Mercurial > mplayer.hg
annotate postproc/postprocess_template.c @ 2475:3369845d92f4
3dnow dering
author | michael |
---|---|
date | Thu, 25 Oct 2001 22:50:03 +0000 |
parents | 94a0265c408c |
children | a6c5a537f30a |
rev | line source |
---|---|
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2 Copyright (C) 2001 Michael Niedermayer (michaelni@gmx.at) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
4 This program is free software; you can redistribute it and/or modify |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
5 it under the terms of the GNU General Public License as published by |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
6 the Free Software Foundation; either version 2 of the License, or |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
7 (at your option) any later version. |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
8 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
9 This program is distributed in the hope that it will be useful, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
10 but WITHOUT ANY WARRANTY; without even the implied warranty of |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
12 GNU General Public License for more details. |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
13 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
14 You should have received a copy of the GNU General Public License |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
15 along with this program; if not, write to the Free Software |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
16 Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
17 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
18 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
19 /* |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
20 C MMX MMX2 3DNow |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
21 isVertDC Ec Ec |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
22 isVertMinMaxOk Ec Ec |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
23 doVertLowPass E e e |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
24 doVertDefFilter Ec Ec Ec |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
25 isHorizDC Ec Ec |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
26 isHorizMinMaxOk a E |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
27 doHorizLowPass E e e |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
28 doHorizDefFilter Ec Ec Ec |
2475 | 29 deRing e e* |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
30 Vertical RKAlgo1 E a a |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
31 Horizontal RKAlgo1 a a |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
32 Vertical X1 a E E |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
33 Horizontal X1 a E E |
2246 | 34 LinIpolDeinterlace e E E* |
35 CubicIpolDeinterlace a e e* | |
36 LinBlendDeinterlace e E E* | |
2221 | 37 MedianDeinterlace Ec Ec |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
38 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
39 |
2159 | 40 * i dont have a 3dnow CPU -> its untested |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
41 E = Exact implementation |
2246 | 42 e = allmost exact implementation (slightly different rounding,...) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
43 a = alternative / approximate impl |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
44 c = checked against the other implementations (-vo md5) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
45 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
46 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
47 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
48 TODO: |
2179 | 49 verify that everything workes as it should (how?) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
50 reduce the time wasted on the mem transfer |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
51 implement dering |
2159 | 52 implement everything in C at least (done at the moment but ...) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
53 unroll stuff if instructions depend too much on the prior one |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
54 we use 8x8 blocks for the horizontal filters, opendivx seems to use 8x4? |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
55 move YScale thing to the end instead of fixing QP |
2159 | 56 write a faster and higher quality deblocking filter :) |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
57 do something about the speed of the horizontal filters |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
58 make the mainloop more flexible (variable number of blocks at once |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
59 (the if/else stuff per block is slowing things down) |
2179 | 60 compare the quality & speed of all filters |
61 split this huge file | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
62 fix warnings (unused vars, ...) |
2221 | 63 noise reduction filters |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
64 border remover |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
65 ... |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
66 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
67 Notes: |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
68 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
69 |
2221 | 70 //Changelog: use the CVS log |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
71 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
72 #include <inttypes.h> |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
73 #include <stdio.h> |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
74 #include <stdlib.h> |
2389 | 75 #include <string.h> |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
76 #include "../config.h" |
2159 | 77 //#undef HAVE_MMX2 |
78 //#define HAVE_3DNOW | |
79 //#undef HAVE_MMX | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
80 #include "postprocess.h" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
81 |
2189 | 82 #define MIN(a,b) ((a) > (b) ? (b) : (a)) |
83 #define MAX(a,b) ((a) < (b) ? (b) : (a)) | |
84 #define ABS(a) ((a) > 0 ? (a) : (-(a))) | |
85 #define SIGN(a) ((a) > 0 ? 1 : -1) | |
86 | |
87 #ifdef HAVE_MMX2 | |
88 #define PAVGB(a,b) "pavgb " #a ", " #b " \n\t" | |
89 #elif defined (HAVE_3DNOW) | |
90 #define PAVGB(a,b) "pavgusb " #a ", " #b " \n\t" | |
91 #endif | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
92 |
2389 | 93 #define GET_MODE_BUFFER_SIZE 500 |
94 #define OPTIONS_ARRAY_SIZE 10 | |
95 | |
96 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
97 static uint64_t packedYOffset= 0x0000000000000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
98 static uint64_t packedYScale= 0x0100010001000100LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
99 static uint64_t w05= 0x0005000500050005LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
100 static uint64_t w20= 0x0020002000200020LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
101 static uint64_t w1400= 0x1400140014001400LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
102 static uint64_t bm00000001= 0x00000000000000FFLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
103 static uint64_t bm00010000= 0x000000FF00000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
104 static uint64_t bm00001000= 0x00000000FF000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
105 static uint64_t bm10000000= 0xFF00000000000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
106 static uint64_t bm10000001= 0xFF000000000000FFLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
107 static uint64_t bm11000011= 0xFFFF00000000FFFFLL; |
2159 | 108 static uint64_t bm00000011= 0x000000000000FFFFLL; |
2179 | 109 static uint64_t bm11111110= 0xFFFFFFFFFFFFFF00LL; |
2159 | 110 static uint64_t bm11000000= 0xFFFF000000000000LL; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
111 static uint64_t bm00011000= 0x000000FFFF000000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
112 static uint64_t bm00110011= 0x0000FFFF0000FFFFLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
113 static uint64_t bm11001100= 0xFFFF0000FFFF0000LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
114 static uint64_t b00= 0x0000000000000000LL; |
2179 | 115 static uint64_t b01= 0x0101010101010101LL; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
116 static uint64_t b02= 0x0202020202020202LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
117 static uint64_t b0F= 0x0F0F0F0F0F0F0F0FLL; |
2473 | 118 static uint64_t b04= 0x0404040404040404LL; |
119 static uint64_t b08= 0x0808080808080808LL; | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
120 static uint64_t bFF= 0xFFFFFFFFFFFFFFFFLL; |
2159 | 121 static uint64_t b20= 0x2020202020202020LL; |
122 static uint64_t b80= 0x8080808080808080LL; | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
123 static uint64_t b7E= 0x7E7E7E7E7E7E7E7ELL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
124 static uint64_t b7C= 0x7C7C7C7C7C7C7C7CLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
125 static uint64_t b3F= 0x3F3F3F3F3F3F3F3FLL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
126 static uint64_t temp0=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
127 static uint64_t temp1=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
128 static uint64_t temp2=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
129 static uint64_t temp3=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
130 static uint64_t temp4=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
131 static uint64_t temp5=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
132 static uint64_t pQPb=0; |
2473 | 133 static uint64_t pQPb2=0; |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
134 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
|
135 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
136 int hFlatnessThreshold= 56 - 16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
137 int vFlatnessThreshold= 56 - 16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
138 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
139 //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
|
140 double maxClippedThreshold= 0.01; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
141 |
2389 | 142 int maxAllowedY=234; |
2285 | 143 int minAllowedY=16; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
144 |
2389 | 145 static struct PPFilter filters[]= |
146 { | |
147 {"hb", "hdeblock", 1, 1, 3, H_DEBLOCK}, | |
148 {"vb", "vdeblock", 1, 2, 4, V_DEBLOCK}, | |
149 {"vr", "rkvdeblock", 1, 2, 4, H_RK1_FILTER}, | |
150 {"h1", "x1hdeblock", 1, 1, 3, H_X1_FILTER}, | |
151 {"v1", "x1vdeblock", 1, 2, 4, V_X1_FILTER}, | |
152 {"dr", "dering", 1, 5, 6, DERING}, | |
153 {"al", "autolevels", 0, 1, 2, LEVEL_FIX}, | |
154 {"lb", "linblenddeint", 0, 1, 6, LINEAR_BLEND_DEINT_FILTER}, | |
155 {"li", "linipoldeint", 0, 1, 6, LINEAR_IPOL_DEINT_FILTER}, | |
156 {"ci", "cubicipoldeint", 0, 1, 6, CUBIC_IPOL_DEINT_FILTER}, | |
157 {"md", "mediandeint", 0, 1, 6, MEDIAN_DEINT_FILTER}, | |
158 {NULL, NULL,0,0,0,0} //End Marker | |
159 }; | |
160 | |
161 static char *replaceTable[]= | |
162 { | |
163 "default", "hdeblock:a,vdeblock:a,dering:a,autolevels", | |
164 "de", "hdeblock:a,vdeblock:a,dering:a,autolevels", | |
165 "fast", "x1hdeblock:a,x1vdeblock:a,dering:a,autolevels", | |
166 "fa", "x1hdeblock:a,x1vdeblock:a,dering:a,autolevels", | |
167 NULL //End Marker | |
168 }; | |
169 | |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
170 static inline void unusedVariableWarningFixer() |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
171 { |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
172 if( |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
173 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
|
174 + bm00001000 + bm10000000 + bm10000001 + bm11000011 + bm00000011 + bm11111110 |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
175 + 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
|
176 + 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
|
177 + temp5 + pQPb== 0) b00=0; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
178 } |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
179 |
2221 | 180 #ifdef TIMING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
181 static inline long long rdtsc() |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
182 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
183 long long l; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
184 asm volatile( "rdtsc\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
185 : "=A" (l) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
186 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
187 // printf("%d\n", int(l/1000)); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
188 return l; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
189 } |
2185 | 190 #endif |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
191 |
2185 | 192 #ifdef HAVE_MMX2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
193 static inline void prefetchnta(void *p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
194 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
195 asm volatile( "prefetchnta (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
196 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
197 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
198 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
199 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
200 static inline void prefetcht0(void *p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
201 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
202 asm volatile( "prefetcht0 (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
203 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
204 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
205 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
206 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
207 static inline void prefetcht1(void *p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
208 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
209 asm volatile( "prefetcht1 (%0)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
210 : : "r" (p) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
211 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
212 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
213 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
214 static inline void prefetcht2(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( "prefetcht2 (%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 } |
2185 | 220 #endif |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
221 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
222 //FIXME? |255-0| = 1 (shouldnt be a problem ...) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
223 /** |
2246 | 224 * 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
|
225 */ |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
226 static inline int isVertDC(uint8_t src[], int stride){ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
227 int numEq= 0; |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
228 #ifndef HAVE_MMX |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
229 int y; |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
230 #endif |
2246 | 231 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
|
232 #ifdef HAVE_MMX |
2413 | 233 asm volatile( |
234 "leal (%1, %2), %%eax \n\t" | |
235 "leal (%%eax, %2, 4), %%ebx \n\t" | |
236 // 0 1 2 3 4 5 6 7 8 9 | |
237 // %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
|
238 "movq b7E, %%mm7 \n\t" // mm7 = 0x7F |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
239 "movq b7C, %%mm6 \n\t" // mm6 = 0x7D |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
240 "movq (%1), %%mm0 \n\t" |
2413 | 241 "movq (%%eax), %%mm1 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
242 "psubb %%mm1, %%mm0 \n\t" // mm0 = differnece |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
243 "paddb %%mm7, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
244 "pcmpgtb %%mm6, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
245 |
2413 | 246 "movq (%%eax,%2), %%mm2 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
247 "psubb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
248 "paddb %%mm7, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
249 "pcmpgtb %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
250 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
251 |
2413 | 252 "movq (%%eax, %2, 2), %%mm1 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
253 "psubb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
254 "paddb %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
255 "pcmpgtb %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
256 "paddb %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
257 |
2413 | 258 "movq (%1, %2, 4), %%mm2 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
259 "psubb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
260 "paddb %%mm7, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
261 "pcmpgtb %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
262 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
263 |
2413 | 264 "movq (%%ebx), %%mm1 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
265 "psubb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
266 "paddb %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
267 "pcmpgtb %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
268 "paddb %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
269 |
2413 | 270 "movq (%%ebx, %2), %%mm2 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
271 "psubb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
272 "paddb %%mm7, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
273 "pcmpgtb %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
274 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
275 |
2413 | 276 "movq (%%ebx, %2, 2), %%mm1 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
277 "psubb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
278 "paddb %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
279 "pcmpgtb %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
280 "paddb %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
281 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
282 " \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
283 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
284 "psrlw $8, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
285 "paddb %%mm1, %%mm0 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
286 #ifdef HAVE_MMX2 |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
287 "pshufw $0xF9, %%mm0, %%mm1 \n\t" |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
288 "paddb %%mm1, %%mm0 \n\t" |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
289 "pshufw $0xFE, %%mm0, %%mm1 \n\t" |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
290 #else |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
291 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
292 "psrlq $16, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
293 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
294 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
295 "psrlq $32, %%mm0 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
296 #endif |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
297 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
298 "movd %%mm0, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
299 : "=r" (numEq) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
300 : "r" (src), "r" (stride) |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
301 : "%eax", "%ebx" |
2158
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 |
2413 | 304 numEq= (256 - numEq) &0xFF; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
305 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
306 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
307 for(y=0; y<BLOCK_SIZE-1; y++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
308 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
309 if(((src[0] - src[0+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
310 if(((src[1] - src[1+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
311 if(((src[2] - src[2+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
312 if(((src[3] - src[3+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
313 if(((src[4] - src[4+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
314 if(((src[5] - src[5+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
315 if(((src[6] - src[6+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
316 if(((src[7] - src[7+stride] + 1)&0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
317 src+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
318 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
319 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
320 /* if(abs(numEq - asmEq) > 0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
321 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
322 printf("\nasm:%d c:%d\n", asmEq, numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
323 for(int y=0; y<8; y++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
324 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
325 for(int x=0; x<8; x++) |
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 printf("%d ", temp[x + y*stride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
328 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
329 printf("\n"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
330 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
331 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
332 */ |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
333 // 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
|
334 return (numEq > vFlatnessThreshold) ? 1 : 0; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
335 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
336 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
337 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
|
338 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
339 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
340 int isOk; |
2246 | 341 src+= stride*3; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
342 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
343 // "int $3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
344 "movq (%1, %2), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
345 "movq (%1, %2, 8), %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
346 "movq %%mm0, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
347 "psubusb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
348 "psubusb %%mm2, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
349 "por %%mm1, %%mm0 \n\t" // ABS Diff |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
350 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
351 "movq pQPb, %%mm7 \n\t" // QP,..., QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
352 "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
353 "psubusb %%mm7, %%mm0 \n\t" // Diff <= 2QP -> 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
354 "pcmpeqd b00, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
355 "psrlq $16, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
356 "pcmpeqd bFF, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
357 // "movd %%mm0, (%1, %2, 4)\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
358 "movd %%mm0, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
359 : "=r" (isOk) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
360 : "r" (src), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
361 ); |
2416 | 362 return isOk; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
363 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
364 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
365 int isOk2= 1; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
366 int x; |
2246 | 367 src+= stride*3; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
368 for(x=0; x<BLOCK_SIZE; x++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
369 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
370 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
|
371 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
372 /* if(isOk && !isOk2 || !isOk && isOk2) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
373 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
374 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
|
375 for(int y=0; y<9; y++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
376 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
377 for(int x=0; x<8; x++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
378 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
379 printf("%d ", src[x + y*stride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
380 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
381 printf("\n"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
382 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
383 } */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
384 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
385 return isOk2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
386 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
387 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
388 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
389 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
390 /** |
2246 | 391 * Do a vertical low pass filter on the 8x16 block (only write to the 8x8 block in the middle) |
2221 | 392 * 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
|
393 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
394 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
|
395 { |
2159 | 396 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2246 | 397 src+= stride*3; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
398 asm volatile( //"movv %0 %1 %2\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
399 "pushl %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
400 "movq pQPb, %%mm0 \n\t" // QP,..., QP |
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 "movq (%0), %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
403 "movq (%0, %1), %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
404 "movq %%mm5, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
405 "movq %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
406 "psubusb %%mm6, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
407 "psubusb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
408 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
409 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
410 "pcmpeqb b00, %%mm2 \n\t" // diff <= QP -> FF |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
411 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
412 "pand %%mm2, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
413 "pandn %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
414 "por %%mm2, %%mm6 \n\t"// First Line to Filter |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
415 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
416 "movq (%0, %1, 8), %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
417 "leal (%0, %1, 4), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
418 "leal (%0, %1, 8), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
419 "subl %1, %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
420 "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
|
421 "movq (%0, %1, 8), %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
422 "movq %%mm5, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
423 "movq %%mm7, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
424 "psubusb %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
425 "psubusb %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
426 "por %%mm5, %%mm2 \n\t" // ABS Diff of lines |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
427 "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
428 "pcmpeqb b00, %%mm2 \n\t" // diff <= QP -> FF |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
429 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
430 "pand %%mm2, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
431 "pandn %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
432 "por %%mm2, %%mm7 \n\t" // First Line to Filter |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
433 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
434 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
435 // 1 2 3 4 5 6 7 8 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
436 // %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
|
437 // 6 4 2 2 1 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
438 // 6 4 4 2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
439 // 6 8 2 |
2246 | 440 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
441 "movq (%0, %1), %%mm0 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
442 "movq %%mm0, %%mm1 \n\t" // 1 |
2159 | 443 PAVGB(%%mm6, %%mm0) //1 1 /2 |
444 PAVGB(%%mm6, %%mm0) //3 1 /4 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
445 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
446 "movq (%0, %1, 4), %%mm2 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
447 "movq %%mm2, %%mm5 \n\t" // 1 |
2159 | 448 PAVGB((%%eax), %%mm2) // 11 /2 |
449 PAVGB((%0, %1, 2), %%mm2) // 211 /4 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
450 "movq %%mm2, %%mm3 \n\t" // 211 /4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
451 "movq (%0), %%mm4 \n\t" // 1 |
2159 | 452 PAVGB(%%mm4, %%mm3) // 4 211 /8 |
453 PAVGB(%%mm0, %%mm3) //642211 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
454 "movq %%mm3, (%0) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
455 // 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
|
456 "movq %%mm1, %%mm0 \n\t" // 1 |
2159 | 457 PAVGB(%%mm6, %%mm0) //1 1 /2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
458 "movq %%mm4, %%mm3 \n\t" // 1 |
2159 | 459 PAVGB((%0,%1,2), %%mm3) // 1 1 /2 |
460 PAVGB((%%eax,%1,2), %%mm5) // 11 /2 | |
461 PAVGB((%%eax), %%mm5) // 211 /4 | |
462 PAVGB(%%mm5, %%mm3) // 2 2211 /8 | |
463 PAVGB(%%mm0, %%mm3) //4242211 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
464 "movq %%mm3, (%0,%1) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
465 // mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9 |
2159 | 466 PAVGB(%%mm4, %%mm6) //11 /2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
467 "movq (%%ebx), %%mm0 \n\t" // 1 |
2159 | 468 PAVGB((%%eax, %1, 2), %%mm0) // 11/2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
469 "movq %%mm0, %%mm3 \n\t" // 11/2 |
2159 | 470 PAVGB(%%mm1, %%mm0) // 2 11/4 |
471 PAVGB(%%mm6, %%mm0) //222 11/8 | |
472 PAVGB(%%mm2, %%mm0) //22242211/16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
473 "movq (%0, %1, 2), %%mm2 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
474 "movq %%mm0, (%0, %1, 2) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
475 // 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
|
476 "movq (%%eax, %1, 4), %%mm0 \n\t" // 1 |
2159 | 477 PAVGB((%%ebx), %%mm0) // 11 /2 |
478 PAVGB(%%mm0, %%mm6) //11 11 /4 | |
479 PAVGB(%%mm1, %%mm4) // 11 /2 | |
480 PAVGB(%%mm2, %%mm1) // 11 /2 | |
481 PAVGB(%%mm1, %%mm6) //1122 11 /8 | |
482 PAVGB(%%mm5, %%mm6) //112242211 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
483 "movq (%%eax), %%mm5 \n\t" // 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
484 "movq %%mm6, (%%eax) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
485 // 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
|
486 "movq (%%eax, %1, 4), %%mm6 \n\t" // 1 |
2159 | 487 PAVGB(%%mm7, %%mm6) // 11 /2 |
488 PAVGB(%%mm4, %%mm6) // 11 11 /4 | |
489 PAVGB(%%mm3, %%mm6) // 11 2211 /8 | |
490 PAVGB(%%mm5, %%mm2) // 11 /2 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
491 "movq (%0, %1, 4), %%mm4 \n\t" // 1 |
2159 | 492 PAVGB(%%mm4, %%mm2) // 112 /4 |
493 PAVGB(%%mm2, %%mm6) // 112242211 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
494 "movq %%mm6, (%0, %1, 4) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
495 // mm0=7(11) mm1=2(11) mm2=3(112) mm3=6(11) mm4=5 mm5=4 mm7=9 |
2159 | 496 PAVGB(%%mm7, %%mm1) // 11 2 /4 |
497 PAVGB(%%mm4, %%mm5) // 11 /2 | |
498 PAVGB(%%mm5, %%mm0) // 11 11 /4 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
499 "movq (%%eax, %1, 2), %%mm6 \n\t" // 1 |
2159 | 500 PAVGB(%%mm6, %%mm1) // 11 4 2 /8 |
501 PAVGB(%%mm0, %%mm1) // 11224222 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
502 "movq %%mm1, (%%eax, %1, 2) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
503 // mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9 |
2159 | 504 PAVGB((%%ebx), %%mm2) // 112 4 /8 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
505 "movq (%%eax, %1, 4), %%mm0 \n\t" // 1 |
2159 | 506 PAVGB(%%mm0, %%mm6) // 1 1 /2 |
507 PAVGB(%%mm7, %%mm6) // 1 12 /4 | |
508 PAVGB(%%mm2, %%mm6) // 1122424 /4 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
509 "movq %%mm6, (%%ebx) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
510 // mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9 |
2159 | 511 PAVGB(%%mm7, %%mm5) // 11 2 /4 |
512 PAVGB(%%mm7, %%mm5) // 11 6 /8 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
513 |
2159 | 514 PAVGB(%%mm3, %%mm0) // 112 /4 |
515 PAVGB(%%mm0, %%mm5) // 112246 /16 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
516 "movq %%mm5, (%%eax, %1, 4) \n\t" // X |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
517 "popl %0\n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
518 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
519 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
520 : "r" (src), "r" (stride) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
521 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
522 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
523 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
524 const int l1= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
525 const int l2= stride + l1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
526 const int l3= stride + l2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
527 const int l4= stride + l3; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
528 const int l5= stride + l4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
529 const int l6= stride + l5; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
530 const int l7= stride + l6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
531 const int l8= stride + l7; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
532 const int l9= stride + l8; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
533 int x; |
2246 | 534 src+= stride*3; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
535 for(x=0; x<BLOCK_SIZE; x++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
536 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
537 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
|
538 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
|
539 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
540 int sums[9]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
541 sums[0] = first + src[l1]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
542 sums[1] = src[l1] + src[l2]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
543 sums[2] = src[l2] + src[l3]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
544 sums[3] = src[l3] + src[l4]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
545 sums[4] = src[l4] + src[l5]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
546 sums[5] = src[l5] + src[l6]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
547 sums[6] = src[l6] + src[l7]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
548 sums[7] = src[l7] + src[l8]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
549 sums[8] = src[l8] + last; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
550 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
551 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
|
552 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
|
553 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
|
554 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
|
555 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
|
556 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
|
557 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
|
558 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
|
559 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
560 src++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
561 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
562 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
563 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
564 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
565 |
2159 | 566 /** |
567 * Experimental implementation of the filter (Algorithm 1) described in a paper from Ramkishor & Karandikar | |
568 * values are correctly clipped (MMX2) | |
569 * values are wraparound (C) | |
570 * conclusion: its fast, but introduces ugly horizontal patterns if there is a continious gradient | |
571 0 8 16 24 | |
572 x = 8 | |
573 x/2 = 4 | |
574 x/8 = 1 | |
575 1 12 12 23 | |
576 */ | |
2179 | 577 static inline void vertRK1Filter(uint8_t *src, int stride, int QP) |
2159 | 578 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
579 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2246 | 580 src+= stride*3; |
2159 | 581 // FIXME rounding |
582 asm volatile( | |
583 "pxor %%mm7, %%mm7 \n\t" // 0 | |
584 "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE | |
585 "leal (%0, %1), %%eax \n\t" | |
586 "leal (%%eax, %1, 4), %%ebx \n\t" | |
587 // 0 1 2 3 4 5 6 7 8 9 | |
588 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 | |
589 "movq pQPb, %%mm0 \n\t" // QP,..., QP | |
590 "movq %%mm0, %%mm1 \n\t" // QP,..., QP | |
591 "paddusb b02, %%mm0 \n\t" | |
592 "psrlw $2, %%mm0 \n\t" | |
593 "pand b3F, %%mm0 \n\t" // QP/4,..., QP/4 | |
594 "paddusb %%mm1, %%mm0 \n\t" // QP*1.25 ... | |
595 "movq (%0, %1, 4), %%mm2 \n\t" // line 4 | |
596 "movq (%%ebx), %%mm3 \n\t" // line 5 | |
597 "movq %%mm2, %%mm4 \n\t" // line 4 | |
598 "pcmpeqb %%mm5, %%mm5 \n\t" // -1 | |
599 "pxor %%mm2, %%mm5 \n\t" // -line 4 - 1 | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
600 PAVGB(%%mm3, %%mm5) |
2159 | 601 "paddb %%mm6, %%mm5 \n\t" // (l5-l4)/2 |
602 "psubusb %%mm3, %%mm4 \n\t" | |
603 "psubusb %%mm2, %%mm3 \n\t" | |
604 "por %%mm3, %%mm4 \n\t" // |l4 - l5| | |
605 "psubusb %%mm0, %%mm4 \n\t" | |
606 "pcmpeqb %%mm7, %%mm4 \n\t" | |
607 "pand %%mm4, %%mm5 \n\t" // d/2 | |
608 | |
609 // "paddb %%mm6, %%mm2 \n\t" // line 4 + 0x80 | |
610 "paddb %%mm5, %%mm2 \n\t" | |
611 // "psubb %%mm6, %%mm2 \n\t" | |
612 "movq %%mm2, (%0,%1, 4) \n\t" | |
613 | |
614 "movq (%%ebx), %%mm2 \n\t" | |
615 // "paddb %%mm6, %%mm2 \n\t" // line 5 + 0x80 | |
616 "psubb %%mm5, %%mm2 \n\t" | |
617 // "psubb %%mm6, %%mm2 \n\t" | |
618 "movq %%mm2, (%%ebx) \n\t" | |
619 | |
620 "paddb %%mm6, %%mm5 \n\t" | |
621 "psrlw $2, %%mm5 \n\t" | |
622 "pand b3F, %%mm5 \n\t" | |
623 "psubb b20, %%mm5 \n\t" // (l5-l4)/8 | |
624 | |
625 "movq (%%eax, %1, 2), %%mm2 \n\t" | |
626 "paddb %%mm6, %%mm2 \n\t" // line 3 + 0x80 | |
627 "paddsb %%mm5, %%mm2 \n\t" | |
628 "psubb %%mm6, %%mm2 \n\t" | |
629 "movq %%mm2, (%%eax, %1, 2) \n\t" | |
630 | |
631 "movq (%%ebx, %1), %%mm2 \n\t" | |
632 "paddb %%mm6, %%mm2 \n\t" // line 6 + 0x80 | |
633 "psubsb %%mm5, %%mm2 \n\t" | |
634 "psubb %%mm6, %%mm2 \n\t" | |
635 "movq %%mm2, (%%ebx, %1) \n\t" | |
636 | |
637 : | |
638 : "r" (src), "r" (stride) | |
639 : "%eax", "%ebx" | |
640 ); | |
641 #else | |
642 const int l1= stride; | |
643 const int l2= stride + l1; | |
644 const int l3= stride + l2; | |
645 const int l4= stride + l3; | |
646 const int l5= stride + l4; | |
647 const int l6= stride + l5; | |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
648 // const int l7= stride + l6; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
649 // const int l8= stride + l7; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
650 // const int l9= stride + l8; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
651 int x; |
2246 | 652 src+= stride*3; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
653 for(x=0; x<BLOCK_SIZE; x++) |
2159 | 654 { |
655 if(ABS(src[l4]-src[l5]) < QP + QP/4) | |
656 { | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
657 int v = (src[l5] - src[l4]); |
2159 | 658 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
659 src[l3] +=v/8; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
660 src[l4] +=v/2; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
661 src[l5] -=v/2; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
662 src[l6] -=v/8; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
663 |
2159 | 664 } |
665 src++; | |
666 } | |
667 | |
668 #endif | |
669 } | |
670 | |
671 /** | |
672 * Experimental Filter 1 | |
2179 | 673 * will not damage linear gradients |
674 * 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
|
675 * 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
|
676 * MMX2 version does correct clipping C version doesnt |
2159 | 677 */ |
678 static inline void vertX1Filter(uint8_t *src, int stride, int QP) | |
679 { | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
680 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2246 | 681 src+= stride*3; |
682 | |
2159 | 683 asm volatile( |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
684 "pxor %%mm7, %%mm7 \n\t" // 0 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
685 // "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
686 "leal (%0, %1), %%eax \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
687 "leal (%%eax, %1, 4), %%ebx \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
688 // 0 1 2 3 4 5 6 7 8 9 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
689 // %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
|
690 "movq (%%eax, %1, 2), %%mm0 \n\t" // line 3 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
691 "movq (%0, %1, 4), %%mm1 \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
692 "movq %%mm1, %%mm2 \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
693 "psubusb %%mm0, %%mm1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
694 "psubusb %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
695 "por %%mm1, %%mm0 \n\t" // |l2 - l3| |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
696 "movq (%%ebx), %%mm3 \n\t" // line 5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
697 "movq (%%ebx, %1), %%mm4 \n\t" // line 6 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
698 "movq %%mm3, %%mm5 \n\t" // line 5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
699 "psubusb %%mm4, %%mm3 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
700 "psubusb %%mm5, %%mm4 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
701 "por %%mm4, %%mm3 \n\t" // |l5 - l6| |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
702 PAVGB(%%mm3, %%mm0) // (|l2 - l3| + |l5 - l6|)/2 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
703 "movq %%mm2, %%mm1 \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
704 "psubusb %%mm5, %%mm2 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
705 "movq %%mm2, %%mm4 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
706 "pcmpeqb %%mm7, %%mm2 \n\t" // (l4 - l5) <= 0 ? -1 : 0 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
707 "psubusb %%mm1, %%mm5 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
708 "por %%mm5, %%mm4 \n\t" // |l4 - l5| |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
709 "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
|
710 "movq %%mm4, %%mm3 \n\t" // d |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
711 "psubusb pQPb, %%mm4 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
712 "pcmpeqb %%mm7, %%mm4 \n\t" // d <= QP ? -1 : 0 |
2179 | 713 "psubusb b01, %%mm3 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
714 "pand %%mm4, %%mm3 \n\t" // d <= QP ? d : 0 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
715 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
716 PAVGB(%%mm7, %%mm3) // d/2 |
2179 | 717 "movq %%mm3, %%mm1 \n\t" // d/2 |
718 PAVGB(%%mm7, %%mm3) // d/4 | |
719 PAVGB(%%mm1, %%mm3) // 3*d/8 | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
720 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
721 "movq (%0, %1, 4), %%mm0 \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
722 "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
|
723 "psubusb %%mm3, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
724 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
725 "movq %%mm0, (%0, %1, 4) \n\t" // line 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
726 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
727 "movq (%%ebx), %%mm0 \n\t" // line 5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
728 "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
|
729 "paddusb %%mm3, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
730 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
731 "movq %%mm0, (%%ebx) \n\t" // line 5 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
732 |
2179 | 733 PAVGB(%%mm7, %%mm1) // d/4 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
734 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
735 "movq (%%eax, %1, 2), %%mm0 \n\t" // line 3 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
736 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4 |
2179 | 737 "psubusb %%mm1, %%mm0 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
738 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
739 "movq %%mm0, (%%eax, %1, 2) \n\t" // line 3 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
740 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
741 "movq (%%ebx, %1), %%mm0 \n\t" // line 6 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
742 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5 |
2179 | 743 "paddusb %%mm1, %%mm0 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
744 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
745 "movq %%mm0, (%%ebx, %1) \n\t" // line 6 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
746 |
2179 | 747 PAVGB(%%mm7, %%mm1) // d/8 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
748 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
749 "movq (%%eax, %1), %%mm0 \n\t" // line 2 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
750 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l2-1 : l2 |
2179 | 751 "psubusb %%mm1, %%mm0 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
752 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
753 "movq %%mm0, (%%eax, %1) \n\t" // line 2 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
754 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
755 "movq (%%ebx, %1, 2), %%mm0 \n\t" // line 7 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
756 "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l7-1 : l7 |
2179 | 757 "paddusb %%mm1, %%mm0 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
758 "pxor %%mm2, %%mm0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
759 "movq %%mm0, (%%ebx, %1, 2) \n\t" // line 7 |
2159 | 760 |
761 : | |
762 : "r" (src), "r" (stride) | |
763 : "%eax", "%ebx" | |
764 ); | |
765 #else | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
766 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
767 const int l1= stride; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
768 const int l2= stride + l1; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
769 const int l3= stride + l2; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
770 const int l4= stride + l3; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
771 const int l5= stride + l4; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
772 const int l6= stride + l5; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
773 const int l7= stride + l6; |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
774 // const int l8= stride + l7; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
775 // const int l9= stride + l8; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
776 int x; |
2246 | 777 |
778 src+= stride*3; | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
779 for(x=0; x<BLOCK_SIZE; x++) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
780 { |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
781 int a= src[l3] - src[l4]; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
782 int b= src[l4] - src[l5]; |
2179 | 783 int c= src[l5] - src[l6]; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
784 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
785 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
|
786 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
787 if(d < QP) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
788 { |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
789 int v = d * SIGN(-b); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
790 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
791 src[l2] +=v/8; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
792 src[l3] +=v/4; |
2179 | 793 src[l4] +=3*v/8; |
794 src[l5] -=3*v/8; | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
795 src[l6] -=v/4; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
796 src[l7] -=v/8; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
797 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
798 } |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
799 src++; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
800 } |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
801 /* |
2159 | 802 const int l1= stride; |
803 const int l2= stride + l1; | |
804 const int l3= stride + l2; | |
805 const int l4= stride + l3; | |
806 const int l5= stride + l4; | |
807 const int l6= stride + l5; | |
808 const int l7= stride + l6; | |
809 const int l8= stride + l7; | |
810 const int l9= stride + l8; | |
811 for(int x=0; x<BLOCK_SIZE; x++) | |
812 { | |
813 int v2= src[l2]; | |
814 int v3= src[l3]; | |
815 int v4= src[l4]; | |
816 int v5= src[l5]; | |
817 int v6= src[l6]; | |
818 int v7= src[l7]; | |
819 | |
820 if(ABS(v4-v5)<QP && ABS(v4-v5) - (ABS(v3-v4) + ABS(v5-v6))>0 ) | |
821 { | |
822 src[l3] = (6*v2 + 4*v3 + 3*v4 + 2*v5 + v6 )/16; | |
823 src[l4] = (3*v2 + 3*v3 + 4*v4 + 3*v5 + 2*v6 + v7 )/16; | |
824 src[l5] = (1*v2 + 2*v3 + 3*v4 + 4*v5 + 3*v6 + 3*v7)/16; | |
825 src[l6] = ( 1*v3 + 2*v4 + 3*v5 + 4*v6 + 6*v7)/16; | |
826 } | |
827 src++; | |
828 } | |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
829 */ |
2159 | 830 #endif |
831 } | |
832 | |
2179 | 833 /** |
834 * Experimental Filter 1 (Horizontal) | |
835 * will not damage linear gradients | |
836 * Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter | |
837 * can only smooth blocks at the expected locations (it cant smooth them if they did move) | |
838 * MMX2 version does correct clipping C version doesnt | |
839 * not identical with the vertical one | |
840 */ | |
841 static inline void horizX1Filter(uint8_t *src, int stride, int QP) | |
842 { | |
843 int y; | |
844 static uint64_t *lut= NULL; | |
845 if(lut==NULL) | |
846 { | |
847 int i; | |
848 lut= (uint64_t*)memalign(8, 256*8); | |
849 for(i=0; i<256; i++) | |
850 { | |
851 int v= i < 128 ? 2*i : 2*(i-256); | |
852 /* | |
853 //Simulate 112242211 9-Tap filter | |
854 uint64_t a= (v/16) & 0xFF; | |
855 uint64_t b= (v/8) & 0xFF; | |
856 uint64_t c= (v/4) & 0xFF; | |
857 uint64_t d= (3*v/8) & 0xFF; | |
858 */ | |
859 //Simulate piecewise linear interpolation | |
860 uint64_t a= (v/16) & 0xFF; | |
861 uint64_t b= (v*3/16) & 0xFF; | |
862 uint64_t c= (v*5/16) & 0xFF; | |
863 uint64_t d= (7*v/16) & 0xFF; | |
864 uint64_t A= (0x100 - a)&0xFF; | |
865 uint64_t B= (0x100 - b)&0xFF; | |
866 uint64_t C= (0x100 - c)&0xFF; | |
867 uint64_t D= (0x100 - c)&0xFF; | |
868 | |
869 lut[i] = (a<<56) | (b<<48) | (c<<40) | (d<<32) | | |
870 (D<<24) | (C<<16) | (B<<8) | (A); | |
871 //lut[i] = (v<<32) | (v<<24); | |
872 } | |
873 } | |
874 | |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
875 #if 0 |
2179 | 876 asm volatile( |
877 "pxor %%mm7, %%mm7 \n\t" // 0 | |
878 // "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE | |
879 "leal (%0, %1), %%eax \n\t" | |
880 "leal (%%eax, %1, 4), %%ebx \n\t" | |
881 | |
882 "movq b80, %%mm6 \n\t" | |
2180 | 883 "movd pQPb, %%mm5 \n\t" // QP |
2179 | 884 "movq %%mm5, %%mm4 \n\t" |
885 "paddusb %%mm5, %%mm5 \n\t" // 2QP | |
886 "paddusb %%mm5, %%mm4 \n\t" // 3QP | |
887 "pxor %%mm5, %%mm5 \n\t" // 0 | |
888 "psubb %%mm4, %%mm5 \n\t" // -3QP | |
889 "por bm11111110, %%mm5 \n\t" // ...,FF,FF,-3QP | |
890 "psllq $24, %%mm5 \n\t" | |
891 | |
892 // 0 1 2 3 4 5 6 7 8 9 | |
893 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 | |
894 | |
895 #define HX1old(a) \ | |
896 "movd " #a ", %%mm0 \n\t"\ | |
897 "movd 4" #a ", %%mm1 \n\t"\ | |
898 "punpckldq %%mm1, %%mm0 \n\t"\ | |
899 "movq %%mm0, %%mm1 \n\t"\ | |
900 "movq %%mm0, %%mm2 \n\t"\ | |
901 "psrlq $8, %%mm1 \n\t"\ | |
902 "psubusb %%mm1, %%mm2 \n\t"\ | |
903 "psubusb %%mm0, %%mm1 \n\t"\ | |
904 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ | |
905 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
906 "pshufw $0x00, %%mm1, %%mm3 \n\t" /* p´5 = |p1 - p2| */\ | |
907 PAVGB(%%mm1, %%mm3) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\ | |
908 "psrlq $16, %%mm3 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\ | |
909 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\ | |
910 "paddb %%mm5, %%mm1 \n\t"\ | |
911 "psubusb %%mm5, %%mm1 \n\t"\ | |
912 PAVGB(%%mm7, %%mm1)\ | |
913 "pxor %%mm2, %%mm1 \n\t"\ | |
914 "psubb %%mm2, %%mm1 \n\t"\ | |
915 "psrlq $24, %%mm1 \n\t"\ | |
916 "movd %%mm1, %%ecx \n\t"\ | |
917 "paddb %%mm6, %%mm0 \n\t"\ | |
918 "paddsb (%3, %%ecx, 8), %%mm0 \n\t"\ | |
919 "paddb %%mm6, %%mm0 \n\t"\ | |
920 "movq %%mm0, " #a " \n\t"\ | |
921 | |
922 /* | |
923 HX1old((%0)) | |
924 HX1old((%%eax)) | |
925 HX1old((%%eax, %1)) | |
926 HX1old((%%eax, %1, 2)) | |
927 HX1old((%0, %1, 4)) | |
928 HX1old((%%ebx)) | |
929 HX1old((%%ebx, %1)) | |
930 HX1old((%%ebx, %1, 2)) | |
931 */ | |
932 | |
933 //FIXME add some comments, its unreadable ... | |
934 #define HX1b(a, c, b, d) \ | |
935 "movd " #a ", %%mm0 \n\t"\ | |
936 "movd 4" #a ", %%mm1 \n\t"\ | |
937 "punpckldq %%mm1, %%mm0 \n\t"\ | |
938 "movd " #b ", %%mm4 \n\t"\ | |
939 "movq %%mm0, %%mm1 \n\t"\ | |
940 "movq %%mm0, %%mm2 \n\t"\ | |
941 "psrlq $8, %%mm1 \n\t"\ | |
942 "movd 4" #b ", %%mm3 \n\t"\ | |
943 "psubusb %%mm1, %%mm2 \n\t"\ | |
944 "psubusb %%mm0, %%mm1 \n\t"\ | |
945 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ | |
946 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
947 "punpckldq %%mm3, %%mm4 \n\t"\ | |
948 "movq %%mm1, %%mm3 \n\t"\ | |
949 "psllq $32, %%mm3 \n\t" /* p´5 = |p1 - p2| */\ | |
950 PAVGB(%%mm1, %%mm3) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\ | |
951 "paddb %%mm6, %%mm0 \n\t"\ | |
952 "psrlq $16, %%mm3 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\ | |
953 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\ | |
954 "movq %%mm4, %%mm3 \n\t"\ | |
955 "paddb %%mm5, %%mm1 \n\t"\ | |
956 "psubusb %%mm5, %%mm1 \n\t"\ | |
957 "psrlq $8, %%mm3 \n\t"\ | |
958 PAVGB(%%mm7, %%mm1)\ | |
959 "pxor %%mm2, %%mm1 \n\t"\ | |
960 "psubb %%mm2, %%mm1 \n\t"\ | |
961 "movq %%mm4, %%mm2 \n\t"\ | |
962 "psrlq $24, %%mm1 \n\t"\ | |
963 "psubusb %%mm3, %%mm2 \n\t"\ | |
964 "movd %%mm1, %%ecx \n\t"\ | |
965 "psubusb %%mm4, %%mm3 \n\t"\ | |
2180 | 966 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\ |
2179 | 967 "por %%mm2, %%mm3 \n\t" /* p´x = |px - p(x+1)| */\ |
968 "paddb %%mm6, %%mm0 \n\t"\ | |
969 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
970 "movq %%mm3, %%mm1 \n\t"\ | |
971 "psllq $32, %%mm1 \n\t" /* p´5 = |p1 - p2| */\ | |
972 "movq %%mm0, " #a " \n\t"\ | |
973 PAVGB(%%mm3, %%mm1) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\ | |
974 "paddb %%mm6, %%mm4 \n\t"\ | |
975 "psrlq $16, %%mm1 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\ | |
976 "psubusb %%mm1, %%mm3 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\ | |
977 "paddb %%mm5, %%mm3 \n\t"\ | |
978 "psubusb %%mm5, %%mm3 \n\t"\ | |
979 PAVGB(%%mm7, %%mm3)\ | |
980 "pxor %%mm2, %%mm3 \n\t"\ | |
981 "psubb %%mm2, %%mm3 \n\t"\ | |
982 "psrlq $24, %%mm3 \n\t"\ | |
983 "movd " #c ", %%mm0 \n\t"\ | |
984 "movd 4" #c ", %%mm1 \n\t"\ | |
985 "punpckldq %%mm1, %%mm0 \n\t"\ | |
986 "paddb %%mm6, %%mm0 \n\t"\ | |
2180 | 987 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\ |
2179 | 988 "paddb %%mm6, %%mm0 \n\t"\ |
989 "movq %%mm0, " #c " \n\t"\ | |
990 "movd %%mm3, %%ecx \n\t"\ | |
991 "movd " #d ", %%mm0 \n\t"\ | |
2180 | 992 "paddsb (%2, %%ecx, 8), %%mm4 \n\t"\ |
2179 | 993 "movd 4" #d ", %%mm1 \n\t"\ |
994 "paddb %%mm6, %%mm4 \n\t"\ | |
995 "punpckldq %%mm1, %%mm0 \n\t"\ | |
996 "movq %%mm4, " #b " \n\t"\ | |
997 "paddb %%mm6, %%mm0 \n\t"\ | |
2180 | 998 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\ |
2179 | 999 "paddb %%mm6, %%mm0 \n\t"\ |
1000 "movq %%mm0, " #d " \n\t"\ | |
1001 | |
1002 HX1b((%0),(%%eax),(%%eax, %1),(%%eax, %1, 2)) | |
1003 HX1b((%0, %1, 4),(%%ebx),(%%ebx, %1),(%%ebx, %1, 2)) | |
1004 | |
1005 | |
1006 : | |
2180 | 1007 : "r" (src), "r" (stride), "r" (lut) |
2179 | 1008 : "%eax", "%ebx", "%ecx" |
1009 ); | |
1010 #else | |
1011 | |
1012 //FIXME (has little in common with the mmx2 version) | |
1013 for(y=0; y<BLOCK_SIZE; y++) | |
1014 { | |
1015 int a= src[1] - src[2]; | |
1016 int b= src[3] - src[4]; | |
1017 int c= src[5] - src[6]; | |
1018 | |
1019 int d= MAX(ABS(b) - (ABS(a) + ABS(c))/2, 0); | |
1020 | |
1021 if(d < QP) | |
1022 { | |
1023 int v = d * SIGN(-b); | |
1024 | |
1025 src[1] +=v/8; | |
1026 src[2] +=v/4; | |
1027 src[3] +=3*v/8; | |
1028 src[4] -=3*v/8; | |
1029 src[5] -=v/4; | |
1030 src[6] -=v/8; | |
1031 | |
1032 } | |
1033 src+=stride; | |
1034 } | |
1035 #endif | |
1036 } | |
1037 | |
2159 | 1038 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1039 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
|
1040 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1041 #ifdef HAVE_MMX |
2246 | 1042 src+= stride*4; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1043 //FIXME try pmul for *5 stuff |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1044 // src[0]=0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1045 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1046 "pxor %%mm7, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1047 "leal (%0, %1), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1048 "leal (%%eax, %1, 4), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1049 // 0 1 2 3 4 5 6 7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1050 // %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
|
1051 // %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
|
1052 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1053 "movq (%0), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1054 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1055 "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1056 "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1057 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1058 "movq (%%eax), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1059 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1060 "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1061 "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1062 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1063 "movq (%%eax, %1), %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1064 "movq %%mm4, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1065 "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1066 "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1067 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1068 "paddw %%mm0, %%mm0 \n\t" // 2L0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1069 "paddw %%mm1, %%mm1 \n\t" // 2H0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1070 "psubw %%mm4, %%mm2 \n\t" // L1 - L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1071 "psubw %%mm5, %%mm3 \n\t" // H1 - H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1072 "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1073 "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1074 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1075 "psllw $2, %%mm2 \n\t" // 4L1 - 4L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1076 "psllw $2, %%mm3 \n\t" // 4H1 - 4H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1077 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1078 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1079 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1080 "movq (%%eax, %1, 2), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1081 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1082 "punpcklbw %%mm7, %%mm2 \n\t" // L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1083 "punpckhbw %%mm7, %%mm3 \n\t" // H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1084 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1085 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1086 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1087 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1088 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1089 "movq %%mm0, temp0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1090 "movq %%mm1, temp1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1091 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1092 "movq (%0, %1, 4), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1093 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1094 "punpcklbw %%mm7, %%mm0 \n\t" // L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1095 "punpckhbw %%mm7, %%mm1 \n\t" // H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1096 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1097 "psubw %%mm0, %%mm2 \n\t" // L3 - L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1098 "psubw %%mm1, %%mm3 \n\t" // H3 - H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1099 "movq %%mm2, temp2 \n\t" // L3 - L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1100 "movq %%mm3, temp3 \n\t" // H3 - H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1101 "paddw %%mm4, %%mm4 \n\t" // 2L2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1102 "paddw %%mm5, %%mm5 \n\t" // 2H2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1103 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1104 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4 |
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 "psllw $2, %%mm2 \n\t" // 4L3 - 4L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1107 "psllw $2, %%mm3 \n\t" // 4H3 - 4H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1108 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1109 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1110 //50 opcodes so far |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1111 "movq (%%ebx), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1112 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1113 "punpcklbw %%mm7, %%mm2 \n\t" // L5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1114 "punpckhbw %%mm7, %%mm3 \n\t" // H5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1115 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1116 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1117 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1118 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1119 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1120 "movq (%%ebx, %1), %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1121 "punpcklbw %%mm7, %%mm6 \n\t" // L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1122 "psubw %%mm6, %%mm2 \n\t" // L5 - L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1123 "movq (%%ebx, %1), %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1124 "punpckhbw %%mm7, %%mm6 \n\t" // H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1125 "psubw %%mm6, %%mm3 \n\t" // H5 - H6 |
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 "paddw %%mm0, %%mm0 \n\t" // 2L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1128 "paddw %%mm1, %%mm1 \n\t" // 2H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1129 "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1130 "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1131 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1132 "psllw $2, %%mm2 \n\t" // 4L5 - 4L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1133 "psllw $2, %%mm3 \n\t" // 4H5 - 4H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1134 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1135 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1136 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1137 "movq (%%ebx, %1, 2), %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1138 "movq %%mm2, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1139 "punpcklbw %%mm7, %%mm2 \n\t" // L7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1140 "punpckhbw %%mm7, %%mm3 \n\t" // H7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1141 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1142 "paddw %%mm2, %%mm2 \n\t" // 2L7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1143 "paddw %%mm3, %%mm3 \n\t" // 2H7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1144 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1145 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1146 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1147 "movq temp0, %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1148 "movq temp1, %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1149 //FIXME pxor, psubw, pmax for abs |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1150 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1151 "pcmpgtw %%mm0, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1152 "pxor %%mm6, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1153 "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1154 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1155 "pcmpgtw %%mm1, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1156 "pxor %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1157 "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1158 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1159 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1160 "pcmpgtw %%mm2, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1161 "pxor %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1162 "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1163 "movq %%mm7, %%mm6 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1164 "pcmpgtw %%mm3, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1165 "pxor %%mm6, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1166 "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3| |
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 #ifdef HAVE_MMX2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1169 "pminsw %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1170 "pminsw %%mm3, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1171 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1172 "movq %%mm0, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1173 "psubusw %%mm2, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1174 "psubw %%mm6, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1175 "movq %%mm1, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1176 "psubusw %%mm3, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1177 "psubw %%mm6, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1178 #endif |
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 %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1182 "pxor %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1183 "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1184 "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1185 "pxor %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1186 "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1187 // 100 opcodes |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1188 "movd %2, %%mm2 \n\t" // QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1189 "punpcklwd %%mm2, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1190 "punpcklwd %%mm2, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1191 "psllw $3, %%mm2 \n\t" // 8QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1192 "movq %%mm2, %%mm3 \n\t" // 8QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1193 "pcmpgtw %%mm4, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1194 "pcmpgtw %%mm5, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1195 "pand %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1196 "pand %%mm3, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1197 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1198 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1199 "psubusw %%mm0, %%mm4 \n\t" // hd |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1200 "psubusw %%mm1, %%mm5 \n\t" // ld |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1201 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1202 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1203 "movq w05, %%mm2 \n\t" // 5 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1204 "pmullw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1205 "pmullw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1206 "movq w20, %%mm2 \n\t" // 32 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1207 "paddw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1208 "paddw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1209 "psrlw $6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1210 "psrlw $6, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1211 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1212 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1213 "movq w06, %%mm2 \n\t" // 6 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1214 "paddw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1215 "paddw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1216 "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
|
1217 //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
|
1218 "pmulhw %%mm2, %%mm4 \n\t" // hd/13 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1219 "pmulhw %%mm2, %%mm5 \n\t" // ld/13 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1220 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1221 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1222 "movq temp2, %%mm0 \n\t" // L3 - L4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1223 "movq temp3, %%mm1 \n\t" // H3 - H4 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1224 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1225 "pxor %%mm2, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1226 "pxor %%mm3, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1227 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1228 "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1229 "pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1230 "pxor %%mm2, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1231 "pxor %%mm3, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1232 "psubw %%mm2, %%mm0 \n\t" // |L3-L4| |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1233 "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
|
1234 "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
|
1235 "psrlw $1, %%mm1 \n\t" // |H3 - H4|/2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1236 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1237 "pxor %%mm6, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1238 "pxor %%mm7, %%mm3 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1239 "pand %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1240 "pand %%mm3, %%mm5 \n\t" |
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 #ifdef HAVE_MMX2 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1243 "pminsw %%mm0, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1244 "pminsw %%mm1, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1245 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1246 "movq %%mm4, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1247 "psubusw %%mm0, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1248 "psubw %%mm2, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1249 "movq %%mm5, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1250 "psubusw %%mm1, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1251 "psubw %%mm2, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1252 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1253 "pxor %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1254 "pxor %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1255 "psubw %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1256 "psubw %%mm7, %%mm5 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1257 "packsswb %%mm5, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1258 "movq (%%eax, %1, 2), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1259 "paddb %%mm4, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1260 "movq %%mm0, (%%eax, %1, 2) \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1261 "movq (%0, %1, 4), %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1262 "psubb %%mm4, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1263 "movq %%mm0, (%0, %1, 4) \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1264 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1265 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1266 : "r" (src), "r" (stride), "r" (QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1267 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1268 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1269 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1270 const int l1= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1271 const int l2= stride + l1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1272 const int l3= stride + l2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1273 const int l4= stride + l3; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1274 const int l5= stride + l4; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1275 const int l6= stride + l5; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1276 const int l7= stride + l6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1277 const int l8= stride + l7; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1278 // const int l9= stride + l8; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1279 int x; |
2246 | 1280 src+= stride*3; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1281 for(x=0; x<BLOCK_SIZE; x++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1282 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1283 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
|
1284 if(ABS(middleEnergy) < 8*QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1285 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1286 const int q=(src[l4] - src[l5])/2; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1287 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
|
1288 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
|
1289 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1290 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1291 d= MAX(d, 0); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1292 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1293 d= (5*d + 32) >> 6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1294 d*= SIGN(-middleEnergy); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1295 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1296 if(q>0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1297 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1298 d= d<0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1299 d= d>q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1300 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1301 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1302 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1303 d= d>0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1304 d= d<q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1305 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1306 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1307 src[l4]-= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1308 src[l5]+= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1309 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1310 src++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1311 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1312 #endif |
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 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1315 //FIXME? |255-0| = 1 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1316 /** |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1317 * Check if the given 8x8 Block is mostly "flat" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1318 */ |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1319 static inline int isHorizDC(uint8_t src[], int stride) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1320 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1321 // src++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1322 int numEq= 0; |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1323 #if 0 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1324 asm volatile ( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1325 // "int $3 \n\t" |
2416 | 1326 "leal (%1, %2), %%ecx \n\t" |
1327 "leal (%%ecx, %2, 4), %%ebx \n\t" | |
1328 // 0 1 2 3 4 5 6 7 8 9 | |
1329 // %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
|
1330 "movq b7E, %%mm7 \n\t" // mm7 = 0x7F |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1331 "movq b7C, %%mm6 \n\t" // mm6 = 0x7D |
2416 | 1332 "pxor %%mm0, %%mm0 \n\t" |
1333 "movl %1, %%eax \n\t" | |
1334 "andl $0x1F, %%eax \n\t" | |
1335 "cmpl $24, %%eax \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1336 "leal tempBlock, %%eax \n\t" |
2416 | 1337 "jb 1f \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1338 |
2416 | 1339 #define HDC_CHECK_AND_CPY(src, dst) \ |
1340 "movd " #src ", %%mm2 \n\t"\ | |
1341 "punpckldq 4" #src ", %%mm2 \n\t" /* (%1) */\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1342 "movq %%mm2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1343 "psrlq $8, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1344 "psubb %%mm1, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1345 "paddb %%mm7, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1346 "pcmpgtb %%mm6, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1347 "paddb %%mm2, %%mm0 \n\t"\ |
2416 | 1348 "movq %%mm1," #dst "(%%eax) \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1349 |
2416 | 1350 HDC_CHECK_AND_CPY((%1),0) |
1351 HDC_CHECK_AND_CPY((%%ecx),8) | |
1352 HDC_CHECK_AND_CPY((%%ecx, %2),16) | |
1353 HDC_CHECK_AND_CPY((%%ecx, %2, 2),24) | |
1354 HDC_CHECK_AND_CPY((%1, %2, 4),32) | |
1355 HDC_CHECK_AND_CPY((%%ebx),40) | |
1356 HDC_CHECK_AND_CPY((%%ebx, %2),48) | |
1357 HDC_CHECK_AND_CPY((%%ebx, %2, 2),56) | |
1358 "jmp 2f \n\t" | |
1359 "1: \n\t" | |
1360 // src does not cross a 32 byte cache line so dont waste time with alignment | |
1361 #define HDC_CHECK_AND_CPY2(src, dst) \ | |
1362 "movq " #src ", %%mm2 \n\t"\ | |
1363 "movq " #src ", %%mm1 \n\t"\ | |
1364 "psrlq $8, %%mm2 \n\t"\ | |
1365 "psubb %%mm1, %%mm2 \n\t"\ | |
1366 "paddb %%mm7, %%mm2 \n\t"\ | |
1367 "pcmpgtb %%mm6, %%mm2 \n\t"\ | |
1368 "paddb %%mm2, %%mm0 \n\t"\ | |
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_CPY2((%1),0) |
1372 HDC_CHECK_AND_CPY2((%%ecx),8) | |
1373 HDC_CHECK_AND_CPY2((%%ecx, %2),16) | |
1374 HDC_CHECK_AND_CPY2((%%ecx, %2, 2),24) | |
1375 HDC_CHECK_AND_CPY2((%1, %2, 4),32) | |
1376 HDC_CHECK_AND_CPY2((%%ebx),40) | |
1377 HDC_CHECK_AND_CPY2((%%ebx, %2),48) | |
1378 HDC_CHECK_AND_CPY2((%%ebx, %2, 2),56) | |
1379 "2: \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1380 "psllq $8, %%mm0 \n\t" // remove dummy value |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1381 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1382 "psrlw $8, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1383 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1384 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1385 "psrlq $16, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1386 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1387 "movq %%mm0, %%mm1 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1388 "psrlq $32, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1389 "paddb %%mm1, %%mm0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1390 "movd %%mm0, %0 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1391 : "=r" (numEq) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1392 : "r" (src), "r" (stride) |
2416 | 1393 : "%eax", "%ebx", "%ecx" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1394 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1395 // printf("%d\n", numEq); |
2416 | 1396 numEq= (256 - numEq) &0xFF; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1397 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1398 int y; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1399 for(y=0; y<BLOCK_SIZE; y++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1400 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1401 if(((src[0] - src[1] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1402 if(((src[1] - src[2] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1403 if(((src[2] - src[3] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1404 if(((src[3] - src[4] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1405 if(((src[4] - src[5] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1406 if(((src[5] - src[6] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1407 if(((src[6] - src[7] + 1) & 0xFFFF) < 3) numEq++; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1408 src+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1409 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1410 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1411 /* if(abs(numEq - asmEq) > 0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1412 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1413 // printf("\nasm:%d c:%d\n", asmEq, numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1414 for(int y=0; y<8; y++) |
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 for(int x=0; x<8; x++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1417 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1418 printf("%d ", src[x + y*stride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1419 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1420 printf("\n"); |
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 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1423 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1424 // printf("%d\n", numEq); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1425 return numEq > hFlatnessThreshold; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1426 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1427 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1428 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
|
1429 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1430 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
|
1431 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1432 return 1; |
2158
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 |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1435 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
|
1436 { |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1437 #if 0 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1438 asm volatile( |
2428 | 1439 "leal (%0, %1), %%ecx \n\t" |
1440 "leal (%%ecx, %1, 4), %%ebx \n\t" | |
1441 // 0 1 2 3 4 5 6 7 8 9 | |
1442 // %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
|
1443 "pxor %%mm7, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1444 "movq bm00001000, %%mm6 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1445 "movd %2, %%mm5 \n\t" // QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1446 "movq %%mm5, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1447 "paddusb %%mm5, %%mm5 \n\t" // 2QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1448 "paddusb %%mm5, %%mm4 \n\t" // 3QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1449 "psllq $24, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1450 "pxor %%mm5, %%mm5 \n\t" // 0 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1451 "psubb %%mm4, %%mm5 \n\t" // -QP |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1452 "leal tempBlock, %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1453 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1454 //FIXME? "unroll by 2" and mix |
2159 | 1455 #ifdef HAVE_MMX2 |
2428 | 1456 #define HDF(src, dst) \ |
1457 "movq " #src "(%%eax), %%mm0 \n\t"\ | |
1458 "movq " #src "(%%eax), %%mm1 \n\t"\ | |
1459 "movq " #src "(%%eax), %%mm2 \n\t"\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1460 "psrlq $8, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1461 "psubusb %%mm1, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1462 "psubusb %%mm0, %%mm1 \n\t"\ |
2159 | 1463 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ |
1464 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
1465 "pshufw $0x00, %%mm1, %%mm3 \n\t" /* p´5 = |p1 - p2| */\ | |
1466 "pminub %%mm1, %%mm3 \n\t" /* p´5 = min(|p2-p1|, |p6-p5|)*/\ | |
1467 "psrlq $16, %%mm3 \n\t" /* p´3 = min(|p2-p1|, |p6-p5|)*/\ | |
1468 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-min(|p1-p2|,|p5-p6|) */\ | |
1469 "paddb %%mm5, %%mm1 \n\t"\ | |
1470 "psubusb %%mm5, %%mm1 \n\t"\ | |
1471 "psrlw $2, %%mm1 \n\t"\ | |
1472 "pxor %%mm2, %%mm1 \n\t"\ | |
1473 "psubb %%mm2, %%mm1 \n\t"\ | |
1474 "pand %%mm6, %%mm1 \n\t"\ | |
1475 "psubb %%mm1, %%mm0 \n\t"\ | |
1476 "psllq $8, %%mm1 \n\t"\ | |
1477 "paddb %%mm1, %%mm0 \n\t"\ | |
2428 | 1478 "movd %%mm0, " #dst" \n\t"\ |
2159 | 1479 "psrlq $32, %%mm0 \n\t"\ |
2428 | 1480 "movd %%mm0, 4" #dst" \n\t" |
2159 | 1481 #else |
2428 | 1482 #define HDF(src, dst)\ |
1483 "movq " #src "(%%eax), %%mm0 \n\t"\ | |
2159 | 1484 "movq %%mm0, %%mm1 \n\t"\ |
1485 "movq %%mm0, %%mm2 \n\t"\ | |
1486 "psrlq $8, %%mm1 \n\t"\ | |
1487 "psubusb %%mm1, %%mm2 \n\t"\ | |
1488 "psubusb %%mm0, %%mm1 \n\t"\ | |
1489 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\ | |
1490 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\ | |
1491 "movq %%mm1, %%mm3 \n\t"\ | |
1492 "psllq $32, %%mm3 \n\t"\ | |
1493 "movq %%mm3, %%mm4 \n\t"\ | |
1494 "psubusb %%mm1, %%mm4 \n\t"\ | |
1495 "psubb %%mm4, %%mm3 \n\t"\ | |
1496 "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
|
1497 "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
|
1498 "paddb %%mm5, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1499 "psubusb %%mm5, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1500 "psrlw $2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1501 "pxor %%mm2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1502 "psubb %%mm2, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1503 "pand %%mm6, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1504 "psubb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1505 "psllq $8, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1506 "paddb %%mm1, %%mm0 \n\t"\ |
2428 | 1507 "movd %%mm0, " #dst " \n\t"\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1508 "psrlq $32, %%mm0 \n\t"\ |
2428 | 1509 "movd %%mm0, 4" #dst " \n\t" |
2159 | 1510 #endif |
2428 | 1511 HDF(0,(%0)) |
1512 HDF(8,(%%ecx)) | |
1513 HDF(16,(%%ecx, %1)) | |
1514 HDF(24,(%%ecx, %1, 2)) | |
1515 HDF(32,(%0, %1, 4)) | |
1516 HDF(40,(%%ebx)) | |
1517 HDF(48,(%%ebx, %1)) | |
1518 HDF(56,(%%ebx, %1, 2)) | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1519 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1520 : "r" (dst), "r" (stride), "r" (QP) |
2428 | 1521 : "%eax", "%ebx", "%ecx" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1522 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1523 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1524 int y; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1525 for(y=0; y<BLOCK_SIZE; y++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1526 { |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1527 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
|
1528 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1529 if(ABS(middleEnergy) < 8*QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1530 { |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1531 const int q=(dst[3] - dst[4])/2; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1532 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
|
1533 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
|
1534 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1535 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1536 d= MAX(d, 0); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1537 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1538 d= (5*d + 32) >> 6; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1539 d*= SIGN(-middleEnergy); |
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 if(q>0) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1542 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1543 d= d<0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1544 d= d>q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1545 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1546 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1547 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1548 d= d>0 ? 0 : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1549 d= d<q ? q : d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1550 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1551 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1552 dst[3]-= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1553 dst[4]+= d; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1554 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1555 dst+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1556 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1557 #endif |
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 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1560 /** |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1561 * Do a horizontal low pass filter on the 10x8 block (dst points to middle 8x8 Block) |
2221 | 1562 * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 (C version) |
1563 * 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
|
1564 */ |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1565 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
|
1566 { |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1567 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1568 #if 0 |
2428 | 1569 asm volatile( |
1570 "leal (%0, %1), %%ecx \n\t" | |
1571 "leal (%%ecx, %1, 4), %%ebx \n\t" | |
1572 // 0 1 2 3 4 5 6 7 8 9 | |
1573 // %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
|
1574 "pxor %%mm7, %%mm7 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1575 "leal tempBlock, %%eax \n\t" |
2159 | 1576 /* |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1577 #define HLP1 "movq (%0), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1578 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1579 "psllq $8, %%mm0 \n\t"\ |
2159 | 1580 PAVGB(%%mm1, %%mm0)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1581 "psrlw $8, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1582 "pxor %%mm1, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1583 "packuswb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1584 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1585 "movq %%mm0, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1586 "psllq $32, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1587 "paddb %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1588 "psllq $16, %%mm2 \n\t"\ |
2159 | 1589 PAVGB(%%mm2, %%mm0)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1590 "movq %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1591 "pand bm11001100, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1592 "paddusb %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1593 "psrlq $8, %%mm3 \n\t"\ |
2159 | 1594 PAVGB(%%mm1, %%mm4)\ |
1595 PAVGB(%%mm3, %%mm2)\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1596 "psrlq $16, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1597 "punpcklbw %%mm2, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1598 "movq %%mm2, (%0) \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1599 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1600 #define HLP2 "movq (%0), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1601 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1602 "psllq $8, %%mm0 \n\t"\ |
2159 | 1603 PAVGB(%%mm1, %%mm0)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1604 "psrlw $8, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1605 "pxor %%mm1, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1606 "packuswb %%mm1, %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1607 "movq %%mm0, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1608 "psllq $32, %%mm0 \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(%%mm3, %%mm2)\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1616 "psrlq $16, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1617 "punpcklbw %%mm2, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1618 "movq %%mm2, (%0) \n\t"\ |
2159 | 1619 */ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1620 // 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
|
1621 /* |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1622 Implemented Exact 7-Tap |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1623 9421 A321 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1624 36421 64321 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1625 334321 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1626 1234321 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1627 1234321 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1628 123433 = |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1629 12463 12346 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1630 1249 123A |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1631 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1632 */ |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1633 |
2159 | 1634 #ifdef HAVE_MMX2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1635 #define HLP3(i) "movq " #i "(%%eax), %%mm0 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1636 "movq %%mm0, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1637 "movq %%mm0, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1638 "movq %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1639 "movq %%mm0, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1640 "psllq $8, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1641 "psrlq $8, %%mm2 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1642 "pand bm00000001, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1643 "pand bm10000000, %%mm4 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1644 "por %%mm3, %%mm1 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1645 "por %%mm4, %%mm2 \n\t"\ |
2159 | 1646 PAVGB(%%mm2, %%mm1)\ |
1647 PAVGB(%%mm1, %%mm0)\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1648 \ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1649 "pshufw $0xF9, %%mm0, %%mm3 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1650 "pshufw $0x90, %%mm0, %%mm4 \n\t"\ |
2159 | 1651 PAVGB(%%mm3, %%mm4)\ |
1652 PAVGB(%%mm4, %%mm0)\ | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1653 "movd %%mm0, (%0) \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1654 "psrlq $32, %%mm0 \n\t"\ |
2159 | 1655 "movd %%mm0, 4(%0) \n\t" |
1656 #else | |
1657 #define HLP3(i) "movq " #i "(%%eax), %%mm0 \n\t"\ | |
1658 "movq %%mm0, %%mm1 \n\t"\ | |
1659 "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
|
1660 "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
|
1661 "movd 8(%0), %%mm4 \n\t" /*0001000*/\ |
2159 | 1662 "psllq $8, %%mm1 \n\t"\ |
1663 "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
|
1664 "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
|
1665 "psllq $56, %%mm4 \n\t"\ |
2159 | 1666 "por %%mm3, %%mm1 \n\t"\ |
1667 "por %%mm4, %%mm2 \n\t"\ | |
1668 PAVGB(%%mm2, %%mm1)\ | |
1669 PAVGB(%%mm1, %%mm0)\ | |
1670 \ | |
1671 "movq %%mm0, %%mm3 \n\t"\ | |
1672 "movq %%mm0, %%mm4 \n\t"\ | |
1673 "movq %%mm0, %%mm5 \n\t"\ | |
1674 "psrlq $16, %%mm3 \n\t"\ | |
1675 "psllq $16, %%mm4 \n\t"\ | |
1676 "pand bm11000000, %%mm5 \n\t"\ | |
1677 "por %%mm5, %%mm3 \n\t"\ | |
1678 "movq %%mm0, %%mm5 \n\t"\ | |
1679 "pand bm00000011, %%mm5 \n\t"\ | |
1680 "por %%mm5, %%mm4 \n\t"\ | |
1681 PAVGB(%%mm3, %%mm4)\ | |
1682 PAVGB(%%mm4, %%mm0)\ | |
1683 "movd %%mm0, (%0) \n\t"\ | |
1684 "psrlq $32, %%mm0 \n\t"\ | |
1685 "movd %%mm0, 4(%0) \n\t" | |
1686 #endif | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1687 |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1688 /* uses the 7-Tap Filter: 1112111 */ |
2428 | 1689 #define NEW_HLP(src, dst)\ |
1690 "movq " #src "(%%eax), %%mm1 \n\t"\ | |
1691 "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
|
1692 "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
|
1693 "psrlq $8, %%mm2 \n\t"\ |
2428 | 1694 "movd -4" #dst ", %%mm3 \n\t" /*0001000*/\ |
1695 "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
|
1696 "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
|
1697 "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
|
1698 "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
|
1699 "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
|
1700 "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
|
1701 PAVGB(%%mm2, %%mm1)\ |
2428 | 1702 "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
|
1703 PAVGB(%%mm1, %%mm0)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1704 "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
|
1705 "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
|
1706 "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
|
1707 "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
|
1708 "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
|
1709 PAVGB(%%mm2, %%mm5)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1710 "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
|
1711 "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
|
1712 "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
|
1713 "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
|
1714 PAVGB(%%mm2, %%mm1)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1715 PAVGB(%%mm1, %%mm5)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1716 PAVGB(%%mm5, %%mm0)\ |
2428 | 1717 "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
|
1718 "psrlq $32, %%mm0 \n\t"\ |
2428 | 1719 "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
|
1720 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1721 /* 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
|
1722 #define NEW_HLP2(i)\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1723 "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
|
1724 "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
|
1725 "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
|
1726 "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
|
1727 "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
|
1728 "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
|
1729 "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
|
1730 "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
|
1731 "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
|
1732 "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
|
1733 "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
|
1734 "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
|
1735 PAVGB(%%mm2, %%mm1) /*0010100*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1736 PAVGB(%%mm1, %%mm0) /*0012100*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1737 "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
|
1738 "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
|
1739 "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
|
1740 "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
|
1741 "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
|
1742 PAVGB(%%mm2, %%mm5) /*0100010*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1743 "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
|
1744 "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
|
1745 "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
|
1746 "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
|
1747 "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
|
1748 PAVGB(%%mm2, %%mm1) /*1000001*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1749 "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
|
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 "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
|
1752 "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
|
1753 PAVGB(%%mm2, %%mm6) /*100000001*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1754 PAVGB(%%mm6, %%mm1) /*110000011*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1755 PAVGB(%%mm1, %%mm5) /*112000211*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1756 PAVGB(%%mm5, %%mm0) /*112242211*/\ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
1757 "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
|
1758 "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
|
1759 "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
|
1760 |
2428 | 1761 #define HLP(src, dst) NEW_HLP(src, dst) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1762 |
2428 | 1763 HLP(0, (%0)) |
1764 HLP(8, (%%ecx)) | |
1765 HLP(16, (%%ecx, %1)) | |
1766 HLP(24, (%%ecx, %1, 2)) | |
1767 HLP(32, (%0, %1, 4)) | |
1768 HLP(40, (%%ebx)) | |
1769 HLP(48, (%%ebx, %1)) | |
1770 HLP(56, (%%ebx, %1, 2)) | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1771 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1772 : |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1773 : "r" (dst), "r" (stride) |
2428 | 1774 : "%eax", "%ebx", "%ecx" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1775 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1776 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1777 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1778 int y; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
1779 for(y=0; y<BLOCK_SIZE; y++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1780 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1781 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
|
1782 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
|
1783 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1784 int sums[9]; |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1785 sums[0] = first + dst[0]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1786 sums[1] = dst[0] + dst[1]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1787 sums[2] = dst[1] + dst[2]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1788 sums[3] = dst[2] + dst[3]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1789 sums[4] = dst[3] + dst[4]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1790 sums[5] = dst[4] + dst[5]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1791 sums[6] = dst[5] + dst[6]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1792 sums[7] = dst[6] + dst[7]; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
1793 sums[8] = dst[7] + last; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1794 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1795 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
|
1796 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
|
1797 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
|
1798 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
|
1799 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
|
1800 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
|
1801 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
|
1802 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
|
1803 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1804 dst+= stride; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1805 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1806 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1807 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1808 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1809 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
|
1810 { |
2475 | 1811 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1812 asm volatile( |
2473 | 1813 "movq pQPb, %%mm0 \n\t" |
1814 "paddusb %%mm0, %%mm0 \n\t" | |
1815 "movq %%mm0, pQPb2 \n\t" | |
1816 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1817 "leal (%0, %1), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1818 "leal (%%eax, %1, 4), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1819 // 0 1 2 3 4 5 6 7 8 9 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1820 // %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
|
1821 |
2473 | 1822 "pcmpeqb %%mm6, %%mm6 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1823 "pxor %%mm7, %%mm7 \n\t" |
2475 | 1824 #ifdef HAVE_MMX2 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1825 #define FIND_MIN_MAX(addr)\ |
2473 | 1826 "movq " #addr ", %%mm0 \n\t"\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1827 "pminub %%mm0, %%mm6 \n\t"\ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1828 "pmaxub %%mm0, %%mm7 \n\t" |
2475 | 1829 #else |
1830 #define FIND_MIN_MAX(addr)\ | |
1831 "movq " #addr ", %%mm0 \n\t"\ | |
1832 "movq %%mm6, %%mm1 \n\t"\ | |
1833 "psubusb %%mm0, %%mm7 \n\t"\ | |
1834 "paddb %%mm0, %%mm7 \n\t"\ | |
1835 "psubusb %%mm0, %%mm1 \n\t"\ | |
1836 "psubb %%mm1, %%mm6 \n\t" | |
1837 #endif | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1838 |
2473 | 1839 FIND_MIN_MAX((%%eax)) |
1840 FIND_MIN_MAX((%%eax, %1)) | |
1841 FIND_MIN_MAX((%%eax, %1, 2)) | |
1842 FIND_MIN_MAX((%0, %1, 4)) | |
1843 FIND_MIN_MAX((%%ebx)) | |
1844 FIND_MIN_MAX((%%ebx, %1)) | |
1845 FIND_MIN_MAX((%%ebx, %1, 2)) | |
1846 FIND_MIN_MAX((%0, %1, 8)) | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1847 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1848 "movq %%mm6, %%mm4 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1849 "psrlq $8, %%mm6 \n\t" |
2475 | 1850 #ifdef HAVE_MMX2 |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1851 "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
|
1852 "pshufw $0xF9, %%mm6, %%mm4 \n\t" |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1853 "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
|
1854 "pshufw $0xFE, %%mm6, %%mm4 \n\t" |
2475 | 1855 "pminub %%mm4, %%mm6 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1856 #else |
2475 | 1857 "movq %%mm6, %%mm1 \n\t" |
1858 "psubusb %%mm4, %%mm1 \n\t" | |
1859 "psubb %%mm1, %%mm6 \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1860 "movq %%mm6, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1861 "psrlq $16, %%mm6 \n\t" |
2475 | 1862 "movq %%mm6, %%mm1 \n\t" |
1863 "psubusb %%mm4, %%mm1 \n\t" | |
1864 "psubb %%mm1, %%mm6 \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1865 "movq %%mm6, %%mm4 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1866 "psrlq $32, %%mm6 \n\t" |
2475 | 1867 "movq %%mm6, %%mm1 \n\t" |
1868 "psubusb %%mm4, %%mm1 \n\t" | |
1869 "psubb %%mm1, %%mm6 \n\t" | |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1870 #endif |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1871 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1872 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1873 "movq %%mm7, %%mm4 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1874 "psrlq $8, %%mm7 \n\t" |
2475 | 1875 #ifdef HAVE_MMX2 |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1876 "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
|
1877 "pshufw $0xF9, %%mm7, %%mm4 \n\t" |
2475 | 1878 "pmaxub %%mm4, %%mm7 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1879 "pshufw $0xFE, %%mm7, %%mm4 \n\t" |
2475 | 1880 "pmaxub %%mm4, %%mm7 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1881 #else |
2475 | 1882 "psubusb %%mm4, %%mm7 \n\t" |
1883 "paddb %%mm4, %%mm7 \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1884 "movq %%mm7, %%mm4 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1885 "psrlq $16, %%mm7 \n\t" |
2475 | 1886 "psubusb %%mm4, %%mm7 \n\t" |
1887 "paddb %%mm4, %%mm7 \n\t" | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
1888 "movq %%mm7, %%mm4 \n\t" |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1889 "psrlq $32, %%mm7 \n\t" |
2475 | 1890 "psubusb %%mm4, %%mm7 \n\t" |
1891 "paddb %%mm4, %%mm7 \n\t" | |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1892 #endif |
2473 | 1893 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
|
1894 "punpcklbw %%mm7, %%mm7 \n\t" |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1895 "punpcklbw %%mm7, %%mm7 \n\t" |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
1896 "punpcklbw %%mm7, %%mm7 \n\t" |
2473 | 1897 "movq %%mm7, temp0 \n\t" |
1898 | |
1899 "movq (%0), %%mm0 \n\t" // L10 | |
1900 "movq %%mm0, %%mm1 \n\t" // L10 | |
1901 "movq %%mm0, %%mm2 \n\t" // L10 | |
1902 "psllq $8, %%mm1 \n\t" | |
1903 "psrlq $8, %%mm2 \n\t" | |
1904 "movd -4(%0), %%mm3 \n\t" | |
1905 "movd 8(%0), %%mm4 \n\t" | |
1906 "psrlq $24, %%mm3 \n\t" | |
1907 "psllq $56, %%mm4 \n\t" | |
1908 "por %%mm3, %%mm1 \n\t" // L00 | |
1909 "por %%mm4, %%mm2 \n\t" // L20 | |
1910 "movq %%mm1, %%mm3 \n\t" // L00 | |
1911 PAVGB(%%mm2, %%mm1) // (L20 + L00)/2 | |
1912 PAVGB(%%mm0, %%mm1) // (L20 + L00 + 2L10)/4 | |
1913 "psubusb %%mm7, %%mm0 \n\t" | |
1914 "psubusb %%mm7, %%mm2 \n\t" | |
1915 "psubusb %%mm7, %%mm3 \n\t" | |
1916 "pcmpeqb b00, %%mm0 \n\t" // L10 > a ? 0 : -1 | |
1917 "pcmpeqb b00, %%mm2 \n\t" // L20 > a ? 0 : -1 | |
1918 "pcmpeqb b00, %%mm3 \n\t" // L00 > a ? 0 : -1 | |
1919 "paddb %%mm2, %%mm0 \n\t" | |
1920 "paddb %%mm3, %%mm0 \n\t" | |
1921 | |
1922 "movq (%%eax), %%mm2 \n\t" // L11 | |
1923 "movq %%mm2, %%mm3 \n\t" // L11 | |
1924 "movq %%mm2, %%mm4 \n\t" // L11 | |
1925 "psllq $8, %%mm3 \n\t" | |
1926 "psrlq $8, %%mm4 \n\t" | |
1927 "movd -4(%%eax), %%mm5 \n\t" | |
1928 "movd 8(%%eax), %%mm6 \n\t" | |
1929 "psrlq $24, %%mm5 \n\t" | |
1930 "psllq $56, %%mm6 \n\t" | |
1931 "por %%mm5, %%mm3 \n\t" // L01 | |
1932 "por %%mm6, %%mm4 \n\t" // L21 | |
1933 "movq %%mm3, %%mm5 \n\t" // L01 | |
1934 PAVGB(%%mm4, %%mm3) // (L21 + L01)/2 | |
1935 PAVGB(%%mm2, %%mm3) // (L21 + L01 + 2L11)/4 | |
1936 "psubusb %%mm7, %%mm2 \n\t" | |
1937 "psubusb %%mm7, %%mm4 \n\t" | |
1938 "psubusb %%mm7, %%mm5 \n\t" | |
1939 "pcmpeqb b00, %%mm2 \n\t" // L11 > a ? 0 : -1 | |
1940 "pcmpeqb b00, %%mm4 \n\t" // L21 > a ? 0 : -1 | |
1941 "pcmpeqb b00, %%mm5 \n\t" // L01 > a ? 0 : -1 | |
1942 "paddb %%mm4, %%mm2 \n\t" | |
1943 "paddb %%mm5, %%mm2 \n\t" | |
1944 // 0, 2, 3, 1 | |
1945 #define DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \ | |
1946 "movq " #src ", " #sx " \n\t" /* src[0] */\ | |
1947 "movq " #sx ", " #lx " \n\t" /* src[0] */\ | |
1948 "movq " #sx ", " #t0 " \n\t" /* src[0] */\ | |
1949 "psllq $8, " #lx " \n\t"\ | |
1950 "psrlq $8, " #t0 " \n\t"\ | |
1951 "movd -4" #src ", " #t1 " \n\t"\ | |
1952 "psrlq $24, " #t1 " \n\t"\ | |
1953 "por " #t1 ", " #lx " \n\t" /* src[-1] */\ | |
1954 "movd 8" #src ", " #t1 " \n\t"\ | |
1955 "psllq $56, " #t1 " \n\t"\ | |
1956 "por " #t1 ", " #t0 " \n\t" /* src[+1] */\ | |
1957 "movq " #lx ", " #t1 " \n\t" /* src[-1] */\ | |
1958 PAVGB(t0, lx) /* (src[-1] + src[+1])/2 */\ | |
1959 PAVGB(sx, lx) /* (src[-1] + 2src[0] + src[+1])/4 */\ | |
1960 "psubusb temp0, " #t1 " \n\t"\ | |
1961 "psubusb temp0, " #t0 " \n\t"\ | |
1962 "psubusb temp0, " #sx " \n\t"\ | |
1963 "pcmpeqb b00, " #t1 " \n\t" /* src[-1] > a ? 0 : -1*/\ | |
1964 "pcmpeqb b00, " #t0 " \n\t" /* src[+1] > a ? 0 : -1*/\ | |
1965 "pcmpeqb b00, " #sx " \n\t" /* src[0] > a ? 0 : -1*/\ | |
1966 "paddb " #t1 ", " #t0 " \n\t"\ | |
1967 "paddb " #t0 ", " #sx " \n\t"\ | |
1968 \ | |
1969 PAVGB(lx, pplx) \ | |
1970 PAVGB(plx, pplx) /* filtered */\ | |
1971 "movq " #dst ", " #t0 " \n\t" /* dst */\ | |
1972 "movq " #pplx ", " #t1 " \n\t"\ | |
1973 "psubusb " #t0 ", " #pplx " \n\t"\ | |
1974 "psubusb " #t1 ", " #t0 " \n\t"\ | |
1975 "por " #t0 ", " #pplx " \n\t" /* |filtered - dst| */\ | |
1976 "psubusb pQPb2, " #pplx " \n\t"\ | |
1977 "pcmpeqb b00, " #pplx " \n\t"\ | |
1978 "paddb " #sx ", " #ppsx " \n\t"\ | |
1979 "paddb " #psx ", " #ppsx " \n\t"\ | |
1980 "#paddb b02, " #ppsx " \n\t"\ | |
1981 "pand b08, " #ppsx " \n\t"\ | |
1982 "pcmpeqb b00, " #ppsx " \n\t"\ | |
1983 "pand " #pplx ", " #ppsx " \n\t"\ | |
1984 "pand " #ppsx ", " #t1 " \n\t"\ | |
1985 "pandn " #dst ", " #ppsx " \n\t"\ | |
1986 "por " #t1 ", " #ppsx " \n\t"\ | |
1987 "movq " #ppsx ", " #dst " \n\t" | |
1988 /* | |
1989 0000000 | |
1990 1111111 | |
1991 | |
1992 1111110 | |
1993 1111101 | |
1994 1111100 | |
1995 1111011 | |
1996 1111010 | |
1997 1111001 | |
1998 | |
1999 1111000 | |
2000 1110111 | |
2001 | |
2002 */ | |
2003 //DERING_CORE(dst,src ,ppsx ,psx ,sx ,pplx ,plx ,lx ,t0 ,t1) | |
2004 DERING_CORE((%%eax),(%%eax, %1) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7) | |
2005 DERING_CORE((%%eax, %1),(%%eax, %1, 2) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7) | |
2006 DERING_CORE((%%eax, %1, 2),(%0, %1, 4) ,%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7) | |
2007 DERING_CORE((%0, %1, 4),(%%ebx) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7) | |
2008 DERING_CORE((%%ebx),(%%ebx, %1) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7) | |
2009 DERING_CORE((%%ebx, %1), (%%ebx, %1, 2),%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7) | |
2010 DERING_CORE((%%ebx, %1, 2),(%0, %1, 8) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7) | |
2011 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
|
2012 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2013 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2014 : : "r" (src), "r" (stride), "r" (QP) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2015 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2016 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2017 #else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2018 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2019 //FIXME |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2020 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2021 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2022 |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2023 /** |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2024 * Deinterlaces the given block |
2246 | 2025 * 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
|
2026 */ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2027 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
|
2028 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2029 #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
|
2030 asm volatile( |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2031 "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
|
2032 "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
|
2033 // 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
|
2034 // %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
|
2035 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2036 "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
|
2037 "movq (%%eax, %1), %%mm1 \n\t" |
2246 | 2038 PAVGB(%%mm1, %%mm0) |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2039 "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
|
2040 "movq (%0, %1, 4), %%mm0 \n\t" |
2246 | 2041 PAVGB(%%mm0, %%mm1) |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2042 "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
|
2043 "movq (%%ebx, %1), %%mm1 \n\t" |
2246 | 2044 PAVGB(%%mm1, %%mm0) |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2045 "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
|
2046 "movq (%0, %1, 8), %%mm0 \n\t" |
2246 | 2047 PAVGB(%%mm0, %%mm1) |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2048 "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
|
2049 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2050 : : "r" (src), "r" (stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2051 : "%eax", "%ebx" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2052 ); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2053 #else |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2054 int x; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2055 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
|
2056 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2057 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
|
2058 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
|
2059 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
|
2060 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
|
2061 src++; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2062 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2063 #endif |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2064 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2065 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2066 /** |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2067 * Deinterlaces the given block |
2246 | 2068 * will be called for every 8x8 block, and can read & write into an 8x16 block |
2069 * 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
|
2070 */ |
2246 | 2071 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
|
2072 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2073 #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
|
2074 asm volatile( |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2075 "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
|
2076 "leal (%%eax, %1, 4), %%ebx \n\t" |
2246 | 2077 "leal (%%ebx, %1, 4), %%ecx \n\t" |
2078 "addl %1, %%ecx \n\t" | |
2079 "pxor %%mm7, %%mm7 \n\t" | |
2080 // 0 1 2 3 4 5 6 7 8 9 10 | |
2081 // %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
|
2082 |
2246 | 2083 #define DEINT_CUBIC(a,b,c,d,e)\ |
2084 "movq " #a ", %%mm0 \n\t"\ | |
2085 "movq " #b ", %%mm1 \n\t"\ | |
2086 "movq " #d ", %%mm2 \n\t"\ | |
2087 "movq " #e ", %%mm3 \n\t"\ | |
2088 PAVGB(%%mm2, %%mm1) /* (b+d) /2 */\ | |
2089 PAVGB(%%mm3, %%mm0) /* a(a+e) /2 */\ | |
2090 "movq %%mm0, %%mm2 \n\t"\ | |
2091 "punpcklbw %%mm7, %%mm0 \n\t"\ | |
2092 "punpckhbw %%mm7, %%mm2 \n\t"\ | |
2093 "movq %%mm1, %%mm3 \n\t"\ | |
2094 "punpcklbw %%mm7, %%mm1 \n\t"\ | |
2095 "punpckhbw %%mm7, %%mm3 \n\t"\ | |
2096 "psubw %%mm1, %%mm0 \n\t" /* L(a+e - (b+d))/2 */\ | |
2097 "psubw %%mm3, %%mm2 \n\t" /* H(a+e - (b+d))/2 */\ | |
2098 "psraw $3, %%mm0 \n\t" /* L(a+e - (b+d))/16 */\ | |
2099 "psraw $3, %%mm2 \n\t" /* H(a+e - (b+d))/16 */\ | |
2100 "psubw %%mm0, %%mm1 \n\t" /* L(9b + 9d - a - e)/16 */\ | |
2101 "psubw %%mm2, %%mm3 \n\t" /* H(9b + 9d - a - e)/16 */\ | |
2102 "packuswb %%mm3, %%mm1 \n\t"\ | |
2103 "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
|
2104 |
2246 | 2105 DEINT_CUBIC((%0), (%%eax, %1), (%%eax, %1, 2), (%0, %1, 4), (%%ebx, %1)) |
2106 DEINT_CUBIC((%%eax, %1), (%0, %1, 4), (%%ebx), (%%ebx, %1), (%0, %1, 8)) | |
2107 DEINT_CUBIC((%0, %1, 4), (%%ebx, %1), (%%ebx, %1, 2), (%0, %1, 8), (%%ecx)) | |
2108 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
|
2109 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2110 : : "r" (src), "r" (stride) |
2246 | 2111 : "%eax", "%ebx", "ecx" |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2112 ); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2113 #else |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2114 int x; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2115 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
|
2116 { |
2246 | 2117 src[stride*3] = (-src[0] + 9*src[stride*2] + 9*src[stride*4] - src[stride*6])>>4; |
2118 src[stride*5] = (-src[stride*2] + 9*src[stride*4] + 9*src[stride*6] - src[stride*8])>>4; | |
2119 src[stride*7] = (-src[stride*4] + 9*src[stride*6] + 9*src[stride*8] - src[stride*10])>>4; | |
2120 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
|
2121 src++; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2122 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2123 #endif |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2124 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2125 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2126 /** |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2127 * Deinterlaces the given block |
2246 | 2128 * 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
|
2129 * 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
|
2130 */ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2131 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
|
2132 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2133 #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
|
2134 asm volatile( |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2135 "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
|
2136 "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
|
2137 // 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
|
2138 // %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
|
2139 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2140 "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
|
2141 "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
|
2142 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
|
2143 "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
|
2144 PAVGB(%%mm2, %%mm0) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2145 "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
|
2146 "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
|
2147 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
|
2148 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
|
2149 "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
|
2150 "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
|
2151 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
|
2152 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
|
2153 "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
|
2154 "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
|
2155 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
|
2156 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
|
2157 "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
|
2158 "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
|
2159 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
|
2160 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
|
2161 "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
|
2162 "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
|
2163 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
|
2164 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
|
2165 "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
|
2166 "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
|
2167 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
|
2168 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
|
2169 "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
|
2170 "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
|
2171 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
|
2172 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
|
2173 "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
|
2174 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2175 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2176 : : "r" (src), "r" (stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2177 : "%eax", "%ebx" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2178 ); |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2179 #else |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2180 int x; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2181 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
|
2182 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2183 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
|
2184 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
|
2185 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
|
2186 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
|
2187 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
|
2188 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
|
2189 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
|
2190 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
|
2191 src++; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2192 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2193 #endif |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2194 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2195 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2196 /** |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2197 * Deinterlaces the given block |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2198 * 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
|
2199 */ |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2200 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
|
2201 { |
2221 | 2202 #ifdef HAVE_MMX |
2203 #ifdef HAVE_MMX2 | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2204 asm volatile( |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2205 "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
|
2206 "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
|
2207 // 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
|
2208 // %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
|
2209 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2210 "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
|
2211 "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
|
2212 "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
|
2213 "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
|
2214 "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
|
2215 "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
|
2216 "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
|
2217 "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
|
2218 "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
|
2219 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2220 "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
|
2221 "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
|
2222 "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
|
2223 "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
|
2224 "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
|
2225 "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
|
2226 "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
|
2227 "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
|
2228 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2229 "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
|
2230 "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
|
2231 "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
|
2232 "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
|
2233 "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
|
2234 "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
|
2235 "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
|
2236 "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
|
2237 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2238 "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
|
2239 "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
|
2240 "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
|
2241 "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
|
2242 "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
|
2243 "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
|
2244 "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
|
2245 "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
|
2246 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2247 |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2248 : : "r" (src), "r" (stride) |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2249 : "%eax", "%ebx" |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2250 ); |
2221 | 2251 |
2252 #else // MMX without MMX2 | |
2253 asm volatile( | |
2254 "leal (%0, %1), %%eax \n\t" | |
2255 "leal (%%eax, %1, 4), %%ebx \n\t" | |
2256 // 0 1 2 3 4 5 6 7 8 9 | |
2257 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 | |
2258 "pxor %%mm7, %%mm7 \n\t" | |
2259 | |
2260 #define MEDIAN(a,b,c)\ | |
2261 "movq " #a ", %%mm0 \n\t"\ | |
2262 "movq " #b ", %%mm2 \n\t"\ | |
2263 "movq " #c ", %%mm1 \n\t"\ | |
2264 "movq %%mm0, %%mm3 \n\t"\ | |
2265 "movq %%mm1, %%mm4 \n\t"\ | |
2266 "movq %%mm2, %%mm5 \n\t"\ | |
2267 "psubusb %%mm1, %%mm3 \n\t"\ | |
2268 "psubusb %%mm2, %%mm4 \n\t"\ | |
2269 "psubusb %%mm0, %%mm5 \n\t"\ | |
2270 "pcmpeqb %%mm7, %%mm3 \n\t"\ | |
2271 "pcmpeqb %%mm7, %%mm4 \n\t"\ | |
2272 "pcmpeqb %%mm7, %%mm5 \n\t"\ | |
2273 "movq %%mm3, %%mm6 \n\t"\ | |
2274 "pxor %%mm4, %%mm3 \n\t"\ | |
2275 "pxor %%mm5, %%mm4 \n\t"\ | |
2276 "pxor %%mm6, %%mm5 \n\t"\ | |
2277 "por %%mm3, %%mm1 \n\t"\ | |
2278 "por %%mm4, %%mm2 \n\t"\ | |
2279 "por %%mm5, %%mm0 \n\t"\ | |
2280 "pand %%mm2, %%mm0 \n\t"\ | |
2281 "pand %%mm1, %%mm0 \n\t"\ | |
2282 "movq %%mm0, " #b " \n\t" | |
2283 | |
2284 MEDIAN((%0), (%%eax), (%%eax, %1)) | |
2285 MEDIAN((%%eax, %1), (%%eax, %1, 2), (%0, %1, 4)) | |
2286 MEDIAN((%0, %1, 4), (%%ebx), (%%ebx, %1)) | |
2287 MEDIAN((%%ebx, %1), (%%ebx, %1, 2), (%0, %1, 8)) | |
2288 | |
2289 : : "r" (src), "r" (stride) | |
2290 : "%eax", "%ebx" | |
2291 ); | |
2292 #endif // MMX | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2293 #else |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2294 //FIXME |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2295 int x; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2296 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
|
2297 { |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2298 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
|
2299 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
|
2300 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
|
2301 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
|
2302 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
|
2303 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
|
2304 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
|
2305 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
|
2306 src++; |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2307 } |
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
2308 #endif |
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 |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2311 #ifdef HAVE_MMX |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2312 /** |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2313 * 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
|
2314 */ |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2315 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
|
2316 { |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2317 asm( |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2318 "leal (%0, %1), %%eax \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2319 "leal (%%eax, %1, 4), %%ebx \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2320 // 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
|
2321 // %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
|
2322 "movq (%0), %%mm0 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2323 "movq (%%eax), %%mm1 \n\t" // abcdefgh |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2324 "movq %%mm0, %%mm2 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2325 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2326 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2327 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2328 "movq (%%eax, %1), %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2329 "movq (%%eax, %1, 2), %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2330 "movq %%mm1, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2331 "punpcklbw %%mm3, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2332 "punpckhbw %%mm3, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2333 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2334 "movq %%mm0, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2335 "punpcklwd %%mm1, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2336 "punpckhwd %%mm1, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2337 "movq %%mm2, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2338 "punpcklwd %%mm4, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2339 "punpckhwd %%mm4, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2340 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2341 "movd %%mm0, 128(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2342 "psrlq $32, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2343 "movd %%mm0, 144(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2344 "movd %%mm3, 160(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2345 "psrlq $32, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2346 "movd %%mm3, 176(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2347 "movd %%mm3, 48(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2348 "movd %%mm2, 192(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2349 "movd %%mm2, 64(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2350 "psrlq $32, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2351 "movd %%mm2, 80(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2352 "movd %%mm1, 96(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2353 "psrlq $32, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2354 "movd %%mm1, 112(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2355 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2356 "movq (%0, %1, 4), %%mm0 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2357 "movq (%%ebx), %%mm1 \n\t" // abcdefgh |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2358 "movq %%mm0, %%mm2 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2359 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2360 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2361 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2362 "movq (%%ebx, %1), %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2363 "movq (%%ebx, %1, 2), %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2364 "movq %%mm1, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2365 "punpcklbw %%mm3, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2366 "punpckhbw %%mm3, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2367 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2368 "movq %%mm0, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2369 "punpcklwd %%mm1, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2370 "punpckhwd %%mm1, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2371 "movq %%mm2, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2372 "punpcklwd %%mm4, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2373 "punpckhwd %%mm4, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2374 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2375 "movd %%mm0, 132(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2376 "psrlq $32, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2377 "movd %%mm0, 148(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2378 "movd %%mm3, 164(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2379 "psrlq $32, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2380 "movd %%mm3, 180(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2381 "movd %%mm3, 52(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2382 "movd %%mm2, 196(%2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2383 "movd %%mm2, 68(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2384 "psrlq $32, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2385 "movd %%mm2, 84(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2386 "movd %%mm1, 100(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2387 "psrlq $32, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2388 "movd %%mm1, 116(%3) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2389 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2390 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2391 :: "r" (src), "r" (srcStride), "r" (dst1), "r" (dst2) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2392 : "%eax", "%ebx" |
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 } |
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 /** |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2397 * transposes the given 8x8 block |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2398 */ |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2399 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
|
2400 { |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2401 asm( |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2402 "leal (%0, %1), %%eax \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2403 "leal (%%eax, %1, 4), %%ebx \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2404 // 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
|
2405 // %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
|
2406 "movq (%2), %%mm0 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2407 "movq 16(%2), %%mm1 \n\t" // abcdefgh |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2408 "movq %%mm0, %%mm2 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2409 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2410 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2411 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2412 "movq 32(%2), %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2413 "movq 48(%2), %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2414 "movq %%mm1, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2415 "punpcklbw %%mm3, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2416 "punpckhbw %%mm3, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2417 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2418 "movq %%mm0, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2419 "punpcklwd %%mm1, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2420 "punpckhwd %%mm1, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2421 "movq %%mm2, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2422 "punpcklwd %%mm4, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2423 "punpckhwd %%mm4, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2424 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2425 "movd %%mm0, (%0) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2426 "psrlq $32, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2427 "movd %%mm0, (%%eax) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2428 "movd %%mm3, (%%eax, %1) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2429 "psrlq $32, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2430 "movd %%mm3, (%%eax, %1, 2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2431 "movd %%mm2, (%0, %1, 4) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2432 "psrlq $32, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2433 "movd %%mm2, (%%ebx) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2434 "movd %%mm1, (%%ebx, %1) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2435 "psrlq $32, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2436 "movd %%mm1, (%%ebx, %1, 2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2437 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2438 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2439 "movq 64(%2), %%mm0 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2440 "movq 80(%2), %%mm1 \n\t" // abcdefgh |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2441 "movq %%mm0, %%mm2 \n\t" // 12345678 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2442 "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2443 "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2444 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2445 "movq 96(%2), %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2446 "movq 112(%2), %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2447 "movq %%mm1, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2448 "punpcklbw %%mm3, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2449 "punpckhbw %%mm3, %%mm4 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2450 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2451 "movq %%mm0, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2452 "punpcklwd %%mm1, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2453 "punpckhwd %%mm1, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2454 "movq %%mm2, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2455 "punpcklwd %%mm4, %%mm2 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2456 "punpckhwd %%mm4, %%mm1 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2457 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2458 "movd %%mm0, 4(%0) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2459 "psrlq $32, %%mm0 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2460 "movd %%mm0, 4(%%eax) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2461 "movd %%mm3, 4(%%eax, %1) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2462 "psrlq $32, %%mm3 \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2463 "movd %%mm3, 4(%%eax, %1, 2) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2464 "movd %%mm2, 4(%0, %1, 4) \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, 4(%%ebx) \n\t" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2467 "movd %%mm1, 4(%%ebx, %1) \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, 4(%%ebx, %1, 2) \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 :: "r" (dst), "r" (dstStride), "r" (src) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2472 : "%eax", "%ebx" |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2473 ); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2474 } |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2475 #endif |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
2476 |
2185 | 2477 #ifdef HAVE_ODIVX_POSTPROCESS |
2478 #include "../opendivx/postprocess.h" | |
2479 int use_old_pp=0; | |
2480 #endif | |
2159 | 2481 |
2185 | 2482 static void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height, |
2483 QP_STORE_T QPs[], int QPStride, int isColor, int mode); | |
2159 | 2484 |
2389 | 2485 /* -pp Command line Help |
2486 NOTE/FIXME: put this at an appropriate place (--help, html docs, man mplayer)? | |
2487 | |
2488 -pp <filterName>[:<option>[:<option>...]][,[-]<filterName>[:<option>...]]... | |
2489 | |
2490 long form example: | |
2491 -pp vdeblock:autoq,hdeblock:autoq,linblenddeint -pp default,-vdeblock | |
2492 short form example: | |
2493 -pp vb:a,hb:a,lb -pp de,-vb | |
2494 | |
2495 Filters Options | |
2496 short long name short long option Description | |
2497 * * a autoq cpu power dependant enabler | |
2498 c chrom chrominance filtring enabled | |
2499 y nochrom chrominance filtring disabled | |
2500 hb hdeblock horizontal deblocking filter | |
2501 vb vdeblock vertical deblocking filter | |
2502 vr rkvdeblock | |
2503 h1 x1hdeblock Experimental horizontal deblock filter 1 | |
2504 v1 x1vdeblock Experimental vertical deblock filter 1 | |
2505 dr dering not implemented yet | |
2506 al autolevels automatic brightness / contrast fixer | |
2507 f fullyrange stretch luminance range to (0..255) | |
2508 lb linblenddeint linear blend deinterlacer | |
2509 li linipoldeint linear interpolating deinterlacer | |
2510 ci cubicipoldeint cubic interpolating deinterlacer | |
2511 md mediandeint median deinterlacer | |
2512 de default hdeblock:a,vdeblock:a,dering:a,autolevels | |
2513 fa fast x1hdeblock:a,x1vdeblock:a,dering:a,autolevels | |
2514 */ | |
2515 | |
2516 /** | |
2517 * returns a PPMode struct which will have a non 0 error variable if an error occured | |
2518 * name is the string after "-pp" on the command line | |
2519 * quality is a number from 0 to GET_PP_QUALITY_MAX | |
2520 */ | |
2521 struct PPMode getPPModeByNameAndQuality(char *name, int quality) | |
2522 { | |
2523 char temp[GET_MODE_BUFFER_SIZE]; | |
2524 char *p= temp; | |
2525 char *filterDelimiters= ","; | |
2526 char *optionDelimiters= ":"; | |
2527 struct PPMode ppMode= {0,0,0,0,0,0}; | |
2528 char *filterToken; | |
2529 | |
2530 strncpy(temp, name, GET_MODE_BUFFER_SIZE); | |
2531 | |
2532 for(;;){ | |
2533 char *filterName; | |
2534 int q= GET_PP_QUALITY_MAX; | |
2535 int chrom=-1; | |
2536 char *option; | |
2537 char *options[OPTIONS_ARRAY_SIZE]; | |
2538 int i; | |
2539 int filterNameOk=0; | |
2540 int numOfUnknownOptions=0; | |
2541 int enable=1; //does the user want us to enabled or disabled the filter | |
2542 | |
2543 filterToken= strtok(p, filterDelimiters); | |
2544 if(filterToken == NULL) break; | |
2545 p+= strlen(filterToken) + 1; | |
2546 filterName= strtok(filterToken, optionDelimiters); | |
2547 printf("%s::%s\n", filterToken, filterName); | |
2548 | |
2549 if(*filterName == '-') | |
2550 { | |
2551 enable=0; | |
2552 filterName++; | |
2553 } | |
2554 for(;;){ //for all options | |
2555 option= strtok(NULL, optionDelimiters); | |
2556 if(option == NULL) break; | |
2557 | |
2558 printf("%s\n", option); | |
2559 if(!strcmp("autoq", option) || !strcmp("a", option)) q= quality; | |
2560 else if(!strcmp("nochrom", option) || !strcmp("y", option)) chrom=0; | |
2561 else if(!strcmp("chrom", option) || !strcmp("c", option)) chrom=1; | |
2562 else | |
2563 { | |
2564 options[numOfUnknownOptions] = option; | |
2565 numOfUnknownOptions++; | |
2566 options[numOfUnknownOptions] = NULL; | |
2567 } | |
2568 if(numOfUnknownOptions >= OPTIONS_ARRAY_SIZE-1) break; | |
2569 } | |
2570 | |
2571 /* replace stuff from the replace Table */ | |
2572 for(i=0; replaceTable[2*i]!=NULL; i++) | |
2573 { | |
2574 if(!strcmp(replaceTable[2*i], filterName)) | |
2575 { | |
2576 int newlen= strlen(replaceTable[2*i + 1]); | |
2577 int plen; | |
2578 int spaceLeft; | |
2579 | |
2580 if(p==NULL) p= temp, *p=0; //last filter | |
2581 else p--, *p=','; //not last filter | |
2582 | |
2583 plen= strlen(p); | |
2584 spaceLeft= (int)p - (int)temp + plen; | |
2585 if(spaceLeft + newlen >= GET_MODE_BUFFER_SIZE) | |
2586 { | |
2587 ppMode.error++; | |
2588 break; | |
2589 } | |
2590 memmove(p + newlen, p, plen+1); | |
2591 memcpy(p, replaceTable[2*i + 1], newlen); | |
2592 filterNameOk=1; | |
2593 } | |
2594 } | |
2595 | |
2596 for(i=0; filters[i].shortName!=NULL; i++) | |
2597 { | |
2598 if( !strcmp(filters[i].longName, filterName) | |
2599 || !strcmp(filters[i].shortName, filterName)) | |
2600 { | |
2601 ppMode.lumMode &= ~filters[i].mask; | |
2602 ppMode.chromMode &= ~filters[i].mask; | |
2603 | |
2604 filterNameOk=1; | |
2605 if(!enable) break; // user wants to disable it | |
2606 | |
2607 if(q >= filters[i].minLumQuality) | |
2608 ppMode.lumMode|= filters[i].mask; | |
2609 if(chrom==1 || (chrom==-1 && filters[i].chromDefault)) | |
2610 if(q >= filters[i].minChromQuality) | |
2611 ppMode.chromMode|= filters[i].mask; | |
2612 | |
2613 if(filters[i].mask == LEVEL_FIX) | |
2614 { | |
2615 int o; | |
2616 ppMode.minAllowedY= 16; | |
2617 ppMode.maxAllowedY= 234; | |
2618 for(o=0; options[o]!=NULL; o++) | |
2619 if( !strcmp(options[o],"fullyrange") | |
2620 ||!strcmp(options[o],"f")) | |
2621 { | |
2622 ppMode.minAllowedY= 0; | |
2623 ppMode.maxAllowedY= 255; | |
2624 numOfUnknownOptions--; | |
2625 } | |
2626 } | |
2627 } | |
2628 } | |
2629 if(!filterNameOk) ppMode.error++; | |
2630 ppMode.error += numOfUnknownOptions; | |
2631 } | |
2632 | |
2633 if(ppMode.lumMode & H_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_Y_H; | |
2634 if(ppMode.lumMode & V_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_Y_V; | |
2635 if(ppMode.chromMode & H_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_C_H; | |
2636 if(ppMode.chromMode & V_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_C_V; | |
2637 if(ppMode.lumMode & DERING) ppMode.oldMode |= PP_DERING_Y; | |
2638 if(ppMode.chromMode & DERING) ppMode.oldMode |= PP_DERING_C; | |
2639 | |
2640 return ppMode; | |
2641 } | |
2642 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2643 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2644 * ... |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2645 */ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2646 void postprocess(unsigned char * src[], int src_stride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2647 unsigned char * dst[], int dst_stride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2648 int horizontal_size, int vertical_size, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2649 QP_STORE_T *QP_store, int QP_stride, |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2650 int mode) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2651 { |
2389 | 2652 /* |
2653 static int qual=0; | |
2654 | |
2655 struct PPMode ppMode= getPPModeByNameAndQuality("fast,default,-hdeblock,-vdeblock", qual); | |
2656 qual++; | |
2657 qual%=7; | |
2658 printf("\n%d %d %d %d\n", ppMode.lumMode, ppMode.chromMode, ppMode.oldMode, ppMode.error); | |
2659 postprocess2(src, src_stride, dst, dst_stride, | |
2660 horizontal_size, vertical_size, QP_store, QP_stride, &ppMode); | |
2661 | |
2662 return; | |
2663 */ | |
2159 | 2664 |
2185 | 2665 #ifdef HAVE_ODIVX_POSTPROCESS |
2666 // Note: I could make this shit outside of this file, but it would mean one | |
2667 // more function call... | |
2668 if(use_old_pp){ | |
2669 odivx_postprocess(src,src_stride,dst,dst_stride,horizontal_size,vertical_size,QP_store,QP_stride,mode); | |
2670 return; | |
2671 } | |
2672 #endif | |
2673 | |
2159 | 2674 postProcess(src[0], src_stride, dst[0], dst_stride, |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2675 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
|
2676 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2677 horizontal_size >>= 1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2678 vertical_size >>= 1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2679 src_stride >>= 1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2680 dst_stride >>= 1; |
2179 | 2681 mode= ((mode&0xFF)>>4) | (mode&0xFFFFFF00); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2682 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2683 if(1) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2684 { |
2159 | 2685 postProcess(src[1], src_stride, dst[1], dst_stride, |
2179 | 2686 horizontal_size, vertical_size, QP_store, QP_stride, 1, mode); |
2159 | 2687 postProcess(src[2], src_stride, dst[2], dst_stride, |
2389 | 2688 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
|
2689 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2690 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2691 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2692 memcpy(dst[1], src[1], src_stride*horizontal_size); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2693 memcpy(dst[2], src[2], src_stride*horizontal_size); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2694 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2695 } |
2185 | 2696 |
2389 | 2697 void postprocess2(unsigned char * src[], int src_stride, |
2698 unsigned char * dst[], int dst_stride, | |
2699 int horizontal_size, int vertical_size, | |
2700 QP_STORE_T *QP_store, int QP_stride, | |
2701 struct PPMode *mode) | |
2702 { | |
2703 | |
2704 #ifdef HAVE_ODIVX_POSTPROCESS | |
2705 // Note: I could make this shit outside of this file, but it would mean one | |
2706 // more function call... | |
2707 if(use_old_pp){ | |
2708 odivx_postprocess(src,src_stride,dst,dst_stride,horizontal_size,vertical_size,QP_store,QP_stride, | |
2709 mode->oldMode); | |
2710 return; | |
2711 } | |
2712 #endif | |
2713 | |
2714 postProcess(src[0], src_stride, dst[0], dst_stride, | |
2715 horizontal_size, vertical_size, QP_store, QP_stride, 0, mode->lumMode); | |
2716 | |
2717 horizontal_size >>= 1; | |
2718 vertical_size >>= 1; | |
2719 src_stride >>= 1; | |
2720 dst_stride >>= 1; | |
2721 | |
2722 postProcess(src[1], src_stride, dst[1], dst_stride, | |
2723 horizontal_size, vertical_size, QP_store, QP_stride, 1, mode->chromMode); | |
2724 postProcess(src[2], src_stride, dst[2], dst_stride, | |
2725 horizontal_size, vertical_size, QP_store, QP_stride, 2, mode->chromMode); | |
2726 } | |
2727 | |
2728 | |
2159 | 2729 /** |
2730 * gets the mode flags for a given quality (larger values mean slower but better postprocessing) | |
2185 | 2731 * 0 <= quality <= 6 |
2159 | 2732 */ |
2185 | 2733 int getPpModeForQuality(int quality){ |
2734 int modes[1+GET_PP_QUALITY_MAX]= { | |
2735 0, | |
2736 #if 1 | |
2737 // horizontal filters first | |
2738 LUM_H_DEBLOCK, | |
2739 LUM_H_DEBLOCK | LUM_V_DEBLOCK, | |
2740 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK, | |
2741 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK, | |
2742 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK | LUM_DERING, | |
2743 LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK | LUM_DERING | CHROM_DERING | |
2744 #else | |
2745 // vertical filters first | |
2159 | 2746 LUM_V_DEBLOCK, |
2747 LUM_V_DEBLOCK | LUM_H_DEBLOCK, | |
2748 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK, | |
2749 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK, | |
2750 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING, | |
2751 LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING | CHROM_DERING | |
2185 | 2752 #endif |
2753 }; | |
2159 | 2754 |
2185 | 2755 #ifdef HAVE_ODIVX_POSTPROCESS |
2756 int odivx_modes[1+GET_PP_QUALITY_MAX]= { | |
2757 0, | |
2758 PP_DEBLOCK_Y_H, | |
2759 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V, | |
2760 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H, | |
2761 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V, | |
2762 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V|PP_DERING_Y, | |
2763 PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V|PP_DERING_Y|PP_DERING_C | |
2764 }; | |
2765 if(use_old_pp) return odivx_modes[quality]; | |
2766 #endif | |
2767 return modes[quality]; | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2768 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2769 |
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 * 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
|
2772 * numLines must be a multiple of 4 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2773 * levelFix == 0 -> dont touch the brighness & contrast |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2774 */ |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2775 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
|
2776 int numLines, int levelFix) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2777 { |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2778 #ifndef HAVE_MMX |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2779 int i; |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2780 #endif |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2781 if(levelFix) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2782 { |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2783 #ifdef HAVE_MMX |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2784 asm volatile( |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2785 "leal (%2,%2), %%eax \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2786 "leal (%3,%3), %%ebx \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2787 "movq packedYOffset, %%mm2 \n\t" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2788 "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
|
2789 "pxor %%mm4, %%mm4 \n\t" |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2790 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2791 #define SCALED_CPY \ |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2792 "movq (%0), %%mm0 \n\t"\ |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2793 "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
|
2794 "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
|
2795 "punpckhbw %%mm4, %%mm5 \n\t"\ |
2394 | 2796 "psubw %%mm2, %%mm0 \n\t"\ |
2797 "psubw %%mm2, %%mm5 \n\t"\ | |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2798 "movq (%0,%2), %%mm1 \n\t"\ |
2394 | 2799 "psllw $6, %%mm0 \n\t"\ |
2800 "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
|
2801 "pmulhw %%mm3, %%mm0 \n\t"\ |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2802 "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
|
2803 "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
|
2804 "punpcklbw %%mm4, %%mm1 \n\t"\ |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2805 "punpckhbw %%mm4, %%mm6 \n\t"\ |
2394 | 2806 "psubw %%mm2, %%mm1 \n\t"\ |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2807 "psubw %%mm2, %%mm6 \n\t"\ |
2394 | 2808 "psllw $6, %%mm1 \n\t"\ |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2809 "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
|
2810 "pmulhw %%mm3, %%mm1 \n\t"\ |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2811 "pmulhw %%mm3, %%mm6 \n\t"\ |
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2812 "addl %%eax, %0 \n\t"\ |
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2813 "packuswb %%mm5, %%mm0 \n\t"\ |
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2814 "packuswb %%mm6, %%mm1 \n\t"\ |
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2815 "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
|
2816 "movq %%mm1, (%1, %3) \n\t"\ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2817 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2818 SCALED_CPY |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2819 "addl %%ebx, %1 \n\t" |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2820 SCALED_CPY |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2821 "addl %%ebx, %1 \n\t" |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2822 SCALED_CPY |
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2823 "addl %%ebx, %1 \n\t" |
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2824 SCALED_CPY |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2825 |
2413 | 2826 : "+r"(src), |
2827 "+r"(dst) | |
2828 :"r" (srcStride), | |
2401
bc69d7c0e1dc
brightness / contrast fix/copy optimizations +2% speedup
michael
parents:
2394
diff
changeset
|
2829 "r" (dstStride) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2830 : "%eax", "%ebx" |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2831 ); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2832 #else |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2833 for(i=0; i<numLines; i++) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2834 memcpy( &(dst[dstStride*i]), |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2835 &(src[srcStride*i]), BLOCK_SIZE); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2836 #endif |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2837 } |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2838 else |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2839 { |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2840 #ifdef HAVE_MMX |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2841 asm volatile( |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2842 "movl %4, %%eax \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2843 "movl %%eax, temp0\n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2844 "pushl %0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2845 "pushl %1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2846 "leal (%2,%2), %%eax \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2847 "leal (%3,%3), %%ebx \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2848 "movq packedYOffset, %%mm2 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2849 "movq packedYScale, %%mm3 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2850 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2851 #define SIMPLE_CPY \ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2852 "movq (%0), %%mm0 \n\t"\ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2853 "movq (%0,%2), %%mm1 \n\t"\ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2854 "movq %%mm0, (%1) \n\t"\ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2855 "movq %%mm1, (%1, %3) \n\t"\ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2856 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2857 "1: \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2858 SIMPLE_CPY |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2859 "addl %%eax, %0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2860 "addl %%ebx, %1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2861 SIMPLE_CPY |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2862 "addl %%eax, %0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2863 "addl %%ebx, %1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2864 "decl temp0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2865 "jnz 1b \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2866 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2867 "popl %1 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2868 "popl %0 \n\t" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2869 : : "r" (src), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2870 "r" (dst), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2871 "r" (srcStride), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2872 "r" (dstStride), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2873 "m" (numLines>>2) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2874 : "%eax", "%ebx" |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2875 ); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2876 #else |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2877 for(i=0; i<numLines; i++) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2878 memcpy( &(dst[dstStride*i]), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2879 &(src[srcStride*i]), BLOCK_SIZE); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2880 #endif |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2881 } |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2882 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2883 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2884 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2885 /** |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2886 * Filters array of bytes (Y or U or V values) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2887 */ |
2185 | 2888 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
|
2889 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
|
2890 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2891 int x,y; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2892 /* 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
|
2893 after watching a black picture for 5 hours*/ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2894 static uint64_t *yHistogram= NULL; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2895 int black=0, white=255; // blackest black and whitest white in the picture |
2437 | 2896 int QPCorrecture= 256; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2897 |
2246 | 2898 /* Temporary buffers for handling the last row(s) */ |
2899 static uint8_t *tempDst= NULL; | |
2900 static uint8_t *tempSrc= NULL; | |
2901 | |
2285 | 2902 /* Temporary buffers for handling the last block */ |
2903 static uint8_t *tempDstBlock= NULL; | |
2904 static uint8_t *tempSrcBlock= NULL; | |
2905 | |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2906 #ifdef PP_FUNNY_STRIDE |
2285 | 2907 uint8_t *dstBlockPtrBackup; |
2908 uint8_t *srcBlockPtrBackup; | |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2909 #endif |
2285 | 2910 |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2911 #ifdef MORE_TIMING |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2912 long long T0, T1, diffTime=0; |
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2913 #endif |
2221 | 2914 #ifdef TIMING |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
2915 long long memcpyTime=0, vertTime=0, horizTime=0, sumTime; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2916 sumTime= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2917 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2918 |
2246 | 2919 if(tempDst==NULL) |
2920 { | |
2921 tempDst= (uint8_t*)memalign(8, 1024*24); | |
2922 tempSrc= (uint8_t*)memalign(8, 1024*24); | |
2285 | 2923 tempDstBlock= (uint8_t*)memalign(8, 1024*24); |
2924 tempSrcBlock= (uint8_t*)memalign(8, 1024*24); | |
2246 | 2925 } |
2926 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2927 if(!yHistogram) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2928 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2929 int i; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2930 yHistogram= (uint64_t*)malloc(8*256); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2931 for(i=0; i<256; i++) yHistogram[i]= width*height/64*15/256; |
2285 | 2932 |
2933 if(mode & FULL_Y_RANGE) | |
2934 { | |
2935 maxAllowedY=255; | |
2936 minAllowedY=0; | |
2937 } | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2938 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2939 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2940 if(!isColor) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2941 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2942 uint64_t sum= 0; |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2943 int i; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2944 static int framenum= -1; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2945 uint64_t maxClipped; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2946 uint64_t clipped; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2947 double scale; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2948 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2949 framenum++; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2950 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
|
2951 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2952 for(i=0; i<256; i++) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2953 { |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2954 sum+= yHistogram[i]; |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2955 // printf("%d ", yHistogram[i]); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2956 } |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2957 // printf("\n\n"); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2958 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2959 /* we allways get a completly black picture first */ |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2960 maxClipped= (uint64_t)(sum * maxClippedThreshold); |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2961 |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2962 clipped= sum; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2963 for(black=255; black>0; black--) |
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 if(clipped < maxClipped) break; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2966 clipped-= yHistogram[black]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2967 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2968 |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2969 clipped= sum; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2970 for(white=0; white<256; white++) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2971 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2972 if(clipped < maxClipped) break; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2973 clipped-= yHistogram[white]; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2974 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2975 |
2394 | 2976 packedYOffset= (black - minAllowedY) & 0xFFFF; |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2977 packedYOffset|= packedYOffset<<32; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2978 packedYOffset|= packedYOffset<<16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2979 |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2980 scale= (double)(maxAllowedY - minAllowedY) / (double)(white-black); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2981 |
2394 | 2982 packedYScale= (uint16_t)(scale*1024.0 + 0.5); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2983 packedYScale|= packedYScale<<32; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2984 packedYScale|= packedYScale<<16; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2985 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2986 else |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2987 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2988 packedYScale= 0x0100010001000100LL; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2989 packedYOffset= 0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2990 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
2991 |
2437 | 2992 if(mode & LEVEL_FIX) QPCorrecture= packedYScale &0xFFFF; |
2993 else QPCorrecture= 256; | |
2994 | |
2246 | 2995 /* copy first row of 8x8 blocks */ |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2996 for(x=0; x<width; x+=BLOCK_SIZE) |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
2997 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
|
2998 |
2246 | 2999 for(y=0; y<height; y+=BLOCK_SIZE) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3000 { |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3001 //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
|
3002 uint8_t *srcBlock= &(src[y*srcStride]); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3003 uint8_t *dstBlock= &(dst[y*dstStride]); |
2437 | 3004 #ifdef ARCH_X86 |
3005 int *QPptr= isColor ? &QPs[(y>>3)*QPStride] :&QPs[(y>>4)*QPStride]; | |
3006 int QPDelta= isColor ? 1<<(32-3) : 1<<(32-4); | |
3007 int QPFrac= QPDelta; | |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3008 uint8_t *tempBlock1= tempBlocks; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3009 uint8_t *tempBlock2= tempBlocks + 8; |
2437 | 3010 #endif |
2473 | 3011 /* can we mess with a 8x16 block from srcBlock/dstBlock downwards and 1 line upwards |
3012 if not than use a temporary buffer */ | |
2246 | 3013 if(y+15 >= height) |
3014 { | |
2473 | 3015 /* copy from line 5 to 12 of src, these will be copied with |
2246 | 3016 blockcopy to dst later */ |
3017 memcpy(tempSrc + srcStride*5, srcBlock + srcStride*5, | |
3018 srcStride*MAX(height-y-5, 0) ); | |
3019 | |
3020 /* duplicate last line to fill the void upto line 12 */ | |
3021 if(y+12 >= height) | |
3022 { | |
3023 int i; | |
3024 for(i=height-y; i<=12; i++) | |
3025 memcpy(tempSrc + srcStride*i, | |
3026 src + srcStride*(height-1), srcStride); | |
3027 } | |
3028 | |
3029 | |
2473 | 3030 /* copy up to 6 lines of dst */ |
3031 memcpy(tempDst, dstBlock - dstStride, dstStride*MIN(height-y+1, 6) ); | |
3032 dstBlock= tempDst + dstStride; | |
2246 | 3033 srcBlock= tempSrc; |
3034 } | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3035 |
2285 | 3036 // 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
|
3037 // 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
|
3038 // 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
|
3039 for(x=0; x<width; x+=BLOCK_SIZE) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3040 { |
2168
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3041 const int stride= dstStride; |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3042 uint8_t *tmpXchg; |
2437 | 3043 #ifdef ARCH_X86 |
3044 int QP= *QPptr; | |
3045 asm volatile( | |
3046 "addl %2, %1 \n\t" | |
3047 "sbbl %%eax, %%eax \n\t" | |
3048 "shll $2, %%eax \n\t" | |
3049 "subl %%eax, %0 \n\t" | |
3050 : "+r" (QPptr), "+m" (QPFrac) | |
3051 : "r" (QPDelta) | |
3052 : "%eax" | |
3053 ); | |
3054 #else | |
3055 int QP= isColor ? | |
3056 QPs[(y>>3)*QPStride + (x>>3)]: | |
3057 QPs[(y>>4)*QPStride + (x>>4)]; | |
3058 #endif | |
3059 if(!isColor) | |
2428 | 3060 { |
2437 | 3061 QP= (QP* QPCorrecture)>>8; |
3062 yHistogram[ srcBlock[srcStride*4 + 4] ]++; | |
2428 | 3063 } |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3064 #ifdef HAVE_MMX |
2246 | 3065 asm volatile( |
3066 "movd %0, %%mm7 \n\t" | |
3067 "packuswb %%mm7, %%mm7 \n\t" // 0, 0, 0, QP, 0, 0, 0, QP | |
3068 "packuswb %%mm7, %%mm7 \n\t" // 0,QP, 0, QP, 0,QP, 0, QP | |
3069 "packuswb %%mm7, %%mm7 \n\t" // QP,..., QP | |
3070 "movq %%mm7, pQPb \n\t" | |
3071 : : "r" (QP) | |
3072 ); | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3073 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3074 |
2221 | 3075 #ifdef MORE_TIMING |
2246 | 3076 T0= rdtsc(); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3077 #endif |
2159 | 3078 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3079 #ifdef HAVE_MMX2 |
2437 | 3080 /* |
2246 | 3081 prefetchnta(srcBlock + (((x>>3)&3) + 5)*srcStride + 32); |
3082 prefetchnta(srcBlock + (((x>>3)&3) + 9)*srcStride + 32); | |
3083 prefetcht0(dstBlock + (((x>>3)&3) + 5)*dstStride + 32); | |
3084 prefetcht0(dstBlock + (((x>>3)&3) + 9)*dstStride + 32); | |
2437 | 3085 */ |
3086 /* | |
3087 prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32); | |
3088 prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32); | |
3089 prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32); | |
3090 prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32); | |
3091 */ | |
3092 | |
3093 asm( | |
3094 "movl %4, %%eax \n\t" | |
3095 "shrl $2, %%eax \n\t" | |
3096 "andl $6, %%eax \n\t" | |
3097 "addl $5, %%eax \n\t" | |
3098 "movl %%eax, %%ebx \n\t" | |
3099 "imul %1, %%eax \n\t" | |
3100 "imul %3, %%ebx \n\t" | |
3101 "prefetchnta 32(%%eax, %0) \n\t" | |
3102 "prefetcht0 32(%%ebx, %2) \n\t" | |
3103 "addl %1, %%eax \n\t" | |
3104 "addl %3, %%ebx \n\t" | |
3105 "prefetchnta 32(%%eax, %0) \n\t" | |
3106 "prefetcht0 32(%%ebx, %2) \n\t" | |
3107 :: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride), | |
3108 "m" (x) | |
3109 : "%eax", "%ebx" | |
3110 ); | |
3111 | |
2159 | 3112 #elif defined(HAVE_3DNOW) |
3113 //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ... | |
2246 | 3114 /* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32); |
3115 prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32); | |
3116 prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32); | |
3117 prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32); | |
2159 | 3118 */ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3119 #endif |
2246 | 3120 |
2286 | 3121 #ifdef PP_FUNNY_STRIDE |
2285 | 3122 //can we mess with a 8x16 block, if not use a temp buffer, yes again |
3123 if(x+7 >= width) | |
3124 { | |
3125 int i; | |
3126 dstBlockPtrBackup= dstBlock; | |
3127 srcBlockPtrBackup= srcBlock; | |
3128 | |
3129 for(i=0;i<BLOCK_SIZE*2; i++) | |
3130 { | |
3131 memcpy(tempSrcBlock+i*srcStride, srcBlock+i*srcStride, width-x); | |
3132 memcpy(tempDstBlock+i*dstStride, dstBlock+i*dstStride, width-x); | |
3133 } | |
3134 | |
3135 dstBlock= tempDstBlock; | |
3136 srcBlock= tempSrcBlock; | |
3137 } | |
2286 | 3138 #endif |
2285 | 3139 |
2246 | 3140 blockCopy(dstBlock + dstStride*5, dstStride, |
3141 srcBlock + srcStride*5, srcStride, 8, mode & LEVEL_FIX); | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3142 |
2246 | 3143 if(mode & LINEAR_IPOL_DEINT_FILTER) |
3144 deInterlaceInterpolateLinear(dstBlock, dstStride); | |
3145 else if(mode & LINEAR_BLEND_DEINT_FILTER) | |
3146 deInterlaceBlendLinear(dstBlock, dstStride); | |
3147 else if(mode & MEDIAN_DEINT_FILTER) | |
3148 deInterlaceMedian(dstBlock, dstStride); | |
3149 else if(mode & CUBIC_IPOL_DEINT_FILTER) | |
3150 deInterlaceInterpolateCubic(dstBlock, dstStride); | |
3151 /* else if(mode & CUBIC_BLEND_DEINT_FILTER) | |
3152 deInterlaceBlendCubic(dstBlock, dstStride); | |
2203
f90b6e259dc8
rewrote the horizontal lowpass filter to fix a bug which caused a blocky look
michael
parents:
2195
diff
changeset
|
3153 */ |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3154 |
2246 | 3155 /* only deblock if we have 2 blocks */ |
3156 if(y + 8 < height) | |
3157 { | |
2221 | 3158 #ifdef MORE_TIMING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3159 T1= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3160 memcpyTime+= T1-T0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3161 T0=T1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3162 #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
|
3163 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
|
3164 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
|
3165 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
|
3166 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
|
3167 else if(mode & V_DEBLOCK) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3168 { |
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
|
3169 if( isVertDC(dstBlock, stride)) |
2159 | 3170 { |
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
|
3171 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
|
3172 doVertLowPass(dstBlock, stride, QP); |
2159 | 3173 } |
2300
e10f7dc4938f
more logic behavior if the altenative deblock filters are used (turning a alt filter on without turning the deblock filter on uses the alt filter instead of using no filter now)
michael
parents:
2286
diff
changeset
|
3174 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
|
3175 doVertDefFilter(dstBlock, stride, QP); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3176 } |
2221 | 3177 #ifdef MORE_TIMING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3178 T1= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3179 vertTime+= T1-T0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3180 T0=T1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3181 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3182 } |
2473 | 3183 |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3184 #ifdef HAVE_MMX |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3185 transpose1(tempBlock1, tempBlock2, dstBlock, dstStride); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3186 #endif |
2246 | 3187 /* check if we have a previous block to deblock it with dstBlock */ |
2285 | 3188 if(x - 8 >= 0) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3189 { |
2221 | 3190 #ifdef MORE_TIMING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3191 T0= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3192 #endif |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3193 #ifdef HAVE_MMX |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3194 if(mode & H_RK1_FILTER) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3195 vertRK1Filter(tempBlock1, 16, QP); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3196 else if(mode & H_X1_FILTER) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3197 vertX1Filter(tempBlock1, 16, QP); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3198 else if(mode & H_DEBLOCK) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3199 { |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3200 if( isVertDC(tempBlock1, 16)) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3201 { |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3202 if(isVertMinMaxOk(tempBlock1, 16, QP)) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3203 doVertLowPass(tempBlock1, 16, QP); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3204 } |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3205 else |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3206 doVertDefFilter(tempBlock1, 16, QP); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3207 } |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3208 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3209 transpose2(dstBlock-4, dstStride, tempBlock1 + 4*16); |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3210 |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3211 #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
|
3212 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
|
3213 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
|
3214 else if(mode & H_DEBLOCK) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3215 { |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3216 if( isHorizDC(dstBlock-4, stride)) |
2159 | 3217 { |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3218 if(isHorizMinMaxOk(dstBlock-4, stride, QP)) |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3219 doHorizLowPass(dstBlock-4, stride, QP); |
2159 | 3220 } |
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
|
3221 else |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3222 doHorizDefFilter(dstBlock-4, stride, QP); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3223 } |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3224 #endif |
2221 | 3225 #ifdef MORE_TIMING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3226 T1= rdtsc(); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3227 horizTime+= T1-T0; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3228 T0=T1; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3229 #endif |
2473 | 3230 if(mode & DERING) |
3231 { | |
3232 //FIXME filter first line | |
3233 if(y>0) dering(dstBlock - stride - 8, stride, QP); | |
3234 } | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3235 } |
2473 | 3236 else if(mode & DERING) |
3237 { | |
3238 //FIXME y+15 is required cuz of the tempBuffer thing -> bottom right block isnt filtered | |
3239 if(y > 8 && y+15 < height) dering(dstBlock - stride*9 + width - 8, stride, QP); | |
3240 } | |
3241 | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3242 |
2286 | 3243 #ifdef PP_FUNNY_STRIDE |
2285 | 3244 /* did we use a tmp-block buffer */ |
3245 if(x+7 >= width) | |
3246 { | |
3247 int i; | |
3248 dstBlock= dstBlockPtrBackup; | |
3249 srcBlock= srcBlockPtrBackup; | |
3250 | |
3251 for(i=0;i<BLOCK_SIZE*2; i++) | |
3252 { | |
3253 memcpy(dstBlock+i*dstStride, tempDstBlock+i*dstStride, width-x); | |
3254 } | |
3255 } | |
2286 | 3256 #endif |
2285 | 3257 |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3258 dstBlock+=8; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3259 srcBlock+=8; |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3260 |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
3261 #ifdef HAVE_MMX |
2454
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3262 tmpXchg= tempBlock1; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3263 tempBlock1= tempBlock2; |
b74c2a08eac9
much better horizontal filters (transpose & use the vertical ones) :)
michael
parents:
2437
diff
changeset
|
3264 tempBlock2 = tmpXchg; |
2461
60f16575bece
fixed difference with -vo md5 between doVertDefFilter() C and MMX / MMX2 versions
michael
parents:
2454
diff
changeset
|
3265 #endif |
2246 | 3266 } |
3267 | |
3268 /* did we use a tmp buffer */ | |
2285 | 3269 if(y+15 >= height) |
2246 | 3270 { |
3271 uint8_t *dstBlock= &(dst[y*dstStride]); | |
2473 | 3272 memcpy(dstBlock, tempDst + dstStride, dstStride*(height-y) ); |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3273 } |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3274 } |
2159 | 3275 #ifdef HAVE_3DNOW |
3276 asm volatile("femms"); | |
3277 #elif defined (HAVE_MMX) | |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3278 asm volatile("emms"); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3279 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3280 |
2221 | 3281 #ifdef TIMING |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3282 // 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
|
3283 sumTime= rdtsc() - sumTime; |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3284 if(!isColor) |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3285 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
|
3286 (int)(memcpyTime/1000), (int)(vertTime/1000), (int)(horizTime/1000), |
21a8f158d19f
bugfixes: last 3 lines not brightness/contrast corrected
michael
parents:
2159
diff
changeset
|
3287 (int)(sumTime/1000), (int)((sumTime-memcpyTime-vertTime-horizTime)/1000) |
2158
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3288 , black, white); |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3289 #endif |
508468a75be0
new postprocess code by Michael Niedermayer (michaelni@gmx.at)
arpi
parents:
diff
changeset
|
3290 } |