mm3 \n\t" // d <= QP ? d : 0 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
736 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
737 PAVGB(%%mm7, %%mm3) // d/2 |
2179 | 738 "movq %%mm3, %%mm1 \n\t" // d/2 |
739 PAVGB(%%mm7, %%mm3) // d/4 | |
740 PAVGB(%%mm1, %%mm3) // 3*d/8 | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
741 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
742 "movq (%0, %1, 4), %%mm0 \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
743 "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
|
744 "psubusb %%mm3, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
745 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
746 "movq %%mm0, (%0, %1, 4) \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
747 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
748 "movq (%%ebx), %%mm0 \n\t" // line 5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
749 "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
|
750 "paddusb %%mm3, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
751 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
752 "movq %%mm0, (%%ebx) \n\t" // line 5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
753 |
2179 | 754 PAVGB(%%mm7, %%mm1) // d/4 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
755 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
756 "movq (%%eax, %1, 2), %%mm0 \n\t" // line 3 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
757 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4 |
2179 | 758 "psubusb %%mm1, %%mm0 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
759 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
760 "movq %%mm0, (%%eax, %1, 2) \n\t" // line 3 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
761 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
762 "movq (%%ebx, %1), %%mm0 \n\t" // line 6 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
763 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5 |
2179 | 764 "paddusb %%mm1, %%mm0 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
765 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
766 "movq %%mm0, (%%ebx, %1) \n\t" // line 6 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
767 |
2179 | 768 PAVGB(%%mm7, %%mm1) // d/8 |
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 "movq (%%eax, %1), %%mm0 \n\t" // line 2 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
771 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l2-1 : l2 |
2179 | 772 "psubusb %%mm1, %%mm0 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
773 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
774 "movq %%mm0, (%%eax, %1) \n\t" // line 2 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
775 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
776 "movq (%%ebx, %1, 2), %%mm0 \n\t" // line 7 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
777 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l7-1 : l7 |
2179 | 778 "paddusb %%mm1, %%mm0 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
779 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
780 "movq %%mm0, (%%ebx, %1, 2) \n\t" // line 7 |
2159 | 781 |
782 : | |
783 : "r" (src), "r" (stride) | |
784 : "%eax", "%ebx" | |
785 ); | |
786 #else | |
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 const int l1= stride; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
789 const int l2= stride + l1; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
790 const int l3= stride + l2; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
791 const int l4= stride + l3; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
792 const int l5= stride + l4; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
793 const int l6= stride + l5; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
794 const int l7= stride + l6; |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
795 // const int l8= stride + l7; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
796 // const int l9= stride + l8; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
797 int x; |
2246 | 798 |
799 src+= stride*3; | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
800 for(x=0; x<BLOCK_SIZE; x++) |
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 int a= src[l3] - src[l4]; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
803 int b= src[l4] - src[l5]; |
2179 | 804 int c= src[l5] - src[l6]; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
805 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
806 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
|
807 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
808 if(d < QP) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
809 { |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
810 int v = d * SIGN(-b); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
811 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
812 src[l2] +=v/8; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
813 src[l3] +=v/4; |
2179 | 814 src[l4] +=3*v/8; |
815 src[l5] -=3*v/8; | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
816 src[l6] -=v/4; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
817 src[l7] -=v/8; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
818 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
819 } |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
820 src++; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
821 } |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
822 /* |
2159 | 823 const int l1= stride; |
824 const int l2= stride + l1; | |
825 const int l3= stride + l2; | |
826 const int l4= stride + l3; | |
827 const int l5= stride + l4; | |
828 const int l6= stride + l5; | |
829 const int l7= stride + l6; | |
830 const int l8= stride + l7; | |
831 const int l9= stride + l8; | |
832 for(int x=0; x<BLOCK_SIZE; x++) | |
833 { | |
834 int v2= src[l2]; | |
835 int v3= src[l3]; | |
836 int v4= src[l4]; | |
837 int v5= src[l5]; | |
838 int v6= src[l6]; | |
839 int v7= src[l7]; | |
840 | |
841 if(ABS(v4-v5)<QP && ABS(v4-v5) - (ABS(v3-v4) + ABS(v5-v6))>0 ) | |
842 { | |
843 src[l3] = (6*v2 + 4*v3 + 3*v4 + 2*v5 + v6 )/16; | |
844 src[l4] = (3*v2 + 3*v3 + 4*v4 + 3*v5 + 2*v6 + v7 )/16; | |
845 src[l5] = (1*v2 + 2*v3 + 3*v4 + 4*v5 + 3*v6 + 3*v7)/16; | |
846 src[l6] = ( 1*v3 + 2*v4 + 3*v5 + 4*v6 + 6*v7)/16; | |
847 } | |
848 src++; | |
849 } | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
850 */ |
2159 | 851 #endif |
852 } | |
853 | |
2179 | 854 /** |
855 * Experimental Filter 1 (Horizontal) | |
856 * will not damage linear gradients | |
857 * Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter | |
858 * can only smooth blocks at the expected locations (it cant smooth them if they did move) | |
859 * MMX2 version does correct clipping C version doesnt | |
860 * not identical with the vertical one | |
861 */ | |
862 static inline void horizX1Filter(uint8_t *src, int stride, int QP) | |
863 { | |
864 int y; | |
865 static uint64_t *lut= NULL; | |
866 if(lut==NULL) | |
867 { | |
868 int i; | |
869 lut= (uint64_t*)memalign(8, 256*8); | |
870 for(i=0; i<256; i++) | |
871 { | |
872 int v= i < 128 ? 2*i : 2*(i-256); | |
873 /* | |
874 //Simulate 112242211 9-Tap filter | |
875 uint64_t a= (v/16) & 0xFF; | |
876 uint64_t b= (v/8) & 0xFF; | |
877 uint64_t c= (v/4) & 0xFF; | |
878 uint64_t d= (3*v/8) & 0xFF; | |
879 */ | |
880 //Simulate piecewise linear interpolation | |
881 uint64_t a= (v/16) & 0xFF; | |
882 uint64_t b= (v*3/16) & 0xFF; | |
883 uint64_t c= (v*5/16) & 0xFF; | |
884 uint64_t d= (7*v/16) & 0xFF; | |
885 uint64_t A= (0x100 - a)&0xFF; | |
886 uint64_t B= (0x100 - b)&0xFF; | |
887 uint64_t C= (0x100 - c)&0xFF; | |
888 uint64_t D= (0x100 - c)&0xFF; | |
889 | |
890 lut[i] = (a<<56) | (b<<48) | (c<<40) | (d<<32) | | |
891 (D<<24) | (C<<16) | (B<<8) | (A); | |
892 //lut[i] = (v<<32) | (v<<24); | |
893 } | |
894 } | |
895 | |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
896 #if 0 |
2179 | 897 asm volatile( |
898 "pxor %%mm7, %%mm7 \n\t" // 0 | |
899 // "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE | |
900 "leal (%0, %1), %%eax \n\t" | |
901 "leal (%%eax, %1, 4), %%ebx \n\t" | |
902 | |
903 "movq b80, %%mm6 \n\t" | |
2180 | 904 "movd pQPb, %%mm5 \n\t" // QP |
2179 | 905 "movq %%mm5, %%mm4 \n\t" |
906 "paddusb %%mm5, %%mm5 \n\t" // 2QP | |
907 "paddusb %%mm5, %%mm4 \n\t" // 3QP | |
908 "pxor %%mm5, %%mm5 \n\t" // 0 | |
909 "psubb %%mm4, %%mm5 \n\t" // -3QP | |
910 "por bm11111110, %%mm5 \n\t" // ...,FF,FF,-3QP | |
911 "psllq $24, %%mm5 \n\t" | |
912 | |
913 // 0 1 2 3 4 5 6 7 8 9 | |
914 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 | |
915 | |
916 #define HX1old(a) \ | |
917 "movd " #a ", %%mm0 \n\t"\ | |
918 "movd 4" #a ", %%mm1 \n\t"\ | |
919 "punpckldq %%mm1, %%mm0 \n\t"\ | |
920 "movq %%mm0, %%mm1 \n\t"\ | |
921 "movq %%mm0, %%mm2 \n\t"\ | |
922 "psrlq $8, %%mm1 \n\t"\ | |
923 "psubusb %%mm1, %%mm2 \n\t"\ | |
924 "psubusb %%mm0, %%mm1 \n\t"\ | |
925 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ | |
926 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
927 "pshufw $0x00, %%mm1, %%mm3 \n\t" /* p´5 = |p1 - p2| */\ | |
928 PAVGB(%%mm1, %%mm3) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\ | |
929 "psrlq $16, %%mm3 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\ | |
930 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\ | |
931 "paddb %%mm5, %%mm1 \n\t"\ | |
932 "psubusb %%mm5, %%mm1 \n\t"\ | |
933 PAVGB(%%mm7, %%mm1)\ | |
934 "pxor %%mm2, %%mm1 \n\t"\ | |
935 "psubb %%mm2, %%mm1 \n\t"\ | |
936 "psrlq $24, %%mm1 \n\t"\ | |
937 "movd %%mm1, %%ecx \n\t"\ | |
938 "paddb %%mm6, %%mm0 \n\t"\ | |
939 "paddsb (%3, %%ecx, 8), %%mm0 \n\t"\ | |
940 "paddb %%mm6, %%mm0 \n\t"\ | |
941 "movq %%mm0, " #a " \n\t"\ | |
942 | |
943 /* | |
944 HX1old((%0)) | |
945 HX1old((%%eax)) | |
946 HX1old((%%eax, %1)) | |
947 HX1old((%%eax, %1, 2)) | |
948 HX1old((%0, %1, 4)) | |
949 HX1old((%%ebx)) | |
950 HX1old((%%ebx, %1)) | |
951 HX1old((%%ebx, %1, 2)) | |
952 */ | |
953 | |
954 //FIXME add some comments, its unreadable ... | |
955 #define HX1b(a, c, b, d) \ | |
956 "movd " #a ", %%mm0 \n\t"\ | |
957 "movd 4" #a ", %%mm1 \n\t"\ | |
958 "punpckldq %%mm1, %%mm0 \n\t"\ | |
959 "movd " #b ", %%mm4 \n\t"\ | |
960 "movq %%mm0, %%mm1 \n\t"\ | |
961 "movq %%mm0, %%mm2 \n\t"\ | |
962 "psrlq $8, %%mm1 \n\t"\ | |
963 "movd 4" #b ", %%mm3 \n\t"\ | |
964 "psubusb %%mm1, %%mm2 \n\t"\ | |
965 "psubusb %%mm0, %%mm1 \n\t"\ | |
966 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ | |
967 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
968 "punpckldq %%mm3, %%mm4 \n\t"\ | |
969 "movq %%mm1, %%mm3 \n\t"\ | |
970 "psllq $32, %%mm3 \n\t" /* p´5 = |p1 - p2| */\ | |
971 PAVGB(%%mm1, %%mm3) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\ | |
972 "paddb %%mm6, %%mm0 \n\t"\ | |
973 "psrlq $16, %%mm3 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\ | |
974 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\ | |
975 "movq %%mm4, %%mm3 \n\t"\ | |
976 "paddb %%mm5, %%mm1 \n\t"\ | |
977 "psubusb %%mm5, %%mm1 \n\t"\ | |
978 "psrlq $8, %%mm3 \n\t"\ | |
979 PAVGB(%%mm7, %%mm1)\ | |
980 "pxor %%mm2, %%mm1 \n\t"\ | |
981 "psubb %%mm2, %%mm1 \n\t"\ | |
982 "movq %%mm4, %%mm2 \n\t"\ | |
983 "psrlq $24, %%mm1 \n\t"\ | |
984 "psubusb %%mm3, %%mm2 \n\t"\ | |
985 "movd %%mm1, %%ecx \n\t"\ | |
986 "psubusb %%mm4, %%mm3 \n\t"\ | |
2180 | 987 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\ |
2179 | 988 "por %%mm2, %%mm3 \n\t" /* p´x = |px - p(x+1)| */\ |
989 "paddb %%mm6, %%mm0 \n\t"\ | |
990 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
991 "movq %%mm3, %%mm1 \n\t"\ | |
992 "psllq $32, %%mm1 \n\t" /* p´5 = |p1 - p2| */\ | |
993 "movq %%mm0, " #a " \n\t"\ | |
994 PAVGB(%%mm3, %%mm1) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\ | |
995 "paddb %%mm6, %%mm4 \n\t"\ | |
996 "psrlq $16, %%mm1 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\ | |
997 "psubusb %%mm1, %%mm3 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\ | |
998 "paddb %%mm5, %%mm3 \n\t"\ | |
999 "psubusb %%mm5, %%mm3 \n\t"\ | |
1000 PAVGB(%%mm7, %%mm3)\ | |
1001 "pxor %%mm2, %%mm3 \n\t"\ | |
1002 "psubb %%mm2, %%mm3 \n\t"\ | |
1003 "psrlq $24, %%mm3 \n\t"\ | |
1004 "movd " #c ", %%mm0 \n\t"\ | |
1005 "movd 4" #c ", %%mm1 \n\t"\ | |
1006 "punpckldq %%mm1, %%mm0 \n\t"\ | |
1007 "paddb %%mm6, %%mm0 \n\t"\ | |
2180 | 1008 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\ |
2179 | 1009 "paddb %%mm6, %%mm0 \n\t"\ |
1010 "movq %%mm0, " #c " \n\t"\ | |
1011 "movd %%mm3, %%ecx \n\t"\ | |
1012 "movd " #d ", %%mm0 \n\t"\ | |
2180 | 1013 "paddsb (%2, %%ecx, 8), %%mm4 \n\t"\ |
2179 | 1014 "movd 4" #d ", %%mm1 \n\t"\ |
1015 "paddb %%mm6, %%mm4 \n\t"\ | |
1016 "punpckldq %%mm1, %%mm0 \n\t"\ | |
1017 "movq %%mm4, " #b " \n\t"\ | |
1018 "paddb %%mm6, %%mm0 \n\t"\ | |
2180 | 1019 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\ |
2179 | 1020 "paddb %%mm6, %%mm0 \n\t"\ |
1021 "movq %%mm0, " #d " \n\t"\ | |
1022 | |
1023 HX1b((%0),(%%eax),(%%eax, %1),(%%eax, %1, 2)) | |
1024 HX1b((%0, %1, 4),(%%ebx),(%%ebx, %1),(%%ebx, %1, 2)) | |
1025 | |
1026 | |
1027 : | |
2180 | 1028 : "r" (src), "r" (stride), "r" (lut) |
2179 | 1029 : "%eax", "%ebx", "%ecx" |
1030 ); | |
1031 #else | |
1032 | |
1033 //FIXME (has little in common with the mmx2 version) | |
1034 for(y=0; y<BLOCK_SIZE; y++) | |
1035 { | |
1036 int a= src[1] - src[2]; | |
1037 int b= src[3] - src[4]; | |
1038 int c= src[5] - src[6]; | |
1039 | |
1040 int d= MAX(ABS(b) - (ABS(a) + ABS(c))/2, 0); | |
1041 | |
1042 if(d < QP) | |
1043 { | |
1044 int v = d * SIGN(-b); | |
1045 | |
1046 src[1] +=v/8; | |
1047 src[2] +=v/4; | |
1048 src[3] +=3*v/8; | |
1049 src[4] -=3*v/8; | |
1050 src[5] -=v/4; | |
1051 src[6] -=v/8; | |
1052 | |
1053 } | |
1054 src+=stride; | |
1055 } | |
1056 #endif | |
1057 } | |
1058 | |
2159 | 1059 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1060 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
|
1061 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1062 #ifdef HAVE_MMX |
2246 | 1063 src+= stride*4; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1064 //FIXME try pmul for *5 stuff |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1065 // src[0]=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1066 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1067 "pxor %%mm7, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1068 "leal (%0, %1), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1069 "leal (%%eax, %1, 4), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1070 // 0 1 2 3 4 5 6 7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1071 // %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
|
1072 // %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
|
1073 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1074 "movq (%0), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1075 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1076 "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1077 "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1078 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1079 "movq (%%eax), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1080 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1081 "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1082 "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1083 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1084 "movq (%%eax, %1), %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1085 "movq %%mm4, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1086 "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1087 "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1088 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1089 "paddw %%mm0, %%mm0 \n\t" // 2L0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1090 "paddw %%mm1, %%mm1 \n\t" // 2H0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1091 "psubw %%mm4, %%mm2 \n\t" // L1 - L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1092 "psubw %%mm5, %%mm3 \n\t" // H1 - H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1093 "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1094 "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1095 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1096 "psllw $2, %%mm2 \n\t" // 4L1 - 4L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1097 "psllw $2, %%mm3 \n\t" // 4H1 - 4H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1098 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1099 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1100 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1101 "movq (%%eax, %1, 2), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1102 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1103 "punpcklbw %%mm7, %%mm2 \n\t" // L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1104 "punpckhbw %%mm7, %%mm3 \n\t" // H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1105 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1106 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1107 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1108 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1109 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1110 "movq %%mm0, temp0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1111 "movq %%mm1, temp1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1112 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1113 "movq (%0, %1, 4), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1114 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1115 "punpcklbw %%mm7, %%mm0 \n\t" // L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1116 "punpckhbw %%mm7, %%mm1 \n\t" // H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1117 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1118 "psubw %%mm0, %%mm2 \n\t" // L3 - L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1119 "psubw %%mm1, %%mm3 \n\t" // H3 - H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1120 "movq %%mm2, temp2 \n\t" // L3 - L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1121 "movq %%mm3, temp3 \n\t" // H3 - H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1122 "paddw %%mm4, %%mm4 \n\t" // 2L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1123 "paddw %%mm5, %%mm5 \n\t" // 2H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1124 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1125 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1126 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1127 "psllw $2, %%mm2 \n\t" // 4L3 - 4L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1128 "psllw $2, %%mm3 \n\t" // 4H3 - 4H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1129 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1130 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1131 //50 opcodes so far |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1132 "movq (%%ebx), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1133 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1134 "punpcklbw %%mm7, %%mm2 \n\t" // L5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1135 "punpckhbw %%mm7, %%mm3 \n\t" // H5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1136 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1137 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1138 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1139 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1140 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1141 "movq (%%ebx, %1), %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1142 "punpcklbw %%mm7, %%mm6 \n\t" // L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1143 "psubw %%mm6, %%mm2 \n\t" // L5 - L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1144 "movq (%%ebx, %1), %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1145 "punpckhbw %%mm7, %%mm6 \n\t" // H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1146 "psubw %%mm6, %%mm3 \n\t" // H5 - H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1147 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1148 "paddw %%mm0, %%mm0 \n\t" // 2L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1149 "paddw %%mm1, %%mm1 \n\t" // 2H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1150 "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1151 "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1152 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1153 "psllw $2, %%mm2 \n\t" // 4L5 - 4L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1154 "psllw $2, %%mm3 \n\t" // 4H5 - 4H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1155 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1156 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1157 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1158 "movq (%%ebx, %1, 2), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1159 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1160 "punpcklbw %%mm7, %%mm2 \n\t" // L7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1161 "punpckhbw %%mm7, %%mm3 \n\t" // H7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1162 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1163 "paddw %%mm2, %%mm2 \n\t" // 2L7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1164 "paddw %%mm3, %%mm3 \n\t" // 2H7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1165 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1166 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1167 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1168 "movq temp0, %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1169 "movq temp1, %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1170 //FIXME pxor, psubw, pmax for abs |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1171 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1172 "pcmpgtw %%mm0, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1173 "pxor %%mm6, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1174 "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1175 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1176 "pcmpgtw %%mm1, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1177 "pxor %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1178 "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1179 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1180 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1181 "pcmpgtw %%mm2, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1182 "pxor %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1183 "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1184 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1185 "pcmpgtw %%mm3, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1186 "pxor %%mm6, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1187 "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1188 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1189 #ifdef HAVE_MMX2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1190 "pminsw %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1191 "pminsw %%mm3, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1192 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1193 "movq %%mm0, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1194 "psubusw %%mm2, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1195 "psubw %%mm6, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1196 "movq %%mm1, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1197 "psubusw %%mm3, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1198 "psubw %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1199 #endif |
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 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1202 "pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1203 "pxor %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1204 "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1205 "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1206 "pxor %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1207 "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1208 // 100 opcodes |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1209 "movd %2, %%mm2 \n\t" // QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1210 "punpcklwd %%mm2, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1211 "punpcklwd %%mm2, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1212 "psllw $3, %%mm2 \n\t" // 8QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1213 "movq %%mm2, %%mm3 \n\t" // 8QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1214 "pcmpgtw %%mm4, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1215 "pcmpgtw %%mm5, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1216 "pand %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1217 "pand %%mm3, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1218 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1219 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1220 "psubusw %%mm0, %%mm4 \n\t" // hd |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1221 "psubusw %%mm1, %%mm5 \n\t" // ld |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1222 |
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 "movq w05, %%mm2 \n\t" // 5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1225 "pmullw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1226 "pmullw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1227 "movq w20, %%mm2 \n\t" // 32 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1228 "paddw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1229 "paddw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1230 "psrlw $6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1231 "psrlw $6, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1232 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1233 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1234 "movq w06, %%mm2 \n\t" // 6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1235 "paddw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1236 "paddw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1237 "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
|
1238 //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
|
1239 "pmulhw %%mm2, %%mm4 \n\t" // hd/13 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1240 "pmulhw %%mm2, %%mm5 \n\t" // ld/13 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1241 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1242 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1243 "movq temp2, %%mm0 \n\t" // L3 - L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1244 "movq temp3, %%mm1 \n\t" // H3 - H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1245 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1246 "pxor %%mm2, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1247 "pxor %%mm3, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1248 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1249 "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1250 "pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1251 "pxor %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1252 "pxor %%mm3, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1253 "psubw %%mm2, %%mm0 \n\t" // |L3-L4| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1254 "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
|
1255 "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
|
1256 "psrlw $1, %%mm1 \n\t" // |H3 - H4|/2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1257 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1258 "pxor %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1259 "pxor %%mm7, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1260 "pand %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1261 "pand %%mm3, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1262 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1263 #ifdef HAVE_MMX2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1264 "pminsw %%mm0, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1265 "pminsw %%mm1, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1266 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1267 "movq %%mm4, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1268 "psubusw %%mm0, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1269 "psubw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1270 "movq %%mm5, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1271 "psubusw %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1272 "psubw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1273 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1274 "pxor %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1275 "pxor %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1276 "psubw %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1277 "psubw %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1278 "packsswb %%mm5, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1279 "movq (%%eax, %1, 2), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1280 "paddb %%mm4, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1281 "movq %%mm0, (%%eax, %1, 2) \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1282 "movq (%0, %1, 4), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1283 "psubb %%mm4, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1284 "movq %%mm0, (%0, %1, 4) \n\t" |
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 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1287 : "r" (src), "r" (stride), "r" (QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1288 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1289 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1290 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1291 const int l1= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1292 const int l2= stride + l1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1293 const int l3= stride + l2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1294 const int l4= stride + l3; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1295 const int l5= stride + l4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1296 const int l6= stride + l5; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1297 const int l7= stride + l6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1298 const int l8= stride + l7; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1299 // const int l9= stride + l8; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1300 int x; |
2246 | 1301 src+= stride*3; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1302 for(x=0; x<BLOCK_SIZE; x++) |
2158
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 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
|
1305 if(ABS(middleEnergy) < 8*QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1306 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1307 const int q=(src[l4] - src[l5])/2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1308 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
|
1309 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
|
1310 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1311 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1312 d= MAX(d, 0); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1313 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1314 d= (5*d + 32) >> 6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1315 d*= SIGN(-middleEnergy); |
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 if(q>0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1318 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1319 d= d<0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1320 d= d>q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1321 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1322 else |
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 d= d>0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1325 d= d<q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1326 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1327 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1328 src[l4]-= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1329 src[l5]+= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1330 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1331 src++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1332 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1333 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1334 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1335 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1336 //FIXME? |255-0| = 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1337 /** |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1338 * Check if the given 8x8 Block is mostly "flat" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1339 */ |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1340 static inline int isHorizDC(uint8_t src[], int stride) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1341 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1342 // src++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1343 int numEq= 0; |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1344 #if 0 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1345 asm volatile ( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1346 // "int $3 \n\t" |
2416 | 1347 "leal (%1, %2), %%ecx \n\t" |
1348 "leal (%%ecx, %2, 4), %%ebx \n\t" | |
1349 // 0 1 2 3 4 5 6 7 8 9 | |
1350 // %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
|
1351 "movq b7E, %%mm7 \n\t" // mm7 = 0x7F |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1352 "movq b7C, %%mm6 \n\t" // mm6 = 0x7D |
2416 | 1353 "pxor %%mm0, %%mm0 \n\t" |
1354 "movl %1, %%eax \n\t" | |
1355 "andl $0x1F, %%eax \n\t" | |
1356 "cmpl $24, %%eax \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1357 "leal tempBlock, %%eax \n\t" |
2416 | 1358 "jb 1f \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1359 |
2416 | 1360 #define HDC_CHECK_AND_CPY(src, dst) \ |
1361 "movd " #src ", %%mm2 \n\t"\ | |
1362 "punpckldq 4" #src ", %%mm2 \n\t" /* (%1) */\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1363 "movq %%mm2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1364 "psrlq $8, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1365 "psubb %%mm1, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1366 "paddb %%mm7, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1367 "pcmpgtb %%mm6, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1368 "paddb %%mm2, %%mm0 \n\t"\ |
2416 | 1369 "movq %%mm1," #dst "(%%eax) \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1370 |
2416 | 1371 HDC_CHECK_AND_CPY((%1),0) |
1372 HDC_CHECK_AND_CPY((%%ecx),8) | |
1373 HDC_CHECK_AND_CPY((%%ecx, %2),16) | |
1374 HDC_CHECK_AND_CPY((%%ecx, %2, 2),24) | |
1375 HDC_CHECK_AND_CPY((%1, %2, 4),32) | |
1376 HDC_CHECK_AND_CPY((%%ebx),40) | |
1377 HDC_CHECK_AND_CPY((%%ebx, %2),48) | |
1378 HDC_CHECK_AND_CPY((%%ebx, %2, 2),56) | |
1379 "jmp 2f \n\t" | |
1380 "1: \n\t" | |
1381 // src does not cross a 32 byte cache line so dont waste time with alignment | |
1382 #define HDC_CHECK_AND_CPY2(src, dst) \ | |
1383 "movq " #src ", %%mm2 \n\t"\ | |
1384 "movq " #src ", %%mm1 \n\t"\ | |
1385 "psrlq $8, %%mm2 \n\t"\ | |
1386 "psubb %%mm1, %%mm2 \n\t"\ | |
1387 "paddb %%mm7, %%mm2 \n\t"\ | |
1388 "pcmpgtb %%mm6, %%mm2 \n\t"\ | |
1389 "paddb %%mm2, %%mm0 \n\t"\ | |
1390 "movq %%mm1," #dst "(%%eax) \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1391 |
2416 | 1392 HDC_CHECK_AND_CPY2((%1),0) |
1393 HDC_CHECK_AND_CPY2((%%ecx),8) | |
1394 HDC_CHECK_AND_CPY2((%%ecx, %2),16) | |
1395 HDC_CHECK_AND_CPY2((%%ecx, %2, 2),24) | |
1396 HDC_CHECK_AND_CPY2((%1, %2, 4),32) | |
1397 HDC_CHECK_AND_CPY2((%%ebx),40) | |
1398 HDC_CHECK_AND_CPY2((%%ebx, %2),48) | |
1399 HDC_CHECK_AND_CPY2((%%ebx, %2, 2),56) | |
1400 "2: \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1401 "psllq $8, %%mm0 \n\t" // remove dummy value |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1402 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1403 "psrlw $8, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1404 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1405 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1406 "psrlq $16, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1407 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1408 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1409 "psrlq $32, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1410 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1411 "movd %%mm0, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1412 : "=r" (numEq) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1413 : "r" (src), "r" (stride) |
2416 | 1414 : "%eax", "%ebx", "%ecx" |
2158
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("%d\n", numEq); |
2416 | 1417 numEq= (256 - numEq) &0xFF; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1418 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1419 int y; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1420 for(y=0; y<BLOCK_SIZE; y++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1421 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1422 if(((src[0] - src[1] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1423 if(((src[1] - src[2] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1424 if(((src[2] - src[3] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1425 if(((src[3] - src[4] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1426 if(((src[4] - src[5] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1427 if(((src[5] - src[6] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1428 if(((src[6] - src[7] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1429 src+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1430 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1431 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1432 /* if(abs(numEq - asmEq) > 0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1433 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1434 // printf("\nasm:%d c:%d\n", asmEq, numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1435 for(int y=0; y<8; y++) |
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 for(int x=0; x<8; x++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1438 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1439 printf("%d ", src[x + y*stride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1440 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1441 printf("\n"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1442 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1443 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1444 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1445 // printf("%d\n", numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1446 return numEq > hFlatnessThreshold; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1447 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1448 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1449 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
|
1450 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1451 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
|
1452 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1453 return 1; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1454 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1455 |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1456 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
|
1457 { |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1458 #if 0 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1459 asm volatile( |
2428 | 1460 "leal (%0, %1), %%ecx \n\t" |
1461 "leal (%%ecx, %1, 4), %%ebx \n\t" | |
1462 // 0 1 2 3 4 5 6 7 8 9 | |
1463 // %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
|
1464 "pxor %%mm7, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1465 "movq bm00001000, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1466 "movd %2, %%mm5 \n\t" // QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1467 "movq %%mm5, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1468 "paddusb %%mm5, %%mm5 \n\t" // 2QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1469 "paddusb %%mm5, %%mm4 \n\t" // 3QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1470 "psllq $24, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1471 "pxor %%mm5, %%mm5 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1472 "psubb %%mm4, %%mm5 \n\t" // -QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1473 "leal tempBlock, %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1474 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1475 //FIXME? "unroll by 2" and mix |
2159 | 1476 #ifdef HAVE_MMX2 |
2428 | 1477 #define HDF(src, dst) \ |
1478 "movq " #src "(%%eax), %%mm0 \n\t"\ | |
1479 "movq " #src "(%%eax), %%mm1 \n\t"\ | |
1480 "movq " #src "(%%eax), %%mm2 \n\t"\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1481 "psrlq $8, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1482 "psubusb %%mm1, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1483 "psubusb %%mm0, %%mm1 \n\t"\ |
2159 | 1484 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ |
1485 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
1486 "pshufw $0x00, %%mm1, %%mm3 \n\t" /* p´5 = |p1 - p2| */\ | |
1487 "pminub %%mm1, %%mm3 \n\t" /* p´5 = min(|p2-p1|, |p6-p5|)*/\ | |
1488 "psrlq $16, %%mm3 \n\t" /* p´3 = min(|p2-p1|, |p6-p5|)*/\ | |
1489 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-min(|p1-p2|,|p5-p6|) */\ | |
1490 "paddb %%mm5, %%mm1 \n\t"\ | |
1491 "psubusb %%mm5, %%mm1 \n\t"\ | |
1492 "psrlw $2, %%mm1 \n\t"\ | |
1493 "pxor %%mm2, %%mm1 \n\t"\ | |
1494 "psubb %%mm2, %%mm1 \n\t"\ | |
1495 "pand %%mm6, %%mm1 \n\t"\ | |
1496 "psubb %%mm1, %%mm0 \n\t"\ | |
1497 "psllq $8, %%mm1 \n\t"\ | |
1498 "paddb %%mm1, %%mm0 \n\t"\ | |
2428 | 1499 "movd %%mm0, " #dst" \n\t"\ |
2159 | 1500 "psrlq $32, %%mm0 \n\t"\ |
2428 | 1501 "movd %%mm0, 4" #dst" \n\t" |
2159 | 1502 #else |
2428 | 1503 #define HDF(src, dst)\ |
1504 "movq " #src "(%%eax), %%mm0 \n\t"\ | |
2159 | 1505 "movq %%mm0, %%mm1 \n\t"\ |
1506 "movq %%mm0, %%mm2 \n\t"\ | |
1507 "psrlq $8, %%mm1 \n\t"\ | |
1508 "psubusb %%mm1, %%mm2 \n\t"\ | |
1509 "psubusb %%mm0, %%mm1 \n\t"\ | |
1510 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ | |
1511 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
1512 "movq %%mm1, %%mm3 \n\t"\ | |
1513 "psllq $32, %%mm3 \n\t"\ | |
1514 "movq %%mm3, %%mm4 \n\t"\ | |
1515 "psubusb %%mm1, %%mm4 \n\t"\ | |
1516 "psubb %%mm4, %%mm3 \n\t"\ | |
1517 "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
|
1518 "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
|
1519 "paddb %%mm5, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1520 "psubusb %%mm5, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1521 "psrlw $2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1522 "pxor %%mm2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1523 "psubb %%mm2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1524 "pand %%mm6, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1525 "psubb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1526 "psllq $8, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1527 "paddb %%mm1, %%mm0 \n\t"\ |
2428 | 1528 "movd %%mm0, " #dst " \n\t"\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1529 "psrlq $32, %%mm0 \n\t"\ |
2428 | 1530 "movd %%mm0, 4" #dst " \n\t" |
2159 | 1531 #endif |
2428 | 1532 HDF(0,(%0)) |
1533 HDF(8,(%%ecx)) | |
1534 HDF(16,(%%ecx, %1)) | |
1535 HDF(24,(%%ecx, %1, 2)) | |
1536 HDF(32,(%0, %1, 4)) | |
1537 HDF(40,(%%ebx)) | |
1538 HDF(48,(%%ebx, %1)) | |
1539 HDF(56,(%%ebx, %1, 2)) | |
2158
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 : "r" (dst), "r" (stride), "r" (QP) |
2428 | 1542 : "%eax", "%ebx", "%ecx" |
2158
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 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1545 int y; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1546 for(y=0; y<BLOCK_SIZE; y++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1547 { |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1548 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
|
1549 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1550 if(ABS(middleEnergy) < 8*QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1551 { |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1552 const int q=(dst[3] - dst[4])/2; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1553 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
|
1554 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
|
1555 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1556 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1557 d= MAX(d, 0); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1558 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1559 d= (5*d + 32) >> 6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1560 d*= SIGN(-middleEnergy); |
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 if(q>0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1563 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1564 d= d<0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1565 d= d>q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1566 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1567 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1568 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1569 d= d>0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1570 d= d<q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1571 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1572 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1573 dst[3]-= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1574 dst[4]+= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1575 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1576 dst+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1577 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1578 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1579 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1580 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1581 /** |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1582 * Do a horizontal low pass filter on the 10x8 block (dst points to middle 8x8 Block) |
2221 | 1583 * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 (C version) |
1584 * 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
|
1585 */ |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1586 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
|
1587 { |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1588 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1589 #if 0 |
2428 | 1590 asm volatile( |
1591 "leal (%0, %1), %%ecx \n\t" | |
1592 "leal (%%ecx, %1, 4), %%ebx \n\t" | |
1593 // 0 1 2 3 4 5 6 7 8 9 | |
1594 // %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
|
1595 "pxor %%mm7, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1596 "leal tempBlock, %%eax \n\t" |
2159 | 1597 /* |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1598 #define HLP1 "movq (%0), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1599 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1600 "psllq $8, %%mm0 \n\t"\ |
2159 | 1601 PAVGB(%%mm1, %%mm0)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1602 "psrlw $8, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1603 "pxor %%mm1, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1604 "packuswb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1605 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1606 "movq %%mm0, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1607 "psllq $32, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1608 "paddb %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1609 "psllq $16, %%mm2 \n\t"\ |
2159 | 1610 PAVGB(%%mm2, %%mm0)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1611 "movq %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1612 "pand bm11001100, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1613 "paddusb %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1614 "psrlq $8, %%mm3 \n\t"\ |
2159 | 1615 PAVGB(%%mm1, %%mm4)\ |
1616 PAVGB(%%mm3, %%mm2)\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1617 "psrlq $16, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1618 "punpcklbw %%mm2, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1619 "movq %%mm2, (%0) \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1620 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1621 #define HLP2 "movq (%0), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1622 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1623 "psllq $8, %%mm0 \n\t"\ |
2159 | 1624 PAVGB(%%mm1, %%mm0)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1625 "psrlw $8, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1626 "pxor %%mm1, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1627 "packuswb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1628 "movq %%mm0, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1629 "psllq $32, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1630 "psllq $16, %%mm2 \n\t"\ |
2159 | 1631 PAVGB(%%mm2, %%mm0)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1632 "movq %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1633 "pand bm11001100, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1634 "paddusb %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1635 "psrlq $8, %%mm3 \n\t"\ |
2159 | 1636 PAVGB(%%mm3, %%mm2)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1637 "psrlq $16, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1638 "punpcklbw %%mm2, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1639 "movq %%mm2, (%0) \n\t"\ |
2159 | 1640 */ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1641 // approximately a 7-Tap Filter with Vector (1,2,3,4,3,2,1)/16 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1642 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1643 Implemented Exact 7-Tap |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1644 9421 A321 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1645 36421 64321 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1646 334321 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1647 1234321 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1648 1234321 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1649 123433 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1650 12463 12346 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1651 1249 123A |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1652 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1653 */ |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1654 |
2159 | 1655 #ifdef HAVE_MMX2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1656 #define HLP3(i) "movq " #i "(%%eax), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1657 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1658 "movq %%mm0, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1659 "movq %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1660 "movq %%mm0, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1661 "psllq $8, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1662 "psrlq $8, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1663 "pand bm00000001, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1664 "pand bm10000000, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1665 "por %%mm3, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1666 "por %%mm4, %%mm2 \n\t"\ |
2159 | 1667 PAVGB(%%mm2, %%mm1)\ |
1668 PAVGB(%%mm1, %%mm0)\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1669 \ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1670 "pshufw $0xF9, %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1671 "pshufw $0x90, %%mm0, %%mm4 \n\t"\ |
2159 | 1672 PAVGB(%%mm3, %%mm4)\ |
1673 PAVGB(%%mm4, %%mm0)\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1674 "movd %%mm0, (%0) \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1675 "psrlq $32, %%mm0 \n\t"\ |
2159 | 1676 "movd %%mm0, 4(%0) \n\t" |
1677 #else | |
1678 #define HLP3(i) "movq " #i "(%%eax), %%mm0 \n\t"\ | |
1679 "movq %%mm0, %%mm1 \n\t"\ | |
1680 "movq %%mm0, %%mm2 \n\t"\ | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1681 "movd -4(%0), %%mm3 \n\t" /*0001000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1682 "movd 8(%0), %%mm4 \n\t" /*0001000*/\ |
2159 | 1683 "psllq $8, %%mm1 \n\t"\ |
1684 "psrlq $8, %%mm2 \n\t"\ | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1685 "psrlq $24, %%mm3 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1686 "psllq $56, %%mm4 \n\t"\ |
2159 | 1687 "por %%mm3, %%mm1 \n\t"\ |
1688 "por %%mm4, %%mm2 \n\t"\ | |
1689 PAVGB(%%mm2, %%mm1)\ | |
1690 PAVGB(%%mm1, %%mm0)\ | |
1691 \ | |
1692 "movq %%mm0, %%mm3 \n\t"\ | |
1693 "movq %%mm0, %%mm4 \n\t"\ | |
1694 "movq %%mm0, %%mm5 \n\t"\ | |
1695 "psrlq $16, %%mm3 \n\t"\ | |
1696 "psllq $16, %%mm4 \n\t"\ | |
1697 "pand bm11000000, %%mm5 \n\t"\ | |
1698 "por %%mm5, %%mm3 \n\t"\ | |
1699 "movq %%mm0, %%mm5 \n\t"\ | |
1700 "pand bm00000011, %%mm5 \n\t"\ | |
1701 "por %%mm5, %%mm4 \n\t"\ | |
1702 PAVGB(%%mm3, %%mm4)\ | |
1703 PAVGB(%%mm4, %%mm0)\ | |
1704 "movd %%mm0, (%0) \n\t"\ | |
1705 "psrlq $32, %%mm0 \n\t"\ | |
1706 "movd %%mm0, 4(%0) \n\t" | |
1707 #endif | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1708 |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1709 /* uses the 7-Tap Filter: 1112111 */ |
2428 | 1710 #define NEW_HLP(src, dst)\ |
1711 "movq " #src "(%%eax), %%mm1 \n\t"\ | |
1712 "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
|
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"\ |
2428 | 1715 "movd -4" #dst ", %%mm3 \n\t" /*0001000*/\ |
1716 "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
|
1717 "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
|
1718 "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
|
1719 "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
|
1720 "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
|
1721 "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
|
1722 PAVGB(%%mm2, %%mm1)\ |
2428 | 1723 "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
|
1724 PAVGB(%%mm1, %%mm0)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1725 "psllq $8, %%mm5 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1726 "psrlq $8, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1727 "por %%mm3, %%mm5 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1728 "por %%mm4, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1729 "movq %%mm5, %%mm1 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1730 PAVGB(%%mm2, %%mm5)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1731 "psllq $8, %%mm1 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1732 "psrlq $8, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1733 "por %%mm3, %%mm1 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1734 "por %%mm4, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1735 PAVGB(%%mm2, %%mm1)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1736 PAVGB(%%mm1, %%mm5)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1737 PAVGB(%%mm5, %%mm0)\ |
2428 | 1738 "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
|
1739 "psrlq $32, %%mm0 \n\t"\ |
2428 | 1740 "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
|
1741 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1742 /* uses the 9-Tap Filter: 112242211 */ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1743 #define NEW_HLP2(i)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1744 "movq " #i "(%%eax), %%mm0 \n\t" /*0001000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1745 "movq %%mm0, %%mm1 \n\t" /*0001000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1746 "movq %%mm0, %%mm2 \n\t" /*0001000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1747 "movd -4(%0), %%mm3 \n\t" /*0001000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1748 "movd 8(%0), %%mm4 \n\t" /*0001000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1749 "psllq $8, %%mm1 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1750 "psrlq $8, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1751 "psrlq $24, %%mm3 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1752 "psllq $56, %%mm4 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1753 "por %%mm3, %%mm1 \n\t" /*0010000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1754 "por %%mm4, %%mm2 \n\t" /*0000100*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1755 "movq %%mm1, %%mm5 \n\t" /*0010000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1756 PAVGB(%%mm2, %%mm1) /*0010100*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1757 PAVGB(%%mm1, %%mm0) /*0012100*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1758 "psllq $8, %%mm5 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1759 "psrlq $8, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1760 "por %%mm3, %%mm5 \n\t" /*0100000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1761 "por %%mm4, %%mm2 \n\t" /*0000010*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1762 "movq %%mm5, %%mm1 \n\t" /*0100000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1763 PAVGB(%%mm2, %%mm5) /*0100010*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1764 "psllq $8, %%mm1 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1765 "psrlq $8, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1766 "por %%mm3, %%mm1 \n\t" /*1000000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1767 "por %%mm4, %%mm2 \n\t" /*0000001*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1768 "movq %%mm1, %%mm6 \n\t" /*1000000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1769 PAVGB(%%mm2, %%mm1) /*1000001*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1770 "psllq $8, %%mm6 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1771 "psrlq $8, %%mm2 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1772 "por %%mm3, %%mm6 \n\t"/*100000000*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1773 "por %%mm4, %%mm2 \n\t"/*000000001*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1774 PAVGB(%%mm2, %%mm6) /*100000001*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1775 PAVGB(%%mm6, %%mm1) /*110000011*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1776 PAVGB(%%mm1, %%mm5) /*112000211*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1777 PAVGB(%%mm5, %%mm0) /*112242211*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1778 "movd %%mm0, (%0) \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1779 "psrlq $32, %%mm0 \n\t"\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1780 "movd %%mm0, 4(%0) \n\t" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1781 |
2428 | 1782 #define HLP(src, dst) NEW_HLP(src, dst) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1783 |
2428 | 1784 HLP(0, (%0)) |
1785 HLP(8, (%%ecx)) | |
1786 HLP(16, (%%ecx, %1)) | |
1787 HLP(24, (%%ecx, %1, 2)) | |
1788 HLP(32, (%0, %1, 4)) | |
1789 HLP(40, (%%ebx)) | |
1790 HLP(48, (%%ebx, %1)) | |
1791 HLP(56, (%%ebx, %1, 2)) | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1792 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1793 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1794 : "r" (dst), "r" (stride) |
2428 | 1795 : "%eax", "%ebx", "%ecx" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1796 ); |
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 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1799 int y; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1800 for(y=0; y<BLOCK_SIZE; y++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1801 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1802 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
|
1803 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
|
1804 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1805 int sums[9]; |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1806 sums[0] = first + dst[0]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1807 sums[1] = dst[0] + dst[1]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1808 sums[2] = dst[1] + dst[2]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1809 sums[3] = dst[2] + dst[3]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1810 sums[4] = dst[3] + dst[4]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1811 sums[5] = dst[4] + dst[5]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1812 sums[6] = dst[5] + dst[6]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1813 sums[7] = dst[6] + dst[7]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1814 sums[8] = dst[7] + last; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1815 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1816 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
|
1817 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
|
1818 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
|
1819 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
|
1820 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
|
1821 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
|
1822 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
|
1823 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
|
1824 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1825 dst+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1826 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1827 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1828 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1829 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1830 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
|
1831 { |
2475 | 1832 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1833 asm volatile( |
2473 | 1834 "movq pQPb, %%mm0 \n\t" |
1835 "paddusb %%mm0, %%mm0 \n\t" | |
1836 "movq %%mm0, pQPb2 \n\t" | |
1837 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1838 "leal (%0, %1), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1839 "leal (%%eax, %1, 4), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1840 // 0 1 2 3 4 5 6 7 8 9 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1841 // %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
|
1842 |
2473 | 1843 "pcmpeqb %%mm6, %%mm6 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1844 "pxor %%mm7, %%mm7 \n\t" |
2475 | 1845 #ifdef HAVE_MMX2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1846 #define FIND_MIN_MAX(addr)\ |
2473 | 1847 "movq " #addr ", %%mm0 \n\t"\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1848 "pminub %%mm0, %%mm6 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1849 "pmaxub %%mm0, %%mm7 \n\t" |
2475 | 1850 #else |
1851 #define FIND_MIN_MAX(addr)\ | |
1852 "movq " #addr ", %%mm0 \n\t"\ | |
1853 "movq %%mm6, %%mm1 \n\t"\ | |
1854 "psubusb %%mm0, %%mm7 \n\t"\ | |
1855 "paddb %%mm0, %%mm7 \n\t"\ | |
1856 "psubusb %%mm0, %%mm1 \n\t"\ | |
1857 "psubb %%mm1, %%mm6 \n\t" | |
1858 #endif | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1859 |
2473 | 1860 FIND_MIN_MAX((%%eax)) |
1861 FIND_MIN_MAX((%%eax, %1)) | |
1862 FIND_MIN_MAX((%%eax, %1, 2)) | |
1863 FIND_MIN_MAX((%0, %1, 4)) | |
1864 FIND_MIN_MAX((%%ebx)) | |
1865 FIND_MIN_MAX((%%ebx, %1)) | |
1866 FIND_MIN_MAX((%%ebx, %1, 2)) | |
1867 FIND_MIN_MAX((%0, %1, 8)) | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1868 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1869 "movq %%mm6, %%mm4 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1870 "psrlq $8, %%mm6 \n\t" |
2475 | 1871 #ifdef HAVE_MMX2 |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1872 "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
|
1873 "pshufw $0xF9, %%mm6, %%mm4 \n\t" |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1874 "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
|
1875 "pshufw $0xFE, %%mm6, %%mm4 \n\t" |
2475 | 1876 "pminub %%mm4, %%mm6 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1877 #else |
2475 | 1878 "movq %%mm6, %%mm1 \n\t" |
1879 "psubusb %%mm4, %%mm1 \n\t" | |
1880 "psubb %%mm1, %%mm6 \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1881 "movq %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1882 "psrlq $16, %%mm6 \n\t" |
2475 | 1883 "movq %%mm6, %%mm1 \n\t" |
1884 "psubusb %%mm4, %%mm1 \n\t" | |
1885 "psubb %%mm1, %%mm6 \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1886 "movq %%mm6, %%mm4 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1887 "psrlq $32, %%mm6 \n\t" |
2475 | 1888 "movq %%mm6, %%mm1 \n\t" |
1889 "psubusb %%mm4, %%mm1 \n\t" | |
1890 "psubb %%mm1, %%mm6 \n\t" | |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1891 #endif |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1892 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1893 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1894 "movq %%mm7, %%mm4 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1895 "psrlq $8, %%mm7 \n\t" |
2475 | 1896 #ifdef HAVE_MMX2 |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1897 "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
|
1898 "pshufw $0xF9, %%mm7, %%mm4 \n\t" |
2475 | 1899 "pmaxub %%mm4, %%mm7 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1900 "pshufw $0xFE, %%mm7, %%mm4 \n\t" |
2475 | 1901 "pmaxub %%mm4, %%mm7 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1902 #else |
2475 | 1903 "psubusb %%mm4, %%mm7 \n\t" |
1904 "paddb %%mm4, %%mm7 \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1905 "movq %%mm7, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1906 "psrlq $16, %%mm7 \n\t" |
2475 | 1907 "psubusb %%mm4, %%mm7 \n\t" |
1908 "paddb %%mm4, %%mm7 \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1909 "movq %%mm7, %%mm4 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1910 "psrlq $32, %%mm7 \n\t" |
2475 | 1911 "psubusb %%mm4, %%mm7 \n\t" |
1912 "paddb %%mm4, %%mm7 \n\t" | |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1913 #endif |
2473 | 1914 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
|
1915 "punpcklbw %%mm7, %%mm7 \n\t" |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1916 "punpcklbw %%mm7, %%mm7 \n\t" |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1917 "punpcklbw %%mm7, %%mm7 \n\t" |
2473 | 1918 "movq %%mm7, temp0 \n\t" |
1919 | |
1920 "movq (%0), %%mm0 \n\t" // L10 | |
1921 "movq %%mm0, %%mm1 \n\t" // L10 | |
1922 "movq %%mm0, %%mm2 \n\t" // L10 | |
1923 "psllq $8, %%mm1 \n\t" | |
1924 "psrlq $8, %%mm2 \n\t" | |
1925 "movd -4(%0), %%mm3 \n\t" | |
1926 "movd 8(%0), %%mm4 \n\t" | |
1927 "psrlq $24, %%mm3 \n\t" | |
1928 "psllq $56, %%mm4 \n\t" | |
1929 "por %%mm3, %%mm1 \n\t" // L00 | |
1930 "por %%mm4, %%mm2 \n\t" // L20 | |
1931 "movq %%mm1, %%mm3 \n\t" // L00 | |
1932 PAVGB(%%mm2, %%mm1) // (L20 + L00)/2 | |
1933 PAVGB(%%mm0, %%mm1) // (L20 + L00 + 2L10)/4 | |
1934 "psubusb %%mm7, %%mm0 \n\t" | |
1935 "psubusb %%mm7, %%mm2 \n\t" | |
1936 "psubusb %%mm7, %%mm3 \n\t" | |
1937 "pcmpeqb b00, %%mm0 \n\t" // L10 > a ? 0 : -1 | |
1938 "pcmpeqb b00, %%mm2 \n\t" // L20 > a ? 0 : -1 | |
1939 "pcmpeqb b00, %%mm3 \n\t" // L00 > a ? 0 : -1 | |
1940 "paddb %%mm2, %%mm0 \n\t" | |
1941 "paddb %%mm3, %%mm0 \n\t" | |
1942 | |
1943 "movq (%%eax), %%mm2 \n\t" // L11 | |
1944 "movq %%mm2, %%mm3 \n\t" // L11 | |
1945 "movq %%mm2, %%mm4 \n\t" // L11 | |
1946 "psllq $8, %%mm3 \n\t" | |
1947 "psrlq $8, %%mm4 \n\t" | |
1948 "movd -4(%%eax), %%mm5 \n\t" | |
1949 "movd 8(%%eax), %%mm6 \n\t" | |
1950 "psrlq $24, %%mm5 \n\t" | |
1951 "psllq $56, %%mm6 \n\t" | |
1952 "por %%mm5, %%mm3 \n\t" // L01 | |
1953 "por %%mm6, %%mm4 \n\t" // L21 | |
1954 "movq %%mm3, %%mm5 \n\t" // L01 | |
1955 PAVGB(%%mm4, %%mm3) // (L21 + L01)/2 | |
1956 PAVGB(%%mm2, %%mm3) // (L21 + L01 + 2L11)/4 | |
1957 "psubusb %%mm7, %%mm2 \n\t" | |
1958 "psubusb %%mm7, %%mm4 \n\t" | |
1959 "psubusb %%mm7, %%mm5 \n\t" | |
1960 "pcmpeqb b00, %%mm2 \n\t" // L11 > a ? 0 : -1 | |
1961 "pcmpeqb b00, %%mm4 \n\t" // L21 > a ? 0 : -1 | |
1962 "pcmpeqb b00, %%mm5 \n\t" // L01 > a ? 0 : -1 | |
1963 "paddb %%mm4, %%mm2 \n\t" | |
1964 "paddb %%mm5, %%mm2 \n\t" | |
1965 // 0, 2, 3, 1 | |
1966 #define DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \ | |
1967 "movq " #src ", " #sx " \n\t" /* src[0] */\ | |
1968 "movq " #sx ", " #lx " \n\t" /* src[0] */\ | |
1969 "movq " #sx ", " #t0 " \n\t" /* src[0] */\ | |
1970 "psllq $8, " #lx " \n\t"\ | |
1971 "psrlq $8, " #t0 " \n\t"\ | |
1972 "movd -4" #src ", " #t1 " \n\t"\ | |
1973 "psrlq $24, " #t1 " \n\t"\ | |
1974 "por " #t1 ", " #lx " \n\t" /* src[-1] */\ | |
1975 "movd 8" #src ", " #t1 " \n\t"\ | |
1976 "psllq $56, " #t1 " \n\t"\ | |
1977 "por " #t1 ", " #t0 " \n\t" /* src[+1] */\ | |
1978 "movq " #lx ", " #t1 " \n\t" /* src[-1] */\ | |
1979 PAVGB(t0, lx) /* (src[-1] + src[+1])/2 */\ | |
1980 PAVGB(sx, lx) /* (src[-1] + 2src[0] + src[+1])/4 */\ | |
2478 | 1981 PAVGB(lx, pplx) \ |
1982 "movq " #lx ", temp1 \n\t"\ | |
1983 "movq temp0, " #lx " \n\t"\ | |
1984 "psubusb " #lx ", " #t1 " \n\t"\ | |
1985 "psubusb " #lx ", " #t0 " \n\t"\ | |
1986 "psubusb " #lx ", " #sx " \n\t"\ | |
1987 "movq b00, " #lx " \n\t"\ | |
1988 "pcmpeqb " #lx ", " #t1 " \n\t" /* src[-1] > a ? 0 : -1*/\ | |
1989 "pcmpeqb " #lx ", " #t0 " \n\t" /* src[+1] > a ? 0 : -1*/\ | |
1990 "pcmpeqb " #lx ", " #sx " \n\t" /* src[0] > a ? 0 : -1*/\ | |
2473 | 1991 "paddb " #t1 ", " #t0 " \n\t"\ |
1992 "paddb " #t0 ", " #sx " \n\t"\ | |
1993 \ | |
1994 PAVGB(plx, pplx) /* filtered */\ | |
1995 "movq " #dst ", " #t0 " \n\t" /* dst */\ | |
2477 | 1996 "movq " #t0 ", " #t1 " \n\t" /* dst */\ |
1997 "psubusb pQPb2, " #t0 " \n\t"\ | |
1998 "paddusb pQPb2, " #t1 " \n\t"\ | |
1999 PMAXUB(t0, pplx)\ | |
2000 PMINUB(t1, pplx, t0)\ | |
2473 | 2001 "paddb " #sx ", " #ppsx " \n\t"\ |
2002 "paddb " #psx ", " #ppsx " \n\t"\ | |
2003 "#paddb b02, " #ppsx " \n\t"\ | |
2004 "pand b08, " #ppsx " \n\t"\ | |
2478 | 2005 "pcmpeqb " #lx ", " #ppsx " \n\t"\ |
2477 | 2006 "pand " #ppsx ", " #pplx " \n\t"\ |
2473 | 2007 "pandn " #dst ", " #ppsx " \n\t"\ |
2477 | 2008 "por " #pplx ", " #ppsx " \n\t"\ |
2478 | 2009 "movq " #ppsx ", " #dst " \n\t"\ |
2010 "movq temp1, " #lx " \n\t" | |
2477 | 2011 |
2473 | 2012 /* |
2013 0000000 | |
2014 1111111 | |
2015 | |
2016 1111110 | |
2017 1111101 | |
2018 1111100 | |
2019 1111011 | |
2020 1111010 | |
2021 1111001 | |
2022 | |
2023 1111000 | |
2024 1110111 | |
2025 | |
2026 */ | |
2027 //DERING_CORE(dst,src ,ppsx ,psx ,sx ,pplx ,plx ,lx ,t0 ,t1) | |
2028 DERING_CORE((%%eax),(%%eax, %1) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7) | |
2029 DERING_CORE((%%eax, %1),(%%eax, %1, 2) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7) | |
2030 DERING_CORE((%%eax, %1, 2),(%0, %1, 4) ,%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7) | |
2031 DERING_CORE((%0, %1, 4),(%%ebx) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7) | |
2032 DERING_CORE((%%ebx),(%%ebx, %1) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7) | |
2033 DERING_CORE((%%ebx, %1), (%%ebx, %1, 2),%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7) | |
2034 DERING_CORE((%%ebx, %1, 2),(%0, %1, 8) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7) | |
2035 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
|
2036 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2037 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2038 : : "r" (src), "r" (stride), "r" (QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2039 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2040 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2041 #else |
2477 | 2042 int y; |
2043 int min=255; | |
2044 int max=0; | |
2045 int avg; | |
2046 uint8_t *p; | |
2047 int s[10]; | |
2048 | |
2049 for(y=1; y<9; y++) | |
2050 { | |
2051 int x; | |
2052 p= src + stride*y; | |
2053 for(x=1; x<9; x++) | |
2054 { | |
2055 p++; | |
2056 if(*p > max) max= *p; | |
2057 if(*p < min) min= *p; | |
2058 } | |
2059 } | |
2060 avg= (min + max + 1)/2; | |
2061 | |
2062 for(y=0; y<10; y++) | |
2063 { | |
2064 int x; | |
2065 int t = 0; | |
2066 p= src + stride*y; | |
2067 for(x=0; x<10; x++) | |
2068 { | |
2069 if(*p > avg) t |= (1<<x); | |
2070 p++; | |
2071 } | |
2072 t |= (~t)<<16; | |
2073 t &= (t<<1) & (t>>1); | |
2074 s[y] = t; | |
2075 } | |
2076 | |
2077 for(y=1; y<9; y++) | |
2078 { | |
2079 int x; | |
2080 int t = s[y-1] & s[y] & s[y+1]; | |
2081 t|= t>>16; | |
2082 | |
2083 p= src + stride*y; | |
2084 for(x=1; x<9; x++) | |
2085 { | |
2086 p++; | |
2087 if(t & (1<<x)) | |
2088 { | |
2089 int f= (*(p-stride-1)) + 2*(*(p-stride)) + (*(p-stride+1)) | |
2090 +2*(*(p -1)) + 4*(*p ) + 2*(*(p +1)) | |
2091 +(*(p+stride-1)) + 2*(*(p+stride)) + (*(p+stride+1)); | |
2092 f= (f + 8)>>4; | |
2093 | |
2094 if (*p + 2*QP < f) *p= *p + 2*QP; | |
2095 else if(*p - 2*QP > f) *p= *p - 2*QP; | |
2096 else *p=f; | |
2097 } | |
2098 } | |
2099 } | |
2100 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2101 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2102 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2103 |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2104 /** |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2105 * Deinterlaces the given block |
2246 | 2106 * 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
|
2107 */ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2108 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
|
2109 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2110 #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
|
2111 asm volatile( |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2112 "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
|
2113 "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
|
2114 // 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
|
2115 // %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
|
2116 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2117 "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
|
2118 "movq (%%eax, %1), %%mm1 \n\t" |
2246 | 2119 PAVGB(%%mm1, %%mm0) |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2120 "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
|
2121 "movq (%0, %1, 4), %%mm0 \n\t" |
2246 | 2122 PAVGB(%%mm0, %%mm1) |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2123 "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
|
2124 "movq (%%ebx, %1), %%mm1 \n\t" |
2246 | 2125 PAVGB(%%mm1, %%mm0) |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2126 "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
|
2127 "movq (%0, %1, 8), %%mm0 \n\t" |
2246 | 2128 PAVGB(%%mm0, %%mm1) |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2129 "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
|
2130 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2131 : : "r" (src), "r" (stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2132 : "%eax", "%ebx" |
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 #else |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2135 int x; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2136 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
|
2137 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2138 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
|
2139 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
|
2140 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
|
2141 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
|
2142 src++; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2143 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2144 #endif |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2145 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2146 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2147 /** |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2148 * Deinterlaces the given block |
2246 | 2149 * will be called for every 8x8 block, and can read & write into an 8x16 block |
2150 * 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
|
2151 */ |
2246 | 2152 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
|
2153 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2154 #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
|
2155 asm volatile( |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2156 "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
|
2157 "leal (%%eax, %1, 4), %%ebx \n\t" |
2246 | 2158 "leal (%%ebx, %1, 4), %%ecx \n\t" |
2159 "addl %1, %%ecx \n\t" | |
2160 "pxor %%mm7, %%mm7 \n\t" | |
2161 // 0 1 2 3 4 5 6 7 8 9 10 | |
2162 // %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
|
2163 |
2246 | 2164 #define DEINT_CUBIC(a,b,c,d,e)\ |
2165 "movq " #a ", %%mm0 \n\t"\ | |
2166 "movq " #b ", %%mm1 \n\t"\ | |
2167 "movq " #d ", %%mm2 \n\t"\ | |
2168 "movq " #e ", %%mm3 \n\t"\ | |
2169 PAVGB(%%mm2, %%mm1) /* (b+d) /2 */\ | |
2170 PAVGB(%%mm3, %%mm0) /* a(a+e) /2 */\ | |
2171 "movq %%mm0, %%mm2 \n\t"\ | |
2172 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
2173 "punpckhbw %%mm7, %%mm2 \n\t"\ | |
2174 "movq %%mm1, %%mm3 \n\t"\ | |
2175 "punpcklbw %%mm7, %%mm1 \n\t"\ | |
2176 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
2177 "psubw %%mm1, %%mm0 \n\t" /* L(a+e - (b+d))/2 */\ | |
2178 "psubw %%mm3, %%mm2 \n\t" /* H(a+e - (b+d))/2 */\ | |
2179 "psraw $3, %%mm0 \n\t" /* L(a+e - (b+d))/16 */\ | |
2180 "psraw $3, %%mm2 \n\t" /* H(a+e - (b+d))/16 */\ | |
2181 "psubw %%mm0, %%mm1 \n\t" /* L(9b + 9d - a - e)/16 */\ | |
2182 "psubw %%mm2, %%mm3 \n\t" /* H(9b + 9d - a - e)/16 */\ | |
2183 "packuswb %%mm3, %%mm1 \n\t"\ | |
2184 "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
|
2185 |
2246 | 2186 DEINT_CUBIC((%0), (%%eax, %1), (%%eax, %1, 2), (%0, %1, 4), (%%ebx, %1)) |
2187 DEINT_CUBIC((%%eax, %1), (%0, %1, 4), (%%ebx), (%%ebx, %1), (%0, %1, 8)) | |
2188 DEINT_CUBIC((%0, %1, 4), (%%ebx, %1), (%%ebx, %1, 2), (%0, %1, 8), (%%ecx)) | |
2189 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
|
2190 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2191 : : "r" (src), "r" (stride) |
2246 | 2192 : "%eax", "%ebx", "ecx" |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2193 ); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2194 #else |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2195 int x; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2196 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
|
2197 { |
2246 | 2198 src[stride*3] = (-src[0] + 9*src[stride*2] + 9*src[stride*4] - src[stride*6])>>4; |
2199 src[stride*5] = (-src[stride*2] + 9*src[stride*4] + 9*src[stride*6] - src[stride*8])>>4; | |
2200 src[stride*7] = (-src[stride*4] + 9*src[stride*6] + 9*src[stride*8] - src[stride*10])>>4; | |
2201 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
|
2202 src++; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2203 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2204 #endif |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2205 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2206 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2207 /** |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2208 * Deinterlaces the given block |
2246 | 2209 * 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
|
2210 * 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
|
2211 */ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2212 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
|
2213 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2214 #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
|
2215 asm volatile( |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2216 "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
|
2217 "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
|
2218 // 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
|
2219 // %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
|
2220 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2221 "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
|
2222 "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
|
2223 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
|
2224 "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
|
2225 PAVGB(%%mm2, %%mm0) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2226 "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
|
2227 "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
|
2228 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
|
2229 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
|
2230 "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
|
2231 "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
|
2232 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
|
2233 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
|
2234 "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
|
2235 "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
|
2236 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
|
2237 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
|
2238 "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
|
2239 "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
|
2240 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
|
2241 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
|
2242 "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
|
2243 "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
|
2244 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
|
2245 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
|
2246 "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
|
2247 "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
|
2248 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
|
2249 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
|
2250 "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
|
2251 "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
|
2252 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
|
2253 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
|
2254 "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
|
2255 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2256 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2257 : : "r" (src), "r" (stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2258 : "%eax", "%ebx" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2259 ); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2260 #else |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2261 int x; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2262 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
|
2263 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2264 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
|
2265 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
|
2266 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
|
2267 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
|
2268 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
|
2269 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
|
2270 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
|
2271 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
|
2272 src++; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2273 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2274 #endif |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2275 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2276 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2277 /** |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2278 * Deinterlaces the given block |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2279 * 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
|
2280 */ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2281 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
|
2282 { |
2221 | 2283 #ifdef HAVE_MMX |
2284 #ifdef HAVE_MMX2 | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2285 asm volatile( |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2286 "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
|
2287 "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
|
2288 // 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
|
2289 // %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
|
2290 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2291 "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
|
2292 "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
|
2293 "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
|
2294 "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
|
2295 "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
|
2296 "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
|
2297 "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
|
2298 "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
|
2299 "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
|
2300 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2301 "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
|
2302 "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
|
2303 "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
|
2304 "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
|
2305 "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
|
2306 "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
|
2307 "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
|
2308 "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
|
2309 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2310 "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
|
2311 "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
|
2312 "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
|
2313 "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
|
2314 "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
|
2315 "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
|
2316 "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
|
2317 "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
|
2318 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2319 "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
|
2320 "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
|
2321 "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
|
2322 "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
|
2323 "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
|
2324 "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
|
2325 "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
|
2326 "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
|
2327 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2328 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2329 : : "r" (src), "r" (stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2330 : "%eax", "%ebx" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2331 ); |
2221 | 2332 |
2333 #else // MMX without MMX2 | |
2334 asm volatile( | |
2335 "leal (%0, %1), %%eax \n\t" | |
2336 "leal (%%eax, %1, 4), %%ebx \n\t" | |
2337 // 0 1 2 3 4 5 6 7 8 9 | |
2338 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 | |
2339 "pxor %%mm7, %%mm7 \n\t" | |
2340 | |
2341 #define MEDIAN(a,b,c)\ | |
2342 "movq " #a ", %%mm0 \n\t"\ | |
2343 "movq " #b ", %%mm2 \n\t"\ | |
2344 "movq " #c ", %%mm1 \n\t"\ | |
2345 "movq %%mm0, %%mm3 \n\t"\ | |
2346 "movq %%mm1, %%mm4 \n\t"\ | |
2347 "movq %%mm2, %%mm5 \n\t"\ | |
2348 "psubusb %%mm1, %%mm3 \n\t"\ | |
2349 "psubusb %%mm2, %%mm4 \n\t"\ | |
2350 "psubusb %%mm0, %%mm5 \n\t"\ | |
2351 "pcmpeqb %%mm7, %%mm3 \n\t"\ | |
2352 "pcmpeqb %%mm7, %%mm4 \n\t"\ | |
2353 "pcmpeqb %%mm7, %%mm5 \n\t"\ | |
2354 "movq %%mm3, %%mm6 \n\t"\ | |
2355 "pxor %%mm4, %%mm3 \n\t"\ | |
2356 "pxor %%mm5, %%mm4 \n\t"\ | |
2357 "pxor %%mm6, %%mm5 \n\t"\ | |
2358 "por %%mm3, %%mm1 \n\t"\ | |
2359 "por %%mm4, %%mm2 \n\t"\ | |
2360 "por %%mm5, %%mm0 \n\t"\ | |
2361 "pand %%mm2, %%mm0 \n\t"\ | |
2362 "pand %%mm1, %%mm0 \n\t"\ | |
2363 "movq %%mm0, " #b " \n\t" | |
2364 | |
2365 MEDIAN((%0), (%%eax), (%%eax, %1)) | |
2366 MEDIAN((%%eax, %1), (%%eax, %1, 2), (%0, %1, 4)) | |
2367 MEDIAN((%0, %1, 4), (%%ebx), (%%ebx, %1)) | |
2368 MEDIAN((%%ebx, %1), (%%ebx, %1, 2), (%0, %1, 8)) | |
2369 | |
2370 : : "r" (src), "r" (stride) | |
2371 : "%eax", "%ebx" | |
2372 ); | |
2373 #endif // MMX | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2374 #else |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2375 //FIXME |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2376 int x; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2377 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
|
2378 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2379 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
|
2380 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
|
2381 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
|
2382 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
|
2383 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
|
2384 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
|
2385 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
|
2386 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
|
2387 src++; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2388 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2389 #endif |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2390 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2391 |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2392 #ifdef HAVE_MMX |
2454
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 * 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
|
2395 */ |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2396 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
|
2397 { |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2398 asm( |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2399 "leal (%0, %1), %%eax \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2400 "leal (%%eax, %1, 4), %%ebx \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2401 // 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
|
2402 // %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
|
2403 "movq (%0), %%mm0 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2404 "movq (%%eax), %%mm1 \n\t" // abcdefgh |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2405 "movq %%mm0, %%mm2 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2406 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2407 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2408 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2409 "movq (%%eax, %1), %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2410 "movq (%%eax, %1, 2), %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2411 "movq %%mm1, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2412 "punpcklbw %%mm3, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2413 "punpckhbw %%mm3, %%mm4 \n\t" |
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 %%mm0, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2416 "punpcklwd %%mm1, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2417 "punpckhwd %%mm1, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2418 "movq %%mm2, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2419 "punpcklwd %%mm4, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2420 "punpckhwd %%mm4, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2421 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2422 "movd %%mm0, 128(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2423 "psrlq $32, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2424 "movd %%mm0, 144(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2425 "movd %%mm3, 160(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2426 "psrlq $32, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2427 "movd %%mm3, 176(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2428 "movd %%mm3, 48(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2429 "movd %%mm2, 192(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2430 "movd %%mm2, 64(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2431 "psrlq $32, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2432 "movd %%mm2, 80(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2433 "movd %%mm1, 96(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2434 "psrlq $32, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2435 "movd %%mm1, 112(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2436 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2437 "movq (%0, %1, 4), %%mm0 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2438 "movq (%%ebx), %%mm1 \n\t" // abcdefgh |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2439 "movq %%mm0, %%mm2 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2440 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2441 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2442 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2443 "movq (%%ebx, %1), %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2444 "movq (%%ebx, %1, 2), %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2445 "movq %%mm1, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2446 "punpcklbw %%mm3, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2447 "punpckhbw %%mm3, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2448 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2449 "movq %%mm0, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2450 "punpcklwd %%mm1, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2451 "punpckhwd %%mm1, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2452 "movq %%mm2, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2453 "punpcklwd %%mm4, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2454 "punpckhwd %%mm4, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2455 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2456 "movd %%mm0, 132(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2457 "psrlq $32, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2458 "movd %%mm0, 148(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2459 "movd %%mm3, 164(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2460 "psrlq $32, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2461 "movd %%mm3, 180(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2462 "movd %%mm3, 52(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2463 "movd %%mm2, 196(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2464 "movd %%mm2, 68(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2465 "psrlq $32, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2466 "movd %%mm2, 84(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2467 "movd %%mm1, 100(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2468 "psrlq $32, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2469 "movd %%mm1, 116(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2470 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2471 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2472 :: "r" (src), "r" (srcStride), "r" (dst1), "r" (dst2) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2473 : "%eax", "%ebx" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2474 ); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2475 } |
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 /** |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2478 * transposes the given 8x8 block |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2479 */ |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2480 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
|
2481 { |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2482 asm( |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2483 "leal (%0, %1), %%eax \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2484 "leal (%%eax, %1, 4), %%ebx \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2485 // 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
|
2486 // %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
|
2487 "movq (%2), %%mm0 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2488 "movq 16(%2), %%mm1 \n\t" // abcdefgh |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2489 "movq %%mm0, %%mm2 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2490 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2491 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2492 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2493 "movq 32(%2), %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2494 "movq 48(%2), %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2495 "movq %%mm1, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2496 "punpcklbw %%mm3, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2497 "punpckhbw %%mm3, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2498 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2499 "movq %%mm0, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2500 "punpcklwd %%mm1, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2501 "punpckhwd %%mm1, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2502 "movq %%mm2, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2503 "punpcklwd %%mm4, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2504 "punpckhwd %%mm4, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2505 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2506 "movd %%mm0, (%0) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2507 "psrlq $32, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2508 "movd %%mm0, (%%eax) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2509 "movd %%mm3, (%%eax, %1) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2510 "psrlq $32, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2511 "movd %%mm3, (%%eax, %1, 2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2512 "movd %%mm2, (%0, %1, 4) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2513 "psrlq $32, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2514 "movd %%mm2, (%%ebx) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2515 "movd %%mm1, (%%ebx, %1) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2516 "psrlq $32, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2517 "movd %%mm1, (%%ebx, %1, 2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2518 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2519 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2520 "movq 64(%2), %%mm0 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2521 "movq 80(%2), %%mm1 \n\t" // abcdefgh |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2522 "movq %%mm0, %%mm2 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2523 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2524 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2525 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2526 "movq 96(%2), %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2527 "movq 112(%2), %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2528 "movq %%mm1, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2529 "punpcklbw %%mm3, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2530 "punpckhbw %%mm3, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2531 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2532 "movq %%mm0, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2533 "punpcklwd %%mm1, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2534 "punpckhwd %%mm1, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2535 "movq %%mm2, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2536 "punpcklwd %%mm4, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2537 "punpckhwd %%mm4, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2538 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2539 "movd %%mm0, 4(%0) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2540 "psrlq $32, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2541 "movd %%mm0, 4(%%eax) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2542 "movd %%mm3, 4(%%eax, %1) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2543 "psrlq $32, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2544 "movd %%mm3, 4(%%eax, %1, 2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2545 "movd %%mm2, 4(%0, %1, 4) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2546 "psrlq $32, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2547 "movd %%mm2, 4(%%ebx) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2548 "movd %%mm1, 4(%%ebx, %1) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2549 "psrlq $32, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2550 "movd %%mm1, 4(%%ebx, %1, 2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2551 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2552 :: "r" (dst), "r" (dstStride), "r" (src) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2553 : "%eax", "%ebx" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2554 ); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2555 } |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2556 #endif |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2557 |
2185 | 2558 #ifdef HAVE_ODIVX_POSTPROCESS |
2559 #include "../opendivx/postprocess.h" | |
2560 int use_old_pp=0; | |
2561 #endif | |
2159 | 2562 |
2185 | 2563 static void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height, |
2564 QP_STORE_T QPs[], int QPStride, int isColor, int mode); | |
2159 | 2565 |
2389 | 2566 /* -pp Command line Help |
2567 NOTE/FIXME: put this at an appropriate place (--help, html docs, man mplayer)? | |
2568 | |
2569 -pp <filterName>[:<option>[:<option>...]][,[-]<filterName>[:<option>...]]... | |
2570 | |
2571 long form example: | |
2572 -pp vdeblock:autoq,hdeblock:autoq,linblenddeint -pp default,-vdeblock | |
2573 short form example: | |
2574 -pp vb:a,hb:a,lb -pp de,-vb | |
2575 | |
2576 Filters Options | |
2577 short long name short long option Description | |
2578 * * a autoq cpu power dependant enabler | |
2579 c chrom chrominance filtring enabled | |
2580 y nochrom chrominance filtring disabled | |
2581 hb hdeblock horizontal deblocking filter | |
2582 vb vdeblock vertical deblocking filter | |
2583 vr rkvdeblock | |
2584 h1 x1hdeblock Experimental horizontal deblock filter 1 | |
2585 v1 x1vdeblock Experimental vertical deblock filter 1 | |
2586 dr dering not implemented yet | |
2587 al autolevels automatic brightness / contrast fixer | |
2588 f fullyrange stretch luminance range to (0..255) | |
2589 lb linblenddeint linear blend deinterlacer | |
2590 li linipoldeint linear interpolating deinterlacer | |
2591 ci cubicipoldeint cubic interpolating deinterlacer | |
2592 md mediandeint median deinterlacer | |
2593 de default hdeblock:a,vdeblock:a,dering:a,autolevels | |
2594 fa fast x1hdeblock:a,x1vdeblock:a,dering:a,autolevels | |
2595 */ | |
2596 | |
2597 /** | |
2598 * returns a PPMode struct which will have a non 0 error variable if an error occured | |
2599 * name is the string after "-pp" on the command line | |
2600 * quality is a number from 0 to GET_PP_QUALITY_MAX | |
2601 */ | |
2602 struct PPMode getPPModeByNameAndQuality(char *name, int quality) | |
2603 { | |
2604 char temp[GET_MODE_BUFFER_SIZE]; | |
2605 char *p= temp; | |
2606 char *filterDelimiters= ","; | |
2607 char *optionDelimiters= ":"; | |
2608 struct PPMode ppMode= {0,0,0,0,0,0}; | |
2609 char *filterToken; | |
2610 | |
2611 strncpy(temp, name, GET_MODE_BUFFER_SIZE); | |
2612 | |
2613 for(;;){ | |
2614 char *filterName; | |
2615 int q= GET_PP_QUALITY_MAX; | |
2616 int chrom=-1; | |
2617 char *option; | |
2618 char *options[OPTIONS_ARRAY_SIZE]; | |
2619 int i; | |
2620 int filterNameOk=0; | |
2621 int numOfUnknownOptions=0; | |
2622 int enable=1; //does the user want us to enabled or disabled the filter | |
2623 | |
2624 filterToken= strtok(p, filterDelimiters); | |
2625 if(filterToken == NULL) break; | |
2626 p+= strlen(filterToken) + 1; | |
2627 filterName= strtok(filterToken, optionDelimiters); | |
2628 printf("%s::%s\n", filterToken, filterName); | |
2629 | |
2630 if(*filterName == '-') | |
2631 { | |
2632 enable=0; | |
2633 filterName++; | |
2634 } | |
2635 for(;;){ //for all options | |
2636 option= strtok(NULL, optionDelimiters); | |
2637 if(option == NULL) break; | |
2638 | |
2639 printf("%s\n", option); | |
2640 if(!strcmp("autoq", option) || !strcmp("a", option)) q= quality; | |
2641 else if(!strcmp("nochrom", option) || !strcmp("y", option)) chrom=0; | |
2642 else if(!strcmp("chrom", option) || !strcmp("c", option)) chrom=1; | |
2643 else | |
2644 { | |
2645 options[numOfUnknownOptions] = option; | |
2646 numOfUnknownOptions++; | |
2647 options[numOfUnknownOptions] = NULL; | |
2648 } | |
2649 if(numOfUnknownOptions >= OPTIONS_ARRAY_SIZE-1) break; | |
2650 } | |
2651 | |
2652 /* replace stuff from the replace Table */ | |
2653 for(i=0; replaceTable[2*i]!=NULL; i++) | |
2654 { | |
2655 if(!strcmp(replaceTable[2*i], filterName)) | |
2656 { | |
2657 int newlen= strlen(replaceTable[2*i + 1]); | |
2658 int plen; | |
2659 int spaceLeft; | |
2660 | |
2661 if(p==NULL) p= temp, *p=0; //last filter | |
2662 else p--, *p=','; //not last filter | |
2663 | |
2664 plen= strlen(p); | |
2665 spaceLeft= (int)p - (int)temp + plen; | |
2666 if(spaceLeft + newlen >= GET_MODE_BUFFER_SIZE) | |
2667 { | |
2668 ppMode.error++; | |
2669 break; | |
2670 } | |
2671 memmove(p + newlen, p, plen+1); | |
2672 memcpy(p, replaceTable[2*i + 1], newlen); | |
2673 filterNameOk=1; | |
2674 } | |
2675 } | |
2676 | |
2677 for(i=0; filters[i].shortName!=NULL; i++) | |
2678 { | |
2679 if( !strcmp(filters[i].longName, filterName) | |
2680 || !strcmp(filters[i].shortName, filterName)) | |
2681 { | |
2682 ppMode.lumMode &= ~filters[i].mask; | |
2683 ppMode.chromMode &= ~filters[i].mask; | |
2684 | |
2685 filterNameOk=1; | |
2686 if(!enable) break; // user wants to disable it | |
2687 | |
2688 if(q >= filters[i].minLumQuality) | |
2689 ppMode.lumMode|= filters[i].mask; | |
2690 if(chrom==1 || (chrom==-1 && filters[i].chromDefault)) | |
2691 if(q >= filters[i].minChromQuality) | |
2692 ppMode.chromMode|= filters[i].mask; | |
2693 | |
2694 if(filters[i].mask == LEVEL_FIX) | |
2695 { | |
2696 int o; | |
2697 ppMode.minAllowedY= 16; | |
2698 ppMode.maxAllowedY= 234; | |
2699 for(o=0; options[o]!=NULL; o++) | |
2700 if( !strcmp(options[o],"fullyrange") | |
2701 ||!strcmp(options[o],"f")) | |
2702 { | |
2703 ppMode.minAllowedY= 0; | |
2704 ppMode.maxAllowedY= 255; | |
2705 numOfUnknownOptions--; | |
2706 } | |
2707 } | |
2708 } | |
2709 } | |
2710 if(!filterNameOk) ppMode.error++; | |
2711 ppMode.error += numOfUnknownOptions; | |
2712 } | |
2713 | |
2714 if(ppMode.lumMode & H_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_Y_H; | |
2715 if(ppMode.lumMode & V_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_Y_V; | |
2716 if(ppMode.chromMode & H_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_C_H; | |
2717 if(ppMode.chromMode & V_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_C_V; | |
2718 if(ppMode.lumMode & DERING) ppMode.oldMode |= PP_DERING_Y; | |
2719 if(ppMode.chromMode & DERING) ppMode.oldMode |= PP_DERING_C; | |
2720 | |
2721 return ppMode; | |
2722 } | |
2723 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2724 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2725 * ... |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2726 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2727 void postprocess(unsigned char * src[], int src_stride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2728 unsigned char * dst[], int dst_stride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2729 int horizontal_size, int vertical_size, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2730 QP_STORE_T *QP_store, int QP_stride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2731 int mode) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2732 { |
2389 | 2733 /* |
2734 static int qual=0; | |
2735 | |
2736 struct PPMode ppMode= getPPModeByNameAndQuality("fast,default,-hdeblock,-vdeblock", qual); | |
2737 qual++; | |
2738 qual%=7; | |
2739 printf("\n%d %d %d %d\n", ppMode.lumMode, ppMode.chromMode, ppMode.oldMode, ppMode.error); | |
2740 postprocess2(src, src_stride, dst, dst_stride, | |
2741 horizontal_size, vertical_size, QP_store, QP_stride, &ppMode); | |
2742 | |
2743 return; | |
2744 */ | |
2159 | 2745 |
2185 | 2746 #ifdef HAVE_ODIVX_POSTPROCESS |
2747 // Note: I could make this shit outside of this file, but it would mean one | |
2748 // more function call... | |
2749 if(use_old_pp){ | |
2750 odivx_postprocess(src,src_stride,dst,dst_stride,horizontal_size,vertical_size,QP_store,QP_stride,mode); | |
2751 return; | |
2752 } | |
2753 #endif | |
2754 | |
2159 | 2755 postProcess(src[0], src_stride, dst[0], dst_stride, |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2756 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
|
2757 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2758 horizontal_size >>= 1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2759 vertical_size >>= 1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2760 src_stride >>= 1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2761 dst_stride >>= 1; |
2179 | 2762 mode= ((mode&0xFF)>>4) | (mode&0xFFFFFF00); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2763 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2764 if(1) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2765 { |
2159 | 2766 postProcess(src[1], src_stride, dst[1], dst_stride, |
2179 | 2767 horizontal_size, vertical_size, QP_store, QP_stride, 1, mode); |
2159 | 2768 postProcess(src[2], src_stride, dst[2], dst_stride, |
2389 | 2769 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
|
2770 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2771 else |
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 memcpy(dst[1], src[1], src_stride*horizontal_size); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2774 memcpy(dst[2], src[2], src_stride*horizontal_size); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2775 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2776 } |
2185 | 2777 |
2389 | 2778 void postprocess2(unsigned char * src[], int src_stride, |
2779 unsigned char * dst[], int dst_stride, | |
2780 int horizontal_size, int vertical_size, | |
2781 QP_STORE_T *QP_store, int QP_stride, | |
2782 struct PPMode *mode) | |
2783 { | |
2784 | |
2785 #ifdef HAVE_ODIVX_POSTPROCESS | |
2786 // Note: I could make this shit outside of this file, but it would mean one | |
2787 // more function call... | |
2788 if(use_old_pp){ | |
2789 odivx_postprocess(src,src_stride,dst,dst_stride,horizontal_size,vertical_size,QP_store,QP_stride, | |
2790 mode->oldMode); | |
2791 return; | |
2792 } | |
2793 #endif | |
2794 | |
2795 postProcess(src[0], src_stride, dst[0], dst_stride, | |
2796 horizontal_size, vertical_size, QP_store, QP_stride, 0, mode->lumMode); | |
2797 | |
2798 horizontal_size >>= 1; | |
2799 vertical_size >>= 1; | |
2800 src_stride >>= 1; | |
2801 dst_stride >>= 1; | |
2802 | |
2803 postProcess(src[1], src_stride, dst[1], dst_stride, | |
2804 horizontal_size, vertical_size, QP_store, QP_stride, 1, mode->chromMode); | |
2805 postProcess(src[2], src_stride, dst[2], dst_stride, | |
2806 horizontal_size, vertical_size, QP_store, QP_stride, 2, mode->chromMode); | |
2807 } | |
2808 | |
2809 | |
2159 | 2810 /** |
2811 * gets the mode flags for a given quality (larger values mean slower but better postprocessing) | |
2185 | 2812 * 0 <= quality <= 6 |
2159 | 2813 */ |
2185 | 2814 int getPpModeForQuality(int quality){ |
2815 int modes[1+GET_PP_QUALITY_MAX]= { | |
2816 0, | |
2817 #if 1 | |
2818 // horizontal filters first | |
2819 LUM_H_DEBLOCK, | |
2820 LUM_H_DEBLOCK | LUM_V_DEBLOCK, | |
2821 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK, | |
2822 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK, | |
2823 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK | LUM_DERING, | |
2824 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK | LUM_DERING | CHROM_DERING | |
2825 #else | |
2826 // vertical filters first | |
2159 | 2827 LUM_V_DEBLOCK, |
2828 LUM_V_DEBLOCK | LUM_H_DEBLOCK, | |
2829 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK, | |
2830 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK, | |
2831 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING, | |
2832 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING | CHROM_DERING | |
2185 | 2833 #endif |
2834 }; | |
2159 | 2835 |
2185 | 2836 #ifdef HAVE_ODIVX_POSTPROCESS |
2837 int odivx_modes[1+GET_PP_QUALITY_MAX]= { | |
2838 0, | |
2839 PP_DEBLOCK_Y_H, | |
2840 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V, | |
2841 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H, | |
2842 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V, | |
2843 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V|PP_DERING_Y, | |
2844 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V|PP_DERING_Y|PP_DERING_C | |
2845 }; | |
2846 if(use_old_pp) return odivx_modes[quality]; | |
2847 #endif | |
2848 return modes[quality]; | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2849 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2850 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2851 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2852 * 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
|
2853 * numLines must be a multiple of 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2854 * levelFix == 0 -> dont touch the brighness & contrast |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2855 */ |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2856 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
|
2857 int numLines, int levelFix) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2858 { |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2859 #ifndef HAVE_MMX |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2860 int i; |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2861 #endif |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2862 if(levelFix) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2863 { |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2864 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2865 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2866 "leal (%2,%2), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2867 "leal (%3,%3), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2868 "movq packedYOffset, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2869 "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
|
2870 "pxor %%mm4, %%mm4 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2871 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2872 #define SCALED_CPY \ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2873 "movq (%0), %%mm0 \n\t"\ |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2874 "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
|
2875 "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
|
2876 "punpckhbw %%mm4, %%mm5 \n\t"\ |
2394 | 2877 "psubw %%mm2, %%mm0 \n\t"\ |
2878 "psubw %%mm2, %%mm5 \n\t"\ | |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2879 "movq (%0,%2), %%mm1 \n\t"\ |
2394 | 2880 "psllw $6, %%mm0 \n\t"\ |
2881 "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
|
2882 "pmulhw %%mm3, %%mm0 \n\t"\ |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2883 "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
|
2884 "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
|
2885 "punpcklbw %%mm4, %%mm1 \n\t"\ |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2886 "punpckhbw %%mm4, %%mm6 \n\t"\ |
2394 | 2887 "psubw %%mm2, %%mm1 \n\t"\ |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2888 "psubw %%mm2, %%mm6 \n\t"\ |
2394 | 2889 "psllw $6, %%mm1 \n\t"\ |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2890 "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
|
2891 "pmulhw %%mm3, %%mm1 \n\t"\ |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2892 "pmulhw %%mm3, %%mm6 \n\t"\ |
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2893 "addl %%eax, %0 \n\t"\ |
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2894 "packuswb %%mm5, %%mm0 \n\t"\ |
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2895 "packuswb %%mm6, %%mm1 \n\t"\ |
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2896 "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
|
2897 "movq %%mm1, (%1, %3) \n\t"\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2898 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2899 SCALED_CPY |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2900 "addl %%ebx, %1 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2901 SCALED_CPY |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2902 "addl %%ebx, %1 \n\t" |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2903 SCALED_CPY |
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2904 "addl %%ebx, %1 \n\t" |
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2905 SCALED_CPY |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2906 |
2413 | 2907 : "+r"(src), |
2908 "+r"(dst) | |
2909 :"r" (srcStride), | |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2910 "r" (dstStride) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2911 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2912 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2913 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2914 for(i=0; i<numLines; i++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2915 memcpy( &(dst[dstStride*i]), |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2916 &(src[srcStride*i]), BLOCK_SIZE); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2917 #endif |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2918 } |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2919 else |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2920 { |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2921 #ifdef HAVE_MMX |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2922 asm volatile( |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2923 "movl %4, %%eax \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2924 "movl %%eax, temp0\n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2925 "pushl %0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2926 "pushl %1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2927 "leal (%2,%2), %%eax \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2928 "leal (%3,%3), %%ebx \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2929 "movq packedYOffset, %%mm2 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2930 "movq packedYScale, %%mm3 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2931 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2932 #define SIMPLE_CPY \ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2933 "movq (%0), %%mm0 \n\t"\ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2934 "movq (%0,%2), %%mm1 \n\t"\ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2935 "movq %%mm0, (%1) \n\t"\ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2936 "movq %%mm1, (%1, %3) \n\t"\ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2937 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2938 "1: \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2939 SIMPLE_CPY |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2940 "addl %%eax, %0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2941 "addl %%ebx, %1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2942 SIMPLE_CPY |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2943 "addl %%eax, %0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2944 "addl %%ebx, %1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2945 "decl temp0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2946 "jnz 1b \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2947 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2948 "popl %1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2949 "popl %0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2950 : : "r" (src), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2951 "r" (dst), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2952 "r" (srcStride), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2953 "r" (dstStride), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2954 "m" (numLines>>2) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2955 : "%eax", "%ebx" |
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 #else |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2958 for(i=0; i<numLines; i++) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2959 memcpy( &(dst[dstStride*i]), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2960 &(src[srcStride*i]), BLOCK_SIZE); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2961 #endif |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2962 } |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2963 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2964 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2965 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2966 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2967 * Filters array of bytes (Y or U or V values) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2968 */ |
2185 | 2969 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
|
2970 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
|
2971 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2972 int x,y; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2973 /* 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
|
2974 after watching a black picture for 5 hours*/ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2975 static uint64_t *yHistogram= NULL; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2976 int black=0, white=255; // blackest black and whitest white in the picture |
2437 | 2977 int QPCorrecture= 256; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2978 |
2246 | 2979 /* Temporary buffers for handling the last row(s) */ |
2980 static uint8_t *tempDst= NULL; | |
2981 static uint8_t *tempSrc= NULL; | |
2982 | |
2285 | 2983 /* Temporary buffers for handling the last block */ |
2984 static uint8_t *tempDstBlock= NULL; | |
2985 static uint8_t *tempSrcBlock= NULL; | |
2986 | |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2987 #ifdef PP_FUNNY_STRIDE |
2285 | 2988 uint8_t *dstBlockPtrBackup; |
2989 uint8_t *srcBlockPtrBackup; | |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2990 #endif |
2285 | 2991 |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2992 #ifdef MORE_TIMING |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2993 long long T0, T1, diffTime=0; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2994 #endif |
2221 | 2995 #ifdef TIMING |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2996 long long memcpyTime=0, vertTime=0, horizTime=0, sumTime; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2997 sumTime= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2998 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2999 |
2246 | 3000 if(tempDst==NULL) |
3001 { | |
3002 tempDst= (uint8_t*)memalign(8, 1024*24); | |
3003 tempSrc= (uint8_t*)memalign(8, 1024*24); | |
2285 | 3004 tempDstBlock= (uint8_t*)memalign(8, 1024*24); |
3005 tempSrcBlock= (uint8_t*)memalign(8, 1024*24); | |
2246 | 3006 } |
3007 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3008 if(!yHistogram) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3009 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3010 int i; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3011 yHistogram= (uint64_t*)malloc(8*256); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3012 for(i=0; i<256; i++) yHistogram[i]= width*height/64*15/256; |
2285 | 3013 |
3014 if(mode & FULL_Y_RANGE) | |
3015 { | |
3016 maxAllowedY=255; | |
3017 minAllowedY=0; | |
3018 } | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3019 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3020 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3021 if(!isColor) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3022 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3023 uint64_t sum= 0; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3024 int i; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3025 static int framenum= -1; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3026 uint64_t maxClipped; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3027 uint64_t clipped; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3028 double scale; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3029 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3030 framenum++; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3031 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
|
3032 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3033 for(i=0; i<256; i++) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3034 { |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3035 sum+= yHistogram[i]; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3036 // printf("%d ", yHistogram[i]); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3037 } |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3038 // printf("\n\n"); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3039 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3040 /* we allways get a completly black picture first */ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3041 maxClipped= (uint64_t)(sum * maxClippedThreshold); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3042 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3043 clipped= sum; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3044 for(black=255; black>0; black--) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3045 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3046 if(clipped < maxClipped) break; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3047 clipped-= yHistogram[black]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3048 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3049 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3050 clipped= sum; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3051 for(white=0; white<256; white++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3052 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3053 if(clipped < maxClipped) break; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3054 clipped-= yHistogram[white]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3055 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3056 |
2394 | 3057 packedYOffset= (black - minAllowedY) & 0xFFFF; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3058 packedYOffset|= packedYOffset<<32; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3059 packedYOffset|= packedYOffset<<16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3060 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3061 scale= (double)(maxAllowedY - minAllowedY) / (double)(white-black); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3062 |
2394 | 3063 packedYScale= (uint16_t)(scale*1024.0 + 0.5); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3064 packedYScale|= packedYScale<<32; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3065 packedYScale|= packedYScale<<16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3066 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3067 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3068 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3069 packedYScale= 0x0100010001000100LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3070 packedYOffset= 0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3071 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3072 |
2437 | 3073 if(mode & LEVEL_FIX) QPCorrecture= packedYScale &0xFFFF; |
3074 else QPCorrecture= 256; | |
3075 | |
2246 | 3076 /* copy first row of 8x8 blocks */ |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3077 for(x=0; x<width; x+=BLOCK_SIZE) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3078 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
|
3079 |
2246 | 3080 for(y=0; y<height; y+=BLOCK_SIZE) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3081 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3082 //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
|
3083 uint8_t *srcBlock= &(src[y*srcStride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3084 uint8_t *dstBlock= &(dst[y*dstStride]); |
2437 | 3085 #ifdef ARCH_X86 |
3086 int *QPptr= isColor ? &QPs[(y>>3)*QPStride] :&QPs[(y>>4)*QPStride]; | |
3087 int QPDelta= isColor ? 1<<(32-3) : 1<<(32-4); | |
3088 int QPFrac= QPDelta; | |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3089 uint8_t *tempBlock1= tempBlocks; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3090 uint8_t *tempBlock2= tempBlocks + 8; |
2437 | 3091 #endif |
2473 | 3092 /* can we mess with a 8x16 block from srcBlock/dstBlock downwards and 1 line upwards |
3093 if not than use a temporary buffer */ | |
2246 | 3094 if(y+15 >= height) |
3095 { | |
2473 | 3096 /* copy from line 5 to 12 of src, these will be copied with |
2246 | 3097 blockcopy to dst later */ |
3098 memcpy(tempSrc + srcStride*5, srcBlock + srcStride*5, | |
3099 srcStride*MAX(height-y-5, 0) ); | |
3100 | |
3101 /* duplicate last line to fill the void upto line 12 */ | |
3102 if(y+12 >= height) | |
3103 { | |
3104 int i; | |
3105 for(i=height-y; i<=12; i++) | |
3106 memcpy(tempSrc + srcStride*i, | |
3107 src + srcStride*(height-1), srcStride); | |
3108 } | |
3109 | |
3110 | |
2473 | 3111 /* copy up to 6 lines of dst */ |
3112 memcpy(tempDst, dstBlock - dstStride, dstStride*MIN(height-y+1, 6) ); | |
3113 dstBlock= tempDst + dstStride; | |
2246 | 3114 srcBlock= tempSrc; |
3115 } | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3116 |
2285 | 3117 // 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
|
3118 // 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
|
3119 // 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
|
3120 for(x=0; x<width; x+=BLOCK_SIZE) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3121 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3122 const int stride= dstStride; |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3123 uint8_t *tmpXchg; |
2437 | 3124 #ifdef ARCH_X86 |
3125 int QP= *QPptr; | |
3126 asm volatile( | |
3127 "addl %2, %1 \n\t" | |
3128 "sbbl %%eax, %%eax \n\t" | |
3129 "shll $2, %%eax \n\t" | |
3130 "subl %%eax, %0 \n\t" | |
3131 : "+r" (QPptr), "+m" (QPFrac) | |
3132 : "r" (QPDelta) | |
3133 : "%eax" | |
3134 ); | |
3135 #else | |
3136 int QP= isColor ? | |
3137 QPs[(y>>3)*QPStride + (x>>3)]: | |
3138 QPs[(y>>4)*QPStride + (x>>4)]; | |
3139 #endif | |
3140 if(!isColor) | |
2428 | 3141 { |
2437 | 3142 QP= (QP* QPCorrecture)>>8; |
3143 yHistogram[ srcBlock[srcStride*4 + 4] ]++; | |
2428 | 3144 } |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3145 #ifdef HAVE_MMX |
2246 | 3146 asm volatile( |
3147 "movd %0, %%mm7 \n\t" | |
3148 "packuswb %%mm7, %%mm7 \n\t" // 0, 0, 0, QP, 0, 0, 0, QP | |
3149 "packuswb %%mm7, %%mm7 \n\t" // 0,QP, 0, QP, 0,QP, 0, QP | |
3150 "packuswb %%mm7, %%mm7 \n\t" // QP,..., QP | |
3151 "movq %%mm7, pQPb \n\t" | |
3152 : : "r" (QP) | |
3153 ); | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3154 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3155 |
2221 | 3156 #ifdef MORE_TIMING |
2246 | 3157 T0= rdtsc(); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3158 #endif |
2159 | 3159 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3160 #ifdef HAVE_MMX2 |
2437 | 3161 /* |
2246 | 3162 prefetchnta(srcBlock + (((x>>3)&3) + 5)*srcStride + 32); |
3163 prefetchnta(srcBlock + (((x>>3)&3) + 9)*srcStride + 32); | |
3164 prefetcht0(dstBlock + (((x>>3)&3) + 5)*dstStride + 32); | |
3165 prefetcht0(dstBlock + (((x>>3)&3) + 9)*dstStride + 32); | |
2437 | 3166 */ |
3167 /* | |
3168 prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32); | |
3169 prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32); | |
3170 prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32); | |
3171 prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32); | |
3172 */ | |
3173 | |
3174 asm( | |
3175 "movl %4, %%eax \n\t" | |
3176 "shrl $2, %%eax \n\t" | |
3177 "andl $6, %%eax \n\t" | |
3178 "addl $5, %%eax \n\t" | |
3179 "movl %%eax, %%ebx \n\t" | |
3180 "imul %1, %%eax \n\t" | |
3181 "imul %3, %%ebx \n\t" | |
3182 "prefetchnta 32(%%eax, %0) \n\t" | |
3183 "prefetcht0 32(%%ebx, %2) \n\t" | |
3184 "addl %1, %%eax \n\t" | |
3185 "addl %3, %%ebx \n\t" | |
3186 "prefetchnta 32(%%eax, %0) \n\t" | |
3187 "prefetcht0 32(%%ebx, %2) \n\t" | |
3188 :: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride), | |
3189 "m" (x) | |
3190 : "%eax", "%ebx" | |
3191 ); | |
3192 | |
2159 | 3193 #elif defined(HAVE_3DNOW) |
3194 //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ... | |
2246 | 3195 /* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32); |
3196 prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32); | |
3197 prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32); | |
3198 prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32); | |
2159 | 3199 */ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3200 #endif |
2246 | 3201 |
2286 | 3202 #ifdef PP_FUNNY_STRIDE |
2285 | 3203 //can we mess with a 8x16 block, if not use a temp buffer, yes again |
3204 if(x+7 >= width) | |
3205 { | |
3206 int i; | |
3207 dstBlockPtrBackup= dstBlock; | |
3208 srcBlockPtrBackup= srcBlock; | |
3209 | |
3210 for(i=0;i<BLOCK_SIZE*2; i++) | |
3211 { | |
3212 memcpy(tempSrcBlock+i*srcStride, srcBlock+i*srcStride, width-x); | |
3213 memcpy(tempDstBlock+i*dstStride, dstBlock+i*dstStride, width-x); | |
3214 } | |
3215 | |
3216 dstBlock= tempDstBlock; | |
3217 srcBlock= tempSrcBlock; | |
3218 } | |
2286 | 3219 #endif |
2285 | 3220 |
2246 | 3221 blockCopy(dstBlock + dstStride*5, dstStride, |
3222 srcBlock + srcStride*5, srcStride, 8, mode & LEVEL_FIX); | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3223 |
2246 | 3224 if(mode & LINEAR_IPOL_DEINT_FILTER) |
3225 deInterlaceInterpolateLinear(dstBlock, dstStride); | |
3226 else if(mode & LINEAR_BLEND_DEINT_FILTER) | |
3227 deInterlaceBlendLinear(dstBlock, dstStride); | |
3228 else if(mode & MEDIAN_DEINT_FILTER) | |
3229 deInterlaceMedian(dstBlock, dstStride); | |
3230 else if(mode & CUBIC_IPOL_DEINT_FILTER) | |
3231 deInterlaceInterpolateCubic(dstBlock, dstStride); | |
3232 /* else if(mode & CUBIC_BLEND_DEINT_FILTER) | |
3233 deInterlaceBlendCubic(dstBlock, dstStride); | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
3234 */ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3235 |
2246 | 3236 /* only deblock if we have 2 blocks */ |
3237 if(y + 8 < height) | |
3238 { | |
2221 | 3239 #ifdef MORE_TIMING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3240 T1= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3241 memcpyTime+= T1-T0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3242 T0=T1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3243 #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
|
3244 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
|
3245 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
|
3246 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
|
3247 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
|
3248 else if(mode & V_DEBLOCK) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3249 { |
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
|
3250 if( isVertDC(dstBlock, stride)) |
2159 | 3251 { |
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
|
3252 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
|
3253 doVertLowPass(dstBlock, stride, QP); |
2159 | 3254 } |
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
|
3255 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
|
3256 doVertDefFilter(dstBlock, stride, QP); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3257 } |
2221 | 3258 #ifdef MORE_TIMING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3259 T1= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3260 vertTime+= T1-T0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3261 T0=T1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3262 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3263 } |
2473 | 3264 |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3265 #ifdef HAVE_MMX |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3266 transpose1(tempBlock1, tempBlock2, dstBlock, dstStride); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3267 #endif |
2246 | 3268 /* check if we have a previous block to deblock it with dstBlock */ |
2285 | 3269 if(x - 8 >= 0) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3270 { |
2221 | 3271 #ifdef MORE_TIMING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3272 T0= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3273 #endif |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3274 #ifdef HAVE_MMX |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3275 if(mode & H_RK1_FILTER) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3276 vertRK1Filter(tempBlock1, 16, QP); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3277 else if(mode & H_X1_FILTER) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3278 vertX1Filter(tempBlock1, 16, QP); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3279 else if(mode & H_DEBLOCK) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3280 { |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3281 if( isVertDC(tempBlock1, 16)) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3282 { |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3283 if(isVertMinMaxOk(tempBlock1, 16, QP)) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3284 doVertLowPass(tempBlock1, 16, QP); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3285 } |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3286 else |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3287 doVertDefFilter(tempBlock1, 16, QP); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3288 } |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3289 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3290 transpose2(dstBlock-4, dstStride, tempBlock1 + 4*16); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3291 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3292 #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
|
3293 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
|
3294 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
|
3295 else if(mode & H_DEBLOCK) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3296 { |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3297 if( isHorizDC(dstBlock-4, stride)) |
2159 | 3298 { |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3299 if(isHorizMinMaxOk(dstBlock-4, stride, QP)) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3300 doHorizLowPass(dstBlock-4, stride, QP); |
2159 | 3301 } |
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
|
3302 else |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3303 doHorizDefFilter(dstBlock-4, stride, QP); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3304 } |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3305 #endif |
2221 | 3306 #ifdef MORE_TIMING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3307 T1= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3308 horizTime+= T1-T0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3309 T0=T1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3310 #endif |
2473 | 3311 if(mode & DERING) |
3312 { | |
3313 //FIXME filter first line | |
3314 if(y>0) dering(dstBlock - stride - 8, stride, QP); | |
3315 } | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3316 } |
2473 | 3317 else if(mode & DERING) |
3318 { | |
3319 //FIXME y+15 is required cuz of the tempBuffer thing -> bottom right block isnt filtered | |
3320 if(y > 8 && y+15 < height) dering(dstBlock - stride*9 + width - 8, stride, QP); | |
3321 } | |
3322 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3323 |
2286 | 3324 #ifdef PP_FUNNY_STRIDE |
2285 | 3325 /* did we use a tmp-block buffer */ |
3326 if(x+7 >= width) | |
3327 { | |
3328 int i; | |
3329 dstBlock= dstBlockPtrBackup; | |
3330 srcBlock= srcBlockPtrBackup; | |
3331 | |
3332 for(i=0;i<BLOCK_SIZE*2; i++) | |
3333 { | |
3334 memcpy(dstBlock+i*dstStride, tempDstBlock+i*dstStride, width-x); | |
3335 } | |
3336 } | |
2286 | 3337 #endif |
2285 | 3338 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3339 dstBlock+=8; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3340 srcBlock+=8; |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3341 |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
3342 #ifdef HAVE_MMX |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3343 tmpXchg= tempBlock1; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3344 tempBlock1= tempBlock2; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3345 tempBlock2 = tmpXchg; |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
3346 #endif |
2246 | 3347 } |
3348 | |
3349 /* did we use a tmp buffer */ | |
2285 | 3350 if(y+15 >= height) |
2246 | 3351 { |
3352 uint8_t *dstBlock= &(dst[y*dstStride]); | |
2473 | 3353 memcpy(dstBlock, tempDst + dstStride, dstStride*(height-y) ); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3354 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3355 } |
2159 | 3356 #ifdef HAVE_3DNOW |
3357 asm volatile("femms"); | |
3358 #elif defined (HAVE_MMX) | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3359 asm volatile("emms"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3360 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3361 |
2221 | 3362 #ifdef TIMING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3363 // 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
|
3364 sumTime= rdtsc() - sumTime; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3365 if(!isColor) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3366 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
|
3367 (int)(memcpyTime/1000), (int)(vertTime/1000), (int)(horizTime/1000), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3368 (int)(sumTime/1000), (int)((sumTime-memcpyTime-vertTime-horizTime)/1000) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3369 , black, white); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3370 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3371 } |